fs/btrfs/transaction.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/fs.h>
  20 #include <linux/slab.h>
  21 #include <linux/sched.h>
  22 #include <linux/writeback.h>
  23 #include <linux/pagemap.h>
  24 #include <linux/blkdev.h>
  25 #include "ctree.h"
  26 #include "disk-io.h"
  27 #include "transaction.h"
  28 #include "locking.h"
  29 #include "tree-log.h"
  30 #include "inode-map.h"
  31
  32 #define BTRFS_ROOT_TRANS_TAG 0
  33
  34 static noinline void put_transaction(struct btrfs_transaction *transaction)
  35 {
  36         WARN_ON(atomic_read(&transaction->use_count) == 0);
  37         if (atomic_dec_and_test(&transaction->use_count)) {
  38                 BUG_ON(!list_empty(&transaction->list));
  39                 memset(transaction, 0, sizeof(*transaction));
  40                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  41         }
  42 }
  43
  44 static noinline void switch_commit_root(struct btrfs_root *root)
  45 {
  46         free_extent_buffer(root->commit_root);
  47         root->commit_root = btrfs_root_node(root);
  48 }
  49
  50 /*
  51  * either allocate a new transaction or hop into the existing one
  52  */
  53 static noinline int join_transaction(struct btrfs_root *root, int nofail)
  54 {
  55         struct btrfs_transaction *cur_trans;
  56
  57         spin_lock(&root->fs_info->trans_lock);
  58         if (root->fs_info->trans_no_join) {
  59                 if (!nofail) {
  60                         spin_unlock(&root->fs_info->trans_lock);
  61                         return -EBUSY;
  62                 }
  63         }
  64
  65         cur_trans = root->fs_info->running_transaction;
  66         if (cur_trans) {
  67                 atomic_inc(&cur_trans->use_count);
  68                 atomic_inc(&cur_trans->num_writers);
  69                 cur_trans->num_joined++;
  70                 spin_unlock(&root->fs_info->trans_lock);
  71                 return 0;
  72         }
  73         spin_unlock(&root->fs_info->trans_lock);
  74
  75         cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
  76         if (!cur_trans)
  77                 return -ENOMEM;
  78         spin_lock(&root->fs_info->trans_lock);
  79         if (root->fs_info->running_transaction) {
  80                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
  81                 cur_trans = root->fs_info->running_transaction;
  82                 atomic_inc(&cur_trans->use_count);
  83                 atomic_inc(&cur_trans->num_writers);
  84                 cur_trans->num_joined++;
  85                 spin_unlock(&root->fs_info->trans_lock);
  86                 return 0;
  87         }
  88         atomic_set(&cur_trans->num_writers, 1);
  89         cur_trans->num_joined = 0;
  90         init_waitqueue_head(&cur_trans->writer_wait);
  91         init_waitqueue_head(&cur_trans->commit_wait);
  92         cur_trans->in_commit = 0;
  93         cur_trans->blocked = 0;
  94         /*
  95          * One for this trans handle, one so it will live on until we
  96          * commit the transaction.
  97          */
  98         atomic_set(&cur_trans->use_count, 2);
  99         cur_trans->commit_done = 0;
 100         cur_trans->start_time = get_seconds();
 101
 102         cur_trans->delayed_refs.root = RB_ROOT;
 103         cur_trans->delayed_refs.num_entries = 0;
 104         cur_trans->delayed_refs.num_heads_ready = 0;
 105         cur_trans->delayed_refs.num_heads = 0;
 106         cur_trans->delayed_refs.flushing = 0;
 107         cur_trans->delayed_refs.run_delayed_start = 0;
 108         spin_lock_init(&cur_trans->commit_lock);
 109         spin_lock_init(&cur_trans->delayed_refs.lock);
 110
 111         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 112         list_add_tail(&cur_trans->list, &root->fs_info->trans_list);
 113         extent_io_tree_init(&cur_trans->dirty_pages,
 114                              root->fs_info->btree_inode->i_mapping);
 115         root->fs_info->generation++;
 116         cur_trans->transid = root->fs_info->generation;
 117         root->fs_info->running_transaction = cur_trans;
 118         spin_unlock(&root->fs_info->trans_lock);
 119
 120         return 0;
 121 }
 122
 123 /*
 124  * this does all the record keeping required to make sure that a reference
 125  * counted root is properly recorded in a given transaction.  This is required
 126  * to make sure the old root from before we joined the transaction is deleted
 127  * when the transaction commits
 128  */
 129 static int record_root_in_trans(struct btrfs_trans_handle *trans,
 130                                struct btrfs_root *root)
 131 {
 132         if (root->ref_cows && root->last_trans < trans->transid) {
 133                 WARN_ON(root == root->fs_info->extent_root);
 134                 WARN_ON(root->commit_root != root->node);
 135
 136                 /*
 137                  * see below for in_trans_setup usage rules
 138                  * we have the reloc mutex held now, so there
 139                  * is only one writer in this function
 140                  */
 141                 root->in_trans_setup = 1;
 142
 143                 /* make sure readers find in_trans_setup before
 144                  * they find our root->last_trans update
 145                  */
 146                 smp_wmb();
 147
 148                 spin_lock(&root->fs_info->fs_roots_radix_lock);
 149                 if (root->last_trans == trans->transid) {
 150                         spin_unlock(&root->fs_info->fs_roots_radix_lock);
 151                         return 0;
 152                 }
 153                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 154                            (unsigned long)root->root_key.objectid,
 155                            BTRFS_ROOT_TRANS_TAG);
 156                 spin_unlock(&root->fs_info->fs_roots_radix_lock);
 157                 root->last_trans = trans->transid;
 158
 159                 /* this is pretty tricky.  We don't want to
 160                  * take the relocation lock in btrfs_record_root_in_trans
 161                  * unless we're really doing the first setup for this root in
 162                  * this transaction.
 163                  *
 164                  * Normally we'd use root->last_trans as a flag to decide
 165                  * if we want to take the expensive mutex.
 166                  *
 167                  * But, we have to set root->last_trans before we
 168                  * init the relocation root, otherwise, we trip over warnings
 169                  * in ctree.c.  The solution used here is to flag ourselves
 170                  * with root->in_trans_setup.  When this is 1, we're still
 171                  * fixing up the reloc trees and everyone must wait.
