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