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/sched.h>
  21 #include <linux/writeback.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/blkdev.h>
  24 #include "ctree.h"
  25 #include "disk-io.h"
  26 #include "transaction.h"
  27 #include "locking.h"
  28 #include "tree-log.h"
  29
  30 #define BTRFS_ROOT_TRANS_TAG 0
  31
  32 static noinline void put_transaction(struct btrfs_transaction *transaction)
  33 {
  34         WARN_ON(transaction->use_count == 0);
  35         transaction->use_count--;
  36         if (transaction->use_count == 0) {
  37                 list_del_init(&transaction->list);
  38                 memset(transaction, 0, sizeof(*transaction));
  39                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  40         }
  41 }
  42
  43 static noinline void switch_commit_root(struct btrfs_root *root)
  44 {
  45         free_extent_buffer(root->commit_root);
  46         root->commit_root = btrfs_root_node(root);
  47 }
  48
  49 /*
  50  * either allocate a new transaction or hop into the existing one
  51  */
  52 static noinline int join_transaction(struct btrfs_root *root)
  53 {
  54         struct btrfs_transaction *cur_trans;
  55         cur_trans = root->fs_info->running_transaction;
  56         if (!cur_trans) {
  57                 cur_trans = kmem_cache_alloc(btrfs_transaction_cachep,
  58                                              GFP_NOFS);
  59                 BUG_ON(!cur_trans);
  60                 root->fs_info->generation++;
  61                 cur_trans->num_writers = 1;
  62                 cur_trans->num_joined = 0;
  63                 cur_trans->transid = root->fs_info->generation;
  64                 init_waitqueue_head(&cur_trans->writer_wait);
  65                 init_waitqueue_head(&cur_trans->commit_wait);
  66                 cur_trans->in_commit = 0;
  67                 cur_trans->blocked = 0;
  68                 cur_trans->use_count = 1;
  69                 cur_trans->commit_done = 0;
  70                 cur_trans->start_time = get_seconds();
  71
  72                 cur_trans->delayed_refs.root.rb_node = NULL;
  73                 cur_trans->delayed_refs.num_entries = 0;
  74                 cur_trans->delayed_refs.num_heads_ready = 0;
  75                 cur_trans->delayed_refs.num_heads = 0;
  76                 cur_trans->delayed_refs.flushing = 0;
  77                 cur_trans->delayed_refs.run_delayed_start = 0;
  78                 spin_lock_init(&cur_trans->delayed_refs.lock);
  79
  80                 INIT_LIST_HEAD(&cur_trans->pending_snapshots);
  81                 list_add_tail(&cur_trans->list, &root->fs_info->trans_list);
  82                 extent_io_tree_init(&cur_trans->dirty_pages,
  83                                      root->fs_info->btree_inode->i_mapping,
  84                                      GFP_NOFS);
  85                 spin_lock(&root->fs_info->new_trans_lock);
  86                 root->fs_info->running_transaction = cur_trans;
  87                 spin_unlock(&root->fs_info->new_trans_lock);
  88         } else {
  89                 cur_trans->num_writers++;
  90                 cur_trans->num_joined++;
  91         }
  92
  93         return 0;
  94 }
  95
  96 /*
  97  * this does all the record keeping required to make sure that a reference
  98  * counted root is properly recorded in a given transaction.  This is required
  99  * to make sure the old root from before we joined the transaction is deleted
 100  * when the transaction commits
 101  */
 102 static noinline int record_root_in_trans(struct btrfs_trans_handle *trans,
 103                                          struct btrfs_root *root)
 104 {
 105         if (root->ref_cows && root->last_trans < trans->transid) {
 106                 WARN_ON(root == root->fs_info->extent_root);
 107                 WARN_ON(root->commit_root != root->node);
 108
 109                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 110                            (unsigned long)root->root_key.objectid,
 111                            BTRFS_ROOT_TRANS_TAG);
 112                 root->last_trans = trans->transid;
 113                 btrfs_init_reloc_root(trans, root);
 114         }
 115         return 0;
 116 }
 117
 118 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 119                                struct btrfs_root *root)
 120 {
 121         if (!root->ref_cows)
 122                 return 0;
 123
 124         mutex_lock(&root->fs_info->trans_mutex);
 125         if (root->last_trans == trans->transid) {
 126                 mutex_unlock(&root->fs_info->trans_mutex);
 127                 return 0;
 128         }
 129
 130         record_root_in_trans(trans, root);
 131         mutex_unlock(&root->fs_info->trans_mutex);
 132         return 0;
 133 }
 134
 135 /* wait for commit against the current transaction to become unblocked
 136  * when this is done, it is safe to start a new transaction, but the current
 137  * transaction might not be fully on disk.
