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 "ctree.h"
  26 #include "disk-io.h"
  27 #include "transaction.h"
  28 #include "locking.h"
  29 #include "tree-log.h"
  30 #include "inode-map.h"
  31
  32 #define BTRFS_ROOT_TRANS_TAG 0
  33
  34 static noinline void put_transaction(struct btrfs_transaction *transaction)
  35 {
  36         WARN_ON(atomic_read(&transaction->use_count) == 0);
  37         if (atomic_dec_and_test(&transaction->use_count)) {
  38                 BUG_ON(!list_empty(&transaction->list));
  39                 memset(transaction, 0, sizeof(*transaction));
  40                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  41         }
  42 }
  43
  44 static noinline void switch_commit_root(struct btrfs_root *root)
  45 {
  46         free_extent_buffer(root->commit_root);
  47         root->commit_root = btrfs_root_node(root);
  48 }
  49
  50 /*
  51  * either allocate a new transaction or hop into the existing one
  52  */
  53 static noinline int join_transaction(struct btrfs_root *root, int nofail)
  54 {
  55         struct btrfs_transaction *cur_trans;
  56
  57         spin_lock(&root->fs_info->trans_lock);
  58         if (root->fs_info->trans_no_join) {
  59                 if (!nofail) {
  60                         spin_unlock(&root->fs_info->trans_lock);
  61                         return -EBUSY;
  62                 }
  63         }
  64
  65         cur_trans = root->fs_info->running_transaction;
  66         if (cur_trans) {
  67                 atomic_inc(&cur_trans->use_count);
  68                 atomic_inc(&cur_trans->num_writers);
  69                 cur_trans->num_joined++;
  70                 spin_unlock(&root->fs_info->trans_lock);
  71                 return 0;
  72         }
  73         spin_unlock(&root->fs_info->trans_lock);
  74
  75         cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
  76         if (!cur_trans)
  77                 return -ENOMEM;
  78         spin_lock(&root->fs_info->trans_lock);
  79         if (root->fs_info->running_transaction) {
  80                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
  81                 cur_trans = root->fs_info->running_transaction;
  82                 atomic_inc(&cur_trans->use_count);
  83                 atomic_inc(&cur_trans->num_writers);
  84                 cur_trans->num_joined++;
  85                 spin_unlock(&root->fs_info->trans_lock);
  86                 return 0;
  87         }
  88         atomic_set(&cur_trans->num_writers, 1);
  89         cur_trans->num_joined = 0;
  90         init_waitqueue_head(&cur_trans->writer_wait);
  91         init_waitqueue_head(&cur_trans->commit_wait);
  92         cur_trans->in_commit = 0;
  93         cur_trans->blocked = 0;
  94         /*
  95          * One for this trans handle, one so it will live on until we
  96          * commit the transaction.
  97          */
  98         atomic_set(&cur_trans->use_count, 2);
  99         cur_trans->commit_done = 0;
 100         cur_trans->start_time = get_seconds();
 101
 102         cur_trans->delayed_refs.root = RB_ROOT;
 103         cur_trans->delayed_refs.num_entries = 0;
 104         cur_trans->delayed_refs.num_heads_ready = 0;
 105         cur_trans->delayed_refs.num_heads = 0;
 106         cur_trans->delayed_refs.flushing = 0;
 107         cur_trans->delayed_refs.run_delayed_start = 0;
 108         spin_lock_init(&cur_trans->commit_lock);
 109         spin_lock_init(&cur_trans->delayed_refs.lock);
 110
 111         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 112         list_add_tail(&cur_trans->list, &root->fs_info->trans_list);
 113         extent_io_tree_init(&cur_trans->dirty_pages,
 114                              root->fs_info->btree_inode->i_mapping);
 115         root->fs_info->generation++;
 116         cur_trans->transid = root->fs_info->generation;
 117         root->fs_info->running_transaction = cur_trans;
 118         spin_unlock(&root->fs_info->trans_lock);
 119
 120         return 0;
 121 }
 122
 123 /*
 124  * this does all the record keeping required to make sure that a reference
 125  * counted root is properly recorded in a given transaction.  This is required
 126  * to make sure the old root from before we joined the transaction is deleted
 127  * when the transaction commits
 128  */
 129 static int record_root_in_trans(struct btrfs_trans_handle *trans,
 130                                struct btrfs_root *root)
 131 {
 132         if (root->ref_cows && root->last_trans < trans->transid) {
 133                 WARN_ON(root == root->fs_info->extent_root);
 134                 WARN_ON(root->commit_root != root->node);
 135
 136                 /*
 137                  * see below for in_trans_setup usage rules
 138                  * we have the reloc mutex held now, so there
 139                  * is only one writer in this function
 140                  */
 141                 root->in_trans_setup = 1;
 142
 143                 /* make sure readers find in_trans_setup before
 144                  * they find our root->last_trans update
 145                  */
 146                 smp_wmb();
 147
 148                 spin_lock(&root->fs_info->fs_roots_radix_lock);
 149                 if (root->last_trans == trans->transid) {
 150                         spin_unlock(&root->fs_info->fs_roots_radix_lock);
 151                         return 0;
 152                 }
 153                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 154                            (unsigned long)root->root_key.objectid,
 155                            BTRFS_ROOT_TRANS_TAG);
 156                 spin_unlock(&root->fs_info->fs_roots_radix_lock);
 157                 root->last_trans = trans->transid;
 158
 159                 /* this is pretty tricky.  We don't want to
 160                  * take the relocation lock in btrfs_record_root_in_trans
 161                  * unless we're really doing the first setup for this root in
 162                  * this transaction.
 163                  *
 164                  * Normally we'd use root->last_trans as a flag to decide
 165                  * if we want to take the expensive mutex.
 166                  *
 167                  * But, we have to set root->last_trans before we
 168                  * init the relocation root, otherwise, we trip over warnings
 169                  * in ctree.c.  The solution used here is to flag ourselves
 170                  * with root->in_trans_setup.  When this is 1, we're still
 171                  * fixing up the reloc trees and everyone must wait.
