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 retries = 0;
 183         int ret;
 184 again:
 185         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 186         if (!h)
 187                 return ERR_PTR(-ENOMEM);
 188
 189         mutex_lock(&root->fs_info->trans_mutex);
 190         if (may_wait_transaction(root, type))
 191                 wait_current_trans(root);
 192
 193         ret = join_transaction(root);
 194         BUG_ON(ret);
 195
 196         cur_trans = root->fs_info->running_transaction;
 197         cur_trans->use_count++;
 198         mutex_unlock(&root->fs_info->trans_mutex);
 199
 200         h->transid = cur_trans->transid;
 201         h->transaction = cur_trans;
 202         h->blocks_used = 0;
 203         h->block_group = 0;
 204         h->bytes_reserved = 0;
 205         h->delayed_ref_updates = 0;
 206         h->block_rsv = NULL;
 207
 208         smp_mb();
 209         if (cur_trans->blocked && may_wait_transaction(root, type)) {
 210                 btrfs_commit_transaction(h, root);
 211                 goto again;
 212         }
 213
 214         if (num_items > 0) {
 215                 ret = btrfs_trans_reserve_metadata(h, root, num_items,
 216                                                    &retries);
 217                 if (ret == -EAGAIN) {
 218                         btrfs_commit_transaction(h, root);
 219                         goto again;
 220                 }
 221                 if (ret < 0) {
 222                         btrfs_end_transaction(h, root);
 223                         return ERR_PTR(ret);
 224                 }
 225         }
 226
 227         mutex_lock(&root->fs_info->trans_mutex);
 228         record_root_in_trans(h, root);
 229         mutex_unlock(&root->fs_info->trans_mutex);
 230
 231         if (!current->journal_info && type != TRANS_USERSPACE)
 232                 current->journal_info = h;
 233         return h;
 234 }
 235
 236 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 237                                                    int num_items)
 238 {
 239         return start_transaction(root, num_items, TRANS_START);
 240 }
 241 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root,
 242                                                    int num_blocks)
 243 {
 244         return start_transaction(root, 0, TRANS_JOIN);
 245 }
 246
 247 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *r,
 248                                                          int num_blocks)
 249 {
 250         return start_transaction(r, 0, TRANS_USERSPACE);
 251 }
 252
 253 /* wait for a transaction commit to be fully complete */
 254 static noinline int wait_for_commit(struct btrfs_root *root,
 255                                     struct btrfs_transaction *commit)
 256 {
 257         DEFINE_WAIT(wait);
 258         mutex_lock(&root->fs_info->trans_mutex);
 259         while (!commit->commit_done) {
 260                 prepare_to_wait(&commit->commit_wait, &wait,
 261                                 TASK_UNINTERRUPTIBLE);
 262                 if (commit->commit_done)
 263                         break;
 264                 mutex_unlock(&root->fs_info->trans_mutex);
 265                 schedule();
 266                 mutex_lock(&root->fs_info->trans_mutex);
 267         }
 268         mutex_unlock(&root->fs_info->trans_mutex);
 269         finish_wait(&commit->commit_wait, &wait);
 270         return 0;
 271 }
 272
 273 #if 0
 274 /*
 275  * rate limit against the drop_snapshot code.  This helps to slow down new
 276  * operations if the drop_snapshot code isn't able to keep up.