 138  */
 139 static void wait_current_trans(struct btrfs_root *root)
 140 {
 141         struct btrfs_transaction *cur_trans;
 142
 143         cur_trans = root->fs_info->running_transaction;
 144         if (cur_trans && cur_trans->blocked) {
 145                 DEFINE_WAIT(wait);
 146                 cur_trans->use_count++;
 147                 while (1) {
 148                         prepare_to_wait(&root->fs_info->transaction_wait, &wait,
 149                                         TASK_UNINTERRUPTIBLE);
 150                         if (cur_trans->blocked) {
 151                                 mutex_unlock(&root->fs_info->trans_mutex);
 152                                 schedule();
 153                                 mutex_lock(&root->fs_info->trans_mutex);
 154                                 finish_wait(&root->fs_info->transaction_wait,
 155                                             &wait);
 156                         } else {
 157                                 finish_wait(&root->fs_info->transaction_wait,
 158                                             &wait);
 159                                 break;
 160                         }
 161                 }
 162                 put_transaction(cur_trans);
 163         }
 164 }
 165
 166 enum btrfs_trans_type {
 167         TRANS_START,
 168         TRANS_JOIN,
 169         TRANS_USERSPACE,
 170 };
 171
 172 static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
 173                                              int num_blocks, int type)
 174 {
 175         struct btrfs_trans_handle *h =
 176                 kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 177         int ret;
 178
 179         mutex_lock(&root->fs_info->trans_mutex);
 180         if (!root->fs_info->log_root_recovering &&
 181             ((type == TRANS_START && !root->fs_info->open_ioctl_trans) ||
 182              type == TRANS_USERSPACE))
 183                 wait_current_trans(root);
 184         ret = join_transaction(root);
 185         BUG_ON(ret);
 186
 187         h->transid = root->fs_info->running_transaction->transid;
 188         h->transaction = root->fs_info->running_transaction;
 189         h->blocks_reserved = num_blocks;
 190         h->blocks_used = 0;
 191         h->block_group = 0;
 192         h->alloc_exclude_nr = 0;
 193         h->alloc_exclude_start = 0;
 194         h->delayed_ref_updates = 0;
 195
 196         if (!current->journal_info && type != TRANS_USERSPACE)
 197                 current->journal_info = h;
 198
 199         root->fs_info->running_transaction->use_count++;
 200         record_root_in_trans(h, root);
 201         mutex_unlock(&root->fs_info->trans_mutex);
 202         return h;
 203 }
 204
 205 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 206                                                    int num_blocks)
 207 {
 208         return start_transaction(root, num_blocks, TRANS_START);
 209 }
 210 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root,
 211                                                    int num_blocks)
 212 {
 213         return start_transaction(root, num_blocks, TRANS_JOIN);
 214 }
 215
 216 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *r,
 217                                                          int num_blocks)
 218 {
 219         return start_transaction(r, num_blocks, TRANS_USERSPACE);
 220 }
 221
 222 /* wait for a transaction commit to be fully complete */
 223 static noinline int wait_for_commit(struct btrfs_root *root,
 224                                     struct btrfs_transaction *commit)
 225 {
 226         DEFINE_WAIT(wait);
 227         mutex_lock(&root->fs_info->trans_mutex);
 228         while (!commit->commit_done) {
 229                 prepare_to_wait(&commit->commit_wait, &wait,
 230                                 TASK_UNINTERRUPTIBLE);
 231                 if (commit->commit_done)
 232                         break;
 233                 mutex_unlock(&root->fs_info->trans_mutex);
 234                 schedule();
 235                 mutex_lock(&root->fs_info->trans_mutex);
 236         }
 237         mutex_unlock(&root->fs_info->trans_mutex);
 238         finish_wait(&commit->commit_wait, &wait);
 239         return 0;
 240 }
 241
 242 #if 0
 243 /*
 244  * rate limit against the drop_snapshot code.  This helps to slow down new
 245  * operations if the drop_snapshot code isn't able to keep up.