 172                  *
 173                  * When this is zero, they can trust root->last_trans and fly
 174                  * through btrfs_record_root_in_trans without having to take the
 175                  * lock.  smp_wmb() makes sure that all the writes above are
 176                  * done before we pop in the zero below
 177                  */
 178                 btrfs_init_reloc_root(trans, root);
 179                 smp_wmb();
 180                 root->in_trans_setup = 0;
 181         }
 182         return 0;
 183 }
 184
 185
 186 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 187                                struct btrfs_root *root)
 188 {
 189         if (!root->ref_cows)
 190                 return 0;
 191
 192         /*
 193          * see record_root_in_trans for comments about in_trans_setup usage
 194          * and barriers
 195          */
 196         smp_rmb();
 197         if (root->last_trans == trans->transid &&
 198             !root->in_trans_setup)
 199                 return 0;
 200
 201         mutex_lock(&root->fs_info->reloc_mutex);
 202         record_root_in_trans(trans, root);
 203         mutex_unlock(&root->fs_info->reloc_mutex);
 204
 205         return 0;
 206 }
 207
 208 /* wait for commit against the current transaction to become unblocked
 209  * when this is done, it is safe to start a new transaction, but the current
 210  * transaction might not be fully on disk.
 211  */
 212 static void wait_current_trans(struct btrfs_root *root)
 213 {
 214         struct btrfs_transaction *cur_trans;
 215
 216         spin_lock(&root->fs_info->trans_lock);
 217         cur_trans = root->fs_info->running_transaction;
 218         if (cur_trans && cur_trans->blocked) {
 219                 DEFINE_WAIT(wait);
 220                 atomic_inc(&cur_trans->use_count);
 221                 spin_unlock(&root->fs_info->trans_lock);
 222                 while (1) {
 223                         prepare_to_wait(&root->fs_info->transaction_wait, &wait,
 224                                         TASK_UNINTERRUPTIBLE);
 225                         if (!cur_trans->blocked)
 226                                 break;
 227                         schedule();
 228                 }
 229                 finish_wait(&root->fs_info->transaction_wait, &wait);
 230                 put_transaction(cur_trans);
 231         } else {
 232                 spin_unlock(&root->fs_info->trans_lock);
 233         }
 234 }
 235
 236 enum btrfs_trans_type {
 237         TRANS_START,
 238         TRANS_JOIN,
 239         TRANS_USERSPACE,
 240         TRANS_JOIN_NOLOCK,
 241 };
 242
 243 static int may_wait_transaction(struct btrfs_root *root, int type)
 244 {
 245         if (root->fs_info->log_root_recovering)
 246                 return 0;
 247
 248         if (type == TRANS_USERSPACE)
 249                 return 1;
 250
 251         if (type == TRANS_START &&
 252             !atomic_read(&root->fs_info->open_ioctl_trans))
 253                 return 1;
 254
 255         return 0;
 256 }
 257
 258 static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
 259                                                     u64 num_items, int type)
 260 {
 261         struct btrfs_trans_handle *h;
 262         struct btrfs_transaction *cur_trans;
 263         int retries = 0;
 264         int ret;
 265
 266         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR)
 267                 return ERR_PTR(-EROFS);
 268
 269         if (current->journal_info) {
 270                 WARN_ON(type != TRANS_JOIN && type != TRANS_JOIN_NOLOCK);
 271                 h = current->journal_info;
 272                 h->use_count++;
 273                 h->orig_rsv = h->block_rsv;
 274                 h->block_rsv = NULL;
 275                 goto got_it;
 276         }
 277 again:
 278         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 279         if (!h)
 280                 return ERR_PTR(-ENOMEM);
 281
 282         if (may_wait_transaction(root, type))
 283                 wait_current_trans(root);
 284
 285         do {
 286                 ret = join_transaction(root, type == TRANS_JOIN_NOLOCK);
 287                 if (ret == -EBUSY)
 288                         wait_current_trans(root);
 289         } while (ret == -EBUSY);
 290
 291         if (ret < 0) {
 292                 kmem_cache_free(btrfs_trans_handle_cachep, h);
 293                 return ERR_PTR(ret);
 294         }
 295
 296         cur_trans = root->fs_info->running_transaction;
 297
 298         h->transid = cur_trans->transid;
 299         h->transaction = cur_trans;
 300         h->blocks_used = 0;
 301         h->bytes_reserved = 0;
 302         h->delayed_ref_updates = 0;
 303         h->use_count = 1;
 304         h->block_rsv = NULL;
 305         h->orig_rsv = NULL;
 306
 307         smp_mb();
 308         if (cur_trans->blocked && may_wait_transaction(root, type)) {
 309                 btrfs_commit_transaction(h, root);
 310                 goto again;
 311         }
 312
 313         if (num_items > 0) {
 314                 ret = btrfs_trans_reserve_metadata(h, root, num_items);
 315                 if (ret == -EAGAIN && !retries) {
 316                         retries++;
 317                         btrfs_commit_transaction(h, root);
 318                         goto again;
 319                 } else if (ret == -EAGAIN) {
 320                         /*
 321                          * We have already retried and got EAGAIN, so really we
 322                          * don't have space, so set ret to -ENOSPC.
