fs/btrfs/transaction.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/fs.h>
  20 #include <linux/slab.h>
  21 #include <linux/sched.h>
  22 #include <linux/writeback.h>
  23 #include <linux/pagemap.h>
  24 #include <linux/blkdev.h>
  25 #include <linux/uuid.h>
  26 #include "ctree.h"
  27 #include "disk-io.h"
  28 #include "transaction.h"
  29 #include "locking.h"
  30 #include "tree-log.h"
  31 #include "inode-map.h"
  32 #include "volumes.h"
  33 #include "dev-replace.h"
  34 #include "qgroup.h"
  35
  36 #define BTRFS_ROOT_TRANS_TAG 0
  37
  38 static const unsigned int btrfs_blocked_trans_types[TRANS_STATE_MAX] = {
  39         [TRANS_STATE_RUNNING]           = 0U,
  40         [TRANS_STATE_BLOCKED]           = (__TRANS_USERSPACE |
  41                                            __TRANS_START),
  42         [TRANS_STATE_COMMIT_START]      = (__TRANS_USERSPACE |
  43                                            __TRANS_START |
  44                                            __TRANS_ATTACH),
  45         [TRANS_STATE_COMMIT_DOING]      = (__TRANS_USERSPACE |
  46                                            __TRANS_START |
  47                                            __TRANS_ATTACH |
  48                                            __TRANS_JOIN),
  49         [TRANS_STATE_UNBLOCKED]         = (__TRANS_USERSPACE |
  50                                            __TRANS_START |
  51                                            __TRANS_ATTACH |
  52                                            __TRANS_JOIN |
  53                                            __TRANS_JOIN_NOLOCK),
  54         [TRANS_STATE_COMPLETED]         = (__TRANS_USERSPACE |
  55                                            __TRANS_START |
  56                                            __TRANS_ATTACH |
  57                                            __TRANS_JOIN |
  58                                            __TRANS_JOIN_NOLOCK),
  59 };
  60
  61 void btrfs_put_transaction(struct btrfs_transaction *transaction)
  62 {
  63         WARN_ON(atomic_read(&transaction->use_count) == 0);
  64         if (atomic_dec_and_test(&transaction->use_count)) {
  65                 BUG_ON(!list_empty(&transaction->list));
  66                 WARN_ON(!RB_EMPTY_ROOT(&transaction->delayed_refs.href_root));
  67                 if (transaction->delayed_refs.pending_csums)
  68                         printk(KERN_ERR "pending csums is %llu\n",
  69                                transaction->delayed_refs.pending_csums);
  70                 while (!list_empty(&transaction->pending_chunks)) {
  71                         struct extent_map *em;
  72
  73                         em = list_first_entry(&transaction->pending_chunks,
  74                                               struct extent_map, list);
  75                         list_del_init(&em->list);
  76                         free_extent_map(em);
  77                 }
  78                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  79         }
  80 }
  81
  82 static void clear_btree_io_tree(struct extent_io_tree *tree)
  83 {
  84         spin_lock(&tree->lock);
  85         /*
  86          * Do a single barrier for the waitqueue_active check here, the state
  87          * of the waitqueue should not change once clear_btree_io_tree is
  88          * called.
  89          */
  90         smp_mb();
  91         while (!RB_EMPTY_ROOT(&tree->state)) {
  92                 struct rb_node *node;
  93                 struct extent_state *state;
  94
  95                 node = rb_first(&tree->state);
  96                 state = rb_entry(node, struct extent_state, rb_node);
  97                 rb_erase(&state->rb_node, &tree->state);
  98                 RB_CLEAR_NODE(&state->rb_node);
  99                 /*
 100                  * btree io trees aren't supposed to have tasks waiting for
 101                  * changes in the flags of extent states ever.
 102                  */
 103                 ASSERT(!waitqueue_active(&state->wq));
 104                 free_extent_state(state);
 105
 106                 cond_resched_lock(&tree->lock);
 107         }
 108         spin_unlock(&tree->lock);
 109 }
 110
 111 static noinline void switch_commit_roots(struct btrfs_transaction *trans,
 112                                          struct btrfs_fs_info *fs_info)
 113 {
 114         struct btrfs_root *root, *tmp;
 115
 116         down_write(&fs_info->commit_root_sem);
 117         list_for_each_entry_safe(root, tmp, &trans->switch_commits,
 118                                  dirty_list) {
 119                 list_del_init(&root->dirty_list);
 120                 free_extent_buffer(root->commit_root);
 121                 root->commit_root = btrfs_root_node(root);
 122                 if (is_fstree(root->objectid))
 123                         btrfs_unpin_free_ino(root);
 124                 clear_btree_io_tree(&root->dirty_log_pages);
 125         }
 126
 127         /* We can free old roots now. */
 128         spin_lock(&trans->dropped_roots_lock);
 129         while (!list_empty(&trans->dropped_roots)) {
 130                 root = list_first_entry(&trans->dropped_roots,
 131                                         struct btrfs_root, root_list);
 132                 list_del_init(&root->root_list);
 133                 spin_unlock(&trans->dropped_roots_lock);
 134                 btrfs_drop_and_free_fs_root(fs_info, root);
 135                 spin_lock(&trans->dropped_roots_lock);
 136         }
 137         spin_unlock(&trans->dropped_roots_lock);
 138         up_write(&fs_info->commit_root_sem);
 139 }
 140
 141 static inline void extwriter_counter_inc(struct btrfs_transaction *trans,
 142                                          unsigned int type)
 143 {
 144         if (type & TRANS_EXTWRITERS)
 145                 atomic_inc(&trans->num_extwriters);
 146 }
 147
 148 static inline void extwriter_counter_dec(struct btrfs_transaction *trans,
 149                                          unsigned int type)
 150 {
 151         if (type & TRANS_EXTWRITERS)
 152                 atomic_dec(&trans->num_extwriters);
 153 }
 154
 155 static inline void extwriter_counter_init(struct btrfs_transaction *trans,
 156                                           unsigned int type)
 157 {
 158         atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0));
 159 }
 160
 161 static inline int extwriter_counter_read(struct btrfs_transaction *trans)
 162 {
 163         return atomic_read(&trans->num_extwriters);
 164 }
 165
 166 /*
 167  * either allocate a new transaction or hop into the existing one
 168  */
 169 static noinline int join_transaction(struct btrfs_root *root, unsigned int type)
 170 {
 171         struct btrfs_transaction *cur_trans;
 172         struct btrfs_fs_info *fs_info = root->fs_info;
 173
 174         spin_lock(&fs_info->trans_lock);
 175 loop:
 176         /* The file system has been taken offline. No new transactions. */
 177         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
 178                 spin_unlock(&fs_info->trans_lock);
 179                 return -EROFS;
 180         }
 181
 182         cur_trans = fs_info->running_transaction;
 183         if (cur_trans) {
 184                 if (cur_trans->aborted) {
 185                         spin_unlock(&fs_info->trans_lock);
 186                         return cur_trans->aborted;
 187                 }
 188                 if (btrfs_blocked_trans_types[cur_trans->state] & type) {
 189                         spin_unlock(&fs_info->trans_lock);
 190                         return -EBUSY;
 191                 }
 192                 atomic_inc(&cur_trans->use_count);
 193                 atomic_inc(&cur_trans->num_writers);
 194                 extwriter_counter_inc(cur_trans, type);
 195                 spin_unlock(&fs_info->trans_lock);
 196                 return 0;
 197         }
 198         spin_unlock(&fs_info->trans_lock);
 199
 200         /*
 201          * If we are ATTACH, we just want to catch the current transaction,
 202          * and commit it. If there is no transaction, just return ENOENT.
 203          */
 204         if (type == TRANS_ATTACH)
 205                 return -ENOENT;
 206
 207         /*
 208          * JOIN_NOLOCK only happens during the transaction commit, so
 209          * it is impossible that ->running_transaction is NULL
 210          */
 211         BUG_ON(type == TRANS_JOIN_NOLOCK);
 212
 213         cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
 214         if (!cur_trans)
 215                 return -ENOMEM;
 216
 217         spin_lock(&fs_info->trans_lock);
 218         if (fs_info->running_transaction) {
 219                 /*
 220                  * someone started a transaction after we unlocked.  Make sure
 221                  * to redo the checks above
 222                  */
 223                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 224                 goto loop;
 225         } else if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
 226                 spin_unlock(&fs_info->trans_lock);
 227                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 228                 return -EROFS;
 229         }
 230
 231         atomic_set(&cur_trans->num_writers, 1);
 232         extwriter_counter_init(cur_trans, type);
 233         init_waitqueue_head(&cur_trans->writer_wait);
 234         init_waitqueue_head(&cur_trans->commit_wait);
 235         cur_trans->state = TRANS_STATE_RUNNING;
 236         /*
 237          * One for this trans handle, one so it will live on until we
 238          * commit the transaction.
 239          */
 240         atomic_set(&cur_trans->use_count, 2);
 241         cur_trans->have_free_bgs = 0;
 242         cur_trans->start_time = get_seconds();
 243         cur_trans->dirty_bg_run = 0;
 244
 245         cur_trans->delayed_refs.href_root = RB_ROOT;
 246         cur_trans->delayed_refs.dirty_extent_root = RB_ROOT;
 247         atomic_set(&cur_trans->delayed_refs.num_entries, 0);
 248         cur_trans->delayed_refs.num_heads_ready = 0;
 249         cur_trans->delayed_refs.pending_csums = 0;
 250         cur_trans->delayed_refs.num_heads = 0;
 251         cur_trans->delayed_refs.flushing = 0;
 252         cur_trans->delayed_refs.run_delayed_start = 0;
 253         cur_trans->delayed_refs.qgroup_to_skip = 0;
 254
 255         /*
 256          * although the tree mod log is per file system and not per transaction,
 257          * the log must never go across transaction boundaries.
