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                 memset(transaction, 0, sizeof(*transaction));
  39                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  40         }
  41 }
  42
  43 static noinline void switch_commit_root(struct btrfs_root *root)
  44 {
  45         free_extent_buffer(root->commit_root);
  46         root->commit_root = btrfs_root_node(root);
  47 }
  48
  49 /*
  50  * either allocate a new transaction or hop into the existing one
  51  */
  52 static noinline int join_transaction(struct btrfs_root *root)
  53 {
  54         struct btrfs_transaction *cur_trans;
  55         cur_trans = root->fs_info->running_transaction;
  56         if (!cur_trans) {
  57                 cur_trans = kmem_cache_alloc(btrfs_transaction_cachep,
  58                                              GFP_NOFS);
  59                 if (!cur_trans)
  60                         return -ENOMEM;
  61                 root->fs_info->generation++;
  62                 atomic_set(&cur_trans->num_writers, 1);
  63                 cur_trans->num_joined = 0;
  64                 cur_trans->transid = root->fs_info->generation;
  65                 init_waitqueue_head(&cur_trans->writer_wait);
  66                 init_waitqueue_head(&cur_trans->commit_wait);
  67                 cur_trans->in_commit = 0;
  68                 cur_trans->blocked = 0;
  69                 atomic_set(&cur_trans->use_count, 1);
  70                 cur_trans->commit_done = 0;
  71                 cur_trans->start_time = get_seconds();
  72
  73                 cur_trans->delayed_refs.root = RB_ROOT;
  74                 cur_trans->delayed_refs.num_entries = 0;
  75                 cur_trans->delayed_refs.num_heads_ready = 0;
  76                 cur_trans->delayed_refs.num_heads = 0;
  77                 cur_trans->delayed_refs.flushing = 0;
  78                 cur_trans->delayed_refs.run_delayed_start = 0;
  79                 spin_lock_init(&cur_trans->delayed_refs.lock);
  80
  81                 INIT_LIST_HEAD(&cur_trans->pending_snapshots);
  82                 list_add_tail(&cur_trans->list, &root->fs_info->trans_list);
  83                 extent_io_tree_init(&cur_trans->dirty_pages,
  84                                      root->fs_info->btree_inode->i_mapping,
  85                                      GFP_NOFS);
  86                 spin_lock(&root->fs_info->new_trans_lock);
  87                 root->fs_info->running_transaction = cur_trans;
  88                 spin_unlock(&root->fs_info->new_trans_lock);
  89         } else {
  90                 atomic_inc(&cur_trans->num_writers);
  91                 cur_trans->num_joined++;
  92         }
  93
  94         return 0;
  95 }
  96
  97 /*
  98  * this does all the record keeping required to make sure that a reference
  99  * counted root is properly recorded in a given transaction.  This is required
 100  * to make sure the old root from before we joined the transaction is deleted
 101  * when the transaction commits
 102  */
 103 static noinline int record_root_in_trans(struct btrfs_trans_handle *trans,
 104                                          struct btrfs_root *root)
 105 {
 106         if (root->ref_cows && root->last_trans < trans->transid) {
 107                 WARN_ON(root == root->fs_info->extent_root);
 108                 WARN_ON(root->commit_root != root->node);
 109
 110                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 111                            (unsigned long)root->root_key.objectid,
 112                            BTRFS_ROOT_TRANS_TAG);
 113                 root->last_trans = trans->transid;
 114                 btrfs_init_reloc_root(trans, root);
 115         }
 116         return 0;
 117 }
 118
 119 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 120                                struct btrfs_root *root)
 121 {
 122         if (!root->ref_cows)
 123                 return 0;
 124
 125         mutex_lock(&root->fs_info->trans_mutex);
 126         if (root->last_trans == trans->transid) {
 127                 mutex_unlock(&root->fs_info->trans_mutex);
 128                 return 0;
 129         }
 130
 131         record_root_in_trans(trans, root);
 132         mutex_unlock(&root->fs_info->trans_mutex);
 133         return 0;
 134 }
 135
 136 /* wait for commit against the current transaction to become unblocked
 137  * when this is done, it is safe to start a new transaction, but the current
 138  * transaction might not be fully on disk.
 139  */
 140 static void wait_current_trans(struct btrfs_root *root)
 141 {
 142         struct btrfs_transaction *cur_trans;
 143
 144         cur_trans = root->fs_info->running_transaction;
 145         if (cur_trans && cur_trans->blocked) {
 146                 DEFINE_WAIT(wait);
 147                 atomic_inc(&cur_trans->use_count);
 148                 while (1) {
 149                         prepare_to_wait(&root->fs_info->transaction_wait, &wait,
 150                                         TASK_UNINTERRUPTIBLE);
 151                         if (!cur_trans->blocked)
 152                                 break;
 153                         mutex_unlock(&root->fs_info->trans_mutex);
 154                         schedule();
 155                         mutex_lock(&root->fs_info->trans_mutex);
 156                 }
 157                 finish_wait(&root->fs_info->transaction_wait, &wait);
 158                 put_transaction(cur_trans);
 159         }
 160 }
 161
 162 enum btrfs_trans_type {
 163         TRANS_START,
 164         TRANS_JOIN,
 165         TRANS_USERSPACE,
 166         TRANS_JOIN_NOLOCK,
 167 };
 168
 169 static int may_wait_transaction(struct btrfs_root *root, int type)
 170 {
 171         if (!root->fs_info->log_root_recovering &&
 172             ((type == TRANS_START && !root->fs_info->open_ioctl_trans) ||
 173              type == TRANS_USERSPACE))
 174                 return 1;
 175         return 0;
 176 }
 177
 178 static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
 179                                                     u64 num_items, int type)
 180 {
 181         struct btrfs_trans_handle *h;
 182         struct btrfs_transaction *cur_trans;
 183         int retries = 0;
 184         int ret;
 185
 186         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR)
 187                 return ERR_PTR(-EROFS);
 188 again:
 189         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 190         if (!h)
 191                 return ERR_PTR(-ENOMEM);
 192
 193         if (type != TRANS_JOIN_NOLOCK)
 194                 mutex_lock(&root->fs_info->trans_mutex);
 195         if (may_wait_transaction(root, type))
 196                 wait_current_trans(root);
 197
 198         ret = join_transaction(root);
 199         if (ret < 0) {
 200                 kmem_cache_free(btrfs_trans_handle_cachep, h);
 201                 if (type != TRANS_JOIN_NOLOCK)
 202                         mutex_unlock(&root->fs_info->trans_mutex);
 203                 return ERR_PTR(ret);
 204         }
 205
 206         cur_trans = root->fs_info->running_transaction;
 207         atomic_inc(&cur_trans->use_count);
 208         if (type != TRANS_JOIN_NOLOCK)
 209                 mutex_unlock(&root->fs_info->trans_mutex);
 210
 211         h->transid = cur_trans->transid;
 212         h->transaction = cur_trans;
 213         h->blocks_used = 0;
 214         h->block_group = 0;
 215         h->bytes_reserved = 0;
 216         h->delayed_ref_updates = 0;
 217         h->block_rsv = NULL;
 218
 219         smp_mb();
 220         if (cur_trans->blocked && may_wait_transaction(root, type)) {
 221                 btrfs_commit_transaction(h, root);
 222                 goto again;
 223         }
 224
 225         if (num_items > 0) {
 226                 ret = btrfs_trans_reserve_metadata(h, root, num_items);
 227                 if (ret == -EAGAIN && !retries) {
 228                         retries++;
 229                         btrfs_commit_transaction(h, root);
 230                         goto again;
 231                 } else if (ret == -EAGAIN) {
 232                         /*
 233                          * We have already retried and got EAGAIN, so really we
 234                          * don't have space, so set ret to -ENOSPC.