 323                          */
 324                         ret = -ENOSPC;
 325                 }
 326
 327                 if (ret < 0) {
 328                         btrfs_end_transaction(h, root);
 329                         return ERR_PTR(ret);
 330                 }
 331         }
 332
 333 got_it:
 334         btrfs_record_root_in_trans(h, root);
 335
 336         if (!current->journal_info && type != TRANS_USERSPACE)
 337                 current->journal_info = h;
 338         return h;
 339 }
 340
 341 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 342                                                    int num_items)
 343 {
 344         return start_transaction(root, num_items, TRANS_START);
 345 }
 346 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 347 {
 348         return start_transaction(root, 0, TRANS_JOIN);
 349 }
 350
 351 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 352 {
 353         return start_transaction(root, 0, TRANS_JOIN_NOLOCK);
 354 }
 355
 356 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 357 {
 358         return start_transaction(root, 0, TRANS_USERSPACE);
 359 }
 360
 361 /* wait for a transaction commit to be fully complete */
 362 static noinline int wait_for_commit(struct btrfs_root *root,
 363                                     struct btrfs_transaction *commit)
 364 {
 365         DEFINE_WAIT(wait);
 366         while (!commit->commit_done) {
 367                 prepare_to_wait(&commit->commit_wait, &wait,
 368                                 TASK_UNINTERRUPTIBLE);
 369                 if (commit->commit_done)
 370                         break;
 371                 schedule();
 372         }
 373         finish_wait(&commit->commit_wait, &wait);
 374         return 0;
 375 }
 376
 377 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 378 {
 379         struct btrfs_transaction *cur_trans = NULL, *t;
 380         int ret;
 381
 382         ret = 0;
 383         if (transid) {
 384                 if (transid <= root->fs_info->last_trans_committed)
 385                         goto out;
 386
 387                 /* find specified transaction */
 388                 spin_lock(&root->fs_info->trans_lock);
 389                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 390                         if (t->transid == transid) {
 391                                 cur_trans = t;
 392                                 atomic_inc(&cur_trans->use_count);
 393                                 break;
 394                         }
 395                         if (t->transid > transid)
 396                                 break;
 397                 }
 398                 spin_unlock(&root->fs_info->trans_lock);
 399                 ret = -EINVAL;
 400                 if (!cur_trans)
 401                         goto out;  /* bad transid */
 402         } else {
 403                 /* find newest transaction that is committing | committed */
 404                 spin_lock(&root->fs_info->trans_lock);
 405                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 406                                             list) {
 407                         if (t->in_commit) {
 408                                 if (t->commit_done)
 409                                         break;
 410                                 cur_trans = t;
 411                                 atomic_inc(&cur_trans->use_count);
 412                                 break;
 413                         }
 414                 }
 415                 spin_unlock(&root->fs_info->trans_lock);
 416                 if (!cur_trans)
 417                         goto out;  /* nothing committing|committed */
 418         }
 419
 420         wait_for_commit(root, cur_trans);
 421
 422         put_transaction(cur_trans);
 423         ret = 0;
 424 out:
 425         return ret;
 426 }
 427
 428 void btrfs_throttle(struct btrfs_root *root)
 429 {
 430         if (!atomic_read(&root->fs_info->open_ioctl_trans))
 431                 wait_current_trans(root);
 432 }
 433
 434 static int should_end_transaction(struct btrfs_trans_handle *trans,
 435                                   struct btrfs_root *root)
 436 {
 437         int ret;
 438         ret = btrfs_block_rsv_check(trans, root,
 439                                     &root->fs_info->global_block_rsv, 0, 5);
 440         return ret ? 1 : 0;
 441 }
 442
 443 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 444                                  struct btrfs_root *root)
 445 {
 446         struct btrfs_transaction *cur_trans = trans->transaction;
 447         int updates;
 448
 449         smp_mb();
 450         if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
 451                 return 1;
 452
 453         updates = trans->delayed_ref_updates;
 454         trans->delayed_ref_updates = 0;
 455         if (updates)
 456                 btrfs_run_delayed_refs(trans, root, updates);
 457
 458         return should_end_transaction(trans, root);
 459 }
 460
 461 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 462                           struct btrfs_root *root, int throttle, int lock)
 463 {
 464         struct btrfs_transaction *cur_trans = trans->transaction;
 465         struct btrfs_fs_info *info = root->fs_info;
 466         int count = 0;
 467
 468         if (--trans->use_count) {
 469                 trans->block_rsv = trans->orig_rsv;
 470                 return 0;
 471         }
 472
 473         while (count < 4) {
 474                 unsigned long cur = trans->delayed_ref_updates;
 475                 trans->delayed_ref_updates = 0;
 476                 if (cur &&
 477                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 478                         trans->delayed_ref_updates = 0;
 479
 480                         /*
 481                          * do a full flush if the transaction is trying
 482                          * to close
 483                          */
 484                         if (trans->transaction->delayed_refs.flushing)
 485                                 cur = 0;
 486                         btrfs_run_delayed_refs(trans, root, cur);
 487                 } else {
 488                         break;
 489                 }
 490                 count++;
 491         }
 492
 493         btrfs_trans_release_metadata(trans, root);
 494
 495         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
 496             should_end_transaction(trans, root)) {
 497                 trans->transaction->blocked = 1;
 498                 smp_wmb();
 499         }
 500
 501         if (lock && cur_trans->blocked && !cur_trans->in_commit) {
 502                 if (throttle)
 503                         return btrfs_commit_transaction(trans, root);
 504                 else
 505                         wake_up_process(info->transaction_kthread);
 506         }
 507
 508         WARN_ON(cur_trans != info->running_transaction);
 509         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 510         atomic_dec(&cur_trans->num_writers);
 511
 512         smp_mb();
 513         if (waitqueue_active(&cur_trans->writer_wait))
 514                 wake_up(&cur_trans->writer_wait);
 515         put_transaction(cur_trans);
 516
 517         if (current->journal_info == trans)
 518                 current->journal_info = NULL;
