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
  35 #define BTRFS_ROOT_TRANS_TAG 0
  36
  37 void put_transaction(struct btrfs_transaction *transaction)
  38 {
  39         WARN_ON(atomic_read(&transaction->use_count) == 0);
  40         if (atomic_dec_and_test(&transaction->use_count)) {
  41                 BUG_ON(!list_empty(&transaction->list));
  42                 WARN_ON(transaction->delayed_refs.root.rb_node);
  43                 memset(transaction, 0, sizeof(*transaction));
  44                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  45         }
  46 }
  47
  48 static noinline void switch_commit_root(struct btrfs_root *root)
  49 {
  50         free_extent_buffer(root->commit_root);
  51         root->commit_root = btrfs_root_node(root);
  52 }
  53
  54 /*
  55  * either allocate a new transaction or hop into the existing one
  56  */
  57 static noinline int join_transaction(struct btrfs_root *root, int type)
  58 {
  59         struct btrfs_transaction *cur_trans;
  60         struct btrfs_fs_info *fs_info = root->fs_info;
  61
  62         spin_lock(&fs_info->trans_lock);
  63 loop:
  64         /* The file system has been taken offline. No new transactions. */
  65         if (fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
  66                 spin_unlock(&fs_info->trans_lock);
  67                 return -EROFS;
  68         }
  69
  70         if (fs_info->trans_no_join) {
  71                 /*
  72                  * If we are JOIN_NOLOCK we're already committing a current
  73                  * transaction, we just need a handle to deal with something
  74                  * when committing the transaction, such as inode cache and
  75                  * space cache. It is a special case.
  76                  */
  77                 if (type != TRANS_JOIN_NOLOCK) {
  78                         spin_unlock(&fs_info->trans_lock);
  79                         return -EBUSY;
  80                 }
  81         }
  82
  83         cur_trans = fs_info->running_transaction;
  84         if (cur_trans) {
  85                 if (cur_trans->aborted) {
  86                         spin_unlock(&fs_info->trans_lock);
  87                         return cur_trans->aborted;
  88                 }
  89                 atomic_inc(&cur_trans->use_count);
  90                 atomic_inc(&cur_trans->num_writers);
  91                 cur_trans->num_joined++;
  92                 spin_unlock(&fs_info->trans_lock);
  93                 return 0;
  94         }
  95         spin_unlock(&fs_info->trans_lock);
  96
  97         /*
  98          * If we are ATTACH, we just want to catch the current transaction,
  99          * and commit it. If there is no transaction, just return ENOENT.
 100          */
 101         if (type == TRANS_ATTACH)
 102                 return -ENOENT;
 103
 104         cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
 105         if (!cur_trans)
 106                 return -ENOMEM;
 107
 108         spin_lock(&fs_info->trans_lock);
 109         if (fs_info->running_transaction) {
 110                 /*
 111                  * someone started a transaction after we unlocked.  Make sure
 112                  * to redo the trans_no_join checks above
 113                  */
 114                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 115                 cur_trans = fs_info->running_transaction;
 116                 goto loop;
 117         } else if (fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
 118                 spin_unlock(&fs_info->trans_lock);
 119                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 120                 return -EROFS;
 121         }
 122
 123         atomic_set(&cur_trans->num_writers, 1);
 124         cur_trans->num_joined = 0;
 125         init_waitqueue_head(&cur_trans->writer_wait);
 126         init_waitqueue_head(&cur_trans->commit_wait);
 127         cur_trans->in_commit = 0;
 128         cur_trans->blocked = 0;
 129         /*
 130          * One for this trans handle, one so it will live on until we
 131          * commit the transaction.
 132          */
 133         atomic_set(&cur_trans->use_count, 2);
 134         cur_trans->commit_done = 0;
 135         cur_trans->start_time = get_seconds();
 136
 137         cur_trans->delayed_refs.root = RB_ROOT;
 138         cur_trans->delayed_refs.num_entries = 0;
 139         cur_trans->delayed_refs.num_heads_ready = 0;
 140         cur_trans->delayed_refs.num_heads = 0;
 141         cur_trans->delayed_refs.flushing = 0;
 142         cur_trans->delayed_refs.run_delayed_start = 0;
 143
 144         /*
 145          * although the tree mod log is per file system and not per transaction,
 146          * the log must never go across transaction boundaries.
 147          */
 148         smp_mb();
 149         if (!list_empty(&fs_info->tree_mod_seq_list))
 150                 WARN(1, KERN_ERR "btrfs: tree_mod_seq_list not empty when "
 151                         "creating a fresh transaction\n");
 152         if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
 153                 WARN(1, KERN_ERR "btrfs: tree_mod_log rb tree not empty when "
 154                         "creating a fresh transaction\n");
 155         atomic_set(&fs_info->tree_mod_seq, 0);
 156
 157         spin_lock_init(&cur_trans->commit_lock);
 158         spin_lock_init(&cur_trans->delayed_refs.lock);
 159
 160         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 161         list_add_tail(&cur_trans->list, &fs_info->trans_list);
 162         extent_io_tree_init(&cur_trans->dirty_pages,
 163                              fs_info->btree_inode->i_mapping);
 164         fs_info->generation++;
 165         cur_trans->transid = fs_info->generation;
 166         fs_info->running_transaction = cur_trans;
 167         cur_trans->aborted = 0;
 168         spin_unlock(&fs_info->trans_lock);
 169
 170         return 0;
 171 }
 172
 173 /*
 174  * this does all the record keeping required to make sure that a reference
 175  * counted root is properly recorded in a given transaction.  This is required
 176  * to make sure the old root from before we joined the transaction is deleted
 177  * when the transaction commits
 178  */
 179 static int record_root_in_trans(struct btrfs_trans_handle *trans,
 180                                struct btrfs_root *root)
 181 {
 182         if (root->ref_cows && root->last_trans < trans->transid) {
 183                 WARN_ON(root == root->fs_info->extent_root);
 184                 WARN_ON(root->commit_root != root->node);
 185
 186                 /*
 187                  * see below for in_trans_setup usage rules
 188                  * we have the reloc mutex held now, so there
 189                  * is only one writer in this function
 190                  */
 191                 root->in_trans_setup = 1;
 192
 193                 /* make sure readers find in_trans_setup before
 194                  * they find our root->last_trans update
 195                  */
 196                 smp_wmb();
 197
 198                 spin_lock(&root->fs_info->fs_roots_radix_lock);
 199                 if (root->last_trans == trans->transid) {
 200                         spin_unlock(&root->fs_info->fs_roots_radix_lock);
 201                         return 0;
 202                 }
 203                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 204                            (unsigned long)root->root_key.objectid,
 205                            BTRFS_ROOT_TRANS_TAG);
 206                 spin_unlock(&root->fs_info->fs_roots_radix_lock);
 207                 root->last_trans = trans->transid;
 208
 209                 /* this is pretty tricky.  We don't want to
 210                  * take the relocation lock in btrfs_record_root_in_trans
 211                  * unless we're really doing the first setup for this root in
 212                  * this transaction.
 213                  *
 214                  * Normally we'd use root->last_trans as a flag to decide
 215                  * if we want to take the expensive mutex.
 216                  *
 217                  * But, we have to set root->last_trans before we
 218                  * init the relocation root, otherwise, we trip over warnings
 219                  * in ctree.c.  The solution used here is to flag ourselves
 220                  * with root->in_trans_setup.  When this is 1, we're still
 221                  * fixing up the reloc trees and everyone must wait.