 235                          */
 236                         ret = -ENOSPC;
 237                 }
 238
 239                 if (ret < 0) {
 240                         btrfs_end_transaction(h, root);
 241                         return ERR_PTR(ret);
 242                 }
 243         }
 244
 245         if (type != TRANS_JOIN_NOLOCK)
 246                 mutex_lock(&root->fs_info->trans_mutex);
 247         record_root_in_trans(h, root);
 248         if (type != TRANS_JOIN_NOLOCK)
 249                 mutex_unlock(&root->fs_info->trans_mutex);
 250
 251         if (!current->journal_info && type != TRANS_USERSPACE)
 252                 current->journal_info = h;
 253         return h;
 254 }
 255
 256 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 257                                                    int num_items)
 258 {
 259         return start_transaction(root, num_items, TRANS_START);
 260 }
 261 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root,
 262                                                    int num_blocks)
 263 {
 264         return start_transaction(root, 0, TRANS_JOIN);
 265 }
 266
 267 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root,
 268                                                           int num_blocks)
 269 {
 270         return start_transaction(root, 0, TRANS_JOIN_NOLOCK);
 271 }
 272
 273 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *r,
 274                                                          int num_blocks)
 275 {
 276         return start_transaction(r, 0, TRANS_USERSPACE);
 277 }
 278
 279 /* wait for a transaction commit to be fully complete */
 280 static noinline int wait_for_commit(struct btrfs_root *root,
 281                                     struct btrfs_transaction *commit)
 282 {
 283         DEFINE_WAIT(wait);
 284         mutex_lock(&root->fs_info->trans_mutex);
 285         while (!commit->commit_done) {
 286                 prepare_to_wait(&commit->commit_wait, &wait,
 287                                 TASK_UNINTERRUPTIBLE);
 288                 if (commit->commit_done)
 289                         break;
 290                 mutex_unlock(&root->fs_info->trans_mutex);
 291                 schedule();
 292                 mutex_lock(&root->fs_info->trans_mutex);
 293         }
 294         mutex_unlock(&root->fs_info->trans_mutex);
 295         finish_wait(&commit->commit_wait, &wait);
 296         return 0;
 297 }
 298
 299 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 300 {
 301         struct btrfs_transaction *cur_trans = NULL, *t;
 302         int ret;
 303
 304         mutex_lock(&root->fs_info->trans_mutex);
 305
 306         ret = 0;
 307         if (transid) {
 308                 if (transid <= root->fs_info->last_trans_committed)
 309                         goto out_unlock;
 310
 311                 /* find specified transaction */
 312                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 313                         if (t->transid == transid) {
 314                                 cur_trans = t;
 315                                 break;
 316                         }
 317                         if (t->transid > transid)
 318                                 break;
 319                 }
 320                 ret = -EINVAL;
 321                 if (!cur_trans)
 322                         goto out_unlock;  /* bad transid */
 323         } else {
 324                 /* find newest transaction that is committing | committed */
 325                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 326                                             list) {
 327                         if (t->in_commit) {
 328                                 if (t->commit_done)
 329                                         goto out_unlock;
 330                                 cur_trans = t;
 331                                 break;
 332                         }
 333                 }
 334                 if (!cur_trans)
 335                         goto out_unlock;  /* nothing committing|committed */
 336         }
 337
 338         atomic_inc(&cur_trans->use_count);
 339         mutex_unlock(&root->fs_info->trans_mutex);
 340
 341         wait_for_commit(root, cur_trans);
 342
 343         mutex_lock(&root->fs_info->trans_mutex);
 344         put_transaction(cur_trans);
 345         ret = 0;
 346 out_unlock:
 347         mutex_unlock(&root->fs_info->trans_mutex);
 348         return ret;
 349 }
 350
 351 #if 0
 352 /*
 353  * rate limit against the drop_snapshot code.  This helps to slow down new
 354  * operations if the drop_snapshot code isn't able to keep up.
