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 struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
 169                                              int num_blocks, int type)
 170 {
 171         struct btrfs_trans_handle *h =
 172                 kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 173         int ret;
 174
 175         mutex_lock(&root->fs_info->trans_mutex);
 176         if (!root->fs_info->log_root_recovering &&
 177             ((type == TRANS_START && !root->fs_info->open_ioctl_trans) ||
 178              type == TRANS_USERSPACE))
 179                 wait_current_trans(root);
 180         ret = join_transaction(root);
 181         BUG_ON(ret);
 182
 183         h->transid = root->fs_info->running_transaction->transid;
 184         h->transaction = root->fs_info->running_transaction;
 185         h->blocks_reserved = num_blocks;
 186         h->blocks_used = 0;
 187         h->block_group = 0;
 188         h->alloc_exclude_nr = 0;
 189         h->alloc_exclude_start = 0;
 190         h->delayed_ref_updates = 0;
 191
 192         if (!current->journal_info && type != TRANS_USERSPACE)
 193                 current->journal_info = h;
 194
 195         root->fs_info->running_transaction->use_count++;
 196         record_root_in_trans(h, root);
 197         mutex_unlock(&root->fs_info->trans_mutex);
 198         return h;
 199 }
 200
 201 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 202                                                    int num_blocks)
 203 {
 204         return start_transaction(root, num_blocks, TRANS_START);
 205 }
 206 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root,
 207                                                    int num_blocks)
 208 {
 209         return start_transaction(root, num_blocks, TRANS_JOIN);
 210 }
 211
 212 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *r,
 213                                                          int num_blocks)
 214 {
 215         return start_transaction(r, num_blocks, TRANS_USERSPACE);
 216 }
 217
 218 /* wait for a transaction commit to be fully complete */
 219 static noinline int wait_for_commit(struct btrfs_root *root,
 220                                     struct btrfs_transaction *commit)
 221 {
 222         DEFINE_WAIT(wait);
 223         mutex_lock(&root->fs_info->trans_mutex);
 224         while (!commit->commit_done) {
 225                 prepare_to_wait(&commit->commit_wait, &wait,
 226                                 TASK_UNINTERRUPTIBLE);
 227                 if (commit->commit_done)
 228                         break;
 229                 mutex_unlock(&root->fs_info->trans_mutex);
 230                 schedule();
 231                 mutex_lock(&root->fs_info->trans_mutex);
 232         }
 233         mutex_unlock(&root->fs_info->trans_mutex);
 234         finish_wait(&commit->commit_wait, &wait);
 235         return 0;
 236 }
 237
 238 #if 0
 239 /*
 240  * rate limit against the drop_snapshot code.  This helps to slow down new
 241  * operations if the drop_snapshot code isn't able to keep up.
 242  */
 243 static void throttle_on_drops(struct btrfs_root *root)
 244 {
 245         struct btrfs_fs_info *info = root->fs_info;
 246         int harder_count = 0;
 247
 248 harder:
 249         if (atomic_read(&info->throttles)) {
 250                 DEFINE_WAIT(wait);
 251                 int thr;
 252                 thr = atomic_read(&info->throttle_gen);
 253
 254                 do {
 255                         prepare_to_wait(&info->transaction_throttle,
 256                                         &wait, TASK_UNINTERRUPTIBLE);
 257                         if (!atomic_read(&info->throttles)) {
 258                                 finish_wait(&info->transaction_throttle, &wait);
 259                                 break;
 260                         }
 261                         schedule();
 262                         finish_wait(&info->transaction_throttle, &wait);
 263                 } while (thr == atomic_read(&info->throttle_gen));
 264                 harder_count++;
 265
 266                 if (root->fs_info->total_ref_cache_size > 1 * 1024 * 1024 &&
 267                     harder_count < 2)
 268                         goto harder;
 269
 270                 if (root->fs_info->total_ref_cache_size > 5 * 1024 * 1024 &&
 271                     harder_count < 10)
 272                         goto harder;
 273
 274                 if (root->fs_info->total_ref_cache_size > 10 * 1024 * 1024 &&
 275                     harder_count < 20)
 276                         goto harder;
 277         }
 278 }
 279 #endif
 280
 281 void btrfs_throttle(struct btrfs_root *root)
 282 {
 283         mutex_lock(&root->fs_info->trans_mutex);
 284         if (!root->fs_info->open_ioctl_trans)
 285                 wait_current_trans(root);
 286         mutex_unlock(&root->fs_info->trans_mutex);
 287 }
 288
 289 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 290                           struct btrfs_root *root, int throttle)
