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