fs/btrfs/transaction.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/fs.h>
   7 #include <linux/slab.h>
   8 #include <linux/sched.h>
   9 #include <linux/writeback.h>
  10 #include <linux/pagemap.h>
  11 #include <linux/blkdev.h>
  12 #include <linux/uuid.h>
  13 #include "misc.h"
  14 #include "ctree.h"
  15 #include "disk-io.h"
  16 #include "transaction.h"
  17 #include "locking.h"
  18 #include "tree-log.h"
  19 #include "volumes.h"
  20 #include "dev-replace.h"
  21 #include "qgroup.h"
  22 #include "block-group.h"
  23 #include "space-info.h"
  24 #include "zoned.h"
  25
  26 #define BTRFS_ROOT_TRANS_TAG 0
  27
  28 /*
  29  * Transaction states and transitions
  30  *
  31  * No running transaction (fs tree blocks are not modified)
  32  * |
  33  * | To next stage:
  34  * |  Call start_transaction() variants. Except btrfs_join_transaction_nostart().
  35  * V
  36  * Transaction N [[TRANS_STATE_RUNNING]]
  37  * |
  38  * | New trans handles can be attached to transaction N by calling all
  39  * | start_transaction() variants.
  40  * |
  41  * | To next stage:
  42  * |  Call btrfs_commit_transaction() on any trans handle attached to
  43  * |  transaction N
  44  * V
  45  * Transaction N [[TRANS_STATE_COMMIT_START]]
  46  * |
  47  * | Will wait for previous running transaction to completely finish if there
  48  * | is one
  49  * |
  50  * | Then one of the following happes:
  51  * | - Wait for all other trans handle holders to release.
  52  * |   The btrfs_commit_transaction() caller will do the commit work.
  53  * | - Wait for current transaction to be committed by others.
  54  * |   Other btrfs_commit_transaction() caller will do the commit work.
  55  * |
  56  * | At this stage, only btrfs_join_transaction*() variants can attach
  57  * | to this running transaction.
  58  * | All other variants will wait for current one to finish and attach to
  59  * | transaction N+1.
  60  * |
  61  * | To next stage:
  62  * |  Caller is chosen to commit transaction N, and all other trans handle
  63  * |  haven been released.
  64  * V
  65  * Transaction N [[TRANS_STATE_COMMIT_DOING]]
  66  * |
  67  * | The heavy lifting transaction work is started.
  68  * | From running delayed refs (modifying extent tree) to creating pending
  69  * | snapshots, running qgroups.
  70  * | In short, modify supporting trees to reflect modifications of subvolume
  71  * | trees.
  72  * |
  73  * | At this stage, all start_transaction() calls will wait for this
  74  * | transaction to finish and attach to transaction N+1.
  75  * |
  76  * | To next stage:
  77  * |  Until all supporting trees are updated.
  78  * V
  79  * Transaction N [[TRANS_STATE_UNBLOCKED]]
  80  * |                                                Transaction N+1
  81  * | All needed trees are modified, thus we only    [[TRANS_STATE_RUNNING]]
  82  * | need to write them back to disk and update     |
  83  * | super blocks.                                  |
  84  * |                                                |
  85  * | At this stage, new transaction is allowed to   |
  86  * | start.                                         |
  87  * | All new start_transaction() calls will be      |
  88  * | attached to transid N+1.                       |
  89  * |                                                |
  90  * | To next stage:                                 |
  91  * |  Until all tree blocks are super blocks are    |
  92  * |  written to block devices                      |
  93  * V                                                |
  94  * Transaction N [[TRANS_STATE_COMPLETED]]          V
  95  *   All tree blocks and super blocks are written.  Transaction N+1
  96  *   This transaction is finished and all its       [[TRANS_STATE_COMMIT_START]]
  97  *   data structures will be cleaned up.            | Life goes on
  98  */
  99 static const unsigned int btrfs_blocked_trans_types[TRANS_STATE_MAX] = {
 100         [TRANS_STATE_RUNNING]           = 0U,
 101         [TRANS_STATE_COMMIT_START]      = (__TRANS_START | __TRANS_ATTACH),
 102         [TRANS_STATE_COMMIT_DOING]      = (__TRANS_START |
 103                                            __TRANS_ATTACH |
 104                                            __TRANS_JOIN |
 105                                            __TRANS_JOIN_NOSTART),
 106         [TRANS_STATE_UNBLOCKED]         = (__TRANS_START |
 107                                            __TRANS_ATTACH |
 108                                            __TRANS_JOIN |
 109                                            __TRANS_JOIN_NOLOCK |
 110                                            __TRANS_JOIN_NOSTART),
 111         [TRANS_STATE_SUPER_COMMITTED]   = (__TRANS_START |
 112                                            __TRANS_ATTACH |
 113                                            __TRANS_JOIN |
 114                                            __TRANS_JOIN_NOLOCK |
 115                                            __TRANS_JOIN_NOSTART),
 116         [TRANS_STATE_COMPLETED]         = (__TRANS_START |
 117                                            __TRANS_ATTACH |
 118                                            __TRANS_JOIN |
 119                                            __TRANS_JOIN_NOLOCK |
 120                                            __TRANS_JOIN_NOSTART),
 121 };
 122
 123 void btrfs_put_transaction(struct btrfs_transaction *transaction)
 124 {
 125         WARN_ON(refcount_read(&transaction->use_count) == 0);
 126         if (refcount_dec_and_test(&transaction->use_count)) {
 127                 BUG_ON(!list_empty(&transaction->list));
 128                 WARN_ON(!RB_EMPTY_ROOT(
 129                                 &transaction->delayed_refs.href_root.rb_root));
 130                 WARN_ON(!RB_EMPTY_ROOT(
 131                                 &transaction->delayed_refs.dirty_extent_root));
 132                 if (transaction->delayed_refs.pending_csums)
 133                         btrfs_err(transaction->fs_info,
 134                                   "pending csums is %llu",
 135                                   transaction->delayed_refs.pending_csums);
 136                 /*
 137                  * If any block groups are found in ->deleted_bgs then it's
 138                  * because the transaction was aborted and a commit did not
 139                  * happen (things failed before writing the new superblock
 140                  * and calling btrfs_finish_extent_commit()), so we can not
 141                  * discard the physical locations of the block groups.
 142                  */
 143                 while (!list_empty(&transaction->deleted_bgs)) {
 144                         struct btrfs_block_group *cache;
 145
 146                         cache = list_first_entry(&transaction->deleted_bgs,
 147                                                  struct btrfs_block_group,
 148                                                  bg_list);
 149                         list_del_init(&cache->bg_list);
 150                         btrfs_unfreeze_block_group(cache);
 151                         btrfs_put_block_group(cache);
 152                 }
 153                 WARN_ON(!list_empty(&transaction->dev_update_list));
 154                 kfree(transaction);
 155         }
 156 }
 157
 158 static noinline void switch_commit_roots(struct btrfs_trans_handle *trans)
 159 {
 160         struct btrfs_transaction *cur_trans = trans->transaction;
 161         struct btrfs_fs_info *fs_info = trans->fs_info;
 162         struct btrfs_root *root, *tmp;
 163         struct btrfs_caching_control *caching_ctl, *next;
 164
 165         down_write(&fs_info->commit_root_sem);
 166         list_for_each_entry_safe(root, tmp, &cur_trans->switch_commits,
 167                                  dirty_list) {
 168                 list_del_init(&root->dirty_list);
 169                 free_extent_buffer(root->commit_root);
 170                 root->commit_root = btrfs_root_node(root);
 171                 extent_io_tree_release(&root->dirty_log_pages);
 172                 btrfs_qgroup_clean_swapped_blocks(root);
 173         }
 174
 175         /* We can free old roots now. */
 176         spin_lock(&cur_trans->dropped_roots_lock);
 177         while (!list_empty(&cur_trans->dropped_roots)) {
 178                 root = list_first_entry(&cur_trans->dropped_roots,
 179                                         struct btrfs_root, root_list);
 180                 list_del_init(&root->root_list);
 181                 spin_unlock(&cur_trans->dropped_roots_lock);
 182                 btrfs_free_log(trans, root);
 183                 btrfs_drop_and_free_fs_root(fs_info, root);
 184                 spin_lock(&cur_trans->dropped_roots_lock);
 185         }
 186         spin_unlock(&cur_trans->dropped_roots_lock);
 187
 188         /*
 189          * We have to update the last_byte_to_unpin under the commit_root_sem,
 190          * at the same time we swap out the commit roots.
 191          *
 192          * This is because we must have a real view of the last spot the caching
 193          * kthreads were while caching.  Consider the following views of the
 194          * extent tree for a block group
 195          *
 196          * commit root
 197          * +----+----+----+----+----+----+----+
 198          * |\\\\|    |\\\\|\\\\|    |\\\\|\\\\|
 199          * +----+----+----+----+----+----+----+
 200          * 0    1    2    3    4    5    6    7
 201          *
 202          * new commit root
 203          * +----+----+----+----+----+----+----+
 204          * |    |    |    |\\\\|    |    |\\\\|
 205          * +----+----+----+----+----+----+----+
 206          * 0    1    2    3    4    5    6    7
 207          *
 208          * If the cache_ctl->progress was at 3, then we are only allowed to
 209          * unpin [0,1) and [2,3], because the caching thread has already
 210          * processed those extents.  We are not allowed to unpin [5,6), because
 211          * the caching thread will re-start it's search from 3, and thus find
 212          * the hole from [4,6) to add to the free space cache.
 213          */
 214         spin_lock(&fs_info->block_group_cache_lock);
 215         list_for_each_entry_safe(caching_ctl, next,
 216                                  &fs_info->caching_block_groups, list) {
 217                 struct btrfs_block_group *cache = caching_ctl->block_group;
 218
 219                 if (btrfs_block_group_done(cache)) {
 220                         cache->last_byte_to_unpin = (u64)-1;
 221                         list_del_init(&caching_ctl->list);
 222                         btrfs_put_caching_control(caching_ctl);
 223                 } else {
 224                         cache->last_byte_to_unpin = caching_ctl->progress;
 225                 }
 226         }
 227         spin_unlock(&fs_info->block_group_cache_lock);
 228         up_write(&fs_info->commit_root_sem);
 229 }
 230
 231 static inline void extwriter_counter_inc(struct btrfs_transaction *trans,
 232                                          unsigned int type)
 233 {
 234         if (type & TRANS_EXTWRITERS)
 235                 atomic_inc(&trans->num_extwriters);
 236 }
 237
 238 static inline void extwriter_counter_dec(struct btrfs_transaction *trans,
 239                                          unsigned int type)
 240 {
 241         if (type & TRANS_EXTWRITERS)
 242                 atomic_dec(&trans->num_extwriters);
 243 }
 244
 245 static inline void extwriter_counter_init(struct btrfs_transaction *trans,
 246                                           unsigned int type)
 247 {
 248         atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0));
 249 }
 250
 251 static inline int extwriter_counter_read(struct btrfs_transaction *trans)
 252 {
 253         return atomic_read(&trans->num_extwriters);
 254 }
 255
 256 /*
 257  * To be called after all the new block groups attached to the transaction
 258  * handle have been created (btrfs_create_pending_block_groups()).
