fs/btrfs/transaction.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/fs.h>
  20 #include <linux/slab.h>
  21 #include <linux/sched.h>
  22 #include <linux/writeback.h>
  23 #include <linux/pagemap.h>
  24 #include <linux/blkdev.h>
  25 #include <linux/uuid.h>
  26 #include "ctree.h"
  27 #include "disk-io.h"
  28 #include "transaction.h"
  29 #include "locking.h"
  30 #include "tree-log.h"
  31 #include "inode-map.h"
  32 #include "volumes.h"
  33 #include "dev-replace.h"
  34
  35 #define BTRFS_ROOT_TRANS_TAG 0
  36
  37 void put_transaction(struct btrfs_transaction *transaction)
  38 {
  39         WARN_ON(atomic_read(&transaction->use_count) == 0);
  40         if (atomic_dec_and_test(&transaction->use_count)) {
  41                 BUG_ON(!list_empty(&transaction->list));
  42                 WARN_ON(transaction->delayed_refs.root.rb_node);
  43                 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 static inline int can_join_transaction(struct btrfs_transaction *trans,
  54                                        int type)
  55 {
  56         return !(trans->in_commit &&
  57                  type != TRANS_JOIN &&
  58                  type != TRANS_JOIN_NOLOCK);
  59 }
  60
  61 /*
  62  * either allocate a new transaction or hop into the existing one
  63  */
  64 static noinline int join_transaction(struct btrfs_root *root, int type)
  65 {
  66         struct btrfs_transaction *cur_trans;
  67         struct btrfs_fs_info *fs_info = root->fs_info;
  68
  69         spin_lock(&fs_info->trans_lock);
  70 loop:
  71         /* The file system has been taken offline. No new transactions. */
  72         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
  73                 spin_unlock(&fs_info->trans_lock);
  74                 return -EROFS;
  75         }
  76
  77         if (fs_info->trans_no_join) {
  78                 /*
  79                  * If we are JOIN_NOLOCK we're already committing a current
  80                  * transaction, we just need a handle to deal with something
  81                  * when committing the transaction, such as inode cache and
  82                  * space cache. It is a special case.
  83                  */
  84                 if (type != TRANS_JOIN_NOLOCK) {
  85                         spin_unlock(&fs_info->trans_lock);
  86                         return -EBUSY;
  87                 }
  88         }
  89
  90         cur_trans = fs_info->running_transaction;
  91         if (cur_trans) {
  92                 if (cur_trans->aborted) {
  93                         spin_unlock(&fs_info->trans_lock);
  94                         return cur_trans->aborted;
  95                 }
  96                 if (!can_join_transaction(cur_trans, type)) {
  97                         spin_unlock(&fs_info->trans_lock);
  98                         return -EBUSY;
  99                 }
 100                 atomic_inc(&cur_trans->use_count);
 101                 atomic_inc(&cur_trans->num_writers);
 102                 cur_trans->num_joined++;
 103                 spin_unlock(&fs_info->trans_lock);
 104                 return 0;
 105         }
 106         spin_unlock(&fs_info->trans_lock);
 107
 108         /*
 109          * If we are ATTACH, we just want to catch the current transaction,
 110          * and commit it. If there is no transaction, just return ENOENT.
 111          */
 112         if (type == TRANS_ATTACH)
 113                 return -ENOENT;
 114
 115         cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
 116         if (!cur_trans)
 117                 return -ENOMEM;
 118
 119         spin_lock(&fs_info->trans_lock);
 120         if (fs_info->running_transaction) {
 121                 /*
 122                  * someone started a transaction after we unlocked.  Make sure
 123                  * to redo the trans_no_join checks above
 124                  */
 125                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 126                 cur_trans = fs_info->running_transaction;
 127                 goto loop;
 128         } else if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
 129                 spin_unlock(&fs_info->trans_lock);
 130                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 131                 return -EROFS;
 132         }
 133
 134         atomic_set(&cur_trans->num_writers, 1);
 135         cur_trans->num_joined = 0;
 136         init_waitqueue_head(&cur_trans->writer_wait);
 137         init_waitqueue_head(&cur_trans->commit_wait);
 138         cur_trans->in_commit = 0;
 139         cur_trans->blocked = 0;
 140         /*
 141          * One for this trans handle, one so it will live on until we
 142          * commit the transaction.
 143          */
 144         atomic_set(&cur_trans->use_count, 2);
 145         cur_trans->commit_done = 0;
 146         cur_trans->start_time = get_seconds();
 147
 148         cur_trans->delayed_refs.root = RB_ROOT;
 149         cur_trans->delayed_refs.num_entries = 0;
 150         cur_trans->delayed_refs.num_heads_ready = 0;
 151         cur_trans->delayed_refs.num_heads = 0;
 152         cur_trans->delayed_refs.flushing = 0;
 153         cur_trans->delayed_refs.run_delayed_start = 0;
 154
 155         /*
 156          * although the tree mod log is per file system and not per transaction,
 157          * the log must never go across transaction boundaries.
 158          */
 159         smp_mb();
 160         if (!list_empty(&fs_info->tree_mod_seq_list))
 161                 WARN(1, KERN_ERR "btrfs: tree_mod_seq_list not empty when "
 162                         "creating a fresh transaction\n");
 163         if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
 164                 WARN(1, KERN_ERR "btrfs: tree_mod_log rb tree not empty when "
 165                         "creating a fresh transaction\n");
 166         atomic_set(&fs_info->tree_mod_seq, 0);
 167
 168         spin_lock_init(&cur_trans->commit_lock);
 169         spin_lock_init(&cur_trans->delayed_refs.lock);
 170
 171         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 172         INIT_LIST_HEAD(&cur_trans->ordered_operations);
 173         list_add_tail(&cur_trans->list, &fs_info->trans_list);
 174         extent_io_tree_init(&cur_trans->dirty_pages,
 175                              fs_info->btree_inode->i_mapping);
 176         fs_info->generation++;
 177         cur_trans->transid = fs_info->generation;
 178         fs_info->running_transaction = cur_trans;
 179         cur_trans->aborted = 0;
 180         spin_unlock(&fs_info->trans_lock);
 181
 182         return 0;
 183 }
 184
 185 /*
 186  * this does all the record keeping required to make sure that a reference
 187  * counted root is properly recorded in a given transaction.  This is required
 188  * to make sure the old root from before we joined the transaction is deleted
 189  * when the transaction commits
 190  */
 191 static int record_root_in_trans(struct btrfs_trans_handle *trans,
 192                                struct btrfs_root *root)
 193 {
 194         if (root->ref_cows && root->last_trans < trans->transid) {
 195                 WARN_ON(root == root->fs_info->extent_root);
 196                 WARN_ON(root->commit_root != root->node);
 197
 198                 /*
 199                  * see below for in_trans_setup usage rules
 200                  * we have the reloc mutex held now, so there
 201                  * is only one writer in this function
 202                  */
 203                 root->in_trans_setup = 1;
 204
 205                 /* make sure readers find in_trans_setup before
 206                  * they find our root->last_trans update
 207                  */
 208                 smp_wmb();
 209
 210                 spin_lock(&root->fs_info->fs_roots_radix_lock);
 211                 if (root->last_trans == trans->transid) {
 212                         spin_unlock(&root->fs_info->fs_roots_radix_lock);
 213                         return 0;
 214                 }
 215                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 216                            (unsigned long)root->root_key.objectid,
 217                            BTRFS_ROOT_TRANS_TAG);
 218                 spin_unlock(&root->fs_info->fs_roots_radix_lock);
 219                 root->last_trans = trans->transid;
 220
 221                 /* this is pretty tricky.  We don't want to
 222                  * take the relocation lock in btrfs_record_root_in_trans
 223                  * unless we're really doing the first setup for this root in
 224                  * this transaction.
 225                  *
 226                  * Normally we'd use root->last_trans as a flag to decide
 227                  * if we want to take the expensive mutex.