 172                  *
 173                  * When this is zero, they can trust root->last_trans and fly
 174                  * through btrfs_record_root_in_trans without having to take the
 175                  * lock.  smp_wmb() makes sure that all the writes above are
 176                  * done before we pop in the zero below
 177                  */
 178                 btrfs_init_reloc_root(trans, root);
 179                 smp_wmb();
 180                 root->in_trans_setup = 0;
 181         }
 182         return 0;
 183 }
 184
 185
 186 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 187                                struct btrfs_root *root)
 188 {
 189         if (!root->ref_cows)
 190                 return 0;
 191
 192         /*
 193          * see record_root_in_trans for comments about in_trans_setup usage
 194          * and barriers
 195          */
 196         smp_rmb();
 197         if (root->last_trans == trans->transid &&
 198             !root->in_trans_setup)
 199                 return 0;
 200
 201         mutex_lock(&root->fs_info->reloc_mutex);
 202         record_root_in_trans(trans, root);
 203         mutex_unlock(&root->fs_info->reloc_mutex);
 204
 205         return 0;
 206 }
 207
 208 /* wait for commit against the current transaction to become unblocked
 209  * when this is done, it is safe to start a new transaction, but the current
 210  * transaction might not be fully on disk.
 211  */
 212 static void wait_current_trans(struct btrfs_root *root)
 213 {
 214         struct btrfs_transaction *cur_trans;
 215
 216         spin_lock(&root->fs_info->trans_lock);
 217         cur_trans = root->fs_info->running_transaction;
 218         if (cur_trans && cur_trans->blocked) {
 219                 DEFINE_WAIT(wait);
 220                 atomic_inc(&cur_trans->use_count);
 221                 spin_unlock(&root->fs_info->trans_lock);
 222                 while (1) {
 223                         prepare_to_wait(&root->fs_info->transaction_wait, &wait,
 224                                         TASK_UNINTERRUPTIBLE);
 225                         if (!cur_trans->blocked)
 226                                 break;
 227                         schedule();
 228                 }
 229                 finish_wait(&root->fs_info->transaction_wait, &wait);
 230                 put_transaction(cur_trans);
 231         } else {
 232                 spin_unlock(&root->fs_info->trans_lock);
 233         }
 234 }
 235
 236 enum btrfs_trans_type {
 237         TRANS_START,
 238         TRANS_JOIN,
 239         TRANS_USERSPACE,
 240         TRANS_JOIN_NOLOCK,
 241 };
 242
 243 static int may_wait_transaction(struct btrfs_root *root, int type)
 244 {
 245         if (root->fs_info->log_root_recovering)
 246                 return 0;
 247
 248         if (type == TRANS_USERSPACE)
 249                 return 1;
 250
 251         if (type == TRANS_START &&
 252             !atomic_read(&root->fs_info->open_ioctl_trans))
 253                 return 1;
 254
 255         return 0;
 256 }
 257
 258 static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
 259                                                     u64 num_items, int type)
 260 {
 261         struct btrfs_trans_handle *h;
 262         struct btrfs_transaction *cur_trans;
 263         u64 num_bytes = 0;
 264         int ret;
 265
 266         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR)
 267                 return ERR_PTR(-EROFS);
 268
 269         if (current->journal_info) {
 270                 WARN_ON(type != TRANS_JOIN && type != TRANS_JOIN_NOLOCK);
 271                 h = current->journal_info;
 272                 h->use_count++;
 273                 h->orig_rsv = h->block_rsv;
 274                 h->block_rsv = NULL;
 275                 goto got_it;
 276         }
 277
 278         /*
 279          * Do the reservation before we join the transaction so we can do all
 280          * the appropriate flushing if need be.
 281          */
 282         if (num_items > 0 && root != root->fs_info->chunk_root) {
 283                 num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
 284                 ret = btrfs_block_rsv_add(NULL, root,
 285                                           &root->fs_info->trans_block_rsv,
 286                                           num_bytes);
 287                 if (ret)
 288                         return ERR_PTR(ret);
 289         }
 290 again:
 291         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 292         if (!h)
 293                 return ERR_PTR(-ENOMEM);
 294
 295         if (may_wait_transaction(root, type))
 296                 wait_current_trans(root);
 297
 298         do {
 299                 ret = join_transaction(root, type == TRANS_JOIN_NOLOCK);
 300                 if (ret == -EBUSY)
 301                         wait_current_trans(root);
 302         } while (ret == -EBUSY);
 303
 304         if (ret < 0) {
 305                 kmem_cache_free(btrfs_trans_handle_cachep, h);
 306                 return ERR_PTR(ret);
 307         }
 308
 309         cur_trans = root->fs_info->running_transaction;
 310
 311         h->transid = cur_trans->transid;
 312         h->transaction = cur_trans;
 313         h->blocks_used = 0;
 314         h->bytes_reserved = 0;
 315         h->delayed_ref_updates = 0;
 316         h->use_count = 1;
 317         h->block_rsv = NULL;
 318         h->orig_rsv = NULL;
 319
 320         smp_mb();
 321         if (cur_trans->blocked && may_wait_transaction(root, type)) {
 322                 btrfs_commit_transaction(h, root);
 323                 goto again;
 324         }
 325
 326         if (num_bytes) {
 327                 h->block_rsv = &root->fs_info->trans_block_rsv;
 328                 h->bytes_reserved = num_bytes;
 329         }
 330
 331 got_it:
 332         btrfs_record_root_in_trans(h, root);
 333
 334         if (!current->journal_info && type != TRANS_USERSPACE)
 335                 current->journal_info = h;
 336         return h;
 337 }
 338
 339 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 340                                                    int num_items)
 341 {
 342         return start_transaction(root, num_items, TRANS_START);