 258          */
 259         smp_mb();
 260         if (!list_empty(&fs_info->tree_mod_seq_list))
 261                 WARN(1, KERN_ERR "BTRFS: tree_mod_seq_list not empty when "
 262                         "creating a fresh transaction\n");
 263         if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
 264                 WARN(1, KERN_ERR "BTRFS: tree_mod_log rb tree not empty when "
 265                         "creating a fresh transaction\n");
 266         atomic64_set(&fs_info->tree_mod_seq, 0);
 267
 268         spin_lock_init(&cur_trans->delayed_refs.lock);
 269
 270         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 271         INIT_LIST_HEAD(&cur_trans->pending_chunks);
 272         INIT_LIST_HEAD(&cur_trans->switch_commits);
 273         INIT_LIST_HEAD(&cur_trans->pending_ordered);
 274         INIT_LIST_HEAD(&cur_trans->dirty_bgs);
 275         INIT_LIST_HEAD(&cur_trans->io_bgs);
 276         INIT_LIST_HEAD(&cur_trans->dropped_roots);
 277         mutex_init(&cur_trans->cache_write_mutex);
 278         cur_trans->num_dirty_bgs = 0;
 279         spin_lock_init(&cur_trans->dirty_bgs_lock);
 280         INIT_LIST_HEAD(&cur_trans->deleted_bgs);
 281         spin_lock_init(&cur_trans->deleted_bgs_lock);
 282         spin_lock_init(&cur_trans->dropped_roots_lock);
 283         list_add_tail(&cur_trans->list, &fs_info->trans_list);
 284         extent_io_tree_init(&cur_trans->dirty_pages,
 285                              fs_info->btree_inode->i_mapping);
 286         fs_info->generation++;
 287         cur_trans->transid = fs_info->generation;
 288         fs_info->running_transaction = cur_trans;
 289         cur_trans->aborted = 0;
 290         spin_unlock(&fs_info->trans_lock);
 291
 292         return 0;
 293 }
 294
 295 /*
 296  * this does all the record keeping required to make sure that a reference
 297  * counted root is properly recorded in a given transaction.  This is required
 298  * to make sure the old root from before we joined the transaction is deleted
 299  * when the transaction commits
 300  */
 301 static int record_root_in_trans(struct btrfs_trans_handle *trans,
 302                                struct btrfs_root *root)
 303 {
 304         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
 305             root->last_trans < trans->transid) {
 306                 WARN_ON(root == root->fs_info->extent_root);
 307                 WARN_ON(root->commit_root != root->node);
 308
 309                 /*
 310                  * see below for IN_TRANS_SETUP usage rules
 311                  * we have the reloc mutex held now, so there
 312                  * is only one writer in this function
 313                  */
 314                 set_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
 315
 316                 /* make sure readers find IN_TRANS_SETUP before
 317                  * they find our root->last_trans update
 318                  */
 319                 smp_wmb();
 320
 321                 spin_lock(&root->fs_info->fs_roots_radix_lock);
 322                 if (root->last_trans == trans->transid) {
 323                         spin_unlock(&root->fs_info->fs_roots_radix_lock);
 324                         return 0;
 325                 }
 326                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 327                            (unsigned long)root->root_key.objectid,
 328                            BTRFS_ROOT_TRANS_TAG);
 329                 spin_unlock(&root->fs_info->fs_roots_radix_lock);
 330                 root->last_trans = trans->transid;
 331
 332                 /* this is pretty tricky.  We don't want to
 333                  * take the relocation lock in btrfs_record_root_in_trans
 334                  * unless we're really doing the first setup for this root in
 335                  * this transaction.
 336                  *
 337                  * Normally we'd use root->last_trans as a flag to decide
 338                  * if we want to take the expensive mutex.
 339                  *
 340                  * But, we have to set root->last_trans before we
 341                  * init the relocation root, otherwise, we trip over warnings
 342                  * in ctree.c.  The solution used here is to flag ourselves
 343                  * with root IN_TRANS_SETUP.  When this is 1, we're still
 344                  * fixing up the reloc trees and everyone must wait.
 345                  *
 346                  * When this is zero, they can trust root->last_trans and fly
 347                  * through btrfs_record_root_in_trans without having to take the
 348                  * lock.  smp_wmb() makes sure that all the writes above are
 349                  * done before we pop in the zero below
 350                  */
 351                 btrfs_init_reloc_root(trans, root);
 352                 smp_mb__before_atomic();
 353                 clear_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
 354         }
 355         return 0;
 356 }
 357
 358
 359 void btrfs_add_dropped_root(struct btrfs_trans_handle *trans,
 360                             struct btrfs_root *root)
 361 {
 362         struct btrfs_transaction *cur_trans = trans->transaction;
 363
 364         /* Add ourselves to the transaction dropped list */
 365         spin_lock(&cur_trans->dropped_roots_lock);
 366         list_add_tail(&root->root_list, &cur_trans->dropped_roots);
 367         spin_unlock(&cur_trans->dropped_roots_lock);
 368
 369         /* Make sure we don't try to update the root at commit time */
 370         spin_lock(&root->fs_info->fs_roots_radix_lock);
 371         radix_tree_tag_clear(&root->fs_info->fs_roots_radix,
 372                              (unsigned long)root->root_key.objectid,
 373                              BTRFS_ROOT_TRANS_TAG);
 374         spin_unlock(&root->fs_info->fs_roots_radix_lock);
 375 }
 376
 377 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 378                                struct btrfs_root *root)
 379 {
 380         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
 381                 return 0;
 382
 383         /*
 384          * see record_root_in_trans for comments about IN_TRANS_SETUP usage
 385          * and barriers
 386          */
 387         smp_rmb();
 388         if (root->last_trans == trans->transid &&
 389             !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state))
 390                 return 0;
 391
 392         mutex_lock(&root->fs_info->reloc_mutex);
 393         record_root_in_trans(trans, root);
 394         mutex_unlock(&root->fs_info->reloc_mutex);
 395
 396         return 0;
 397 }
 398
 399 static inline int is_transaction_blocked(struct btrfs_transaction *trans)
 400 {
 401         return (trans->state >= TRANS_STATE_BLOCKED &&
 402                 trans->state < TRANS_STATE_UNBLOCKED &&
 403                 !trans->aborted);
 404 }
 405
 406 /* wait for commit against the current transaction to become unblocked
 407  * when this is done, it is safe to start a new transaction, but the current
 408  * transaction might not be fully on disk.
 409  */
 410 static void wait_current_trans(struct btrfs_root *root)
 411 {
 412         struct btrfs_transaction *cur_trans;
 413
 414         spin_lock(&root->fs_info->trans_lock);
 415         cur_trans = root->fs_info->running_transaction;
 416         if (cur_trans && is_transaction_blocked(cur_trans)) {
 417                 atomic_inc(&cur_trans->use_count);
 418                 spin_unlock(&root->fs_info->trans_lock);
 419
 420                 wait_event(root->fs_info->transaction_wait,
 421                            cur_trans->state >= TRANS_STATE_UNBLOCKED ||
 422                            cur_trans->aborted);
 423                 btrfs_put_transaction(cur_trans);
 424         } else {
 425                 spin_unlock(&root->fs_info->trans_lock);
 426         }
 427 }
 428
 429 static int may_wait_transaction(struct btrfs_root *root, int type)
 430 {
 431         if (root->fs_info->log_root_recovering)
 432                 return 0;
 433
 434         if (type == TRANS_USERSPACE)
 435                 return 1;
 436
 437         if (type == TRANS_START &&
 438             !atomic_read(&root->fs_info->open_ioctl_trans))
 439                 return 1;
 440
 441         return 0;
 442 }
 443
 444 static inline bool need_reserve_reloc_root(struct btrfs_root *root)
 445 {
 446         if (!root->fs_info->reloc_ctl ||
 447             !test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
 448             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 449             root->reloc_root)
 450                 return false;
 451
 452         return true;
 453 }
 454
 455 static struct btrfs_trans_handle *
 456 start_transaction(struct btrfs_root *root, u64 num_items, unsigned int type,
 457                   enum btrfs_reserve_flush_enum flush)
 458 {
 459         struct btrfs_trans_handle *h;
 460         struct btrfs_transaction *cur_trans;
 461         u64 num_bytes = 0;
 462         u64 qgroup_reserved = 0;
 463         bool reloc_reserved = false;
 464         int ret;
 465
 466         /* Send isn't supposed to start transactions. */
 467         ASSERT(current->journal_info != BTRFS_SEND_TRANS_STUB);
 468
 469         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
 470                 return ERR_PTR(-EROFS);
 471
 472         if (current->journal_info) {
 473                 WARN_ON(type & TRANS_EXTWRITERS);
 474                 h = current->journal_info;
 475                 h->use_count++;
 476                 WARN_ON(h->use_count > 2);
 477                 h->orig_rsv = h->block_rsv;
 478                 h->block_rsv = NULL;
 479                 goto got_it;
 480         }
 481
 482         /*
 483          * Do the reservation before we join the transaction so we can do all
 484          * the appropriate flushing if need be.
 485          */
 486         if (num_items > 0 && root != root->fs_info->chunk_root) {
 487                 if (root->fs_info->quota_enabled &&
 488                     is_fstree(root->root_key.objectid)) {
 489                         qgroup_reserved = num_items * root->nodesize;
 490                         ret = btrfs_qgroup_reserve(root, qgroup_reserved);
 491                         if (ret)
 492                                 return ERR_PTR(ret);
 493                 }
 494
 495                 num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
 496                 /*
 497                  * Do the reservation for the relocation root creation
 498                  */
 499                 if (need_reserve_reloc_root(root)) {
 500                         num_bytes += root->nodesize;
 501                         reloc_reserved = true;
 502                 }
 503
 504                 ret = btrfs_block_rsv_add(root,
 505                                           &root->fs_info->trans_block_rsv,
 506                                           num_bytes, flush);
 507                 if (ret)
 508                         goto reserve_fail;
 509         }
 510 again:
 511         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 512         if (!h) {
 513                 ret = -ENOMEM;
 514                 goto alloc_fail;
 515         }
 516
 517         /*
 518          * If we are JOIN_NOLOCK we're already committing a transaction and
 519          * waiting on this guy, so we don't need to do the sb_start_intwrite
 520          * because we're already holding a ref.  We need this because we could
 521          * have raced in and did an fsync() on a file which can kick a commit
 522          * and then we deadlock with somebody doing a freeze.
 523          *
 524          * If we are ATTACH, it means we just want to catch the current
 525          * transaction and commit it, so we needn't do sb_start_intwrite().
