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
  31 #define BTRFS_ROOT_TRANS_TAG 0
  32
  33 static noinline void put_transaction(struct btrfs_transaction *transaction)
  34 {
  35         WARN_ON(transaction->use_count == 0);
  36         transaction->use_count--;
  37         if (transaction->use_count == 0) {
  38                 list_del_init(&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)
  54 {
  55         struct btrfs_transaction *cur_trans;
  56         cur_trans = root->fs_info->running_transaction;
  57         if (!cur_trans) {
  58                 cur_trans = kmem_cache_alloc(btrfs_transaction_cachep,
  59                                              GFP_NOFS);
  60                 if (!cur_trans)
  61                         return -ENOMEM;
  62                 root->fs_info->generation++;
  63                 cur_trans->num_writers = 1;
  64                 cur_trans->num_joined = 0;
  65                 cur_trans->transid = root->fs_info->generation;
  66                 init_waitqueue_head(&cur_trans->writer_wait);
  67                 init_waitqueue_head(&cur_trans->commit_wait);
  68                 cur_trans->in_commit = 0;
  69                 cur_trans->blocked = 0;
  70                 cur_trans->use_count = 1;
  71                 cur_trans->commit_done = 0;
  72                 cur_trans->start_time = get_seconds();
  73
  74                 cur_trans->delayed_refs.root = RB_ROOT;
  75                 cur_trans->delayed_refs.num_entries = 0;
  76                 cur_trans->delayed_refs.num_heads_ready = 0;
  77                 cur_trans->delayed_refs.num_heads = 0;
  78                 cur_trans->delayed_refs.flushing = 0;
  79                 cur_trans->delayed_refs.run_delayed_start = 0;
  80                 spin_lock_init(&cur_trans->delayed_refs.lock);
  81
  82                 INIT_LIST_HEAD(&cur_trans->pending_snapshots);
  83                 list_add_tail(&cur_trans->list, &root->fs_info->trans_list);
  84                 extent_io_tree_init(&cur_trans->dirty_pages,
  85                                      root->fs_info->btree_inode->i_mapping,
  86                                      GFP_NOFS);
  87                 spin_lock(&root->fs_info->new_trans_lock);
  88                 root->fs_info->running_transaction = cur_trans;
  89                 spin_unlock(&root->fs_info->new_trans_lock);
  90         } else {
  91                 cur_trans->num_writers++;
  92                 cur_trans->num_joined++;
  93         }
  94
  95         return 0;
  96 }
  97
  98 /*
  99  * this does all the record keeping required to make sure that a reference
 100  * counted root is properly recorded in a given transaction.  This is required
 101  * to make sure the old root from before we joined the transaction is deleted
 102  * when the transaction commits
 103  */
 104 static noinline int record_root_in_trans(struct btrfs_trans_handle *trans,
 105                                          struct btrfs_root *root)
 106 {
 107         if (root->ref_cows && root->last_trans < trans->transid) {
 108                 WARN_ON(root == root->fs_info->extent_root);
 109                 WARN_ON(root->commit_root != root->node);
 110
 111                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 112                            (unsigned long)root->root_key.objectid,
 113                            BTRFS_ROOT_TRANS_TAG);
 114                 root->last_trans = trans->transid;
 115                 btrfs_init_reloc_root(trans, root);
 116         }
 117         return 0;
 118 }
 119
 120 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 121                                struct btrfs_root *root)
 122 {
 123         if (!root->ref_cows)
 124                 return 0;
 125
 126         mutex_lock(&root->fs_info->trans_mutex);
 127         if (root->last_trans == trans->transid) {
 128                 mutex_unlock(&root->fs_info->trans_mutex);
 129                 return 0;
 130         }
 131
 132         record_root_in_trans(trans, root);
 133         mutex_unlock(&root->fs_info->trans_mutex);
 134         return 0;
 135 }
 136
 137 /* wait for commit against the current transaction to become unblocked
 138  * when this is done, it is safe to start a new transaction, but the current
 139  * transaction might not be fully on disk.
 140  */
 141 static void wait_current_trans(struct btrfs_root *root)
 142 {
 143         struct btrfs_transaction *cur_trans;
 144
 145         cur_trans = root->fs_info->running_transaction;
 146         if (cur_trans && cur_trans->blocked) {
 147                 DEFINE_WAIT(wait);
 148                 cur_trans->use_count++;
 149                 while (1) {
 150                         prepare_to_wait(&root->fs_info->transaction_wait, &wait,
 151                                         TASK_UNINTERRUPTIBLE);
 152                         if (!cur_trans->blocked)
 153                                 break;
 154                         mutex_unlock(&root->fs_info->trans_mutex);
 155                         schedule();
 156                         mutex_lock(&root->fs_info->trans_mutex);
 157                 }
 158                 finish_wait(&root->fs_info->transaction_wait, &wait);
 159                 put_transaction(cur_trans);
 160         }
 161 }
 162
 163 enum btrfs_trans_type {
 164         TRANS_START,
 165         TRANS_JOIN,
 166         TRANS_USERSPACE,
 167         TRANS_JOIN_NOLOCK,
 168 };
 169
 170 static int may_wait_transaction(struct btrfs_root *root, int type)
 171 {
 172         if (!root->fs_info->log_root_recovering &&
 173             ((type == TRANS_START && !root->fs_info->open_ioctl_trans) ||
 174              type == TRANS_USERSPACE))
 175                 return 1;
 176         return 0;
 177 }
 178
 179 static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
 180                                                     u64 num_items, int type)
 181 {
 182         struct btrfs_trans_handle *h;
 183         struct btrfs_transaction *cur_trans;
 184         int retries = 0;
 185         int ret;
 186
 187         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR)
 188                 return ERR_PTR(-EROFS);
 189 again:
 190         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 191         if (!h)
 192                 return ERR_PTR(-ENOMEM);
 193
 194         if (type != TRANS_JOIN_NOLOCK)
 195                 mutex_lock(&root->fs_info->trans_mutex);
 196         if (may_wait_transaction(root, type))
 197                 wait_current_trans(root);
 198
 199         ret = join_transaction(root);
 200         if (ret < 0) {
 201                 kmem_cache_free(btrfs_trans_handle_cachep, h);
 202                 if (type != TRANS_JOIN_NOLOCK)
 203                         mutex_unlock(&root->fs_info->trans_mutex);
 204                 return ERR_PTR(ret);
 205         }
 206
 207         cur_trans = root->fs_info->running_transaction;
 208         cur_trans->use_count++;
 209         if (type != TRANS_JOIN_NOLOCK)
 210                 mutex_unlock(&root->fs_info->trans_mutex);
 211
 212         h->transid = cur_trans->transid;
 213         h->transaction = cur_trans;
 214         h->blocks_used = 0;
 215         h->block_group = 0;
 216         h->bytes_reserved = 0;
 217         h->delayed_ref_updates = 0;
 218         h->block_rsv = NULL;
 219
 220         smp_mb();
 221         if (cur_trans->blocked && may_wait_transaction(root, type)) {
 222                 btrfs_commit_transaction(h, root);
 223                 goto again;
 224         }
 225
 226         if (num_items > 0) {
 227                 ret = btrfs_trans_reserve_metadata(h, root, num_items);
 228                 if (ret == -EAGAIN && !retries) {
 229                         retries++;
 230                         btrfs_commit_transaction(h, root);
 231                         goto again;
 232                 } else if (ret == -EAGAIN) {
 233                         /*
 234                          * We have already retried and got EAGAIN, so really we
 235                          * don't have space, so set ret to -ENOSPC.
