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