 343 }
 344 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 345 {
 346         return start_transaction(root, 0, TRANS_JOIN);
 347 }
 348
 349 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 350 {
 351         return start_transaction(root, 0, TRANS_JOIN_NOLOCK);
 352 }
 353
 354 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 355 {
 356         return start_transaction(root, 0, TRANS_USERSPACE);
 357 }
 358
 359 /* wait for a transaction commit to be fully complete */
 360 static noinline int wait_for_commit(struct btrfs_root *root,
 361                                     struct btrfs_transaction *commit)
 362 {
 363         DEFINE_WAIT(wait);
 364         while (!commit->commit_done) {
 365                 prepare_to_wait(&commit->commit_wait, &wait,
 366                                 TASK_UNINTERRUPTIBLE);
 367                 if (commit->commit_done)
 368                         break;
 369                 schedule();
 370         }
 371         finish_wait(&commit->commit_wait, &wait);
 372         return 0;
 373 }
 374
 375 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 376 {
 377         struct btrfs_transaction *cur_trans = NULL, *t;
 378         int ret;
 379
 380         ret = 0;
 381         if (transid) {
 382                 if (transid <= root->fs_info->last_trans_committed)
 383                         goto out;
 384
 385                 /* find specified transaction */
 386                 spin_lock(&root->fs_info->trans_lock);
 387                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 388                         if (t->transid == transid) {
 389                                 cur_trans = t;
 390                                 atomic_inc(&cur_trans->use_count);
 391                                 break;
 392                         }
 393                         if (t->transid > transid)
 394                                 break;
 395                 }
 396                 spin_unlock(&root->fs_info->trans_lock);
 397                 ret = -EINVAL;
 398                 if (!cur_trans)
 399                         goto out;  /* bad transid */
 400         } else {
 401                 /* find newest transaction that is committing | committed */
 402                 spin_lock(&root->fs_info->trans_lock);
 403                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 404                                             list) {
 405                         if (t->in_commit) {
 406                                 if (t->commit_done)
 407                                         break;
 408                                 cur_trans = t;
 409                                 atomic_inc(&cur_trans->use_count);
 410                                 break;
 411                         }
 412                 }
 413                 spin_unlock(&root->fs_info->trans_lock);
 414                 if (!cur_trans)
 415                         goto out;  /* nothing committing|committed */
 416         }
 417
 418         wait_for_commit(root, cur_trans);
 419
 420         put_transaction(cur_trans);
 421         ret = 0;
 422 out:
 423         return ret;
 424 }
 425
 426 void btrfs_throttle(struct btrfs_root *root)
 427 {
 428         if (!atomic_read(&root->fs_info->open_ioctl_trans))
 429                 wait_current_trans(root);
 430 }
 431
 432 static int should_end_transaction(struct btrfs_trans_handle *trans,
 433                                   struct btrfs_root *root)
 434 {
 435         int ret;
 436         ret = btrfs_block_rsv_check(trans, root,
 437                                     &root->fs_info->global_block_rsv, 0, 5);
 438         return ret ? 1 : 0;
 439 }
 440
 441 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 442                                  struct btrfs_root *root)
 443 {
 444         struct btrfs_transaction *cur_trans = trans->transaction;
 445         int updates;
 446
 447         smp_mb();
 448         if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
 449                 return 1;
 450
 451         updates = trans->delayed_ref_updates;
 452         trans->delayed_ref_updates = 0;
 453         if (updates)
 454                 btrfs_run_delayed_refs(trans, root, updates);
 455
 456         return should_end_transaction(trans, root);
 457 }
 458
 459 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 460                           struct btrfs_root *root, int throttle, int lock)
 461 {
 462         struct btrfs_transaction *cur_trans = trans->transaction;
 463         struct btrfs_fs_info *info = root->fs_info;
 464         int count = 0;
 465
 466         if (--trans->use_count) {
 467                 trans->block_rsv = trans->orig_rsv;
 468                 return 0;
 469         }
 470
 471         while (count < 4) {
 472                 unsigned long cur = trans->delayed_ref_updates;
 473                 trans->delayed_ref_updates = 0;
 474                 if (cur &&
 475                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 476                         trans->delayed_ref_updates = 0;
 477
 478                         /*
 479                          * do a full flush if the transaction is trying
 480                          * to close
 481                          */
 482                         if (trans->transaction->delayed_refs.flushing)
 483                                 cur = 0;
 484                         btrfs_run_delayed_refs(trans, root, cur);
 485                 } else {
 486                         break;
 487                 }
 488                 count++;
 489         }
 490
 491         btrfs_trans_release_metadata(trans, root);
 492
 493         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
 494             should_end_transaction(trans, root)) {
 495                 trans->transaction->blocked = 1;
 496                 smp_wmb();
 497         }
 498
 499         if (lock && cur_trans->blocked && !cur_trans->in_commit) {
 500                 if (throttle) {
 501                         /*
 502                          * We may race with somebody else here so end up having
 503                          * to call end_transaction on ourselves again, so inc
 504                          * our use_count.