 236                          */
 237                         ret = -ENOSPC;
 238                 }
 239
 240                 if (ret < 0) {
 241                         btrfs_end_transaction(h, root);
 242                         return ERR_PTR(ret);
 243                 }
 244         }
 245
 246         if (type != TRANS_JOIN_NOLOCK)
 247                 mutex_lock(&root->fs_info->trans_mutex);
 248         record_root_in_trans(h, root);
 249         if (type != TRANS_JOIN_NOLOCK)
 250                 mutex_unlock(&root->fs_info->trans_mutex);
 251
 252         if (!current->journal_info && type != TRANS_USERSPACE)
 253                 current->journal_info = h;
 254         return h;
 255 }
 256
 257 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 258                                                    int num_items)
 259 {
 260         return start_transaction(root, num_items, TRANS_START);
 261 }
 262 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root,
 263                                                    int num_blocks)
 264 {
 265         return start_transaction(root, 0, TRANS_JOIN);
 266 }
 267
 268 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root,
 269                                                           int num_blocks)
 270 {
 271         return start_transaction(root, 0, TRANS_JOIN_NOLOCK);
 272 }
 273
 274 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *r,
 275                                                          int num_blocks)
 276 {
 277         return start_transaction(r, 0, TRANS_USERSPACE);
 278 }
 279
 280 /* wait for a transaction commit to be fully complete */
 281 static noinline int wait_for_commit(struct btrfs_root *root,
 282                                     struct btrfs_transaction *commit)
 283 {
 284         DEFINE_WAIT(wait);
 285         mutex_lock(&root->fs_info->trans_mutex);
 286         while (!commit->commit_done) {
 287                 prepare_to_wait(&commit->commit_wait, &wait,
 288                                 TASK_UNINTERRUPTIBLE);
 289                 if (commit->commit_done)
 290                         break;
 291                 mutex_unlock(&root->fs_info->trans_mutex);
 292                 schedule();
 293                 mutex_lock(&root->fs_info->trans_mutex);
 294         }
 295         mutex_unlock(&root->fs_info->trans_mutex);
 296         finish_wait(&commit->commit_wait, &wait);
 297         return 0;
 298 }
 299
 300 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 301 {
 302         struct btrfs_transaction *cur_trans = NULL, *t;
 303         int ret;
 304
 305         mutex_lock(&root->fs_info->trans_mutex);
 306
 307         ret = 0;
 308         if (transid) {
 309                 if (transid <= root->fs_info->last_trans_committed)
 310                         goto out_unlock;
 311
 312                 /* find specified transaction */
 313                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 314                         if (t->transid == transid) {
 315                                 cur_trans = t;
 316                                 break;
 317                         }
 318                         if (t->transid > transid)
 319                                 break;
 320                 }
 321                 ret = -EINVAL;
 322                 if (!cur_trans)
 323                         goto out_unlock;  /* bad transid */
 324         } else {
 325                 /* find newest transaction that is committing | committed */
 326                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 327                                             list) {
 328                         if (t->in_commit) {
 329                                 if (t->commit_done)
 330                                         goto out_unlock;
 331                                 cur_trans = t;
 332                                 break;
 333                         }
 334                 }
 335                 if (!cur_trans)
 336                         goto out_unlock;  /* nothing committing|committed */
 337         }
 338
 339         cur_trans->use_count++;
 340         mutex_unlock(&root->fs_info->trans_mutex);
 341
 342         wait_for_commit(root, cur_trans);
 343
 344         mutex_lock(&root->fs_info->trans_mutex);
 345         put_transaction(cur_trans);
 346         ret = 0;
 347 out_unlock:
 348         mutex_unlock(&root->fs_info->trans_mutex);
 349         return ret;
 350 }
 351
 352 #if 0
 353 /*
 354  * rate limit against the drop_snapshot code.  This helps to slow down new
 355  * operations if the drop_snapshot code isn't able to keep up.
