fs/btrfs/transaction.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/fs.h>
  20 #include <linux/slab.h>
  21 #include <linux/sched.h>
  22 #include <linux/writeback.h>
  23 #include <linux/pagemap.h>
  24 #include <linux/blkdev.h>
  25 #include "ctree.h"
  26 #include "disk-io.h"
  27 #include "transaction.h"
  28 #include "locking.h"
  29 #include "tree-log.h"
  30
  31 #define BTRFS_ROOT_TRANS_TAG 0
  32
  33 static noinline void put_transaction(struct btrfs_transaction *transaction)
  34 {
  35         WARN_ON(transaction->use_count == 0);
  36         transaction->use_count--;
  37         if (transaction->use_count == 0) {
  38                 list_del_init(&transaction->list);
  39                 memset(transaction, 0, sizeof(*transaction));
  40                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  41         }
  42 }
  43
  44 static noinline void switch_commit_root(struct btrfs_root *root)
  45 {
  46         free_extent_buffer(root->commit_root);
  47         root->commit_root = btrfs_root_node(root);
  48 }
  49
  50 /*
  51  * either allocate a new transaction or hop into the existing one
  52  */
  53 static noinline int join_transaction(struct btrfs_root *root)
  54 {
  55         struct btrfs_transaction *cur_trans;
  56         cur_trans = root->fs_info->running_transaction;
  57         if (!cur_trans) {
  58                 cur_trans = kmem_cache_alloc(btrfs_transaction_cachep,
  59                                              GFP_NOFS);
  60                 BUG_ON(!cur_trans);
  61                 root->fs_info->generation++;
  62                 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                 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                 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                 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 };
 167
 168 static int may_wait_transaction(struct btrfs_root *root, int type)
 169 {
 170         if (!root->fs_info->log_root_recovering &&
 171             ((type == TRANS_START && !root->fs_info->open_ioctl_trans) ||
 172              type == TRANS_USERSPACE))
 173                 return 1;
 174         return 0;
 175 }
 176
 177 static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
 178                                                     u64 num_items, int type)
 179 {
 180         struct btrfs_trans_handle *h;
 181         struct btrfs_transaction *cur_trans;
 182         int ret;
 183 again:
 184         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 185         if (!h)
 186                 return ERR_PTR(-ENOMEM);
 187
 188         mutex_lock(&root->fs_info->trans_mutex);
 189         if (may_wait_transaction(root, type))
 190                 wait_current_trans(root);
 191
 192         ret = join_transaction(root);
 193         BUG_ON(ret);
 194
 195         cur_trans = root->fs_info->running_transaction;
 196         cur_trans->use_count++;
 197         mutex_unlock(&root->fs_info->trans_mutex);
 198
 199         h->transid = cur_trans->transid;
 200         h->transaction = cur_trans;
 201         h->blocks_used = 0;
 202         h->block_group = 0;
 203         h->bytes_reserved = 0;
 204         h->delayed_ref_updates = 0;
 205         h->block_rsv = NULL;
 206
 207         smp_mb();
 208         if (cur_trans->blocked && may_wait_transaction(root, type)) {
 209                 btrfs_commit_transaction(h, root);
 210                 goto again;
 211         }
 212
 213         if (num_items > 0) {
 214                 ret = btrfs_trans_reserve_metadata(h, root, num_items);
 215                 if (ret == -EAGAIN) {
 216                         btrfs_commit_transaction(h, root);
 217                         goto again;
 218                 }
 219                 if (ret < 0) {
 220                         btrfs_end_transaction(h, root);
 221                         return ERR_PTR(ret);
 222                 }
 223         }
 224
 225         mutex_lock(&root->fs_info->trans_mutex);
 226         record_root_in_trans(h, root);
 227         mutex_unlock(&root->fs_info->trans_mutex);
 228
 229         if (!current->journal_info && type != TRANS_USERSPACE)
 230                 current->journal_info = h;
 231         return h;
 232 }
 233
 234 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 235                                                    int num_items)
 236 {
 237         return start_transaction(root, num_items, TRANS_START);
 238 }
 239 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root,
 240                                                    int num_blocks)
 241 {
 242         return start_transaction(root, 0, TRANS_JOIN);
 243 }
 244
 245 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *r,
 246                                                          int num_blocks)
 247 {
 248         return start_transaction(r, 0, TRANS_USERSPACE);
 249 }
 250
 251 /* wait for a transaction commit to be fully complete */
 252 static noinline int wait_for_commit(struct btrfs_root *root,
 253                                     struct btrfs_transaction *commit)
 254 {
 255         DEFINE_WAIT(wait);
 256         mutex_lock(&root->fs_info->trans_mutex);
 257         while (!commit->commit_done) {
 258                 prepare_to_wait(&commit->commit_wait, &wait,
 259                                 TASK_UNINTERRUPTIBLE);
 260                 if (commit->commit_done)
 261                         break;
 262                 mutex_unlock(&root->fs_info->trans_mutex);
 263                 schedule();
 264                 mutex_lock(&root->fs_info->trans_mutex);
 265         }
 266         mutex_unlock(&root->fs_info->trans_mutex);
 267         finish_wait(&commit->commit_wait, &wait);
 268         return 0;
 269 }
 270
 271 #if 0
 272 /*
 273  * rate limit against the drop_snapshot code.  This helps to slow down new
 274  * operations if the drop_snapshot code isn't able to keep up.