 526          */
 527         if (type & __TRANS_FREEZABLE)
 528                 sb_start_intwrite(root->fs_info->sb);
 529
 530         if (may_wait_transaction(root, type))
 531                 wait_current_trans(root);
 532
 533         do {
 534                 ret = join_transaction(root, type);
 535                 if (ret == -EBUSY) {
 536                         wait_current_trans(root);
 537                         if (unlikely(type == TRANS_ATTACH))
 538                                 ret = -ENOENT;
 539                 }
 540         } while (ret == -EBUSY);
 541
 542         if (ret < 0) {
 543                 /* We must get the transaction if we are JOIN_NOLOCK. */
 544                 BUG_ON(type == TRANS_JOIN_NOLOCK);
 545                 goto join_fail;
 546         }
 547
 548         cur_trans = root->fs_info->running_transaction;
 549
 550         h->transid = cur_trans->transid;
 551         h->transaction = cur_trans;
 552         h->blocks_used = 0;
 553         h->bytes_reserved = 0;
 554         h->chunk_bytes_reserved = 0;
 555         h->root = root;
 556         h->delayed_ref_updates = 0;
 557         h->use_count = 1;
 558         h->adding_csums = 0;
 559         h->block_rsv = NULL;
 560         h->orig_rsv = NULL;
 561         h->aborted = 0;
 562         h->qgroup_reserved = 0;
 563         h->delayed_ref_elem.seq = 0;
 564         h->type = type;
 565         h->allocating_chunk = false;
 566         h->can_flush_pending_bgs = true;
 567         h->reloc_reserved = false;
 568         h->sync = false;
 569         INIT_LIST_HEAD(&h->qgroup_ref_list);
 570         INIT_LIST_HEAD(&h->new_bgs);
 571         INIT_LIST_HEAD(&h->ordered);
 572
 573         smp_mb();
 574         if (cur_trans->state >= TRANS_STATE_BLOCKED &&
 575             may_wait_transaction(root, type)) {
 576                 current->journal_info = h;
 577                 btrfs_commit_transaction(h, root);
 578                 goto again;
 579         }
 580
 581         if (num_bytes) {
 582                 trace_btrfs_space_reservation(root->fs_info, "transaction",
 583                                               h->transid, num_bytes, 1);
 584                 h->block_rsv = &root->fs_info->trans_block_rsv;
 585                 h->bytes_reserved = num_bytes;
 586                 h->reloc_reserved = reloc_reserved;
 587         }
 588         h->qgroup_reserved = qgroup_reserved;
 589
 590 got_it:
 591         btrfs_record_root_in_trans(h, root);
 592
 593         if (!current->journal_info && type != TRANS_USERSPACE)
 594                 current->journal_info = h;
 595         return h;
 596
 597 join_fail:
 598         if (type & __TRANS_FREEZABLE)
 599                 sb_end_intwrite(root->fs_info->sb);
 600         kmem_cache_free(btrfs_trans_handle_cachep, h);
 601 alloc_fail:
 602         if (num_bytes)
 603                 btrfs_block_rsv_release(root, &root->fs_info->trans_block_rsv,
 604                                         num_bytes);
 605 reserve_fail:
 606         if (qgroup_reserved)
 607                 btrfs_qgroup_free(root, qgroup_reserved);
 608         return ERR_PTR(ret);
 609 }
 610
 611 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 612                                                    int num_items)
 613 {
 614         return start_transaction(root, num_items, TRANS_START,
 615                                  BTRFS_RESERVE_FLUSH_ALL);
 616 }
 617
 618 struct btrfs_trans_handle *btrfs_start_transaction_lflush(
 619                                         struct btrfs_root *root, int num_items)
 620 {
 621         return start_transaction(root, num_items, TRANS_START,
 622                                  BTRFS_RESERVE_FLUSH_LIMIT);
 623 }
 624
 625 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 626 {
 627         return start_transaction(root, 0, TRANS_JOIN, 0);
 628 }
 629
 630 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 631 {
 632         return start_transaction(root, 0, TRANS_JOIN_NOLOCK, 0);
 633 }
 634
 635 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 636 {
 637         return start_transaction(root, 0, TRANS_USERSPACE, 0);
 638 }
 639
 640 /*
 641  * btrfs_attach_transaction() - catch the running transaction
 642  *
 643  * It is used when we want to commit the current the transaction, but
 644  * don't want to start a new one.
 645  *
 646  * Note: If this function return -ENOENT, it just means there is no
 647  * running transaction. But it is possible that the inactive transaction
 648  * is still in the memory, not fully on disk. If you hope there is no
 649  * inactive transaction in the fs when -ENOENT is returned, you should
 650  * invoke
 651  *     btrfs_attach_transaction_barrier()
 652  */
 653 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
 654 {
 655         return start_transaction(root, 0, TRANS_ATTACH, 0);
 656 }
 657
 658 /*
 659  * btrfs_attach_transaction_barrier() - catch the running transaction
 660  *
 661  * It is similar to the above function, the differentia is this one
 662  * will wait for all the inactive transactions until they fully
 663  * complete.
 664  */
 665 struct btrfs_trans_handle *
 666 btrfs_attach_transaction_barrier(struct btrfs_root *root)
 667 {
 668         struct btrfs_trans_handle *trans;
 669
 670         trans = start_transaction(root, 0, TRANS_ATTACH, 0);
 671         if (IS_ERR(trans) && PTR_ERR(trans) == -ENOENT)
 672                 btrfs_wait_for_commit(root, 0);
 673
 674         return trans;
 675 }
 676
 677 /* wait for a transaction commit to be fully complete */
 678 static noinline void wait_for_commit(struct btrfs_root *root,
 679                                     struct btrfs_transaction *commit)
 680 {
 681         wait_event(commit->commit_wait, commit->state == TRANS_STATE_COMPLETED);
 682 }
 683
 684 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 685 {
 686         struct btrfs_transaction *cur_trans = NULL, *t;
 687         int ret = 0;
 688
 689         if (transid) {
 690                 if (transid <= root->fs_info->last_trans_committed)
 691                         goto out;
 692
 693                 /* find specified transaction */
 694                 spin_lock(&root->fs_info->trans_lock);
 695                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 696                         if (t->transid == transid) {
 697                                 cur_trans = t;
 698                                 atomic_inc(&cur_trans->use_count);
 699                                 ret = 0;
 700                                 break;
 701                         }
 702                         if (t->transid > transid) {
 703                                 ret = 0;
 704                                 break;
 705                         }
 706                 }
 707                 spin_unlock(&root->fs_info->trans_lock);
 708
 709                 /*
 710                  * The specified transaction doesn't exist, or we
 711                  * raced with btrfs_commit_transaction
 712                  */
 713                 if (!cur_trans) {
 714                         if (transid > root->fs_info->last_trans_committed)
 715                                 ret = -EINVAL;
 716                         goto out;
 717                 }
 718         } else {
 719                 /* find newest transaction that is committing | committed */
 720                 spin_lock(&root->fs_info->trans_lock);
 721                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 722                                             list) {
 723                         if (t->state >= TRANS_STATE_COMMIT_START) {
 724                                 if (t->state == TRANS_STATE_COMPLETED)
 725                                         break;
 726                                 cur_trans = t;
 727                                 atomic_inc(&cur_trans->use_count);
 728                                 break;
 729                         }
 730                 }
 731                 spin_unlock(&root->fs_info->trans_lock);
 732                 if (!cur_trans)
 733                         goto out;  /* nothing committing|committed */
 734         }
 735
 736         wait_for_commit(root, cur_trans);
 737         btrfs_put_transaction(cur_trans);
 738 out:
 739         return ret;
 740 }
 741
 742 void btrfs_throttle(struct btrfs_root *root)
 743 {
 744         if (!atomic_read(&root->fs_info->open_ioctl_trans))
 745                 wait_current_trans(root);
 746 }
 747
 748 static int should_end_transaction(struct btrfs_trans_handle *trans,
 749                                   struct btrfs_root *root)
 750 {
 751         if (root->fs_info->global_block_rsv.space_info->full &&
 752             btrfs_check_space_for_delayed_refs(trans, root))
 753                 return 1;
 754
 755         return !!btrfs_block_rsv_check(root, &root->fs_info->global_block_rsv, 5);
 756 }
 757
 758 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 759                                  struct btrfs_root *root)
 760 {
 761         struct btrfs_transaction *cur_trans = trans->transaction;
 762         int updates;
 763         int err;
 764
 765         smp_mb();
 766         if (cur_trans->state >= TRANS_STATE_BLOCKED ||
 767             cur_trans->delayed_refs.flushing)
 768                 return 1;
 769
 770         updates = trans->delayed_ref_updates;
 771         trans->delayed_ref_updates = 0;
 772         if (updates) {
 773                 err = btrfs_run_delayed_refs(trans, root, updates * 2);
 774                 if (err) /* Error code will also eval true */
 775                         return err;
 776         }
 777
 778         return should_end_transaction(trans, root);
 779 }
 780
 781 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 782                           struct btrfs_root *root, int throttle)
 783 {
 784         struct btrfs_transaction *cur_trans = trans->transaction;
 785         struct btrfs_fs_info *info = root->fs_info;
 786         unsigned long cur = trans->delayed_ref_updates;
 787         int lock = (trans->type != TRANS_JOIN_NOLOCK);
 788         int err = 0;
 789         int must_run_delayed_refs = 0;
 790
 791         if (trans->use_count > 1) {
 792                 trans->use_count--;
 793                 trans->block_rsv = trans->orig_rsv;
 794                 return 0;
 795         }
 796
 797         btrfs_trans_release_metadata(trans, root);
 798         trans->block_rsv = NULL;
 799
 800         if (!list_empty(&trans->new_bgs))
 801                 btrfs_create_pending_block_groups(trans, root);
 802
 803         if (!list_empty(&trans->ordered)) {
 804                 spin_lock(&info->trans_lock);
 805                 list_splice_init(&trans->ordered, &cur_trans->pending_ordered);
 806                 spin_unlock(&info->trans_lock);
 807         }
 808
 809         trans->delayed_ref_updates = 0;
 810         if (!trans->sync) {
 811                 must_run_delayed_refs =
 812                         btrfs_should_throttle_delayed_refs(trans, root);
 813                 cur = max_t(unsigned long, cur, 32);
 814
 815                 /*
 816                  * don't make the caller wait if they are from a NOLOCK
 817                  * or ATTACH transaction, it will deadlock with commit
 818                  */
 819                 if (must_run_delayed_refs == 1 &&
 820                     (trans->type & (__TRANS_JOIN_NOLOCK | __TRANS_ATTACH)))
 821                         must_run_delayed_refs = 2;
 822         }
 823
 824         if (trans->qgroup_reserved) {
 825                 /*
 826                  * the same root has to be passed here between start_transaction
 827                  * and end_transaction. Subvolume quota depends on this.