 356  */
 357 static void throttle_on_drops(struct btrfs_root *root)
 358 {
 359         struct btrfs_fs_info *info = root->fs_info;
 360         int harder_count = 0;
 361
 362 harder:
 363         if (atomic_read(&info->throttles)) {
 364                 DEFINE_WAIT(wait);
 365                 int thr;
 366                 thr = atomic_read(&info->throttle_gen);
 367
 368                 do {
 369                         prepare_to_wait(&info->transaction_throttle,
 370                                         &wait, TASK_UNINTERRUPTIBLE);
 371                         if (!atomic_read(&info->throttles)) {
 372                                 finish_wait(&info->transaction_throttle, &wait);
 373                                 break;
 374                         }
 375                         schedule();
 376                         finish_wait(&info->transaction_throttle, &wait);
 377                 } while (thr == atomic_read(&info->throttle_gen));
 378                 harder_count++;
 379
 380                 if (root->fs_info->total_ref_cache_size > 1 * 1024 * 1024 &&
 381                     harder_count < 2)
 382                         goto harder;
 383
 384                 if (root->fs_info->total_ref_cache_size > 5 * 1024 * 1024 &&
 385                     harder_count < 10)
 386                         goto harder;
 387
 388                 if (root->fs_info->total_ref_cache_size > 10 * 1024 * 1024 &&
 389                     harder_count < 20)
 390                         goto harder;
 391         }
 392 }
 393 #endif
 394
 395 void btrfs_throttle(struct btrfs_root *root)
 396 {
 397         mutex_lock(&root->fs_info->trans_mutex);
 398         if (!root->fs_info->open_ioctl_trans)
 399                 wait_current_trans(root);
 400         mutex_unlock(&root->fs_info->trans_mutex);
 401 }
 402
 403 static int should_end_transaction(struct btrfs_trans_handle *trans,
 404                                   struct btrfs_root *root)
 405 {
 406         int ret;
 407         ret = btrfs_block_rsv_check(trans, root,
 408                                     &root->fs_info->global_block_rsv, 0, 5);
 409         return ret ? 1 : 0;
 410 }
 411
 412 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 413                                  struct btrfs_root *root)
 414 {
 415         struct btrfs_transaction *cur_trans = trans->transaction;
 416         int updates;
 417
 418         if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
 419                 return 1;
 420
 421         updates = trans->delayed_ref_updates;
 422         trans->delayed_ref_updates = 0;
 423         if (updates)
 424                 btrfs_run_delayed_refs(trans, root, updates);
 425
 426         return should_end_transaction(trans, root);
 427 }
 428
 429 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 430                           struct btrfs_root *root, int throttle, int lock)
 431 {
 432         struct btrfs_transaction *cur_trans = trans->transaction;
 433         struct btrfs_fs_info *info = root->fs_info;
 434         int count = 0;
 435
 436         while (count < 4) {
 437                 unsigned long cur = trans->delayed_ref_updates;
 438                 trans->delayed_ref_updates = 0;
 439                 if (cur &&
 440                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 441                         trans->delayed_ref_updates = 0;
 442
 443                         /*
 444                          * do a full flush if the transaction is trying
 445                          * to close
 446                          */
 447                         if (trans->transaction->delayed_refs.flushing)
 448                                 cur = 0;
 449                         btrfs_run_delayed_refs(trans, root, cur);
 450                 } else {
 451                         break;
 452                 }
 453                 count++;
 454         }
 455
 456         btrfs_trans_release_metadata(trans, root);
 457
 458         if (lock && !root->fs_info->open_ioctl_trans &&
 459             should_end_transaction(trans, root))
 460                 trans->transaction->blocked = 1;
 461
 462         if (lock && cur_trans->blocked && !cur_trans->in_commit) {
 463                 if (throttle)
 464                         return btrfs_commit_transaction(trans, root);
 465                 else
 466                         wake_up_process(info->transaction_kthread);
 467         }
 468
 469         if (lock)
 470                 mutex_lock(&info->trans_mutex);
 471         WARN_ON(cur_trans != info->running_transaction);
 472         WARN_ON(cur_trans->num_writers < 1);
 473         cur_trans->num_writers--;
 474
 475         smp_mb();
 476         if (waitqueue_active(&cur_trans->writer_wait))
 477                 wake_up(&cur_trans->writer_wait);
 478         put_transaction(cur_trans);
 479         if (lock)
 480                 mutex_unlock(&info->trans_mutex);
 481
 482         if (current->journal_info == trans)
 483                 current->journal_info = NULL;
 484         memset(trans, 0, sizeof(*trans));
 485         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 486
 487         if (throttle)
 488                 btrfs_run_delayed_iputs(root);
 489
 490         return 0;
 491 }
 492
 493 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 494                           struct btrfs_root *root)
 495 {
 496         return __btrfs_end_transaction(trans, root, 0, 1);
 497 }
 498
 499 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 500                                    struct btrfs_root *root)
 501 {
 502         return __btrfs_end_transaction(trans, root, 1, 1);
 503 }
 504
 505 int btrfs_end_transaction_nolock(struct btrfs_trans_handle *trans,
 506                                  struct btrfs_root *root)
 507 {
 508         return __btrfs_end_transaction(trans, root, 0, 0);
 509 }
 510
 511 /*
 512  * when btree blocks are allocated, they have some corresponding bits set for
 513  * them in one of two extent_io trees.  This is used to make sure all of
 514  * those extents are sent to disk but does not wait on them
 515  */
 516 int btrfs_write_marked_extents(struct btrfs_root *root,
 517                                struct extent_io_tree *dirty_pages, int mark)
 518 {
 519         int ret;
 520         int err = 0;
 521         int werr = 0;
 522         struct page *page;
 523         struct inode *btree_inode = root->fs_info->btree_inode;
 524         u64 start = 0;
 525         u64 end;
 526         unsigned long index;
 527
 528         while (1) {
 529                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 530                                             mark);
 531                 if (ret)
 532                         break;
 533                 while (start <= end) {
 534                         cond_resched();
 535
 536                         index = start >> PAGE_CACHE_SHIFT;
 537                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 538                         page = find_get_page(btree_inode->i_mapping, index);
 539                         if (!page)
 540                                 continue;
 541
 542                         btree_lock_page_hook(page);
 543                         if (!page->mapping) {
 544                                 unlock_page(page);
 545                                 page_cache_release(page);
 546                                 continue;
 547                         }
 548
 549                         if (PageWriteback(page)) {
 550                                 if (PageDirty(page))
 551                                         wait_on_page_writeback(page);
 552                                 else {
 553                                         unlock_page(page);
 554                                         page_cache_release(page);
 555                                         continue;
 556                                 }
 557                         }
 558                         err = write_one_page(page, 0);
 559                         if (err)
 560                                 werr = err;
 561                         page_cache_release(page);
 562                 }
 563         }
 564         if (err)
 565                 werr = err;
 566         return werr;
 567 }
 568
 569 /*
 570  * when btree blocks are allocated, they have some corresponding bits set for
 571  * them in one of two extent_io trees.  This is used to make sure all of
 572  * those extents are on disk for transaction or log commit.  We wait
 573  * on all the pages and clear them from the dirty pages state tree