 228                  *
 229                  * But, we have to set root->last_trans before we
 230                  * init the relocation root, otherwise, we trip over warnings
 231                  * in ctree.c.  The solution used here is to flag ourselves
 232                  * with root->in_trans_setup.  When this is 1, we're still
 233                  * fixing up the reloc trees and everyone must wait.
 234                  *
 235                  * When this is zero, they can trust root->last_trans and fly
 236                  * through btrfs_record_root_in_trans without having to take the
 237                  * lock.  smp_wmb() makes sure that all the writes above are
 238                  * done before we pop in the zero below
 239                  */
 240                 btrfs_init_reloc_root(trans, root);
 241                 smp_wmb();
 242                 root->in_trans_setup = 0;
 243         }
 244         return 0;
 245 }
 246
 247
 248 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 249                                struct btrfs_root *root)
 250 {
 251         if (!root->ref_cows)
 252                 return 0;
 253
 254         /*
 255          * see record_root_in_trans for comments about in_trans_setup usage
 256          * and barriers
 257          */
 258         smp_rmb();
 259         if (root->last_trans == trans->transid &&
 260             !root->in_trans_setup)
 261                 return 0;
 262
 263         mutex_lock(&root->fs_info->reloc_mutex);
 264         record_root_in_trans(trans, root);
 265         mutex_unlock(&root->fs_info->reloc_mutex);
 266
 267         return 0;
 268 }
 269
 270 /* wait for commit against the current transaction to become unblocked
 271  * when this is done, it is safe to start a new transaction, but the current
 272  * transaction might not be fully on disk.
 273  */
 274 static void wait_current_trans(struct btrfs_root *root)
 275 {
 276         struct btrfs_transaction *cur_trans;
 277
 278         spin_lock(&root->fs_info->trans_lock);
 279         cur_trans = root->fs_info->running_transaction;
 280         if (cur_trans && cur_trans->blocked) {
 281                 atomic_inc(&cur_trans->use_count);
 282                 spin_unlock(&root->fs_info->trans_lock);
 283
 284                 wait_event(root->fs_info->transaction_wait,
 285                            !cur_trans->blocked);
 286                 put_transaction(cur_trans);
 287         } else {
 288                 spin_unlock(&root->fs_info->trans_lock);
 289         }
 290 }
 291
 292 static int may_wait_transaction(struct btrfs_root *root, int type)
 293 {
 294         if (root->fs_info->log_root_recovering)
 295                 return 0;
 296
 297         if (type == TRANS_USERSPACE)
 298                 return 1;
 299
 300         if (type == TRANS_START &&
 301             !atomic_read(&root->fs_info->open_ioctl_trans))
 302                 return 1;
 303
 304         return 0;
 305 }
 306
 307 static struct btrfs_trans_handle *
 308 start_transaction(struct btrfs_root *root, u64 num_items, int type,
 309                   enum btrfs_reserve_flush_enum flush)
 310 {
 311         struct btrfs_trans_handle *h;
 312         struct btrfs_transaction *cur_trans;
 313         u64 num_bytes = 0;
 314         int ret;
 315         u64 qgroup_reserved = 0;
 316
 317         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
 318                 return ERR_PTR(-EROFS);
 319
 320         if (current->journal_info) {
 321                 WARN_ON(type != TRANS_JOIN && type != TRANS_JOIN_NOLOCK);
 322                 h = current->journal_info;
 323                 h->use_count++;
 324                 WARN_ON(h->use_count > 2);
 325                 h->orig_rsv = h->block_rsv;
 326                 h->block_rsv = NULL;
 327                 goto got_it;
 328         }
 329
 330         /*
 331          * Do the reservation before we join the transaction so we can do all
 332          * the appropriate flushing if need be.
 333          */
 334         if (num_items > 0 && root != root->fs_info->chunk_root) {
 335                 if (root->fs_info->quota_enabled &&
 336                     is_fstree(root->root_key.objectid)) {
 337                         qgroup_reserved = num_items * root->leafsize;
 338                         ret = btrfs_qgroup_reserve(root, qgroup_reserved);
 339                         if (ret)
 340                                 return ERR_PTR(ret);
 341                 }
 342
 343                 num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
 344                 ret = btrfs_block_rsv_add(root,
 345                                           &root->fs_info->trans_block_rsv,
 346                                           num_bytes, flush);
 347                 if (ret)
 348                         goto reserve_fail;
 349         }
 350 again:
 351         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 352         if (!h) {
 353                 ret = -ENOMEM;
 354                 goto alloc_fail;
 355         }
 356
 357         /*
 358          * If we are JOIN_NOLOCK we're already committing a transaction and
 359          * waiting on this guy, so we don't need to do the sb_start_intwrite
 360          * because we're already holding a ref.  We need this because we could
 361          * have raced in and did an fsync() on a file which can kick a commit
 362          * and then we deadlock with somebody doing a freeze.
 363          *
 364          * If we are ATTACH, it means we just want to catch the current
 365          * transaction and commit it, so we needn't do sb_start_intwrite().
 366          */
 367         if (type < TRANS_JOIN_NOLOCK)
 368                 sb_start_intwrite(root->fs_info->sb);
 369
 370         if (may_wait_transaction(root, type))
 371                 wait_current_trans(root);
 372
 373         do {
 374                 ret = join_transaction(root, type);
 375                 if (ret == -EBUSY) {
 376                         wait_current_trans(root);
 377                         if (unlikely(type == TRANS_ATTACH))
 378                                 ret = -ENOENT;
 379                 }
 380         } while (ret == -EBUSY);
 381
 382         if (ret < 0) {
 383                 /* We must get the transaction if we are JOIN_NOLOCK. */
 384                 BUG_ON(type == TRANS_JOIN_NOLOCK);
 385                 goto join_fail;
 386         }
 387
 388         cur_trans = root->fs_info->running_transaction;
 389
 390         h->transid = cur_trans->transid;
 391         h->transaction = cur_trans;
 392         h->blocks_used = 0;
 393         h->bytes_reserved = 0;
 394         h->root = root;
 395         h->delayed_ref_updates = 0;
 396         h->use_count = 1;
 397         h->adding_csums = 0;
 398         h->block_rsv = NULL;
 399         h->orig_rsv = NULL;
 400         h->aborted = 0;
 401         h->qgroup_reserved = 0;
 402         h->delayed_ref_elem.seq = 0;
 403         h->type = type;
 404         h->allocating_chunk = false;
 405         INIT_LIST_HEAD(&h->qgroup_ref_list);
 406         INIT_LIST_HEAD(&h->new_bgs);
 407
 408         smp_mb();
 409         if (cur_trans->blocked && may_wait_transaction(root, type)) {
 410                 btrfs_commit_transaction(h, root);
 411                 goto again;
 412         }
 413
 414         if (num_bytes) {
 415                 trace_btrfs_space_reservation(root->fs_info, "transaction",
 416                                               h->transid, num_bytes, 1);
 417                 h->block_rsv = &root->fs_info->trans_block_rsv;
 418                 h->bytes_reserved = num_bytes;
 419         }
 420         h->qgroup_reserved = qgroup_reserved;
 421
 422 got_it:
 423         btrfs_record_root_in_trans(h, root);
 424
 425         if (!current->journal_info && type != TRANS_USERSPACE)
 426                 current->journal_info = h;
 427         return h;
 428
 429 join_fail:
 430         if (type < TRANS_JOIN_NOLOCK)
 431                 sb_end_intwrite(root->fs_info->sb);
 432         kmem_cache_free(btrfs_trans_handle_cachep, h);
 433 alloc_fail:
 434         if (num_bytes)
 435                 btrfs_block_rsv_release(root, &root->fs_info->trans_block_rsv,
 436                                         num_bytes);
 437 reserve_fail:
 438         if (qgroup_reserved)
 439                 btrfs_qgroup_free(root, qgroup_reserved);
 440         return ERR_PTR(ret);
 441 }
 442
 443 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 444                                                    int num_items)
 445 {
 446         return start_transaction(root, num_items, TRANS_START,
 447                                  BTRFS_RESERVE_FLUSH_ALL);
 448 }
 449
 450 struct btrfs_trans_handle *btrfs_start_transaction_lflush(
 451                                         struct btrfs_root *root, int num_items)
 452 {
 453         return start_transaction(root, num_items, TRANS_START,
 454                                  BTRFS_RESERVE_FLUSH_LIMIT);
 455 }
 456
 457 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 458 {
 459         return start_transaction(root, 0, TRANS_JOIN, 0);
 460 }
 461
 462 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 463 {
 464         return start_transaction(root, 0, TRANS_JOIN_NOLOCK, 0);
 465 }
 466
 467 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 468 {
 469         return start_transaction(root, 0, TRANS_USERSPACE, 0);
 470 }
 471
 472 /*
 473  * btrfs_attach_transaction() - catch the running transaction
 474  *
 475  * It is used when we want to commit the current the transaction, but
 476  * don't want to start a new one.