 355  */
 356 static void throttle_on_drops(struct btrfs_root *root)
 357 {
 358         struct btrfs_fs_info *info = root->fs_info;
 359         int harder_count = 0;
 360
 361 harder:
 362         if (atomic_read(&info->throttles)) {
 363                 DEFINE_WAIT(wait);
 364                 int thr;
 365                 thr = atomic_read(&info->throttle_gen);
 366
 367                 do {
 368                         prepare_to_wait(&info->transaction_throttle,
 369                                         &wait, TASK_UNINTERRUPTIBLE);
 370                         if (!atomic_read(&info->throttles)) {
 371                                 finish_wait(&info->transaction_throttle, &wait);
 372                                 break;
 373                         }
 374                         schedule();
 375                         finish_wait(&info->transaction_throttle, &wait);
 376                 } while (thr == atomic_read(&info->throttle_gen));
 377                 harder_count++;
 378
 379                 if (root->fs_info->total_ref_cache_size > 1 * 1024 * 1024 &&
 380                     harder_count < 2)
 381                         goto harder;
 382
 383                 if (root->fs_info->total_ref_cache_size > 5 * 1024 * 1024 &&
 384                     harder_count < 10)
 385                         goto harder;
 386
 387                 if (root->fs_info->total_ref_cache_size > 10 * 1024 * 1024 &&
 388                     harder_count < 20)
 389                         goto harder;
 390         }
 391 }
 392 #endif
 393
 394 void btrfs_throttle(struct btrfs_root *root)
 395 {
 396         mutex_lock(&root->fs_info->trans_mutex);
 397         if (!root->fs_info->open_ioctl_trans)
 398                 wait_current_trans(root);
 399         mutex_unlock(&root->fs_info->trans_mutex);
 400 }
 401
 402 static int should_end_transaction(struct btrfs_trans_handle *trans,
 403                                   struct btrfs_root *root)
 404 {
 405         int ret;
 406         ret = btrfs_block_rsv_check(trans, root,
 407                                     &root->fs_info->global_block_rsv, 0, 5);
 408         return ret ? 1 : 0;
 409 }
 410
 411 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 412                                  struct btrfs_root *root)
 413 {
 414         struct btrfs_transaction *cur_trans = trans->transaction;
 415         int updates;
 416
 417         if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
 418                 return 1;
 419
 420         updates = trans->delayed_ref_updates;
 421         trans->delayed_ref_updates = 0;
 422         if (updates)
 423                 btrfs_run_delayed_refs(trans, root, updates);
 424
 425         return should_end_transaction(trans, root);
 426 }
 427
 428 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 429                           struct btrfs_root *root, int throttle, int lock)
 430 {
 431         struct btrfs_transaction *cur_trans = trans->transaction;
 432         struct btrfs_fs_info *info = root->fs_info;
 433         int count = 0;
 434
 435         while (count < 4) {
 436                 unsigned long cur = trans->delayed_ref_updates;
 437                 trans->delayed_ref_updates = 0;
 438                 if (cur &&
 439                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 440                         trans->delayed_ref_updates = 0;
 441
 442                         /*
 443                          * do a full flush if the transaction is trying
 444                          * to close
 445                          */
 446                         if (trans->transaction->delayed_refs.flushing)
 447                                 cur = 0;
 448                         btrfs_run_delayed_refs(trans, root, cur);
 449                 } else {
 450                         break;
 451                 }
 452                 count++;
 453         }
 454
 455         btrfs_trans_release_metadata(trans, root);
 456
 457         if (lock && !root->fs_info->open_ioctl_trans &&
 458             should_end_transaction(trans, root))
 459                 trans->transaction->blocked = 1;
 460
 461         if (lock && cur_trans->blocked && !cur_trans->in_commit) {
 462                 if (throttle)
 463                         return btrfs_commit_transaction(trans, root);
 464                 else
 465                         wake_up_process(info->transaction_kthread);
 466         }
 467
 468         WARN_ON(cur_trans != info->running_transaction);
 469         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 470         atomic_dec(&cur_trans->num_writers);
 471
 472         smp_mb();
 473         if (waitqueue_active(&cur_trans->writer_wait))
 474                 wake_up(&cur_trans->writer_wait);
 475         put_transaction(cur_trans);
 476
 477         if (current->journal_info == trans)
 478                 current->journal_info = NULL;
 479         memset(trans, 0, sizeof(*trans));
 480         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 481
 482         if (throttle)
 483                 btrfs_run_delayed_iputs(root);
 484
 485         return 0;
 486 }
 487
 488 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 489                           struct btrfs_root *root)
 490 {
 491         return __btrfs_end_transaction(trans, root, 0, 1);
 492 }
 493
 494 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 495                                    struct btrfs_root *root)
 496 {
 497         return __btrfs_end_transaction(trans, root, 1, 1);
 498 }
 499
 500 int btrfs_end_transaction_nolock(struct btrfs_trans_handle *trans,
 501                                  struct btrfs_root *root)
 502 {
 503         return __btrfs_end_transaction(trans, root, 0, 0);
 504 }
 505
 506 /*
 507  * when btree blocks are allocated, they have some corresponding bits set for
 508  * them in one of two extent_io trees.  This is used to make sure all of
 509  * those extents are sent to disk but does not wait on them
 510  */
 511 int btrfs_write_marked_extents(struct btrfs_root *root,
 512                                struct extent_io_tree *dirty_pages, int mark)
 513 {
 514         int ret;
 515         int err = 0;
 516         int werr = 0;
 517         struct page *page;
 518         struct inode *btree_inode = root->fs_info->btree_inode;
 519         u64 start = 0;
 520         u64 end;
 521         unsigned long index;
 522
 523         while (1) {
 524                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 525                                             mark);
 526                 if (ret)
 527                         break;
 528                 while (start <= end) {
 529                         cond_resched();
 530
 531                         index = start >> PAGE_CACHE_SHIFT;
 532                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 533                         page = find_get_page(btree_inode->i_mapping, index);
 534                         if (!page)
 535                                 continue;
 536
 537                         btree_lock_page_hook(page);
 538                         if (!page->mapping) {
 539                                 unlock_page(page);
 540                                 page_cache_release(page);
 541                                 continue;
 542                         }
 543
 544                         if (PageWriteback(page)) {
 545                                 if (PageDirty(page))
 546                                         wait_on_page_writeback(page);
 547                                 else {
 548                                         unlock_page(page);
 549                                         page_cache_release(page);
 550                                         continue;
 551                                 }
 552                         }
 553                         err = write_one_page(page, 0);
 554                         if (err)
 555                                 werr = err;
 556                         page_cache_release(page);
 557                 }
 558         }
 559         if (err)
 560                 werr = err;
 561         return werr;
 562 }
 563
 564 /*
 565  * when btree blocks are allocated, they have some corresponding bits set for
 566  * them in one of two extent_io trees.  This is used to make sure all of
 567  * those extents are on disk for transaction or log commit.  We wait
 568  * on all the pages and clear them from the dirty pages state tree
 569  */
 570 int btrfs_wait_marked_extents(struct btrfs_root *root,
 571                               struct extent_io_tree *dirty_pages, int mark)
 572 {
 573         int ret;