 246  */
 247 static void throttle_on_drops(struct btrfs_root *root)
 248 {
 249         struct btrfs_fs_info *info = root->fs_info;
 250         int harder_count = 0;
 251
 252 harder:
 253         if (atomic_read(&info->throttles)) {
 254                 DEFINE_WAIT(wait);
 255                 int thr;
 256                 thr = atomic_read(&info->throttle_gen);
 257
 258                 do {
 259                         prepare_to_wait(&info->transaction_throttle,
 260                                         &wait, TASK_UNINTERRUPTIBLE);
 261                         if (!atomic_read(&info->throttles)) {
 262                                 finish_wait(&info->transaction_throttle, &wait);
 263                                 break;
 264                         }
 265                         schedule();
 266                         finish_wait(&info->transaction_throttle, &wait);
 267                 } while (thr == atomic_read(&info->throttle_gen));
 268                 harder_count++;
 269
 270                 if (root->fs_info->total_ref_cache_size > 1 * 1024 * 1024 &&
 271                     harder_count < 2)
 272                         goto harder;
 273
 274                 if (root->fs_info->total_ref_cache_size > 5 * 1024 * 1024 &&
 275                     harder_count < 10)
 276                         goto harder;
 277
 278                 if (root->fs_info->total_ref_cache_size > 10 * 1024 * 1024 &&
 279                     harder_count < 20)
 280                         goto harder;
 281         }
 282 }
 283 #endif
 284
 285 void btrfs_throttle(struct btrfs_root *root)
 286 {
 287         mutex_lock(&root->fs_info->trans_mutex);
 288         if (!root->fs_info->open_ioctl_trans)
 289                 wait_current_trans(root);
 290         mutex_unlock(&root->fs_info->trans_mutex);
 291 }
 292
 293 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 294                           struct btrfs_root *root, int throttle)
 295 {
 296         struct btrfs_transaction *cur_trans;
 297         struct btrfs_fs_info *info = root->fs_info;
 298         int count = 0;
 299
 300         while (count < 4) {
 301                 unsigned long cur = trans->delayed_ref_updates;
 302                 trans->delayed_ref_updates = 0;
 303                 if (cur &&
 304                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 305                         trans->delayed_ref_updates = 0;
 306
 307                         /*
 308                          * do a full flush if the transaction is trying
 309                          * to close
 310                          */
 311                         if (trans->transaction->delayed_refs.flushing)
 312                                 cur = 0;
 313                         btrfs_run_delayed_refs(trans, root, cur);
 314                 } else {
 315                         break;
 316                 }
 317                 count++;
 318         }
 319
 320         mutex_lock(&info->trans_mutex);
 321         cur_trans = info->running_transaction;
 322         WARN_ON(cur_trans != trans->transaction);
 323         WARN_ON(cur_trans->num_writers < 1);
 324         cur_trans->num_writers--;
 325
 326         if (waitqueue_active(&cur_trans->writer_wait))
 327                 wake_up(&cur_trans->writer_wait);
 328         put_transaction(cur_trans);
 329         mutex_unlock(&info->trans_mutex);
 330
 331         if (current->journal_info == trans)
 332                 current->journal_info = NULL;
 333         memset(trans, 0, sizeof(*trans));
 334         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 335
 336         if (throttle)
 337                 btrfs_run_delayed_iputs(root);
 338
 339         return 0;
 340 }
 341
 342 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 343                           struct btrfs_root *root)
 344 {
 345         return __btrfs_end_transaction(trans, root, 0);
 346 }
 347
 348 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 349                                    struct btrfs_root *root)
 350 {
 351         return __btrfs_end_transaction(trans, root, 1);
 352 }
 353
 354 /*
 355  * when btree blocks are allocated, they have some corresponding bits set for
 356  * them in one of two extent_io trees.  This is used to make sure all of
 357  * those extents are sent to disk but does not wait on them
 358  */
 359 int btrfs_write_marked_extents(struct btrfs_root *root,
 360                                struct extent_io_tree *dirty_pages, int mark)
 361 {
 362         int ret;
 363         int err = 0;
 364         int werr = 0;
 365         struct page *page;
 366         struct inode *btree_inode = root->fs_info->btree_inode;
 367         u64 start = 0;
 368         u64 end;
 369         unsigned long index;
 370
 371         while (1) {
 372                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 373                                             mark);
 374                 if (ret)
 375                         break;
 376                 while (start <= end) {
 377                         cond_resched();
 378
 379                         index = start >> PAGE_CACHE_SHIFT;
 380                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 381                         page = find_get_page(btree_inode->i_mapping, index);
 382                         if (!page)
 383                                 continue;
 384
 385                         btree_lock_page_hook(page);
 386                         if (!page->mapping) {
 387                                 unlock_page(page);
 388                                 page_cache_release(page);
 389                                 continue;
 390                         }
 391
 392                         if (PageWriteback(page)) {
 393                                 if (PageDirty(page))
 394                                         wait_on_page_writeback(page);
 395                                 else {
 396                                         unlock_page(page);
 397                                         page_cache_release(page);
 398                                         continue;
 399                                 }
 400                         }
 401                         err = write_one_page(page, 0);
 402                         if (err)
 403                                 werr = err;
 404                         page_cache_release(page);
 405                 }
 406         }
 407         if (err)
 408                 werr = err;
 409         return werr;
 410 }
 411
 412 /*
 413  * when btree blocks are allocated, they have some corresponding bits set for
 414  * them in one of two extent_io trees.  This is used to make sure all of
 415  * those extents are on disk for transaction or log commit.  We wait
 416  * on all the pages and clear them from the dirty pages state tree
 417  */
 418 int btrfs_wait_marked_extents(struct btrfs_root *root,
 419                               struct extent_io_tree *dirty_pages, int mark)
 420 {
 421         int ret;
 422         int err = 0;
 423         int werr = 0;
 424         struct page *page;
 425         struct inode *btree_inode = root->fs_info->btree_inode;
 426         u64 start = 0;
 427         u64 end;
 428         unsigned long index;
 429
 430         while (1) {
 431                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 432                                             mark);
 433                 if (ret)
 434                         break;
 435
 436                 clear_extent_bits(dirty_pages, start, end, mark, GFP_NOFS);