 275  */
 276 static void throttle_on_drops(struct btrfs_root *root)
 277 {
 278         struct btrfs_fs_info *info = root->fs_info;
 279         int harder_count = 0;
 280
 281 harder:
 282         if (atomic_read(&info->throttles)) {
 283                 DEFINE_WAIT(wait);
 284                 int thr;
 285                 thr = atomic_read(&info->throttle_gen);
 286
 287                 do {
 288                         prepare_to_wait(&info->transaction_throttle,
 289                                         &wait, TASK_UNINTERRUPTIBLE);
 290                         if (!atomic_read(&info->throttles)) {
 291                                 finish_wait(&info->transaction_throttle, &wait);
 292                                 break;
 293                         }
 294                         schedule();
 295                         finish_wait(&info->transaction_throttle, &wait);
 296                 } while (thr == atomic_read(&info->throttle_gen));
 297                 harder_count++;
 298
 299                 if (root->fs_info->total_ref_cache_size > 1 * 1024 * 1024 &&
 300                     harder_count < 2)
 301                         goto harder;
 302
 303                 if (root->fs_info->total_ref_cache_size > 5 * 1024 * 1024 &&
 304                     harder_count < 10)
 305                         goto harder;
 306
 307                 if (root->fs_info->total_ref_cache_size > 10 * 1024 * 1024 &&
 308                     harder_count < 20)
 309                         goto harder;
 310         }
 311 }
 312 #endif
 313
 314 void btrfs_throttle(struct btrfs_root *root)
 315 {
 316         mutex_lock(&root->fs_info->trans_mutex);
 317         if (!root->fs_info->open_ioctl_trans)
 318                 wait_current_trans(root);
 319         mutex_unlock(&root->fs_info->trans_mutex);
 320 }
 321
 322 static int should_end_transaction(struct btrfs_trans_handle *trans,
 323                                   struct btrfs_root *root)
 324 {
 325         int ret;
 326         ret = btrfs_block_rsv_check(trans, root,
 327                                     &root->fs_info->global_block_rsv, 0, 5);
 328         return ret ? 1 : 0;
 329 }
 330
 331 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 332                                  struct btrfs_root *root)
 333 {
 334         struct btrfs_transaction *cur_trans = trans->transaction;
 335         int updates;
 336
 337         if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
 338                 return 1;
 339
 340         updates = trans->delayed_ref_updates;
 341         trans->delayed_ref_updates = 0;
 342         if (updates)
 343                 btrfs_run_delayed_refs(trans, root, updates);
 344
 345         return should_end_transaction(trans, root);
 346 }
 347
 348 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 349                           struct btrfs_root *root, int throttle)
 350 {
 351         struct btrfs_transaction *cur_trans = trans->transaction;
 352         struct btrfs_fs_info *info = root->fs_info;
 353         int count = 0;
 354
 355         while (count < 4) {
 356                 unsigned long cur = trans->delayed_ref_updates;
 357                 trans->delayed_ref_updates = 0;
 358                 if (cur &&
 359                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 360                         trans->delayed_ref_updates = 0;
 361
 362                         /*
 363                          * do a full flush if the transaction is trying
 364                          * to close
 365                          */
 366                         if (trans->transaction->delayed_refs.flushing)
 367                                 cur = 0;
 368                         btrfs_run_delayed_refs(trans, root, cur);
 369                 } else {
 370                         break;
 371                 }
 372                 count++;
 373         }
 374
 375         btrfs_trans_release_metadata(trans, root);
 376
 377         if (!root->fs_info->open_ioctl_trans &&
 378             should_end_transaction(trans, root))
 379                 trans->transaction->blocked = 1;
 380
 381         if (cur_trans->blocked && !cur_trans->in_commit) {
 382                 if (throttle)
 383                         return btrfs_commit_transaction(trans, root);
 384                 else
 385                         wake_up_process(info->transaction_kthread);
 386         }
 387
 388         mutex_lock(&info->trans_mutex);
 389         WARN_ON(cur_trans != info->running_transaction);
 390         WARN_ON(cur_trans->num_writers < 1);
 391         cur_trans->num_writers--;
 392
 393         if (waitqueue_active(&cur_trans->writer_wait))
 394                 wake_up(&cur_trans->writer_wait);
 395         put_transaction(cur_trans);
 396         mutex_unlock(&info->trans_mutex);
 397
 398         if (current->journal_info == trans)
 399                 current->journal_info = NULL;
 400         memset(trans, 0, sizeof(*trans));
 401         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 402
 403         if (throttle)
 404                 btrfs_run_delayed_iputs(root);
 405
 406         return 0;
 407 }
 408
 409 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 410                           struct btrfs_root *root)
 411 {
 412         return __btrfs_end_transaction(trans, root, 0);
 413 }
 414
 415 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 416                                    struct btrfs_root *root)
 417 {
 418         return __btrfs_end_transaction(trans, root, 1);
 419 }
 420
 421 /*
 422  * when btree blocks are allocated, they have some corresponding bits set for