 519         memset(trans, 0, sizeof(*trans));
 520         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 521
 522         if (throttle)
 523                 btrfs_run_delayed_iputs(root);
 524
 525         return 0;
 526 }
 527
 528 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 529                           struct btrfs_root *root)
 530 {
 531         int ret;
 532
 533         ret = __btrfs_end_transaction(trans, root, 0, 1);
 534         if (ret)
 535                 return ret;
 536         return 0;
 537 }
 538
 539 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 540                                    struct btrfs_root *root)
 541 {
 542         int ret;
 543
 544         ret = __btrfs_end_transaction(trans, root, 1, 1);
 545         if (ret)
 546                 return ret;
 547         return 0;
 548 }
 549
 550 int btrfs_end_transaction_nolock(struct btrfs_trans_handle *trans,
 551                                  struct btrfs_root *root)
 552 {
 553         int ret;
 554
 555         ret = __btrfs_end_transaction(trans, root, 0, 0);
 556         if (ret)
 557                 return ret;
 558         return 0;
 559 }
 560
 561 int btrfs_end_transaction_dmeta(struct btrfs_trans_handle *trans,
 562                                 struct btrfs_root *root)
 563 {
 564         return __btrfs_end_transaction(trans, root, 1, 1);
 565 }
 566
 567 /*
 568  * when btree blocks are allocated, they have some corresponding bits set for
 569  * them in one of two extent_io trees.  This is used to make sure all of
 570  * those extents are sent to disk but does not wait on them
 571  */
 572 int btrfs_write_marked_extents(struct btrfs_root *root,
 573                                struct extent_io_tree *dirty_pages, int mark)
 574 {
 575         int ret;
 576         int err = 0;
 577         int werr = 0;
 578         struct page *page;
 579         struct inode *btree_inode = root->fs_info->btree_inode;
 580         u64 start = 0;
 581         u64 end;
 582         unsigned long index;
 583
 584         while (1) {
 585                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 586                                             mark);
 587                 if (ret)
 588                         break;
 589                 while (start <= end) {
 590                         cond_resched();
 591
 592                         index = start >> PAGE_CACHE_SHIFT;
 593                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 594                         page = find_get_page(btree_inode->i_mapping, index);
 595                         if (!page)
 596                                 continue;
 597
 598                         btree_lock_page_hook(page);
 599                         if (!page->mapping) {
 600                                 unlock_page(page);
 601                                 page_cache_release(page);
 602                                 continue;
 603                         }
 604
 605                         if (PageWriteback(page)) {
 606                                 if (PageDirty(page))
 607                                         wait_on_page_writeback(page);
 608                                 else {
 609                                         unlock_page(page);
 610                                         page_cache_release(page);
 611                                         continue;
 612                                 }
 613                         }
 614                         err = write_one_page(page, 0);
 615                         if (err)
 616                                 werr = err;
 617                         page_cache_release(page);
 618                 }
 619         }
 620         if (err)
 621                 werr = err;
 622         return werr;
 623 }
 624
 625 /*
 626  * when btree blocks are allocated, they have some corresponding bits set for
 627  * them in one of two extent_io trees.  This is used to make sure all of
 628  * those extents are on disk for transaction or log commit.  We wait
 629  * on all the pages and clear them from the dirty pages state tree
 630  */
 631 int btrfs_wait_marked_extents(struct btrfs_root *root,
 632                               struct extent_io_tree *dirty_pages, int mark)
 633 {
 634         int ret;
 635         int err = 0;
 636         int werr = 0;
 637         struct page *page;
 638         struct inode *btree_inode = root->fs_info->btree_inode;
 639         u64 start = 0;
 640         u64 end;
 641         unsigned long index;
 642
 643         while (1) {
 644                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 645                                             mark);
 646                 if (ret)
 647                         break;
 648
 649                 clear_extent_bits(dirty_pages, start, end, mark, GFP_NOFS);
 650                 while (start <= end) {
 651                         index = start >> PAGE_CACHE_SHIFT;
 652                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 653                         page = find_get_page(btree_inode->i_mapping, index);
 654                         if (!page)
 655                                 continue;
 656                         if (PageDirty(page)) {
 657                                 btree_lock_page_hook(page);
 658                                 wait_on_page_writeback(page);
 659                                 err = write_one_page(page, 0);
 660                                 if (err)
 661                                         werr = err;
 662                         }
 663                         wait_on_page_writeback(page);
 664                         page_cache_release(page);
 665                         cond_resched();
 666                 }
 667         }
 668         if (err)
 669                 werr = err;
 670         return werr;
 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 on disk for transaction or log commit
 677  */
 678 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 679                                 struct extent_io_tree *dirty_pages, int mark)
 680 {
 681         int ret;
 682         int ret2;
 683
 684         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 685         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 686         return ret || ret2;
 687 }
 688
 689 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 690                                      struct btrfs_root *root)
 691 {
 692         if (!trans || !trans->transaction) {
 693                 struct inode *btree_inode;
 694                 btree_inode = root->fs_info->btree_inode;
 695                 return filemap_write_and_wait(btree_inode->i_mapping);
 696         }
 697         return btrfs_write_and_wait_marked_extents(root,
 698                                            &trans->transaction->dirty_pages,
 699                                            EXTENT_DIRTY);
 700 }
 701
 702 /*
 703  * this is used to update the root pointer in the tree of tree roots.
 704  *
 705  * But, in the case of the extent allocation tree, updating the root
 706  * pointer may allocate blocks which may change the root of the extent
 707  * allocation tree.
 708  *
 709  * So, this loops and repeats and makes sure the cowonly root didn't
 710  * change while the root pointer was being updated in the metadata.