 437                 while (start <= end) {
 438                         index = start >> PAGE_CACHE_SHIFT;
 439                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 440                         page = find_get_page(btree_inode->i_mapping, index);
 441                         if (!page)
 442                                 continue;
 443                         if (PageDirty(page)) {
 444                                 btree_lock_page_hook(page);
 445                                 wait_on_page_writeback(page);
 446                                 err = write_one_page(page, 0);
 447                                 if (err)
 448                                         werr = err;
 449                         }
 450                         wait_on_page_writeback(page);
 451                         page_cache_release(page);
 452                         cond_resched();
 453                 }
 454         }
 455         if (err)
 456                 werr = err;
 457         return werr;
 458 }
 459
 460 /*
 461  * when btree blocks are allocated, they have some corresponding bits set for
 462  * them in one of two extent_io trees.  This is used to make sure all of
 463  * those extents are on disk for transaction or log commit
 464  */
 465 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 466                                 struct extent_io_tree *dirty_pages, int mark)
 467 {
 468         int ret;
 469         int ret2;
 470
 471         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 472         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 473         return ret || ret2;
 474 }
 475
 476 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 477                                      struct btrfs_root *root)
 478 {
 479         if (!trans || !trans->transaction) {
 480                 struct inode *btree_inode;
 481                 btree_inode = root->fs_info->btree_inode;
 482                 return filemap_write_and_wait(btree_inode->i_mapping);
 483         }
 484         return btrfs_write_and_wait_marked_extents(root,
 485                                            &trans->transaction->dirty_pages,
 486                                            EXTENT_DIRTY);
 487 }
 488
 489 /*
 490  * this is used to update the root pointer in the tree of tree roots.
 491  *
 492  * But, in the case of the extent allocation tree, updating the root
 493  * pointer may allocate blocks which may change the root of the extent
 494  * allocation tree.
 495  *
 496  * So, this loops and repeats and makes sure the cowonly root didn't
 497  * change while the root pointer was being updated in the metadata.
 498  */
 499 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 500                                struct btrfs_root *root)
 501 {
 502         int ret;
 503         u64 old_root_bytenr;
 504         struct btrfs_root *tree_root = root->fs_info->tree_root;
 505
 506         btrfs_write_dirty_block_groups(trans, root);
 507
 508         while (1) {
 509                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 510                 if (old_root_bytenr == root->node->start)
 511                         break;
 512
 513                 btrfs_set_root_node(&root->root_item, root->node);
 514                 ret = btrfs_update_root(trans, tree_root,
 515                                         &root->root_key,
 516                                         &root->root_item);
 517                 BUG_ON(ret);
 518
 519                 ret = btrfs_write_dirty_block_groups(trans, root);
 520                 BUG_ON(ret);
 521         }
 522
 523         if (root != root->fs_info->extent_root)
 524                 switch_commit_root(root);
 525
 526         return 0;
 527 }
 528
 529 /*
 530  * update all the cowonly tree roots on disk
 531  */
 532 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 533                                          struct btrfs_root *root)
 534 {
 535         struct btrfs_fs_info *fs_info = root->fs_info;
 536         struct list_head *next;
 537         struct extent_buffer *eb;
 538         int ret;
 539
 540         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 541         BUG_ON(ret);
 542
 543         eb = btrfs_lock_root_node(fs_info->tree_root);
 544         btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
 545         btrfs_tree_unlock(eb);
 546         free_extent_buffer(eb);
 547
 548         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 549         BUG_ON(ret);
 550
 551         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 552                 next = fs_info->dirty_cowonly_roots.next;
 553                 list_del_init(next);
 554                 root = list_entry(next, struct btrfs_root, dirty_list);
 555
 556                 update_cowonly_root(trans, root);
 557         }
 558
 559         down_write(&fs_info->extent_commit_sem);
 560         switch_commit_root(fs_info->extent_root);
 561         up_write(&fs_info->extent_commit_sem);
 562
 563         return 0;
 564 }
 565
 566 /*
 567  * dead roots are old snapshots that need to be deleted.  This allocates
 568  * a dirty root struct and adds it into the list of dead roots that need to
 569  * be deleted
 570  */
 571 int btrfs_add_dead_root(struct btrfs_root *root)
 572 {
 573         mutex_lock(&root->fs_info->trans_mutex);
 574         list_add(&root->root_list, &root->fs_info->dead_roots);
 575         mutex_unlock(&root->fs_info->trans_mutex);
 576         return 0;
 577 }
 578
 579 /*
 580  * update all the cowonly tree roots on disk
 581  */
 582 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 583                                     struct btrfs_root *root)
 584 {
 585         struct btrfs_root *gang[8];
 586         struct btrfs_fs_info *fs_info = root->fs_info;
 587         int i;
 588         int ret;
 589         int err = 0;
 590
 591         while (1) {
 592                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 593                                                  (void **)gang, 0,
 594                                                  ARRAY_SIZE(gang),
 595                                                  BTRFS_ROOT_TRANS_TAG);
 596                 if (ret == 0)
 597                         break;
 598                 for (i = 0; i < ret; i++) {
 599                         root = gang[i];
 600                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 601                                         (unsigned long)root->root_key.objectid,
 602                                         BTRFS_ROOT_TRANS_TAG);
 603
 604                         btrfs_free_log(trans, root);
 605                         btrfs_update_reloc_root(trans, root);
 606
 607                         if (root->commit_root != root->node) {
 608                                 switch_commit_root(root);
 609                                 btrfs_set_root_node(&root->root_item,
 610                                                     root->node);
 611                         }
 612
 613                         err = btrfs_update_root(trans, fs_info->tree_root,
 614                                                 &root->root_key,
 615                                                 &root->root_item);
 616                         if (err)
 617                                 break;
 618                 }
 619         }
 620         return err;
 621 }
 622
 623 /*
 624  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 625  * otherwise every leaf in the btree is read and defragged.