 477  *
 478  * Note: If this function return -ENOENT, it just means there is no
 479  * running transaction. But it is possible that the inactive transaction
 480  * is still in the memory, not fully on disk. If you hope there is no
 481  * inactive transaction in the fs when -ENOENT is returned, you should
 482  * invoke
 483  *     btrfs_attach_transaction_barrier()
 484  */
 485 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
 486 {
 487         return start_transaction(root, 0, TRANS_ATTACH, 0);
 488 }
 489
 490 /*
 491  * btrfs_attach_transaction() - catch the running transaction
 492  *
 493  * It is similar to the above function, the differentia is this one
 494  * will wait for all the inactive transactions until they fully
 495  * complete.
 496  */
 497 struct btrfs_trans_handle *
 498 btrfs_attach_transaction_barrier(struct btrfs_root *root)
 499 {
 500         struct btrfs_trans_handle *trans;
 501
 502         trans = start_transaction(root, 0, TRANS_ATTACH, 0);
 503         if (IS_ERR(trans) && PTR_ERR(trans) == -ENOENT)
 504                 btrfs_wait_for_commit(root, 0);
 505
 506         return trans;
 507 }
 508
 509 /* wait for a transaction commit to be fully complete */
 510 static noinline void wait_for_commit(struct btrfs_root *root,
 511                                     struct btrfs_transaction *commit)
 512 {
 513         wait_event(commit->commit_wait, commit->commit_done);
 514 }
 515
 516 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 517 {
 518         struct btrfs_transaction *cur_trans = NULL, *t;
 519         int ret = 0;
 520
 521         if (transid) {
 522                 if (transid <= root->fs_info->last_trans_committed)
 523                         goto out;
 524
 525                 ret = -EINVAL;
 526                 /* find specified transaction */
 527                 spin_lock(&root->fs_info->trans_lock);
 528                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 529                         if (t->transid == transid) {
 530                                 cur_trans = t;
 531                                 atomic_inc(&cur_trans->use_count);
 532                                 ret = 0;
 533                                 break;
 534                         }
 535                         if (t->transid > transid) {
 536                                 ret = 0;
 537                                 break;
 538                         }
 539                 }
 540                 spin_unlock(&root->fs_info->trans_lock);
 541                 /* The specified transaction doesn't exist */
 542                 if (!cur_trans)
 543                         goto out;
 544         } else {
 545                 /* find newest transaction that is committing | committed */
 546                 spin_lock(&root->fs_info->trans_lock);
 547                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 548                                             list) {
 549                         if (t->in_commit) {
 550                                 if (t->commit_done)
 551                                         break;
 552                                 cur_trans = t;
 553                                 atomic_inc(&cur_trans->use_count);
 554                                 break;
 555                         }
 556                 }
 557                 spin_unlock(&root->fs_info->trans_lock);
 558                 if (!cur_trans)
 559                         goto out;  /* nothing committing|committed */
 560         }
 561
 562         wait_for_commit(root, cur_trans);
 563         put_transaction(cur_trans);
 564 out:
 565         return ret;
 566 }
 567
 568 void btrfs_throttle(struct btrfs_root *root)
 569 {
 570         if (!atomic_read(&root->fs_info->open_ioctl_trans))
 571                 wait_current_trans(root);
 572 }
 573
 574 static int should_end_transaction(struct btrfs_trans_handle *trans,
 575                                   struct btrfs_root *root)
 576 {
 577         int ret;
 578
 579         ret = btrfs_block_rsv_check(root, &root->fs_info->global_block_rsv, 5);
 580         return ret ? 1 : 0;
 581 }
 582
 583 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 584                                  struct btrfs_root *root)
 585 {
 586         struct btrfs_transaction *cur_trans = trans->transaction;
 587         int updates;
 588         int err;
 589
 590         smp_mb();
 591         if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
 592                 return 1;
 593
 594         updates = trans->delayed_ref_updates;
 595         trans->delayed_ref_updates = 0;
 596         if (updates) {
 597                 err = btrfs_run_delayed_refs(trans, root, updates);
 598                 if (err) /* Error code will also eval true */
 599                         return err;
 600         }
 601
 602         return should_end_transaction(trans, root);
 603 }
 604
 605 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 606                           struct btrfs_root *root, int throttle)
 607 {
 608         struct btrfs_transaction *cur_trans = trans->transaction;
 609         struct btrfs_fs_info *info = root->fs_info;
 610         int count = 0;
 611         int lock = (trans->type != TRANS_JOIN_NOLOCK);
 612         int err = 0;
 613
 614         if (--trans->use_count) {
 615                 trans->block_rsv = trans->orig_rsv;
 616                 return 0;
 617         }
 618
 619         /*
 620          * do the qgroup accounting as early as possible
 621          */
 622         err = btrfs_delayed_refs_qgroup_accounting(trans, info);
 623
 624         btrfs_trans_release_metadata(trans, root);
 625         trans->block_rsv = NULL;
 626         /*
 627          * the same root has to be passed to start_transaction and
 628          * end_transaction. Subvolume quota depends on this.
 629          */
 630         WARN_ON(trans->root != root);
 631
 632         if (trans->qgroup_reserved) {
 633                 btrfs_qgroup_free(root, trans->qgroup_reserved);
 634                 trans->qgroup_reserved = 0;
 635         }
 636
 637         if (!list_empty(&trans->new_bgs))
 638                 btrfs_create_pending_block_groups(trans, root);
 639
 640         while (count < 2) {
 641                 unsigned long cur = trans->delayed_ref_updates;
 642                 trans->delayed_ref_updates = 0;
 643                 if (cur &&
 644                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 645                         trans->delayed_ref_updates = 0;
 646                         btrfs_run_delayed_refs(trans, root, cur);
 647                 } else {
 648                         break;
 649                 }
 650                 count++;
 651         }
 652         btrfs_trans_release_metadata(trans, root);
 653         trans->block_rsv = NULL;
 654
 655         if (!list_empty(&trans->new_bgs))
 656                 btrfs_create_pending_block_groups(trans, root);
 657
 658         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
 659             should_end_transaction(trans, root)) {
 660                 trans->transaction->blocked = 1;
 661                 smp_wmb();
 662         }
 663
 664         if (lock && cur_trans->blocked && !cur_trans->in_commit) {
 665                 if (throttle) {
 666                         /*
 667                          * We may race with somebody else here so end up having
 668                          * to call end_transaction on ourselves again, so inc
 669                          * our use_count.