 574         int err = 0;
 575         int werr = 0;
 576         struct page *page;
 577         struct inode *btree_inode = root->fs_info->btree_inode;
 578         u64 start = 0;
 579         u64 end;
 580         unsigned long index;
 581
 582         while (1) {
 583                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 584                                             mark);
 585                 if (ret)
 586                         break;
 587
 588                 clear_extent_bits(dirty_pages, start, end, mark, GFP_NOFS);
 589                 while (start <= end) {
 590                         index = start >> PAGE_CACHE_SHIFT;
 591                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 592                         page = find_get_page(btree_inode->i_mapping, index);
 593                         if (!page)
 594                                 continue;
 595                         if (PageDirty(page)) {
 596                                 btree_lock_page_hook(page);
 597                                 wait_on_page_writeback(page);
 598                                 err = write_one_page(page, 0);
 599                                 if (err)
 600                                         werr = err;
 601                         }
 602                         wait_on_page_writeback(page);
 603                         page_cache_release(page);
 604                         cond_resched();
 605                 }
 606         }
 607         if (err)
 608                 werr = err;
 609         return werr;
 610 }
 611
 612 /*
 613  * when btree blocks are allocated, they have some corresponding bits set for
 614  * them in one of two extent_io trees.  This is used to make sure all of
 615  * those extents are on disk for transaction or log commit
 616  */
 617 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 618                                 struct extent_io_tree *dirty_pages, int mark)
 619 {
 620         int ret;
 621         int ret2;
 622
 623         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 624         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 625         return ret || ret2;
 626 }
 627
 628 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 629                                      struct btrfs_root *root)
 630 {
 631         if (!trans || !trans->transaction) {
 632                 struct inode *btree_inode;
 633                 btree_inode = root->fs_info->btree_inode;
 634                 return filemap_write_and_wait(btree_inode->i_mapping);
 635         }
 636         return btrfs_write_and_wait_marked_extents(root,
 637                                            &trans->transaction->dirty_pages,
 638                                            EXTENT_DIRTY);
 639 }
 640
 641 /*
 642  * this is used to update the root pointer in the tree of tree roots.
 643  *
 644  * But, in the case of the extent allocation tree, updating the root
 645  * pointer may allocate blocks which may change the root of the extent
 646  * allocation tree.
 647  *
 648  * So, this loops and repeats and makes sure the cowonly root didn't
 649  * change while the root pointer was being updated in the metadata.
 650  */
 651 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 652                                struct btrfs_root *root)
 653 {
 654         int ret;
 655         u64 old_root_bytenr;
 656         u64 old_root_used;
 657         struct btrfs_root *tree_root = root->fs_info->tree_root;
 658
 659         old_root_used = btrfs_root_used(&root->root_item);
 660         btrfs_write_dirty_block_groups(trans, root);
 661
 662         while (1) {
 663                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 664                 if (old_root_bytenr == root->node->start &&
 665                     old_root_used == btrfs_root_used(&root->root_item))
 666                         break;
 667
 668                 btrfs_set_root_node(&root->root_item, root->node);
 669                 ret = btrfs_update_root(trans, tree_root,
 670                                         &root->root_key,
 671                                         &root->root_item);
 672                 BUG_ON(ret);
 673
 674                 old_root_used = btrfs_root_used(&root->root_item);
 675                 ret = btrfs_write_dirty_block_groups(trans, root);
 676                 BUG_ON(ret);
 677         }
 678
 679         if (root != root->fs_info->extent_root)
 680                 switch_commit_root(root);
 681
 682         return 0;
 683 }
 684
 685 /*
 686  * update all the cowonly tree roots on disk
 687  */
 688 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 689                                          struct btrfs_root *root)
 690 {
 691         struct btrfs_fs_info *fs_info = root->fs_info;
 692         struct list_head *next;
 693         struct extent_buffer *eb;
 694         int ret;
 695
 696         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 697         BUG_ON(ret);
 698
 699         eb = btrfs_lock_root_node(fs_info->tree_root);
 700         btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
 701         btrfs_tree_unlock(eb);
 702         free_extent_buffer(eb);
 703
 704         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 705         BUG_ON(ret);
 706
 707         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 708                 next = fs_info->dirty_cowonly_roots.next;
 709                 list_del_init(next);
 710                 root = list_entry(next, struct btrfs_root, dirty_list);
 711
 712                 update_cowonly_root(trans, root);
 713         }
 714
 715         down_write(&fs_info->extent_commit_sem);
 716         switch_commit_root(fs_info->extent_root);
 717         up_write(&fs_info->extent_commit_sem);
 718
 719         return 0;
 720 }
 721
 722 /*
 723  * dead roots are old snapshots that need to be deleted.  This allocates
 724  * a dirty root struct and adds it into the list of dead roots that need to
 725  * be deleted
 726  */
 727 int btrfs_add_dead_root(struct btrfs_root *root)
 728 {
 729         mutex_lock(&root->fs_info->trans_mutex);
 730         list_add(&root->root_list, &root->fs_info->dead_roots);
 731         mutex_unlock(&root->fs_info->trans_mutex);
 732         return 0;
 733 }
 734
 735 /*
 736  * update all the cowonly tree roots on disk
 737  */
 738 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 739                                     struct btrfs_root *root)
 740 {
 741         struct btrfs_root *gang[8];
 742         struct btrfs_fs_info *fs_info = root->fs_info;
 743         int i;
 744         int ret;
 745         int err = 0;
 746
 747         while (1) {
 748                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 749                                                  (void **)gang, 0,
 750                                                  ARRAY_SIZE(gang),
 751                                                  BTRFS_ROOT_TRANS_TAG);
 752                 if (ret == 0)
 753                         break;
 754                 for (i = 0; i < ret; i++) {
 755                         root = gang[i];
 756                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 757                                         (unsigned long)root->root_key.objectid,
 758                                         BTRFS_ROOT_TRANS_TAG);
 759
 760                         btrfs_free_log(trans, root);
 761                         btrfs_update_reloc_root(trans, root);
 762                         btrfs_orphan_commit_root(trans, root);
 763
 764                         btrfs_save_ino_cache(root, trans);
 765
 766                         if (root->commit_root != root->node) {
 767                                 mutex_lock(&root->fs_commit_mutex);
 768                                 switch_commit_root(root);
 769                                 btrfs_unpin_free_ino(root);
 770                                 mutex_unlock(&root->fs_commit_mutex);
 771
 772                                 btrfs_set_root_node(&root->root_item,
 773                                                     root->node);
 774                         }
 775
 776                         err = btrfs_update_root(trans, fs_info->tree_root,
 777                                                 &root->root_key,
 778                                                 &root->root_item);
 779                         if (err)
 780                                 break;
 781                 }
 782         }
 783         return err;
 784 }
 785
 786 /*
 787  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 788  * otherwise every leaf in the btree is read and defragged.