 574  */
 575 int btrfs_wait_marked_extents(struct btrfs_root *root,
 576                               struct extent_io_tree *dirty_pages, int mark)
 577 {
 578         int ret;
 579         int err = 0;
 580         int werr = 0;
 581         struct page *page;
 582         struct inode *btree_inode = root->fs_info->btree_inode;
 583         u64 start = 0;
 584         u64 end;
 585         unsigned long index;
 586
 587         while (1) {
 588                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 589                                             mark);
 590                 if (ret)
 591                         break;
 592
 593                 clear_extent_bits(dirty_pages, start, end, mark, GFP_NOFS);
 594                 while (start <= end) {
 595                         index = start >> PAGE_CACHE_SHIFT;
 596                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 597                         page = find_get_page(btree_inode->i_mapping, index);
 598                         if (!page)
 599                                 continue;
 600                         if (PageDirty(page)) {
 601                                 btree_lock_page_hook(page);
 602                                 wait_on_page_writeback(page);
 603                                 err = write_one_page(page, 0);
 604                                 if (err)
 605                                         werr = err;
 606                         }
 607                         wait_on_page_writeback(page);
 608                         page_cache_release(page);
 609                         cond_resched();
 610                 }
 611         }
 612         if (err)
 613                 werr = err;
 614         return werr;
 615 }
 616
 617 /*
 618  * when btree blocks are allocated, they have some corresponding bits set for
 619  * them in one of two extent_io trees.  This is used to make sure all of
 620  * those extents are on disk for transaction or log commit
 621  */
 622 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 623                                 struct extent_io_tree *dirty_pages, int mark)
 624 {
 625         int ret;
 626         int ret2;
 627
 628         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 629         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 630         return ret || ret2;
 631 }
 632
 633 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 634                                      struct btrfs_root *root)
 635 {
 636         if (!trans || !trans->transaction) {
 637                 struct inode *btree_inode;
 638                 btree_inode = root->fs_info->btree_inode;
 639                 return filemap_write_and_wait(btree_inode->i_mapping);
 640         }
 641         return btrfs_write_and_wait_marked_extents(root,
 642                                            &trans->transaction->dirty_pages,
 643                                            EXTENT_DIRTY);
 644 }
 645
 646 /*
 647  * this is used to update the root pointer in the tree of tree roots.
 648  *
 649  * But, in the case of the extent allocation tree, updating the root
 650  * pointer may allocate blocks which may change the root of the extent
 651  * allocation tree.
 652  *
 653  * So, this loops and repeats and makes sure the cowonly root didn't
 654  * change while the root pointer was being updated in the metadata.
 655  */
 656 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 657                                struct btrfs_root *root)
 658 {
 659         int ret;
 660         u64 old_root_bytenr;
 661         u64 old_root_used;
 662         struct btrfs_root *tree_root = root->fs_info->tree_root;
 663
 664         old_root_used = btrfs_root_used(&root->root_item);
 665         btrfs_write_dirty_block_groups(trans, root);
 666
 667         while (1) {
 668                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 669                 if (old_root_bytenr == root->node->start &&
 670                     old_root_used == btrfs_root_used(&root->root_item))
 671                         break;
 672
 673                 btrfs_set_root_node(&root->root_item, root->node);
 674                 ret = btrfs_update_root(trans, tree_root,
 675                                         &root->root_key,
 676                                         &root->root_item);
 677                 BUG_ON(ret);
 678
 679                 old_root_used = btrfs_root_used(&root->root_item);
 680                 ret = btrfs_write_dirty_block_groups(trans, root);
 681                 BUG_ON(ret);
 682         }
 683
 684         if (root != root->fs_info->extent_root)
 685                 switch_commit_root(root);
 686
 687         return 0;
 688 }
 689
 690 /*
 691  * update all the cowonly tree roots on disk
 692  */
 693 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 694                                          struct btrfs_root *root)
 695 {
 696         struct btrfs_fs_info *fs_info = root->fs_info;
 697         struct list_head *next;
 698         struct extent_buffer *eb;
 699         int ret;
 700
 701         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 702         BUG_ON(ret);
 703
 704         eb = btrfs_lock_root_node(fs_info->tree_root);
 705         btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
 706         btrfs_tree_unlock(eb);
 707         free_extent_buffer(eb);
 708
 709         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 710         BUG_ON(ret);
 711
 712         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 713                 next = fs_info->dirty_cowonly_roots.next;
 714                 list_del_init(next);
 715                 root = list_entry(next, struct btrfs_root, dirty_list);
 716
 717                 update_cowonly_root(trans, root);
 718         }
 719
 720         down_write(&fs_info->extent_commit_sem);
 721         switch_commit_root(fs_info->extent_root);
 722         up_write(&fs_info->extent_commit_sem);
 723
 724         return 0;
 725 }
 726
 727 /*
 728  * dead roots are old snapshots that need to be deleted.  This allocates
 729  * a dirty root struct and adds it into the list of dead roots that need to
 730  * be deleted
 731  */
 732 int btrfs_add_dead_root(struct btrfs_root *root)
 733 {
 734         mutex_lock(&root->fs_info->trans_mutex);
 735         list_add(&root->root_list, &root->fs_info->dead_roots);
 736         mutex_unlock(&root->fs_info->trans_mutex);
 737         return 0;
 738 }
 739
 740 /*
 741  * update all the cowonly tree roots on disk
 742  */
 743 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 744                                     struct btrfs_root *root)
 745 {
 746         struct btrfs_root *gang[8];
 747         struct btrfs_fs_info *fs_info = root->fs_info;
 748         int i;
 749         int ret;
 750         int err = 0;
 751
 752         while (1) {
 753                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 754                                                  (void **)gang, 0,
 755                                                  ARRAY_SIZE(gang),
 756                                                  BTRFS_ROOT_TRANS_TAG);
 757                 if (ret == 0)
 758                         break;
 759                 for (i = 0; i < ret; i++) {
 760                         root = gang[i];
 761                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 762                                         (unsigned long)root->root_key.objectid,
 763                                         BTRFS_ROOT_TRANS_TAG);
 764
 765                         btrfs_free_log(trans, root);
 766                         btrfs_update_reloc_root(trans, root);
 767                         btrfs_orphan_commit_root(trans, root);
 768
 769                         if (root->commit_root != root->node) {
 770                                 switch_commit_root(root);
 771                                 btrfs_set_root_node(&root->root_item,
 772                                                     root->node);
 773                         }
 774
 775                         err = btrfs_update_root(trans, fs_info->tree_root,
 776                                                 &root->root_key,
 777                                                 &root->root_item);
 778                         if (err)
 779                                 break;
 780                 }
 781         }
 782         return err;
 783 }
 784
 785 /*
 786  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 787  * otherwise every leaf in the btree is read and defragged.