 277  */
 278 static void throttle_on_drops(struct btrfs_root *root)
 279 {
 280         struct btrfs_fs_info *info = root->fs_info;
 281         int harder_count = 0;
 282
 283 harder:
 284         if (atomic_read(&info->throttles)) {
 285                 DEFINE_WAIT(wait);
 286                 int thr;
 287                 thr = atomic_read(&info->throttle_gen);
 288
 289                 do {
 290                         prepare_to_wait(&info->transaction_throttle,
 291                                         &wait, TASK_UNINTERRUPTIBLE);
 292                         if (!atomic_read(&info->throttles)) {
 293                                 finish_wait(&info->transaction_throttle, &wait);
 294                                 break;
 295                         }
 296                         schedule();
 297                         finish_wait(&info->transaction_throttle, &wait);
 298                 } while (thr == atomic_read(&info->throttle_gen));
 299                 harder_count++;
 300
 301                 if (root->fs_info->total_ref_cache_size > 1 * 1024 * 1024 &&
 302                     harder_count < 2)
 303                         goto harder;
 304
 305                 if (root->fs_info->total_ref_cache_size > 5 * 1024 * 1024 &&
 306                     harder_count < 10)
 307                         goto harder;
 308
 309                 if (root->fs_info->total_ref_cache_size > 10 * 1024 * 1024 &&
 310                     harder_count < 20)
 311                         goto harder;
 312         }
 313 }
 314 #endif
 315
 316 void btrfs_throttle(struct btrfs_root *root)
 317 {
 318         mutex_lock(&root->fs_info->trans_mutex);
 319         if (!root->fs_info->open_ioctl_trans)
 320                 wait_current_trans(root);
 321         mutex_unlock(&root->fs_info->trans_mutex);
 322 }
 323
 324 static int should_end_transaction(struct btrfs_trans_handle *trans,
 325                                   struct btrfs_root *root)
 326 {
 327         int ret;
 328         ret = btrfs_block_rsv_check(trans, root,
 329                                     &root->fs_info->global_block_rsv, 0, 5);
 330         return ret ? 1 : 0;
 331 }
 332
 333 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 334                                  struct btrfs_root *root)
 335 {
 336         struct btrfs_transaction *cur_trans = trans->transaction;
 337         int updates;
 338
 339         if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
 340                 return 1;
 341
 342         updates = trans->delayed_ref_updates;
 343         trans->delayed_ref_updates = 0;
 344         if (updates)
 345                 btrfs_run_delayed_refs(trans, root, updates);
 346
 347         return should_end_transaction(trans, root);
 348 }
 349
 350 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 351                           struct btrfs_root *root, int throttle)
 352 {
 353         struct btrfs_transaction *cur_trans = trans->transaction;
 354         struct btrfs_fs_info *info = root->fs_info;
 355         int count = 0;
 356
 357         while (count < 4) {
 358                 unsigned long cur = trans->delayed_ref_updates;
 359                 trans->delayed_ref_updates = 0;
 360                 if (cur &&
 361                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 362                         trans->delayed_ref_updates = 0;
 363
 364                         /*
 365                          * do a full flush if the transaction is trying
 366                          * to close
 367                          */
 368                         if (trans->transaction->delayed_refs.flushing)
 369                                 cur = 0;
 370                         btrfs_run_delayed_refs(trans, root, cur);
 371                 } else {
 372                         break;
 373                 }
 374                 count++;
 375         }
 376
 377         btrfs_trans_release_metadata(trans, root);
 378
 379         if (!root->fs_info->open_ioctl_trans &&
 380             should_end_transaction(trans, root))
 381                 trans->transaction->blocked = 1;
 382
 383         if (cur_trans->blocked && !cur_trans->in_commit) {
 384                 if (throttle)
 385                         return btrfs_commit_transaction(trans, root);
 386                 else
 387                         wake_up_process(info->transaction_kthread);
 388         }
 389
 390         mutex_lock(&info->trans_mutex);
 391         WARN_ON(cur_trans != info->running_transaction);
 392         WARN_ON(cur_trans->num_writers < 1);
 393         cur_trans->num_writers--;
 394
 395         if (waitqueue_active(&cur_trans->writer_wait))
 396                 wake_up(&cur_trans->writer_wait);
 397         put_transaction(cur_trans);
 398         mutex_unlock(&info->trans_mutex);
 399
 400         if (current->journal_info == trans)
 401                 current->journal_info = NULL;
 402         memset(trans, 0, sizeof(*trans));
 403         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 404
 405         if (throttle)
 406                 btrfs_run_delayed_iputs(root);
 407
 408         return 0;
 409 }
 410
 411 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 412                           struct btrfs_root *root)
 413 {
 414         return __btrfs_end_transaction(trans, root, 0);
 415 }
 416
 417 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 418                                    struct btrfs_root *root)
 419 {
 420         return __btrfs_end_transaction(trans, root, 1);
 421 }
 422
 423 /*
 424  * when btree blocks are allocated, they have some corresponding bits set for
 425  * them in one of two extent_io trees.  This is used to make sure all of
 426  * those extents are sent to disk but does not wait on them
 427  */