 626  */
 627 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 628 {
 629         struct btrfs_fs_info *info = root->fs_info;
 630         int ret;
 631         struct btrfs_trans_handle *trans;
 632         unsigned long nr;
 633
 634         smp_mb();
 635         if (root->defrag_running)
 636                 return 0;
 637         trans = btrfs_start_transaction(root, 1);
 638         while (1) {
 639                 root->defrag_running = 1;
 640                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 641                 nr = trans->blocks_used;
 642                 btrfs_end_transaction(trans, root);
 643                 btrfs_btree_balance_dirty(info->tree_root, nr);
 644                 cond_resched();
 645
 646                 trans = btrfs_start_transaction(root, 1);
 647                 if (root->fs_info->closing || ret != -EAGAIN)
 648                         break;
 649         }
 650         root->defrag_running = 0;
 651         smp_mb();
 652         btrfs_end_transaction(trans, root);
 653         return 0;
 654 }
 655
 656 #if 0
 657 /*
 658  * when dropping snapshots, we generate a ton of delayed refs, and it makes
 659  * sense not to join the transaction while it is trying to flush the current
 660  * queue of delayed refs out.
 661  *
 662  * This is used by the drop snapshot code only
 663  */
 664 static noinline int wait_transaction_pre_flush(struct btrfs_fs_info *info)
 665 {
 666         DEFINE_WAIT(wait);
 667
 668         mutex_lock(&info->trans_mutex);
 669         while (info->running_transaction &&
 670                info->running_transaction->delayed_refs.flushing) {
 671                 prepare_to_wait(&info->transaction_wait, &wait,
 672                                 TASK_UNINTERRUPTIBLE);
 673                 mutex_unlock(&info->trans_mutex);
 674
 675                 schedule();
 676
 677                 mutex_lock(&info->trans_mutex);
 678                 finish_wait(&info->transaction_wait, &wait);
 679         }
 680         mutex_unlock(&info->trans_mutex);
 681         return 0;
 682 }
 683
 684 /*
 685  * Given a list of roots that need to be deleted, call btrfs_drop_snapshot on
 686  * all of them
 687  */
 688 int btrfs_drop_dead_root(struct btrfs_root *root)
 689 {
 690         struct btrfs_trans_handle *trans;
 691         struct btrfs_root *tree_root = root->fs_info->tree_root;
 692         unsigned long nr;
 693         int ret;
 694
 695         while (1) {
 696                 /*
 697                  * we don't want to jump in and create a bunch of
 698                  * delayed refs if the transaction is starting to close
 699                  */
 700                 wait_transaction_pre_flush(tree_root->fs_info);
 701                 trans = btrfs_start_transaction(tree_root, 1);
 702
 703                 /*
 704                  * we've joined a transaction, make sure it isn't
 705                  * closing right now
 706                  */
 707                 if (trans->transaction->delayed_refs.flushing) {
 708                         btrfs_end_transaction(trans, tree_root);
 709                         continue;
 710                 }
 711
 712                 ret = btrfs_drop_snapshot(trans, root);
 713                 if (ret != -EAGAIN)
 714                         break;
 715
 716                 ret = btrfs_update_root(trans, tree_root,
 717                                         &root->root_key,
 718                                         &root->root_item);
 719                 if (ret)
 720                         break;
 721
 722                 nr = trans->blocks_used;
 723                 ret = btrfs_end_transaction(trans, tree_root);
 724                 BUG_ON(ret);
 725
 726                 btrfs_btree_balance_dirty(tree_root, nr);
 727                 cond_resched();
 728         }
 729         BUG_ON(ret);
 730
 731         ret = btrfs_del_root(trans, tree_root, &root->root_key);
 732         BUG_ON(ret);
 733
 734         nr = trans->blocks_used;
 735         ret = btrfs_end_transaction(trans, tree_root);
 736         BUG_ON(ret);
 737
 738         free_extent_buffer(root->node);
 739         free_extent_buffer(root->commit_root);
 740         kfree(root);
 741
 742         btrfs_btree_balance_dirty(tree_root, nr);
 743         return ret;
 744 }
 745 #endif
 746
 747 /*
 748  * new snapshots need to be created at a very specific time in the
 749  * transaction commit.  This does the actual creation
 750  */
 751 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 752                                    struct btrfs_fs_info *fs_info,
 753                                    struct btrfs_pending_snapshot *pending)
 754 {
 755         struct btrfs_key key;
 756         struct btrfs_root_item *new_root_item;
 757         struct btrfs_root *tree_root = fs_info->tree_root;
 758         struct btrfs_root *root = pending->root;
 759         struct extent_buffer *tmp;
 760         struct extent_buffer *old;
 761         int ret;
 762         u64 objectid;
 763
 764         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 765         if (!new_root_item) {
 766                 ret = -ENOMEM;
 767                 goto fail;
 768         }
 769         ret = btrfs_find_free_objectid(trans, tree_root, 0, &objectid);
 770         if (ret)