 505                          */
 506                         trans->use_count++;
 507                         return btrfs_commit_transaction(trans, root);
 508                 } else {
 509                         wake_up_process(info->transaction_kthread);
 510                 }
 511         }
 512
 513         WARN_ON(cur_trans != info->running_transaction);
 514         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 515         atomic_dec(&cur_trans->num_writers);
 516
 517         smp_mb();
 518         if (waitqueue_active(&cur_trans->writer_wait))
 519                 wake_up(&cur_trans->writer_wait);
 520         put_transaction(cur_trans);
 521
 522         if (current->journal_info == trans)
 523                 current->journal_info = NULL;
 524         memset(trans, 0, sizeof(*trans));
 525         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 526
 527         if (throttle)
 528                 btrfs_run_delayed_iputs(root);
 529
 530         return 0;
 531 }
 532
 533 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 534                           struct btrfs_root *root)
 535 {
 536         int ret;
 537
 538         ret = __btrfs_end_transaction(trans, root, 0, 1);
 539         if (ret)
 540                 return ret;
 541         return 0;
 542 }
 543
 544 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 545                                    struct btrfs_root *root)
 546 {
 547         int ret;
 548
 549         ret = __btrfs_end_transaction(trans, root, 1, 1);
 550         if (ret)
 551                 return ret;
 552         return 0;
 553 }
 554
 555 int btrfs_end_transaction_nolock(struct btrfs_trans_handle *trans,
 556                                  struct btrfs_root *root)
 557 {
 558         int ret;
 559
 560         ret = __btrfs_end_transaction(trans, root, 0, 0);
 561         if (ret)
 562                 return ret;
 563         return 0;
 564 }
 565
 566 int btrfs_end_transaction_dmeta(struct btrfs_trans_handle *trans,
 567                                 struct btrfs_root *root)
 568 {
 569         return __btrfs_end_transaction(trans, root, 1, 1);
 570 }
 571
 572 /*
 573  * when btree blocks are allocated, they have some corresponding bits set for
 574  * them in one of two extent_io trees.  This is used to make sure all of
 575  * those extents are sent to disk but does not wait on them
 576  */
 577 int btrfs_write_marked_extents(struct btrfs_root *root,
 578                                struct extent_io_tree *dirty_pages, int mark)
 579 {
 580         int ret;
 581         int err = 0;
 582         int werr = 0;
 583         struct page *page;
 584         struct inode *btree_inode = root->fs_info->btree_inode;
 585         u64 start = 0;
 586         u64 end;
 587         unsigned long index;
 588
 589         while (1) {
 590                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 591                                             mark);
 592                 if (ret)
 593                         break;
 594                 while (start <= end) {
 595                         cond_resched();
 596
 597                         index = start >> PAGE_CACHE_SHIFT;
 598                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 599                         page = find_get_page(btree_inode->i_mapping, index);
 600                         if (!page)
 601                                 continue;
 602
 603                         btree_lock_page_hook(page);
 604                         if (!page->mapping) {
 605                                 unlock_page(page);
 606                                 page_cache_release(page);
 607                                 continue;
 608                         }
 609
 610                         if (PageWriteback(page)) {
 611                                 if (PageDirty(page))
 612                                         wait_on_page_writeback(page);
 613                                 else {
 614                                         unlock_page(page);
 615                                         page_cache_release(page);
 616                                         continue;
 617                                 }
 618                         }
 619                         err = write_one_page(page, 0);
 620                         if (err)
 621                                 werr = err;
 622                         page_cache_release(page);
 623                 }
 624         }
 625         if (err)
 626                 werr = err;
 627         return werr;
 628 }
 629
 630 /*
 631  * when btree blocks are allocated, they have some corresponding bits set for
 632  * them in one of two extent_io trees.  This is used to make sure all of
 633  * those extents are on disk for transaction or log commit.  We wait
 634  * on all the pages and clear them from the dirty pages state tree
 635  */
 636 int btrfs_wait_marked_extents(struct btrfs_root *root,
 637                               struct extent_io_tree *dirty_pages, int mark)
 638 {
 639         int ret;
 640         int err = 0;
 641         int werr = 0;
 642         struct page *page;
 643         struct inode *btree_inode = root->fs_info->btree_inode;
 644         u64 start = 0;
 645         u64 end;
 646         unsigned long index;
 647
 648         while (1) {
 649                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 650                                             mark);
 651                 if (ret)
 652                         break;
 653
 654                 clear_extent_bits(dirty_pages, start, end, mark, GFP_NOFS);
 655                 while (start <= end) {
 656                         index = start >> PAGE_CACHE_SHIFT;
 657                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 658                         page = find_get_page(btree_inode->i_mapping, index);
 659                         if (!page)
 660                                 continue;
 661                         if (PageDirty(page)) {
 662                                 btree_lock_page_hook(page);
 663                                 wait_on_page_writeback(page);
 664                                 err = write_one_page(page, 0);
 665                                 if (err)
 666                                         werr = err;
 667                         }
 668                         wait_on_page_writeback(page);
 669                         page_cache_release(page);
 670                         cond_resched();
 671                 }
 672         }
 673         if (err)
 674                 werr = err;
 675         return werr;
 676 }
 677
 678 /*
 679  * when btree blocks are allocated, they have some corresponding bits set for
 680  * them in one of two extent_io trees.  This is used to make sure all of
 681  * those extents are on disk for transaction or log commit
 682  */
 683 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 684                                 struct extent_io_tree *dirty_pages, int mark)
 685 {
 686         int ret;
 687         int ret2;
 688
 689         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 690         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 691         return ret || ret2;
 692 }
 693
 694 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 695                                      struct btrfs_root *root)
 696 {
 697         if (!trans || !trans->transaction) {
 698                 struct inode *btree_inode;
 699                 btree_inode = root->fs_info->btree_inode;
 700                 return filemap_write_and_wait(btree_inode->i_mapping);
 701         }
 702         return btrfs_write_and_wait_marked_extents(root,
 703                                            &trans->transaction->dirty_pages,
 704                                            EXTENT_DIRTY);
 705 }
 706
 707 /*
 708  * this is used to update the root pointer in the tree of tree roots.
 709  *
 710  * But, in the case of the extent allocation tree, updating the root
 711  * pointer may allocate blocks which may change the root of the extent
 712  * allocation tree.
 713  *
 714  * So, this loops and repeats and makes sure the cowonly root didn't
 715  * change while the root pointer was being updated in the metadata.