 222                  *
 223                  * When this is zero, they can trust root->last_trans and fly
 224                  * through btrfs_record_root_in_trans without having to take the
 225                  * lock.  smp_wmb() makes sure that all the writes above are
 226                  * done before we pop in the zero below
 227                  */
 228                 btrfs_init_reloc_root(trans, root);
 229                 smp_wmb();
 230                 root->in_trans_setup = 0;
 231         }
 232         return 0;
 233 }
 234
 235
 236 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 237                                struct btrfs_root *root)
 238 {
 239         if (!root->ref_cows)
 240                 return 0;
 241
 242         /*
 243          * see record_root_in_trans for comments about in_trans_setup usage
 244          * and barriers
 245          */
 246         smp_rmb();
 247         if (root->last_trans == trans->transid &&
 248             !root->in_trans_setup)
 249                 return 0;
 250
 251         mutex_lock(&root->fs_info->reloc_mutex);
 252         record_root_in_trans(trans, root);
 253         mutex_unlock(&root->fs_info->reloc_mutex);
 254
 255         return 0;
 256 }
 257
 258 /* wait for commit against the current transaction to become unblocked
 259  * when this is done, it is safe to start a new transaction, but the current
 260  * transaction might not be fully on disk.
 261  */
 262 static void wait_current_trans(struct btrfs_root *root)
 263 {
 264         struct btrfs_transaction *cur_trans;
 265
 266         spin_lock(&root->fs_info->trans_lock);
 267         cur_trans = root->fs_info->running_transaction;
 268         if (cur_trans && cur_trans->blocked) {
 269                 atomic_inc(&cur_trans->use_count);
 270                 spin_unlock(&root->fs_info->trans_lock);
 271
 272                 wait_event(root->fs_info->transaction_wait,
 273                            !cur_trans->blocked);
 274                 put_transaction(cur_trans);
 275         } else {
 276                 spin_unlock(&root->fs_info->trans_lock);
 277         }
 278 }
 279
 280 static int may_wait_transaction(struct btrfs_root *root, int type)
 281 {
 282         if (root->fs_info->log_root_recovering)
 283                 return 0;
 284
 285         if (type == TRANS_USERSPACE)
 286                 return 1;
 287
 288         if (type == TRANS_START &&
 289             !atomic_read(&root->fs_info->open_ioctl_trans))
 290                 return 1;
 291
 292         return 0;
 293 }
 294
 295 static struct btrfs_trans_handle *
 296 start_transaction(struct btrfs_root *root, u64 num_items, int type,
 297                   enum btrfs_reserve_flush_enum flush)
 298 {
 299         struct btrfs_trans_handle *h;
 300         struct btrfs_transaction *cur_trans;
 301         u64 num_bytes = 0;
 302         int ret;
 303         u64 qgroup_reserved = 0;
 304
 305         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR)
 306                 return ERR_PTR(-EROFS);
 307
 308         if (current->journal_info) {
 309                 WARN_ON(type != TRANS_JOIN && type != TRANS_JOIN_NOLOCK);
 310                 h = current->journal_info;
 311                 h->use_count++;
 312                 WARN_ON(h->use_count > 2);
 313                 h->orig_rsv = h->block_rsv;
 314                 h->block_rsv = NULL;
 315                 goto got_it;
 316         }
 317
 318         /*
 319          * Do the reservation before we join the transaction so we can do all
 320          * the appropriate flushing if need be.
 321          */
 322         if (num_items > 0 && root != root->fs_info->chunk_root) {
 323                 if (root->fs_info->quota_enabled &&
 324                     is_fstree(root->root_key.objectid)) {
 325                         qgroup_reserved = num_items * root->leafsize;
 326                         ret = btrfs_qgroup_reserve(root, qgroup_reserved);
 327                         if (ret)
 328                                 return ERR_PTR(ret);
 329                 }
 330
 331                 num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
 332                 ret = btrfs_block_rsv_add(root,
 333                                           &root->fs_info->trans_block_rsv,
 334                                           num_bytes, flush);
 335                 if (ret)
 336                         return ERR_PTR(ret);
 337         }
 338 again:
 339         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 340         if (!h)
 341                 return ERR_PTR(-ENOMEM);
 342
 343         /*
 344          * If we are JOIN_NOLOCK we're already committing a transaction and
 345          * waiting on this guy, so we don't need to do the sb_start_intwrite
 346          * because we're already holding a ref.  We need this because we could
 347          * have raced in and did an fsync() on a file which can kick a commit
 348          * and then we deadlock with somebody doing a freeze.
 349          *
 350          * If we are ATTACH, it means we just want to catch the current
 351          * transaction and commit it, so we needn't do sb_start_intwrite().
 352          */
 353         if (type < TRANS_JOIN_NOLOCK)
 354                 sb_start_intwrite(root->fs_info->sb);
 355
 356         if (may_wait_transaction(root, type))
 357                 wait_current_trans(root);
 358
 359         do {
 360                 ret = join_transaction(root, type);
 361                 if (ret == -EBUSY)
 362                         wait_current_trans(root);
 363         } while (ret == -EBUSY);
 364
 365         if (ret < 0) {
 366                 /* We must get the transaction if we are JOIN_NOLOCK. */
 367                 BUG_ON(type == TRANS_JOIN_NOLOCK);
 368
 369                 if (type < TRANS_JOIN_NOLOCK)
 370                         sb_end_intwrite(root->fs_info->sb);
 371                 kmem_cache_free(btrfs_trans_handle_cachep, h);
 372                 return ERR_PTR(ret);
 373         }
 374
 375         cur_trans = root->fs_info->running_transaction;
 376
 377         h->transid = cur_trans->transid;
 378         h->transaction = cur_trans;
 379         h->blocks_used = 0;
 380         h->bytes_reserved = 0;
 381         h->root = root;
 382         h->delayed_ref_updates = 0;
 383         h->use_count = 1;
 384         h->adding_csums = 0;
 385         h->block_rsv = NULL;
 386         h->orig_rsv = NULL;
 387         h->aborted = 0;
 388         h->qgroup_reserved = qgroup_reserved;
 389         h->delayed_ref_elem.seq = 0;
 390         h->type = type;
 391         INIT_LIST_HEAD(&h->qgroup_ref_list);
 392         INIT_LIST_HEAD(&h->new_bgs);
 393
 394         smp_mb();
 395         if (cur_trans->blocked && may_wait_transaction(root, type)) {
 396                 btrfs_commit_transaction(h, root);
 397                 goto again;
 398         }
 399
 400         if (num_bytes) {
 401                 trace_btrfs_space_reservation(root->fs_info, "transaction",
 402                                               h->transid, num_bytes, 1);
 403                 h->block_rsv = &root->fs_info->trans_block_rsv;
 404                 h->bytes_reserved = num_bytes;
 405         }
 406
 407 got_it:
 408         btrfs_record_root_in_trans(h, root);
 409
 410         if (!current->journal_info && type != TRANS_USERSPACE)
 411                 current->journal_info = h;
 412         return h;