 291 {
 292         struct btrfs_transaction *cur_trans;
 293         struct btrfs_fs_info *info = root->fs_info;
 294         int count = 0;
 295
 296         while (count < 4) {
 297                 unsigned long cur = trans->delayed_ref_updates;
 298                 trans->delayed_ref_updates = 0;
 299                 if (cur &&
 300                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 301                         trans->delayed_ref_updates = 0;
 302
 303                         /*
 304                          * do a full flush if the transaction is trying
 305                          * to close
 306                          */
 307                         if (trans->transaction->delayed_refs.flushing)
 308                                 cur = 0;
 309                         btrfs_run_delayed_refs(trans, root, cur);
 310                 } else {
 311                         break;
 312                 }
 313                 count++;
 314         }
 315
 316         mutex_lock(&info->trans_mutex);
 317         cur_trans = info->running_transaction;
 318         WARN_ON(cur_trans != trans->transaction);
 319         WARN_ON(cur_trans->num_writers < 1);
 320         cur_trans->num_writers--;
 321
 322         if (waitqueue_active(&cur_trans->writer_wait))
 323                 wake_up(&cur_trans->writer_wait);
 324         put_transaction(cur_trans);
 325         mutex_unlock(&info->trans_mutex);
 326
 327         if (current->journal_info == trans)
 328                 current->journal_info = NULL;
 329         memset(trans, 0, sizeof(*trans));
 330         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 331
 332         if (throttle)
 333                 btrfs_run_delayed_iputs(root);
 334
 335         return 0;
 336 }
 337
 338 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 339                           struct btrfs_root *root)
 340 {
 341         return __btrfs_end_transaction(trans, root, 0);
 342 }
 343
 344 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 345                                    struct btrfs_root *root)
 346 {
 347         return __btrfs_end_transaction(trans, root, 1);
 348 }
 349
 350 /*
 351  * when btree blocks are allocated, they have some corresponding bits set for
 352  * them in one of two extent_io trees.  This is used to make sure all of
 353  * those extents are sent to disk but does not wait on them
 354  */
 355 int btrfs_write_marked_extents(struct btrfs_root *root,
 356                                struct extent_io_tree *dirty_pages, int mark)
 357 {
 358         int ret;
 359         int err = 0;
 360         int werr = 0;
 361         struct page *page;
 362         struct inode *btree_inode = root->fs_info->btree_inode;
 363         u64 start = 0;
 364         u64 end;
 365         unsigned long index;
 366
 367         while (1) {
 368                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 369                                             mark);
 370                 if (ret)
 371                         break;
 372                 while (start <= end) {
 373                         cond_resched();
 374
 375                         index = start >> PAGE_CACHE_SHIFT;
 376                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 377                         page = find_get_page(btree_inode->i_mapping, index);
 378                         if (!page)
 379                                 continue;
 380
 381                         btree_lock_page_hook(page);
 382                         if (!page->mapping) {
 383                                 unlock_page(page);
 384                                 page_cache_release(page);
 385                                 continue;
 386                         }
 387
 388                         if (PageWriteback(page)) {
 389                                 if (PageDirty(page))
 390                                         wait_on_page_writeback(page);
 391                                 else {
 392                                         unlock_page(page);
 393                                         page_cache_release(page);
 394                                         continue;
 395                                 }
 396                         }
 397                         err = write_one_page(page, 0);
 398                         if (err)
 399                                 werr = err;
 400                         page_cache_release(page);
 401                 }
 402         }
 403         if (err)
 404                 werr = err;
 405         return werr;
 406 }
 407
 408 /*
 409  * when btree blocks are allocated, they have some corresponding bits set for
 410  * them in one of two extent_io trees.  This is used to make sure all of
 411  * those extents are on disk for transaction or log commit.  We wait
 412  * on all the pages and clear them from the dirty pages state tree
 413  */
 414 int btrfs_wait_marked_extents(struct btrfs_root *root,