 789  */
 790 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 791 {
 792         struct btrfs_fs_info *info = root->fs_info;
 793         struct btrfs_trans_handle *trans;
 794         int ret;
 795         unsigned long nr;
 796
 797         if (xchg(&root->defrag_running, 1))
 798                 return 0;
 799
 800         while (1) {
 801                 trans = btrfs_start_transaction(root, 0);
 802                 if (IS_ERR(trans))
 803                         return PTR_ERR(trans);
 804
 805                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 806
 807                 nr = trans->blocks_used;
 808                 btrfs_end_transaction(trans, root);
 809                 btrfs_btree_balance_dirty(info->tree_root, nr);
 810                 cond_resched();
 811
 812                 if (root->fs_info->closing || ret != -EAGAIN)
 813                         break;
 814         }
 815         root->defrag_running = 0;
 816         return ret;
 817 }
 818
 819 #if 0
 820 /*
 821  * when dropping snapshots, we generate a ton of delayed refs, and it makes
 822  * sense not to join the transaction while it is trying to flush the current
 823  * queue of delayed refs out.
 824  *
 825  * This is used by the drop snapshot code only
 826  */
 827 static noinline int wait_transaction_pre_flush(struct btrfs_fs_info *info)
 828 {
 829         DEFINE_WAIT(wait);
 830
 831         mutex_lock(&info->trans_mutex);
 832         while (info->running_transaction &&
 833                info->running_transaction->delayed_refs.flushing) {
 834                 prepare_to_wait(&info->transaction_wait, &wait,
 835                                 TASK_UNINTERRUPTIBLE);
 836                 mutex_unlock(&info->trans_mutex);
 837
 838                 schedule();
 839
 840                 mutex_lock(&info->trans_mutex);
 841                 finish_wait(&info->transaction_wait, &wait);
 842         }
 843         mutex_unlock(&info->trans_mutex);
 844         return 0;
 845 }
 846
 847 /*
 848  * Given a list of roots that need to be deleted, call btrfs_drop_snapshot on
 849  * all of them
 850  */
 851 int btrfs_drop_dead_root(struct btrfs_root *root)
 852 {
 853         struct btrfs_trans_handle *trans;
 854         struct btrfs_root *tree_root = root->fs_info->tree_root;
 855         unsigned long nr;
 856         int ret;
 857
 858         while (1) {
 859                 /*
 860                  * we don't want to jump in and create a bunch of
 861                  * delayed refs if the transaction is starting to close
 862                  */
 863                 wait_transaction_pre_flush(tree_root->fs_info);
 864                 trans = btrfs_start_transaction(tree_root, 1);
 865
 866                 /*
 867                  * we've joined a transaction, make sure it isn't
 868                  * closing right now
 869                  */
 870                 if (trans->transaction->delayed_refs.flushing) {
 871                         btrfs_end_transaction(trans, tree_root);
 872                         continue;
 873                 }
 874
 875                 ret = btrfs_drop_snapshot(trans, root);
 876                 if (ret != -EAGAIN)
 877                         break;
 878
 879                 ret = btrfs_update_root(trans, tree_root,
 880                                         &root->root_key,
 881                                         &root->root_item);
 882                 if (ret)
 883                         break;
 884
 885                 nr = trans->blocks_used;
 886                 ret = btrfs_end_transaction(trans, tree_root);
 887                 BUG_ON(ret);
 888
 889                 btrfs_btree_balance_dirty(tree_root, nr);
 890                 cond_resched();
 891         }
 892         BUG_ON(ret);
 893
 894         ret = btrfs_del_root(trans, tree_root, &root->root_key);
 895         BUG_ON(ret);
 896
 897         nr = trans->blocks_used;
 898         ret = btrfs_end_transaction(trans, tree_root);
 899         BUG_ON(ret);
 900
 901         free_extent_buffer(root->node);
 902         free_extent_buffer(root->commit_root);
 903         kfree(root);
 904
 905         btrfs_btree_balance_dirty(tree_root, nr);
 906         return ret;
 907 }
 908 #endif
 909
 910 /*
 911  * new snapshots need to be created at a very specific time in the
 912  * transaction commit.  This does the actual creation
 913  */
 914 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 915                                    struct btrfs_fs_info *fs_info,
 916                                    struct btrfs_pending_snapshot *pending)
 917 {
 918         struct btrfs_key key;
 919         struct btrfs_root_item *new_root_item;
 920         struct btrfs_root *tree_root = fs_info->tree_root;
 921         struct btrfs_root *root = pending->root;
 922         struct btrfs_root *parent_root;
 923         struct inode *parent_inode;
 924         struct dentry *parent;
 925         struct dentry *dentry;
 926         struct extent_buffer *tmp;
 927         struct extent_buffer *old;
 928         int ret;
 929         u64 to_reserve = 0;
 930         u64 index = 0;
 931         u64 objectid;
 932         u64 root_flags;
 933
 934         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 935         if (!new_root_item) {
 936                 pending->error = -ENOMEM;
 937                 goto fail;
 938         }
 939
 940         ret = btrfs_find_free_objectid(tree_root, &objectid);
 941         if (ret) {
 942                 pending->error = ret;
 943                 goto fail;
 944         }
 945
 946         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
 947         btrfs_orphan_pre_snapshot(trans, pending, &to_reserve);
 948
 949         if (to_reserve > 0) {
 950                 ret = btrfs_block_rsv_add(trans, root, &pending->block_rsv,
 951                                           to_reserve);
 952                 if (ret) {
 953                         pending->error = ret;
 954                         goto fail;
 955                 }
 956         }
 957
 958         key.objectid = objectid;
 959         key.offset = (u64)-1;
 960         key.type = BTRFS_ROOT_ITEM_KEY;
 961
 962         trans->block_rsv = &pending->block_rsv;
 963