 670                          */
 671                         trans->use_count++;
 672                         return btrfs_commit_transaction(trans, root);
 673                 } else {
 674                         wake_up_process(info->transaction_kthread);
 675                 }
 676         }
 677
 678         if (trans->type < TRANS_JOIN_NOLOCK)
 679                 sb_end_intwrite(root->fs_info->sb);
 680
 681         WARN_ON(cur_trans != info->running_transaction);
 682         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 683         atomic_dec(&cur_trans->num_writers);
 684
 685         smp_mb();
 686         if (waitqueue_active(&cur_trans->writer_wait))
 687                 wake_up(&cur_trans->writer_wait);
 688         put_transaction(cur_trans);
 689
 690         if (current->journal_info == trans)
 691                 current->journal_info = NULL;
 692
 693         if (throttle)
 694                 btrfs_run_delayed_iputs(root);
 695
 696         if (trans->aborted ||
 697             test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
 698                 err = -EIO;
 699         assert_qgroups_uptodate(trans);
 700
 701         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 702         return err;
 703 }
 704
 705 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 706                           struct btrfs_root *root)
 707 {
 708         int ret;
 709
 710         ret = __btrfs_end_transaction(trans, root, 0);
 711         if (ret)
 712                 return ret;
 713         return 0;
 714 }
 715
 716 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 717                                    struct btrfs_root *root)
 718 {
 719         int ret;
 720
 721         ret = __btrfs_end_transaction(trans, root, 1);
 722         if (ret)
 723                 return ret;
 724         return 0;
 725 }
 726
 727 int btrfs_end_transaction_dmeta(struct btrfs_trans_handle *trans,
 728                                 struct btrfs_root *root)
 729 {
 730         return __btrfs_end_transaction(trans, root, 1);
 731 }
 732
 733 /*
 734  * when btree blocks are allocated, they have some corresponding bits set for
 735  * them in one of two extent_io trees.  This is used to make sure all of
 736  * those extents are sent to disk but does not wait on them
 737  */
 738 int btrfs_write_marked_extents(struct btrfs_root *root,
 739                                struct extent_io_tree *dirty_pages, int mark)
 740 {
 741         int err = 0;
 742         int werr = 0;
 743         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 744         struct extent_state *cached_state = NULL;
 745         u64 start = 0;
 746         u64 end;
 747
 748         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 749                                       mark, &cached_state)) {
 750                 convert_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT,
 751                                    mark, &cached_state, GFP_NOFS);
 752                 cached_state = NULL;
 753                 err = filemap_fdatawrite_range(mapping, start, end);
 754                 if (err)
 755                         werr = err;
 756                 cond_resched();
 757                 start = end + 1;
 758         }
 759         if (err)
 760                 werr = err;
 761         return werr;
 762 }
 763
 764 /*
 765  * when btree blocks are allocated, they have some corresponding bits set for
 766  * them in one of two extent_io trees.  This is used to make sure all of
 767  * those extents are on disk for transaction or log commit.  We wait
 768  * on all the pages and clear them from the dirty pages state tree
 769  */
 770 int btrfs_wait_marked_extents(struct btrfs_root *root,
 771                               struct extent_io_tree *dirty_pages, int mark)
 772 {
 773         int err = 0;
 774         int werr = 0;
 775         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 776         struct extent_state *cached_state = NULL;
 777         u64 start = 0;
 778         u64 end;
 779
 780         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 781                                       EXTENT_NEED_WAIT, &cached_state)) {
 782                 clear_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT,
 783                                  0, 0, &cached_state, GFP_NOFS);
 784                 err = filemap_fdatawait_range(mapping, start, end);
 785                 if (err)
 786                         werr = err;
 787                 cond_resched();
 788                 start = end + 1;
 789         }
 790         if (err)
 791                 werr = err;
 792         return werr;
 793 }
 794
 795 /*
 796  * when btree blocks are allocated, they have some corresponding bits set for
 797  * them in one of two extent_io trees.  This is used to make sure all of
 798  * those extents are on disk for transaction or log commit
 799  */
 800 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 801                                 struct extent_io_tree *dirty_pages, int mark)
 802 {
 803         int ret;
 804         int ret2;
 805
 806         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 807         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 808
 809         if (ret)
 810                 return ret;
 811         if (ret2)
 812                 return ret2;
 813         return 0;
 814 }
 815
 816 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 817                                      struct btrfs_root *root)
 818 {
 819         if (!trans || !trans->transaction) {
 820                 struct inode *btree_inode;
 821                 btree_inode = root->fs_info->btree_inode;
 822                 return filemap_write_and_wait(btree_inode->i_mapping);
 823         }
 824         return btrfs_write_and_wait_marked_extents(root,
 825                                            &trans->transaction->dirty_pages,
 826                                            EXTENT_DIRTY);
 827 }
 828
 829 /*
 830  * this is used to update the root pointer in the tree of tree roots.
 831  *
 832  * But, in the case of the extent allocation tree, updating the root
 833  * pointer may allocate blocks which may change the root of the extent
 834  * allocation tree.
 835  *
 836  * So, this loops and repeats and makes sure the cowonly root didn't
 837  * change while the root pointer was being updated in the metadata.
 838  */
 839 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 840                                struct btrfs_root *root)
 841 {
 842         int ret;
 843         u64 old_root_bytenr;
 844         u64 old_root_used;
 845         struct btrfs_root *tree_root = root->fs_info->tree_root;
 846
 847         old_root_used = btrfs_root_used(&root->root_item);
 848         btrfs_write_dirty_block_groups(trans, root);
 849
 850         while (1) {
 851                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 852                 if (old_root_bytenr == root->node->start &&
 853                     old_root_used == btrfs_root_used(&root->root_item))
 854                         break;
 855
 856                 btrfs_set_root_node(&root->root_item, root->node);
 857                 ret = btrfs_update_root(trans, tree_root,
 858                                         &root->root_key,
 859                                         &root->root_item);
 860                 if (ret)
 861                         return ret;
 862
 863                 old_root_used = btrfs_root_used(&root->root_item);
 864                 ret = btrfs_write_dirty_block_groups(trans, root);
 865                 if (ret)
 866                         return ret;
 867         }
 868
 869         if (root != root->fs_info->extent_root)
 870                 switch_commit_root(root);
 871
 872         return 0;
 873 }
 874
 875 /*
 876  * update all the cowonly tree roots on disk
 877  *
 878  * The error handling in this function may not be obvious. Any of the
 879  * failures will cause the file system to go offline. We still need
 880  * to clean up the delayed refs.
 881  */
 882 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 883                                          struct btrfs_root *root)
 884 {
 885         struct btrfs_fs_info *fs_info = root->fs_info;
 886         struct list_head *next;
 887         struct extent_buffer *eb;
 888         int ret;
 889
 890         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 891         if (ret)
 892                 return ret;
 893
 894         eb = btrfs_lock_root_node(fs_info->tree_root);
 895         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
 896                               0, &eb);
 897         btrfs_tree_unlock(eb);
 898         free_extent_buffer(eb);
 899
 900         if (ret)
 901                 return ret;
 902
 903         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 904         if (ret)
 905                 return ret;
 906
 907         ret = btrfs_run_dev_stats(trans, root->fs_info);
 908         WARN_ON(ret);
 909         ret = btrfs_run_dev_replace(trans, root->fs_info);
 910         WARN_ON(ret);
 911
 912         ret = btrfs_run_qgroups(trans, root->fs_info);
 913         BUG_ON(ret);
 914
 915         /* run_qgroups might have added some more refs */
 916         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 917         BUG_ON(ret);
 918
 919         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 920                 next = fs_info->dirty_cowonly_roots.next;
 921                 list_del_init(next);
 922                 root = list_entry(next, struct btrfs_root, dirty_list);
 923
 924                 ret = update_cowonly_root(trans, root);
 925                 if (ret)
 926                         return ret;
 927         }
 928
 929         down_write(&fs_info->extent_commit_sem);
 930         switch_commit_root(fs_info->extent_root);
 931         up_write(&fs_info->extent_commit_sem);
 932
 933         btrfs_after_dev_replace_commit(fs_info);
 934
 935         return 0;
 936 }
 937
 938 /*
 939  * dead roots are old snapshots that need to be deleted.  This allocates
 940  * a dirty root struct and adds it into the list of dead roots that need to
 941  * be deleted
 942  */
 943 int btrfs_add_dead_root(struct btrfs_root *root)
 944 {
 945         spin_lock(&root->fs_info->trans_lock);
 946         list_add(&root->root_list, &root->fs_info->dead_roots);
 947         spin_unlock(&root->fs_info->trans_lock);
 948         return 0;
 949 }
 950
 951 /*
 952  * update all the cowonly tree roots on disk
 953  */
 954 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 955                                     struct btrfs_root *root)
 956 {
 957         struct btrfs_root *gang[8];
 958         struct btrfs_fs_info *fs_info = root->fs_info;
 959         int i;
 960         int ret;
 961         int err = 0;
 962
 963         spin_lock(&fs_info->fs_roots_radix_lock);
 964         while (1) {
 965                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 966                                                  (void **)gang, 0,
 967                                                  ARRAY_SIZE(gang),
 968                                                  BTRFS_ROOT_TRANS_TAG);
 969                 if (ret == 0)
 970                         break;
 971                 for (i = 0; i < ret; i++) {
 972                         root = gang[i];
 973                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 974                                         (unsigned long)root->root_key.objectid,
 975                                         BTRFS_ROOT_TRANS_TAG);
 976                         spin_unlock(&fs_info->fs_roots_radix_lock);
 977
 978                         btrfs_free_log(trans, root);
 979                         btrfs_update_reloc_root(trans, root);
 980                         btrfs_orphan_commit_root(trans, root);
 981
 982                         btrfs_save_ino_cache(root, trans);
 983
 984                         /* see comments in should_cow_block() */
 985                         root->force_cow = 0;
 986                         smp_wmb();
 987
 988                         if (root->commit_root != root->node) {
 989                                 mutex_lock(&root->fs_commit_mutex);
 990                                 switch_commit_root(root);
 991                                 btrfs_unpin_free_ino(root);
 992                                 mutex_unlock(&root->fs_commit_mutex);
 993
 994                                 btrfs_set_root_node(&root->root_item,
 995                                                     root->node);
 996                         }
 997
 998                         err = btrfs_update_root(trans, fs_info->tree_root,
 999                                                 &root->root_key,
1000                                                 &root->root_item);
1001                         spin_lock(&fs_info->fs_roots_radix_lock);
1002                         if (err)
1003                                 break;
1004                 }
1005         }
1006         spin_unlock(&fs_info->fs_roots_radix_lock);
1007         return err;
1008 }
1009
1010 /*
1011  * defrag a given btree.
