fs/btrfs/transaction.c

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