 423  * them in one of two extent_io trees.  This is used to make sure all of
 424  * those extents are sent to disk but does not wait on them
 425  */
 426 int btrfs_write_marked_extents(struct btrfs_root *root,
 427                                struct extent_io_tree *dirty_pages, int mark)
 428 {
 429         int ret;
 430         int err = 0;
 431         int werr = 0;
 432         struct page *page;
 433         struct inode *btree_inode = root->fs_info->btree_inode;
 434         u64 start = 0;
 435         u64 end;
 436         unsigned long index;
 437
 438         while (1) {
 439                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 440                                             mark);
 441                 if (ret)
 442                         break;
 443                 while (start <= end) {
 444                         cond_resched();
 445
 446                         index = start >> PAGE_CACHE_SHIFT;
 447                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 448                         page = find_get_page(btree_inode->i_mapping, index);
 449                         if (!page)
 450                                 continue;
 451
 452                         btree_lock_page_hook(page);
 453                         if (!page->mapping) {
 454                                 unlock_page(page);
 455                                 page_cache_release(page);
 456                                 continue;
 457                         }
 458
 459                         if (PageWriteback(page)) {
 460                                 if (PageDirty(page))
 461                                         wait_on_page_writeback(page);
 462                                 else {
 463                                         unlock_page(page);
 464                                         page_cache_release(page);
 465                                         continue;
 466                                 }
 467                         }
 468                         err = write_one_page(page, 0);
 469                         if (err)
 470                                 werr = err;
 471                         page_cache_release(page);
 472                 }
 473         }
 474         if (err)
 475                 werr = err;
 476         return werr;
 477 }
 478
 479 /*
 480  * when btree blocks are allocated, they have some corresponding bits set for
 481  * them in one of two extent_io trees.  This is used to make sure all of
 482  * those extents are on disk for transaction or log commit.  We wait
 483  * on all the pages and clear them from the dirty pages state tree
 484  */
 485 int btrfs_wait_marked_extents(struct btrfs_root *root,
 486                               struct extent_io_tree *dirty_pages, int mark)
 487 {
 488         int ret;
 489         int err = 0;
 490         int werr = 0;
 491         struct page *page;
 492         struct inode *btree_inode = root->fs_info->btree_inode;
 493         u64 start = 0;
 494         u64 end;
 495         unsigned long index;
 496
 497         while (1) {
 498                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 499                                             mark);
 500                 if (ret)
 501                         break;
 502
 503                 clear_extent_bits(dirty_pages, start, end, mark, GFP_NOFS);
 504                 while (start <= end) {
 505                         index = start >> PAGE_CACHE_SHIFT;
 506                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 507                         page = find_get_page(btree_inode->i_mapping, index);
 508                         if (!page)
 509                                 continue;
 510                         if (PageDirty(page)) {
 511                                 btree_lock_page_hook(page);
 512                                 wait_on_page_writeback(page);
 513                                 err = write_one_page(page, 0);
 514                                 if (err)
 515                                         werr = err;
 516                         }
 517                         wait_on_page_writeback(page);
 518                         page_cache_release(page);
 519                         cond_resched();
 520                 }
 521         }
 522         if (err)
 523                 werr = err;
 524         return werr;
 525 }
 526
 527 /*
 528  * when btree blocks are allocated, they have some corresponding bits set for
 529  * them in one of two extent_io trees.  This is used to make sure all of
 530  * those extents are on disk for transaction or log commit
 531  */
 532 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 533                                 struct extent_io_tree *dirty_pages, int mark)
 534 {
 535         int ret;
 536         int ret2;
 537
 538         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 539         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 540         return ret || ret2;
 541 }
 542
 543 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 544                                      struct btrfs_root *root)
 545 {
 546         if (!trans || !trans->transaction) {
 547                 struct inode *btree_inode;
 548                 btree_inode = root->fs_info->btree_inode;
 549                 return filemap_write_and_wait(btree_inode->i_mapping);
 550         }
 551         return btrfs_write_and_wait_marked_extents(root,
 552                                            &trans->transaction->dirty_pages,
 553                                            EXTENT_DIRTY);
 554 }
 555
 556 /*
 557  * this is used to update the root pointer in the tree of tree roots.