 259  */
 260 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
 261 {
 262         struct btrfs_fs_info *fs_info = trans->fs_info;
 263         struct btrfs_transaction *cur_trans = trans->transaction;
 264
 265         if (!trans->chunk_bytes_reserved)
 266                 return;
 267
 268         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
 269
 270         btrfs_block_rsv_release(fs_info, &fs_info->chunk_block_rsv,
 271                                 trans->chunk_bytes_reserved, NULL);
 272         atomic64_sub(trans->chunk_bytes_reserved, &cur_trans->chunk_bytes_reserved);
 273         cond_wake_up(&cur_trans->chunk_reserve_wait);
 274         trans->chunk_bytes_reserved = 0;
 275 }
 276
 277 /*
 278  * either allocate a new transaction or hop into the existing one
 279  */
 280 static noinline int join_transaction(struct btrfs_fs_info *fs_info,
 281                                      unsigned int type)
 282 {
 283         struct btrfs_transaction *cur_trans;
 284
 285         spin_lock(&fs_info->trans_lock);
 286 loop:
 287         /* The file system has been taken offline. No new transactions. */
 288         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
 289                 spin_unlock(&fs_info->trans_lock);
 290                 return -EROFS;
 291         }
 292
 293         cur_trans = fs_info->running_transaction;
 294         if (cur_trans) {
 295                 if (TRANS_ABORTED(cur_trans)) {
 296                         spin_unlock(&fs_info->trans_lock);
 297                         return cur_trans->aborted;
 298                 }
 299                 if (btrfs_blocked_trans_types[cur_trans->state] & type) {
 300                         spin_unlock(&fs_info->trans_lock);
 301                         return -EBUSY;
 302                 }
 303                 refcount_inc(&cur_trans->use_count);
 304                 atomic_inc(&cur_trans->num_writers);
 305                 extwriter_counter_inc(cur_trans, type);
 306                 spin_unlock(&fs_info->trans_lock);
 307                 return 0;
 308         }
 309         spin_unlock(&fs_info->trans_lock);
 310
 311         /*
 312          * If we are ATTACH, we just want to catch the current transaction,
 313          * and commit it. If there is no transaction, just return ENOENT.
 314          */
 315         if (type == TRANS_ATTACH)
 316                 return -ENOENT;
 317
 318         /*
 319          * JOIN_NOLOCK only happens during the transaction commit, so
 320          * it is impossible that ->running_transaction is NULL
 321          */
 322         BUG_ON(type == TRANS_JOIN_NOLOCK);
 323
 324         cur_trans = kmalloc(sizeof(*cur_trans), GFP_NOFS);
 325         if (!cur_trans)
 326                 return -ENOMEM;
 327
 328         spin_lock(&fs_info->trans_lock);
 329         if (fs_info->running_transaction) {
 330                 /*
 331                  * someone started a transaction after we unlocked.  Make sure
 332                  * to redo the checks above
 333                  */
 334                 kfree(cur_trans);
 335                 goto loop;
 336         } else if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
 337                 spin_unlock(&fs_info->trans_lock);
 338                 kfree(cur_trans);
 339                 return -EROFS;
 340         }
 341
 342         cur_trans->fs_info = fs_info;
 343         atomic_set(&cur_trans->pending_ordered, 0);
 344         init_waitqueue_head(&cur_trans->pending_wait);
 345         atomic_set(&cur_trans->num_writers, 1);
 346         extwriter_counter_init(cur_trans, type);
 347         init_waitqueue_head(&cur_trans->writer_wait);
 348         init_waitqueue_head(&cur_trans->commit_wait);
 349         cur_trans->state = TRANS_STATE_RUNNING;
 350         /*
 351          * One for this trans handle, one so it will live on until we
 352          * commit the transaction.
 353          */
 354         refcount_set(&cur_trans->use_count, 2);
 355         cur_trans->flags = 0;
 356         cur_trans->start_time = ktime_get_seconds();
 357
 358         memset(&cur_trans->delayed_refs, 0, sizeof(cur_trans->delayed_refs));
 359
 360         cur_trans->delayed_refs.href_root = RB_ROOT_CACHED;
 361         cur_trans->delayed_refs.dirty_extent_root = RB_ROOT;
 362         atomic_set(&cur_trans->delayed_refs.num_entries, 0);
 363
 364         /*
 365          * although the tree mod log is per file system and not per transaction,
 366          * the log must never go across transaction boundaries.
 367          */
 368         smp_mb();
 369         if (!list_empty(&fs_info->tree_mod_seq_list))
 370                 WARN(1, KERN_ERR "BTRFS: tree_mod_seq_list not empty when creating a fresh transaction\n");
 371         if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
 372                 WARN(1, KERN_ERR "BTRFS: tree_mod_log rb tree not empty when creating a fresh transaction\n");
 373         atomic64_set(&fs_info->tree_mod_seq, 0);
 374
 375         spin_lock_init(&cur_trans->delayed_refs.lock);
 376
 377         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 378         INIT_LIST_HEAD(&cur_trans->dev_update_list);
 379         INIT_LIST_HEAD(&cur_trans->switch_commits);
 380         INIT_LIST_HEAD(&cur_trans->dirty_bgs);
 381         INIT_LIST_HEAD(&cur_trans->io_bgs);
 382         INIT_LIST_HEAD(&cur_trans->dropped_roots);
 383         mutex_init(&cur_trans->cache_write_mutex);
 384         spin_lock_init(&cur_trans->dirty_bgs_lock);
 385         INIT_LIST_HEAD(&cur_trans->deleted_bgs);
 386         spin_lock_init(&cur_trans->dropped_roots_lock);
 387         INIT_LIST_HEAD(&cur_trans->releasing_ebs);
 388         spin_lock_init(&cur_trans->releasing_ebs_lock);
 389         atomic64_set(&cur_trans->chunk_bytes_reserved, 0);
 390         init_waitqueue_head(&cur_trans->chunk_reserve_wait);
 391         list_add_tail(&cur_trans->list, &fs_info->trans_list);
 392         extent_io_tree_init(fs_info, &cur_trans->dirty_pages,
 393                         IO_TREE_TRANS_DIRTY_PAGES, fs_info->btree_inode);
 394         extent_io_tree_init(fs_info, &cur_trans->pinned_extents,
 395                         IO_TREE_FS_PINNED_EXTENTS, NULL);
 396         fs_info->generation++;
 397         cur_trans->transid = fs_info->generation;
 398         fs_info->running_transaction = cur_trans;
 399         cur_trans->aborted = 0;
 400         spin_unlock(&fs_info->trans_lock);
 401
 402         return 0;
 403 }
 404
 405 /*
 406  * This does all the record keeping required to make sure that a shareable root
 407  * is properly recorded in a given transaction.  This is required to make sure
 408  * the old root from before we joined the transaction is deleted when the
 409  * transaction commits.
 410  */
 411 static int record_root_in_trans(struct btrfs_trans_handle *trans,
 412                                struct btrfs_root *root,
 413                                int force)
 414 {
 415         struct btrfs_fs_info *fs_info = root->fs_info;
 416         int ret = 0;
 417
 418         if ((test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
 419             root->last_trans < trans->transid) || force) {
 420                 WARN_ON(root == fs_info->extent_root);
 421                 WARN_ON(!force && root->commit_root != root->node);
 422
 423                 /*
 424                  * see below for IN_TRANS_SETUP usage rules
 425                  * we have the reloc mutex held now, so there
 426                  * is only one writer in this function
 427                  */
 428                 set_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
 429
 430                 /* make sure readers find IN_TRANS_SETUP before
 431                  * they find our root->last_trans update
 432                  */
 433                 smp_wmb();
 434
 435                 spin_lock(&fs_info->fs_roots_radix_lock);
 436                 if (root->last_trans == trans->transid && !force) {
 437                         spin_unlock(&fs_info->fs_roots_radix_lock);
 438                         return 0;
 439                 }
 440                 radix_tree_tag_set(&fs_info->fs_roots_radix,
 441                                    (unsigned long)root->root_key.objectid,
 442                                    BTRFS_ROOT_TRANS_TAG);
 443                 spin_unlock(&fs_info->fs_roots_radix_lock);
 444                 root->last_trans = trans->transid;
 445
 446                 /* this is pretty tricky.  We don't want to
 447                  * take the relocation lock in btrfs_record_root_in_trans
 448                  * unless we're really doing the first setup for this root in
 449                  * this transaction.
 450                  *
 451                  * Normally we'd use root->last_trans as a flag to decide
 452                  * if we want to take the expensive mutex.
 453                  *
 454                  * But, we have to set root->last_trans before we
 455                  * init the relocation root, otherwise, we trip over warnings
 456                  * in ctree.c.  The solution used here is to flag ourselves
 457                  * with root IN_TRANS_SETUP.  When this is 1, we're still
 458                  * fixing up the reloc trees and everyone must wait.
 459                  *
 460                  * When this is zero, they can trust root->last_trans and fly
 461                  * through btrfs_record_root_in_trans without having to take the
 462                  * lock.  smp_wmb() makes sure that all the writes above are
 463                  * done before we pop in the zero below
 464                  */
 465                 ret = btrfs_init_reloc_root(trans, root);
 466                 smp_mb__before_atomic();
 467                 clear_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
 468         }
 469         return ret;
 470 }
 471
 472
 473 void btrfs_add_dropped_root(struct btrfs_trans_handle *trans,
 474                             struct btrfs_root *root)
 475 {
 476         struct btrfs_fs_info *fs_info = root->fs_info;
 477         struct btrfs_transaction *cur_trans = trans->transaction;
 478
 479         /* Add ourselves to the transaction dropped list */
 480         spin_lock(&cur_trans->dropped_roots_lock);
 481         list_add_tail(&root->root_list, &cur_trans->dropped_roots);
 482         spin_unlock(&cur_trans->dropped_roots_lock);
 483
 484         /* Make sure we don't try to update the root at commit time */
 485         spin_lock(&fs_info->fs_roots_radix_lock);
 486         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 487                              (unsigned long)root->root_key.objectid,
 488                              BTRFS_ROOT_TRANS_TAG);
 489         spin_unlock(&fs_info->fs_roots_radix_lock);
 490 }
 491
 492 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 493                                struct btrfs_root *root)
 494 {
 495         struct btrfs_fs_info *fs_info = root->fs_info;
 496         int ret;
 497
 498         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
 499                 return 0;
 500
 501         /*
 502          * see record_root_in_trans for comments about IN_TRANS_SETUP usage
 503          * and barriers
 504          */
 505         smp_rmb();
 506         if (root->last_trans == trans->transid &&
 507             !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state))
 508                 return 0;
 509
 510         mutex_lock(&fs_info->reloc_mutex);
 511         ret = record_root_in_trans(trans, root, 0);
 512         mutex_unlock(&fs_info->reloc_mutex);
 513
 514         return ret;
 515 }
 516
 517 static inline int is_transaction_blocked(struct btrfs_transaction *trans)
 518 {
 519         return (trans->state >= TRANS_STATE_COMMIT_START &&
 520                 trans->state < TRANS_STATE_UNBLOCKED &&
 521                 !TRANS_ABORTED(trans));
 522 }
 523
 524 /* wait for commit against the current transaction to become unblocked
 525  * when this is done, it is safe to start a new transaction, but the current
 526  * transaction might not be fully on disk.
 527  */
 528 static void wait_current_trans(struct btrfs_fs_info *fs_info)
 529 {
 530         struct btrfs_transaction *cur_trans;
 531
 532         spin_lock(&fs_info->trans_lock);
 533         cur_trans = fs_info->running_transaction;
 534         if (cur_trans && is_transaction_blocked(cur_trans)) {
 535                 refcount_inc(&cur_trans->use_count);
 536                 spin_unlock(&fs_info->trans_lock);
 537
 538                 wait_event(fs_info->transaction_wait,
 539                            cur_trans->state >= TRANS_STATE_UNBLOCKED ||
 540                            TRANS_ABORTED(cur_trans));
 541                 btrfs_put_transaction(cur_trans);
 542         } else {
 543                 spin_unlock(&fs_info->trans_lock);
 544         }
 545 }
 546
 547 static int may_wait_transaction(struct btrfs_fs_info *fs_info, int type)
 548 {
 549         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
 550                 return 0;
 551
 552         if (type == TRANS_START)
 553                 return 1;
 554
 555         return 0;
 556 }
 557
 558 static inline bool need_reserve_reloc_root(struct btrfs_root *root)
 559 {
 560         struct btrfs_fs_info *fs_info = root->fs_info;
 561
 562         if (!fs_info->reloc_ctl ||
 563             !test_bit(BTRFS_ROOT_SHAREABLE, &root->state) ||
 564             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 565             root->reloc_root)
 566                 return false;
 567
 568         return true;
 569 }
 570
 571 static struct btrfs_trans_handle *
 572 start_transaction(struct btrfs_root *root, unsigned int num_items,
 573                   unsigned int type, enum btrfs_reserve_flush_enum flush,
 574                   bool enforce_qgroups)
 575 {
 576         struct btrfs_fs_info *fs_info = root->fs_info;
 577         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
 578         struct btrfs_trans_handle *h;
 579         struct btrfs_transaction *cur_trans;
 580         u64 num_bytes = 0;
 581         u64 qgroup_reserved = 0;
 582         bool reloc_reserved = false;
 583         bool do_chunk_alloc = false;
 584         int ret;
 585
 586         /* Send isn't supposed to start transactions. */
 587         ASSERT(current->journal_info != BTRFS_SEND_TRANS_STUB);
 588
 589         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
 590                 return ERR_PTR(-EROFS);
 591
 592         if (current->journal_info) {
 593                 WARN_ON(type & TRANS_EXTWRITERS);
 594                 h = current->journal_info;
 595                 refcount_inc(&h->use_count);
 596                 WARN_ON(refcount_read(&h->use_count) > 2);
 597                 h->orig_rsv = h->block_rsv;
 598                 h->block_rsv = NULL;
 599                 goto got_it;
 600         }
 601
 602         /*
 603          * Do the reservation before we join the transaction so we can do all
 604          * the appropriate flushing if need be.