 788  */
 789 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 790 {
 791         struct btrfs_fs_info *info = root->fs_info;
 792         struct btrfs_trans_handle *trans;
 793         int ret;
 794         unsigned long nr;
 795
 796         if (xchg(&root->defrag_running, 1))
 797                 return 0;
 798
 799         while (1) {
 800                 trans = btrfs_start_transaction(root, 0);
 801                 if (IS_ERR(trans))
 802                         return PTR_ERR(trans);
 803
 804                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 805
 806                 nr = trans->blocks_used;
 807                 btrfs_end_transaction(trans, root);
 808                 btrfs_btree_balance_dirty(info->tree_root, nr);
 809                 cond_resched();
 810
 811                 if (root->fs_info->closing || ret != -EAGAIN)
 812                         break;
 813         }
 814         root->defrag_running = 0;
 815         return ret;
 816 }
 817
 818 #if 0
 819 /*
 820  * when dropping snapshots, we generate a ton of delayed refs, and it makes
 821  * sense not to join the transaction while it is trying to flush the current
 822  * queue of delayed refs out.
 823  *
 824  * This is used by the drop snapshot code only
 825  */
 826 static noinline int wait_transaction_pre_flush(struct btrfs_fs_info *info)
 827 {
 828         DEFINE_WAIT(wait);
 829
 830         mutex_lock(&info->trans_mutex);
 831         while (info->running_transaction &&
 832                info->running_transaction->delayed_refs.flushing) {
 833                 prepare_to_wait(&info->transaction_wait, &wait,
 834                                 TASK_UNINTERRUPTIBLE);
 835                 mutex_unlock(&info->trans_mutex);
 836
 837                 schedule();
 838
 839                 mutex_lock(&info->trans_mutex);
 840                 finish_wait(&info->transaction_wait, &wait);
 841         }
 842         mutex_unlock(&info->trans_mutex);
 843         return 0;
 844 }
 845
 846 /*
 847  * Given a list of roots that need to be deleted, call btrfs_drop_snapshot on
 848  * all of them
 849  */
 850 int btrfs_drop_dead_root(struct btrfs_root *root)
 851 {
 852         struct btrfs_trans_handle *trans;
 853         struct btrfs_root *tree_root = root->fs_info->tree_root;
 854         unsigned long nr;
 855         int ret;
 856
 857         while (1) {
 858                 /*
 859                  * we don't want to jump in and create a bunch of
 860                  * delayed refs if the transaction is starting to close
 861                  */
 862                 wait_transaction_pre_flush(tree_root->fs_info);
 863                 trans = btrfs_start_transaction(tree_root, 1);
 864
 865                 /*
 866                  * we've joined a transaction, make sure it isn't
 867                  * closing right now
 868                  */
 869                 if (trans->transaction->delayed_refs.flushing) {
 870                         btrfs_end_transaction(trans, tree_root);
 871                         continue;
 872                 }
 873
 874                 ret = btrfs_drop_snapshot(trans, root);
 875                 if (ret != -EAGAIN)
 876                         break;
 877
 878                 ret = btrfs_update_root(trans, tree_root,
 879                                         &root->root_key,
 880                                         &root->root_item);
 881                 if (ret)
 882                         break;
 883
 884                 nr = trans->blocks_used;
 885                 ret = btrfs_end_transaction(trans, tree_root);
 886                 BUG_ON(ret);
 887
 888                 btrfs_btree_balance_dirty(tree_root, nr);
 889                 cond_resched();
 890         }
 891         BUG_ON(ret);
 892
 893         ret = btrfs_del_root(trans, tree_root, &root->root_key);
 894         BUG_ON(ret);
 895
 896         nr = trans->blocks_used;
 897         ret = btrfs_end_transaction(trans, tree_root);
 898         BUG_ON(ret);
 899
 900         free_extent_buffer(root->node);
 901         free_extent_buffer(root->commit_root);
 902         kfree(root);
 903
 904         btrfs_btree_balance_dirty(tree_root, nr);
 905         return ret;
 906 }
 907 #endif
 908
 909 /*
 910  * new snapshots need to be created at a very specific time in the
 911  * transaction commit.  This does the actual creation
 912  */
 913 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 914                                    struct btrfs_fs_info *fs_info,
 915                                    struct btrfs_pending_snapshot *pending)
 916 {
 917         struct btrfs_key key;
 918         struct btrfs_root_item *new_root_item;
 919         struct btrfs_root *tree_root = fs_info->tree_root;
 920         struct btrfs_root *root = pending->root;
 921         struct btrfs_root *parent_root;
 922         struct inode *parent_inode;
 923         struct dentry *parent;
 924         struct dentry *dentry;
 925         struct extent_buffer *tmp;
 926         struct extent_buffer *old;
 927         int ret;
 928         u64 to_reserve = 0;
 929         u64 index = 0;
 930         u64 objectid;
 931         u64 root_flags;
 932
 933         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 934         if (!new_root_item) {
 935                 pending->error = -ENOMEM;
 936                 goto fail;
 937         }
 938
 939         ret = btrfs_find_free_objectid(trans, tree_root, 0, &objectid);
 940         if (ret) {
 941                 pending->error = ret;
 942                 goto fail;
 943         }
 944
 945         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
 946         btrfs_orphan_pre_snapshot(trans, pending, &to_reserve);
 947
 948         if (to_reserve > 0) {
 949                 ret = btrfs_block_rsv_add(trans, root, &pending->block_rsv,
 950                                           to_reserve);
 951                 if (ret) {
 952                         pending->error = ret;
 953                         goto fail;
 954                 }
 955         }
 956
 957         key.objectid = objectid;
 958         key.offset = (u64)-1;
 959         key.type = BTRFS_ROOT_ITEM_KEY;
 960
 961         trans->block_rsv = &pending->block_rsv;