 828                  */
 829                 btrfs_qgroup_free(trans->root, trans->qgroup_reserved);
 830                 trans->qgroup_reserved = 0;
 831         }
 832
 833         btrfs_trans_release_metadata(trans, root);
 834         trans->block_rsv = NULL;
 835
 836         if (!list_empty(&trans->new_bgs))
 837                 btrfs_create_pending_block_groups(trans, root);
 838
 839         btrfs_trans_release_chunk_metadata(trans);
 840
 841         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
 842             should_end_transaction(trans, root) &&
 843             ACCESS_ONCE(cur_trans->state) == TRANS_STATE_RUNNING) {
 844                 spin_lock(&info->trans_lock);
 845                 if (cur_trans->state == TRANS_STATE_RUNNING)
 846                         cur_trans->state = TRANS_STATE_BLOCKED;
 847                 spin_unlock(&info->trans_lock);
 848         }
 849
 850         if (lock && ACCESS_ONCE(cur_trans->state) == TRANS_STATE_BLOCKED) {
 851                 if (throttle)
 852                         return btrfs_commit_transaction(trans, root);
 853                 else
 854                         wake_up_process(info->transaction_kthread);
 855         }
 856
 857         if (trans->type & __TRANS_FREEZABLE)
 858                 sb_end_intwrite(root->fs_info->sb);
 859
 860         WARN_ON(cur_trans != info->running_transaction);
 861         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 862         atomic_dec(&cur_trans->num_writers);
 863         extwriter_counter_dec(cur_trans, trans->type);
 864
 865         /*
 866          * Make sure counter is updated before we wake up waiters.
 867          */
 868         smp_mb();
 869         if (waitqueue_active(&cur_trans->writer_wait))
 870                 wake_up(&cur_trans->writer_wait);
 871         btrfs_put_transaction(cur_trans);
 872
 873         if (current->journal_info == trans)
 874                 current->journal_info = NULL;
 875
 876         if (throttle)
 877                 btrfs_run_delayed_iputs(root);
 878
 879         if (trans->aborted ||
 880             test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state)) {
 881                 wake_up_process(info->transaction_kthread);
 882                 err = -EIO;
 883         }
 884         assert_qgroups_uptodate(trans);
 885
 886         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 887         if (must_run_delayed_refs) {
 888                 btrfs_async_run_delayed_refs(root, cur,
 889                                              must_run_delayed_refs == 1);
 890         }
 891         return err;
 892 }
 893
 894 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 895                           struct btrfs_root *root)
 896 {
 897         return __btrfs_end_transaction(trans, root, 0);
 898 }
 899
 900 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 901                                    struct btrfs_root *root)
 902 {
 903         return __btrfs_end_transaction(trans, root, 1);
 904 }
 905
 906 /*
 907  * when btree blocks are allocated, they have some corresponding bits set for
 908  * them in one of two extent_io trees.  This is used to make sure all of
 909  * those extents are sent to disk but does not wait on them
 910  */
 911 int btrfs_write_marked_extents(struct btrfs_root *root,
 912                                struct extent_io_tree *dirty_pages, int mark)
 913 {
 914         int err = 0;
 915         int werr = 0;
 916         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 917         struct extent_state *cached_state = NULL;
 918         u64 start = 0;
 919         u64 end;
 920
 921         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 922                                       mark, &cached_state)) {
 923                 bool wait_writeback = false;
 924
 925                 err = convert_extent_bit(dirty_pages, start, end,
 926                                          EXTENT_NEED_WAIT,
 927                                          mark, &cached_state, GFP_NOFS);
 928                 /*
 929                  * convert_extent_bit can return -ENOMEM, which is most of the
 930                  * time a temporary error. So when it happens, ignore the error
 931                  * and wait for writeback of this range to finish - because we
 932                  * failed to set the bit EXTENT_NEED_WAIT for the range, a call
 933                  * to btrfs_wait_marked_extents() would not know that writeback
 934                  * for this range started and therefore wouldn't wait for it to
 935                  * finish - we don't want to commit a superblock that points to
 936                  * btree nodes/leafs for which writeback hasn't finished yet
 937                  * (and without errors).
 938                  * We cleanup any entries left in the io tree when committing
 939                  * the transaction (through clear_btree_io_tree()).
 940                  */
 941                 if (err == -ENOMEM) {
 942                         err = 0;
 943                         wait_writeback = true;
 944                 }
 945                 if (!err)
 946                         err = filemap_fdatawrite_range(mapping, start, end);
 947                 if (err)
 948                         werr = err;
 949                 else if (wait_writeback)
 950                         werr = filemap_fdatawait_range(mapping, start, end);
 951                 free_extent_state(cached_state);
 952                 cached_state = NULL;
 953                 cond_resched();
 954                 start = end + 1;
 955         }
 956         return werr;
 957 }
 958
 959 /*
 960  * when btree blocks are allocated, they have some corresponding bits set for
 961  * them in one of two extent_io trees.  This is used to make sure all of
 962  * those extents are on disk for transaction or log commit.  We wait
 963  * on all the pages and clear them from the dirty pages state tree
 964  */
 965 int btrfs_wait_marked_extents(struct btrfs_root *root,
 966                               struct extent_io_tree *dirty_pages, int mark)
 967 {
 968         int err = 0;
 969         int werr = 0;
 970         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 971         struct extent_state *cached_state = NULL;
 972         u64 start = 0;
 973         u64 end;
 974         struct btrfs_inode *btree_ino = BTRFS_I(root->fs_info->btree_inode);
 975         bool errors = false;
 976
 977         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 978                                       EXTENT_NEED_WAIT, &cached_state)) {
 979                 /*
 980                  * Ignore -ENOMEM errors returned by clear_extent_bit().
 981                  * When committing the transaction, we'll remove any entries
 982                  * left in the io tree. For a log commit, we don't remove them
 983                  * after committing the log because the tree can be accessed
 984                  * concurrently - we do it only at transaction commit time when
 985                  * it's safe to do it (through clear_btree_io_tree()).
 986                  */
 987                 err = clear_extent_bit(dirty_pages, start, end,
 988                                        EXTENT_NEED_WAIT,
 989                                        0, 0, &cached_state, GFP_NOFS);
 990                 if (err == -ENOMEM)
 991                         err = 0;
 992                 if (!err)
 993                         err = filemap_fdatawait_range(mapping, start, end);
 994                 if (err)
 995                         werr = err;
 996                 free_extent_state(cached_state);
 997                 cached_state = NULL;
 998                 cond_resched();
 999                 start = end + 1;
1000         }
1001         if (err)
1002                 werr = err;
1003
1004         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
1005                 if ((mark & EXTENT_DIRTY) &&
1006                     test_and_clear_bit(BTRFS_INODE_BTREE_LOG1_ERR,
1007                                        &btree_ino->runtime_flags))
1008                         errors = true;
1009
1010                 if ((mark & EXTENT_NEW) &&
1011                     test_and_clear_bit(BTRFS_INODE_BTREE_LOG2_ERR,
1012                                        &btree_ino->runtime_flags))
1013                         errors = true;
1014         } else {
1015                 if (test_and_clear_bit(BTRFS_INODE_BTREE_ERR,
1016                                        &btree_ino->runtime_flags))
1017                         errors = true;
1018         }
1019
1020         if (errors && !werr)
1021                 werr = -EIO;
1022
1023         return werr;
1024 }
1025
1026 /*
1027  * when btree blocks are allocated, they have some corresponding bits set for
1028  * them in one of two extent_io trees.  This is used to make sure all of
1029  * those extents are on disk for transaction or log commit
1030  */
1031 static int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
1032                                 struct extent_io_tree *dirty_pages, int mark)
1033 {
1034         int ret;
1035         int ret2;
1036         struct blk_plug plug;
1037
1038         blk_start_plug(&plug);
1039         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
1040         blk_finish_plug(&plug);
1041         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
1042
1043         if (ret)
1044                 return ret;
1045         if (ret2)
1046                 return ret2;
1047         return 0;
1048 }
1049
1050 static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
1051                                      struct btrfs_root *root)
1052 {
1053         int ret;
1054
1055         ret = btrfs_write_and_wait_marked_extents(root,
1056                                            &trans->transaction->dirty_pages,
1057                                            EXTENT_DIRTY);
1058         clear_btree_io_tree(&trans->transaction->dirty_pages);
1059
1060         return ret;
1061 }
1062
1063 /*
1064  * this is used to update the root pointer in the tree of tree roots.
1065  *
1066  * But, in the case of the extent allocation tree, updating the root
1067  * pointer may allocate blocks which may change the root of the extent
1068  * allocation tree.
1069  *
1070  * So, this loops and repeats and makes sure the cowonly root didn't
1071  * change while the root pointer was being updated in the metadata.
1072  */
1073 static int update_cowonly_root(struct btrfs_trans_handle *trans,
1074                                struct btrfs_root *root)
1075 {
1076         int ret;
1077         u64 old_root_bytenr;
1078         u64 old_root_used;
1079         struct btrfs_root *tree_root = root->fs_info->tree_root;
1080
1081         old_root_used = btrfs_root_used(&root->root_item);
1082
1083         while (1) {
1084                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1085                 if (old_root_bytenr == root->node->start &&
1086                     old_root_used == btrfs_root_used(&root->root_item))
1087                         break;
1088
1089                 btrfs_set_root_node(&root->root_item, root->node);
1090                 ret = btrfs_update_root(trans, tree_root,
1091                                         &root->root_key,
1092                                         &root->root_item);
1093                 if (ret)
1094                         return ret;
1095
1096                 old_root_used = btrfs_root_used(&root->root_item);
1097         }
1098
1099         return 0;
1100 }
1101
1102 /*
1103  * update all the cowonly tree roots on disk
1104  *
1105  * The error handling in this function may not be obvious. Any of the
1106  * failures will cause the file system to go offline. We still need
1107  * to clean up the delayed refs.
