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