 962
 963         dentry = pending->dentry;
 964         parent = dget_parent(dentry);
 965         parent_inode = parent->d_inode;
 966         parent_root = BTRFS_I(parent_inode)->root;
 967         record_root_in_trans(trans, parent_root);
 968
 969         /*
 970          * insert the directory item
 971          */
 972         ret = btrfs_set_inode_index(parent_inode, &index);
 973         BUG_ON(ret);
 974         ret = btrfs_insert_dir_item(trans, parent_root,
 975                                 dentry->d_name.name, dentry->d_name.len,
 976                                 parent_inode->i_ino, &key,
 977                                 BTRFS_FT_DIR, index);
 978         BUG_ON(ret);
 979
 980         btrfs_i_size_write(parent_inode, parent_inode->i_size +
 981                                          dentry->d_name.len * 2);
 982         ret = btrfs_update_inode(trans, parent_root, parent_inode);
 983         BUG_ON(ret);
 984
 985         record_root_in_trans(trans, root);
 986         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
 987         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
 988         btrfs_check_and_init_root_item(new_root_item);
 989
 990         root_flags = btrfs_root_flags(new_root_item);
 991         if (pending->readonly)
 992                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
 993         else
 994                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
 995         btrfs_set_root_flags(new_root_item, root_flags);
 996
 997         old = btrfs_lock_root_node(root);
 998         btrfs_cow_block(trans, root, old, NULL, 0, &old);
 999         btrfs_set_lock_blocking(old);
1000
1001         btrfs_copy_root(trans, root, old, &tmp, objectid);
1002         btrfs_tree_unlock(old);
1003         free_extent_buffer(old);
1004
1005         btrfs_set_root_node(new_root_item, tmp);
1006         /* record when the snapshot was created in key.offset */
1007         key.offset = trans->transid;
1008         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1009         btrfs_tree_unlock(tmp);
1010         free_extent_buffer(tmp);
1011         BUG_ON(ret);
1012
1013         /*
1014          * insert root back/forward references
1015          */
1016         ret = btrfs_add_root_ref(trans, tree_root, objectid,
1017                                  parent_root->root_key.objectid,
1018                                  parent_inode->i_ino, index,
1019                                  dentry->d_name.name, dentry->d_name.len);
1020         BUG_ON(ret);
1021         dput(parent);
1022
1023         key.offset = (u64)-1;
1024         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1025         BUG_ON(IS_ERR(pending->snap));
1026
1027         btrfs_reloc_post_snapshot(trans, pending);
1028         btrfs_orphan_post_snapshot(trans, pending);
1029 fail:
1030         kfree(new_root_item);
1031         btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
1032         return 0;
1033 }
1034
1035 /*
1036  * create all the snapshots we've scheduled for creation
1037  */
1038 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1039                                              struct btrfs_fs_info *fs_info)
1040 {
1041         struct btrfs_pending_snapshot *pending;
1042         struct list_head *head = &trans->transaction->pending_snapshots;
1043         int ret;
1044
1045         list_for_each_entry(pending, head, list) {
1046                 ret = create_pending_snapshot(trans, fs_info, pending);
1047                 BUG_ON(ret);
1048         }
1049         return 0;
1050 }
1051
1052 static void update_super_roots(struct btrfs_root *root)
1053 {
1054         struct btrfs_root_item *root_item;
1055         struct btrfs_super_block *super;
1056
1057         super = &root->fs_info->super_copy;
1058
1059         root_item = &root->fs_info->chunk_root->root_item;
1060         super->chunk_root = root_item->bytenr;
1061         super->chunk_root_generation = root_item->generation;
1062         super->chunk_root_level = root_item->level;
1063
1064         root_item = &root->fs_info->tree_root->root_item;
1065         super->root = root_item->bytenr;
1066         super->generation = root_item->generation;
1067         super->root_level = root_item->level;
1068         if (super->cache_generation != 0 || btrfs_test_opt(root, SPACE_CACHE))
1069                 super->cache_generation = root_item->generation;
1070 }
1071
1072 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1073 {
1074         int ret = 0;
1075         spin_lock(&info->new_trans_lock);
1076         if (info->running_transaction)
1077                 ret = info->running_transaction->in_commit;
1078         spin_unlock(&info->new_trans_lock);
1079         return ret;
1080 }
1081
1082 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1083 {
1084         int ret = 0;
1085         spin_lock(&info->new_trans_lock);
1086         if (info->running_transaction)
1087                 ret = info->running_transaction->blocked;
1088         spin_unlock(&info->new_trans_lock);
1089         return ret;
1090 }
1091
1092 /*
1093  * wait for the current transaction commit to start and block subsequent
1094  * transaction joins
1095  */
1096 static void wait_current_trans_commit_start(struct btrfs_root *root,
1097                                             struct btrfs_transaction *trans)
1098 {
1099         DEFINE_WAIT(wait);
1100
1101         if (trans->in_commit)
1102                 return;
1103
1104         while (1) {
1105                 prepare_to_wait(&root->fs_info->transaction_blocked_wait, &wait,
1106                                 TASK_UNINTERRUPTIBLE);
1107                 if (trans->in_commit) {
1108                         finish_wait(&root->fs_info->transaction_blocked_wait,
1109                                     &wait);
1110                         break;
1111                 }
1112                 mutex_unlock(&root->fs_info->trans_mutex);
1113                 schedule();
1114                 mutex_lock(&root->fs_info->trans_mutex);
1115                 finish_wait(&root->fs_info->transaction_blocked_wait, &wait);
1116         }
1117 }
1118
1119 /*
1120  * wait for the current transaction to start and then become unblocked.