 711  */
 712 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 713                                struct btrfs_root *root)
 714 {
 715         int ret;
 716         u64 old_root_bytenr;
 717         u64 old_root_used;
 718         struct btrfs_root *tree_root = root->fs_info->tree_root;
 719
 720         old_root_used = btrfs_root_used(&root->root_item);
 721         btrfs_write_dirty_block_groups(trans, root);
 722
 723         while (1) {
 724                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 725                 if (old_root_bytenr == root->node->start &&
 726                     old_root_used == btrfs_root_used(&root->root_item))
 727                         break;
 728
 729                 btrfs_set_root_node(&root->root_item, root->node);
 730                 ret = btrfs_update_root(trans, tree_root,
 731                                         &root->root_key,
 732                                         &root->root_item);
 733                 BUG_ON(ret);
 734
 735                 old_root_used = btrfs_root_used(&root->root_item);
 736                 ret = btrfs_write_dirty_block_groups(trans, root);
 737                 BUG_ON(ret);
 738         }
 739
 740         if (root != root->fs_info->extent_root)
 741                 switch_commit_root(root);
 742
 743         return 0;
 744 }
 745
 746 /*
 747  * update all the cowonly tree roots on disk
 748  */
 749 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 750                                          struct btrfs_root *root)
 751 {
 752         struct btrfs_fs_info *fs_info = root->fs_info;
 753         struct list_head *next;
 754         struct extent_buffer *eb;
 755         int ret;
 756
 757         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 758         BUG_ON(ret);
 759
 760         eb = btrfs_lock_root_node(fs_info->tree_root);
 761         btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
 762         btrfs_tree_unlock(eb);
 763         free_extent_buffer(eb);
 764
 765         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 766         BUG_ON(ret);
 767
 768         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 769                 next = fs_info->dirty_cowonly_roots.next;
 770                 list_del_init(next);
 771                 root = list_entry(next, struct btrfs_root, dirty_list);
 772
 773                 update_cowonly_root(trans, root);
 774         }
 775
 776         down_write(&fs_info->extent_commit_sem);
 777         switch_commit_root(fs_info->extent_root);
 778         up_write(&fs_info->extent_commit_sem);
 779
 780         return 0;
 781 }
 782
 783 /*
 784  * dead roots are old snapshots that need to be deleted.  This allocates
 785  * a dirty root struct and adds it into the list of dead roots that need to
 786  * be deleted
 787  */
 788 int btrfs_add_dead_root(struct btrfs_root *root)
 789 {
 790         spin_lock(&root->fs_info->trans_lock);
 791         list_add(&root->root_list, &root->fs_info->dead_roots);
 792         spin_unlock(&root->fs_info->trans_lock);
 793         return 0;
 794 }
 795
 796 /*
 797  * update all the cowonly tree roots on disk
 798  */
 799 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 800                                     struct btrfs_root *root)
 801 {
 802         struct btrfs_root *gang[8];
 803         struct btrfs_fs_info *fs_info = root->fs_info;
 804         int i;
 805         int ret;
 806         int err = 0;
 807
 808         spin_lock(&fs_info->fs_roots_radix_lock);
 809         while (1) {
 810                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 811                                                  (void **)gang, 0,
 812                                                  ARRAY_SIZE(gang),
 813                                                  BTRFS_ROOT_TRANS_TAG);
 814                 if (ret == 0)
 815                         break;
 816                 for (i = 0; i < ret; i++) {
 817                         root = gang[i];
 818                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 819                                         (unsigned long)root->root_key.objectid,
 820                                         BTRFS_ROOT_TRANS_TAG);
 821                         spin_unlock(&fs_info->fs_roots_radix_lock);
 822
 823                         btrfs_free_log(trans, root);
 824                         btrfs_update_reloc_root(trans, root);
 825                         btrfs_orphan_commit_root(trans, root);
 826
 827                         btrfs_save_ino_cache(root, trans);
 828
 829                         if (root->commit_root != root->node) {
 830                                 mutex_lock(&root->fs_commit_mutex);
 831                                 switch_commit_root(root);
 832                                 btrfs_unpin_free_ino(root);
 833                                 mutex_unlock(&root->fs_commit_mutex);
 834
 835                                 btrfs_set_root_node(&root->root_item,
 836                                                     root->node);
 837                         }
 838
 839                         err = btrfs_update_root(trans, fs_info->tree_root,
 840                                                 &root->root_key,
 841                                                 &root->root_item);
 842                         spin_lock(&fs_info->fs_roots_radix_lock);
 843                         if (err)
 844                                 break;
 845                 }
 846         }
 847         spin_unlock(&fs_info->fs_roots_radix_lock);
 848         return err;
 849 }
 850
 851 /*
 852  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 853  * otherwise every leaf in the btree is read and defragged.
 854  */
 855 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 856 {
 857         struct btrfs_fs_info *info = root->fs_info;
 858         struct btrfs_trans_handle *trans;
 859         int ret;
 860         unsigned long nr;
 861
 862         if (xchg(&root->defrag_running, 1))
 863                 return 0;
 864
 865         while (1) {
 866                 trans = btrfs_start_transaction(root, 0);
 867                 if (IS_ERR(trans))
 868                         return PTR_ERR(trans);
 869
 870                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 871
 872                 nr = trans->blocks_used;
 873                 btrfs_end_transaction(trans, root);
 874                 btrfs_btree_balance_dirty(info->tree_root, nr);
 875                 cond_resched();
 876
 877                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
 878                         break;
 879         }
 880         root->defrag_running = 0;
 881         return ret;
 882 }
 883
 884 /*
 885  * new snapshots need to be created at a very specific time in the
 886  * transaction commit.  This does the actual creation
 887  */
 888 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 889                                    struct btrfs_fs_info *fs_info,
 890                                    struct btrfs_pending_snapshot *pending)
 891 {
 892         struct btrfs_key key;
 893         struct btrfs_root_item *new_root_item;
 894         struct btrfs_root *tree_root = fs_info->tree_root;
 895         struct btrfs_root *root = pending->root;
 896         struct btrfs_root *parent_root;
 897         struct inode *parent_inode;
 898         struct dentry *parent;
 899         struct dentry *dentry;
 900         struct extent_buffer *tmp;
 901         struct extent_buffer *old;
 902         int ret;
 903         u64 to_reserve = 0;
 904         u64 index = 0;
 905         u64 objectid;
 906         u64 root_flags;
 907
 908         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 909         if (!new_root_item) {
 910                 pending->error = -ENOMEM;
 911                 goto fail;
 912         }
 913
 914         ret = btrfs_find_free_objectid(tree_root, &objectid);
 915         if (ret) {
 916                 pending->error = ret;
 917                 goto fail;
 918         }
 919
 920         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