 413 }
 414
 415 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 416                                                    int num_items)
 417 {
 418         return start_transaction(root, num_items, TRANS_START,
 419                                  BTRFS_RESERVE_FLUSH_ALL);
 420 }
 421
 422 struct btrfs_trans_handle *btrfs_start_transaction_lflush(
 423                                         struct btrfs_root *root, int num_items)
 424 {
 425         return start_transaction(root, num_items, TRANS_START,
 426                                  BTRFS_RESERVE_FLUSH_LIMIT);
 427 }
 428
 429 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 430 {
 431         return start_transaction(root, 0, TRANS_JOIN, 0);
 432 }
 433
 434 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 435 {
 436         return start_transaction(root, 0, TRANS_JOIN_NOLOCK, 0);
 437 }
 438
 439 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 440 {
 441         return start_transaction(root, 0, TRANS_USERSPACE, 0);
 442 }
 443
 444 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
 445 {
 446         return start_transaction(root, 0, TRANS_ATTACH, 0);
 447 }
 448
 449 /* wait for a transaction commit to be fully complete */
 450 static noinline void wait_for_commit(struct btrfs_root *root,
 451                                     struct btrfs_transaction *commit)
 452 {
 453         wait_event(commit->commit_wait, commit->commit_done);
 454 }
 455
 456 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 457 {
 458         struct btrfs_transaction *cur_trans = NULL, *t;
 459         int ret;
 460
 461         ret = 0;
 462         if (transid) {
 463                 if (transid <= root->fs_info->last_trans_committed)
 464                         goto out;
 465
 466                 /* find specified transaction */
 467                 spin_lock(&root->fs_info->trans_lock);
 468                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 469                         if (t->transid == transid) {
 470                                 cur_trans = t;
 471                                 atomic_inc(&cur_trans->use_count);
 472                                 break;
 473                         }
 474                         if (t->transid > transid)
 475                                 break;
 476                 }
 477                 spin_unlock(&root->fs_info->trans_lock);
 478                 ret = -EINVAL;
 479                 if (!cur_trans)
 480                         goto out;  /* bad transid */
 481         } else {
 482                 /* find newest transaction that is committing | committed */
 483                 spin_lock(&root->fs_info->trans_lock);
 484                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 485                                             list) {
 486                         if (t->in_commit) {
 487                                 if (t->commit_done)
 488                                         break;
 489                                 cur_trans = t;
 490                                 atomic_inc(&cur_trans->use_count);
 491                                 break;
 492                         }
 493                 }
 494                 spin_unlock(&root->fs_info->trans_lock);
 495                 if (!cur_trans)
 496                         goto out;  /* nothing committing|committed */
 497         }
 498
 499         wait_for_commit(root, cur_trans);
 500
 501         put_transaction(cur_trans);
 502         ret = 0;
 503 out:
 504         return ret;
 505 }
 506
 507 void btrfs_throttle(struct btrfs_root *root)
 508 {
 509         if (!atomic_read(&root->fs_info->open_ioctl_trans))
 510                 wait_current_trans(root);
 511 }
 512
 513 static int should_end_transaction(struct btrfs_trans_handle *trans,
 514                                   struct btrfs_root *root)
 515 {
 516         int ret;
 517
 518         ret = btrfs_block_rsv_check(root, &root->fs_info->global_block_rsv, 5);
 519         return ret ? 1 : 0;
 520 }
 521
 522 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 523                                  struct btrfs_root *root)
 524 {
 525         struct btrfs_transaction *cur_trans = trans->transaction;
 526         int updates;
 527         int err;
 528
 529         smp_mb();
 530         if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
 531                 return 1;
 532
 533         updates = trans->delayed_ref_updates;
 534         trans->delayed_ref_updates = 0;
 535         if (updates) {
 536                 err = btrfs_run_delayed_refs(trans, root, updates);
 537                 if (err) /* Error code will also eval true */
 538                         return err;
 539         }
 540
 541         return should_end_transaction(trans, root);
 542 }
 543
 544 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 545                           struct btrfs_root *root, int throttle)
 546 {
 547         struct btrfs_transaction *cur_trans = trans->transaction;
 548         struct btrfs_fs_info *info = root->fs_info;
 549         int count = 0;
 550         int lock = (trans->type != TRANS_JOIN_NOLOCK);
 551         int err = 0;
 552
 553         if (--trans->use_count) {
 554                 trans->block_rsv = trans->orig_rsv;
 555                 return 0;
 556         }
 557
 558         /*
 559          * do the qgroup accounting as early as possible
 560          */
 561         err = btrfs_delayed_refs_qgroup_accounting(trans, info);
 562
 563         btrfs_trans_release_metadata(trans, root);
 564         trans->block_rsv = NULL;
 565         /*
 566          * the same root has to be passed to start_transaction and
 567          * end_transaction. Subvolume quota depends on this.
 568          */
 569         WARN_ON(trans->root != root);
 570
 571         if (trans->qgroup_reserved) {
 572                 btrfs_qgroup_free(root, trans->qgroup_reserved);
 573                 trans->qgroup_reserved = 0;
 574         }
 575
 576         if (!list_empty(&trans->new_bgs))
 577                 btrfs_create_pending_block_groups(trans, root);
 578
 579         while (count < 2) {
 580                 unsigned long cur = trans->delayed_ref_updates;
 581                 trans->delayed_ref_updates = 0;
 582                 if (cur &&
 583                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 584                         trans->delayed_ref_updates = 0;
 585                         btrfs_run_delayed_refs(trans, root, cur);
 586                 } else {
 587                         break;
 588                 }
 589                 count++;
 590         }
 591         btrfs_trans_release_metadata(trans, root);
 592         trans->block_rsv = NULL;
 593
 594         if (!list_empty(&trans->new_bgs))
 595                 btrfs_create_pending_block_groups(trans, root);
 596
 597         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
 598             should_end_transaction(trans, root)) {
 599                 trans->transaction->blocked = 1;
 600                 smp_wmb();
 601         }
 602
 603         if (lock && cur_trans->blocked && !cur_trans->in_commit) {
 604                 if (throttle) {
 605                         /*
 606                          * We may race with somebody else here so end up having
 607                          * to call end_transaction on ourselves again, so inc
 608                          * our use_count.