 558  *
 559  * But, in the case of the extent allocation tree, updating the root
 560  * pointer may allocate blocks which may change the root of the extent
 561  * allocation tree.
 562  *
 563  * So, this loops and repeats and makes sure the cowonly root didn't
 564  * change while the root pointer was being updated in the metadata.
 565  */
 566 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 567                                struct btrfs_root *root)
 568 {
 569         int ret;
 570         u64 old_root_bytenr;
 571         u64 old_root_used;
 572         struct btrfs_root *tree_root = root->fs_info->tree_root;
 573
 574         old_root_used = btrfs_root_used(&root->root_item);
 575         btrfs_write_dirty_block_groups(trans, root);
 576
 577         while (1) {
 578                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 579                 if (old_root_bytenr == root->node->start &&
 580                     old_root_used == btrfs_root_used(&root->root_item))
 581                         break;
 582
 583                 btrfs_set_root_node(&root->root_item, root->node);
 584                 ret = btrfs_update_root(trans, tree_root,
 585                                         &root->root_key,
 586                                         &root->root_item);
 587                 BUG_ON(ret);
 588
 589                 old_root_used = btrfs_root_used(&root->root_item);
 590                 ret = btrfs_write_dirty_block_groups(trans, root);
 591                 BUG_ON(ret);
 592         }
 593
 594         if (root != root->fs_info->extent_root)
 595                 switch_commit_root(root);
 596
 597         return 0;
 598 }
 599
 600 /*
 601  * update all the cowonly tree roots on disk
 602  */
 603 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 604                                          struct btrfs_root *root)
 605 {
 606         struct btrfs_fs_info *fs_info = root->fs_info;
 607         struct list_head *next;
 608         struct extent_buffer *eb;
 609         int ret;
 610
 611         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 612         BUG_ON(ret);
 613
 614         eb = btrfs_lock_root_node(fs_info->tree_root);
 615         btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
 616         btrfs_tree_unlock(eb);
 617         free_extent_buffer(eb);
 618
 619         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 620         BUG_ON(ret);
 621
 622         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 623                 next = fs_info->dirty_cowonly_roots.next;
 624                 list_del_init(next);
 625                 root = list_entry(next, struct btrfs_root, dirty_list);
 626
 627                 update_cowonly_root(trans, root);
 628         }
 629
 630         down_write(&fs_info->extent_commit_sem);
 631         switch_commit_root(fs_info->extent_root);
 632         up_write(&fs_info->extent_commit_sem);
 633
 634         return 0;
 635 }
 636
 637 /*
 638  * dead roots are old snapshots that need to be deleted.  This allocates
 639  * a dirty root struct and adds it into the list of dead roots that need to
 640  * be deleted
 641  */
 642 int btrfs_add_dead_root(struct btrfs_root *root)
 643 {
 644         mutex_lock(&root->fs_info->trans_mutex);
 645         list_add(&root->root_list, &root->fs_info->dead_roots);
 646         mutex_unlock(&root->fs_info->trans_mutex);
 647         return 0;
 648 }
 649
 650 /*
 651  * update all the cowonly tree roots on disk
 652  */
 653 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 654                                     struct btrfs_root *root)
 655 {
 656         struct btrfs_root *gang[8];
 657         struct btrfs_fs_info *fs_info = root->fs_info;
 658         int i;
 659         int ret;
 660         int err = 0;
 661
 662         while (1) {
 663                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 664                                                  (void **)gang, 0,
 665                                                  ARRAY_SIZE(gang),
 666                                                  BTRFS_ROOT_TRANS_TAG);
 667                 if (ret == 0)
 668                         break;
 669                 for (i = 0; i < ret; i++) {
 670                         root = gang[i];
 671                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 672                                         (unsigned long)root->root_key.objectid,
 673                                         BTRFS_ROOT_TRANS_TAG);
 674
 675                         btrfs_free_log(trans, root);
 676                         btrfs_update_reloc_root(trans, root);
 677                         btrfs_orphan_commit_root(trans, root);
 678
 679                         if (root->commit_root != root->node) {
 680                                 switch_commit_root(root);
 681                                 btrfs_set_root_node(&root->root_item,
 682                                                     root->node);
 683                         }
 684
 685                         err = btrfs_update_root(trans, fs_info->tree_root,
 686                                                 &root->root_key,
 687                                                 &root->root_item);
 688                         if (err)
 689                                 break;
 690                 }
 691         }
 692         return err;
 693 }
 694
 695 /*
 696  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 697  * otherwise every leaf in the btree is read and defragged.