1121  * caller holds ref.
1122  */
1123 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1124                                          struct btrfs_transaction *trans)
1125 {
1126         DEFINE_WAIT(wait);
1127
1128         if (trans->commit_done || (trans->in_commit && !trans->blocked))
1129                 return;
1130
1131         while (1) {
1132                 prepare_to_wait(&root->fs_info->transaction_wait, &wait,
1133                                 TASK_UNINTERRUPTIBLE);
1134                 if (trans->commit_done ||
1135                     (trans->in_commit && !trans->blocked)) {
1136                         finish_wait(&root->fs_info->transaction_wait,
1137                                     &wait);
1138                         break;
1139                 }
1140                 mutex_unlock(&root->fs_info->trans_mutex);
1141                 schedule();
1142                 mutex_lock(&root->fs_info->trans_mutex);
1143                 finish_wait(&root->fs_info->transaction_wait,
1144                             &wait);
1145         }
1146 }
1147
1148 /*
1149  * commit transactions asynchronously. once btrfs_commit_transaction_async
1150  * returns, any subsequent transaction will not be allowed to join.
1151  */
1152 struct btrfs_async_commit {
1153         struct btrfs_trans_handle *newtrans;
1154         struct btrfs_root *root;
1155         struct delayed_work work;
1156 };
1157
1158 static void do_async_commit(struct work_struct *work)
1159 {
1160         struct btrfs_async_commit *ac =
1161                 container_of(work, struct btrfs_async_commit, work.work);
1162
1163         btrfs_commit_transaction(ac->newtrans, ac->root);
1164         kfree(ac);
1165 }
1166
1167 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1168                                    struct btrfs_root *root,
1169                                    int wait_for_unblock)
1170 {
1171         struct btrfs_async_commit *ac;
1172         struct btrfs_transaction *cur_trans;
1173
1174         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1175         if (!ac)
1176                 return -ENOMEM;
1177
1178         INIT_DELAYED_WORK(&ac->work, do_async_commit);
1179         ac->root = root;
1180         ac->newtrans = btrfs_join_transaction(root, 0);
1181         if (IS_ERR(ac->newtrans)) {
1182                 int err = PTR_ERR(ac->newtrans);
1183                 kfree(ac);
1184                 return err;
1185         }
1186
1187         /* take transaction reference */
1188         mutex_lock(&root->fs_info->trans_mutex);
1189         cur_trans = trans->transaction;
1190         cur_trans->use_count++;
1191         mutex_unlock(&root->fs_info->trans_mutex);
1192
1193         btrfs_end_transaction(trans, root);
1194         schedule_delayed_work(&ac->work, 0);
1195
1196         /* wait for transaction to start and unblock */
1197         mutex_lock(&root->fs_info->trans_mutex);
1198         if (wait_for_unblock)
1199                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1200         else
1201                 wait_current_trans_commit_start(root, cur_trans);
1202         put_transaction(cur_trans);
1203         mutex_unlock(&root->fs_info->trans_mutex);
1204
1205         return 0;
1206 }
1207
1208 /*
1209  * btrfs_transaction state sequence:
1210  *    in_commit = 0, blocked = 0  (initial)
1211  *    in_commit = 1, blocked = 1
1212  *    blocked = 0
1213  *    commit_done = 1
1214  */
1215 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1216                              struct btrfs_root *root)
1217 {
1218         unsigned long joined = 0;
1219         struct btrfs_transaction *cur_trans;
1220         struct btrfs_transaction *prev_trans = NULL;
1221         DEFINE_WAIT(wait);
1222         int ret;
1223         int should_grow = 0;
1224         unsigned long now = get_seconds();
1225         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1226
1227         btrfs_run_ordered_operations(root, 0);
1228
1229         /* make a pass through all the delayed refs we have so far
1230          * any runnings procs may add more while we are here
1231          */
1232         ret = btrfs_run_delayed_refs(trans, root, 0);
1233         BUG_ON(ret);
1234
1235         btrfs_trans_release_metadata(trans, root);
1236
1237         cur_trans = trans->transaction;
1238         /*
1239          * set the flushing flag so procs in this transaction have to
1240          * start sending their work down.