1012  * Every leaf in the btree is read and defragged.
1013  */
1014 int btrfs_defrag_root(struct btrfs_root *root)
1015 {
1016         struct btrfs_fs_info *info = root->fs_info;
1017         struct btrfs_trans_handle *trans;
1018         int ret;
1019
1020         if (xchg(&root->defrag_running, 1))
1021                 return 0;
1022
1023         while (1) {
1024                 trans = btrfs_start_transaction(root, 0);
1025                 if (IS_ERR(trans))
1026                         return PTR_ERR(trans);
1027
1028                 ret = btrfs_defrag_leaves(trans, root);
1029
1030                 btrfs_end_transaction(trans, root);
1031                 btrfs_btree_balance_dirty(info->tree_root);
1032                 cond_resched();
1033
1034                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
1035                         break;
1036
1037                 if (btrfs_defrag_cancelled(root->fs_info)) {
1038                         printk(KERN_DEBUG "btrfs: defrag_root cancelled\n");
1039                         ret = -EAGAIN;
1040                         break;
1041                 }
1042         }
1043         root->defrag_running = 0;
1044         return ret;
1045 }
1046
1047 /*
1048  * new snapshots need to be created at a very specific time in the
1049  * transaction commit.  This does the actual creation
1050  */
1051 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1052                                    struct btrfs_fs_info *fs_info,
1053                                    struct btrfs_pending_snapshot *pending)
1054 {
1055         struct btrfs_key key;
1056         struct btrfs_root_item *new_root_item;
1057         struct btrfs_root *tree_root = fs_info->tree_root;
1058         struct btrfs_root *root = pending->root;
1059         struct btrfs_root *parent_root;
1060         struct btrfs_block_rsv *rsv;
1061         struct inode *parent_inode;
1062         struct btrfs_path *path;
1063         struct btrfs_dir_item *dir_item;
1064         struct dentry *parent;
1065         struct dentry *dentry;
1066         struct extent_buffer *tmp;
1067         struct extent_buffer *old;
1068         struct timespec cur_time = CURRENT_TIME;
1069         int ret;
1070         u64 to_reserve = 0;
1071         u64 index = 0;
1072         u64 objectid;
1073         u64 root_flags;
1074         uuid_le new_uuid;
1075
1076         path = btrfs_alloc_path();
1077         if (!path) {
1078                 ret = pending->error = -ENOMEM;
1079                 goto path_alloc_fail;
1080         }
1081
1082         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
1083         if (!new_root_item) {
1084                 ret = pending->error = -ENOMEM;
1085                 goto root_item_alloc_fail;
1086         }
1087
1088         ret = btrfs_find_free_objectid(tree_root, &objectid);
1089         if (ret) {
1090                 pending->error = ret;
1091                 goto no_free_objectid;
1092         }
1093
1094         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
1095
1096         if (to_reserve > 0) {
1097                 ret = btrfs_block_rsv_add(root, &pending->block_rsv,
1098                                           to_reserve,
1099                                           BTRFS_RESERVE_NO_FLUSH);
1100                 if (ret) {
1101                         pending->error = ret;
1102                         goto no_free_objectid;
1103                 }
1104         }
1105
1106         ret = btrfs_qgroup_inherit(trans, fs_info, root->root_key.objectid,
1107                                    objectid, pending->inherit);
1108         if (ret) {
1109                 pending->error = ret;
1110                 goto no_free_objectid;
1111         }
1112
1113         key.objectid = objectid;
1114         key.offset = (u64)-1;
1115         key.type = BTRFS_ROOT_ITEM_KEY;
1116
1117         rsv = trans->block_rsv;
1118         trans->block_rsv = &pending->block_rsv;
1119
1120         dentry = pending->dentry;
1121         parent = dget_parent(dentry);
1122         parent_inode = parent->d_inode;
1123         parent_root = BTRFS_I(parent_inode)->root;
1124         record_root_in_trans(trans, parent_root);
1125
1126         /*
1127          * insert the directory item
1128          */
1129         ret = btrfs_set_inode_index(parent_inode, &index);
1130         BUG_ON(ret); /* -ENOMEM */
1131
1132         /* check if there is a file/dir which has the same name. */
1133         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1134                                          btrfs_ino(parent_inode),
1135                                          dentry->d_name.name,
1136                                          dentry->d_name.len, 0);
1137         if (dir_item != NULL && !IS_ERR(dir_item)) {
1138                 pending->error = -EEXIST;
1139                 goto fail;
1140         } else if (IS_ERR(dir_item)) {
1141                 ret = PTR_ERR(dir_item);
1142                 btrfs_abort_transaction(trans, root, ret);
1143                 goto fail;
1144         }
1145         btrfs_release_path(path);
1146
1147         /*
1148          * pull in the delayed directory update
1149          * and the delayed inode item
1150          * otherwise we corrupt the FS during
1151          * snapshot
1152          */
1153         ret = btrfs_run_delayed_items(trans, root);
1154         if (ret) {      /* Transaction aborted */
1155                 btrfs_abort_transaction(trans, root, ret);
1156                 goto fail;
1157         }
1158
1159         record_root_in_trans(trans, root);
1160         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1161         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1162         btrfs_check_and_init_root_item(new_root_item);
1163
1164         root_flags = btrfs_root_flags(new_root_item);
1165         if (pending->readonly)
1166                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1167         else
1168                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1169         btrfs_set_root_flags(new_root_item, root_flags);
1170
1171         btrfs_set_root_generation_v2(new_root_item,
1172                         trans->transid);
1173         uuid_le_gen(&new_uuid);
1174         memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1175         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1176                         BTRFS_UUID_SIZE);
1177         new_root_item->otime.sec = cpu_to_le64(cur_time.tv_sec);
1178         new_root_item->otime.nsec = cpu_to_le32(cur_time.tv_nsec);
1179         btrfs_set_root_otransid(new_root_item, trans->transid);
1180         memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1181         memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1182         btrfs_set_root_stransid(new_root_item, 0);
1183         btrfs_set_root_rtransid(new_root_item, 0);
1184
1185         old = btrfs_lock_root_node(root);
1186         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1187         if (ret) {
1188                 btrfs_tree_unlock(old);
1189                 free_extent_buffer(old);
1190                 btrfs_abort_transaction(trans, root, ret);
1191                 goto fail;
1192         }
1193
1194         btrfs_set_lock_blocking(old);
1195
1196         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1197         /* clean up in any case */
1198         btrfs_tree_unlock(old);
1199         free_extent_buffer(old);
1200         if (ret) {
1201                 btrfs_abort_transaction(trans, root, ret);
1202                 goto fail;
1203         }
1204
1205         /* see comments in should_cow_block() */
1206         root->force_cow = 1;
1207         smp_wmb();
1208
1209         btrfs_set_root_node(new_root_item, tmp);
