fs/btrfs/transaction.c

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