1108  */
1109 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
1110                                          struct btrfs_root *root)
1111 {
1112         struct btrfs_fs_info *fs_info = root->fs_info;
1113         struct list_head *dirty_bgs = &trans->transaction->dirty_bgs;
1114         struct list_head *io_bgs = &trans->transaction->io_bgs;
1115         struct list_head *next;
1116         struct extent_buffer *eb;
1117         int ret;
1118
1119         eb = btrfs_lock_root_node(fs_info->tree_root);
1120         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
1121                               0, &eb);
1122         btrfs_tree_unlock(eb);
1123         free_extent_buffer(eb);
1124
1125         if (ret)
1126                 return ret;
1127
1128         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1129         if (ret)
1130                 return ret;
1131
1132         ret = btrfs_run_dev_stats(trans, root->fs_info);
1133         if (ret)
1134                 return ret;
1135         ret = btrfs_run_dev_replace(trans, root->fs_info);
1136         if (ret)
1137                 return ret;
1138         ret = btrfs_run_qgroups(trans, root->fs_info);
1139         if (ret)
1140                 return ret;
1141
1142         ret = btrfs_setup_space_cache(trans, root);
1143         if (ret)
1144                 return ret;
1145
1146         /* run_qgroups might have added some more refs */
1147         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1148         if (ret)
1149                 return ret;
1150 again:
1151         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1152                 next = fs_info->dirty_cowonly_roots.next;
1153                 list_del_init(next);
1154                 root = list_entry(next, struct btrfs_root, dirty_list);
1155                 clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1156
1157                 if (root != fs_info->extent_root)
1158                         list_add_tail(&root->dirty_list,
1159                                       &trans->transaction->switch_commits);
1160                 ret = update_cowonly_root(trans, root);
1161                 if (ret)
1162                         return ret;
1163                 ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1164                 if (ret)
1165                         return ret;
1166         }
1167
1168         while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1169                 ret = btrfs_write_dirty_block_groups(trans, root);
1170                 if (ret)
1171                         return ret;
1172                 ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1173                 if (ret)
1174                         return ret;
1175         }
1176
1177         if (!list_empty(&fs_info->dirty_cowonly_roots))
1178                 goto again;
1179
1180         list_add_tail(&fs_info->extent_root->dirty_list,
1181                       &trans->transaction->switch_commits);
1182         btrfs_after_dev_replace_commit(fs_info);
1183
1184         return 0;
1185 }
1186
1187 /*
1188  * dead roots are old snapshots that need to be deleted.  This allocates
1189  * a dirty root struct and adds it into the list of dead roots that need to
1190  * be deleted
1191  */
1192 void btrfs_add_dead_root(struct btrfs_root *root)
1193 {
1194         spin_lock(&root->fs_info->trans_lock);
1195         if (list_empty(&root->root_list))
1196                 list_add_tail(&root->root_list, &root->fs_info->dead_roots);
1197         spin_unlock(&root->fs_info->trans_lock);
1198 }
1199
1200 /*
1201  * update all the cowonly tree roots on disk
1202  */
1203 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
1204                                     struct btrfs_root *root)
1205 {
1206         struct btrfs_root *gang[8];
1207         struct btrfs_fs_info *fs_info = root->fs_info;
1208         int i;
1209         int ret;
1210         int err = 0;
1211
1212         spin_lock(&fs_info->fs_roots_radix_lock);
1213         while (1) {
1214                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1215                                                  (void **)gang, 0,
1216                                                  ARRAY_SIZE(gang),
1217                                                  BTRFS_ROOT_TRANS_TAG);
1218                 if (ret == 0)
1219                         break;
1220                 for (i = 0; i < ret; i++) {
1221                         root = gang[i];
1222                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
1223                                         (unsigned long)root->root_key.objectid,
1224                                         BTRFS_ROOT_TRANS_TAG);
1225                         spin_unlock(&fs_info->fs_roots_radix_lock);
1226
1227                         btrfs_free_log(trans, root);
1228                         btrfs_update_reloc_root(trans, root);
1229                         btrfs_orphan_commit_root(trans, root);
1230
1231                         btrfs_save_ino_cache(root, trans);
1232
1233                         /* see comments in should_cow_block() */
1234                         clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1235                         smp_mb__after_atomic();
1236
1237                         if (root->commit_root != root->node) {
1238                                 list_add_tail(&root->dirty_list,
1239                                         &trans->transaction->switch_commits);
1240                                 btrfs_set_root_node(&root->root_item,
1241                                                     root->node);
1242                         }
1243
1244                         err = btrfs_update_root(trans, fs_info->tree_root,
1245                                                 &root->root_key,
1246                                                 &root->root_item);
1247                         spin_lock(&fs_info->fs_roots_radix_lock);
1248                         if (err)
1249                                 break;
1250                 }
1251         }
1252         spin_unlock(&fs_info->fs_roots_radix_lock);
1253         return err;
1254 }
1255
1256 /*
1257  * defrag a given btree.
1258  * Every leaf in the btree is read and defragged.
1259  */
1260 int btrfs_defrag_root(struct btrfs_root *root)
1261 {
1262         struct btrfs_fs_info *info = root->fs_info;
1263         struct btrfs_trans_handle *trans;
1264         int ret;
1265
1266         if (test_and_set_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state))
1267                 return 0;
1268
1269         while (1) {
1270                 trans = btrfs_start_transaction(root, 0);
1271                 if (IS_ERR(trans))
1272                         return PTR_ERR(trans);
1273
1274                 ret = btrfs_defrag_leaves(trans, root);
1275
1276                 btrfs_end_transaction(trans, root);
1277                 btrfs_btree_balance_dirty(info->tree_root);
1278                 cond_resched();
1279
1280                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
1281                         break;
1282
1283                 if (btrfs_defrag_cancelled(root->fs_info)) {
1284                         pr_debug("BTRFS: defrag_root cancelled\n");
1285                         ret = -EAGAIN;
1286                         break;
1287                 }
1288         }
1289         clear_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state);
1290         return ret;
1291 }
1292
1293 /*
1294  * new snapshots need to be created at a very specific time in the
1295  * transaction commit.  This does the actual creation.
1296  *
1297  * Note:
1298  * If the error which may affect the commitment of the current transaction
1299  * happens, we should return the error number. If the error which just affect
1300  * the creation of the pending snapshots, just return 0.
1301  */
1302 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1303                                    struct btrfs_fs_info *fs_info,
1304                                    struct btrfs_pending_snapshot *pending)
1305 {
1306         struct btrfs_key key;
1307         struct btrfs_root_item *new_root_item;
1308         struct btrfs_root *tree_root = fs_info->tree_root;
1309         struct btrfs_root *root = pending->root;
1310         struct btrfs_root *parent_root;
1311         struct btrfs_block_rsv *rsv;
1312         struct inode *parent_inode;
1313         struct btrfs_path *path;
1314         struct btrfs_dir_item *dir_item;
1315         struct dentry *dentry;
1316         struct extent_buffer *tmp;
1317         struct extent_buffer *old;
1318         struct timespec cur_time = CURRENT_TIME;
1319         int ret = 0;
1320         u64 to_reserve = 0;
1321         u64 index = 0;
1322         u64 objectid;
1323         u64 root_flags;
1324         uuid_le new_uuid;
1325
1326         path = btrfs_alloc_path();
1327         if (!path) {
1328                 pending->error = -ENOMEM;
1329                 return 0;
1330         }
1331
1332         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
1333         if (!new_root_item) {
1334                 pending->error = -ENOMEM;
1335                 goto root_item_alloc_fail;
1336         }
1337
1338         pending->error = btrfs_find_free_objectid(tree_root, &objectid);
1339         if (pending->error)
1340                 goto no_free_objectid;
1341
1342         /*
1343          * Make qgroup to skip current new snapshot's qgroupid, as it is
1344          * accounted by later btrfs_qgroup_inherit().