 609                          */
 610                         trans->use_count++;
 611                         return btrfs_commit_transaction(trans, root);
 612                 } else {
 613                         wake_up_process(info->transaction_kthread);
 614                 }
 615         }
 616
 617         if (trans->type < TRANS_JOIN_NOLOCK)
 618                 sb_end_intwrite(root->fs_info->sb);
 619
 620         WARN_ON(cur_trans != info->running_transaction);
 621         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 622         atomic_dec(&cur_trans->num_writers);
 623
 624         smp_mb();
 625         if (waitqueue_active(&cur_trans->writer_wait))
 626                 wake_up(&cur_trans->writer_wait);
 627         put_transaction(cur_trans);
 628
 629         if (current->journal_info == trans)
 630                 current->journal_info = NULL;
 631
 632         if (throttle)
 633                 btrfs_run_delayed_iputs(root);
 634
 635         if (trans->aborted ||
 636             root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
 637                 err = -EIO;
 638         }
 639         assert_qgroups_uptodate(trans);
 640
 641         memset(trans, 0, sizeof(*trans));
 642         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 643         return err;
 644 }
 645
 646 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 647                           struct btrfs_root *root)
 648 {
 649         int ret;
 650
 651         ret = __btrfs_end_transaction(trans, root, 0);
 652         if (ret)
 653                 return ret;
 654         return 0;
 655 }
 656
 657 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 658                                    struct btrfs_root *root)
 659 {
 660         int ret;
 661
 662         ret = __btrfs_end_transaction(trans, root, 1);
 663         if (ret)
 664                 return ret;
 665         return 0;
 666 }
 667
 668 int btrfs_end_transaction_dmeta(struct btrfs_trans_handle *trans,
 669                                 struct btrfs_root *root)
 670 {
 671         return __btrfs_end_transaction(trans, root, 1);
 672 }
 673
 674 /*
 675  * when btree blocks are allocated, they have some corresponding bits set for
 676  * them in one of two extent_io trees.  This is used to make sure all of
 677  * those extents are sent to disk but does not wait on them
 678  */
 679 int btrfs_write_marked_extents(struct btrfs_root *root,
 680                                struct extent_io_tree *dirty_pages, int mark)
 681 {
 682         int err = 0;
 683         int werr = 0;
 684         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 685         struct extent_state *cached_state = NULL;
 686         u64 start = 0;
 687         u64 end;
 688
 689         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 690                                       mark, &cached_state)) {
 691                 convert_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT,
 692                                    mark, &cached_state, GFP_NOFS);
 693                 cached_state = NULL;
 694                 err = filemap_fdatawrite_range(mapping, start, end);
 695                 if (err)
 696                         werr = err;
 697                 cond_resched();
 698                 start = end + 1;
 699         }
 700         if (err)
 701                 werr = err;
 702         return werr;
 703 }
 704
 705 /*
 706  * when btree blocks are allocated, they have some corresponding bits set for
 707  * them in one of two extent_io trees.  This is used to make sure all of
 708  * those extents are on disk for transaction or log commit.  We wait
 709  * on all the pages and clear them from the dirty pages state tree
 710  */
 711 int btrfs_wait_marked_extents(struct btrfs_root *root,
 712                               struct extent_io_tree *dirty_pages, int mark)
 713 {
 714         int err = 0;
 715         int werr = 0;
 716         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 717         struct extent_state *cached_state = NULL;
 718         u64 start = 0;
 719         u64 end;
 720
 721         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 722                                       EXTENT_NEED_WAIT, &cached_state)) {
 723                 clear_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT,
 724                                  0, 0, &cached_state, GFP_NOFS);
 725                 err = filemap_fdatawait_range(mapping, start, end);
 726                 if (err)
 727                         werr = err;
 728                 cond_resched();
 729                 start = end + 1;
 730         }
 731         if (err)
 732                 werr = err;
 733         return werr;
 734 }
 735
 736 /*
 737  * when btree blocks are allocated, they have some corresponding bits set for
 738  * them in one of two extent_io trees.  This is used to make sure all of
 739  * those extents are on disk for transaction or log commit
 740  */
 741 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 742                                 struct extent_io_tree *dirty_pages, int mark)
 743 {
 744         int ret;
 745         int ret2;
 746
 747         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 748         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 749
 750         if (ret)
 751                 return ret;
 752         if (ret2)
 753                 return ret2;
 754         return 0;
 755 }
 756
 757 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 758                                      struct btrfs_root *root)
 759 {
 760         if (!trans || !trans->transaction) {
 761                 struct inode *btree_inode;
 762                 btree_inode = root->fs_info->btree_inode;
 763                 return filemap_write_and_wait(btree_inode->i_mapping);
 764         }
 765         return btrfs_write_and_wait_marked_extents(root,
 766                                            &trans->transaction->dirty_pages,
 767                                            EXTENT_DIRTY);
 768 }
 769
 770 /*
 771  * this is used to update the root pointer in the tree of tree roots.
 772  *
 773  * But, in the case of the extent allocation tree, updating the root
 774  * pointer may allocate blocks which may change the root of the extent
 775  * allocation tree.
 776  *
 777  * So, this loops and repeats and makes sure the cowonly root didn't
 778  * change while the root pointer was being updated in the metadata.
 779  */
 780 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 781                                struct btrfs_root *root)
 782 {
 783         int ret;
 784         u64 old_root_bytenr;
 785         u64 old_root_used;
 786         struct btrfs_root *tree_root = root->fs_info->tree_root;
 787
 788         old_root_used = btrfs_root_used(&root->root_item);
 789         btrfs_write_dirty_block_groups(trans, root);
 790
 791         while (1) {
 792                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 793                 if (old_root_bytenr == root->node->start &&
 794                     old_root_used == btrfs_root_used(&root->root_item))
 795                         break;
 796
 797                 btrfs_set_root_node(&root->root_item, root->node);
 798                 ret = btrfs_update_root(trans, tree_root,
 799                                         &root->root_key,
 800                                         &root->root_item);
 801                 if (ret)
 802                         return ret;
 803
 804                 old_root_used = btrfs_root_used(&root->root_item);
 805                 ret = btrfs_write_dirty_block_groups(trans, root);
 806                 if (ret)
 807                         return ret;
 808         }
 809
 810         if (root != root->fs_info->extent_root)
 811                 switch_commit_root(root);
 812
 813         return 0;
 814 }
 815
 816 /*
 817  * update all the cowonly tree roots on disk
 818  *
 819  * The error handling in this function may not be obvious. Any of the
 820  * failures will cause the file system to go offline. We still need
 821  * to clean up the delayed refs.
 822  */
 823 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 824                                          struct btrfs_root *root)
 825 {
 826         struct btrfs_fs_info *fs_info = root->fs_info;
 827         struct list_head *next;
 828         struct extent_buffer *eb;
 829         int ret;
 830
 831         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 832         if (ret)
 833                 return ret;
 834
 835         eb = btrfs_lock_root_node(fs_info->tree_root);
 836         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
 837                               0, &eb);
 838         btrfs_tree_unlock(eb);
 839         free_extent_buffer(eb);
 840
 841         if (ret)
 842                 return ret;
 843
 844         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 845         if (ret)
 846                 return ret;
 847
 848         ret = btrfs_run_dev_stats(trans, root->fs_info);
 849         WARN_ON(ret);
 850         ret = btrfs_run_dev_replace(trans, root->fs_info);
 851         WARN_ON(ret);
 852
 853         ret = btrfs_run_qgroups(trans, root->fs_info);
 854         BUG_ON(ret);
 855
 856         /* run_qgroups might have added some more refs */
 857         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 858         BUG_ON(ret);
 859
 860         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 861                 next = fs_info->dirty_cowonly_roots.next;
 862                 list_del_init(next);
 863                 root = list_entry(next, struct btrfs_root, dirty_list);
 864
 865                 ret = update_cowonly_root(trans, root);
 866                 if (ret)
 867                         return ret;
 868         }
 869
 870         down_write(&fs_info->extent_commit_sem);
 871         switch_commit_root(fs_info->extent_root);
 872         up_write(&fs_info->extent_commit_sem);
 873
 874         btrfs_after_dev_replace_commit(fs_info);
 875
 876         return 0;
 877 }
 878
 879 /*
 880  * dead roots are old snapshots that need to be deleted.  This allocates
 881  * a dirty root struct and adds it into the list of dead roots that need to
 882  * be deleted
 883  */
 884 int btrfs_add_dead_root(struct btrfs_root *root)
 885 {
 886         spin_lock(&root->fs_info->trans_lock);
 887         list_add(&root->root_list, &root->fs_info->dead_roots);
 888         spin_unlock(&root->fs_info->trans_lock);
 889         return 0;
 890 }
 891
 892 /*
 893  * update all the cowonly tree roots on disk
 894  */
 895 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 896                                     struct btrfs_root *root)
 897 {
 898         struct btrfs_root *gang[8];
 899         struct btrfs_fs_info *fs_info = root->fs_info;
 900         int i;
 901         int ret;
 902         int err = 0;
 903
 904         spin_lock(&fs_info->fs_roots_radix_lock);
 905         while (1) {
 906                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 907                                                  (void **)gang, 0,
 908                                                  ARRAY_SIZE(gang),
 909                                                  BTRFS_ROOT_TRANS_TAG);
 910                 if (ret == 0)
 911                         break;
 912                 for (i = 0; i < ret; i++) {
 913                         root = gang[i];
 914                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 915                                         (unsigned long)root->root_key.objectid,
 916                                         BTRFS_ROOT_TRANS_TAG);
 917                         spin_unlock(&fs_info->fs_roots_radix_lock);
 918
 919                         btrfs_free_log(trans, root);
 920                         btrfs_update_reloc_root(trans, root);
 921                         btrfs_orphan_commit_root(trans, root);
 922
 923                         btrfs_save_ino_cache(root, trans);
 924
 925                         /* see comments in should_cow_block() */
 926                         root->force_cow = 0;
 927                         smp_wmb();
 928
 929                         if (root->commit_root != root->node) {
 930                                 mutex_lock(&root->fs_commit_mutex);
 931                                 switch_commit_root(root);
 932                                 btrfs_unpin_free_ino(root);
 933                                 mutex_unlock(&root->fs_commit_mutex);
 934
 935                                 btrfs_set_root_node(&root->root_item,
 936                                                     root->node);
 937                         }
 938
 939                         err = btrfs_update_root(trans, fs_info->tree_root,
 940                                                 &root->root_key,
 941                                                 &root->root_item);
 942                         spin_lock(&fs_info->fs_roots_radix_lock);
 943                         if (err)
 944                                 break;
 945                 }
 946         }
 947         spin_unlock(&fs_info->fs_roots_radix_lock);
 948         return err;
 949 }
 950
 951 /*
 952  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 953  * otherwise every leaf in the btree is read and defragged.