 698  */
 699 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 700 {
 701         struct btrfs_fs_info *info = root->fs_info;
 702         struct btrfs_trans_handle *trans;
 703         int ret;
 704         unsigned long nr;
 705
 706         if (xchg(&root->defrag_running, 1))
 707                 return 0;
 708
 709         while (1) {
 710                 trans = btrfs_start_transaction(root, 0);
 711                 if (IS_ERR(trans))
 712                         return PTR_ERR(trans);
 713
 714                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 715
 716                 nr = trans->blocks_used;
 717                 btrfs_end_transaction(trans, root);
 718                 btrfs_btree_balance_dirty(info->tree_root, nr);
 719                 cond_resched();
 720
 721                 if (root->fs_info->closing || ret != -EAGAIN)
 722                         break;
 723         }
 724         root->defrag_running = 0;
 725         return ret;
 726 }
 727
 728 #if 0
 729 /*
 730  * when dropping snapshots, we generate a ton of delayed refs, and it makes
 731  * sense not to join the transaction while it is trying to flush the current
 732  * queue of delayed refs out.
 733  *
 734  * This is used by the drop snapshot code only
 735  */
 736 static noinline int wait_transaction_pre_flush(struct btrfs_fs_info *info)
 737 {
 738         DEFINE_WAIT(wait);
 739
 740         mutex_lock(&info->trans_mutex);
 741         while (info->running_transaction &&
 742                info->running_transaction->delayed_refs.flushing) {
 743                 prepare_to_wait(&info->transaction_wait, &wait,
 744                                 TASK_UNINTERRUPTIBLE);
 745                 mutex_unlock(&info->trans_mutex);
 746
 747                 schedule();
 748
 749                 mutex_lock(&info->trans_mutex);
 750                 finish_wait(&info->transaction_wait, &wait);
 751         }
 752         mutex_unlock(&info->trans_mutex);
 753         return 0;
 754 }
 755
 756 /*
 757  * Given a list of roots that need to be deleted, call btrfs_drop_snapshot on
 758  * all of them
 759  */
 760 int btrfs_drop_dead_root(struct btrfs_root *root)
 761 {
 762         struct btrfs_trans_handle *trans;
 763         struct btrfs_root *tree_root = root->fs_info->tree_root;
 764         unsigned long nr;
 765         int ret;
 766
 767         while (1) {
 768                 /*
 769                  * we don't want to jump in and create a bunch of
 770                  * delayed refs if the transaction is starting to close
 771                  */
 772                 wait_transaction_pre_flush(tree_root->fs_info);
 773                 trans = btrfs_start_transaction(tree_root, 1);
 774
 775                 /*
 776                  * we've joined a transaction, make sure it isn't
 777                  * closing right now
 778                  */
 779                 if (trans->transaction->delayed_refs.flushing) {
 780                         btrfs_end_transaction(trans, tree_root);
 781                         continue;
 782                 }
 783
 784                 ret = btrfs_drop_snapshot(trans, root);
 785                 if (ret != -EAGAIN)
 786                         break;
 787
 788                 ret = btrfs_update_root(trans, tree_root,
 789                                         &root->root_key,
 790                                         &root->root_item);
 791                 if (ret)
 792                         break;
 793
 794                 nr = trans->blocks_used;
 795                 ret = btrfs_end_transaction(trans, tree_root);
 796                 BUG_ON(ret);
 797
 798                 btrfs_btree_balance_dirty(tree_root, nr);
 799                 cond_resched();
 800         }
 801         BUG_ON(ret);
 802
 803         ret = btrfs_del_root(trans, tree_root, &root->root_key);
 804         BUG_ON(ret);
 805
 806         nr = trans->blocks_used;
 807         ret = btrfs_end_transaction(trans, tree_root);
 808         BUG_ON(ret);
 809
 810         free_extent_buffer(root->node);
 811         free_extent_buffer(root->commit_root);
 812         kfree(root);
 813
 814         btrfs_btree_balance_dirty(tree_root, nr);
 815         return ret;
 816 }
 817 #endif
 818
 819 /*
 820  * new snapshots need to be created at a very specific time in the
 821  * transaction commit.  This does the actual creation
 822  */
 823 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 824                                    struct btrfs_fs_info *fs_info,
 825                                    struct btrfs_pending_snapshot *pending)
 826 {
 827         struct btrfs_key key;
 828         struct btrfs_root_item *new_root_item;
 829         struct btrfs_root *tree_root = fs_info->tree_root;
 830         struct btrfs_root *root = pending->root;
 831         struct btrfs_root *parent_root;
 832         struct inode *parent_inode;
 833         struct dentry *dentry;
 834         struct extent_buffer *tmp;
 835         struct extent_buffer *old;
 836         int ret;
 837         u64 to_reserve = 0;
 838         u64 index = 0;
 839         u64 objectid;
 840
 841         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 842         if (!new_root_item) {
 843                 pending->error = -ENOMEM;