 605          */
 606         if (num_items && root != fs_info->chunk_root) {
 607                 struct btrfs_block_rsv *rsv = &fs_info->trans_block_rsv;
 608                 u64 delayed_refs_bytes = 0;
 609
 610                 qgroup_reserved = num_items * fs_info->nodesize;
 611                 ret = btrfs_qgroup_reserve_meta_pertrans(root, qgroup_reserved,
 612                                 enforce_qgroups);
 613                 if (ret)
 614                         return ERR_PTR(ret);
 615
 616                 /*
 617                  * We want to reserve all the bytes we may need all at once, so
 618                  * we only do 1 enospc flushing cycle per transaction start.  We
 619                  * accomplish this by simply assuming we'll do 2 x num_items
 620                  * worth of delayed refs updates in this trans handle, and
 621                  * refill that amount for whatever is missing in the reserve.
 622                  */
 623                 num_bytes = btrfs_calc_insert_metadata_size(fs_info, num_items);
 624                 if (flush == BTRFS_RESERVE_FLUSH_ALL &&
 625                     delayed_refs_rsv->full == 0) {
 626                         delayed_refs_bytes = num_bytes;
 627                         num_bytes <<= 1;
 628                 }
 629
 630                 /*
 631                  * Do the reservation for the relocation root creation
 632                  */
 633                 if (need_reserve_reloc_root(root)) {
 634                         num_bytes += fs_info->nodesize;
 635                         reloc_reserved = true;
 636                 }
 637
 638                 ret = btrfs_block_rsv_add(root, rsv, num_bytes, flush);
 639                 if (ret)
 640                         goto reserve_fail;
 641                 if (delayed_refs_bytes) {
 642                         btrfs_migrate_to_delayed_refs_rsv(fs_info, rsv,
 643                                                           delayed_refs_bytes);
 644                         num_bytes -= delayed_refs_bytes;
 645                 }
 646
 647                 if (rsv->space_info->force_alloc)
 648                         do_chunk_alloc = true;
 649         } else if (num_items == 0 && flush == BTRFS_RESERVE_FLUSH_ALL &&
 650                    !delayed_refs_rsv->full) {
 651                 /*
 652                  * Some people call with btrfs_start_transaction(root, 0)
 653                  * because they can be throttled, but have some other mechanism
 654                  * for reserving space.  We still want these guys to refill the
 655                  * delayed block_rsv so just add 1 items worth of reservation
 656                  * here.
 657                  */
 658                 ret = btrfs_delayed_refs_rsv_refill(fs_info, flush);
 659                 if (ret)
 660                         goto reserve_fail;
 661         }
 662 again:
 663         h = kmem_cache_zalloc(btrfs_trans_handle_cachep, GFP_NOFS);
 664         if (!h) {
 665                 ret = -ENOMEM;
 666                 goto alloc_fail;
 667         }
 668
 669         /*
 670          * If we are JOIN_NOLOCK we're already committing a transaction and
 671          * waiting on this guy, so we don't need to do the sb_start_intwrite
 672          * because we're already holding a ref.  We need this because we could
 673          * have raced in and did an fsync() on a file which can kick a commit
 674          * and then we deadlock with somebody doing a freeze.
 675          *
 676          * If we are ATTACH, it means we just want to catch the current
 677          * transaction and commit it, so we needn't do sb_start_intwrite().
 678          */
 679         if (type & __TRANS_FREEZABLE)
 680                 sb_start_intwrite(fs_info->sb);
 681
 682         if (may_wait_transaction(fs_info, type))
 683                 wait_current_trans(fs_info);
 684
 685         do {
 686                 ret = join_transaction(fs_info, type);
 687                 if (ret == -EBUSY) {
 688                         wait_current_trans(fs_info);
 689                         if (unlikely(type == TRANS_ATTACH ||
 690                                      type == TRANS_JOIN_NOSTART))
 691                                 ret = -ENOENT;
 692                 }
 693         } while (ret == -EBUSY);
 694
 695         if (ret < 0)
 696                 goto join_fail;
 697
 698         cur_trans = fs_info->running_transaction;
 699
 700         h->transid = cur_trans->transid;
 701         h->transaction = cur_trans;
 702         h->root = root;
 703         refcount_set(&h->use_count, 1);
 704         h->fs_info = root->fs_info;
 705
 706         h->type = type;
 707         h->can_flush_pending_bgs = true;
 708         INIT_LIST_HEAD(&h->new_bgs);
 709
 710         smp_mb();
 711         if (cur_trans->state >= TRANS_STATE_COMMIT_START &&
 712             may_wait_transaction(fs_info, type)) {
 713                 current->journal_info = h;
 714                 btrfs_commit_transaction(h);
 715                 goto again;
 716         }
 717
 718         if (num_bytes) {
 719                 trace_btrfs_space_reservation(fs_info, "transaction",
 720                                               h->transid, num_bytes, 1);
 721                 h->block_rsv = &fs_info->trans_block_rsv;
 722                 h->bytes_reserved = num_bytes;
 723                 h->reloc_reserved = reloc_reserved;
 724         }
 725
 726 got_it:
 727         if (!current->journal_info)
 728                 current->journal_info = h;
 729
 730         /*
 731          * If the space_info is marked ALLOC_FORCE then we'll get upgraded to
 732          * ALLOC_FORCE the first run through, and then we won't allocate for
 733          * anybody else who races in later.  We don't care about the return
 734          * value here.
 735          */
 736         if (do_chunk_alloc && num_bytes) {
 737                 u64 flags = h->block_rsv->space_info->flags;
 738
 739                 btrfs_chunk_alloc(h, btrfs_get_alloc_profile(fs_info, flags),
 740                                   CHUNK_ALLOC_NO_FORCE);
 741         }
 742
 743         /*
 744          * btrfs_record_root_in_trans() needs to alloc new extents, and may
 745          * call btrfs_join_transaction() while we're also starting a
 746          * transaction.
 747          *
 748          * Thus it need to be called after current->journal_info initialized,
 749          * or we can deadlock.
 750          */
 751         ret = btrfs_record_root_in_trans(h, root);
 752         if (ret) {
 753                 /*
 754                  * The transaction handle is fully initialized and linked with
 755                  * other structures so it needs to be ended in case of errors,
 756                  * not just freed.
 757                  */
 758                 btrfs_end_transaction(h);
 759                 return ERR_PTR(ret);
 760         }
 761
 762         return h;
 763
 764 join_fail:
 765         if (type & __TRANS_FREEZABLE)
 766                 sb_end_intwrite(fs_info->sb);
 767         kmem_cache_free(btrfs_trans_handle_cachep, h);
 768 alloc_fail:
 769         if (num_bytes)
 770                 btrfs_block_rsv_release(fs_info, &fs_info->trans_block_rsv,
 771                                         num_bytes, NULL);
 772 reserve_fail:
 773         btrfs_qgroup_free_meta_pertrans(root, qgroup_reserved);
 774         return ERR_PTR(ret);
 775 }
 776
 777 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 778                                                    unsigned int num_items)
 779 {
 780         return start_transaction(root, num_items, TRANS_START,
 781                                  BTRFS_RESERVE_FLUSH_ALL, true);
 782 }
 783
 784 struct btrfs_trans_handle *btrfs_start_transaction_fallback_global_rsv(
 785                                         struct btrfs_root *root,
 786                                         unsigned int num_items)
 787 {
 788         return start_transaction(root, num_items, TRANS_START,
 789                                  BTRFS_RESERVE_FLUSH_ALL_STEAL, false);
 790 }
 791
 792 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 793 {
 794         return start_transaction(root, 0, TRANS_JOIN, BTRFS_RESERVE_NO_FLUSH,
 795                                  true);
 796 }
 797
 798 struct btrfs_trans_handle *btrfs_join_transaction_spacecache(struct btrfs_root *root)
 799 {
 800         return start_transaction(root, 0, TRANS_JOIN_NOLOCK,
 801                                  BTRFS_RESERVE_NO_FLUSH, true);
 802 }
 803
 804 /*
 805  * Similar to regular join but it never starts a transaction when none is
 806  * running or after waiting for the current one to finish.
 807  */
 808 struct btrfs_trans_handle *btrfs_join_transaction_nostart(struct btrfs_root *root)
 809 {
 810         return start_transaction(root, 0, TRANS_JOIN_NOSTART,
 811                                  BTRFS_RESERVE_NO_FLUSH, true);
 812 }
 813
 814 /*
 815  * btrfs_attach_transaction() - catch the running transaction
 816  *
 817  * It is used when we want to commit the current the transaction, but
 818  * don't want to start a new one.
 819  *
 820  * Note: If this function return -ENOENT, it just means there is no
 821  * running transaction. But it is possible that the inactive transaction
 822  * is still in the memory, not fully on disk. If you hope there is no
 823  * inactive transaction in the fs when -ENOENT is returned, you should
 824  * invoke
 825  *     btrfs_attach_transaction_barrier()
 826  */
 827 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
 828 {
 829         return start_transaction(root, 0, TRANS_ATTACH,
 830                                  BTRFS_RESERVE_NO_FLUSH, true);
 831 }
 832
 833 /*
 834  * btrfs_attach_transaction_barrier() - catch the running transaction
 835  *
 836  * It is similar to the above function, the difference is this one
 837  * will wait for all the inactive transactions until they fully
 838  * complete.