 771                 goto fail;
 772
 773         record_root_in_trans(trans, root);
 774         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
 775         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
 776
 777         key.objectid = objectid;
 778         /* record when the snapshot was created in key.offset */
 779         key.offset = trans->transid;
 780         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
 781
 782         old = btrfs_lock_root_node(root);
 783         btrfs_cow_block(trans, root, old, NULL, 0, &old);
 784         btrfs_set_lock_blocking(old);
 785
 786         btrfs_copy_root(trans, root, old, &tmp, objectid);
 787         btrfs_tree_unlock(old);
 788         free_extent_buffer(old);
 789
 790         btrfs_set_root_node(new_root_item, tmp);
 791         ret = btrfs_insert_root(trans, root->fs_info->tree_root, &key,
 792                                 new_root_item);
 793         btrfs_tree_unlock(tmp);
 794         free_extent_buffer(tmp);
 795         if (ret)
 796                 goto fail;
 797
 798         key.offset = (u64)-1;
 799         memcpy(&pending->root_key, &key, sizeof(key));
 800 fail:
 801         kfree(new_root_item);
 802         return ret;
 803 }
 804
 805 static noinline int finish_pending_snapshot(struct btrfs_fs_info *fs_info,
 806                                    struct btrfs_pending_snapshot *pending)
 807 {
 808         int ret;
 809         int namelen;
 810         u64 index = 0;
 811         struct btrfs_trans_handle *trans;
 812         struct inode *parent_inode;
 813         struct btrfs_root *parent_root;
 814
 815         parent_inode = pending->dentry->d_parent->d_inode;
 816         parent_root = BTRFS_I(parent_inode)->root;
 817         trans = btrfs_join_transaction(parent_root, 1);
 818
 819         /*
 820          * insert the directory item
 821          */
 822         namelen = strlen(pending->name);
 823         ret = btrfs_set_inode_index(parent_inode, &index);
 824         ret = btrfs_insert_dir_item(trans, parent_root,
 825                             pending->name, namelen,
 826                             parent_inode->i_ino,
 827                             &pending->root_key, BTRFS_FT_DIR, index);
 828
 829         if (ret)
 830                 goto fail;
 831
 832         btrfs_i_size_write(parent_inode, parent_inode->i_size + namelen * 2);
 833         ret = btrfs_update_inode(trans, parent_root, parent_inode);
 834         BUG_ON(ret);
 835
 836         ret = btrfs_add_root_ref(trans, parent_root->fs_info->tree_root,
 837                                  pending->root_key.objectid,
 838                                  parent_root->root_key.objectid,
 839                                  parent_inode->i_ino, index, pending->name,
 840                                  namelen);
 841
 842         BUG_ON(ret);
 843
 844 fail:
 845         btrfs_end_transaction(trans, fs_info->fs_root);
 846         return ret;
 847 }
 848
 849 /*
 850  * create all the snapshots we've scheduled for creation
 851  */
 852 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
 853                                              struct btrfs_fs_info *fs_info)
 854 {
 855         struct btrfs_pending_snapshot *pending;
 856         struct list_head *head = &trans->transaction->pending_snapshots;
 857         int ret;
 858
 859         list_for_each_entry(pending, head, list) {
 860                 ret = create_pending_snapshot(trans, fs_info, pending);
 861                 BUG_ON(ret);
 862         }
 863         return 0;
 864 }
 865
 866 static noinline int finish_pending_snapshots(struct btrfs_trans_handle *trans,
 867                                              struct btrfs_fs_info *fs_info)
 868 {
 869         struct btrfs_pending_snapshot *pending;
 870         struct list_head *head = &trans->transaction->pending_snapshots;
 871         int ret;
 872
 873         while (!list_empty(head)) {
 874                 pending = list_entry(head->next,
 875                                      struct btrfs_pending_snapshot, list);
 876                 ret = finish_pending_snapshot(fs_info, pending);
 877                 BUG_ON(ret);
 878                 list_del(&pending->list);
 879                 kfree(pending->name);
 880                 kfree(pending);
 881         }
 882         return 0;
 883 }
 884
 885 static void update_super_roots(struct btrfs_root *root)
 886 {
 887         struct btrfs_root_item *root_item;
 888         struct btrfs_super_block *super;
 889
 890         super = &root->fs_info->super_copy;
 891
 892         root_item = &root->fs_info->chunk_root->root_item;
 893         super->chunk_root = root_item->bytenr;
 894         super->chunk_root_generation = root_item->generation;
 895         super->chunk_root_level = root_item->level;
 896
 897         root_item = &root->fs_info->tree_root->root_item;
 898         super->root = root_item->bytenr;
 899         super->generation = root_item->generation;
 900         super->root_level = root_item->level;
 901 }
 902
 903 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
 904 {
 905         int ret = 0;
 906         spin_lock(&info->new_trans_lock);
 907         if (info->running_transaction)
 908                 ret = info->running_transaction->in_commit;
 909         spin_unlock(&info->new_trans_lock);
