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