 428 int btrfs_write_marked_extents(struct btrfs_root *root,
 429                                struct extent_io_tree *dirty_pages, int mark)
 430 {
 431         int ret;
 432         int err = 0;
 433         int werr = 0;
 434         struct page *page;
 435         struct inode *btree_inode = root->fs_info->btree_inode;
 436         u64 start = 0;
 437         u64 end;
 438         unsigned long index;
 439
 440         while (1) {
 441                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 442                                             mark);
 443                 if (ret)
 444                         break;
 445                 while (start <= end) {
 446                         cond_resched();
 447
 448                         index = start >> PAGE_CACHE_SHIFT;
 449                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 450                         page = find_get_page(btree_inode->i_mapping, index);
 451                         if (!page)
 452                                 continue;
 453
 454                         btree_lock_page_hook(page);
 455                         if (!page->mapping) {
 456                                 unlock_page(page);
 457                                 page_cache_release(page);
 458                                 continue;
 459                         }
 460
 461                         if (PageWriteback(page)) {
 462                                 if (PageDirty(page))
 463                                         wait_on_page_writeback(page);
 464                                 else {
 465                                         unlock_page(page);
 466                                         page_cache_release(page);
 467                                         continue;
 468                                 }
 469                         }
 470                         err = write_one_page(page, 0);
 471                         if (err)
 472                                 werr = err;
 473                         page_cache_release(page);
 474                 }
 475         }
 476         if (err)
 477                 werr = err;
 478         return werr;
 479 }
 480
 481 /*
 482  * when btree blocks are allocated, they have some corresponding bits set for
 483  * them in one of two extent_io trees.  This is used to make sure all of
 484  * those extents are on disk for transaction or log commit.  We wait
 485  * on all the pages and clear them from the dirty pages state tree
 486  */
 487 int btrfs_wait_marked_extents(struct btrfs_root *root,
 488                               struct extent_io_tree *dirty_pages, int mark)
 489 {
 490         int ret;
 491         int err = 0;
 492         int werr = 0;
 493         struct page *page;
 494         struct inode *btree_inode = root->fs_info->btree_inode;
 495         u64 start = 0;
 496         u64 end;
 497         unsigned long index;
 498
 499         while (1) {
 500                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 501                                             mark);
 502                 if (ret)
 503                         break;
 504
 505                 clear_extent_bits(dirty_pages, start, end, mark, GFP_NOFS);
 506                 while (start <= end) {
 507                         index = start >> PAGE_CACHE_SHIFT;
 508                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 509                         page = find_get_page(btree_inode->i_mapping, index);
 510                         if (!page)
 511                                 continue;
 512                         if (PageDirty(page)) {
 513                                 btree_lock_page_hook(page);
 514                                 wait_on_page_writeback(page);
 515                                 err = write_one_page(page, 0);
 516                                 if (err)
 517                                         werr = err;
 518                         }
 519                         wait_on_page_writeback(page);
 520                         page_cache_release(page);
 521                         cond_resched();
 522                 }
 523         }
 524         if (err)
 525                 werr = err;
 526         return werr;
 527 }
 528
 529 /*
 530  * when btree blocks are allocated, they have some corresponding bits set for
 531  * them in one of two extent_io trees.  This is used to make sure all of
 532  * those extents are on disk for transaction or log commit
 533  */
 534 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 535                                 struct extent_io_tree *dirty_pages, int mark)
 536 {
 537         int ret;
 538         int ret2;
 539
 540         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 541         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 542         return ret || ret2;
 543 }
 544
 545 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 546                                      struct btrfs_root *root)
 547 {
 548         if (!trans || !trans->transaction) {
 549                 struct inode *btree_inode;
 550                 btree_inode = root->fs_info->btree_inode;
 551                 return filemap_write_and_wait(btree_inode->i_mapping);
 552         }
 553         return btrfs_write_and_wait_marked_extents(root,
 554                                            &trans->transaction->dirty_pages,
 555                                            EXTENT_DIRTY);
 556 }
 557
 558 /*
 559  * this is used to update the root pointer in the tree of tree roots.
 560  *
 561  * But, in the case of the extent allocation tree, updating the root
 562  * pointer may allocate blocks which may change the root of the extent
 563  * allocation tree.
 564  *
 565  * So, this loops and repeats and makes sure the cowonly root didn't
 566  * change while the root pointer was being updated in the metadata.