 716  */
 717 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 718                                struct btrfs_root *root)
 719 {
 720         int ret;
 721         u64 old_root_bytenr;
 722         u64 old_root_used;
 723         struct btrfs_root *tree_root = root->fs_info->tree_root;
 724
 725         old_root_used = btrfs_root_used(&root->root_item);
 726         btrfs_write_dirty_block_groups(trans, root);
 727
 728         while (1) {
 729                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 730                 if (old_root_bytenr == root->node->start &&
 731                     old_root_used == btrfs_root_used(&root->root_item))
 732                         break;
 733
 734                 btrfs_set_root_node(&root->root_item, root->node);
 735                 ret = btrfs_update_root(trans, tree_root,
 736                                         &root->root_key,
 737                                         &root->root_item);
 738                 BUG_ON(ret);
 739
 740                 old_root_used = btrfs_root_used(&root->root_item);
 741                 ret = btrfs_write_dirty_block_groups(trans, root);
 742                 BUG_ON(ret);
 743         }
 744
 745         if (root != root->fs_info->extent_root)
 746                 switch_commit_root(root);
 747
 748         return 0;
 749 }
 750
 751 /*
 752  * update all the cowonly tree roots on disk
 753  */
 754 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 755                                          struct btrfs_root *root)
 756 {
 757         struct btrfs_fs_info *fs_info = root->fs_info;
 758         struct list_head *next;
 759         struct extent_buffer *eb;
 760         int ret;
 761
 762         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 763         BUG_ON(ret);
 764
 765         eb = btrfs_lock_root_node(fs_info->tree_root);
 766         btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
 767         btrfs_tree_unlock(eb);
 768         free_extent_buffer(eb);
 769
 770         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 771         BUG_ON(ret);
 772
 773         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 774                 next = fs_info->dirty_cowonly_roots.next;
 775                 list_del_init(next);
 776                 root = list_entry(next, struct btrfs_root, dirty_list);
 777
 778                 update_cowonly_root(trans, root);
 779         }
 780
 781         down_write(&fs_info->extent_commit_sem);
 782         switch_commit_root(fs_info->extent_root);
 783         up_write(&fs_info->extent_commit_sem);
 784
 785         return 0;
 786 }
 787
 788 /*
 789  * dead roots are old snapshots that need to be deleted.  This allocates
 790  * a dirty root struct and adds it into the list of dead roots that need to
 791  * be deleted
 792  */
 793 int btrfs_add_dead_root(struct btrfs_root *root)
 794 {
 795         spin_lock(&root->fs_info->trans_lock);
 796         list_add(&root->root_list, &root->fs_info->dead_roots);
 797         spin_unlock(&root->fs_info->trans_lock);
 798         return 0;
 799 }
 800
 801 /*
 802  * update all the cowonly tree roots on disk
 803  */
 804 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 805                                     struct btrfs_root *root)
 806 {
 807         struct btrfs_root *gang[8];
 808         struct btrfs_fs_info *fs_info = root->fs_info;
 809         int i;
 810         int ret;
 811         int err = 0;
 812
 813         spin_lock(&fs_info->fs_roots_radix_lock);
 814         while (1) {
 815                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 816                                                  (void **)gang, 0,
 817                                                  ARRAY_SIZE(gang),
 818                                                  BTRFS_ROOT_TRANS_TAG);
 819                 if (ret == 0)
 820                         break;
 821                 for (i = 0; i < ret; i++) {
 822                         root = gang[i];
 823                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 824                                         (unsigned long)root->root_key.objectid,
 825                                         BTRFS_ROOT_TRANS_TAG);
 826                         spin_unlock(&fs_info->fs_roots_radix_lock);
 827
 828                         btrfs_free_log(trans, root);
 829                         btrfs_update_reloc_root(trans, root);
 830                         btrfs_orphan_commit_root(trans, root);
 831
 832                         btrfs_save_ino_cache(root, trans);
 833
 834                         if (root->commit_root != root->node) {
 835                                 mutex_lock(&root->fs_commit_mutex);
 836                                 switch_commit_root(root);
 837                                 btrfs_unpin_free_ino(root);
 838                                 mutex_unlock(&root->fs_commit_mutex);
 839
 840                                 btrfs_set_root_node(&root->root_item,
 841                                                     root->node);
 842                         }
 843
 844                         err = btrfs_update_root(trans, fs_info->tree_root,
 845                                                 &root->root_key,
 846                                                 &root->root_item);
 847                         spin_lock(&fs_info->fs_roots_radix_lock);
 848                         if (err)
 849                                 break;
 850                 }
 851         }
 852         spin_unlock(&fs_info->fs_roots_radix_lock);
 853         return err;
 854 }
 855
 856 /*
 857  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 858  * otherwise every leaf in the btree is read and defragged.
 859  */
 860 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 861 {
 862         struct btrfs_fs_info *info = root->fs_info;
 863         struct btrfs_trans_handle *trans;
 864         int ret;
 865         unsigned long nr;
 866
 867         if (xchg(&root->defrag_running, 1))
 868                 return 0;
 869
 870         while (1) {
 871                 trans = btrfs_start_transaction(root, 0);
 872                 if (IS_ERR(trans))
 873                         return PTR_ERR(trans);
 874
 875                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 876
 877                 nr = trans->blocks_used;
 878                 btrfs_end_transaction(trans, root);
 879                 btrfs_btree_balance_dirty(info->tree_root, nr);
 880                 cond_resched();
 881
 882                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
 883                         break;
 884         }
 885         root->defrag_running = 0;
 886         return ret;
 887 }
 888
 889 /*
 890  * new snapshots need to be created at a very specific time in the
 891  * transaction commit.  This does the actual creation
 892  */
 893 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 894                                    struct btrfs_fs_info *fs_info,
 895                                    struct btrfs_pending_snapshot *pending)
 896 {
 897         struct btrfs_key key;
 898         struct btrfs_root_item *new_root_item;
 899         struct btrfs_root *tree_root = fs_info->tree_root;
 900         struct btrfs_root *root = pending->root;
 901         struct btrfs_root *parent_root;
 902         struct inode *parent_inode;
 903         struct dentry *parent;
 904         struct dentry *dentry;
 905         struct extent_buffer *tmp;
 906         struct extent_buffer *old;
 907         int ret;