1210         /* record when the snapshot was created in key.offset */
1211         key.offset = trans->transid;
1212         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1213         btrfs_tree_unlock(tmp);
1214         free_extent_buffer(tmp);
1215         if (ret) {
1216                 btrfs_abort_transaction(trans, root, ret);
1217                 goto fail;
1218         }
1219
1220         /*
1221          * insert root back/forward references
1222          */
1223         ret = btrfs_add_root_ref(trans, tree_root, objectid,
1224                                  parent_root->root_key.objectid,
1225                                  btrfs_ino(parent_inode), index,
1226                                  dentry->d_name.name, dentry->d_name.len);
1227         if (ret) {
1228                 btrfs_abort_transaction(trans, root, ret);
1229                 goto fail;
1230         }
1231
1232         key.offset = (u64)-1;
1233         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1234         if (IS_ERR(pending->snap)) {
1235                 ret = PTR_ERR(pending->snap);
1236                 btrfs_abort_transaction(trans, root, ret);
1237                 goto fail;
1238         }
1239
1240         ret = btrfs_reloc_post_snapshot(trans, pending);
1241         if (ret) {
1242                 btrfs_abort_transaction(trans, root, ret);
1243                 goto fail;
1244         }
1245
1246         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1247         if (ret) {
1248                 btrfs_abort_transaction(trans, root, ret);
1249                 goto fail;
1250         }
1251
1252         ret = btrfs_insert_dir_item(trans, parent_root,
1253                                     dentry->d_name.name, dentry->d_name.len,
1254                                     parent_inode, &key,
1255                                     BTRFS_FT_DIR, index);
1256         /* We have check then name at the beginning, so it is impossible. */
1257         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1258         if (ret) {
1259                 btrfs_abort_transaction(trans, root, ret);
1260                 goto fail;
1261         }
1262
1263         btrfs_i_size_write(parent_inode, parent_inode->i_size +
1264                                          dentry->d_name.len * 2);
1265         parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
1266         ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1267         if (ret)
1268                 btrfs_abort_transaction(trans, root, ret);
1269 fail:
1270         dput(parent);
1271         trans->block_rsv = rsv;
1272 no_free_objectid:
1273         kfree(new_root_item);
1274 root_item_alloc_fail:
1275         btrfs_free_path(path);
1276 path_alloc_fail:
1277         btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
1278         return ret;
1279 }
1280
1281 /*
1282  * create all the snapshots we've scheduled for creation
1283  */
1284 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1285                                              struct btrfs_fs_info *fs_info)
1286 {
1287         struct btrfs_pending_snapshot *pending;
1288         struct list_head *head = &trans->transaction->pending_snapshots;
1289
1290         list_for_each_entry(pending, head, list)
1291                 create_pending_snapshot(trans, fs_info, pending);
1292         return 0;
1293 }
1294
1295 static void update_super_roots(struct btrfs_root *root)
1296 {
1297         struct btrfs_root_item *root_item;
1298         struct btrfs_super_block *super;
1299
1300         super = root->fs_info->super_copy;
1301
1302         root_item = &root->fs_info->chunk_root->root_item;
1303         super->chunk_root = root_item->bytenr;
1304         super->chunk_root_generation = root_item->generation;
1305         super->chunk_root_level = root_item->level;
1306
1307         root_item = &root->fs_info->tree_root->root_item;
1308         super->root = root_item->bytenr;
1309         super->generation = root_item->generation;
1310         super->root_level = root_item->level;
1311         if (btrfs_test_opt(root, SPACE_CACHE))
1312                 super->cache_generation = root_item->generation;
1313 }
1314
1315 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1316 {
1317         int ret = 0;
1318         spin_lock(&info->trans_lock);
1319         if (info->running_transaction)
1320                 ret = info->running_transaction->in_commit;
1321         spin_unlock(&info->trans_lock);
1322         return ret;
1323 }
1324
1325 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1326 {
1327         int ret = 0;
1328         spin_lock(&info->trans_lock);
1329         if (info->running_transaction)
1330                 ret = info->running_transaction->blocked;
1331         spin_unlock(&info->trans_lock);
1332         return ret;
1333 }
1334
1335 /*
1336  * wait for the current transaction commit to start and block subsequent
1337  * transaction joins
1338  */
1339 static void wait_current_trans_commit_start(struct btrfs_root *root,
1340                                             struct btrfs_transaction *trans)
1341 {
1342         wait_event(root->fs_info->transaction_blocked_wait, trans->in_commit);
1343 }
1344
1345 /*
1346  * wait for the current transaction to start and then become unblocked.
1347  * caller holds ref.
1348  */
1349 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1350                                          struct btrfs_transaction *trans)
1351 {
1352         wait_event(root->fs_info->transaction_wait,
1353                    trans->commit_done || (trans->in_commit && !trans->blocked));
1354 }
1355
1356 /*
1357  * commit transactions asynchronously. once btrfs_commit_transaction_async
1358  * returns, any subsequent transaction will not be allowed to join.
1359  */
1360 struct btrfs_async_commit {
1361         struct btrfs_trans_handle *newtrans;
1362         struct btrfs_root *root;
1363         struct work_struct work;
1364 };
1365
1366 static void do_async_commit(struct work_struct *work)
1367 {
1368         struct btrfs_async_commit *ac =
1369                 container_of(work, struct btrfs_async_commit, work);
1370
1371         /*
1372          * We've got freeze protection passed with the transaction.
1373          * Tell lockdep about it.
1374          */
1375         if (ac->newtrans->type < TRANS_JOIN_NOLOCK)
1376                 rwsem_acquire_read(
1377                      &ac->root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
1378                      0, 1, _THIS_IP_);
1379
1380         current->journal_info = ac->newtrans;
1381
1382         btrfs_commit_transaction(ac->newtrans, ac->root);
1383         kfree(ac);
1384 }
1385
1386 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1387                                    struct btrfs_root *root,
1388                                    int wait_for_unblock)
1389 {
1390         struct btrfs_async_commit *ac;
1391         struct btrfs_transaction *cur_trans;
1392
1393         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1394         if (!ac)
1395                 return -ENOMEM;
1396
1397         INIT_WORK(&ac->work, do_async_commit);
1398         ac->root = root;
1399         ac->newtrans = btrfs_join_transaction(root);
1400         if (IS_ERR(ac->newtrans)) {
1401                 int err = PTR_ERR(ac->newtrans);
1402                 kfree(ac);
1403                 return err;
1404         }
1405
1406         /* take transaction reference */
1407         cur_trans = trans->transaction;
1408         atomic_inc(&cur_trans->use_count);
1409
1410         btrfs_end_transaction(trans, root);
1411
1412         /*
1413          * Tell lockdep we've released the freeze rwsem, since the
1414          * async commit thread will be the one to unlock it.