 567  */
 568 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 569                                struct btrfs_root *root)
 570 {
 571         int ret;
 572         u64 old_root_bytenr;
 573         u64 old_root_used;
 574         struct btrfs_root *tree_root = root->fs_info->tree_root;
 575
 576         old_root_used = btrfs_root_used(&root->root_item);
 577         btrfs_write_dirty_block_groups(trans, root);
 578
 579         while (1) {
 580                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 581                 if (old_root_bytenr == root->node->start &&
 582                     old_root_used == btrfs_root_used(&root->root_item))
 583                         break;
 584
 585                 btrfs_set_root_node(&root->root_item, root->node);
 586                 ret = btrfs_update_root(trans, tree_root,
 587                                         &root->root_key,
 588                                         &root->root_item);
 589                 BUG_ON(ret);
 590
 591                 old_root_used = btrfs_root_used(&root->root_item);
 592                 ret = btrfs_write_dirty_block_groups(trans, root);
 593                 BUG_ON(ret);
 594         }
 595
 596         if (root != root->fs_info->extent_root)
 597                 switch_commit_root(root);
 598
 599         return 0;
 600 }
 601
 602 /*
 603  * update all the cowonly tree roots on disk
 604  */
 605 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 606                                          struct btrfs_root *root)
 607 {
 608         struct btrfs_fs_info *fs_info = root->fs_info;
 609         struct list_head *next;
 610         struct extent_buffer *eb;
 611         int ret;
 612
 613         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 614         BUG_ON(ret);
 615
 616         eb = btrfs_lock_root_node(fs_info->tree_root);
 617         btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
 618         btrfs_tree_unlock(eb);
 619         free_extent_buffer(eb);
 620
 621         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 622         BUG_ON(ret);
 623
 624         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 625                 next = fs_info->dirty_cowonly_roots.next;
 626                 list_del_init(next);
 627                 root = list_entry(next, struct btrfs_root, dirty_list);
 628
 629                 update_cowonly_root(trans, root);
 630         }
 631
 632         down_write(&fs_info->extent_commit_sem);
 633         switch_commit_root(fs_info->extent_root);
 634         up_write(&fs_info->extent_commit_sem);
 635
 636         return 0;
 637 }
 638
 639 /*
 640  * dead roots are old snapshots that need to be deleted.  This allocates
 641  * a dirty root struct and adds it into the list of dead roots that need to
 642  * be deleted
 643  */
 644 int btrfs_add_dead_root(struct btrfs_root *root)
 645 {
 646         mutex_lock(&root->fs_info->trans_mutex);
 647         list_add(&root->root_list, &root->fs_info->dead_roots);
 648         mutex_unlock(&root->fs_info->trans_mutex);
 649         return 0;
 650 }
 651
 652 /*
 653  * update all the cowonly tree roots on disk
 654  */
 655 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 656                                     struct btrfs_root *root)
 657 {
 658         struct btrfs_root *gang[8];
 659         struct btrfs_fs_info *fs_info = root->fs_info;
 660         int i;
 661         int ret;
 662         int err = 0;
 663
 664         while (1) {
 665                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 666                                                  (void **)gang, 0,
 667                                                  ARRAY_SIZE(gang),
 668                                                  BTRFS_ROOT_TRANS_TAG);
 669                 if (ret == 0)
 670                         break;
 671                 for (i = 0; i < ret; i++) {
 672                         root = gang[i];
 673                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 674                                         (unsigned long)root->root_key.objectid,
 675                                         BTRFS_ROOT_TRANS_TAG);
 676
 677                         btrfs_free_log(trans, root);
 678                         btrfs_update_reloc_root(trans, root);
 679                         btrfs_orphan_commit_root(trans, root);
 680
 681                         if (root->commit_root != root->node) {
 682                                 switch_commit_root(root);
 683                                 btrfs_set_root_node(&root->root_item,
 684                                                     root->node);
 685                         }
 686
 687                         err = btrfs_update_root(trans, fs_info->tree_root,
 688                                                 &root->root_key,
 689                                                 &root->root_item);
 690                         if (err)
 691                                 break;
 692                 }
 693         }
 694         return err;
 695 }
 696
 697 /*
 698  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 699  * otherwise every leaf in the btree is read and defragged.
 700  */
 701 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 702 {
 703         struct btrfs_fs_info *info = root->fs_info;
 704         struct btrfs_trans_handle *trans;
 705         int ret;
 706         unsigned long nr;
 707
 708         if (xchg(&root->defrag_running, 1))
 709                 return 0;
 710
 711         while (1) {
 712                 trans = btrfs_start_transaction(root, 0);
 713                 if (IS_ERR(trans))
 714                         return PTR_ERR(trans);
 715
 716                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 717
 718                 nr = trans->blocks_used;
 719                 btrfs_end_transaction(trans, root);
 720                 btrfs_btree_balance_dirty(info->tree_root, nr);
 721                 cond_resched();
 722
 723                 if (root->fs_info->closing || ret != -EAGAIN)
 724                         break;
 725         }
 726         root->defrag_running = 0;
 727         return ret;
 728 }
 729
 730 #if 0
 731 /*
 732  * when dropping snapshots, we generate a ton of delayed refs, and it makes
 733  * sense not to join the transaction while it is trying to flush the current
 734  * queue of delayed refs out.