 844                 goto fail;
 845         }
 846
 847         ret = btrfs_find_free_objectid(trans, tree_root, 0, &objectid);
 848         if (ret) {
 849                 pending->error = ret;
 850                 goto fail;
 851         }
 852
 853         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
 854         btrfs_orphan_pre_snapshot(trans, pending, &to_reserve);
 855
 856         if (to_reserve > 0) {
 857                 ret = btrfs_block_rsv_add(trans, root, &pending->block_rsv,
 858                                           to_reserve);
 859                 if (ret) {
 860                         pending->error = ret;
 861                         goto fail;
 862                 }
 863         }
 864
 865         key.objectid = objectid;
 866         key.offset = (u64)-1;
 867         key.type = BTRFS_ROOT_ITEM_KEY;
 868
 869         trans->block_rsv = &pending->block_rsv;
 870
 871         dentry = pending->dentry;
 872         parent_inode = dentry->d_parent->d_inode;
 873         parent_root = BTRFS_I(parent_inode)->root;
 874         record_root_in_trans(trans, parent_root);
 875
 876         /*
 877          * insert the directory item
 878          */
 879         ret = btrfs_set_inode_index(parent_inode, &index);
 880         BUG_ON(ret);
 881         ret = btrfs_insert_dir_item(trans, parent_root,
 882                                 dentry->d_name.name, dentry->d_name.len,
 883                                 parent_inode->i_ino, &key,
 884                                 BTRFS_FT_DIR, index);
 885         BUG_ON(ret);
 886
 887         btrfs_i_size_write(parent_inode, parent_inode->i_size +
 888                                          dentry->d_name.len * 2);
 889         ret = btrfs_update_inode(trans, parent_root, parent_inode);
 890         BUG_ON(ret);
 891
 892         record_root_in_trans(trans, root);
 893         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
 894         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
 895
 896         old = btrfs_lock_root_node(root);
 897         btrfs_cow_block(trans, root, old, NULL, 0, &old);
 898         btrfs_set_lock_blocking(old);
 899
 900         btrfs_copy_root(trans, root, old, &tmp, objectid);
 901         btrfs_tree_unlock(old);
 902         free_extent_buffer(old);
 903
 904         btrfs_set_root_node(new_root_item, tmp);
 905         /* record when the snapshot was created in key.offset */
 906         key.offset = trans->transid;
 907         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
 908         btrfs_tree_unlock(tmp);
 909         free_extent_buffer(tmp);
 910         BUG_ON(ret);
 911
 912         /*
 913          * insert root back/forward references
 914          */
 915         ret = btrfs_add_root_ref(trans, tree_root, objectid,
 916                                  parent_root->root_key.objectid,
 917                                  parent_inode->i_ino, index,
 918                                  dentry->d_name.name, dentry->d_name.len);
 919         BUG_ON(ret);
 920
 921         key.offset = (u64)-1;
 922         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
 923         BUG_ON(IS_ERR(pending->snap));
 924
 925         btrfs_reloc_post_snapshot(trans, pending);
 926         btrfs_orphan_post_snapshot(trans, pending);
 927 fail:
 928         kfree(new_root_item);
 929         btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
 930         return 0;
 931 }
 932
 933 /*
 934  * create all the snapshots we've scheduled for creation
 935  */
 936 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
 937                                              struct btrfs_fs_info *fs_info)
 938 {
 939         struct btrfs_pending_snapshot *pending;
 940         struct list_head *head = &trans->transaction->pending_snapshots;
 941         int ret;
 942
 943         list_for_each_entry(pending, head, list) {
 944                 ret = create_pending_snapshot(trans, fs_info, pending);
 945                 BUG_ON(ret);
 946         }
 947         return 0;
 948 }
 949
 950 static void update_super_roots(struct btrfs_root *root)
 951 {
 952         struct btrfs_root_item *root_item;
 953         struct btrfs_super_block *super;
 954
 955         super = &root->fs_info->super_copy;
 956
 957         root_item = &root->fs_info->chunk_root->root_item;
 958         super->chunk_root = root_item->bytenr;
 959         super->chunk_root_generation = root_item->generation;
 960         super->chunk_root_level = root_item->level;
 961
 962         root_item = &root->fs_info->tree_root->root_item;
 963         super->root = root_item->bytenr;
 964         super->generation = root_item->generation;
 965         super->root_level = root_item->level;
 966 }
 967
 968 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
 969 {
 970         int ret = 0;
 971         spin_lock(&info->new_trans_lock);
 972         if (info->running_transaction)
 973                 ret = info->running_transaction->in_commit;
 974         spin_unlock(&info->new_trans_lock);
 975         return ret;
 976 }
 977
 978 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
 979 {
 980         int ret = 0;
 981         spin_lock(&info->new_trans_lock);
 982         if (info->running_transaction)
 983                 ret = info->running_transaction->blocked;