 910         return ret;
 911 }
 912
 913 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
 914                              struct btrfs_root *root)
 915 {
 916         unsigned long joined = 0;
 917         unsigned long timeout = 1;
 918         struct btrfs_transaction *cur_trans;
 919         struct btrfs_transaction *prev_trans = NULL;
 920         DEFINE_WAIT(wait);
 921         int ret;
 922         int should_grow = 0;
 923         unsigned long now = get_seconds();
 924         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
 925
 926         btrfs_run_ordered_operations(root, 0);
 927
 928         /* make a pass through all the delayed refs we have so far
 929          * any runnings procs may add more while we are here
 930          */
 931         ret = btrfs_run_delayed_refs(trans, root, 0);
 932         BUG_ON(ret);
 933
 934         cur_trans = trans->transaction;
 935         /*
 936          * set the flushing flag so procs in this transaction have to
 937          * start sending their work down.
 938          */
 939         cur_trans->delayed_refs.flushing = 1;
 940
 941         ret = btrfs_run_delayed_refs(trans, root, 0);
 942         BUG_ON(ret);
 943
 944         mutex_lock(&root->fs_info->trans_mutex);
 945         if (cur_trans->in_commit) {
 946                 cur_trans->use_count++;
 947                 mutex_unlock(&root->fs_info->trans_mutex);
 948                 btrfs_end_transaction(trans, root);
 949
 950                 ret = wait_for_commit(root, cur_trans);
 951                 BUG_ON(ret);
 952
 953                 mutex_lock(&root->fs_info->trans_mutex);
 954                 put_transaction(cur_trans);
 955                 mutex_unlock(&root->fs_info->trans_mutex);
 956
 957                 return 0;
 958         }
 959
 960         trans->transaction->in_commit = 1;
 961         trans->transaction->blocked = 1;
 962         if (cur_trans->list.prev != &root->fs_info->trans_list) {
 963                 prev_trans = list_entry(cur_trans->list.prev,
 964                                         struct btrfs_transaction, list);
 965                 if (!prev_trans->commit_done) {
 966                         prev_trans->use_count++;
 967                         mutex_unlock(&root->fs_info->trans_mutex);
 968
 969                         wait_for_commit(root, prev_trans);
 970
 971                         mutex_lock(&root->fs_info->trans_mutex);
 972                         put_transaction(prev_trans);
 973                 }
 974         }
 975
 976         if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
 977                 should_grow = 1;
 978
 979         do {
 980                 int snap_pending = 0;
 981                 joined = cur_trans->num_joined;
 982                 if (!list_empty(&trans->transaction->pending_snapshots))
 983                         snap_pending = 1;
 984
 985                 WARN_ON(cur_trans != trans->transaction);
 986                 prepare_to_wait(&cur_trans->writer_wait, &wait,
 987                                 TASK_UNINTERRUPTIBLE);
 988
 989                 if (cur_trans->num_writers > 1)
 990                         timeout = MAX_SCHEDULE_TIMEOUT;
 991                 else if (should_grow)
 992                         timeout = 1;
 993
 994                 mutex_unlock(&root->fs_info->trans_mutex);
 995
 996                 if (flush_on_commit) {
 997                         btrfs_start_delalloc_inodes(root, 1);
 998                         ret = btrfs_wait_ordered_extents(root, 0, 1);
 999                         BUG_ON(ret);
1000                 } else if (snap_pending) {
1001                         ret = btrfs_wait_ordered_extents(root, 0, 1);
1002                         BUG_ON(ret);
1003                 }
1004
1005                 /*
1006                  * rename don't use btrfs_join_transaction, so, once we
1007                  * set the transaction to blocked above, we aren't going
1008                  * to get any new ordered operations.  We can safely run
1009                  * it here and no for sure that nothing new will be added
1010                  * to the list
1011                  */
1012                 btrfs_run_ordered_operations(root, 1);
1013
1014                 smp_mb();
1015                 if (cur_trans->num_writers > 1 || should_grow)
1016                         schedule_timeout(timeout);
1017
1018                 mutex_lock(&root->fs_info->trans_mutex);
1019                 finish_wait(&cur_trans->writer_wait, &wait);
1020         } while (cur_trans->num_writers > 1 ||
1021                  (should_grow && cur_trans->num_joined != joined));
1022
1023         ret = create_pending_snapshots(trans, root->fs_info);
1024         BUG_ON(ret);
1025
1026         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1027         BUG_ON(ret);
1028
1029         WARN_ON(cur_trans != trans->transaction);
1030
1031         /* btrfs_commit_tree_roots is responsible for getting the
1032          * various roots consistent with each other.  Every pointer
1033          * in the tree of tree roots has to point to the most up to date
1034          * root for every subvolume and other tree.  So, we have to keep
1035          * the tree logging code from jumping in and changing any
1036          * of the trees.