 415                               struct extent_io_tree *dirty_pages, int mark)
 416 {
 417         int ret;
 418         int err = 0;
 419         int werr = 0;
 420         struct page *page;
 421         struct inode *btree_inode = root->fs_info->btree_inode;
 422         u64 start = 0;
 423         u64 end;
 424         unsigned long index;
 425
 426         while (1) {
 427                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 428                                             mark);
 429                 if (ret)
 430                         break;
 431
 432                 clear_extent_bits(dirty_pages, start, end, mark, GFP_NOFS);
 433                 while (start <= end) {
 434                         index = start >> PAGE_CACHE_SHIFT;
 435                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 436                         page = find_get_page(btree_inode->i_mapping, index);
 437                         if (!page)
 438                                 continue;
 439                         if (PageDirty(page)) {
 440                                 btree_lock_page_hook(page);
 441                                 wait_on_page_writeback(page);
 442                                 err = write_one_page(page, 0);
 443                                 if (err)
 444                                         werr = err;
 445                         }
 446                         wait_on_page_writeback(page);
 447                         page_cache_release(page);
 448                         cond_resched();
 449                 }
 450         }
 451         if (err)
 452                 werr = err;
 453         return werr;
 454 }
 455
 456 /*
 457  * when btree blocks are allocated, they have some corresponding bits set for
 458  * them in one of two extent_io trees.  This is used to make sure all of
 459  * those extents are on disk for transaction or log commit
 460  */
 461 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 462                                 struct extent_io_tree *dirty_pages, int mark)
 463 {
 464         int ret;
 465         int ret2;
 466
 467         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 468         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 469         return ret || ret2;
 470 }
 471
 472 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 473                                      struct btrfs_root *root)
 474 {
 475         if (!trans || !trans->transaction) {
 476                 struct inode *btree_inode;
 477                 btree_inode = root->fs_info->btree_inode;
 478                 return filemap_write_and_wait(btree_inode->i_mapping);
 479         }
 480         return btrfs_write_and_wait_marked_extents(root,
 481                                            &trans->transaction->dirty_pages,
 482                                            EXTENT_DIRTY);
 483 }
 484
 485 /*
 486  * this is used to update the root pointer in the tree of tree roots.
 487  *
 488  * But, in the case of the extent allocation tree, updating the root
 489  * pointer may allocate blocks which may change the root of the extent
 490  * allocation tree.
 491  *
 492  * So, this loops and repeats and makes sure the cowonly root didn't
 493  * change while the root pointer was being updated in the metadata.
 494  */
 495 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 496                                struct btrfs_root *root)
 497 {
 498         int ret;
 499         u64 old_root_bytenr;
 500         u64 old_root_used;
 501         struct btrfs_root *tree_root = root->fs_info->tree_root;
 502
 503         old_root_used = btrfs_root_used(&root->root_item);
 504         btrfs_write_dirty_block_groups(trans, root);
 505
 506         while (1) {
 507                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 508                 if (old_root_bytenr == root->node->start &&
 509                     old_root_used == btrfs_root_used(&root->root_item))
 510                         break;
 511
 512                 btrfs_set_root_node(&root->root_item, root->node);
 513                 ret = btrfs_update_root(trans, tree_root,
 514                                         &root->root_key,
 515                                         &root->root_item);
 516                 BUG_ON(ret);
 517
 518                 old_root_used = btrfs_root_used(&root->root_item);
 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 btrfs_root *parent_root;
 760         struct inode *parent_inode;
 761         struct extent_buffer *tmp;
 762         struct extent_buffer *old;
 763         int ret;
 764         u64 objectid;
 765         int namelen;
 766         u64 index = 0;
 767
 768         parent_inode = pending->dentry->d_parent->d_inode;
 769         parent_root = BTRFS_I(parent_inode)->root;
 770
 771         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 772         if (!new_root_item) {
 773                 ret = -ENOMEM;
 774                 goto fail;
 775         }
 776         ret = btrfs_find_free_objectid(trans, tree_root, 0, &objectid);