 735  *
 736  * This is used by the drop snapshot code only
 737  */
 738 static noinline int wait_transaction_pre_flush(struct btrfs_fs_info *info)
 739 {
 740         DEFINE_WAIT(wait);
 741
 742         mutex_lock(&info->trans_mutex);
 743         while (info->running_transaction &&
 744                info->running_transaction->delayed_refs.flushing) {
 745                 prepare_to_wait(&info->transaction_wait, &wait,
 746                                 TASK_UNINTERRUPTIBLE);
 747                 mutex_unlock(&info->trans_mutex);
 748
 749                 schedule();
 750
 751                 mutex_lock(&info->trans_mutex);
 752                 finish_wait(&info->transaction_wait, &wait);
 753         }
 754         mutex_unlock(&info->trans_mutex);
 755         return 0;
 756 }
 757
 758 /*
 759  * Given a list of roots that need to be deleted, call btrfs_drop_snapshot on
 760  * all of them
 761  */
 762 int btrfs_drop_dead_root(struct btrfs_root *root)
 763 {
 764         struct btrfs_trans_handle *trans;
 765         struct btrfs_root *tree_root = root->fs_info->tree_root;
 766         unsigned long nr;
 767         int ret;
 768
 769         while (1) {
 770                 /*
 771                  * we don't want to jump in and create a bunch of
 772                  * delayed refs if the transaction is starting to close
 773                  */
 774                 wait_transaction_pre_flush(tree_root->fs_info);
 775                 trans = btrfs_start_transaction(tree_root, 1);
 776
 777                 /*
 778                  * we've joined a transaction, make sure it isn't
 779                  * closing right now
 780                  */
 781                 if (trans->transaction->delayed_refs.flushing) {
 782                         btrfs_end_transaction(trans, tree_root);
 783                         continue;
 784                 }
 785
 786                 ret = btrfs_drop_snapshot(trans, root);
 787                 if (ret != -EAGAIN)
 788                         break;
 789
 790                 ret = btrfs_update_root(trans, tree_root,
 791                                         &root->root_key,
 792                                         &root->root_item);
 793                 if (ret)
 794                         break;
 795
 796                 nr = trans->blocks_used;
 797                 ret = btrfs_end_transaction(trans, tree_root);
 798                 BUG_ON(ret);
 799
 800                 btrfs_btree_balance_dirty(tree_root, nr);
 801                 cond_resched();
 802         }
 803         BUG_ON(ret);
 804
 805         ret = btrfs_del_root(trans, tree_root, &root->root_key);
 806         BUG_ON(ret);
 807
 808         nr = trans->blocks_used;
 809         ret = btrfs_end_transaction(trans, tree_root);
 810         BUG_ON(ret);
 811
 812         free_extent_buffer(root->node);
 813         free_extent_buffer(root->commit_root);
 814         kfree(root);
 815
 816         btrfs_btree_balance_dirty(tree_root, nr);
 817         return ret;
 818 }
 819 #endif
 820
 821 /*
 822  * new snapshots need to be created at a very specific time in the
 823  * transaction commit.  This does the actual creation
 824  */
 825 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 826                                    struct btrfs_fs_info *fs_info,
 827                                    struct btrfs_pending_snapshot *pending)
 828 {
 829         struct btrfs_key key;
 830         struct btrfs_root_item *new_root_item;
 831         struct btrfs_root *tree_root = fs_info->tree_root;
 832         struct btrfs_root *root = pending->root;
 833         struct btrfs_root *parent_root;
 834         struct inode *parent_inode;
 835         struct dentry *dentry;
 836         struct extent_buffer *tmp;
 837         struct extent_buffer *old;
 838         int ret;
 839         int retries = 0;
 840         u64 to_reserve = 0;
 841         u64 index = 0;
 842         u64 objectid;
 843
 844         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 845         if (!new_root_item) {
 846                 pending->error = -ENOMEM;
 847                 goto fail;
 848         }
 849
 850         ret = btrfs_find_free_objectid(trans, tree_root, 0, &objectid);
 851         if (ret) {
 852                 pending->error = ret;
 853                 goto fail;
 854         }
 855
 856         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
 857         btrfs_orphan_pre_snapshot(trans, pending, &to_reserve);
 858
 859         if (to_reserve > 0) {
 860                 ret = btrfs_block_rsv_add(trans, root, &pending->block_rsv,
 861                                           to_reserve, &retries);
 862                 if (ret) {
 863                         pending->error = ret;
 864                         goto fail;
 865                 }
 866         }
 867
 868         key.objectid = objectid;
 869         key.offset = (u64)-1;
 870         key.type = BTRFS_ROOT_ITEM_KEY;
 871
 872         trans->block_rsv = &pending->block_rsv;
 873
 874         dentry = pending->dentry;
 875         parent_inode = dentry->d_parent->d_inode;
 876         parent_root = BTRFS_I(parent_inode)->root;
 877         record_root_in_trans(trans, parent_root);
 878
 879         /*
 880          * insert the directory item