1345          */
1346         btrfs_set_skip_qgroup(trans, objectid);
1347
1348         btrfs_reloc_pre_snapshot(pending, &to_reserve);
1349
1350         if (to_reserve > 0) {
1351                 pending->error = btrfs_block_rsv_add(root,
1352                                                      &pending->block_rsv,
1353                                                      to_reserve,
1354                                                      BTRFS_RESERVE_NO_FLUSH);
1355                 if (pending->error)
1356                         goto clear_skip_qgroup;
1357         }
1358
1359         key.objectid = objectid;
1360         key.offset = (u64)-1;
1361         key.type = BTRFS_ROOT_ITEM_KEY;
1362
1363         rsv = trans->block_rsv;
1364         trans->block_rsv = &pending->block_rsv;
1365         trans->bytes_reserved = trans->block_rsv->reserved;
1366
1367         dentry = pending->dentry;
1368         parent_inode = pending->dir;
1369         parent_root = BTRFS_I(parent_inode)->root;
1370         record_root_in_trans(trans, parent_root);
1371
1372         /*
1373          * insert the directory item
1374          */
1375         ret = btrfs_set_inode_index(parent_inode, &index);
1376         BUG_ON(ret); /* -ENOMEM */
1377
1378         /* check if there is a file/dir which has the same name. */
1379         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1380                                          btrfs_ino(parent_inode),
1381                                          dentry->d_name.name,
1382                                          dentry->d_name.len, 0);
1383         if (dir_item != NULL && !IS_ERR(dir_item)) {
1384                 pending->error = -EEXIST;
1385                 goto dir_item_existed;
1386         } else if (IS_ERR(dir_item)) {
1387                 ret = PTR_ERR(dir_item);
1388                 btrfs_abort_transaction(trans, root, ret);
1389                 goto fail;
1390         }
1391         btrfs_release_path(path);
1392
1393         /*
1394          * pull in the delayed directory update
1395          * and the delayed inode item
1396          * otherwise we corrupt the FS during
1397          * snapshot
1398          */
1399         ret = btrfs_run_delayed_items(trans, root);
1400         if (ret) {      /* Transaction aborted */
1401                 btrfs_abort_transaction(trans, root, ret);
1402                 goto fail;
1403         }
1404
1405         record_root_in_trans(trans, root);
1406         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1407         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1408         btrfs_check_and_init_root_item(new_root_item);
1409
1410         root_flags = btrfs_root_flags(new_root_item);
1411         if (pending->readonly)
1412                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1413         else
1414                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1415         btrfs_set_root_flags(new_root_item, root_flags);
1416
1417         btrfs_set_root_generation_v2(new_root_item,
1418                         trans->transid);
1419         uuid_le_gen(&new_uuid);
1420         memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1421         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1422                         BTRFS_UUID_SIZE);
1423         if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1424                 memset(new_root_item->received_uuid, 0,
1425                        sizeof(new_root_item->received_uuid));
1426                 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1427                 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1428                 btrfs_set_root_stransid(new_root_item, 0);
1429                 btrfs_set_root_rtransid(new_root_item, 0);
1430         }
1431         btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1432         btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1433         btrfs_set_root_otransid(new_root_item, trans->transid);
1434
1435         old = btrfs_lock_root_node(root);
1436         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1437         if (ret) {
1438                 btrfs_tree_unlock(old);
1439                 free_extent_buffer(old);
1440                 btrfs_abort_transaction(trans, root, ret);
1441                 goto fail;
1442         }
1443
1444         btrfs_set_lock_blocking(old);
1445
1446         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1447         /* clean up in any case */
1448         btrfs_tree_unlock(old);
1449         free_extent_buffer(old);
1450         if (ret) {
1451                 btrfs_abort_transaction(trans, root, ret);
1452                 goto fail;
1453         }
1454         /* see comments in should_cow_block() */
1455         set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1456         smp_wmb();
1457
1458         btrfs_set_root_node(new_root_item, tmp);
1459         /* record when the snapshot was created in key.offset */
1460         key.offset = trans->transid;
1461         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1462         btrfs_tree_unlock(tmp);
1463         free_extent_buffer(tmp);
1464         if (ret) {
1465                 btrfs_abort_transaction(trans, root, ret);
1466                 goto fail;
1467         }
1468
1469         /*
1470          * insert root back/forward references
1471          */
1472         ret = btrfs_add_root_ref(trans, tree_root, objectid,
1473                                  parent_root->root_key.objectid,
1474                                  btrfs_ino(parent_inode), index,
1475                                  dentry->d_name.name, dentry->d_name.len);
1476         if (ret) {
1477                 btrfs_abort_transaction(trans, root, ret);
1478                 goto fail;
1479         }
1480
1481         key.offset = (u64)-1;
1482         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1483         if (IS_ERR(pending->snap)) {
1484                 ret = PTR_ERR(pending->snap);
1485                 btrfs_abort_transaction(trans, root, ret);
1486                 goto fail;
1487         }
1488
1489         ret = btrfs_reloc_post_snapshot(trans, pending);
1490         if (ret) {
1491                 btrfs_abort_transaction(trans, root, ret);
1492                 goto fail;
1493         }
1494
1495         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1496         if (ret) {
1497                 btrfs_abort_transaction(trans, root, ret);
1498                 goto fail;
1499         }
1500
1501         ret = btrfs_insert_dir_item(trans, parent_root,
1502                                     dentry->d_name.name, dentry->d_name.len,
1503                                     parent_inode, &key,
1504                                     BTRFS_FT_DIR, index);
1505         /* We have check then name at the beginning, so it is impossible. */
1506         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1507         if (ret) {
1508                 btrfs_abort_transaction(trans, root, ret);
1509                 goto fail;
1510         }
1511
1512         btrfs_i_size_write(parent_inode, parent_inode->i_size +
1513                                          dentry->d_name.len * 2);
1514         parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
1515         ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1516         if (ret) {
1517                 btrfs_abort_transaction(trans, root, ret);
1518                 goto fail;
1519         }
1520         ret = btrfs_uuid_tree_add(trans, fs_info->uuid_root, new_uuid.b,
1521                                   BTRFS_UUID_KEY_SUBVOL, objectid);
1522         if (ret) {
1523                 btrfs_abort_transaction(trans, root, ret);
1524                 goto fail;
1525         }
1526         if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1527                 ret = btrfs_uuid_tree_add(trans, fs_info->uuid_root,
1528                                           new_root_item->received_uuid,
1529                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1530                                           objectid);
1531                 if (ret && ret != -EEXIST) {
1532                         btrfs_abort_transaction(trans, root, ret);
1533                         goto fail;
1534                 }
1535         }
1536
1537         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1538         if (ret) {
1539                 btrfs_abort_transaction(trans, root, ret);
1540                 goto fail;
1541         }
1542
1543         /*
1544          * account qgroup counters before qgroup_inherit()
1545          */
1546         ret = btrfs_qgroup_prepare_account_extents(trans, fs_info);
1547         if (ret)
1548                 goto fail;
1549         ret = btrfs_qgroup_account_extents(trans, fs_info);
1550         if (ret)
1551                 goto fail;
1552         ret = btrfs_qgroup_inherit(trans, fs_info,
1553                                    root->root_key.objectid,
1554                                    objectid, pending->inherit);
1555         if (ret) {
1556                 btrfs_abort_transaction(trans, root, ret);
1557                 goto fail;
1558         }
1559
1560 fail:
1561         pending->error = ret;
1562 dir_item_existed:
1563         trans->block_rsv = rsv;
1564         trans->bytes_reserved = 0;
1565 clear_skip_qgroup:
1566         btrfs_clear_skip_qgroup(trans);
1567 no_free_objectid:
1568         kfree(new_root_item);
1569 root_item_alloc_fail:
1570         btrfs_free_path(path);
1571         return ret;
1572 }
1573
1574 /*
1575  * create all the snapshots we've scheduled for creation
1576  */
1577 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1578                                              struct btrfs_fs_info *fs_info)
1579 {
1580         struct btrfs_pending_snapshot *pending, *next;
1581         struct list_head *head = &trans->transaction->pending_snapshots;
1582         int ret = 0;
1583
1584         list_for_each_entry_safe(pending, next, head, list) {
1585                 list_del(&pending->list);
1586                 ret = create_pending_snapshot(trans, fs_info, pending);
1587                 if (ret)
1588                         break;
1589         }
1590         return ret;
1591 }
1592
1593 static void update_super_roots(struct btrfs_root *root)
1594 {
1595         struct btrfs_root_item *root_item;
1596         struct btrfs_super_block *super;
1597
1598         super = root->fs_info->super_copy;
1599
1600         root_item = &root->fs_info->chunk_root->root_item;
1601         super->chunk_root = root_item->bytenr;
1602         super->chunk_root_generation = root_item->generation;
1603         super->chunk_root_level = root_item->level;
1604
1605         root_item = &root->fs_info->tree_root->root_item;
1606         super->root = root_item->bytenr;
1607         super->generation = root_item->generation;
1608         super->root_level = root_item->level;
1609         if (btrfs_test_opt(root, SPACE_CACHE))
1610                 super->cache_generation = root_item->generation;
1611         if (root->fs_info->update_uuid_tree_gen)
1612                 super->uuid_tree_generation = root_item->generation;
1613 }
1614
1615 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1616 {
1617         struct btrfs_transaction *trans;
1618         int ret = 0;
1619
1620         spin_lock(&info->trans_lock);
1621         trans = info->running_transaction;
1622         if (trans)
1623                 ret = (trans->state >= TRANS_STATE_COMMIT_START);
1624         spin_unlock(&info->trans_lock);
1625         return ret;
1626 }
1627
1628 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1629 {
1630         struct btrfs_transaction *trans;
1631         int ret = 0;
1632
1633         spin_lock(&info->trans_lock);
1634         trans = info->running_transaction;
1635         if (trans)
1636                 ret = is_transaction_blocked(trans);
1637         spin_unlock(&info->trans_lock);
1638         return ret;
1639 }
1640
1641 /*
1642  * wait for the current transaction commit to start and block subsequent
1643  * transaction joins
1644  */
1645 static void wait_current_trans_commit_start(struct btrfs_root *root,
1646                                             struct btrfs_transaction *trans)
1647 {
1648         wait_event(root->fs_info->transaction_blocked_wait,
1649                    trans->state >= TRANS_STATE_COMMIT_START ||
1650                    trans->aborted);
1651 }
1652
1653 /*
1654  * wait for the current transaction to start and then become unblocked.
1655  * caller holds ref.
1656  */
1657 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1658                                          struct btrfs_transaction *trans)
1659 {
1660         wait_event(root->fs_info->transaction_wait,
1661                    trans->state >= TRANS_STATE_UNBLOCKED ||
1662                    trans->aborted);
1663 }
1664
1665 /*
1666  * commit transactions asynchronously. once btrfs_commit_transaction_async
1667  * returns, any subsequent transaction will not be allowed to join.
1668  */
1669 struct btrfs_async_commit {
1670         struct btrfs_trans_handle *newtrans;
1671         struct btrfs_root *root;
1672         struct work_struct work;
1673 };
1674
1675 static void do_async_commit(struct work_struct *work)
1676 {
1677         struct btrfs_async_commit *ac =
1678                 container_of(work, struct btrfs_async_commit, work);
1679
1680         /*
1681          * We've got freeze protection passed with the transaction.
1682          * Tell lockdep about it.
1683          */
1684         if (ac->newtrans->type & __TRANS_FREEZABLE)
1685                 __sb_writers_acquired(ac->root->fs_info->sb, SB_FREEZE_FS);
1686
1687         current->journal_info = ac->newtrans;
1688
1689         btrfs_commit_transaction(ac->newtrans, ac->root);
1690         kfree(ac);
1691 }
1692
1693 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1694                                    struct btrfs_root *root,
1695                                    int wait_for_unblock)
1696 {
1697         struct btrfs_async_commit *ac;
1698         struct btrfs_transaction *cur_trans;
1699
1700         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1701         if (!ac)
1702                 return -ENOMEM;
1703
1704         INIT_WORK(&ac->work, do_async_commit);
1705         ac->root = root;
1706         ac->newtrans = btrfs_join_transaction(root);
1707         if (IS_ERR(ac->newtrans)) {
1708                 int err = PTR_ERR(ac->newtrans);
1709                 kfree(ac);
1710                 return err;
1711         }
1712
1713         /* take transaction reference */
1714         cur_trans = trans->transaction;
1715         atomic_inc(&cur_trans->use_count);
1716
1717         btrfs_end_transaction(trans, root);
1718
1719         /*
1720          * Tell lockdep we've released the freeze rwsem, since the
1721          * async commit thread will be the one to unlock it.
1722          */
1723         if (ac->newtrans->type & __TRANS_FREEZABLE)
1724                 __sb_writers_release(root->fs_info->sb, SB_FREEZE_FS);
1725
1726         schedule_work(&ac->work);
1727
1728         /* wait for transaction to start and unblock */
1729         if (wait_for_unblock)
1730                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1731         else
1732                 wait_current_trans_commit_start(root, cur_trans);
1733
1734         if (current->journal_info == trans)
1735                 current->journal_info = NULL;
1736
1737         btrfs_put_transaction(cur_trans);
1738         return 0;
1739 }
1740
1741
1742 static void cleanup_transaction(struct btrfs_trans_handle *trans,
1743                                 struct btrfs_root *root, int err)
1744 {
1745         struct btrfs_transaction *cur_trans = trans->transaction;
1746         DEFINE_WAIT(wait);
1747
1748         WARN_ON(trans->use_count > 1);
1749
1750         btrfs_abort_transaction(trans, root, err);
1751
1752         spin_lock(&root->fs_info->trans_lock);
1753
1754         /*
1755          * If the transaction is removed from the list, it means this
1756          * transaction has been committed successfully, so it is impossible
1757          * to call the cleanup function.