 839  */
 840 struct btrfs_trans_handle *
 841 btrfs_attach_transaction_barrier(struct btrfs_root *root)
 842 {
 843         struct btrfs_trans_handle *trans;
 844
 845         trans = start_transaction(root, 0, TRANS_ATTACH,
 846                                   BTRFS_RESERVE_NO_FLUSH, true);
 847         if (trans == ERR_PTR(-ENOENT))
 848                 btrfs_wait_for_commit(root->fs_info, 0);
 849
 850         return trans;
 851 }
 852
 853 /* Wait for a transaction commit to reach at least the given state. */
 854 static noinline void wait_for_commit(struct btrfs_transaction *commit,
 855                                      const enum btrfs_trans_state min_state)
 856 {
 857         wait_event(commit->commit_wait, commit->state >= min_state);
 858 }
 859
 860 int btrfs_wait_for_commit(struct btrfs_fs_info *fs_info, u64 transid)
 861 {
 862         struct btrfs_transaction *cur_trans = NULL, *t;
 863         int ret = 0;
 864
 865         if (transid) {
 866                 if (transid <= fs_info->last_trans_committed)
 867                         goto out;
 868
 869                 /* find specified transaction */
 870                 spin_lock(&fs_info->trans_lock);
 871                 list_for_each_entry(t, &fs_info->trans_list, list) {
 872                         if (t->transid == transid) {
 873                                 cur_trans = t;
 874                                 refcount_inc(&cur_trans->use_count);
 875                                 ret = 0;
 876                                 break;
 877                         }
 878                         if (t->transid > transid) {
 879                                 ret = 0;
 880                                 break;
 881                         }
 882                 }
 883                 spin_unlock(&fs_info->trans_lock);
 884
 885                 /*
 886                  * The specified transaction doesn't exist, or we
 887                  * raced with btrfs_commit_transaction
 888                  */
 889                 if (!cur_trans) {
 890                         if (transid > fs_info->last_trans_committed)
 891                                 ret = -EINVAL;
 892                         goto out;
 893                 }
 894         } else {
 895                 /* find newest transaction that is committing | committed */
 896                 spin_lock(&fs_info->trans_lock);
 897                 list_for_each_entry_reverse(t, &fs_info->trans_list,
 898                                             list) {
 899                         if (t->state >= TRANS_STATE_COMMIT_START) {
 900                                 if (t->state == TRANS_STATE_COMPLETED)
 901                                         break;
 902                                 cur_trans = t;
 903                                 refcount_inc(&cur_trans->use_count);
 904                                 break;
 905                         }
 906                 }
 907                 spin_unlock(&fs_info->trans_lock);
 908                 if (!cur_trans)
 909                         goto out;  /* nothing committing|committed */
 910         }
 911
 912         wait_for_commit(cur_trans, TRANS_STATE_COMPLETED);
 913         btrfs_put_transaction(cur_trans);
 914 out:
 915         return ret;
 916 }
 917
 918 void btrfs_throttle(struct btrfs_fs_info *fs_info)
 919 {
 920         wait_current_trans(fs_info);
 921 }
 922
 923 static bool should_end_transaction(struct btrfs_trans_handle *trans)
 924 {
 925         struct btrfs_fs_info *fs_info = trans->fs_info;
 926
 927         if (btrfs_check_space_for_delayed_refs(fs_info))
 928                 return true;
 929
 930         return !!btrfs_block_rsv_check(&fs_info->global_block_rsv, 5);
 931 }
 932
 933 bool btrfs_should_end_transaction(struct btrfs_trans_handle *trans)
 934 {
 935         struct btrfs_transaction *cur_trans = trans->transaction;
 936
 937         if (cur_trans->state >= TRANS_STATE_COMMIT_START ||
 938             test_bit(BTRFS_DELAYED_REFS_FLUSHING, &cur_trans->delayed_refs.flags))
 939                 return true;
 940
 941         return should_end_transaction(trans);
 942 }
 943
 944 static void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans)
 945
 946 {
 947         struct btrfs_fs_info *fs_info = trans->fs_info;
 948
 949         if (!trans->block_rsv) {
 950                 ASSERT(!trans->bytes_reserved);
 951                 return;
 952         }
 953
 954         if (!trans->bytes_reserved)
 955                 return;
 956
 957         ASSERT(trans->block_rsv == &fs_info->trans_block_rsv);
 958         trace_btrfs_space_reservation(fs_info, "transaction",
 959                                       trans->transid, trans->bytes_reserved, 0);
 960         btrfs_block_rsv_release(fs_info, trans->block_rsv,
 961                                 trans->bytes_reserved, NULL);
 962         trans->bytes_reserved = 0;
 963 }
 964
 965 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 966                                    int throttle)
 967 {
 968         struct btrfs_fs_info *info = trans->fs_info;
 969         struct btrfs_transaction *cur_trans = trans->transaction;
 970         int err = 0;
 971
 972         if (refcount_read(&trans->use_count) > 1) {
 973                 refcount_dec(&trans->use_count);
 974                 trans->block_rsv = trans->orig_rsv;
 975                 return 0;
 976         }
 977
 978         btrfs_trans_release_metadata(trans);
 979         trans->block_rsv = NULL;
 980
 981         btrfs_create_pending_block_groups(trans);
 982
 983         btrfs_trans_release_chunk_metadata(trans);
 984
 985         if (trans->type & __TRANS_FREEZABLE)
 986                 sb_end_intwrite(info->sb);
 987
 988         WARN_ON(cur_trans != info->running_transaction);
 989         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 990         atomic_dec(&cur_trans->num_writers);
 991         extwriter_counter_dec(cur_trans, trans->type);
 992
 993         cond_wake_up(&cur_trans->writer_wait);
 994         btrfs_put_transaction(cur_trans);
 995
 996         if (current->journal_info == trans)
 997                 current->journal_info = NULL;
 998
 999         if (throttle)
1000                 btrfs_run_delayed_iputs(info);
1001
1002         if (TRANS_ABORTED(trans) ||
1003             test_bit(BTRFS_FS_STATE_ERROR, &info->fs_state)) {
1004                 wake_up_process(info->transaction_kthread);
1005                 if (TRANS_ABORTED(trans))
1006                         err = trans->aborted;
1007                 else
1008                         err = -EROFS;
1009         }
1010
1011         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1012         return err;
1013 }
1014
1015 int btrfs_end_transaction(struct btrfs_trans_handle *trans)
1016 {
1017         return __btrfs_end_transaction(trans, 0);
1018 }
1019
1020 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans)
1021 {
1022         return __btrfs_end_transaction(trans, 1);
1023 }
1024
1025 /*
1026  * when btree blocks are allocated, they have some corresponding bits set for
1027  * them in one of two extent_io trees.  This is used to make sure all of
1028  * those extents are sent to disk but does not wait on them
1029  */
1030 int btrfs_write_marked_extents(struct btrfs_fs_info *fs_info,
1031                                struct extent_io_tree *dirty_pages, int mark)
1032 {
1033         int err = 0;
1034         int werr = 0;
1035         struct address_space *mapping = fs_info->btree_inode->i_mapping;
1036         struct extent_state *cached_state = NULL;
1037         u64 start = 0;
1038         u64 end;
1039
1040         atomic_inc(&BTRFS_I(fs_info->btree_inode)->sync_writers);
1041         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
1042                                       mark, &cached_state)) {
1043                 bool wait_writeback = false;
1044
1045                 err = convert_extent_bit(dirty_pages, start, end,
1046                                          EXTENT_NEED_WAIT,
1047                                          mark, &cached_state);
1048                 /*
1049                  * convert_extent_bit can return -ENOMEM, which is most of the
1050                  * time a temporary error. So when it happens, ignore the error
1051                  * and wait for writeback of this range to finish - because we
1052                  * failed to set the bit EXTENT_NEED_WAIT for the range, a call
1053                  * to __btrfs_wait_marked_extents() would not know that
1054                  * writeback for this range started and therefore wouldn't
1055                  * wait for it to finish - we don't want to commit a
1056                  * superblock that points to btree nodes/leafs for which
1057                  * writeback hasn't finished yet (and without errors).
1058                  * We cleanup any entries left in the io tree when committing
1059                  * the transaction (through extent_io_tree_release()).
1060                  */
1061                 if (err == -ENOMEM) {
1062                         err = 0;
1063                         wait_writeback = true;
1064                 }
1065                 if (!err)
1066                         err = filemap_fdatawrite_range(mapping, start, end);
1067                 if (err)
1068                         werr = err;
1069                 else if (wait_writeback)
1070                         werr = filemap_fdatawait_range(mapping, start, end);
1071                 free_extent_state(cached_state);
1072                 cached_state = NULL;
1073                 cond_resched();
1074                 start = end + 1;
1075         }
1076         atomic_dec(&BTRFS_I(fs_info->btree_inode)->sync_writers);
1077         return werr;
1078 }
1079
1080 /*
1081  * when btree blocks are allocated, they have some corresponding bits set for
1082  * them in one of two extent_io trees.  This is used to make sure all of
1083  * those extents are on disk for transaction or log commit.  We wait
1084  * on all the pages and clear them from the dirty pages state tree
1085  */
1086 static int __btrfs_wait_marked_extents(struct btrfs_fs_info *fs_info,
1087                                        struct extent_io_tree *dirty_pages)
1088 {
1089         int err = 0;
1090         int werr = 0;
1091         struct address_space *mapping = fs_info->btree_inode->i_mapping;
1092         struct extent_state *cached_state = NULL;
1093         u64 start = 0;
1094         u64 end;
1095
1096         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
1097                                       EXTENT_NEED_WAIT, &cached_state)) {
1098                 /*
1099                  * Ignore -ENOMEM errors returned by clear_extent_bit().
1100                  * When committing the transaction, we'll remove any entries
1101                  * left in the io tree. For a log commit, we don't remove them
1102                  * after committing the log because the tree can be accessed
1103                  * concurrently - we do it only at transaction commit time when
1104                  * it's safe to do it (through extent_io_tree_release()).
1105                  */
1106                 err = clear_extent_bit(dirty_pages, start, end,
1107                                        EXTENT_NEED_WAIT, 0, 0, &cached_state);
1108                 if (err == -ENOMEM)
1109                         err = 0;
1110                 if (!err)
1111                         err = filemap_fdatawait_range(mapping, start, end);
1112                 if (err)
1113                         werr = err;
1114                 free_extent_state(cached_state);
1115                 cached_state = NULL;
1116                 cond_resched();
1117                 start = end + 1;
1118         }
1119         if (err)
1120                 werr = err;
1121         return werr;
1122 }
1123
1124 static int btrfs_wait_extents(struct btrfs_fs_info *fs_info,
1125                        struct extent_io_tree *dirty_pages)
1126 {
1127         bool errors = false;
1128         int err;
1129
1130         err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1131         if (test_and_clear_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags))
1132                 errors = true;
1133
1134         if (errors && !err)
1135                 err = -EIO;
1136         return err;
1137 }
1138
1139 int btrfs_wait_tree_log_extents(struct btrfs_root *log_root, int mark)
1140 {
1141         struct btrfs_fs_info *fs_info = log_root->fs_info;
1142         struct extent_io_tree *dirty_pages = &log_root->dirty_log_pages;
1143         bool errors = false;
1144         int err;
1145
1146         ASSERT(log_root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
1147
1148         err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1149         if ((mark & EXTENT_DIRTY) &&
1150             test_and_clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags))
1151                 errors = true;
1152
1153         if ((mark & EXTENT_NEW) &&
1154             test_and_clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags))
1155                 errors = true;
1156
1157         if (errors && !err)
1158                 err = -EIO;
1159         return err;
1160 }
1161
1162 /*
1163  * When btree blocks are allocated the corresponding extents are marked dirty.
1164  * This function ensures such extents are persisted on disk for transaction or
1165  * log commit.
1166  *
1167  * @trans: transaction whose dirty pages we'd like to write
1168  */
1169 static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans)
1170 {
1171         int ret;
1172         int ret2;
1173         struct extent_io_tree *dirty_pages = &trans->transaction->dirty_pages;
1174         struct btrfs_fs_info *fs_info = trans->fs_info;
1175         struct blk_plug plug;
1176
1177         blk_start_plug(&plug);
1178         ret = btrfs_write_marked_extents(fs_info, dirty_pages, EXTENT_DIRTY);
1179         blk_finish_plug(&plug);
1180         ret2 = btrfs_wait_extents(fs_info, dirty_pages);
1181
1182         extent_io_tree_release(&trans->transaction->dirty_pages);
1183
1184         if (ret)
1185                 return ret;
1186         else if (ret2)
1187                 return ret2;
1188         else
1189                 return 0;
1190 }
1191
1192 /*
1193  * this is used to update the root pointer in the tree of tree roots.
1194  *
1195  * But, in the case of the extent allocation tree, updating the root
1196  * pointer may allocate blocks which may change the root of the extent
1197  * allocation tree.
1198  *
1199  * So, this loops and repeats and makes sure the cowonly root didn't
1200  * change while the root pointer was being updated in the metadata.