 881          */
 882         ret = btrfs_set_inode_index(parent_inode, &index);
 883         BUG_ON(ret);
 884         ret = btrfs_insert_dir_item(trans, parent_root,
 885                                 dentry->d_name.name, dentry->d_name.len,
 886                                 parent_inode->i_ino, &key,
 887                                 BTRFS_FT_DIR, index);
 888         BUG_ON(ret);
 889
 890         btrfs_i_size_write(parent_inode, parent_inode->i_size +
 891                                          dentry->d_name.len * 2);
 892         ret = btrfs_update_inode(trans, parent_root, parent_inode);
 893         BUG_ON(ret);
 894
 895         record_root_in_trans(trans, root);
 896         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
 897         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
 898
 899         old = btrfs_lock_root_node(root);
 900         btrfs_cow_block(trans, root, old, NULL, 0, &old);
 901         btrfs_set_lock_blocking(old);
 902
 903         btrfs_copy_root(trans, root, old, &tmp, objectid);
 904         btrfs_tree_unlock(old);
 905         free_extent_buffer(old);
 906
 907         btrfs_set_root_node(new_root_item, tmp);
 908         /* record when the snapshot was created in key.offset */
 909         key.offset = trans->transid;
 910         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
 911         btrfs_tree_unlock(tmp);
 912         free_extent_buffer(tmp);
 913         BUG_ON(ret);
 914
 915         /*
 916          * insert root back/forward references
 917          */
 918         ret = btrfs_add_root_ref(trans, tree_root, objectid,
 919                                  parent_root->root_key.objectid,
 920                                  parent_inode->i_ino, index,
 921                                  dentry->d_name.name, dentry->d_name.len);
 922         BUG_ON(ret);
 923
 924         key.offset = (u64)-1;
 925         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
 926         BUG_ON(IS_ERR(pending->snap));
 927
 928         btrfs_reloc_post_snapshot(trans, pending);
 929         btrfs_orphan_post_snapshot(trans, pending);
 930 fail:
 931         kfree(new_root_item);
 932         btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
 933         return 0;
 934 }
 935
 936 /*
 937  * create all the snapshots we've scheduled for creation
 938  */
 939 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
 940                                              struct btrfs_fs_info *fs_info)
 941 {
 942         struct btrfs_pending_snapshot *pending;
 943         struct list_head *head = &trans->transaction->pending_snapshots;
 944         int ret;
 945
 946         list_for_each_entry(pending, head, list) {
 947                 ret = create_pending_snapshot(trans, fs_info, pending);
 948                 BUG_ON(ret);
 949         }
 950         return 0;
 951 }
 952
 953 static void update_super_roots(struct btrfs_root *root)
 954 {
 955         struct btrfs_root_item *root_item;
 956         struct btrfs_super_block *super;
 957
 958         super = &root->fs_info->super_copy;
 959
 960         root_item = &root->fs_info->chunk_root->root_item;
 961         super->chunk_root = root_item->bytenr;
 962         super->chunk_root_generation = root_item->generation;
 963         super->chunk_root_level = root_item->level;
 964
 965         root_item = &root->fs_info->tree_root->root_item;
 966         super->root = root_item->bytenr;
 967         super->generation = root_item->generation;
 968         super->root_level = root_item->level;
 969 }
 970
 971 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
 972 {
 973         int ret = 0;
 974         spin_lock(&info->new_trans_lock);
 975         if (info->running_transaction)
 976                 ret = info->running_transaction->in_commit;
 977         spin_unlock(&info->new_trans_lock);
 978         return ret;
 979 }
 980
 981 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
 982 {
 983         int ret = 0;
 984         spin_lock(&info->new_trans_lock);
 985         if (info->running_transaction)
 986                 ret = info->running_transaction->blocked;
 987         spin_unlock(&info->new_trans_lock);
 988         return ret;
 989 }
 990
 991 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
 992                              struct btrfs_root *root)
 993 {
 994         unsigned long joined = 0;
 995         unsigned long timeout = 1;
 996         struct btrfs_transaction *cur_trans;
 997         struct btrfs_transaction *prev_trans = NULL;
 998         DEFINE_WAIT(wait);
 999         int ret;
1000         int should_grow = 0;
1001         unsigned long now = get_seconds();
1002         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1003
1004         btrfs_run_ordered_operations(root, 0);
1005
1006         /* make a pass through all the delayed refs we have so far
1007          * any runnings procs may add more while we are here
1008          */
1009         ret = btrfs_run_delayed_refs(trans, root, 0);
1010         BUG_ON(ret);
1011
1012         btrfs_trans_release_metadata(trans, root);
1013
1014         cur_trans = trans->transaction;
1015         /*
1016          * set the flushing flag so procs in this transaction have to
1017          * start sending their work down.