 908         u64 to_reserve = 0;
 909         u64 index = 0;
 910         u64 objectid;
 911         u64 root_flags;
 912
 913         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 914         if (!new_root_item) {
 915                 pending->error = -ENOMEM;
 916                 goto fail;
 917         }
 918
 919         ret = btrfs_find_free_objectid(tree_root, &objectid);
 920         if (ret) {
 921                 pending->error = ret;
 922                 goto fail;
 923         }
 924
 925         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
 926         btrfs_orphan_pre_snapshot(trans, pending, &to_reserve);
 927
 928         if (to_reserve > 0) {
 929                 ret = btrfs_block_rsv_add(trans, root, &pending->block_rsv,
 930                                           to_reserve);
 931                 if (ret) {
 932                         pending->error = ret;
 933                         goto fail;
 934                 }
 935         }
 936
 937         key.objectid = objectid;
 938         key.offset = (u64)-1;
 939         key.type = BTRFS_ROOT_ITEM_KEY;
 940
 941         trans->block_rsv = &pending->block_rsv;
 942
 943         dentry = pending->dentry;
 944         parent = dget_parent(dentry);
 945         parent_inode = parent->d_inode;
 946         parent_root = BTRFS_I(parent_inode)->root;
 947         record_root_in_trans(trans, parent_root);
 948
 949         /*
 950          * insert the directory item
 951          */
 952         ret = btrfs_set_inode_index(parent_inode, &index);
 953         BUG_ON(ret);
 954         ret = btrfs_insert_dir_item(trans, parent_root,
 955                                 dentry->d_name.name, dentry->d_name.len,
 956                                 parent_inode, &key,
 957                                 BTRFS_FT_DIR, index);
 958         BUG_ON(ret);
 959
 960         btrfs_i_size_write(parent_inode, parent_inode->i_size +
 961                                          dentry->d_name.len * 2);
 962         ret = btrfs_update_inode(trans, parent_root, parent_inode);
 963         BUG_ON(ret);
 964
 965         /*
 966          * pull in the delayed directory update
 967          * and the delayed inode item
 968          * otherwise we corrupt the FS during
 969          * snapshot
 970          */
 971         ret = btrfs_run_delayed_items(trans, root);
 972         BUG_ON(ret);
 973
 974         record_root_in_trans(trans, root);
 975         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
 976         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
 977         btrfs_check_and_init_root_item(new_root_item);
 978
 979         root_flags = btrfs_root_flags(new_root_item);
 980         if (pending->readonly)
 981                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
 982         else
 983                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
 984         btrfs_set_root_flags(new_root_item, root_flags);
 985
 986         old = btrfs_lock_root_node(root);
 987         btrfs_cow_block(trans, root, old, NULL, 0, &old);
 988         btrfs_set_lock_blocking(old);
 989
 990         btrfs_copy_root(trans, root, old, &tmp, objectid);
 991         btrfs_tree_unlock(old);
 992         free_extent_buffer(old);
 993
 994         btrfs_set_root_node(new_root_item, tmp);
 995         /* record when the snapshot was created in key.offset */
 996         key.offset = trans->transid;
 997         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
 998         btrfs_tree_unlock(tmp);
 999         free_extent_buffer(tmp);
1000         BUG_ON(ret);
1001
1002         /*
1003          * insert root back/forward references
1004          */
1005         ret = btrfs_add_root_ref(trans, tree_root, objectid,
1006                                  parent_root->root_key.objectid,
1007                                  btrfs_ino(parent_inode), index,
1008                                  dentry->d_name.name, dentry->d_name.len);
1009         BUG_ON(ret);
1010         dput(parent);
1011
1012         key.offset = (u64)-1;
1013         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1014         BUG_ON(IS_ERR(pending->snap));
1015
1016         btrfs_reloc_post_snapshot(trans, pending);
1017         btrfs_orphan_post_snapshot(trans, pending);
1018 fail:
1019         kfree(new_root_item);
1020         btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
1021         return 0;
1022 }
1023
1024 /*
1025  * create all the snapshots we've scheduled for creation
1026  */
1027 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1028                                              struct btrfs_fs_info *fs_info)
1029 {
1030         struct btrfs_pending_snapshot *pending;
1031         struct list_head *head = &trans->transaction->pending_snapshots;
1032         int ret;
1033
1034         list_for_each_entry(pending, head, list) {
1035                 ret = create_pending_snapshot(trans, fs_info, pending);
1036                 BUG_ON(ret);
1037         }
1038         return 0;
1039 }
1040
1041 static void update_super_roots(struct btrfs_root *root)
1042 {
1043         struct btrfs_root_item *root_item;
1044         struct btrfs_super_block *super;
1045
1046         super = &root->fs_info->super_copy;
1047
1048         root_item = &root->fs_info->chunk_root->root_item;
1049         super->chunk_root = root_item->bytenr;
1050         super->chunk_root_generation = root_item->generation;
1051         super->chunk_root_level = root_item->level;
1052
1053         root_item = &root->fs_info->tree_root->root_item;
1054         super->root = root_item->bytenr;
1055         super->generation = root_item->generation;
1056         super->root_level = root_item->level;
1057         if (super->cache_generation != 0 || btrfs_test_opt(root, SPACE_CACHE))
1058                 super->cache_generation = root_item->generation;
1059 }
1060
1061 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1062 {
1063         int ret = 0;
1064         spin_lock(&info->trans_lock);
1065         if (info->running_transaction)
1066                 ret = info->running_transaction->in_commit;
1067         spin_unlock(&info->trans_lock);
1068         return ret;
1069 }
1070
1071 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1072 {
1073         int ret = 0;
1074         spin_lock(&info->trans_lock);
1075         if (info->running_transaction)
1076                 ret = info->running_transaction->blocked;
1077         spin_unlock(&info->trans_lock);
1078         return ret;
1079 }
1080
1081 /*
1082  * wait for the current transaction commit to start and block subsequent
1083  * transaction joins
1084  */
1085 static void wait_current_trans_commit_start(struct btrfs_root *root,
1086                                             struct btrfs_transaction *trans)
1087 {
1088         DEFINE_WAIT(wait);
1089
1090         if (trans->in_commit)
1091                 return;
1092
1093         while (1) {
1094                 prepare_to_wait(&root->fs_info->transaction_blocked_wait, &wait,
1095                                 TASK_UNINTERRUPTIBLE);
1096                 if (trans->in_commit) {
1097                         finish_wait(&root->fs_info->transaction_blocked_wait,
1098                                     &wait);
1099                         break;
1100                 }
1101                 schedule();
1102                 finish_wait(&root->fs_info->transaction_blocked_wait, &wait);
1103         }
1104 }
1105
1106 /*
1107  * wait for the current transaction to start and then become unblocked.
1108  * caller holds ref.