1201  */
1202 static int update_cowonly_root(struct btrfs_trans_handle *trans,
1203                                struct btrfs_root *root)
1204 {
1205         int ret;
1206         u64 old_root_bytenr;
1207         u64 old_root_used;
1208         struct btrfs_fs_info *fs_info = root->fs_info;
1209         struct btrfs_root *tree_root = fs_info->tree_root;
1210
1211         old_root_used = btrfs_root_used(&root->root_item);
1212
1213         while (1) {
1214                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1215                 if (old_root_bytenr == root->node->start &&
1216                     old_root_used == btrfs_root_used(&root->root_item))
1217                         break;
1218
1219                 btrfs_set_root_node(&root->root_item, root->node);
1220                 ret = btrfs_update_root(trans, tree_root,
1221                                         &root->root_key,
1222                                         &root->root_item);
1223                 if (ret)
1224                         return ret;
1225
1226                 old_root_used = btrfs_root_used(&root->root_item);
1227         }
1228
1229         return 0;
1230 }
1231
1232 /*
1233  * update all the cowonly tree roots on disk
1234  *
1235  * The error handling in this function may not be obvious. Any of the
1236  * failures will cause the file system to go offline. We still need
1237  * to clean up the delayed refs.
1238  */
1239 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans)
1240 {
1241         struct btrfs_fs_info *fs_info = trans->fs_info;
1242         struct list_head *dirty_bgs = &trans->transaction->dirty_bgs;
1243         struct list_head *io_bgs = &trans->transaction->io_bgs;
1244         struct list_head *next;
1245         struct extent_buffer *eb;
1246         int ret;
1247
1248         eb = btrfs_lock_root_node(fs_info->tree_root);
1249         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
1250                               0, &eb, BTRFS_NESTING_COW);
1251         btrfs_tree_unlock(eb);
1252         free_extent_buffer(eb);
1253
1254         if (ret)
1255                 return ret;
1256
1257         ret = btrfs_run_dev_stats(trans);
1258         if (ret)
1259                 return ret;
1260         ret = btrfs_run_dev_replace(trans);
1261         if (ret)
1262                 return ret;
1263         ret = btrfs_run_qgroups(trans);
1264         if (ret)
1265                 return ret;
1266
1267         ret = btrfs_setup_space_cache(trans);
1268         if (ret)
1269                 return ret;
1270
1271 again:
1272         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1273                 struct btrfs_root *root;
1274                 next = fs_info->dirty_cowonly_roots.next;
1275                 list_del_init(next);
1276                 root = list_entry(next, struct btrfs_root, dirty_list);
1277                 clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1278
1279                 if (root != fs_info->extent_root)
1280                         list_add_tail(&root->dirty_list,
1281                                       &trans->transaction->switch_commits);
1282                 ret = update_cowonly_root(trans, root);
1283                 if (ret)
1284                         return ret;
1285         }
1286
1287         /* Now flush any delayed refs generated by updating all of the roots */
1288         ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1289         if (ret)
1290                 return ret;
1291
1292         while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1293                 ret = btrfs_write_dirty_block_groups(trans);
1294                 if (ret)
1295                         return ret;
1296
1297                 /*
1298                  * We're writing the dirty block groups, which could generate
1299                  * delayed refs, which could generate more dirty block groups,
1300                  * so we want to keep this flushing in this loop to make sure
1301                  * everything gets run.
1302                  */
1303                 ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1304                 if (ret)
1305                         return ret;
1306         }
1307
1308         if (!list_empty(&fs_info->dirty_cowonly_roots))
1309                 goto again;
1310
1311         list_add_tail(&fs_info->extent_root->dirty_list,
1312                       &trans->transaction->switch_commits);
1313
1314         /* Update dev-replace pointer once everything is committed */
1315         fs_info->dev_replace.committed_cursor_left =
1316                 fs_info->dev_replace.cursor_left_last_write_of_item;
1317
1318         return 0;
1319 }
1320
1321 /*
1322  * dead roots are old snapshots that need to be deleted.  This allocates
1323  * a dirty root struct and adds it into the list of dead roots that need to
1324  * be deleted
1325  */
1326 void btrfs_add_dead_root(struct btrfs_root *root)
1327 {
1328         struct btrfs_fs_info *fs_info = root->fs_info;
1329
1330         spin_lock(&fs_info->trans_lock);
1331         if (list_empty(&root->root_list)) {
1332                 btrfs_grab_root(root);
1333                 list_add_tail(&root->root_list, &fs_info->dead_roots);
1334         }
1335         spin_unlock(&fs_info->trans_lock);
1336 }
1337
1338 /*
1339  * update all the cowonly tree roots on disk
1340  */
1341 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans)
1342 {
1343         struct btrfs_fs_info *fs_info = trans->fs_info;
1344         struct btrfs_root *gang[8];
1345         int i;
1346         int ret;
1347
1348         spin_lock(&fs_info->fs_roots_radix_lock);
1349         while (1) {
1350                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1351                                                  (void **)gang, 0,
1352                                                  ARRAY_SIZE(gang),
1353                                                  BTRFS_ROOT_TRANS_TAG);
1354                 if (ret == 0)
1355                         break;
1356                 for (i = 0; i < ret; i++) {
1357                         struct btrfs_root *root = gang[i];
1358                         int ret2;
1359
1360                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
1361                                         (unsigned long)root->root_key.objectid,
1362                                         BTRFS_ROOT_TRANS_TAG);
1363                         spin_unlock(&fs_info->fs_roots_radix_lock);
1364
1365                         btrfs_free_log(trans, root);
1366                         ret2 = btrfs_update_reloc_root(trans, root);
1367                         if (ret2)
1368                                 return ret2;
1369
1370                         /* see comments in should_cow_block() */
1371                         clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1372                         smp_mb__after_atomic();
1373
1374                         if (root->commit_root != root->node) {
1375                                 list_add_tail(&root->dirty_list,
1376                                         &trans->transaction->switch_commits);
1377                                 btrfs_set_root_node(&root->root_item,
1378                                                     root->node);
1379                         }
1380
1381                         ret2 = btrfs_update_root(trans, fs_info->tree_root,
1382                                                 &root->root_key,
1383                                                 &root->root_item);
1384                         if (ret2)
1385                                 return ret2;
1386                         spin_lock(&fs_info->fs_roots_radix_lock);
1387                         btrfs_qgroup_free_meta_all_pertrans(root);
1388                 }
1389         }
1390         spin_unlock(&fs_info->fs_roots_radix_lock);
1391         return 0;
1392 }
1393
1394 /*
1395  * defrag a given btree.
1396  * Every leaf in the btree is read and defragged.
1397  */
1398 int btrfs_defrag_root(struct btrfs_root *root)
1399 {
1400         struct btrfs_fs_info *info = root->fs_info;
1401         struct btrfs_trans_handle *trans;
1402         int ret;
1403
1404         if (test_and_set_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state))
1405                 return 0;
1406
1407         while (1) {
1408                 trans = btrfs_start_transaction(root, 0);
1409                 if (IS_ERR(trans))
1410                         return PTR_ERR(trans);
1411
1412                 ret = btrfs_defrag_leaves(trans, root);
1413
1414                 btrfs_end_transaction(trans);
1415                 btrfs_btree_balance_dirty(info);
1416                 cond_resched();
1417
1418                 if (btrfs_fs_closing(info) || ret != -EAGAIN)
1419                         break;
1420
1421                 if (btrfs_defrag_cancelled(info)) {
1422                         btrfs_debug(info, "defrag_root cancelled");
1423                         ret = -EAGAIN;
1424                         break;
1425                 }
1426         }
1427         clear_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state);
1428         return ret;
1429 }
1430
1431 /*
1432  * Do all special snapshot related qgroup dirty hack.
1433  *
1434  * Will do all needed qgroup inherit and dirty hack like switch commit
1435  * roots inside one transaction and write all btree into disk, to make
1436  * qgroup works.
1437  */
1438 static int qgroup_account_snapshot(struct btrfs_trans_handle *trans,
1439                                    struct btrfs_root *src,
1440                                    struct btrfs_root *parent,
1441                                    struct btrfs_qgroup_inherit *inherit,
1442                                    u64 dst_objectid)
1443 {
1444         struct btrfs_fs_info *fs_info = src->fs_info;
1445         int ret;
1446
1447         /*
1448          * Save some performance in the case that qgroups are not
1449          * enabled. If this check races with the ioctl, rescan will
1450          * kick in anyway.
1451          */
1452         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
1453                 return 0;
1454
1455         /*
1456          * Ensure dirty @src will be committed.  Or, after coming
1457          * commit_fs_roots() and switch_commit_roots(), any dirty but not
1458          * recorded root will never be updated again, causing an outdated root
1459          * item.
1460          */
1461         ret = record_root_in_trans(trans, src, 1);
1462         if (ret)
1463                 return ret;
1464
1465         /*
1466          * btrfs_qgroup_inherit relies on a consistent view of the usage for the
1467          * src root, so we must run the delayed refs here.
1468          *
1469          * However this isn't particularly fool proof, because there's no
1470          * synchronization keeping us from changing the tree after this point
1471          * before we do the qgroup_inherit, or even from making changes while
1472          * we're doing the qgroup_inherit.  But that's a problem for the future,
1473          * for now flush the delayed refs to narrow the race window where the
1474          * qgroup counters could end up wrong.
1475          */
1476         ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1477         if (ret) {
1478                 btrfs_abort_transaction(trans, ret);
1479                 goto out;
1480         }
1481
1482         /*
1483          * We are going to commit transaction, see btrfs_commit_transaction()
1484          * comment for reason locking tree_log_mutex
1485          */
1486         mutex_lock(&fs_info->tree_log_mutex);
1487
1488         ret = commit_fs_roots(trans);
1489         if (ret)
1490                 goto out;
1491         ret = btrfs_qgroup_account_extents(trans);
1492         if (ret < 0)
1493                 goto out;
1494
1495         /* Now qgroup are all updated, we can inherit it to new qgroups */
1496         ret = btrfs_qgroup_inherit(trans, src->root_key.objectid, dst_objectid,
1497                                    inherit);
1498         if (ret < 0)
1499                 goto out;
1500
1501         /*
1502          * Now we do a simplified commit transaction, which will:
1503          * 1) commit all subvolume and extent tree
1504          *    To ensure all subvolume and extent tree have a valid
1505          *    commit_root to accounting later insert_dir_item()
1506          * 2) write all btree blocks onto disk
1507          *    This is to make sure later btree modification will be cowed
1508          *    Or commit_root can be populated and cause wrong qgroup numbers
1509          * In this simplified commit, we don't really care about other trees
1510          * like chunk and root tree, as they won't affect qgroup.
1511          * And we don't write super to avoid half committed status.
1512          */
1513         ret = commit_cowonly_roots(trans);
1514         if (ret)
1515                 goto out;
1516         switch_commit_roots(trans);
1517         ret = btrfs_write_and_wait_transaction(trans);
1518         if (ret)
1519                 btrfs_handle_fs_error(fs_info, ret,
1520                         "Error while writing out transaction for qgroup");
1521
1522 out:
1523         mutex_unlock(&fs_info->tree_log_mutex);
1524
1525         /*
1526          * Force parent root to be updated, as we recorded it before so its
1527          * last_trans == cur_transid.
1528          * Or it won't be committed again onto disk after later
1529          * insert_dir_item()
1530          */
1531         if (!ret)
1532                 ret = record_root_in_trans(trans, parent, 1);
1533         return ret;
1534 }
1535
1536 /*
1537  * new snapshots need to be created at a very specific time in the
1538  * transaction commit.  This does the actual creation.
1539  *
1540  * Note:
1541  * If the error which may affect the commitment of the current transaction
1542  * happens, we should return the error number. If the error which just affect
1543  * the creation of the pending snapshots, just return 0.