1018          */
1019         cur_trans->delayed_refs.flushing = 1;
1020
1021         ret = btrfs_run_delayed_refs(trans, root, 0);
1022         BUG_ON(ret);
1023
1024         mutex_lock(&root->fs_info->trans_mutex);
1025         if (cur_trans->in_commit) {
1026                 cur_trans->use_count++;
1027                 mutex_unlock(&root->fs_info->trans_mutex);
1028                 btrfs_end_transaction(trans, root);
1029
1030                 ret = wait_for_commit(root, cur_trans);
1031                 BUG_ON(ret);
1032
1033                 mutex_lock(&root->fs_info->trans_mutex);
1034                 put_transaction(cur_trans);
1035                 mutex_unlock(&root->fs_info->trans_mutex);
1036
1037                 return 0;
1038         }
1039
1040         trans->transaction->in_commit = 1;
1041         trans->transaction->blocked = 1;
1042         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1043                 prev_trans = list_entry(cur_trans->list.prev,
1044                                         struct btrfs_transaction, list);
1045                 if (!prev_trans->commit_done) {
1046                         prev_trans->use_count++;
1047                         mutex_unlock(&root->fs_info->trans_mutex);
1048
1049                         wait_for_commit(root, prev_trans);
1050
1051                         mutex_lock(&root->fs_info->trans_mutex);
1052                         put_transaction(prev_trans);
1053                 }
1054         }
1055
1056         if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
1057                 should_grow = 1;
1058
1059         do {
1060                 int snap_pending = 0;
1061                 joined = cur_trans->num_joined;
1062                 if (!list_empty(&trans->transaction->pending_snapshots))
1063                         snap_pending = 1;
1064
1065                 WARN_ON(cur_trans != trans->transaction);
1066                 if (cur_trans->num_writers > 1)
1067                         timeout = MAX_SCHEDULE_TIMEOUT;
1068                 else if (should_grow)
1069                         timeout = 1;
1070
1071                 mutex_unlock(&root->fs_info->trans_mutex);
1072
1073                 if (flush_on_commit || snap_pending) {
1074                         btrfs_start_delalloc_inodes(root, 1);
1075                         ret = btrfs_wait_ordered_extents(root, 0, 1);
1076                         BUG_ON(ret);
1077                 }
1078
1079                 /*
1080                  * rename don't use btrfs_join_transaction, so, once we
1081                  * set the transaction to blocked above, we aren't going
1082                  * to get any new ordered operations.  We can safely run
1083                  * it here and no for sure that nothing new will be added
1084                  * to the list
1085                  */
1086                 btrfs_run_ordered_operations(root, 1);
1087
1088                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1089                                 TASK_UNINTERRUPTIBLE);
1090
1091                 smp_mb();
1092                 if (cur_trans->num_writers > 1 || should_grow)
1093                         schedule_timeout(timeout);
1094
1095                 mutex_lock(&root->fs_info->trans_mutex);
1096                 finish_wait(&cur_trans->writer_wait, &wait);
1097         } while (cur_trans->num_writers > 1 ||
1098                  (should_grow && cur_trans->num_joined != joined));
1099
1100         ret = create_pending_snapshots(trans, root->fs_info);
1101         BUG_ON(ret);
1102
1103         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1104         BUG_ON(ret);
1105
1106         WARN_ON(cur_trans != trans->transaction);
1107
1108         /* btrfs_commit_tree_roots is responsible for getting the
1109          * various roots consistent with each other.  Every pointer
1110          * in the tree of tree roots has to point to the most up to date
1111          * root for every subvolume and other tree.  So, we have to keep
1112          * the tree logging code from jumping in and changing any
1113          * of the trees.