 921         btrfs_orphan_pre_snapshot(trans, pending, &to_reserve);
 922
 923         if (to_reserve > 0) {
 924                 ret = btrfs_block_rsv_add(trans, root, &pending->block_rsv,
 925                                           to_reserve);
 926                 if (ret) {
 927                         pending->error = ret;
 928                         goto fail;
 929                 }
 930         }
 931
 932         key.objectid = objectid;
 933         key.offset = (u64)-1;
 934         key.type = BTRFS_ROOT_ITEM_KEY;
 935
 936         trans->block_rsv = &pending->block_rsv;
 937
 938         dentry = pending->dentry;
 939         parent = dget_parent(dentry);
 940         parent_inode = parent->d_inode;
 941         parent_root = BTRFS_I(parent_inode)->root;
 942         record_root_in_trans(trans, parent_root);
 943
 944         /*
 945          * insert the directory item
 946          */
 947         ret = btrfs_set_inode_index(parent_inode, &index);
 948         BUG_ON(ret);
 949         ret = btrfs_insert_dir_item(trans, parent_root,
 950                                 dentry->d_name.name, dentry->d_name.len,
 951                                 parent_inode, &key,
 952                                 BTRFS_FT_DIR, index);
 953         BUG_ON(ret);
 954
 955         btrfs_i_size_write(parent_inode, parent_inode->i_size +
 956                                          dentry->d_name.len * 2);
 957         ret = btrfs_update_inode(trans, parent_root, parent_inode);
 958         BUG_ON(ret);
 959
 960         /*
 961          * pull in the delayed directory update
 962          * and the delayed inode item
 963          * otherwise we corrupt the FS during
 964          * snapshot
 965          */
 966         ret = btrfs_run_delayed_items(trans, root);
 967         BUG_ON(ret);
 968
 969         record_root_in_trans(trans, root);
 970         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
 971         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
 972         btrfs_check_and_init_root_item(new_root_item);
 973
 974         root_flags = btrfs_root_flags(new_root_item);
 975         if (pending->readonly)
 976                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
 977         else
 978                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
 979         btrfs_set_root_flags(new_root_item, root_flags);
 980
 981         old = btrfs_lock_root_node(root);
 982         btrfs_cow_block(trans, root, old, NULL, 0, &old);
 983         btrfs_set_lock_blocking(old);
 984
 985         btrfs_copy_root(trans, root, old, &tmp, objectid);
 986         btrfs_tree_unlock(old);
 987         free_extent_buffer(old);
 988
 989         btrfs_set_root_node(new_root_item, tmp);
 990         /* record when the snapshot was created in key.offset */
 991         key.offset = trans->transid;
 992         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
 993         btrfs_tree_unlock(tmp);
 994         free_extent_buffer(tmp);
 995         BUG_ON(ret);
 996
 997         /*
 998          * insert root back/forward references
 999          */
1000         ret = btrfs_add_root_ref(trans, tree_root, objectid,
1001                                  parent_root->root_key.objectid,
1002                                  btrfs_ino(parent_inode), index,
1003                                  dentry->d_name.name, dentry->d_name.len);
1004         BUG_ON(ret);
1005         dput(parent);
1006
1007         key.offset = (u64)-1;
1008         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1009         BUG_ON(IS_ERR(pending->snap));
1010
1011         btrfs_reloc_post_snapshot(trans, pending);
1012         btrfs_orphan_post_snapshot(trans, pending);
1013 fail:
1014         kfree(new_root_item);
1015         btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
1016         return 0;
1017 }
1018
1019 /*
1020  * create all the snapshots we've scheduled for creation
1021  */
1022 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1023                                              struct btrfs_fs_info *fs_info)
1024 {
1025         struct btrfs_pending_snapshot *pending;
1026         struct list_head *head = &trans->transaction->pending_snapshots;
1027         int ret;
1028
1029         list_for_each_entry(pending, head, list) {
1030                 ret = create_pending_snapshot(trans, fs_info, pending);
1031                 BUG_ON(ret);
1032         }
1033         return 0;
1034 }
1035
1036 static void update_super_roots(struct btrfs_root *root)
1037 {
1038         struct btrfs_root_item *root_item;
1039         struct btrfs_super_block *super;
1040
1041         super = &root->fs_info->super_copy;
1042
1043         root_item = &root->fs_info->chunk_root->root_item;
1044         super->chunk_root = root_item->bytenr;
1045         super->chunk_root_generation = root_item->generation;
1046         super->chunk_root_level = root_item->level;
1047
1048         root_item = &root->fs_info->tree_root->root_item;
1049         super->root = root_item->bytenr;
1050         super->generation = root_item->generation;
1051         super->root_level = root_item->level;
1052         if (super->cache_generation != 0 || btrfs_test_opt(root, SPACE_CACHE))
1053                 super->cache_generation = root_item->generation;
1054 }
1055
1056 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1057 {
1058         int ret = 0;
1059         spin_lock(&info->trans_lock);
1060         if (info->running_transaction)
1061                 ret = info->running_transaction->in_commit;
1062         spin_unlock(&info->trans_lock);
1063         return ret;
1064 }
1065
1066 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1067 {
1068         int ret = 0;
1069         spin_lock(&info->trans_lock);
1070         if (info->running_transaction)
1071                 ret = info->running_transaction->blocked;
1072         spin_unlock(&info->trans_lock);
1073         return ret;
1074 }
1075
1076 /*
1077  * wait for the current transaction commit to start and block subsequent
1078  * transaction joins
1079  */
1080 static void wait_current_trans_commit_start(struct btrfs_root *root,
1081                                             struct btrfs_transaction *trans)
1082 {
1083         DEFINE_WAIT(wait);
1084
1085         if (trans->in_commit)
1086                 return;
1087
1088         while (1) {
1089                 prepare_to_wait(&root->fs_info->transaction_blocked_wait, &wait,
1090                                 TASK_UNINTERRUPTIBLE);
1091                 if (trans->in_commit) {
1092                         finish_wait(&root->fs_info->transaction_blocked_wait,
1093                                     &wait);
1094                         break;
1095                 }
1096                 schedule();
1097                 finish_wait(&root->fs_info->transaction_blocked_wait, &wait);
1098         }
1099 }
1100
1101 /*
1102  * wait for the current transaction to start and then become unblocked.
1103  * caller holds ref.
1104  */
1105 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1106                                          struct btrfs_transaction *trans)
1107 {
1108         DEFINE_WAIT(wait);
1109
1110         if (trans->commit_done || (trans->in_commit && !trans->blocked))
1111                 return;
1112
1113         while (1) {
1114                 prepare_to_wait(&root->fs_info->transaction_wait, &wait,
1115                                 TASK_UNINTERRUPTIBLE);
1116                 if (trans->commit_done ||
1117                     (trans->in_commit && !trans->blocked)) {
1118                         finish_wait(&root->fs_info->transaction_wait,
1119                                     &wait);
1120                         break;
1121                 }
1122                 schedule();
1123                 finish_wait(&root->fs_info->transaction_wait,
1124                             &wait);
1125         }
1126 }
1127
1128 /*
1129  * commit transactions asynchronously. once btrfs_commit_transaction_async
1130  * returns, any subsequent transaction will not be allowed to join.