 964         dentry = pending->dentry;
 965         parent = dget_parent(dentry);
 966         parent_inode = parent->d_inode;
 967         parent_root = BTRFS_I(parent_inode)->root;
 968         record_root_in_trans(trans, parent_root);
 969
 970         /*
 971          * insert the directory item
 972          */
 973         ret = btrfs_set_inode_index(parent_inode, &index);
 974         BUG_ON(ret);
 975         ret = btrfs_insert_dir_item(trans, parent_root,
 976                                 dentry->d_name.name, dentry->d_name.len,
 977                                 btrfs_ino(parent_inode), &key,
 978                                 BTRFS_FT_DIR, index);
 979         BUG_ON(ret);
 980
 981         btrfs_i_size_write(parent_inode, parent_inode->i_size +
 982                                          dentry->d_name.len * 2);
 983         ret = btrfs_update_inode(trans, parent_root, parent_inode);
 984         BUG_ON(ret);
 985
 986         record_root_in_trans(trans, root);
 987         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
 988         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
 989         btrfs_check_and_init_root_item(new_root_item);
 990
 991         root_flags = btrfs_root_flags(new_root_item);
 992         if (pending->readonly)
 993                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
 994         else
 995                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
 996         btrfs_set_root_flags(new_root_item, root_flags);
 997
 998         old = btrfs_lock_root_node(root);
 999         btrfs_cow_block(trans, root, old, NULL, 0, &old);
1000         btrfs_set_lock_blocking(old);
1001
1002         btrfs_copy_root(trans, root, old, &tmp, objectid);
1003         btrfs_tree_unlock(old);
1004         free_extent_buffer(old);
1005
1006         btrfs_set_root_node(new_root_item, tmp);
1007         /* record when the snapshot was created in key.offset */
1008         key.offset = trans->transid;
1009         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1010         btrfs_tree_unlock(tmp);
1011         free_extent_buffer(tmp);
1012         BUG_ON(ret);
1013
1014         /*
1015          * insert root back/forward references
1016          */
1017         ret = btrfs_add_root_ref(trans, tree_root, objectid,
1018                                  parent_root->root_key.objectid,
1019                                  btrfs_ino(parent_inode), index,
1020                                  dentry->d_name.name, dentry->d_name.len);
1021         BUG_ON(ret);
1022         dput(parent);
1023
1024         key.offset = (u64)-1;
1025         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1026         BUG_ON(IS_ERR(pending->snap));
1027
1028         btrfs_reloc_post_snapshot(trans, pending);
1029         btrfs_orphan_post_snapshot(trans, pending);
1030 fail:
1031         kfree(new_root_item);
1032         btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
1033         return 0;
1034 }
1035
1036 /*
1037  * create all the snapshots we've scheduled for creation
1038  */
1039 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1040                                              struct btrfs_fs_info *fs_info)
1041 {
1042         struct btrfs_pending_snapshot *pending;
1043         struct list_head *head = &trans->transaction->pending_snapshots;
1044         int ret;
1045
1046         list_for_each_entry(pending, head, list) {
1047                 ret = create_pending_snapshot(trans, fs_info, pending);
1048                 BUG_ON(ret);
1049         }
1050         return 0;
1051 }
1052
1053 static void update_super_roots(struct btrfs_root *root)
1054 {
1055         struct btrfs_root_item *root_item;
1056         struct btrfs_super_block *super;
1057
1058         super = &root->fs_info->super_copy;
1059
1060         root_item = &root->fs_info->chunk_root->root_item;
1061         super->chunk_root = root_item->bytenr;
1062         super->chunk_root_generation = root_item->generation;
1063         super->chunk_root_level = root_item->level;
1064
1065         root_item = &root->fs_info->tree_root->root_item;
1066         super->root = root_item->bytenr;
1067         super->generation = root_item->generation;
1068         super->root_level = root_item->level;
1069         if (super->cache_generation != 0 || btrfs_test_opt(root, SPACE_CACHE))
1070                 super->cache_generation = root_item->generation;
1071 }
1072
1073 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1074 {
1075         int ret = 0;
1076         spin_lock(&info->new_trans_lock);
1077         if (info->running_transaction)
1078                 ret = info->running_transaction->in_commit;
1079         spin_unlock(&info->new_trans_lock);
1080         return ret;
1081 }
1082
1083 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1084 {
1085         int ret = 0;
1086         spin_lock(&info->new_trans_lock);
1087         if (info->running_transaction)
1088                 ret = info->running_transaction->blocked;
1089         spin_unlock(&info->new_trans_lock);
1090         return ret;
1091 }
1092
1093 /*
1094  * wait for the current transaction commit to start and block subsequent
1095  * transaction joins
1096  */
1097 static void wait_current_trans_commit_start(struct btrfs_root *root,
1098                                             struct btrfs_transaction *trans)
1099 {
1100         DEFINE_WAIT(wait);
1101
1102         if (trans->in_commit)
1103                 return;
1104
1105         while (1) {
1106                 prepare_to_wait(&root->fs_info->transaction_blocked_wait, &wait,
1107                                 TASK_UNINTERRUPTIBLE);
1108                 if (trans->in_commit) {
1109                         finish_wait(&root->fs_info->transaction_blocked_wait,
1110                                     &wait);
1111                         break;
1112                 }
1113                 mutex_unlock(&root->fs_info->trans_mutex);
1114                 schedule();
1115                 mutex_lock(&root->fs_info->trans_mutex);
1116                 finish_wait(&root->fs_info->transaction_blocked_wait, &wait);
1117         }
1118 }
1119
1120 /*
1121  * wait for the current transaction to start and then become unblocked.