1109  */
1110 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1111                                          struct btrfs_transaction *trans)
1112 {
1113         DEFINE_WAIT(wait);
1114
1115         if (trans->commit_done || (trans->in_commit && !trans->blocked))
1116                 return;
1117
1118         while (1) {
1119                 prepare_to_wait(&root->fs_info->transaction_wait, &wait,
1120                                 TASK_UNINTERRUPTIBLE);
1121                 if (trans->commit_done ||
1122                     (trans->in_commit && !trans->blocked)) {
1123                         finish_wait(&root->fs_info->transaction_wait,
1124                                     &wait);
1125                         break;
1126                 }
1127                 schedule();
1128                 finish_wait(&root->fs_info->transaction_wait,
1129                             &wait);
1130         }
1131 }
1132
1133 /*
1134  * commit transactions asynchronously. once btrfs_commit_transaction_async
1135  * returns, any subsequent transaction will not be allowed to join.
1136  */
1137 struct btrfs_async_commit {
1138         struct btrfs_trans_handle *newtrans;
1139         struct btrfs_root *root;
1140         struct delayed_work work;
1141 };
1142
1143 static void do_async_commit(struct work_struct *work)
1144 {
1145         struct btrfs_async_commit *ac =
1146                 container_of(work, struct btrfs_async_commit, work.work);
1147
1148         btrfs_commit_transaction(ac->newtrans, ac->root);
1149         kfree(ac);
1150 }
1151
1152 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1153                                    struct btrfs_root *root,
1154                                    int wait_for_unblock)
1155 {
1156         struct btrfs_async_commit *ac;
1157         struct btrfs_transaction *cur_trans;
1158
1159         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1160         if (!ac)
1161                 return -ENOMEM;
1162
1163         INIT_DELAYED_WORK(&ac->work, do_async_commit);
1164         ac->root = root;
1165         ac->newtrans = btrfs_join_transaction(root);
1166         if (IS_ERR(ac->newtrans)) {
1167                 int err = PTR_ERR(ac->newtrans);
1168                 kfree(ac);
1169                 return err;
1170         }
1171
1172         /* take transaction reference */
1173         cur_trans = trans->transaction;
1174         atomic_inc(&cur_trans->use_count);
1175
1176         btrfs_end_transaction(trans, root);
1177         schedule_delayed_work(&ac->work, 0);
1178
1179         /* wait for transaction to start and unblock */
1180         if (wait_for_unblock)
1181                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1182         else
1183                 wait_current_trans_commit_start(root, cur_trans);
1184
1185         if (current->journal_info == trans)
1186                 current->journal_info = NULL;
1187
1188         put_transaction(cur_trans);
1189         return 0;
1190 }
1191
1192 /*
1193  * btrfs_transaction state sequence:
1194  *    in_commit = 0, blocked = 0  (initial)
1195  *    in_commit = 1, blocked = 1
1196  *    blocked = 0
1197  *    commit_done = 1
1198  */
1199 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1200                              struct btrfs_root *root)
1201 {
1202         unsigned long joined = 0;
1203         struct btrfs_transaction *cur_trans;
1204         struct btrfs_transaction *prev_trans = NULL;
1205         DEFINE_WAIT(wait);
1206         int ret;
1207         int should_grow = 0;
1208         unsigned long now = get_seconds();
1209         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1210
1211         btrfs_run_ordered_operations(root, 0);
1212
1213         /* make a pass through all the delayed refs we have so far
1214          * any runnings procs may add more while we are here
1215          */
1216         ret = btrfs_run_delayed_refs(trans, root, 0);
1217         BUG_ON(ret);
1218
1219         btrfs_trans_release_metadata(trans, root);
1220
1221         cur_trans = trans->transaction;
1222         /*
1223          * set the flushing flag so procs in this transaction have to
1224          * start sending their work down.
1225          */
1226         cur_trans->delayed_refs.flushing = 1;
1227
1228         ret = btrfs_run_delayed_refs(trans, root, 0);
1229         BUG_ON(ret);
1230
1231         spin_lock(&cur_trans->commit_lock);
1232         if (cur_trans->in_commit) {
1233                 spin_unlock(&cur_trans->commit_lock);
1234                 atomic_inc(&cur_trans->use_count);
1235                 btrfs_end_transaction(trans, root);
1236
1237                 ret = wait_for_commit(root, cur_trans);
1238                 BUG_ON(ret);
1239
1240                 put_transaction(cur_trans);
1241
1242                 return 0;
1243         }
1244
1245         trans->transaction->in_commit = 1;
1246         trans->transaction->blocked = 1;
1247         spin_unlock(&cur_trans->commit_lock);
1248         wake_up(&root->fs_info->transaction_blocked_wait);
1249
1250         spin_lock(&root->fs_info->trans_lock);
1251         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1252                 prev_trans = list_entry(cur_trans->list.prev,
1253                                         struct btrfs_transaction, list);
1254                 if (!prev_trans->commit_done) {
1255                         atomic_inc(&prev_trans->use_count);
1256                         spin_unlock(&root->fs_info->trans_lock);
1257
1258                         wait_for_commit(root, prev_trans);
1259
1260                         put_transaction(prev_trans);
1261                 } else {
1262                         spin_unlock(&root->fs_info->trans_lock);
1263                 }
1264         } else {
1265                 spin_unlock(&root->fs_info->trans_lock);
1266         }
1267
1268         if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
1269                 should_grow = 1;
1270
1271         do {
1272                 int snap_pending = 0;
1273
1274                 joined = cur_trans->num_joined;
1275                 if (!list_empty(&trans->transaction->pending_snapshots))
1276                         snap_pending = 1;
1277
1278                 WARN_ON(cur_trans != trans->transaction);
1279
1280                 if (flush_on_commit || snap_pending) {
1281                         btrfs_start_delalloc_inodes(root, 1);
1282                         ret = btrfs_wait_ordered_extents(root, 0, 1);
1283                         BUG_ON(ret);
1284                 }
1285
1286                 ret = btrfs_run_delayed_items(trans, root);
1287                 BUG_ON(ret);
1288
1289                 /*
1290                  * rename don't use btrfs_join_transaction, so, once we
1291                  * set the transaction to blocked above, we aren't going
1292                  * to get any new ordered operations.  We can safely run
1293                  * it here and no for sure that nothing new will be added
1294                  * to the list
1295                  */
1296                 btrfs_run_ordered_operations(root, 1);
1297
1298                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1299                                 TASK_UNINTERRUPTIBLE);
1300
1301                 if (atomic_read(&cur_trans->num_writers) > 1)
1302                         schedule_timeout(MAX_SCHEDULE_TIMEOUT);
1303                 else if (should_grow)
1304                         schedule_timeout(1);
1305
1306                 finish_wait(&cur_trans->writer_wait, &wait);
1307         } while (atomic_read(&cur_trans->num_writers) > 1 ||
1308                  (should_grow && cur_trans->num_joined != joined));
1309
1310         /*
1311          * Ok now we need to make sure to block out any other joins while we
1312          * commit the transaction.  We could have started a join before setting
1313          * no_join so make sure to wait for num_writers to == 1 again.