1037          *
1038          * At this point in the commit, there can't be any tree-log
1039          * writers, but a little lower down we drop the trans mutex
1040          * and let new people in.  By holding the tree_log_mutex
1041          * from now until after the super is written, we avoid races
1042          * with the tree-log code.
1043          */
1044         mutex_lock(&root->fs_info->tree_log_mutex);
1045
1046         ret = commit_fs_roots(trans, root);
1047         BUG_ON(ret);
1048
1049         /* commit_fs_roots gets rid of all the tree log roots, it is now
1050          * safe to free the root of tree log roots
1051          */
1052         btrfs_free_log_root_tree(trans, root->fs_info);
1053
1054         ret = commit_cowonly_roots(trans, root);
1055         BUG_ON(ret);
1056
1057         btrfs_prepare_extent_commit(trans, root);
1058
1059         cur_trans = root->fs_info->running_transaction;
1060         spin_lock(&root->fs_info->new_trans_lock);
1061         root->fs_info->running_transaction = NULL;
1062         spin_unlock(&root->fs_info->new_trans_lock);
1063
1064         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1065                             root->fs_info->tree_root->node);
1066         switch_commit_root(root->fs_info->tree_root);
1067
1068         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1069                             root->fs_info->chunk_root->node);
1070         switch_commit_root(root->fs_info->chunk_root);
1071
1072         update_super_roots(root);
1073
1074         if (!root->fs_info->log_root_recovering) {
1075                 btrfs_set_super_log_root(&root->fs_info->super_copy, 0);
1076                 btrfs_set_super_log_root_level(&root->fs_info->super_copy, 0);
1077         }
1078
1079         memcpy(&root->fs_info->super_for_commit, &root->fs_info->super_copy,
1080                sizeof(root->fs_info->super_copy));
1081
1082         trans->transaction->blocked = 0;
1083
1084         wake_up(&root->fs_info->transaction_wait);
1085
1086         mutex_unlock(&root->fs_info->trans_mutex);
1087         ret = btrfs_write_and_wait_transaction(trans, root);
1088         BUG_ON(ret);
1089         write_ctree_super(trans, root, 0);
1090
1091         /*
1092          * the super is written, we can safely allow the tree-loggers
1093          * to go about their business
1094          */
1095         mutex_unlock(&root->fs_info->tree_log_mutex);
1096
1097         btrfs_finish_extent_commit(trans, root);
1098
1099         /* do the directory inserts of any pending snapshot creations */
1100         finish_pending_snapshots(trans, root->fs_info);
1101
1102         mutex_lock(&root->fs_info->trans_mutex);
1103
1104         cur_trans->commit_done = 1;
1105
1106         root->fs_info->last_trans_committed = cur_trans->transid;
1107
1108         wake_up(&cur_trans->commit_wait);
1109
1110         put_transaction(cur_trans);
1111         put_transaction(cur_trans);
1112
1113         mutex_unlock(&root->fs_info->trans_mutex);
1114
1115         if (current->journal_info == trans)
1116                 current->journal_info = NULL;
1117
1118         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1119
1120         if (current != root->fs_info->transaction_kthread)
1121                 btrfs_run_delayed_iputs(root);
1122
1123         return ret;
1124 }
1125
1126 /*
1127  * interface function to delete all the snapshots we have scheduled for deletion
1128  */
1129 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1130 {
1131         LIST_HEAD(list);
1132         struct btrfs_fs_info *fs_info = root->fs_info;
1133
1134         mutex_lock(&fs_info->trans_mutex);
1135         list_splice_init(&fs_info->dead_roots, &list);
1136         mutex_unlock(&fs_info->trans_mutex);
1137
1138         while (!list_empty(&list)) {
1139                 root = list_entry(list.next, struct btrfs_root, root_list);
1140                 list_del(&root->root_list);
1141
1142                 if (btrfs_header_backref_rev(root->node) <
1143                     BTRFS_MIXED_BACKREF_REV)
1144                         btrfs_drop_snapshot(root, 0);
1145                 else
1146                         btrfs_drop_snapshot(root, 1);
1147         }
1148         return 0;
1149 }