1114          *
1115          * At this point in the commit, there can't be any tree-log
1116          * writers, but a little lower down we drop the trans mutex
1117          * and let new people in.  By holding the tree_log_mutex
1118          * from now until after the super is written, we avoid races
1119          * with the tree-log code.
1120          */
1121         mutex_lock(&root->fs_info->tree_log_mutex);
1122
1123         ret = commit_fs_roots(trans, root);
1124         BUG_ON(ret);
1125
1126         /* commit_fs_roots gets rid of all the tree log roots, it is now
1127          * safe to free the root of tree log roots
1128          */
1129         btrfs_free_log_root_tree(trans, root->fs_info);
1130
1131         ret = commit_cowonly_roots(trans, root);
1132         BUG_ON(ret);
1133
1134         btrfs_prepare_extent_commit(trans, root);
1135
1136         cur_trans = root->fs_info->running_transaction;
1137         spin_lock(&root->fs_info->new_trans_lock);
1138         root->fs_info->running_transaction = NULL;
1139         spin_unlock(&root->fs_info->new_trans_lock);
1140
1141         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1142                             root->fs_info->tree_root->node);
1143         switch_commit_root(root->fs_info->tree_root);
1144
1145         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1146                             root->fs_info->chunk_root->node);
1147         switch_commit_root(root->fs_info->chunk_root);
1148
1149         update_super_roots(root);
1150
1151         if (!root->fs_info->log_root_recovering) {
1152                 btrfs_set_super_log_root(&root->fs_info->super_copy, 0);
1153                 btrfs_set_super_log_root_level(&root->fs_info->super_copy, 0);
1154         }
1155
1156         memcpy(&root->fs_info->super_for_commit, &root->fs_info->super_copy,
1157                sizeof(root->fs_info->super_copy));
1158
1159         trans->transaction->blocked = 0;
1160
1161         wake_up(&root->fs_info->transaction_wait);
1162
1163         mutex_unlock(&root->fs_info->trans_mutex);
1164         ret = btrfs_write_and_wait_transaction(trans, root);
1165         BUG_ON(ret);
1166         write_ctree_super(trans, root, 0);
1167
1168         /*
1169          * the super is written, we can safely allow the tree-loggers
1170          * to go about their business
1171          */
1172         mutex_unlock(&root->fs_info->tree_log_mutex);
1173
1174         btrfs_finish_extent_commit(trans, root);
1175
1176         mutex_lock(&root->fs_info->trans_mutex);
1177
1178         cur_trans->commit_done = 1;
1179
1180         root->fs_info->last_trans_committed = cur_trans->transid;
1181
1182         wake_up(&cur_trans->commit_wait);
1183
1184         put_transaction(cur_trans);
1185         put_transaction(cur_trans);
1186
1187         mutex_unlock(&root->fs_info->trans_mutex);
1188
1189         if (current->journal_info == trans)
1190                 current->journal_info = NULL;
1191
1192         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1193
1194         if (current != root->fs_info->transaction_kthread)
1195                 btrfs_run_delayed_iputs(root);
1196
1197         return ret;
1198 }
1199
1200 /*
1201  * interface function to delete all the snapshots we have scheduled for deletion
1202  */
1203 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1204 {
1205         LIST_HEAD(list);
1206         struct btrfs_fs_info *fs_info = root->fs_info;
1207
1208         mutex_lock(&fs_info->trans_mutex);
1209         list_splice_init(&fs_info->dead_roots, &list);
1210         mutex_unlock(&fs_info->trans_mutex);
1211
1212         while (!list_empty(&list)) {
1213                 root = list_entry(list.next, struct btrfs_root, root_list);
1214                 list_del(&root->root_list);
1215
1216                 if (btrfs_header_backref_rev(root->node) <
1217                     BTRFS_MIXED_BACKREF_REV)
1218                         btrfs_drop_snapshot(root, NULL, 0);
1219                 else
1220                         btrfs_drop_snapshot(root, NULL, 1);
1221         }
1222         return 0;
1223 }