1758          */
1759         BUG_ON(list_empty(&cur_trans->list));
1760
1761         list_del_init(&cur_trans->list);
1762         if (cur_trans == root->fs_info->running_transaction) {
1763                 cur_trans->state = TRANS_STATE_COMMIT_DOING;
1764                 spin_unlock(&root->fs_info->trans_lock);
1765                 wait_event(cur_trans->writer_wait,
1766                            atomic_read(&cur_trans->num_writers) == 1);
1767
1768                 spin_lock(&root->fs_info->trans_lock);
1769         }
1770         spin_unlock(&root->fs_info->trans_lock);
1771
1772         btrfs_cleanup_one_transaction(trans->transaction, root);
1773
1774         spin_lock(&root->fs_info->trans_lock);
1775         if (cur_trans == root->fs_info->running_transaction)
1776                 root->fs_info->running_transaction = NULL;
1777         spin_unlock(&root->fs_info->trans_lock);
1778
1779         if (trans->type & __TRANS_FREEZABLE)
1780                 sb_end_intwrite(root->fs_info->sb);
1781         btrfs_put_transaction(cur_trans);
1782         btrfs_put_transaction(cur_trans);
1783
1784         trace_btrfs_transaction_commit(root);
1785
1786         if (current->journal_info == trans)
1787                 current->journal_info = NULL;
1788         btrfs_scrub_cancel(root->fs_info);
1789
1790         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1791 }
1792
1793 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
1794 {
1795         if (btrfs_test_opt(fs_info->tree_root, FLUSHONCOMMIT))
1796                 return btrfs_start_delalloc_roots(fs_info, 1, -1);
1797         return 0;
1798 }
1799
1800 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
1801 {
1802         if (btrfs_test_opt(fs_info->tree_root, FLUSHONCOMMIT))
1803                 btrfs_wait_ordered_roots(fs_info, -1);
1804 }
1805
1806 static inline void
1807 btrfs_wait_pending_ordered(struct btrfs_transaction *cur_trans,
1808                            struct btrfs_fs_info *fs_info)
1809 {
1810         struct btrfs_ordered_extent *ordered;
1811
1812         spin_lock(&fs_info->trans_lock);
1813         while (!list_empty(&cur_trans->pending_ordered)) {
1814                 ordered = list_first_entry(&cur_trans->pending_ordered,
1815                                            struct btrfs_ordered_extent,
1816                                            trans_list);
1817                 list_del_init(&ordered->trans_list);
1818                 spin_unlock(&fs_info->trans_lock);
1819
1820                 wait_event(ordered->wait, test_bit(BTRFS_ORDERED_COMPLETE,
1821                                                    &ordered->flags));
1822                 btrfs_put_ordered_extent(ordered);
1823                 spin_lock(&fs_info->trans_lock);
1824         }
1825         spin_unlock(&fs_info->trans_lock);
1826 }
1827
1828 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1829                              struct btrfs_root *root)
1830 {
1831         struct btrfs_transaction *cur_trans = trans->transaction;
1832         struct btrfs_transaction *prev_trans = NULL;
1833         struct btrfs_inode *btree_ino = BTRFS_I(root->fs_info->btree_inode);
1834         int ret;
1835
1836         /* Stop the commit early if ->aborted is set */
1837         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1838                 ret = cur_trans->aborted;
1839                 btrfs_end_transaction(trans, root);
1840                 return ret;
1841         }
1842
1843         /* make a pass through all the delayed refs we have so far
1844          * any runnings procs may add more while we are here
1845          */
1846         ret = btrfs_run_delayed_refs(trans, root, 0);
1847         if (ret) {
1848                 btrfs_end_transaction(trans, root);
1849                 return ret;
1850         }
1851
1852         btrfs_trans_release_metadata(trans, root);
1853         trans->block_rsv = NULL;
1854         if (trans->qgroup_reserved) {
1855                 btrfs_qgroup_free(root, trans->qgroup_reserved);
1856                 trans->qgroup_reserved = 0;
1857         }
1858
1859         cur_trans = trans->transaction;
1860
1861         /*
1862          * set the flushing flag so procs in this transaction have to
1863          * start sending their work down.
1864          */
1865         cur_trans->delayed_refs.flushing = 1;
1866         smp_wmb();
1867
1868         if (!list_empty(&trans->new_bgs))
1869                 btrfs_create_pending_block_groups(trans, root);
1870
1871         ret = btrfs_run_delayed_refs(trans, root, 0);
1872         if (ret) {
1873                 btrfs_end_transaction(trans, root);
1874                 return ret;
1875         }
1876
1877         if (!cur_trans->dirty_bg_run) {
1878                 int run_it = 0;
1879
1880                 /* this mutex is also taken before trying to set
1881                  * block groups readonly.  We need to make sure
1882                  * that nobody has set a block group readonly
1883                  * after a extents from that block group have been
1884                  * allocated for cache files.  btrfs_set_block_group_ro
1885                  * will wait for the transaction to commit if it
1886                  * finds dirty_bg_run = 1
1887                  *
1888                  * The dirty_bg_run flag is also used to make sure only
1889                  * one process starts all the block group IO.  It wouldn't
1890                  * hurt to have more than one go through, but there's no
1891                  * real advantage to it either.
1892                  */
1893                 mutex_lock(&root->fs_info->ro_block_group_mutex);
1894                 if (!cur_trans->dirty_bg_run) {
1895                         run_it = 1;
1896                         cur_trans->dirty_bg_run = 1;
1897                 }
1898                 mutex_unlock(&root->fs_info->ro_block_group_mutex);
1899
1900                 if (run_it)
1901                         ret = btrfs_start_dirty_block_groups(trans, root);
1902         }
1903         if (ret) {
1904                 btrfs_end_transaction(trans, root);
1905                 return ret;
1906         }
1907
1908         spin_lock(&root->fs_info->trans_lock);
1909         list_splice_init(&trans->ordered, &cur_trans->pending_ordered);
1910         if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
1911                 spin_unlock(&root->fs_info->trans_lock);
1912                 atomic_inc(&cur_trans->use_count);
1913                 ret = btrfs_end_transaction(trans, root);
1914
1915                 wait_for_commit(root, cur_trans);
1916
1917                 if (unlikely(cur_trans->aborted))
1918                         ret = cur_trans->aborted;
1919
1920                 btrfs_put_transaction(cur_trans);
1921
1922                 return ret;
1923         }
1924
1925         cur_trans->state = TRANS_STATE_COMMIT_START;
1926         wake_up(&root->fs_info->transaction_blocked_wait);
1927
1928         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1929                 prev_trans = list_entry(cur_trans->list.prev,
1930                                         struct btrfs_transaction, list);
1931                 if (prev_trans->state != TRANS_STATE_COMPLETED) {
1932                         atomic_inc(&prev_trans->use_count);
1933                         spin_unlock(&root->fs_info->trans_lock);
1934
1935                         wait_for_commit(root, prev_trans);
1936                         ret = prev_trans->aborted;
1937
1938                         btrfs_put_transaction(prev_trans);
1939                         if (ret)
1940                                 goto cleanup_transaction;
1941                 } else {
1942                         spin_unlock(&root->fs_info->trans_lock);
1943                 }
1944         } else {
1945                 spin_unlock(&root->fs_info->trans_lock);
1946         }
1947
1948         extwriter_counter_dec(cur_trans, trans->type);
1949
1950         ret = btrfs_start_delalloc_flush(root->fs_info);
1951         if (ret)
1952                 goto cleanup_transaction;
1953
1954         ret = btrfs_run_delayed_items(trans, root);
1955         if (ret)
1956                 goto cleanup_transaction;
1957
1958         wait_event(cur_trans->writer_wait,
1959                    extwriter_counter_read(cur_trans) == 0);
1960
1961         /* some pending stuffs might be added after the previous flush. */
1962         ret = btrfs_run_delayed_items(trans, root);
1963         if (ret)
1964                 goto cleanup_transaction;
1965
1966         btrfs_wait_delalloc_flush(root->fs_info);
1967
1968         btrfs_wait_pending_ordered(cur_trans, root->fs_info);
1969
1970         btrfs_scrub_pause(root);
1971         /*
1972          * Ok now we need to make sure to block out any other joins while we
1973          * commit the transaction.  We could have started a join before setting
1974          * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
1975          */
1976         spin_lock(&root->fs_info->trans_lock);
1977         cur_trans->state = TRANS_STATE_COMMIT_DOING;
1978         spin_unlock(&root->fs_info->trans_lock);
1979         wait_event(cur_trans->writer_wait,
1980                    atomic_read(&cur_trans->num_writers) == 1);
1981
1982         /* ->aborted might be set after the previous check, so check it */
1983         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1984                 ret = cur_trans->aborted;
1985                 goto scrub_continue;
1986         }
1987         /*
1988          * the reloc mutex makes sure that we stop
1989          * the balancing code from coming in and moving
1990          * extents around in the middle of the commit
1991          */
1992         mutex_lock(&root->fs_info->reloc_mutex);
1993
1994         /*
1995          * We needn't worry about the delayed items because we will
1996          * deal with them in create_pending_snapshot(), which is the
1997          * core function of the snapshot creation.
1998          */
1999         ret = create_pending_snapshots(trans, root->fs_info);
2000         if (ret) {
2001                 mutex_unlock(&root->fs_info->reloc_mutex);
2002                 goto scrub_continue;
2003         }
2004
2005         /*
2006          * We insert the dir indexes of the snapshots and update the inode
2007          * of the snapshots' parents after the snapshot creation, so there
2008          * are some delayed items which are not dealt with. Now deal with
2009          * them.
2010          *
2011          * We needn't worry that this operation will corrupt the snapshots,
2012          * because all the tree which are snapshoted will be forced to COW
2013          * the nodes and leaves.