1122  * caller holds ref.
1123  */
1124 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1125                                          struct btrfs_transaction *trans)
1126 {
1127         DEFINE_WAIT(wait);
1128
1129         if (trans->commit_done || (trans->in_commit && !trans->blocked))
1130                 return;
1131
1132         while (1) {
1133                 prepare_to_wait(&root->fs_info->transaction_wait, &wait,
1134                                 TASK_UNINTERRUPTIBLE);
1135                 if (trans->commit_done ||
1136                     (trans->in_commit && !trans->blocked)) {
1137                         finish_wait(&root->fs_info->transaction_wait,
1138                                     &wait);
1139                         break;
1140                 }
1141                 mutex_unlock(&root->fs_info->trans_mutex);
1142                 schedule();
1143                 mutex_lock(&root->fs_info->trans_mutex);
1144                 finish_wait(&root->fs_info->transaction_wait,
1145                             &wait);
1146         }
1147 }
1148
1149 /*
1150  * commit transactions asynchronously. once btrfs_commit_transaction_async
1151  * returns, any subsequent transaction will not be allowed to join.
1152  */
1153 struct btrfs_async_commit {
1154         struct btrfs_trans_handle *newtrans;
1155         struct btrfs_root *root;
1156         struct delayed_work work;
1157 };
1158
1159 static void do_async_commit(struct work_struct *work)
1160 {
1161         struct btrfs_async_commit *ac =
1162                 container_of(work, struct btrfs_async_commit, work.work);
1163
1164         btrfs_commit_transaction(ac->newtrans, ac->root);
1165         kfree(ac);
1166 }
1167
1168 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1169                                    struct btrfs_root *root,
1170                                    int wait_for_unblock)
1171 {
1172         struct btrfs_async_commit *ac;
1173         struct btrfs_transaction *cur_trans;
1174
1175         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1176         if (!ac)
1177                 return -ENOMEM;
1178
1179         INIT_DELAYED_WORK(&ac->work, do_async_commit);
1180         ac->root = root;
1181         ac->newtrans = btrfs_join_transaction(root, 0);
1182         if (IS_ERR(ac->newtrans)) {
1183                 int err = PTR_ERR(ac->newtrans);
1184                 kfree(ac);
1185                 return err;
1186         }
1187
1188         /* take transaction reference */
1189         mutex_lock(&root->fs_info->trans_mutex);
1190         cur_trans = trans->transaction;
1191         atomic_inc(&cur_trans->use_count);
1192         mutex_unlock(&root->fs_info->trans_mutex);
1193
1194         btrfs_end_transaction(trans, root);
1195         schedule_delayed_work(&ac->work, 0);
1196
1197         /* wait for transaction to start and unblock */
1198         mutex_lock(&root->fs_info->trans_mutex);
1199         if (wait_for_unblock)
1200                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1201         else
1202                 wait_current_trans_commit_start(root, cur_trans);
1203         put_transaction(cur_trans);
1204         mutex_unlock(&root->fs_info->trans_mutex);
1205
1206         return 0;
1207 }
1208
1209 /*
1210  * btrfs_transaction state sequence:
1211  *    in_commit = 0, blocked = 0  (initial)
1212  *    in_commit = 1, blocked = 1
1213  *    blocked = 0
1214  *    commit_done = 1
1215  */
1216 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1217                              struct btrfs_root *root)
1218 {
1219         unsigned long joined = 0;
1220         struct btrfs_transaction *cur_trans;
1221         struct btrfs_transaction *prev_trans = NULL;
1222         DEFINE_WAIT(wait);
1223         int ret;
1224         int should_grow = 0;
1225         unsigned long now = get_seconds();
1226         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1227
1228         btrfs_run_ordered_operations(root, 0);
1229
1230         /* make a pass through all the delayed refs we have so far
1231          * any runnings procs may add more while we are here
1232          */
1233         ret = btrfs_run_delayed_refs(trans, root, 0);
1234         BUG_ON(ret);
1235
1236         btrfs_trans_release_metadata(trans, root);
1237
1238         cur_trans = trans->transaction;
1239         /*
1240          * set the flushing flag so procs in this transaction have to
1241          * start sending their work down.