1415          */
1416         if (trans->type < TRANS_JOIN_NOLOCK)
1417                 rwsem_release(
1418                         &root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
1419                         1, _THIS_IP_);
1420
1421         schedule_work(&ac->work);
1422
1423         /* wait for transaction to start and unblock */
1424         if (wait_for_unblock)
1425                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1426         else
1427                 wait_current_trans_commit_start(root, cur_trans);
1428
1429         if (current->journal_info == trans)
1430                 current->journal_info = NULL;
1431
1432         put_transaction(cur_trans);
1433         return 0;
1434 }
1435
1436
1437 static void cleanup_transaction(struct btrfs_trans_handle *trans,
1438                                 struct btrfs_root *root, int err)
1439 {
1440         struct btrfs_transaction *cur_trans = trans->transaction;
1441
1442         WARN_ON(trans->use_count > 1);
1443
1444         btrfs_abort_transaction(trans, root, err);
1445
1446         spin_lock(&root->fs_info->trans_lock);
1447         list_del_init(&cur_trans->list);
1448         if (cur_trans == root->fs_info->running_transaction) {
1449                 root->fs_info->running_transaction = NULL;
1450                 root->fs_info->trans_no_join = 0;
1451         }
1452         spin_unlock(&root->fs_info->trans_lock);
1453
1454         btrfs_cleanup_one_transaction(trans->transaction, root);
1455
1456         put_transaction(cur_trans);
1457         put_transaction(cur_trans);
1458
1459         trace_btrfs_transaction_commit(root);
1460
1461         btrfs_scrub_continue(root);
1462
1463         if (current->journal_info == trans)
1464                 current->journal_info = NULL;
1465
1466         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1467 }
1468
1469 static int btrfs_flush_all_pending_stuffs(struct btrfs_trans_handle *trans,
1470                                           struct btrfs_root *root)
1471 {
1472         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1473         int snap_pending = 0;
1474         int ret;
1475
1476         if (!flush_on_commit) {
1477                 spin_lock(&root->fs_info->trans_lock);
1478                 if (!list_empty(&trans->transaction->pending_snapshots))
1479                         snap_pending = 1;
1480                 spin_unlock(&root->fs_info->trans_lock);
1481         }
1482
1483         if (flush_on_commit || snap_pending) {
1484                 ret = btrfs_start_delalloc_inodes(root, 1);
1485                 if (ret)
1486                         return ret;
1487                 btrfs_wait_ordered_extents(root, 1);
1488         }
1489
1490         ret = btrfs_run_delayed_items(trans, root);
1491         if (ret)
1492                 return ret;
1493
1494         /*
1495          * running the delayed items may have added new refs. account
1496          * them now so that they hinder processing of more delayed refs
1497          * as little as possible.
1498          */
1499         btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
1500
1501         /*
1502          * rename don't use btrfs_join_transaction, so, once we
1503          * set the transaction to blocked above, we aren't going
1504          * to get any new ordered operations.  We can safely run
1505          * it here and no for sure that nothing new will be added
1506          * to the list
1507          */
1508         ret = btrfs_run_ordered_operations(trans, root, 1);
1509
1510         return ret;
1511 }
1512
1513 /*
1514  * btrfs_transaction state sequence:
1515  *    in_commit = 0, blocked = 0  (initial)
1516  *    in_commit = 1, blocked = 1
1517  *    blocked = 0
1518  *    commit_done = 1
1519  */
1520 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1521                              struct btrfs_root *root)
1522 {
1523         unsigned long joined = 0;
1524         struct btrfs_transaction *cur_trans = trans->transaction;
1525         struct btrfs_transaction *prev_trans = NULL;
1526         DEFINE_WAIT(wait);
1527         int ret;
1528         int should_grow = 0;
1529         unsigned long now = get_seconds();
1530
1531         ret = btrfs_run_ordered_operations(trans, root, 0);
1532         if (ret) {
1533                 btrfs_abort_transaction(trans, root, ret);
1534                 btrfs_end_transaction(trans, root);
1535                 return ret;
1536         }
1537
1538         /* Stop the commit early if ->aborted is set */
1539         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1540                 ret = cur_trans->aborted;
1541                 btrfs_end_transaction(trans, root);
1542                 return ret;
1543         }
1544
1545         /* make a pass through all the delayed refs we have so far
1546          * any runnings procs may add more while we are here
1547          */
1548         ret = btrfs_run_delayed_refs(trans, root, 0);
1549         if (ret) {
1550                 btrfs_end_transaction(trans, root);
1551                 return ret;
1552         }
1553
1554         btrfs_trans_release_metadata(trans, root);
1555         trans->block_rsv = NULL;
1556         if (trans->qgroup_reserved) {
1557                 btrfs_qgroup_free(root, trans->qgroup_reserved);
1558                 trans->qgroup_reserved = 0;
1559         }
1560
1561         cur_trans = trans->transaction;
1562
1563         /*
1564          * set the flushing flag so procs in this transaction have to
1565          * start sending their work down.
1566          */
1567         cur_trans->delayed_refs.flushing = 1;
1568
1569         if (!list_empty(&trans->new_bgs))
1570                 btrfs_create_pending_block_groups(trans, root);
1571
1572         ret = btrfs_run_delayed_refs(trans, root, 0);
1573         if (ret) {
1574                 btrfs_end_transaction(trans, root);
1575                 return ret;
1576         }
1577
1578         spin_lock(&cur_trans->commit_lock);
1579         if (cur_trans->in_commit) {
1580                 spin_unlock(&cur_trans->commit_lock);
1581                 atomic_inc(&cur_trans->use_count);
1582                 ret = btrfs_end_transaction(trans, root);
1583
1584                 wait_for_commit(root, cur_trans);
1585
1586                 put_transaction(cur_trans);
1587
1588                 return ret;
1589         }
1590
1591         trans->transaction->in_commit = 1;
1592         trans->transaction->blocked = 1;
1593         spin_unlock(&cur_trans->commit_lock);
1594         wake_up(&root->fs_info->transaction_blocked_wait);
1595
1596         spin_lock(&root->fs_info->trans_lock);
1597         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1598                 prev_trans = list_entry(cur_trans->list.prev,
1599                                         struct btrfs_transaction, list);
1600                 if (!prev_trans->commit_done) {
1601                         atomic_inc(&prev_trans->use_count);
1602                         spin_unlock(&root->fs_info->trans_lock);
1603
1604                         wait_for_commit(root, prev_trans);
1605
1606                         put_transaction(prev_trans);
1607                 } else {
1608                         spin_unlock(&root->fs_info->trans_lock);
1609                 }
1610         } else {
1611                 spin_unlock(&root->fs_info->trans_lock);
1612         }
1613
1614         if (!btrfs_test_opt(root, SSD) &&
1615             (now < cur_trans->start_time || now - cur_trans->start_time < 1))
1616                 should_grow = 1;
1617
1618         do {
1619                 joined = cur_trans->num_joined;
1620
1621                 WARN_ON(cur_trans != trans->transaction);
1622
1623                 ret = btrfs_flush_all_pending_stuffs(trans, root);
1624                 if (ret)
1625                         goto cleanup_transaction;
1626
1627                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1628                                 TASK_UNINTERRUPTIBLE);
1629
1630                 if (atomic_read(&cur_trans->num_writers) > 1)
1631                         schedule_timeout(MAX_SCHEDULE_TIMEOUT);
1632                 else if (should_grow)
1633                         schedule_timeout(1);
1634
1635                 finish_wait(&cur_trans->writer_wait, &wait);
1636         } while (atomic_read(&cur_trans->num_writers) > 1 ||
1637                  (should_grow && cur_trans->num_joined != joined));
1638
1639         ret = btrfs_flush_all_pending_stuffs(trans, root);
1640         if (ret)
1641                 goto cleanup_transaction;
1642
1643         /*
1644          * Ok now we need to make sure to block out any other joins while we
1645          * commit the transaction.  We could have started a join before setting
1646          * no_join so make sure to wait for num_writers to == 1 again.
1647          */
1648         spin_lock(&root->fs_info->trans_lock);
1649         root->fs_info->trans_no_join = 1;
1650         spin_unlock(&root->fs_info->trans_lock);
1651         wait_event(cur_trans->writer_wait,
1652                    atomic_read(&cur_trans->num_writers) == 1);
1653
1654         /* ->aborted might be set after the previous check, so check it */
1655         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1656                 ret = cur_trans->aborted;
1657                 goto cleanup_transaction;
1658         }
1659         /*
1660          * the reloc mutex makes sure that we stop
1661          * the balancing code from coming in and moving
1662          * extents around in the middle of the commit
1663          */
1664         mutex_lock(&root->fs_info->reloc_mutex);
1665
1666         /*
1667          * We needn't worry about the delayed items because we will
1668          * deal with them in create_pending_snapshot(), which is the
1669          * core function of the snapshot creation.
