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