1242          */
1243         cur_trans->delayed_refs.flushing = 1;
1244
1245         ret = btrfs_run_delayed_refs(trans, root, 0);
1246         BUG_ON(ret);
1247
1248         mutex_lock(&root->fs_info->trans_mutex);
1249         if (cur_trans->in_commit) {
1250                 atomic_inc(&cur_trans->use_count);
1251                 mutex_unlock(&root->fs_info->trans_mutex);
1252                 btrfs_end_transaction(trans, root);
1253
1254                 ret = wait_for_commit(root, cur_trans);
1255                 BUG_ON(ret);
1256
1257                 mutex_lock(&root->fs_info->trans_mutex);
1258                 put_transaction(cur_trans);
1259                 mutex_unlock(&root->fs_info->trans_mutex);
1260
1261                 return 0;
1262         }
1263
1264         trans->transaction->in_commit = 1;
1265         trans->transaction->blocked = 1;
1266         wake_up(&root->fs_info->transaction_blocked_wait);
1267
1268         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1269                 prev_trans = list_entry(cur_trans->list.prev,
1270                                         struct btrfs_transaction, list);
1271                 if (!prev_trans->commit_done) {
1272                         atomic_inc(&prev_trans->use_count);
1273                         mutex_unlock(&root->fs_info->trans_mutex);
1274
1275                         wait_for_commit(root, prev_trans);
1276
1277                         mutex_lock(&root->fs_info->trans_mutex);
1278                         put_transaction(prev_trans);
1279                 }
1280         }
1281
1282         if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
1283                 should_grow = 1;
1284
1285         do {
1286                 int snap_pending = 0;
1287                 joined = cur_trans->num_joined;
1288                 if (!list_empty(&trans->transaction->pending_snapshots))
1289                         snap_pending = 1;
1290
1291                 WARN_ON(cur_trans != trans->transaction);
1292                 mutex_unlock(&root->fs_info->trans_mutex);
1293
1294                 if (flush_on_commit || snap_pending) {
1295                         btrfs_start_delalloc_inodes(root, 1);
1296                         ret = btrfs_wait_ordered_extents(root, 0, 1);
1297                         BUG_ON(ret);
1298                 }
1299
1300                 /*
1301                  * rename don't use btrfs_join_transaction, so, once we
1302                  * set the transaction to blocked above, we aren't going
1303                  * to get any new ordered operations.  We can safely run
1304                  * it here and no for sure that nothing new will be added
1305                  * to the list
1306                  */
1307                 btrfs_run_ordered_operations(root, 1);
1308
1309                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1310                                 TASK_UNINTERRUPTIBLE);
1311
1312                 smp_mb();
1313                 if (atomic_read(&cur_trans->num_writers) > 1)
1314                         schedule_timeout(MAX_SCHEDULE_TIMEOUT);
1315                 else if (should_grow)
1316                         schedule_timeout(1);
1317
1318                 mutex_lock(&root->fs_info->trans_mutex);
1319                 finish_wait(&cur_trans->writer_wait, &wait);
1320         } while (atomic_read(&cur_trans->num_writers) > 1 ||
1321                  (should_grow && cur_trans->num_joined != joined));
1322
1323         ret = create_pending_snapshots(trans, root->fs_info);
1324         BUG_ON(ret);
1325
1326         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1327         BUG_ON(ret);
1328
1329         WARN_ON(cur_trans != trans->transaction);
1330
1331         /* btrfs_commit_tree_roots is responsible for getting the
1332          * various roots consistent with each other.  Every pointer
1333          * in the tree of tree roots has to point to the most up to date
1334          * root for every subvolume and other tree.  So, we have to keep
1335          * the tree logging code from jumping in and changing any
1336          * of the trees.
1337          *
1338          * At this point in the commit, there can't be any tree-log
1339          * writers, but a little lower down we drop the trans mutex
1340          * and let new people in.  By holding the tree_log_mutex
1341          * from now until after the super is written, we avoid races
1342          * with the tree-log code.
1343          */
1344         mutex_lock(&root->fs_info->tree_log_mutex);
1345
1346         ret = commit_fs_roots(trans, root);
1347         BUG_ON(ret);
1348
1349         /* commit_fs_roots gets rid of all the tree log roots, it is now
1350          * safe to free the root of tree log roots
1351          */
1352         btrfs_free_log_root_tree(trans, root->fs_info);
1353
1354         ret = commit_cowonly_roots(trans, root);
1355         BUG_ON(ret);
1356
1357         btrfs_prepare_extent_commit(trans, root);
1358
1359         cur_trans = root->fs_info->running_transaction;
1360         spin_lock(&root->fs_info->new_trans_lock);
1361         root->fs_info->running_transaction = NULL;
1362         spin_unlock(&root->fs_info->new_trans_lock);
1363
1364         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1365                             root->fs_info->tree_root->node);
1366         switch_commit_root(root->fs_info->tree_root);
1367
1368         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1369                             root->fs_info->chunk_root->node);
1370         switch_commit_root(root->fs_info->chunk_root);
1371
1372         update_super_roots(root);
1373
1374         if (!root->fs_info->log_root_recovering) {
1375                 btrfs_set_super_log_root(&root->fs_info->super_copy, 0);
1376                 btrfs_set_super_log_root_level(&root->fs_info->super_copy, 0);
1377         }
1378
1379         memcpy(&root->fs_info->super_for_commit, &root->fs_info->super_copy,
1380                sizeof(root->fs_info->super_copy));
1381
1382         trans->transaction->blocked = 0;
1383
1384         wake_up(&root->fs_info->transaction_wait);
1385
1386         mutex_unlock(&root->fs_info->trans_mutex);
1387         ret = btrfs_write_and_wait_transaction(trans, root);
1388         BUG_ON(ret);
1389         write_ctree_super(trans, root, 0);
1390
1391         /*
1392          * the super is written, we can safely allow the tree-loggers
1393          * to go about their business
1394          */
1395         mutex_unlock(&root->fs_info->tree_log_mutex);
1396
1397         btrfs_finish_extent_commit(trans, root);
1398
1399         mutex_lock(&root->fs_info->trans_mutex);
1400
1401         cur_trans->commit_done = 1;
1402
1403         root->fs_info->last_trans_committed = cur_trans->transid;
1404
1405         wake_up(&cur_trans->commit_wait);
1406
1407         list_del_init(&cur_trans->list);
1408         put_transaction(cur_trans);
1409         put_transaction(cur_trans);
1410
1411         trace_btrfs_transaction_commit(root);
1412
1413         mutex_unlock(&root->fs_info->trans_mutex);
1414
1415         if (current->journal_info == trans)
1416                 current->journal_info = NULL;
1417
1418         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1419
1420         if (current != root->fs_info->transaction_kthread)
1421                 btrfs_run_delayed_iputs(root);
1422
1423         return ret;
1424 }
1425
1426 /*
1427  * interface function to delete all the snapshots we have scheduled for deletion
1428  */
1429 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1430 {
1431         LIST_HEAD(list);
1432         struct btrfs_fs_info *fs_info = root->fs_info;
1433
1434         mutex_lock(&fs_info->trans_mutex);
1435         list_splice_init(&fs_info->dead_roots, &list);
1436         mutex_unlock(&fs_info->trans_mutex);
1437
1438         while (!list_empty(&list)) {
1439                 root = list_entry(list.next, struct btrfs_root, root_list);
1440                 list_del(&root->root_list);
1441
1442                 if (btrfs_header_backref_rev(root->node) <
1443                     BTRFS_MIXED_BACKREF_REV)
1444                         btrfs_drop_snapshot(root, NULL, 0);
1445                 else
1446                         btrfs_drop_snapshot(root, NULL, 1);
1447         }
1448         return 0;
1449 }