1241          */
1242         cur_trans->delayed_refs.flushing = 1;
1243
1244         ret = btrfs_run_delayed_refs(trans, root, 0);
1245         BUG_ON(ret);
1246
1247         mutex_lock(&root->fs_info->trans_mutex);
1248         if (cur_trans->in_commit) {
1249                 cur_trans->use_count++;
1250                 mutex_unlock(&root->fs_info->trans_mutex);
1251                 btrfs_end_transaction(trans, root);
1252
1253                 ret = wait_for_commit(root, cur_trans);
1254                 BUG_ON(ret);
1255
1256                 mutex_lock(&root->fs_info->trans_mutex);
1257                 put_transaction(cur_trans);
1258                 mutex_unlock(&root->fs_info->trans_mutex);
1259
1260                 return 0;
1261         }
1262
1263         trans->transaction->in_commit = 1;
1264         trans->transaction->blocked = 1;
1265         wake_up(&root->fs_info->transaction_blocked_wait);
1266
1267         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1268                 prev_trans = list_entry(cur_trans->list.prev,
1269                                         struct btrfs_transaction, list);
1270                 if (!prev_trans->commit_done) {
1271                         prev_trans->use_count++;
1272                         mutex_unlock(&root->fs_info->trans_mutex);
1273
1274                         wait_for_commit(root, prev_trans);
1275
1276                         mutex_lock(&root->fs_info->trans_mutex);
1277                         put_transaction(prev_trans);
1278                 }
1279         }
1280
1281         if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
1282                 should_grow = 1;
1283
1284         do {
1285                 int snap_pending = 0;
1286                 joined = cur_trans->num_joined;
1287                 if (!list_empty(&trans->transaction->pending_snapshots))
1288                         snap_pending = 1;
1289
1290                 WARN_ON(cur_trans != trans->transaction);
1291                 mutex_unlock(&root->fs_info->trans_mutex);
1292
1293                 if (flush_on_commit || snap_pending) {
1294                         btrfs_start_delalloc_inodes(root, 1);
1295                         ret = btrfs_wait_ordered_extents(root, 0, 1);
1296                         BUG_ON(ret);
1297                 }
1298
1299                 /*
1300                  * rename don't use btrfs_join_transaction, so, once we
1301                  * set the transaction to blocked above, we aren't going
1302                  * to get any new ordered operations.  We can safely run
1303                  * it here and no for sure that nothing new will be added
1304                  * to the list
1305                  */
1306                 btrfs_run_ordered_operations(root, 1);
1307
1308                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1309                                 TASK_UNINTERRUPTIBLE);
1310
1311                 smp_mb();
1312                 if (cur_trans->num_writers > 1)
1313                         schedule_timeout(MAX_SCHEDULE_TIMEOUT);
1314                 else if (should_grow)
1315                         schedule_timeout(1);
1316
1317                 mutex_lock(&root->fs_info->trans_mutex);
1318                 finish_wait(&cur_trans->writer_wait, &wait);
1319         } while (cur_trans->num_writers > 1 ||
1320                  (should_grow && cur_trans->num_joined != joined));
1321
1322         ret = create_pending_snapshots(trans, root->fs_info);
1323         BUG_ON(ret);
1324
1325         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1326         BUG_ON(ret);
1327
1328         WARN_ON(cur_trans != trans->transaction);
1329
1330         /* btrfs_commit_tree_roots is responsible for getting the
1331          * various roots consistent with each other.  Every pointer
1332          * in the tree of tree roots has to point to the most up to date
1333          * root for every subvolume and other tree.  So, we have to keep
1334          * the tree logging code from jumping in and changing any
1335          * of the trees.
1336          *
1337          * At this point in the commit, there can't be any tree-log
1338          * writers, but a little lower down we drop the trans mutex
1339          * and let new people in.  By holding the tree_log_mutex
1340          * from now until after the super is written, we avoid races
1341          * with the tree-log code.
1342          */
1343         mutex_lock(&root->fs_info->tree_log_mutex);
1344
1345         ret = commit_fs_roots(trans, root);
1346         BUG_ON(ret);
1347
1348         /* commit_fs_roots gets rid of all the tree log roots, it is now
1349          * safe to free the root of tree log roots
1350          */
1351         btrfs_free_log_root_tree(trans, root->fs_info);
1352
1353         ret = commit_cowonly_roots(trans, root);
1354         BUG_ON(ret);
1355
1356         btrfs_prepare_extent_commit(trans, root);
1357
1358         cur_trans = root->fs_info->running_transaction;
1359         spin_lock(&root->fs_info->new_trans_lock);
1360         root->fs_info->running_transaction = NULL;
1361         spin_unlock(&root->fs_info->new_trans_lock);
1362
1363         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1364                             root->fs_info->tree_root->node);
1365         switch_commit_root(root->fs_info->tree_root);
1366
1367         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1368                             root->fs_info->chunk_root->node);
1369         switch_commit_root(root->fs_info->chunk_root);
1370
1371         update_super_roots(root);
1372
1373         if (!root->fs_info->log_root_recovering) {
1374                 btrfs_set_super_log_root(&root->fs_info->super_copy, 0);
1375                 btrfs_set_super_log_root_level(&root->fs_info->super_copy, 0);
1376         }
1377
1378         memcpy(&root->fs_info->super_for_commit, &root->fs_info->super_copy,
1379                sizeof(root->fs_info->super_copy));
1380
1381         trans->transaction->blocked = 0;
1382
1383         wake_up(&root->fs_info->transaction_wait);
1384
1385         mutex_unlock(&root->fs_info->trans_mutex);
1386         ret = btrfs_write_and_wait_transaction(trans, root);
1387         BUG_ON(ret);
1388         write_ctree_super(trans, root, 0);
1389
1390         /*
1391          * the super is written, we can safely allow the tree-loggers
1392          * to go about their business
1393          */
1394         mutex_unlock(&root->fs_info->tree_log_mutex);
1395
1396         btrfs_finish_extent_commit(trans, root);
1397
1398         mutex_lock(&root->fs_info->trans_mutex);
1399
1400         cur_trans->commit_done = 1;
1401
1402         root->fs_info->last_trans_committed = cur_trans->transid;
1403
1404         wake_up(&cur_trans->commit_wait);
1405
1406         put_transaction(cur_trans);
1407         put_transaction(cur_trans);
1408
1409         trace_btrfs_transaction_commit(root);
1410
1411         mutex_unlock(&root->fs_info->trans_mutex);
1412
1413         if (current->journal_info == trans)
1414                 current->journal_info = NULL;
1415
1416         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1417
1418         if (current != root->fs_info->transaction_kthread)
1419                 btrfs_run_delayed_iputs(root);
1420
1421         return ret;
1422 }
1423
1424 /*
1425  * interface function to delete all the snapshots we have scheduled for deletion
1426  */
1427 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1428 {
1429         LIST_HEAD(list);
1430         struct btrfs_fs_info *fs_info = root->fs_info;
1431
1432         mutex_lock(&fs_info->trans_mutex);
1433         list_splice_init(&fs_info->dead_roots, &list);
1434         mutex_unlock(&fs_info->trans_mutex);
1435
1436         while (!list_empty(&list)) {
1437                 root = list_entry(list.next, struct btrfs_root, root_list);
1438                 list_del(&root->root_list);
1439
1440                 if (btrfs_header_backref_rev(root->node) <
1441                     BTRFS_MIXED_BACKREF_REV)
1442                         btrfs_drop_snapshot(root, NULL, 0);
1443                 else
1444                         btrfs_drop_snapshot(root, NULL, 1);
1445         }
1446         return 0;
1447 }