 954  */
 955 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 956 {
 957         struct btrfs_fs_info *info = root->fs_info;
 958         struct btrfs_trans_handle *trans;
 959         int ret;
 960
 961         if (xchg(&root->defrag_running, 1))
 962                 return 0;
 963
 964         while (1) {
 965                 trans = btrfs_start_transaction(root, 0);
 966                 if (IS_ERR(trans))
 967                         return PTR_ERR(trans);
 968
 969                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 970
 971                 btrfs_end_transaction(trans, root);
 972                 btrfs_btree_balance_dirty(info->tree_root);
 973                 cond_resched();
 974
 975                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
 976                         break;
 977         }
 978         root->defrag_running = 0;
 979         return ret;
 980 }
 981
 982 /*
 983  * new snapshots need to be created at a very specific time in the
 984  * transaction commit.  This does the actual creation
 985  */
 986 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 987                                    struct btrfs_fs_info *fs_info,
 988                                    struct btrfs_pending_snapshot *pending)
 989 {
 990         struct btrfs_key key;
 991         struct btrfs_root_item *new_root_item;
 992         struct btrfs_root *tree_root = fs_info->tree_root;
 993         struct btrfs_root *root = pending->root;
 994         struct btrfs_root *parent_root;
 995         struct btrfs_block_rsv *rsv;
 996         struct inode *parent_inode;
 997         struct btrfs_path *path;
 998         struct btrfs_dir_item *dir_item;
 999         struct dentry *parent;
1000         struct dentry *dentry;
1001         struct extent_buffer *tmp;
1002         struct extent_buffer *old;
1003         struct timespec cur_time = CURRENT_TIME;
1004         int ret;
1005         u64 to_reserve = 0;
1006         u64 index = 0;
1007         u64 objectid;
1008         u64 root_flags;
1009         uuid_le new_uuid;
1010
1011         path = btrfs_alloc_path();
1012         if (!path) {
1013                 ret = pending->error = -ENOMEM;
1014                 goto path_alloc_fail;
1015         }
1016
1017         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
1018         if (!new_root_item) {
1019                 ret = pending->error = -ENOMEM;
1020                 goto root_item_alloc_fail;
1021         }
1022
1023         ret = btrfs_find_free_objectid(tree_root, &objectid);
1024         if (ret) {
1025                 pending->error = ret;
1026                 goto no_free_objectid;
1027         }
1028
1029         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
1030
1031         if (to_reserve > 0) {
1032                 ret = btrfs_block_rsv_add(root, &pending->block_rsv,
1033                                           to_reserve,
1034                                           BTRFS_RESERVE_NO_FLUSH);
1035                 if (ret) {
1036                         pending->error = ret;
1037                         goto no_free_objectid;
1038                 }
1039         }
1040
1041         ret = btrfs_qgroup_inherit(trans, fs_info, root->root_key.objectid,
1042                                    objectid, pending->inherit);
1043         if (ret) {
1044                 pending->error = ret;
1045                 goto no_free_objectid;
1046         }
1047
1048         key.objectid = objectid;
1049         key.offset = (u64)-1;
1050         key.type = BTRFS_ROOT_ITEM_KEY;
1051
1052         rsv = trans->block_rsv;
1053         trans->block_rsv = &pending->block_rsv;
1054
1055         dentry = pending->dentry;
1056         parent = dget_parent(dentry);
1057         parent_inode = parent->d_inode;
1058         parent_root = BTRFS_I(parent_inode)->root;
1059         record_root_in_trans(trans, parent_root);
1060
1061         /*
1062          * insert the directory item
1063          */
1064         ret = btrfs_set_inode_index(parent_inode, &index);
1065         BUG_ON(ret); /* -ENOMEM */
1066
1067         /* check if there is a file/dir which has the same name. */
1068         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1069                                          btrfs_ino(parent_inode),
1070                                          dentry->d_name.name,
1071                                          dentry->d_name.len, 0);
1072         if (dir_item != NULL && !IS_ERR(dir_item)) {
1073                 pending->error = -EEXIST;
1074                 goto fail;
1075         } else if (IS_ERR(dir_item)) {
1076                 ret = PTR_ERR(dir_item);
1077                 btrfs_abort_transaction(trans, root, ret);
1078                 goto fail;
1079         }
1080         btrfs_release_path(path);
1081
1082         /*
1083          * pull in the delayed directory update
1084          * and the delayed inode item
1085          * otherwise we corrupt the FS during
1086          * snapshot
1087          */
1088         ret = btrfs_run_delayed_items(trans, root);
1089         if (ret) {      /* Transaction aborted */
1090                 btrfs_abort_transaction(trans, root, ret);
1091                 goto fail;
1092         }
1093
1094         record_root_in_trans(trans, root);
1095         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1096         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1097         btrfs_check_and_init_root_item(new_root_item);
1098
1099         root_flags = btrfs_root_flags(new_root_item);
1100         if (pending->readonly)
1101                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1102         else
1103                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1104         btrfs_set_root_flags(new_root_item, root_flags);
1105
1106         btrfs_set_root_generation_v2(new_root_item,
1107                         trans->transid);
1108         uuid_le_gen(&new_uuid);
1109         memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1110         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1111                         BTRFS_UUID_SIZE);
1112         new_root_item->otime.sec = cpu_to_le64(cur_time.tv_sec);
1113         new_root_item->otime.nsec = cpu_to_le32(cur_time.tv_nsec);
1114         btrfs_set_root_otransid(new_root_item, trans->transid);
1115         memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1116         memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1117         btrfs_set_root_stransid(new_root_item, 0);
1118         btrfs_set_root_rtransid(new_root_item, 0);
1119
1120         old = btrfs_lock_root_node(root);
1121         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1122         if (ret) {
1123                 btrfs_tree_unlock(old);
1124                 free_extent_buffer(old);
1125                 btrfs_abort_transaction(trans, root, ret);
1126                 goto fail;
1127         }
1128
1129         btrfs_set_lock_blocking(old);
1130
1131         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1132         /* clean up in any case */
1133         btrfs_tree_unlock(old);
1134         free_extent_buffer(old);
1135         if (ret) {
1136                 btrfs_abort_transaction(trans, root, ret);
1137                 goto fail;
1138         }
1139
1140         /* see comments in should_cow_block() */
1141         root->force_cow = 1;
1142         smp_wmb();
1143
1144         btrfs_set_root_node(new_root_item, tmp);
1145         /* record when the snapshot was created in key.offset */
1146         key.offset = trans->transid;