1544  */
1545 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1546                                    struct btrfs_pending_snapshot *pending)
1547 {
1548
1549         struct btrfs_fs_info *fs_info = trans->fs_info;
1550         struct btrfs_key key;
1551         struct btrfs_root_item *new_root_item;
1552         struct btrfs_root *tree_root = fs_info->tree_root;
1553         struct btrfs_root *root = pending->root;
1554         struct btrfs_root *parent_root;
1555         struct btrfs_block_rsv *rsv;
1556         struct inode *parent_inode;
1557         struct btrfs_path *path;
1558         struct btrfs_dir_item *dir_item;
1559         struct dentry *dentry;
1560         struct extent_buffer *tmp;
1561         struct extent_buffer *old;
1562         struct timespec64 cur_time;
1563         int ret = 0;
1564         u64 to_reserve = 0;
1565         u64 index = 0;
1566         u64 objectid;
1567         u64 root_flags;
1568
1569         ASSERT(pending->path);
1570         path = pending->path;
1571
1572         ASSERT(pending->root_item);
1573         new_root_item = pending->root_item;
1574
1575         pending->error = btrfs_get_free_objectid(tree_root, &objectid);
1576         if (pending->error)
1577                 goto no_free_objectid;
1578
1579         /*
1580          * Make qgroup to skip current new snapshot's qgroupid, as it is
1581          * accounted by later btrfs_qgroup_inherit().
1582          */
1583         btrfs_set_skip_qgroup(trans, objectid);
1584
1585         btrfs_reloc_pre_snapshot(pending, &to_reserve);
1586
1587         if (to_reserve > 0) {
1588                 pending->error = btrfs_block_rsv_add(root,
1589                                                      &pending->block_rsv,
1590                                                      to_reserve,
1591                                                      BTRFS_RESERVE_NO_FLUSH);
1592                 if (pending->error)
1593                         goto clear_skip_qgroup;
1594         }
1595
1596         key.objectid = objectid;
1597         key.offset = (u64)-1;
1598         key.type = BTRFS_ROOT_ITEM_KEY;
1599
1600         rsv = trans->block_rsv;
1601         trans->block_rsv = &pending->block_rsv;
1602         trans->bytes_reserved = trans->block_rsv->reserved;
1603         trace_btrfs_space_reservation(fs_info, "transaction",
1604                                       trans->transid,
1605                                       trans->bytes_reserved, 1);
1606         dentry = pending->dentry;
1607         parent_inode = pending->dir;
1608         parent_root = BTRFS_I(parent_inode)->root;
1609         ret = record_root_in_trans(trans, parent_root, 0);
1610         if (ret)
1611                 goto fail;
1612         cur_time = current_time(parent_inode);
1613
1614         /*
1615          * insert the directory item
1616          */
1617         ret = btrfs_set_inode_index(BTRFS_I(parent_inode), &index);
1618         BUG_ON(ret); /* -ENOMEM */
1619
1620         /* check if there is a file/dir which has the same name. */
1621         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1622                                          btrfs_ino(BTRFS_I(parent_inode)),
1623                                          dentry->d_name.name,
1624                                          dentry->d_name.len, 0);
1625         if (dir_item != NULL && !IS_ERR(dir_item)) {
1626                 pending->error = -EEXIST;
1627                 goto dir_item_existed;
1628         } else if (IS_ERR(dir_item)) {
1629                 ret = PTR_ERR(dir_item);
1630                 btrfs_abort_transaction(trans, ret);
1631                 goto fail;
1632         }
1633         btrfs_release_path(path);
1634
1635         /*
1636          * pull in the delayed directory update
1637          * and the delayed inode item
1638          * otherwise we corrupt the FS during
1639          * snapshot
1640          */
1641         ret = btrfs_run_delayed_items(trans);
1642         if (ret) {      /* Transaction aborted */
1643                 btrfs_abort_transaction(trans, ret);
1644                 goto fail;
1645         }
1646
1647         ret = record_root_in_trans(trans, root, 0);
1648         if (ret) {
1649                 btrfs_abort_transaction(trans, ret);
1650                 goto fail;
1651         }
1652         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1653         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1654         btrfs_check_and_init_root_item(new_root_item);
1655
1656         root_flags = btrfs_root_flags(new_root_item);
1657         if (pending->readonly)
1658                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1659         else
1660                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1661         btrfs_set_root_flags(new_root_item, root_flags);
1662
1663         btrfs_set_root_generation_v2(new_root_item,
1664                         trans->transid);
1665         generate_random_guid(new_root_item->uuid);
1666         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1667                         BTRFS_UUID_SIZE);
1668         if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1669                 memset(new_root_item->received_uuid, 0,
1670                        sizeof(new_root_item->received_uuid));
1671                 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1672                 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1673                 btrfs_set_root_stransid(new_root_item, 0);
1674                 btrfs_set_root_rtransid(new_root_item, 0);
1675         }
1676         btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1677         btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1678         btrfs_set_root_otransid(new_root_item, trans->transid);
1679
1680         old = btrfs_lock_root_node(root);
1681         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old,
1682                               BTRFS_NESTING_COW);
1683         if (ret) {
1684                 btrfs_tree_unlock(old);
1685                 free_extent_buffer(old);
1686                 btrfs_abort_transaction(trans, ret);
1687                 goto fail;
1688         }
1689
1690         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1691         /* clean up in any case */
1692         btrfs_tree_unlock(old);
1693         free_extent_buffer(old);
1694         if (ret) {
1695                 btrfs_abort_transaction(trans, ret);
1696                 goto fail;
1697         }
1698         /* see comments in should_cow_block() */
1699         set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1700         smp_wmb();
1701
1702         btrfs_set_root_node(new_root_item, tmp);
1703         /* record when the snapshot was created in key.offset */
1704         key.offset = trans->transid;
1705         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1706         btrfs_tree_unlock(tmp);
1707         free_extent_buffer(tmp);
1708         if (ret) {
1709                 btrfs_abort_transaction(trans, ret);
1710                 goto fail;
1711         }
1712
1713         /*
1714          * insert root back/forward references
1715          */
1716         ret = btrfs_add_root_ref(trans, objectid,
1717                                  parent_root->root_key.objectid,
1718                                  btrfs_ino(BTRFS_I(parent_inode)), index,
1719                                  dentry->d_name.name, dentry->d_name.len);
1720         if (ret) {
1721                 btrfs_abort_transaction(trans, ret);
1722                 goto fail;
1723         }
1724
1725         key.offset = (u64)-1;
1726         pending->snap = btrfs_get_new_fs_root(fs_info, objectid, pending->anon_dev);
1727         if (IS_ERR(pending->snap)) {
1728                 ret = PTR_ERR(pending->snap);
1729                 pending->snap = NULL;
1730                 btrfs_abort_transaction(trans, ret);
1731                 goto fail;
1732         }
1733
1734         ret = btrfs_reloc_post_snapshot(trans, pending);
1735         if (ret) {
1736                 btrfs_abort_transaction(trans, ret);
1737                 goto fail;
1738         }
1739
1740         /*
1741          * Do special qgroup accounting for snapshot, as we do some qgroup
1742          * snapshot hack to do fast snapshot.
1743          * To co-operate with that hack, we do hack again.
1744          * Or snapshot will be greatly slowed down by a subtree qgroup rescan
1745          */
1746         ret = qgroup_account_snapshot(trans, root, parent_root,
1747                                       pending->inherit, objectid);
1748         if (ret < 0)
1749                 goto fail;
1750
1751         ret = btrfs_insert_dir_item(trans, dentry->d_name.name,
1752                                     dentry->d_name.len, BTRFS_I(parent_inode),
1753                                     &key, BTRFS_FT_DIR, index);
1754         /* We have check then name at the beginning, so it is impossible. */
1755         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1756         if (ret) {
1757                 btrfs_abort_transaction(trans, ret);
1758                 goto fail;
1759         }
1760
1761         btrfs_i_size_write(BTRFS_I(parent_inode), parent_inode->i_size +
1762                                          dentry->d_name.len * 2);
1763         parent_inode->i_mtime = parent_inode->i_ctime =
1764                 current_time(parent_inode);
1765         ret = btrfs_update_inode_fallback(trans, parent_root, BTRFS_I(parent_inode));
1766         if (ret) {
1767                 btrfs_abort_transaction(trans, ret);
1768                 goto fail;
1769         }
1770         ret = btrfs_uuid_tree_add(trans, new_root_item->uuid,
1771                                   BTRFS_UUID_KEY_SUBVOL,
1772                                   objectid);
1773         if (ret) {
1774                 btrfs_abort_transaction(trans, ret);
1775                 goto fail;
1776         }
1777         if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1778                 ret = btrfs_uuid_tree_add(trans, new_root_item->received_uuid,
1779                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1780                                           objectid);
1781                 if (ret && ret != -EEXIST) {
1782                         btrfs_abort_transaction(trans, ret);
1783                         goto fail;
1784                 }
1785         }
1786
1787 fail:
1788         pending->error = ret;
1789 dir_item_existed:
1790         trans->block_rsv = rsv;
1791         trans->bytes_reserved = 0;
1792 clear_skip_qgroup:
1793         btrfs_clear_skip_qgroup(trans);
1794 no_free_objectid:
1795         kfree(new_root_item);
1796         pending->root_item = NULL;
1797         btrfs_free_path(path);
1798         pending->path = NULL;
1799
1800         return ret;
1801 }
1802
1803 /*
1804  * create all the snapshots we've scheduled for creation
1805  */
1806 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans)
1807 {
1808         struct btrfs_pending_snapshot *pending, *next;
1809         struct list_head *head = &trans->transaction->pending_snapshots;
1810         int ret = 0;
1811
1812         list_for_each_entry_safe(pending, next, head, list) {
1813                 list_del(&pending->list);
1814                 ret = create_pending_snapshot(trans, pending);
1815                 if (ret)
1816                         break;
1817         }
1818         return ret;
1819 }
1820
1821 static void update_super_roots(struct btrfs_fs_info *fs_info)
1822 {
1823         struct btrfs_root_item *root_item;
1824         struct btrfs_super_block *super;
1825
1826         super = fs_info->super_copy;
1827
1828         root_item = &fs_info->chunk_root->root_item;
1829         super->chunk_root = root_item->bytenr;
1830         super->chunk_root_generation = root_item->generation;
1831         super->chunk_root_level = root_item->level;
1832
1833         root_item = &fs_info->tree_root->root_item;
1834         super->root = root_item->bytenr;
1835         super->generation = root_item->generation;
1836         super->root_level = root_item->level;
1837         if (btrfs_test_opt(fs_info, SPACE_CACHE))
1838                 super->cache_generation = root_item->generation;
1839         else if (test_bit(BTRFS_FS_CLEANUP_SPACE_CACHE_V1, &fs_info->flags))
1840                 super->cache_generation = 0;
1841         if (test_bit(BTRFS_FS_UPDATE_UUID_TREE_GEN, &fs_info->flags))
1842                 super->uuid_tree_generation = root_item->generation;
1843 }
1844
1845 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1846 {
1847         struct btrfs_transaction *trans;
1848         int ret = 0;
1849
1850         spin_lock(&info->trans_lock);
1851         trans = info->running_transaction;
1852         if (trans)
1853                 ret = (trans->state >= TRANS_STATE_COMMIT_START);
1854         spin_unlock(&info->trans_lock);
1855         return ret;
1856 }
1857
1858 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1859 {
1860         struct btrfs_transaction *trans;
1861         int ret = 0;
1862
1863         spin_lock(&info->trans_lock);
1864         trans = info->running_transaction;
1865         if (trans)
1866                 ret = is_transaction_blocked(trans);
1867         spin_unlock(&info->trans_lock);
1868         return ret;
1869 }
1870
1871 /*
1872  * wait for the current transaction commit to start and block subsequent
1873  * transaction joins
1874  */
1875 static void wait_current_trans_commit_start(struct btrfs_fs_info *fs_info,
1876                                             struct btrfs_transaction *trans)
1877 {
1878         wait_event(fs_info->transaction_blocked_wait,
1879                    trans->state >= TRANS_STATE_COMMIT_START ||
1880                    TRANS_ABORTED(trans));
1881 }
1882
1883 /*
1884  * wait for the current transaction to start and then become unblocked.
1885  * caller holds ref.