 984         spin_unlock(&info->new_trans_lock);
 985         return ret;
 986 }
 987
 988 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
 989                              struct btrfs_root *root)
 990 {
 991         unsigned long joined = 0;
 992         unsigned long timeout = 1;
 993         struct btrfs_transaction *cur_trans;
 994         struct btrfs_transaction *prev_trans = NULL;
 995         DEFINE_WAIT(wait);
 996         int ret;
 997         int should_grow = 0;
 998         unsigned long now = get_seconds();
 999         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1000
1001         btrfs_run_ordered_operations(root, 0);
1002
1003         /* make a pass through all the delayed refs we have so far
1004          * any runnings procs may add more while we are here
1005          */
1006         ret = btrfs_run_delayed_refs(trans, root, 0);
1007         BUG_ON(ret);
1008
1009         btrfs_trans_release_metadata(trans, root);
1010
1011         cur_trans = trans->transaction;
1012         /*
1013          * set the flushing flag so procs in this transaction have to
1014          * start sending their work down.
1015          */
1016         cur_trans->delayed_refs.flushing = 1;
1017
1018         ret = btrfs_run_delayed_refs(trans, root, 0);
1019         BUG_ON(ret);
1020
1021         mutex_lock(&root->fs_info->trans_mutex);
1022         if (cur_trans->in_commit) {
1023                 cur_trans->use_count++;
1024                 mutex_unlock(&root->fs_info->trans_mutex);
1025                 btrfs_end_transaction(trans, root);
1026
1027                 ret = wait_for_commit(root, cur_trans);
1028                 BUG_ON(ret);
1029
1030                 mutex_lock(&root->fs_info->trans_mutex);
1031                 put_transaction(cur_trans);
1032                 mutex_unlock(&root->fs_info->trans_mutex);
1033
1034                 return 0;
1035         }
1036
1037         trans->transaction->in_commit = 1;
1038         trans->transaction->blocked = 1;
1039         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1040                 prev_trans = list_entry(cur_trans->list.prev,
1041                                         struct btrfs_transaction, list);
1042                 if (!prev_trans->commit_done) {
1043                         prev_trans->use_count++;
1044                         mutex_unlock(&root->fs_info->trans_mutex);
1045
1046                         wait_for_commit(root, prev_trans);
1047
1048                         mutex_lock(&root->fs_info->trans_mutex);
1049                         put_transaction(prev_trans);
1050                 }
1051         }
1052
1053         if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
1054                 should_grow = 1;
1055
1056         do {
1057                 int snap_pending = 0;
1058                 joined = cur_trans->num_joined;
1059                 if (!list_empty(&trans->transaction->pending_snapshots))
1060                         snap_pending = 1;
1061
1062                 WARN_ON(cur_trans != trans->transaction);
1063                 if (cur_trans->num_writers > 1)
1064                         timeout = MAX_SCHEDULE_TIMEOUT;
1065                 else if (should_grow)
1066                         timeout = 1;
1067
1068                 mutex_unlock(&root->fs_info->trans_mutex);
1069
1070                 if (flush_on_commit || snap_pending) {
1071                         btrfs_start_delalloc_inodes(root, 1);
1072                         ret = btrfs_wait_ordered_extents(root, 0, 1);
1073                         BUG_ON(ret);
1074                 }
1075
1076                 /*
1077                  * rename don't use btrfs_join_transaction, so, once we
1078                  * set the transaction to blocked above, we aren't going
1079                  * to get any new ordered operations.  We can safely run
1080                  * it here and no for sure that nothing new will be added
1081                  * to the list
1082                  */
1083                 btrfs_run_ordered_operations(root, 1);
1084
1085                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1086                                 TASK_UNINTERRUPTIBLE);
1087
1088                 smp_mb();
1089                 if (cur_trans->num_writers > 1 || should_grow)
1090                         schedule_timeout(timeout);
1091
1092                 mutex_lock(&root->fs_info->trans_mutex);
1093                 finish_wait(&cur_trans->writer_wait, &wait);
1094         } while (cur_trans->num_writers > 1 ||
1095                  (should_grow && cur_trans->num_joined != joined));
1096
1097         ret = create_pending_snapshots(trans, root->fs_info);
1098         BUG_ON(ret);
1099
1100         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1101         BUG_ON(ret);
1102
1103         WARN_ON(cur_trans != trans->transaction);
1104
1105         /* btrfs_commit_tree_roots is responsible for getting the
1106          * various roots consistent with each other.  Every pointer
1107          * in the tree of tree roots has to point to the most up to date
1108          * root for every subvolume and other tree.  So, we have to keep
1109          * the tree logging code from jumping in and changing any
1110          * of the trees.