1131  */
1132 struct btrfs_async_commit {
1133         struct btrfs_trans_handle *newtrans;
1134         struct btrfs_root *root;
1135         struct delayed_work work;
1136 };
1137
1138 static void do_async_commit(struct work_struct *work)
1139 {
1140         struct btrfs_async_commit *ac =
1141                 container_of(work, struct btrfs_async_commit, work.work);
1142
1143         btrfs_commit_transaction(ac->newtrans, ac->root);
1144         kfree(ac);
1145 }
1146
1147 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1148                                    struct btrfs_root *root,
1149                                    int wait_for_unblock)
1150 {
1151         struct btrfs_async_commit *ac;
1152         struct btrfs_transaction *cur_trans;
1153
1154         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1155         if (!ac)
1156                 return -ENOMEM;
1157
1158         INIT_DELAYED_WORK(&ac->work, do_async_commit);
1159         ac->root = root;
1160         ac->newtrans = btrfs_join_transaction(root);
1161         if (IS_ERR(ac->newtrans)) {
1162                 int err = PTR_ERR(ac->newtrans);
1163                 kfree(ac);
1164                 return err;
1165         }
1166
1167         /* take transaction reference */
1168         cur_trans = trans->transaction;
1169         atomic_inc(&cur_trans->use_count);
1170
1171         btrfs_end_transaction(trans, root);
1172         schedule_delayed_work(&ac->work, 0);
1173
1174         /* wait for transaction to start and unblock */
1175         if (wait_for_unblock)
1176                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1177         else
1178                 wait_current_trans_commit_start(root, cur_trans);
1179
1180         if (current->journal_info == trans)
1181                 current->journal_info = NULL;
1182
1183         put_transaction(cur_trans);
1184         return 0;
1185 }
1186
1187 /*
1188  * btrfs_transaction state sequence:
1189  *    in_commit = 0, blocked = 0  (initial)
1190  *    in_commit = 1, blocked = 1
1191  *    blocked = 0
1192  *    commit_done = 1
1193  */
1194 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1195                              struct btrfs_root *root)
1196 {
1197         unsigned long joined = 0;
1198         struct btrfs_transaction *cur_trans;
1199         struct btrfs_transaction *prev_trans = NULL;
1200         DEFINE_WAIT(wait);
1201         int ret;
1202         int should_grow = 0;
1203         unsigned long now = get_seconds();
1204         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1205
1206         btrfs_run_ordered_operations(root, 0);
1207
1208         /* make a pass through all the delayed refs we have so far
1209          * any runnings procs may add more while we are here
1210          */
1211         ret = btrfs_run_delayed_refs(trans, root, 0);
1212         BUG_ON(ret);
1213
1214         btrfs_trans_release_metadata(trans, root);
1215
1216         cur_trans = trans->transaction;
1217         /*
1218          * set the flushing flag so procs in this transaction have to
1219          * start sending their work down.
1220          */
1221         cur_trans->delayed_refs.flushing = 1;
1222
1223         ret = btrfs_run_delayed_refs(trans, root, 0);
1224         BUG_ON(ret);
1225
1226         spin_lock(&cur_trans->commit_lock);
1227         if (cur_trans->in_commit) {
1228                 spin_unlock(&cur_trans->commit_lock);
1229                 atomic_inc(&cur_trans->use_count);
1230                 btrfs_end_transaction(trans, root);
1231
1232                 ret = wait_for_commit(root, cur_trans);
1233                 BUG_ON(ret);
1234
1235                 put_transaction(cur_trans);
1236
1237                 return 0;
1238         }
1239
1240         trans->transaction->in_commit = 1;
1241         trans->transaction->blocked = 1;
1242         spin_unlock(&cur_trans->commit_lock);
1243         wake_up(&root->fs_info->transaction_blocked_wait);
1244
1245         spin_lock(&root->fs_info->trans_lock);
1246         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1247                 prev_trans = list_entry(cur_trans->list.prev,
1248                                         struct btrfs_transaction, list);
1249                 if (!prev_trans->commit_done) {
1250                         atomic_inc(&prev_trans->use_count);
1251                         spin_unlock(&root->fs_info->trans_lock);
1252
1253                         wait_for_commit(root, prev_trans);
1254
1255                         put_transaction(prev_trans);
1256                 } else {
1257                         spin_unlock(&root->fs_info->trans_lock);
1258                 }
1259         } else {
1260                 spin_unlock(&root->fs_info->trans_lock);
1261         }
1262
1263         if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
1264                 should_grow = 1;
1265
1266         do {
1267                 int snap_pending = 0;
1268
1269                 joined = cur_trans->num_joined;
1270                 if (!list_empty(&trans->transaction->pending_snapshots))
1271                         snap_pending = 1;
1272
1273                 WARN_ON(cur_trans != trans->transaction);
1274
1275                 if (flush_on_commit || snap_pending) {
1276                         btrfs_start_delalloc_inodes(root, 1);
1277                         ret = btrfs_wait_ordered_extents(root, 0, 1);
1278                         BUG_ON(ret);
1279                 }
1280
1281                 ret = btrfs_run_delayed_items(trans, root);
1282                 BUG_ON(ret);
1283
1284                 /*
1285                  * rename don't use btrfs_join_transaction, so, once we
1286                  * set the transaction to blocked above, we aren't going
1287                  * to get any new ordered operations.  We can safely run
1288                  * it here and no for sure that nothing new will be added
1289                  * to the list
1290                  */
1291                 btrfs_run_ordered_operations(root, 1);
1292
1293                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1294                                 TASK_UNINTERRUPTIBLE);
1295
1296                 if (atomic_read(&cur_trans->num_writers) > 1)
1297                         schedule_timeout(MAX_SCHEDULE_TIMEOUT);
1298                 else if (should_grow)
1299                         schedule_timeout(1);
1300
1301                 finish_wait(&cur_trans->writer_wait, &wait);
1302         } while (atomic_read(&cur_trans->num_writers) > 1 ||
1303                  (should_grow && cur_trans->num_joined != joined));
1304
1305         /*
1306          * Ok now we need to make sure to block out any other joins while we
1307          * commit the transaction.  We could have started a join before setting
1308          * no_join so make sure to wait for num_writers to == 1 again.