2014          */
2015         ret = btrfs_run_delayed_items(trans, root);
2016         if (ret) {
2017                 mutex_unlock(&root->fs_info->reloc_mutex);
2018                 goto scrub_continue;
2019         }
2020
2021         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
2022         if (ret) {
2023                 mutex_unlock(&root->fs_info->reloc_mutex);
2024                 goto scrub_continue;
2025         }
2026
2027         /* Reocrd old roots for later qgroup accounting */
2028         ret = btrfs_qgroup_prepare_account_extents(trans, root->fs_info);
2029         if (ret) {
2030                 mutex_unlock(&root->fs_info->reloc_mutex);
2031                 goto scrub_continue;
2032         }
2033
2034         /*
2035          * make sure none of the code above managed to slip in a
2036          * delayed item
2037          */
2038         btrfs_assert_delayed_root_empty(root);
2039
2040         WARN_ON(cur_trans != trans->transaction);
2041
2042         /* btrfs_commit_tree_roots is responsible for getting the
2043          * various roots consistent with each other.  Every pointer
2044          * in the tree of tree roots has to point to the most up to date
2045          * root for every subvolume and other tree.  So, we have to keep
2046          * the tree logging code from jumping in and changing any
2047          * of the trees.
2048          *
2049          * At this point in the commit, there can't be any tree-log
2050          * writers, but a little lower down we drop the trans mutex
2051          * and let new people in.  By holding the tree_log_mutex
2052          * from now until after the super is written, we avoid races
2053          * with the tree-log code.
2054          */
2055         mutex_lock(&root->fs_info->tree_log_mutex);
2056
2057         ret = commit_fs_roots(trans, root);
2058         if (ret) {
2059                 mutex_unlock(&root->fs_info->tree_log_mutex);
2060                 mutex_unlock(&root->fs_info->reloc_mutex);
2061                 goto scrub_continue;
2062         }
2063
2064         /*
2065          * Since the transaction is done, we can apply the pending changes
2066          * before the next transaction.
2067          */
2068         btrfs_apply_pending_changes(root->fs_info);
2069
2070         /* commit_fs_roots gets rid of all the tree log roots, it is now
2071          * safe to free the root of tree log roots
2072          */
2073         btrfs_free_log_root_tree(trans, root->fs_info);
2074
2075         /*
2076          * Since fs roots are all committed, we can get a quite accurate
2077          * new_roots. So let's do quota accounting.
2078          */
2079         ret = btrfs_qgroup_account_extents(trans, root->fs_info);
2080         if (ret < 0) {
2081                 mutex_unlock(&root->fs_info->tree_log_mutex);
2082                 mutex_unlock(&root->fs_info->reloc_mutex);
2083                 goto scrub_continue;
2084         }
2085
2086         ret = commit_cowonly_roots(trans, root);
2087         if (ret) {
2088                 mutex_unlock(&root->fs_info->tree_log_mutex);
2089                 mutex_unlock(&root->fs_info->reloc_mutex);
2090                 goto scrub_continue;
2091         }
2092
2093         /*
2094          * The tasks which save the space cache and inode cache may also
2095          * update ->aborted, check it.
2096          */
2097         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
2098                 ret = cur_trans->aborted;
2099                 mutex_unlock(&root->fs_info->tree_log_mutex);
2100                 mutex_unlock(&root->fs_info->reloc_mutex);
2101                 goto scrub_continue;
2102         }
2103
2104         btrfs_prepare_extent_commit(trans, root);
2105
2106         cur_trans = root->fs_info->running_transaction;
2107
2108         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
2109                             root->fs_info->tree_root->node);
2110         list_add_tail(&root->fs_info->tree_root->dirty_list,
2111                       &cur_trans->switch_commits);
2112
2113         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
2114                             root->fs_info->chunk_root->node);
2115         list_add_tail(&root->fs_info->chunk_root->dirty_list,
2116                       &cur_trans->switch_commits);
2117
2118         switch_commit_roots(cur_trans, root->fs_info);
2119
2120         assert_qgroups_uptodate(trans);
2121         ASSERT(list_empty(&cur_trans->dirty_bgs));
2122         ASSERT(list_empty(&cur_trans->io_bgs));
2123         update_super_roots(root);
2124
2125         btrfs_set_super_log_root(root->fs_info->super_copy, 0);
2126         btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
2127         memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
2128                sizeof(*root->fs_info->super_copy));
2129
2130         btrfs_update_commit_device_size(root->fs_info);
2131         btrfs_update_commit_device_bytes_used(root, cur_trans);
2132
2133         clear_bit(BTRFS_INODE_BTREE_LOG1_ERR, &btree_ino->runtime_flags);
2134         clear_bit(BTRFS_INODE_BTREE_LOG2_ERR, &btree_ino->runtime_flags);
2135
2136         btrfs_trans_release_chunk_metadata(trans);
2137
2138         spin_lock(&root->fs_info->trans_lock);
2139         cur_trans->state = TRANS_STATE_UNBLOCKED;
2140         root->fs_info->running_transaction = NULL;
2141         spin_unlock(&root->fs_info->trans_lock);
2142         mutex_unlock(&root->fs_info->reloc_mutex);
2143
2144         wake_up(&root->fs_info->transaction_wait);
2145
2146         ret = btrfs_write_and_wait_transaction(trans, root);
2147         if (ret) {
2148                 btrfs_std_error(root->fs_info, ret,
2149                             "Error while writing out transaction");
2150                 mutex_unlock(&root->fs_info->tree_log_mutex);
2151                 goto scrub_continue;
2152         }
2153
2154         ret = write_ctree_super(trans, root, 0);
2155         if (ret) {
2156                 mutex_unlock(&root->fs_info->tree_log_mutex);
2157                 goto scrub_continue;
2158         }
2159
2160         /*
2161          * the super is written, we can safely allow the tree-loggers
2162          * to go about their business
2163          */
2164         mutex_unlock(&root->fs_info->tree_log_mutex);
2165
2166         btrfs_finish_extent_commit(trans, root);
2167
2168         if (cur_trans->have_free_bgs)
2169                 btrfs_clear_space_info_full(root->fs_info);
2170
2171         root->fs_info->last_trans_committed = cur_trans->transid;
2172         /*
2173          * We needn't acquire the lock here because there is no other task
2174          * which can change it.
2175          */
2176         cur_trans->state = TRANS_STATE_COMPLETED;
2177         wake_up(&cur_trans->commit_wait);
2178
2179         spin_lock(&root->fs_info->trans_lock);
2180         list_del_init(&cur_trans->list);
2181         spin_unlock(&root->fs_info->trans_lock);
2182
2183         btrfs_put_transaction(cur_trans);
2184         btrfs_put_transaction(cur_trans);
2185
2186         if (trans->type & __TRANS_FREEZABLE)
2187                 sb_end_intwrite(root->fs_info->sb);
2188
2189         trace_btrfs_transaction_commit(root);
2190
2191         btrfs_scrub_continue(root);
2192
2193         if (current->journal_info == trans)
2194                 current->journal_info = NULL;
2195
2196         kmem_cache_free(btrfs_trans_handle_cachep, trans);
2197
2198         if (current != root->fs_info->transaction_kthread &&
2199             current != root->fs_info->cleaner_kthread)
2200                 btrfs_run_delayed_iputs(root);
2201
2202         return ret;
2203
2204 scrub_continue:
2205         btrfs_scrub_continue(root);
2206 cleanup_transaction:
2207         btrfs_trans_release_metadata(trans, root);
2208         btrfs_trans_release_chunk_metadata(trans);
2209         trans->block_rsv = NULL;
2210         if (trans->qgroup_reserved) {
2211                 btrfs_qgroup_free(root, trans->qgroup_reserved);
2212                 trans->qgroup_reserved = 0;
2213         }
2214         btrfs_warn(root->fs_info, "Skipping commit of aborted transaction.");
2215         if (current->journal_info == trans)
2216                 current->journal_info = NULL;
2217         cleanup_transaction(trans, root, ret);
2218
2219         return ret;
2220 }
2221
2222 /*
2223  * return < 0 if error
2224  * 0 if there are no more dead_roots at the time of call
2225  * 1 there are more to be processed, call me again
2226  *
2227  * The return value indicates there are certainly more snapshots to delete, but
2228  * if there comes a new one during processing, it may return 0. We don't mind,
2229  * because btrfs_commit_super will poke cleaner thread and it will process it a
2230  * few seconds later.
2231  */
2232 int btrfs_clean_one_deleted_snapshot(struct btrfs_root *root)
2233 {
2234         int ret;
2235         struct btrfs_fs_info *fs_info = root->fs_info;
2236
2237         spin_lock(&fs_info->trans_lock);
2238         if (list_empty(&fs_info->dead_roots)) {
2239                 spin_unlock(&fs_info->trans_lock);
2240                 return 0;
2241         }
2242         root = list_first_entry(&fs_info->dead_roots,
2243                         struct btrfs_root, root_list);
2244         list_del_init(&root->root_list);
2245         spin_unlock(&fs_info->trans_lock);
2246
2247         pr_debug("BTRFS: cleaner removing %llu\n", root->objectid);
2248
2249         btrfs_kill_all_delayed_nodes(root);
2250
2251         if (btrfs_header_backref_rev(root->node) <
2252                         BTRFS_MIXED_BACKREF_REV)
2253                 ret = btrfs_drop_snapshot(root, NULL, 0, 0);
2254         else
2255                 ret = btrfs_drop_snapshot(root, NULL, 1, 0);
2256
2257         return (ret < 0) ? 0 : 1;
2258 }
2259
2260 void btrfs_apply_pending_changes(struct btrfs_fs_info *fs_info)
2261 {
2262         unsigned long prev;
2263         unsigned long bit;
2264
2265         prev = xchg(&fs_info->pending_changes, 0);
2266         if (!prev)
2267                 return;
2268
2269         bit = 1 << BTRFS_PENDING_SET_INODE_MAP_CACHE;
2270         if (prev & bit)
2271                 btrfs_set_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2272         prev &= ~bit;
2273
2274         bit = 1 << BTRFS_PENDING_CLEAR_INODE_MAP_CACHE;
2275         if (prev & bit)
2276                 btrfs_clear_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2277         prev &= ~bit;
2278
2279         bit = 1 << BTRFS_PENDING_COMMIT;
2280         if (prev & bit)
2281                 btrfs_debug(fs_info, "pending commit done");
2282         prev &= ~bit;
2283
2284         if (prev)
2285                 btrfs_warn(fs_info,
2286                         "unknown pending changes left 0x%lx, ignoring", prev);
2287 }