 777         if (ret)
 778                 goto fail;
 779
 780         key.objectid = objectid;
 781         /* record when the snapshot was created in key.offset */
 782         key.offset = trans->transid;
 783         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
 784
 785         memcpy(&pending->root_key, &key, sizeof(key));
 786         pending->root_key.offset = (u64)-1;
 787
 788         record_root_in_trans(trans, parent_root);
 789         /*
 790          * insert the directory item
 791          */
 792         namelen = strlen(pending->name);
 793         ret = btrfs_set_inode_index(parent_inode, &index);
 794         BUG_ON(ret);
 795         ret = btrfs_insert_dir_item(trans, parent_root,
 796                             pending->name, namelen,
 797                             parent_inode->i_ino,
 798                             &pending->root_key, BTRFS_FT_DIR, index);
 799         BUG_ON(ret);
 800
 801         btrfs_i_size_write(parent_inode, parent_inode->i_size + namelen * 2);
 802         ret = btrfs_update_inode(trans, parent_root, parent_inode);
 803         BUG_ON(ret);
 804
 805         record_root_in_trans(trans, root);
 806         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
 807         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
 808
 809         old = btrfs_lock_root_node(root);
 810         btrfs_cow_block(trans, root, old, NULL, 0, &old);
 811         btrfs_set_lock_blocking(old);
 812
 813         btrfs_copy_root(trans, root, old, &tmp, objectid);
 814         btrfs_tree_unlock(old);
 815         free_extent_buffer(old);
 816
 817         btrfs_set_root_node(new_root_item, tmp);
 818         ret = btrfs_insert_root(trans, root->fs_info->tree_root, &key,
 819                                 new_root_item);
 820         BUG_ON(ret);
 821         btrfs_tree_unlock(tmp);
 822         free_extent_buffer(tmp);
 823
 824         ret = btrfs_add_root_ref(trans, parent_root->fs_info->tree_root,
 825                                  pending->root_key.objectid,
 826                                  parent_root->root_key.objectid,
 827                                  parent_inode->i_ino, index, pending->name,
 828                                  namelen);
 829         BUG_ON(ret);
 830
 831 fail:
 832         kfree(new_root_item);
 833         return ret;
 834 }
 835
 836 /*
 837  * create all the snapshots we've scheduled for creation
 838  */
 839 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
 840                                              struct btrfs_fs_info *fs_info)
 841 {
 842         struct btrfs_pending_snapshot *pending;
 843         struct list_head *head = &trans->transaction->pending_snapshots;
 844         int ret;
 845
 846         list_for_each_entry(pending, head, list) {
 847                 ret = create_pending_snapshot(trans, fs_info, pending);
 848                 BUG_ON(ret);
 849         }
 850         return 0;
 851 }
 852
 853 static void update_super_roots(struct btrfs_root *root)
 854 {
 855         struct btrfs_root_item *root_item;
 856         struct btrfs_super_block *super;
 857
 858         super = &root->fs_info->super_copy;
 859
 860         root_item = &root->fs_info->chunk_root->root_item;
 861         super->chunk_root = root_item->bytenr;
 862         super->chunk_root_generation = root_item->generation;
 863         super->chunk_root_level = root_item->level;
 864
 865         root_item = &root->fs_info->tree_root->root_item;
 866         super->root = root_item->bytenr;
 867         super->generation = root_item->generation;
 868         super->root_level = root_item->level;
 869 }
 870
 871 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
 872 {
 873         int ret = 0;
 874         spin_lock(&info->new_trans_lock);
 875         if (info->running_transaction)
 876                 ret = info->running_transaction->in_commit;
 877         spin_unlock(&info->new_trans_lock);
 878         return ret;
 879 }
 880
 881 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
 882                              struct btrfs_root *root)
 883 {
 884         unsigned long joined = 0;
 885         unsigned long timeout = 1;
 886         struct btrfs_transaction *cur_trans;
 887         struct btrfs_transaction *prev_trans = NULL;
 888         DEFINE_WAIT(wait);
 889         int ret;
 890         int should_grow = 0;
 891         unsigned long now = get_seconds();
 892         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
 893
 894         btrfs_run_ordered_operations(root, 0);
 895
 896         /* make a pass through all the delayed refs we have so far
 897          * any runnings procs may add more while we are here
 898          */
 899         ret = btrfs_run_delayed_refs(trans, root, 0);
 900         BUG_ON(ret);
 901
 902         cur_trans = trans->transaction;
 903         /*
 904          * set the flushing flag so procs in this transaction have to
 905          * start sending their work down.