1147         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1148         btrfs_tree_unlock(tmp);
1149         free_extent_buffer(tmp);
1150         if (ret) {
1151                 btrfs_abort_transaction(trans, root, ret);
1152                 goto fail;
1153         }
1154
1155         /*
1156          * insert root back/forward references
1157          */
1158         ret = btrfs_add_root_ref(trans, tree_root, objectid,
1159                                  parent_root->root_key.objectid,
1160                                  btrfs_ino(parent_inode), index,
1161                                  dentry->d_name.name, dentry->d_name.len);
1162         if (ret) {
1163                 btrfs_abort_transaction(trans, root, ret);
1164                 goto fail;
1165         }
1166
1167         key.offset = (u64)-1;
1168         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1169         if (IS_ERR(pending->snap)) {
1170                 ret = PTR_ERR(pending->snap);
1171                 btrfs_abort_transaction(trans, root, ret);
1172                 goto fail;
1173         }
1174
1175         ret = btrfs_reloc_post_snapshot(trans, pending);
1176         if (ret) {
1177                 btrfs_abort_transaction(trans, root, ret);
1178                 goto fail;
1179         }
1180
1181         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1182         if (ret) {
1183                 btrfs_abort_transaction(trans, root, ret);
1184                 goto fail;
1185         }
1186
1187         ret = btrfs_insert_dir_item(trans, parent_root,
1188                                     dentry->d_name.name, dentry->d_name.len,
1189                                     parent_inode, &key,
1190                                     BTRFS_FT_DIR, index);
1191         /* We have check then name at the beginning, so it is impossible. */
1192         BUG_ON(ret == -EEXIST);
1193         if (ret) {
1194                 btrfs_abort_transaction(trans, root, ret);
1195                 goto fail;
1196         }
1197
1198         btrfs_i_size_write(parent_inode, parent_inode->i_size +
1199                                          dentry->d_name.len * 2);
1200         parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
1201         ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1202         if (ret)
1203                 btrfs_abort_transaction(trans, root, ret);
1204 fail:
1205         dput(parent);
1206         trans->block_rsv = rsv;
1207 no_free_objectid:
1208         kfree(new_root_item);
1209 root_item_alloc_fail:
1210         btrfs_free_path(path);
1211 path_alloc_fail:
1212         btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
1213         return ret;
1214 }
1215
1216 /*
1217  * create all the snapshots we've scheduled for creation
1218  */
1219 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1220                                              struct btrfs_fs_info *fs_info)
1221 {
1222         struct btrfs_pending_snapshot *pending;
1223         struct list_head *head = &trans->transaction->pending_snapshots;
1224
1225         list_for_each_entry(pending, head, list)
1226                 create_pending_snapshot(trans, fs_info, pending);
1227         return 0;
1228 }
1229
1230 static void update_super_roots(struct btrfs_root *root)
1231 {
1232         struct btrfs_root_item *root_item;
1233         struct btrfs_super_block *super;
1234
1235         super = root->fs_info->super_copy;
1236
1237         root_item = &root->fs_info->chunk_root->root_item;
1238         super->chunk_root = root_item->bytenr;
1239         super->chunk_root_generation = root_item->generation;
1240         super->chunk_root_level = root_item->level;
1241
1242         root_item = &root->fs_info->tree_root->root_item;
1243         super->root = root_item->bytenr;
1244         super->generation = root_item->generation;
1245         super->root_level = root_item->level;
1246         if (btrfs_test_opt(root, SPACE_CACHE))
1247                 super->cache_generation = root_item->generation;
1248 }
1249
1250 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1251 {
1252         int ret = 0;
1253         spin_lock(&info->trans_lock);
1254         if (info->running_transaction)
1255                 ret = info->running_transaction->in_commit;
1256         spin_unlock(&info->trans_lock);
1257         return ret;
1258 }
1259
1260 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1261 {
1262         int ret = 0;
1263         spin_lock(&info->trans_lock);
1264         if (info->running_transaction)
1265                 ret = info->running_transaction->blocked;
1266         spin_unlock(&info->trans_lock);
1267         return ret;
1268 }
1269
1270 /*
1271  * wait for the current transaction commit to start and block subsequent
1272  * transaction joins
1273  */
1274 static void wait_current_trans_commit_start(struct btrfs_root *root,
1275                                             struct btrfs_transaction *trans)
1276 {
1277         wait_event(root->fs_info->transaction_blocked_wait, trans->in_commit);
1278 }
1279
1280 /*
1281  * wait for the current transaction to start and then become unblocked.
1282  * caller holds ref.
1283  */
1284 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1285                                          struct btrfs_transaction *trans)
1286 {
1287         wait_event(root->fs_info->transaction_wait,
1288                    trans->commit_done || (trans->in_commit && !trans->blocked));
1289 }
1290
1291 /*
1292  * commit transactions asynchronously. once btrfs_commit_transaction_async
1293  * returns, any subsequent transaction will not be allowed to join.
1294  */
1295 struct btrfs_async_commit {
1296         struct btrfs_trans_handle *newtrans;
1297         struct btrfs_root *root;
1298         struct delayed_work work;
1299 };
1300
1301 static void do_async_commit(struct work_struct *work)
1302 {
1303         struct btrfs_async_commit *ac =
1304                 container_of(work, struct btrfs_async_commit, work.work);
1305
1306         /*
1307          * We've got freeze protection passed with the transaction.
1308          * Tell lockdep about it.
1309          */
1310         rwsem_acquire_read(
1311                 &ac->root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
1312                 0, 1, _THIS_IP_);
1313
1314         current->journal_info = ac->newtrans;
1315
1316         btrfs_commit_transaction(ac->newtrans, ac->root);
1317         kfree(ac);
1318 }
1319
1320 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1321                                    struct btrfs_root *root,
1322                                    int wait_for_unblock)
1323 {
1324         struct btrfs_async_commit *ac;
1325         struct btrfs_transaction *cur_trans;
1326
1327         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1328         if (!ac)
1329                 return -ENOMEM;
1330
1331         INIT_DELAYED_WORK(&ac->work, do_async_commit);
1332         ac->root = root;
1333         ac->newtrans = btrfs_join_transaction(root);
1334         if (IS_ERR(ac->newtrans)) {
1335                 int err = PTR_ERR(ac->newtrans);
1336                 kfree(ac);
1337                 return err;
1338         }
1339
1340         /* take transaction reference */
1341         cur_trans = trans->transaction;
1342         atomic_inc(&cur_trans->use_count);
1343
1344         btrfs_end_transaction(trans, root);
1345
1346         /*
1347          * Tell lockdep we've released the freeze rwsem, since the
1348          * async commit thread will be the one to unlock it.