1670          */
1671         ret = create_pending_snapshots(trans, root->fs_info);
1672         if (ret) {
1673                 mutex_unlock(&root->fs_info->reloc_mutex);
1674                 goto cleanup_transaction;
1675         }
1676
1677         /*
1678          * We insert the dir indexes of the snapshots and update the inode
1679          * of the snapshots' parents after the snapshot creation, so there
1680          * are some delayed items which are not dealt with. Now deal with
1681          * them.
1682          *
1683          * We needn't worry that this operation will corrupt the snapshots,
1684          * because all the tree which are snapshoted will be forced to COW
1685          * the nodes and leaves.
1686          */
1687         ret = btrfs_run_delayed_items(trans, root);
1688         if (ret) {
1689                 mutex_unlock(&root->fs_info->reloc_mutex);
1690                 goto cleanup_transaction;
1691         }
1692
1693         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1694         if (ret) {
1695                 mutex_unlock(&root->fs_info->reloc_mutex);
1696                 goto cleanup_transaction;
1697         }
1698
1699         /*
1700          * make sure none of the code above managed to slip in a
1701          * delayed item
1702          */
1703         btrfs_assert_delayed_root_empty(root);
1704
1705         WARN_ON(cur_trans != trans->transaction);
1706
1707         btrfs_scrub_pause(root);
1708         /* btrfs_commit_tree_roots is responsible for getting the
1709          * various roots consistent with each other.  Every pointer
1710          * in the tree of tree roots has to point to the most up to date
1711          * root for every subvolume and other tree.  So, we have to keep
1712          * the tree logging code from jumping in and changing any
1713          * of the trees.
1714          *
1715          * At this point in the commit, there can't be any tree-log
1716          * writers, but a little lower down we drop the trans mutex
1717          * and let new people in.  By holding the tree_log_mutex
1718          * from now until after the super is written, we avoid races
1719          * with the tree-log code.
1720          */
1721         mutex_lock(&root->fs_info->tree_log_mutex);
1722
1723         ret = commit_fs_roots(trans, root);
1724         if (ret) {
1725                 mutex_unlock(&root->fs_info->tree_log_mutex);
1726                 mutex_unlock(&root->fs_info->reloc_mutex);
1727                 goto cleanup_transaction;
1728         }
1729
1730         /* commit_fs_roots gets rid of all the tree log roots, it is now
1731          * safe to free the root of tree log roots
1732          */
1733         btrfs_free_log_root_tree(trans, root->fs_info);
1734
1735         ret = commit_cowonly_roots(trans, root);
1736         if (ret) {
1737                 mutex_unlock(&root->fs_info->tree_log_mutex);
1738                 mutex_unlock(&root->fs_info->reloc_mutex);
1739                 goto cleanup_transaction;
1740         }
1741
1742         /*
1743          * The tasks which save the space cache and inode cache may also
1744          * update ->aborted, check it.
1745          */
1746         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1747                 ret = cur_trans->aborted;
1748                 mutex_unlock(&root->fs_info->tree_log_mutex);
1749                 mutex_unlock(&root->fs_info->reloc_mutex);
1750                 goto cleanup_transaction;
1751         }
1752
1753         btrfs_prepare_extent_commit(trans, root);
1754
1755         cur_trans = root->fs_info->running_transaction;
1756
1757         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1758                             root->fs_info->tree_root->node);
1759         switch_commit_root(root->fs_info->tree_root);
1760
1761         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1762                             root->fs_info->chunk_root->node);
1763         switch_commit_root(root->fs_info->chunk_root);
1764
1765         assert_qgroups_uptodate(trans);
1766         update_super_roots(root);
1767
1768         if (!root->fs_info->log_root_recovering) {
1769                 btrfs_set_super_log_root(root->fs_info->super_copy, 0);
1770                 btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
1771         }
1772
1773         memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
1774                sizeof(*root->fs_info->super_copy));
1775
1776         trans->transaction->blocked = 0;
1777         spin_lock(&root->fs_info->trans_lock);
1778         root->fs_info->running_transaction = NULL;
1779         root->fs_info->trans_no_join = 0;
1780         spin_unlock(&root->fs_info->trans_lock);
1781         mutex_unlock(&root->fs_info->reloc_mutex);
1782
1783         wake_up(&root->fs_info->transaction_wait);
1784
1785         ret = btrfs_write_and_wait_transaction(trans, root);
1786         if (ret) {
1787                 btrfs_error(root->fs_info, ret,
1788                             "Error while writing out transaction.");
1789                 mutex_unlock(&root->fs_info->tree_log_mutex);
1790                 goto cleanup_transaction;
1791         }
1792
1793         ret = write_ctree_super(trans, root, 0);
1794         if (ret) {
1795                 mutex_unlock(&root->fs_info->tree_log_mutex);
1796                 goto cleanup_transaction;
1797         }
1798
1799         /*
1800          * the super is written, we can safely allow the tree-loggers
1801          * to go about their business
1802          */
1803         mutex_unlock(&root->fs_info->tree_log_mutex);
1804
1805         btrfs_finish_extent_commit(trans, root);
1806
1807         cur_trans->commit_done = 1;
1808
1809         root->fs_info->last_trans_committed = cur_trans->transid;
1810
1811         wake_up(&cur_trans->commit_wait);
1812
1813         spin_lock(&root->fs_info->trans_lock);
1814         list_del_init(&cur_trans->list);
1815         spin_unlock(&root->fs_info->trans_lock);
1816
1817         put_transaction(cur_trans);
1818         put_transaction(cur_trans);
1819
1820         if (trans->type < TRANS_JOIN_NOLOCK)
1821                 sb_end_intwrite(root->fs_info->sb);
1822
1823         trace_btrfs_transaction_commit(root);
1824
1825         btrfs_scrub_continue(root);
1826
1827         if (current->journal_info == trans)
1828                 current->journal_info = NULL;
1829
1830         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1831
1832         if (current != root->fs_info->transaction_kthread)
1833                 btrfs_run_delayed_iputs(root);
1834
1835         return ret;
1836
1837 cleanup_transaction:
1838         btrfs_trans_release_metadata(trans, root);
1839         trans->block_rsv = NULL;
1840         if (trans->qgroup_reserved) {
1841                 btrfs_qgroup_free(root, trans->qgroup_reserved);
1842                 trans->qgroup_reserved = 0;
1843         }
1844         btrfs_printk(root->fs_info, "Skipping commit of aborted transaction.\n");
1845 //      WARN_ON(1);
1846         if (current->journal_info == trans)
1847                 current->journal_info = NULL;
1848         cleanup_transaction(trans, root, ret);
1849
1850         return ret;
1851 }
1852
1853 /*
1854  * interface function to delete all the snapshots we have scheduled for deletion
1855  */
1856 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1857 {
1858         LIST_HEAD(list);
1859         struct btrfs_fs_info *fs_info = root->fs_info;
1860
1861         spin_lock(&fs_info->trans_lock);
1862         list_splice_init(&fs_info->dead_roots, &list);
1863         spin_unlock(&fs_info->trans_lock);
1864
1865         while (!list_empty(&list)) {
1866                 int ret;
1867
1868                 root = list_entry(list.next, struct btrfs_root, root_list);
1869                 list_del(&root->root_list);
1870
1871                 btrfs_kill_all_delayed_nodes(root);
1872
1873                 if (btrfs_header_backref_rev(root->node) <
1874                     BTRFS_MIXED_BACKREF_REV)
1875                         ret = btrfs_drop_snapshot(root, NULL, 0, 0);
1876                 else
1877                         ret =btrfs_drop_snapshot(root, NULL, 1, 0);
1878                 BUG_ON(ret < 0);
1879         }
1880         return 0;
1881 }