1886  */
1887 static void wait_current_trans_commit_start_and_unblock(
1888                                         struct btrfs_fs_info *fs_info,
1889                                         struct btrfs_transaction *trans)
1890 {
1891         wait_event(fs_info->transaction_wait,
1892                    trans->state >= TRANS_STATE_UNBLOCKED ||
1893                    TRANS_ABORTED(trans));
1894 }
1895
1896 /*
1897  * commit transactions asynchronously. once btrfs_commit_transaction_async
1898  * returns, any subsequent transaction will not be allowed to join.
1899  */
1900 struct btrfs_async_commit {
1901         struct btrfs_trans_handle *newtrans;
1902         struct work_struct work;
1903 };
1904
1905 static void do_async_commit(struct work_struct *work)
1906 {
1907         struct btrfs_async_commit *ac =
1908                 container_of(work, struct btrfs_async_commit, work);
1909
1910         /*
1911          * We've got freeze protection passed with the transaction.
1912          * Tell lockdep about it.
1913          */
1914         if (ac->newtrans->type & __TRANS_FREEZABLE)
1915                 __sb_writers_acquired(ac->newtrans->fs_info->sb, SB_FREEZE_FS);
1916
1917         current->journal_info = ac->newtrans;
1918
1919         btrfs_commit_transaction(ac->newtrans);
1920         kfree(ac);
1921 }
1922
1923 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1924                                    int wait_for_unblock)
1925 {
1926         struct btrfs_fs_info *fs_info = trans->fs_info;
1927         struct btrfs_async_commit *ac;
1928         struct btrfs_transaction *cur_trans;
1929
1930         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1931         if (!ac)
1932                 return -ENOMEM;
1933
1934         INIT_WORK(&ac->work, do_async_commit);
1935         ac->newtrans = btrfs_join_transaction(trans->root);
1936         if (IS_ERR(ac->newtrans)) {
1937                 int err = PTR_ERR(ac->newtrans);
1938                 kfree(ac);
1939                 return err;
1940         }
1941
1942         /* take transaction reference */
1943         cur_trans = trans->transaction;
1944         refcount_inc(&cur_trans->use_count);
1945
1946         btrfs_end_transaction(trans);
1947
1948         /*
1949          * Tell lockdep we've released the freeze rwsem, since the
1950          * async commit thread will be the one to unlock it.
1951          */
1952         if (ac->newtrans->type & __TRANS_FREEZABLE)
1953                 __sb_writers_release(fs_info->sb, SB_FREEZE_FS);
1954
1955         schedule_work(&ac->work);
1956
1957         /* wait for transaction to start and unblock */
1958         if (wait_for_unblock)
1959                 wait_current_trans_commit_start_and_unblock(fs_info, cur_trans);
1960         else
1961                 wait_current_trans_commit_start(fs_info, cur_trans);
1962
1963         if (current->journal_info == trans)
1964                 current->journal_info = NULL;
1965
1966         btrfs_put_transaction(cur_trans);
1967         return 0;
1968 }
1969
1970
1971 static void cleanup_transaction(struct btrfs_trans_handle *trans, int err)
1972 {
1973         struct btrfs_fs_info *fs_info = trans->fs_info;
1974         struct btrfs_transaction *cur_trans = trans->transaction;
1975
1976         WARN_ON(refcount_read(&trans->use_count) > 1);
1977
1978         btrfs_abort_transaction(trans, err);
1979
1980         spin_lock(&fs_info->trans_lock);
1981
1982         /*
1983          * If the transaction is removed from the list, it means this
1984          * transaction has been committed successfully, so it is impossible
1985          * to call the cleanup function.
1986          */
1987         BUG_ON(list_empty(&cur_trans->list));
1988
1989         if (cur_trans == fs_info->running_transaction) {
1990                 cur_trans->state = TRANS_STATE_COMMIT_DOING;
1991                 spin_unlock(&fs_info->trans_lock);
1992                 wait_event(cur_trans->writer_wait,
1993                            atomic_read(&cur_trans->num_writers) == 1);
1994
1995                 spin_lock(&fs_info->trans_lock);
1996         }
1997
1998         /*
1999          * Now that we know no one else is still using the transaction we can
2000          * remove the transaction from the list of transactions. This avoids
2001          * the transaction kthread from cleaning up the transaction while some
2002          * other task is still using it, which could result in a use-after-free
2003          * on things like log trees, as it forces the transaction kthread to
2004          * wait for this transaction to be cleaned up by us.
2005          */
2006         list_del_init(&cur_trans->list);
2007
2008         spin_unlock(&fs_info->trans_lock);
2009
2010         btrfs_cleanup_one_transaction(trans->transaction, fs_info);
2011
2012         spin_lock(&fs_info->trans_lock);
2013         if (cur_trans == fs_info->running_transaction)
2014                 fs_info->running_transaction = NULL;
2015         spin_unlock(&fs_info->trans_lock);
2016
2017         if (trans->type & __TRANS_FREEZABLE)
2018                 sb_end_intwrite(fs_info->sb);
2019         btrfs_put_transaction(cur_trans);
2020         btrfs_put_transaction(cur_trans);
2021
2022         trace_btrfs_transaction_commit(trans->root);
2023
2024         if (current->journal_info == trans)
2025                 current->journal_info = NULL;
2026         btrfs_scrub_cancel(fs_info);
2027
2028         kmem_cache_free(btrfs_trans_handle_cachep, trans);
2029 }
2030
2031 /*
2032  * Release reserved delayed ref space of all pending block groups of the
2033  * transaction and remove them from the list
2034  */
2035 static void btrfs_cleanup_pending_block_groups(struct btrfs_trans_handle *trans)
2036 {
2037        struct btrfs_fs_info *fs_info = trans->fs_info;
2038        struct btrfs_block_group *block_group, *tmp;
2039
2040        list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
2041                btrfs_delayed_refs_rsv_release(fs_info, 1);
2042                list_del_init(&block_group->bg_list);
2043        }
2044 }
2045
2046 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
2047 {
2048         /*
2049          * We use writeback_inodes_sb here because if we used
2050          * btrfs_start_delalloc_roots we would deadlock with fs freeze.
2051          * Currently are holding the fs freeze lock, if we do an async flush
2052          * we'll do btrfs_join_transaction() and deadlock because we need to
2053          * wait for the fs freeze lock.  Using the direct flushing we benefit
2054          * from already being in a transaction and our join_transaction doesn't
2055          * have to re-take the fs freeze lock.
2056          */
2057         if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
2058                 writeback_inodes_sb(fs_info->sb, WB_REASON_SYNC);
2059         return 0;
2060 }
2061
2062 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
2063 {
2064         if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
2065                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0, (u64)-1);
2066 }
2067
2068 int btrfs_commit_transaction(struct btrfs_trans_handle *trans)
2069 {
2070         struct btrfs_fs_info *fs_info = trans->fs_info;
2071         struct btrfs_transaction *cur_trans = trans->transaction;
2072         struct btrfs_transaction *prev_trans = NULL;
2073         int ret;
2074
2075         ASSERT(refcount_read(&trans->use_count) == 1);
2076
2077         /*
2078          * Some places just start a transaction to commit it.  We need to make
2079          * sure that if this commit fails that the abort code actually marks the
2080          * transaction as failed, so set trans->dirty to make the abort code do
2081          * the right thing.
2082          */
2083         trans->dirty = true;
2084
2085         /* Stop the commit early if ->aborted is set */
2086         if (TRANS_ABORTED(cur_trans)) {
2087                 ret = cur_trans->aborted;
2088                 btrfs_end_transaction(trans);
2089                 return ret;
2090         }
2091
2092         btrfs_trans_release_metadata(trans);
2093         trans->block_rsv = NULL;
2094
2095         /*
2096          * We only want one transaction commit doing the flushing so we do not
2097          * waste a bunch of time on lock contention on the extent root node.
2098          */
2099         if (!test_and_set_bit(BTRFS_DELAYED_REFS_FLUSHING,
2100                               &cur_trans->delayed_refs.flags)) {
2101                 /*
2102                  * Make a pass through all the delayed refs we have so far.
2103                  * Any running threads may add more while we are here.
2104                  */
2105                 ret = btrfs_run_delayed_refs(trans, 0);
2106                 if (ret) {
2107                         btrfs_end_transaction(trans);
2108                         return ret;
2109                 }
2110         }
2111
2112         btrfs_create_pending_block_groups(trans);
2113
2114         if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) {
2115                 int run_it = 0;
2116
2117                 /* this mutex is also taken before trying to set
2118                  * block groups readonly.  We need to make sure
2119                  * that nobody has set a block group readonly
2120                  * after a extents from that block group have been
2121                  * allocated for cache files.  btrfs_set_block_group_ro
2122                  * will wait for the transaction to commit if it
2123                  * finds BTRFS_TRANS_DIRTY_BG_RUN set.
2124                  *
2125                  * The BTRFS_TRANS_DIRTY_BG_RUN flag is also used to make sure
2126                  * only one process starts all the block group IO.  It wouldn't
2127                  * hurt to have more than one go through, but there's no
2128                  * real advantage to it either.
2129                  */
2130                 mutex_lock(&fs_info->ro_block_group_mutex);
2131                 if (!test_and_set_bit(BTRFS_TRANS_DIRTY_BG_RUN,
2132                                       &cur_trans->flags))
2133                         run_it = 1;
2134                 mutex_unlock(&fs_info->ro_block_group_mutex);
2135
2136                 if (run_it) {
2137                         ret = btrfs_start_dirty_block_groups(trans);
2138                         if (ret) {
2139                                 btrfs_end_transaction(trans);
2140                                 return ret;
2141                         }
2142                 }
2143         }
2144
2145         spin_lock(&fs_info->trans_lock);
2146         if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
2147                 enum btrfs_trans_state want_state = TRANS_STATE_COMPLETED;
2148
2149                 spin_unlock(&fs_info->trans_lock);
2150                 refcount_inc(&cur_trans->use_count);
2151
2152                 if (trans->in_fsync)
2153                         want_state = TRANS_STATE_SUPER_COMMITTED;
2154                 ret = btrfs_end_transaction(trans);
2155                 wait_for_commit(cur_trans, want_state);
2156
2157                 if (TRANS_ABORTED(cur_trans))
2158                         ret = cur_trans->aborted;
2159
2160                 btrfs_put_transaction(cur_trans);
2161
2162                 return ret;
2163         }
2164
2165         cur_trans->state = TRANS_STATE_COMMIT_START;
2166         wake_up(&fs_info->transaction_blocked_wait);
2167
2168         if (cur_trans->list.prev != &fs_info->trans_list) {
2169                 enum btrfs_trans_state want_state = TRANS_STATE_COMPLETED;
2170
2171                 if (trans->in_fsync)
2172                         want_state = TRANS_STATE_SUPER_COMMITTED;
2173
2174                 prev_trans = list_entry(cur_trans->list.prev,
2175                                         struct btrfs_transaction, list);
2176                 if (prev_trans->state < want_state) {
2177                         refcount_inc(&prev_trans->use_count);
2178                         spin_unlock(&fs_info->trans_lock);
2179
2180                         wait_for_commit(prev_trans, want_state);
2181
2182                         ret = READ_ONCE(prev_trans->aborted);
2183
2184                         btrfs_put_transaction(prev_trans);
2185                         if (ret)
2186                                 goto cleanup_transaction;
2187                 } else {
2188                         spin_unlock(&fs_info->trans_lock);
2189                 }
2190         } else {
2191                 spin_unlock(&fs_info->trans_lock);
2192                 /*
2193                  * The previous transaction was aborted and was already removed
2194                  * from the list of transactions at fs_info->trans_list. So we
2195                  * abort to prevent writing a new superblock that reflects a
2196                  * corrupt state (pointing to trees with unwritten nodes/leafs).