1111          *
1112          * At this point in the commit, there can't be any tree-log
1113          * writers, but a little lower down we drop the trans mutex
1114          * and let new people in.  By holding the tree_log_mutex
1115          * from now until after the super is written, we avoid races
1116          * with the tree-log code.
1117          */
1118         mutex_lock(&root->fs_info->tree_log_mutex);
1119
1120         ret = commit_fs_roots(trans, root);
1121         BUG_ON(ret);
1122
1123         /* commit_fs_roots gets rid of all the tree log roots, it is now
1124          * safe to free the root of tree log roots
1125          */
1126         btrfs_free_log_root_tree(trans, root->fs_info);
1127
1128         ret = commit_cowonly_roots(trans, root);
1129         BUG_ON(ret);
1130
1131         btrfs_prepare_extent_commit(trans, root);
1132
1133         cur_trans = root->fs_info->running_transaction;
1134         spin_lock(&root->fs_info->new_trans_lock);
1135         root->fs_info->running_transaction = NULL;
1136         spin_unlock(&root->fs_info->new_trans_lock);
1137
1138         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1139                             root->fs_info->tree_root->node);
1140         switch_commit_root(root->fs_info->tree_root);
1141
1142         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1143                             root->fs_info->chunk_root->node);
1144         switch_commit_root(root->fs_info->chunk_root);
1145
1146         update_super_roots(root);
1147
1148         if (!root->fs_info->log_root_recovering) {
1149                 btrfs_set_super_log_root(&root->fs_info->super_copy, 0);
1150                 btrfs_set_super_log_root_level(&root->fs_info->super_copy, 0);
1151         }
1152
1153         memcpy(&root->fs_info->super_for_commit, &root->fs_info->super_copy,
1154                sizeof(root->fs_info->super_copy));
1155
1156         trans->transaction->blocked = 0;
1157
1158         wake_up(&root->fs_info->transaction_wait);
1159
1160         mutex_unlock(&root->fs_info->trans_mutex);
1161         ret = btrfs_write_and_wait_transaction(trans, root);
1162         BUG_ON(ret);
1163         write_ctree_super(trans, root, 0);
1164
1165         /*
1166          * the super is written, we can safely allow the tree-loggers
1167          * to go about their business
1168          */
1169         mutex_unlock(&root->fs_info->tree_log_mutex);
1170
1171         btrfs_finish_extent_commit(trans, root);
1172
1173         mutex_lock(&root->fs_info->trans_mutex);
1174
1175         cur_trans->commit_done = 1;
1176
1177         root->fs_info->last_trans_committed = cur_trans->transid;
1178
1179         wake_up(&cur_trans->commit_wait);
1180
1181         put_transaction(cur_trans);
1182         put_transaction(cur_trans);
1183
1184         mutex_unlock(&root->fs_info->trans_mutex);
1185
1186         if (current->journal_info == trans)
1187                 current->journal_info = NULL;
1188
1189         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1190
1191         if (current != root->fs_info->transaction_kthread)
1192                 btrfs_run_delayed_iputs(root);
1193
1194         return ret;
1195 }
1196
1197 /*
1198  * interface function to delete all the snapshots we have scheduled for deletion
1199  */
1200 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1201 {
1202         LIST_HEAD(list);
1203         struct btrfs_fs_info *fs_info = root->fs_info;
1204
1205         mutex_lock(&fs_info->trans_mutex);
1206         list_splice_init(&fs_info->dead_roots, &list);
1207         mutex_unlock(&fs_info->trans_mutex);
1208
1209         while (!list_empty(&list)) {
1210                 root = list_entry(list.next, struct btrfs_root, root_list);
1211                 list_del(&root->root_list);
1212
1213                 if (btrfs_header_backref_rev(root->node) <
1214                     BTRFS_MIXED_BACKREF_REV)
1215                         btrfs_drop_snapshot(root, NULL, 0);
1216                 else
1217                         btrfs_drop_snapshot(root, NULL, 1);
1218         }
1219         return 0;
1220 }