1349          */
1350         rwsem_release(&root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
1351                       1, _THIS_IP_);
1352
1353         schedule_delayed_work(&ac->work, 0);
1354
1355         /* wait for transaction to start and unblock */
1356         if (wait_for_unblock)
1357                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1358         else
1359                 wait_current_trans_commit_start(root, cur_trans);
1360
1361         if (current->journal_info == trans)
1362                 current->journal_info = NULL;
1363
1364         put_transaction(cur_trans);
1365         return 0;
1366 }
1367
1368
1369 static void cleanup_transaction(struct btrfs_trans_handle *trans,
1370                                 struct btrfs_root *root, int err)
1371 {
1372         struct btrfs_transaction *cur_trans = trans->transaction;
1373
1374         WARN_ON(trans->use_count > 1);
1375
1376         btrfs_abort_transaction(trans, root, err);
1377
1378         spin_lock(&root->fs_info->trans_lock);
1379         list_del_init(&cur_trans->list);
1380         if (cur_trans == root->fs_info->running_transaction) {
1381                 root->fs_info->running_transaction = NULL;
1382                 root->fs_info->trans_no_join = 0;
1383         }
1384         spin_unlock(&root->fs_info->trans_lock);
1385
1386         btrfs_cleanup_one_transaction(trans->transaction, root);
1387
1388         put_transaction(cur_trans);
1389         put_transaction(cur_trans);
1390
1391         trace_btrfs_transaction_commit(root);
1392
1393         btrfs_scrub_continue(root);
1394
1395         if (current->journal_info == trans)
1396                 current->journal_info = NULL;
1397
1398         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1399 }
1400
1401 static int btrfs_flush_all_pending_stuffs(struct btrfs_trans_handle *trans,
1402                                           struct btrfs_root *root)
1403 {
1404         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1405         int snap_pending = 0;
1406         int ret;
1407
1408         if (!flush_on_commit) {
1409                 spin_lock(&root->fs_info->trans_lock);
1410                 if (!list_empty(&trans->transaction->pending_snapshots))
1411                         snap_pending = 1;
1412                 spin_unlock(&root->fs_info->trans_lock);
1413         }
1414
1415         if (flush_on_commit || snap_pending) {
1416                 btrfs_start_delalloc_inodes(root, 1);
1417                 btrfs_wait_ordered_extents(root, 1);
1418         }
1419
1420         ret = btrfs_run_delayed_items(trans, root);
1421         if (ret)
1422                 return ret;
1423
1424         /*
1425          * running the delayed items may have added new refs. account
1426          * them now so that they hinder processing of more delayed refs
1427          * as little as possible.
1428          */
1429         btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
1430
1431         /*
1432          * rename don't use btrfs_join_transaction, so, once we
1433          * set the transaction to blocked above, we aren't going
1434          * to get any new ordered operations.  We can safely run
1435          * it here and no for sure that nothing new will be added
1436          * to the list
1437          */
1438         btrfs_run_ordered_operations(root, 1);
1439
1440         return 0;
1441 }
1442
1443 /*
1444  * btrfs_transaction state sequence:
1445  *    in_commit = 0, blocked = 0  (initial)
1446  *    in_commit = 1, blocked = 1
1447  *    blocked = 0
1448  *    commit_done = 1
1449  */
1450 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1451                              struct btrfs_root *root)
1452 {
1453         unsigned long joined = 0;
1454         struct btrfs_transaction *cur_trans = trans->transaction;
1455         struct btrfs_transaction *prev_trans = NULL;
1456         DEFINE_WAIT(wait);
1457         int ret;
1458         int should_grow = 0;
1459         unsigned long now = get_seconds();
1460
1461         ret = btrfs_run_ordered_operations(root, 0);
1462         if (ret) {
1463                 btrfs_abort_transaction(trans, root, ret);
1464                 goto cleanup_transaction;
1465         }
1466
1467         if (cur_trans->aborted) {
1468                 ret = cur_trans->aborted;
1469                 goto cleanup_transaction;
1470         }
1471
1472         /* make a pass through all the delayed refs we have so far
1473          * any runnings procs may add more while we are here
1474          */
1475         ret = btrfs_run_delayed_refs(trans, root, 0);
1476         if (ret)
1477                 goto cleanup_transaction;
1478
1479         btrfs_trans_release_metadata(trans, root);
1480         trans->block_rsv = NULL;
1481
1482         cur_trans = trans->transaction;
1483
1484         /*
1485          * set the flushing flag so procs in this transaction have to
1486          * start sending their work down.
1487          */
1488         cur_trans->delayed_refs.flushing = 1;
1489
1490         if (!list_empty(&trans->new_bgs))
1491                 btrfs_create_pending_block_groups(trans, root);
1492
1493         ret = btrfs_run_delayed_refs(trans, root, 0);
1494         if (ret)
1495                 goto cleanup_transaction;
1496
1497         spin_lock(&cur_trans->commit_lock);
1498         if (cur_trans->in_commit) {
1499                 spin_unlock(&cur_trans->commit_lock);
1500                 atomic_inc(&cur_trans->use_count);
1501                 ret = btrfs_end_transaction(trans, root);
1502
1503                 wait_for_commit(root, cur_trans);
1504
1505                 put_transaction(cur_trans);
1506
1507                 return ret;
1508         }
1509
1510         trans->transaction->in_commit = 1;
1511         trans->transaction->blocked = 1;
1512         spin_unlock(&cur_trans->commit_lock);
1513         wake_up(&root->fs_info->transaction_blocked_wait);
1514
1515         spin_lock(&root->fs_info->trans_lock);
1516         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1517                 prev_trans = list_entry(cur_trans->list.prev,
1518                                         struct btrfs_transaction, list);
1519                 if (!prev_trans->commit_done) {
1520                         atomic_inc(&prev_trans->use_count);
1521                         spin_unlock(&root->fs_info->trans_lock);
1522
1523                         wait_for_commit(root, prev_trans);
1524
1525                         put_transaction(prev_trans);
1526                 } else {
1527                         spin_unlock(&root->fs_info->trans_lock);
1528                 }
1529         } else {
1530                 spin_unlock(&root->fs_info->trans_lock);
1531         }
1532
1533         if (!btrfs_test_opt(root, SSD) &&
1534             (now < cur_trans->start_time || now - cur_trans->start_time < 1))
1535                 should_grow = 1;
1536
1537         do {
1538                 joined = cur_trans->num_joined;
1539
1540                 WARN_ON(cur_trans != trans->transaction);
1541
1542                 ret = btrfs_flush_all_pending_stuffs(trans, root);
1543                 if (ret)
1544                         goto cleanup_transaction;
1545
1546                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1547                                 TASK_UNINTERRUPTIBLE);
1548
1549                 if (atomic_read(&cur_trans->num_writers) > 1)
1550                         schedule_timeout(MAX_SCHEDULE_TIMEOUT);
1551                 else if (should_grow)
1552                         schedule_timeout(1);
1553
1554                 finish_wait(&cur_trans->writer_wait, &wait);
1555         } while (atomic_read(&cur_trans->num_writers) > 1 ||
1556                  (should_grow && cur_trans->num_joined != joined));
1557
1558         ret = btrfs_flush_all_pending_stuffs(trans, root);
1559         if (ret)
1560                 goto cleanup_transaction;
1561
1562         /*
1563          * Ok now we need to make sure to block out any other joins while we
1564          * commit the transaction.  We could have started a join before setting
1565          * no_join so make sure to wait for num_writers to == 1 again.