2197                  */
2198                 if (test_bit(BTRFS_FS_STATE_TRANS_ABORTED, &fs_info->fs_state)) {
2199                         ret = -EROFS;
2200                         goto cleanup_transaction;
2201                 }
2202         }
2203
2204         extwriter_counter_dec(cur_trans, trans->type);
2205
2206         ret = btrfs_start_delalloc_flush(fs_info);
2207         if (ret)
2208                 goto cleanup_transaction;
2209
2210         ret = btrfs_run_delayed_items(trans);
2211         if (ret)
2212                 goto cleanup_transaction;
2213
2214         wait_event(cur_trans->writer_wait,
2215                    extwriter_counter_read(cur_trans) == 0);
2216
2217         /* some pending stuffs might be added after the previous flush. */
2218         ret = btrfs_run_delayed_items(trans);
2219         if (ret)
2220                 goto cleanup_transaction;
2221
2222         btrfs_wait_delalloc_flush(fs_info);
2223
2224         /*
2225          * Wait for all ordered extents started by a fast fsync that joined this
2226          * transaction. Otherwise if this transaction commits before the ordered
2227          * extents complete we lose logged data after a power failure.
2228          */
2229         wait_event(cur_trans->pending_wait,
2230                    atomic_read(&cur_trans->pending_ordered) == 0);
2231
2232         btrfs_scrub_pause(fs_info);
2233         /*
2234          * Ok now we need to make sure to block out any other joins while we
2235          * commit the transaction.  We could have started a join before setting
2236          * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
2237          */
2238         spin_lock(&fs_info->trans_lock);
2239         cur_trans->state = TRANS_STATE_COMMIT_DOING;
2240         spin_unlock(&fs_info->trans_lock);
2241         wait_event(cur_trans->writer_wait,
2242                    atomic_read(&cur_trans->num_writers) == 1);
2243
2244         if (TRANS_ABORTED(cur_trans)) {
2245                 ret = cur_trans->aborted;
2246                 goto scrub_continue;
2247         }
2248         /*
2249          * the reloc mutex makes sure that we stop
2250          * the balancing code from coming in and moving
2251          * extents around in the middle of the commit
2252          */
2253         mutex_lock(&fs_info->reloc_mutex);
2254
2255         /*
2256          * We needn't worry about the delayed items because we will
2257          * deal with them in create_pending_snapshot(), which is the
2258          * core function of the snapshot creation.
2259          */
2260         ret = create_pending_snapshots(trans);
2261         if (ret)
2262                 goto unlock_reloc;
2263
2264         /*
2265          * We insert the dir indexes of the snapshots and update the inode
2266          * of the snapshots' parents after the snapshot creation, so there
2267          * are some delayed items which are not dealt with. Now deal with
2268          * them.
2269          *
2270          * We needn't worry that this operation will corrupt the snapshots,
2271          * because all the tree which are snapshoted will be forced to COW
2272          * the nodes and leaves.
2273          */
2274         ret = btrfs_run_delayed_items(trans);
2275         if (ret)
2276                 goto unlock_reloc;
2277
2278         ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
2279         if (ret)
2280                 goto unlock_reloc;
2281
2282         /*
2283          * make sure none of the code above managed to slip in a
2284          * delayed item
2285          */
2286         btrfs_assert_delayed_root_empty(fs_info);
2287
2288         WARN_ON(cur_trans != trans->transaction);
2289
2290         /* btrfs_commit_tree_roots is responsible for getting the
2291          * various roots consistent with each other.  Every pointer
2292          * in the tree of tree roots has to point to the most up to date
2293          * root for every subvolume and other tree.  So, we have to keep
2294          * the tree logging code from jumping in and changing any
2295          * of the trees.
2296          *
2297          * At this point in the commit, there can't be any tree-log
2298          * writers, but a little lower down we drop the trans mutex
2299          * and let new people in.  By holding the tree_log_mutex
2300          * from now until after the super is written, we avoid races
2301          * with the tree-log code.
2302          */
2303         mutex_lock(&fs_info->tree_log_mutex);
2304
2305         ret = commit_fs_roots(trans);
2306         if (ret)
2307                 goto unlock_tree_log;
2308
2309         /*
2310          * Since the transaction is done, we can apply the pending changes
2311          * before the next transaction.
2312          */
2313         btrfs_apply_pending_changes(fs_info);
2314
2315         /* commit_fs_roots gets rid of all the tree log roots, it is now
2316          * safe to free the root of tree log roots
2317          */
2318         btrfs_free_log_root_tree(trans, fs_info);
2319
2320         /*
2321          * Since fs roots are all committed, we can get a quite accurate
2322          * new_roots. So let's do quota accounting.
2323          */
2324         ret = btrfs_qgroup_account_extents(trans);
2325         if (ret < 0)
2326                 goto unlock_tree_log;
2327
2328         ret = commit_cowonly_roots(trans);
2329         if (ret)
2330                 goto unlock_tree_log;
2331
2332         /*
2333          * The tasks which save the space cache and inode cache may also
2334          * update ->aborted, check it.
2335          */
2336         if (TRANS_ABORTED(cur_trans)) {
2337                 ret = cur_trans->aborted;
2338                 goto unlock_tree_log;
2339         }
2340
2341         cur_trans = fs_info->running_transaction;
2342
2343         btrfs_set_root_node(&fs_info->tree_root->root_item,
2344                             fs_info->tree_root->node);
2345         list_add_tail(&fs_info->tree_root->dirty_list,
2346                       &cur_trans->switch_commits);
2347
2348         btrfs_set_root_node(&fs_info->chunk_root->root_item,
2349                             fs_info->chunk_root->node);
2350         list_add_tail(&fs_info->chunk_root->dirty_list,
2351                       &cur_trans->switch_commits);
2352
2353         switch_commit_roots(trans);
2354
2355         ASSERT(list_empty(&cur_trans->dirty_bgs));
2356         ASSERT(list_empty(&cur_trans->io_bgs));
2357         update_super_roots(fs_info);
2358
2359         btrfs_set_super_log_root(fs_info->super_copy, 0);
2360         btrfs_set_super_log_root_level(fs_info->super_copy, 0);
2361         memcpy(fs_info->super_for_commit, fs_info->super_copy,
2362                sizeof(*fs_info->super_copy));
2363
2364         btrfs_commit_device_sizes(cur_trans);
2365
2366         clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
2367         clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
2368
2369         btrfs_trans_release_chunk_metadata(trans);
2370
2371         spin_lock(&fs_info->trans_lock);
2372         cur_trans->state = TRANS_STATE_UNBLOCKED;
2373         fs_info->running_transaction = NULL;
2374         spin_unlock(&fs_info->trans_lock);
2375         mutex_unlock(&fs_info->reloc_mutex);
2376
2377         wake_up(&fs_info->transaction_wait);
2378
2379         ret = btrfs_write_and_wait_transaction(trans);
2380         if (ret) {
2381                 btrfs_handle_fs_error(fs_info, ret,
2382                                       "Error while writing out transaction");
2383                 /*
2384                  * reloc_mutex has been unlocked, tree_log_mutex is still held
2385                  * but we can't jump to unlock_tree_log causing double unlock
2386                  */
2387                 mutex_unlock(&fs_info->tree_log_mutex);
2388                 goto scrub_continue;
2389         }
2390
2391         /*
2392          * At this point, we should have written all the tree blocks allocated
2393          * in this transaction. So it's now safe to free the redirtyied extent
2394          * buffers.
2395          */
2396         btrfs_free_redirty_list(cur_trans);
2397
2398         ret = write_all_supers(fs_info, 0);
2399         /*
2400          * the super is written, we can safely allow the tree-loggers
2401          * to go about their business
2402          */
2403         mutex_unlock(&fs_info->tree_log_mutex);
2404         if (ret)
2405                 goto scrub_continue;
2406
2407         /*
2408          * We needn't acquire the lock here because there is no other task
2409          * which can change it.
2410          */
2411         cur_trans->state = TRANS_STATE_SUPER_COMMITTED;
2412         wake_up(&cur_trans->commit_wait);
2413
2414         btrfs_finish_extent_commit(trans);
2415
2416         if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &cur_trans->flags))
2417                 btrfs_clear_space_info_full(fs_info);
2418
2419         fs_info->last_trans_committed = cur_trans->transid;
2420         /*
2421          * We needn't acquire the lock here because there is no other task
2422          * which can change it.
2423          */
2424         cur_trans->state = TRANS_STATE_COMPLETED;
2425         wake_up(&cur_trans->commit_wait);
2426
2427         spin_lock(&fs_info->trans_lock);
2428         list_del_init(&cur_trans->list);
2429         spin_unlock(&fs_info->trans_lock);
2430
2431         btrfs_put_transaction(cur_trans);
2432         btrfs_put_transaction(cur_trans);
2433
2434         if (trans->type & __TRANS_FREEZABLE)
2435                 sb_end_intwrite(fs_info->sb);
2436
2437         trace_btrfs_transaction_commit(trans->root);
2438
2439         btrfs_scrub_continue(fs_info);
2440
2441         if (current->journal_info == trans)
2442                 current->journal_info = NULL;
2443
2444         kmem_cache_free(btrfs_trans_handle_cachep, trans);
2445
2446         return ret;
2447
2448 unlock_tree_log:
2449         mutex_unlock(&fs_info->tree_log_mutex);
2450 unlock_reloc:
2451         mutex_unlock(&fs_info->reloc_mutex);
2452 scrub_continue:
2453         btrfs_scrub_continue(fs_info);
2454 cleanup_transaction:
2455         btrfs_trans_release_metadata(trans);
2456         btrfs_cleanup_pending_block_groups(trans);
2457         btrfs_trans_release_chunk_metadata(trans);
2458         trans->block_rsv = NULL;
2459         btrfs_warn(fs_info, "Skipping commit of aborted transaction.");
2460         if (current->journal_info == trans)
2461                 current->journal_info = NULL;
2462         cleanup_transaction(trans, ret);
2463
2464         return ret;
2465 }
2466
2467 /*
2468  * return < 0 if error
2469  * 0 if there are no more dead_roots at the time of call
2470  * 1 there are more to be processed, call me again
2471  *
2472  * The return value indicates there are certainly more snapshots to delete, but
2473  * if there comes a new one during processing, it may return 0. We don't mind,
2474  * because btrfs_commit_super will poke cleaner thread and it will process it a
2475  * few seconds later.
2476  */
2477 int btrfs_clean_one_deleted_snapshot(struct btrfs_root *root)
2478 {
2479         int ret;
2480         struct btrfs_fs_info *fs_info = root->fs_info;
2481
2482         spin_lock(&fs_info->trans_lock);
2483         if (list_empty(&fs_info->dead_roots)) {
2484                 spin_unlock(&fs_info->trans_lock);
2485                 return 0;
2486         }
2487         root = list_first_entry(&fs_info->dead_roots,
2488                         struct btrfs_root, root_list);
2489         list_del_init(&root->root_list);
2490         spin_unlock(&fs_info->trans_lock);
2491
2492         btrfs_debug(fs_info, "cleaner removing %llu", root->root_key.objectid);
2493
2494         btrfs_kill_all_delayed_nodes(root);
2495
2496         if (btrfs_header_backref_rev(root->node) <
2497                         BTRFS_MIXED_BACKREF_REV)
2498                 ret = btrfs_drop_snapshot(root, 0, 0);
2499         else
2500                 ret = btrfs_drop_snapshot(root, 1, 0);
2501
2502         btrfs_put_root(root);
2503         return (ret < 0) ? 0 : 1;
2504 }
2505
2506 void btrfs_apply_pending_changes(struct btrfs_fs_info *fs_info)
2507 {
2508         unsigned long prev;
2509         unsigned long bit;
2510
2511         prev = xchg(&fs_info->pending_changes, 0);
2512         if (!prev)
2513                 return;
2514
2515         bit = 1 << BTRFS_PENDING_COMMIT;
2516         if (prev & bit)
2517                 btrfs_debug(fs_info, "pending commit done");
2518         prev &= ~bit;
2519
2520         if (prev)
2521                 btrfs_warn(fs_info,
2522                         "unknown pending changes left 0x%lx, ignoring", prev);
2523 }