1314          */
1315         spin_lock(&root->fs_info->trans_lock);
1316         root->fs_info->trans_no_join = 1;
1317         spin_unlock(&root->fs_info->trans_lock);
1318         wait_event(cur_trans->writer_wait,
1319                    atomic_read(&cur_trans->num_writers) == 1);
1320
1321         /*
1322          * the reloc mutex makes sure that we stop
1323          * the balancing code from coming in and moving
1324          * extents around in the middle of the commit
1325          */
1326         mutex_lock(&root->fs_info->reloc_mutex);
1327
1328         ret = btrfs_run_delayed_items(trans, root);
1329         BUG_ON(ret);
1330
1331         ret = create_pending_snapshots(trans, root->fs_info);
1332         BUG_ON(ret);
1333
1334         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1335         BUG_ON(ret);
1336
1337         /*
1338          * make sure none of the code above managed to slip in a
1339          * delayed item
1340          */
1341         btrfs_assert_delayed_root_empty(root);
1342
1343         WARN_ON(cur_trans != trans->transaction);
1344
1345         btrfs_scrub_pause(root);
1346         /* btrfs_commit_tree_roots is responsible for getting the
1347          * various roots consistent with each other.  Every pointer
1348          * in the tree of tree roots has to point to the most up to date
1349          * root for every subvolume and other tree.  So, we have to keep
1350          * the tree logging code from jumping in and changing any
1351          * of the trees.
1352          *
1353          * At this point in the commit, there can't be any tree-log
1354          * writers, but a little lower down we drop the trans mutex
1355          * and let new people in.  By holding the tree_log_mutex
1356          * from now until after the super is written, we avoid races
1357          * with the tree-log code.
1358          */
1359         mutex_lock(&root->fs_info->tree_log_mutex);
1360
1361         ret = commit_fs_roots(trans, root);
1362         BUG_ON(ret);
1363
1364         /* commit_fs_roots gets rid of all the tree log roots, it is now
1365          * safe to free the root of tree log roots
1366          */
1367         btrfs_free_log_root_tree(trans, root->fs_info);
1368
1369         ret = commit_cowonly_roots(trans, root);
1370         BUG_ON(ret);
1371
1372         btrfs_prepare_extent_commit(trans, root);
1373
1374         cur_trans = root->fs_info->running_transaction;
1375
1376         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1377                             root->fs_info->tree_root->node);
1378         switch_commit_root(root->fs_info->tree_root);
1379
1380         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1381                             root->fs_info->chunk_root->node);
1382         switch_commit_root(root->fs_info->chunk_root);
1383
1384         update_super_roots(root);
1385
1386         if (!root->fs_info->log_root_recovering) {
1387                 btrfs_set_super_log_root(&root->fs_info->super_copy, 0);
1388                 btrfs_set_super_log_root_level(&root->fs_info->super_copy, 0);
1389         }
1390
1391         memcpy(&root->fs_info->super_for_commit, &root->fs_info->super_copy,
1392                sizeof(root->fs_info->super_copy));
1393
1394         trans->transaction->blocked = 0;
1395         spin_lock(&root->fs_info->trans_lock);
1396         root->fs_info->running_transaction = NULL;
1397         root->fs_info->trans_no_join = 0;
1398         spin_unlock(&root->fs_info->trans_lock);
1399         mutex_unlock(&root->fs_info->reloc_mutex);
1400
1401         wake_up(&root->fs_info->transaction_wait);
1402
1403         ret = btrfs_write_and_wait_transaction(trans, root);
1404         BUG_ON(ret);
1405         write_ctree_super(trans, root, 0);
1406
1407         /*
1408          * the super is written, we can safely allow the tree-loggers
1409          * to go about their business
1410          */
1411         mutex_unlock(&root->fs_info->tree_log_mutex);
1412
1413         btrfs_finish_extent_commit(trans, root);
1414
1415         cur_trans->commit_done = 1;
1416
1417         root->fs_info->last_trans_committed = cur_trans->transid;
1418
1419         wake_up(&cur_trans->commit_wait);
1420
1421         spin_lock(&root->fs_info->trans_lock);
1422         list_del_init(&cur_trans->list);
1423         spin_unlock(&root->fs_info->trans_lock);
1424
1425         put_transaction(cur_trans);
1426         put_transaction(cur_trans);
1427
1428         trace_btrfs_transaction_commit(root);
1429
1430         btrfs_scrub_continue(root);
1431
1432         if (current->journal_info == trans)
1433                 current->journal_info = NULL;
1434
1435         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1436
1437         if (current != root->fs_info->transaction_kthread)
1438                 btrfs_run_delayed_iputs(root);
1439
1440         return ret;
1441 }
1442
1443 /*
1444  * interface function to delete all the snapshots we have scheduled for deletion
1445  */
1446 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1447 {
1448         LIST_HEAD(list);
1449         struct btrfs_fs_info *fs_info = root->fs_info;
1450
1451         spin_lock(&fs_info->trans_lock);
1452         list_splice_init(&fs_info->dead_roots, &list);
1453         spin_unlock(&fs_info->trans_lock);
1454
1455         while (!list_empty(&list)) {
1456                 root = list_entry(list.next, struct btrfs_root, root_list);
1457                 list_del(&root->root_list);
1458
1459                 btrfs_kill_all_delayed_nodes(root);
1460
1461                 if (btrfs_header_backref_rev(root->node) <
1462                     BTRFS_MIXED_BACKREF_REV)
1463                         btrfs_drop_snapshot(root, NULL, 0);
1464                 else
1465                         btrfs_drop_snapshot(root, NULL, 1);
1466         }
1467         return 0;
1468 }