1566          */
1567         spin_lock(&root->fs_info->trans_lock);
1568         root->fs_info->trans_no_join = 1;
1569         spin_unlock(&root->fs_info->trans_lock);
1570         wait_event(cur_trans->writer_wait,
1571                    atomic_read(&cur_trans->num_writers) == 1);
1572
1573         /*
1574          * the reloc mutex makes sure that we stop
1575          * the balancing code from coming in and moving
1576          * extents around in the middle of the commit
1577          */
1578         mutex_lock(&root->fs_info->reloc_mutex);
1579
1580         /*
1581          * We needn't worry about the delayed items because we will
1582          * deal with them in create_pending_snapshot(), which is the
1583          * core function of the snapshot creation.
1584          */
1585         ret = create_pending_snapshots(trans, root->fs_info);
1586         if (ret) {
1587                 mutex_unlock(&root->fs_info->reloc_mutex);
1588                 goto cleanup_transaction;
1589         }
1590
1591         /*
1592          * We insert the dir indexes of the snapshots and update the inode
1593          * of the snapshots' parents after the snapshot creation, so there
1594          * are some delayed items which are not dealt with. Now deal with
1595          * them.
1596          *
1597          * We needn't worry that this operation will corrupt the snapshots,
1598          * because all the tree which are snapshoted will be forced to COW
1599          * the nodes and leaves.
1600          */
1601         ret = btrfs_run_delayed_items(trans, root);
1602         if (ret) {
1603                 mutex_unlock(&root->fs_info->reloc_mutex);
1604                 goto cleanup_transaction;
1605         }
1606
1607         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1608         if (ret) {
1609                 mutex_unlock(&root->fs_info->reloc_mutex);
1610                 goto cleanup_transaction;
1611         }
1612
1613         /*
1614          * make sure none of the code above managed to slip in a
1615          * delayed item
1616          */
1617         btrfs_assert_delayed_root_empty(root);
1618
1619         WARN_ON(cur_trans != trans->transaction);
1620
1621         btrfs_scrub_pause(root);
1622         /* btrfs_commit_tree_roots is responsible for getting the
1623          * various roots consistent with each other.  Every pointer
1624          * in the tree of tree roots has to point to the most up to date
1625          * root for every subvolume and other tree.  So, we have to keep
1626          * the tree logging code from jumping in and changing any
1627          * of the trees.
1628          *
1629          * At this point in the commit, there can't be any tree-log
1630          * writers, but a little lower down we drop the trans mutex
1631          * and let new people in.  By holding the tree_log_mutex
1632          * from now until after the super is written, we avoid races
1633          * with the tree-log code.
1634          */
1635         mutex_lock(&root->fs_info->tree_log_mutex);
1636
1637         ret = commit_fs_roots(trans, root);
1638         if (ret) {
1639                 mutex_unlock(&root->fs_info->tree_log_mutex);
1640                 mutex_unlock(&root->fs_info->reloc_mutex);
1641                 goto cleanup_transaction;
1642         }
1643
1644         /* commit_fs_roots gets rid of all the tree log roots, it is now
1645          * safe to free the root of tree log roots
1646          */
1647         btrfs_free_log_root_tree(trans, root->fs_info);
1648
1649         ret = commit_cowonly_roots(trans, root);
1650         if (ret) {
1651                 mutex_unlock(&root->fs_info->tree_log_mutex);
1652                 mutex_unlock(&root->fs_info->reloc_mutex);
1653                 goto cleanup_transaction;
1654         }
1655
1656         btrfs_prepare_extent_commit(trans, root);
1657
1658         cur_trans = root->fs_info->running_transaction;
1659
1660         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1661                             root->fs_info->tree_root->node);
1662         switch_commit_root(root->fs_info->tree_root);
1663
1664         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1665                             root->fs_info->chunk_root->node);
1666         switch_commit_root(root->fs_info->chunk_root);
1667
1668         assert_qgroups_uptodate(trans);
1669         update_super_roots(root);
1670
1671         if (!root->fs_info->log_root_recovering) {
1672                 btrfs_set_super_log_root(root->fs_info->super_copy, 0);
1673                 btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
1674         }
1675
1676         memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
1677                sizeof(*root->fs_info->super_copy));
1678
1679         trans->transaction->blocked = 0;
1680         spin_lock(&root->fs_info->trans_lock);
1681         root->fs_info->running_transaction = NULL;
1682         root->fs_info->trans_no_join = 0;
1683         spin_unlock(&root->fs_info->trans_lock);
1684         mutex_unlock(&root->fs_info->reloc_mutex);
1685
1686         wake_up(&root->fs_info->transaction_wait);
1687
1688         ret = btrfs_write_and_wait_transaction(trans, root);
1689         if (ret) {
1690                 btrfs_error(root->fs_info, ret,
1691                             "Error while writing out transaction.");
1692                 mutex_unlock(&root->fs_info->tree_log_mutex);
1693                 goto cleanup_transaction;
1694         }
1695
1696         ret = write_ctree_super(trans, root, 0);
1697         if (ret) {
1698                 mutex_unlock(&root->fs_info->tree_log_mutex);
1699                 goto cleanup_transaction;
1700         }
1701
1702         /*
1703          * the super is written, we can safely allow the tree-loggers
1704          * to go about their business
1705          */
1706         mutex_unlock(&root->fs_info->tree_log_mutex);
1707
1708         btrfs_finish_extent_commit(trans, root);
1709
1710         cur_trans->commit_done = 1;
1711
1712         root->fs_info->last_trans_committed = cur_trans->transid;
1713
1714         wake_up(&cur_trans->commit_wait);
1715
1716         spin_lock(&root->fs_info->trans_lock);
1717         list_del_init(&cur_trans->list);
1718         spin_unlock(&root->fs_info->trans_lock);
1719
1720         put_transaction(cur_trans);
1721         put_transaction(cur_trans);
1722
1723         if (trans->type < TRANS_JOIN_NOLOCK)
1724                 sb_end_intwrite(root->fs_info->sb);
1725
1726         trace_btrfs_transaction_commit(root);
1727
1728         btrfs_scrub_continue(root);
1729
1730         if (current->journal_info == trans)
1731                 current->journal_info = NULL;
1732
1733         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1734
1735         if (current != root->fs_info->transaction_kthread)
1736                 btrfs_run_delayed_iputs(root);
1737
1738         return ret;
1739
1740 cleanup_transaction:
1741         btrfs_trans_release_metadata(trans, root);
1742         trans->block_rsv = NULL;
1743         btrfs_printk(root->fs_info, "Skipping commit of aborted transaction.\n");
1744 //      WARN_ON(1);
1745         if (current->journal_info == trans)
1746                 current->journal_info = NULL;
1747         cleanup_transaction(trans, root, ret);
1748
1749         return ret;
1750 }
1751
1752 /*
1753  * interface function to delete all the snapshots we have scheduled for deletion
1754  */
1755 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1756 {
1757         LIST_HEAD(list);
1758         struct btrfs_fs_info *fs_info = root->fs_info;
1759
1760         spin_lock(&fs_info->trans_lock);
1761         list_splice_init(&fs_info->dead_roots, &list);
1762         spin_unlock(&fs_info->trans_lock);
1763
1764         while (!list_empty(&list)) {
1765                 int ret;
1766
1767                 root = list_entry(list.next, struct btrfs_root, root_list);
1768                 list_del(&root->root_list);
1769
1770                 btrfs_kill_all_delayed_nodes(root);
1771
1772                 if (btrfs_header_backref_rev(root->node) <
1773                     BTRFS_MIXED_BACKREF_REV)
1774                         ret = btrfs_drop_snapshot(root, NULL, 0, 0);
1775                 else
1776                         ret =btrfs_drop_snapshot(root, NULL, 1, 0);
1777                 BUG_ON(ret < 0);
1778         }
1779         return 0;
1780 }