1309          */
1310         spin_lock(&root->fs_info->trans_lock);
1311         root->fs_info->trans_no_join = 1;
1312         spin_unlock(&root->fs_info->trans_lock);
1313         wait_event(cur_trans->writer_wait,
1314                    atomic_read(&cur_trans->num_writers) == 1);
1315
1316         /*
1317          * the reloc mutex makes sure that we stop
1318          * the balancing code from coming in and moving
1319          * extents around in the middle of the commit
1320          */
1321         mutex_lock(&root->fs_info->reloc_mutex);
1322
1323         ret = btrfs_run_delayed_items(trans, root);
1324         BUG_ON(ret);
1325
1326         ret = create_pending_snapshots(trans, root->fs_info);
1327         BUG_ON(ret);
1328
1329         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1330         BUG_ON(ret);
1331
1332         /*
1333          * make sure none of the code above managed to slip in a
1334          * delayed item
1335          */
1336         btrfs_assert_delayed_root_empty(root);
1337
1338         WARN_ON(cur_trans != trans->transaction);
1339
1340         btrfs_scrub_pause(root);
1341         /* btrfs_commit_tree_roots is responsible for getting the
1342          * various roots consistent with each other.  Every pointer
1343          * in the tree of tree roots has to point to the most up to date
1344          * root for every subvolume and other tree.  So, we have to keep
1345          * the tree logging code from jumping in and changing any
1346          * of the trees.
1347          *
1348          * At this point in the commit, there can't be any tree-log
1349          * writers, but a little lower down we drop the trans mutex
1350          * and let new people in.  By holding the tree_log_mutex
1351          * from now until after the super is written, we avoid races
1352          * with the tree-log code.
1353          */
1354         mutex_lock(&root->fs_info->tree_log_mutex);
1355
1356         ret = commit_fs_roots(trans, root);
1357         BUG_ON(ret);
1358
1359         /* commit_fs_roots gets rid of all the tree log roots, it is now
1360          * safe to free the root of tree log roots
1361          */
1362         btrfs_free_log_root_tree(trans, root->fs_info);
1363
1364         ret = commit_cowonly_roots(trans, root);
1365         BUG_ON(ret);
1366
1367         btrfs_prepare_extent_commit(trans, root);
1368
1369         cur_trans = root->fs_info->running_transaction;
1370
1371         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1372                             root->fs_info->tree_root->node);
1373         switch_commit_root(root->fs_info->tree_root);
1374
1375         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1376                             root->fs_info->chunk_root->node);
1377         switch_commit_root(root->fs_info->chunk_root);
1378
1379         update_super_roots(root);
1380
1381         if (!root->fs_info->log_root_recovering) {
1382                 btrfs_set_super_log_root(&root->fs_info->super_copy, 0);
1383                 btrfs_set_super_log_root_level(&root->fs_info->super_copy, 0);
1384         }
1385
1386         memcpy(&root->fs_info->super_for_commit, &root->fs_info->super_copy,
1387                sizeof(root->fs_info->super_copy));
1388
1389         trans->transaction->blocked = 0;
1390         spin_lock(&root->fs_info->trans_lock);
1391         root->fs_info->running_transaction = NULL;
1392         root->fs_info->trans_no_join = 0;
1393         spin_unlock(&root->fs_info->trans_lock);
1394         mutex_unlock(&root->fs_info->reloc_mutex);
1395
1396         wake_up(&root->fs_info->transaction_wait);
1397
1398         ret = btrfs_write_and_wait_transaction(trans, root);
1399         BUG_ON(ret);
1400         write_ctree_super(trans, root, 0);
1401
1402         /*
1403          * the super is written, we can safely allow the tree-loggers
1404          * to go about their business
1405          */
1406         mutex_unlock(&root->fs_info->tree_log_mutex);
1407
1408         btrfs_finish_extent_commit(trans, root);
1409
1410         cur_trans->commit_done = 1;
1411
1412         root->fs_info->last_trans_committed = cur_trans->transid;
1413
1414         wake_up(&cur_trans->commit_wait);
1415
1416         spin_lock(&root->fs_info->trans_lock);
1417         list_del_init(&cur_trans->list);
1418         spin_unlock(&root->fs_info->trans_lock);
1419
1420         put_transaction(cur_trans);
1421         put_transaction(cur_trans);
1422
1423         trace_btrfs_transaction_commit(root);
1424
1425         btrfs_scrub_continue(root);
1426
1427         if (current->journal_info == trans)
1428                 current->journal_info = NULL;
1429
1430         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1431
1432         if (current != root->fs_info->transaction_kthread)
1433                 btrfs_run_delayed_iputs(root);
1434
1435         return ret;
1436 }
1437
1438 /*
1439  * interface function to delete all the snapshots we have scheduled for deletion
1440  */
1441 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1442 {
1443         LIST_HEAD(list);
1444         struct btrfs_fs_info *fs_info = root->fs_info;
1445
1446         spin_lock(&fs_info->trans_lock);
1447         list_splice_init(&fs_info->dead_roots, &list);
1448         spin_unlock(&fs_info->trans_lock);
1449
1450         while (!list_empty(&list)) {
1451                 root = list_entry(list.next, struct btrfs_root, root_list);
1452                 list_del(&root->root_list);
1453
1454                 btrfs_kill_all_delayed_nodes(root);
1455
1456                 if (btrfs_header_backref_rev(root->node) <
1457                     BTRFS_MIXED_BACKREF_REV)
1458                         btrfs_drop_snapshot(root, NULL, 0);
1459                 else
1460                         btrfs_drop_snapshot(root, NULL, 1);
1461         }
1462         return 0;
1463 }