 906          */
 907         cur_trans->delayed_refs.flushing = 1;
 908
 909         ret = btrfs_run_delayed_refs(trans, root, 0);
 910         BUG_ON(ret);
 911
 912         mutex_lock(&root->fs_info->trans_mutex);
 913         if (cur_trans->in_commit) {
 914                 cur_trans->use_count++;
 915                 mutex_unlock(&root->fs_info->trans_mutex);
 916                 btrfs_end_transaction(trans, root);
 917
 918                 ret = wait_for_commit(root, cur_trans);
 919                 BUG_ON(ret);
 920
 921                 mutex_lock(&root->fs_info->trans_mutex);
 922                 put_transaction(cur_trans);
 923                 mutex_unlock(&root->fs_info->trans_mutex);
 924
 925                 return 0;
 926         }
 927
 928         trans->transaction->in_commit = 1;
 929         trans->transaction->blocked = 1;
 930         if (cur_trans->list.prev != &root->fs_info->trans_list) {
 931                 prev_trans = list_entry(cur_trans->list.prev,
 932                                         struct btrfs_transaction, list);
 933                 if (!prev_trans->commit_done) {
 934                         prev_trans->use_count++;
 935                         mutex_unlock(&root->fs_info->trans_mutex);
 936
 937                         wait_for_commit(root, prev_trans);
 938
 939                         mutex_lock(&root->fs_info->trans_mutex);
 940                         put_transaction(prev_trans);
 941                 }
 942         }
 943
 944         if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
 945                 should_grow = 1;
 946
 947         do {
 948                 int snap_pending = 0;
 949                 joined = cur_trans->num_joined;
 950                 if (!list_empty(&trans->transaction->pending_snapshots))
 951                         snap_pending = 1;
 952
 953                 WARN_ON(cur_trans != trans->transaction);
 954                 prepare_to_wait(&cur_trans->writer_wait, &wait,
 955                                 TASK_UNINTERRUPTIBLE);
 956
 957                 if (cur_trans->num_writers > 1)
 958                         timeout = MAX_SCHEDULE_TIMEOUT;
 959                 else if (should_grow)
 960                         timeout = 1;
 961
 962                 mutex_unlock(&root->fs_info->trans_mutex);
 963
 964                 if (flush_on_commit || snap_pending) {
 965                         btrfs_start_delalloc_inodes(root, 1);
 966                         ret = btrfs_wait_ordered_extents(root, 0, 1);
 967                         BUG_ON(ret);
 968                 }
 969
 970                 /*
 971                  * rename don't use btrfs_join_transaction, so, once we
 972                  * set the transaction to blocked above, we aren't going
 973                  * to get any new ordered operations.  We can safely run
 974                  * it here and no for sure that nothing new will be added
 975                  * to the list
 976                  */
 977                 btrfs_run_ordered_operations(root, 1);
 978
 979                 smp_mb();
 980                 if (cur_trans->num_writers > 1 || should_grow)
 981                         schedule_timeout(timeout);
 982
 983                 mutex_lock(&root->fs_info->trans_mutex);
 984                 finish_wait(&cur_trans->writer_wait, &wait);
 985         } while (cur_trans->num_writers > 1 ||
 986                  (should_grow && cur_trans->num_joined != joined));
 987
 988         ret = create_pending_snapshots(trans, root->fs_info);
 989         BUG_ON(ret);
 990
 991         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 992         BUG_ON(ret);
 993
 994         WARN_ON(cur_trans != trans->transaction);
 995
 996         /* btrfs_commit_tree_roots is responsible for getting the
 997          * various roots consistent with each other.  Every pointer
 998          * in the tree of tree roots has to point to the most up to date
 999          * root for every subvolume and other tree.  So, we have to keep
1000          * the tree logging code from jumping in and changing any
1001          * of the trees.
1002          *
1003          * At this point in the commit, there can't be any tree-log
1004          * writers, but a little lower down we drop the trans mutex
1005          * and let new people in.  By holding the tree_log_mutex
1006          * from now until after the super is written, we avoid races
1007          * with the tree-log code.
1008          */
1009         mutex_lock(&root->fs_info->tree_log_mutex);
1010
1011         ret = commit_fs_roots(trans, root);
1012         BUG_ON(ret);
1013
1014         /* commit_fs_roots gets rid of all the tree log roots, it is now
1015          * safe to free the root of tree log roots
1016          */
1017         btrfs_free_log_root_tree(trans, root->fs_info);
1018
1019         ret = commit_cowonly_roots(trans, root);
1020         BUG_ON(ret);
1021
1022         btrfs_prepare_extent_commit(trans, root);
1023
1024         cur_trans = root->fs_info->running_transaction;
1025         spin_lock(&root->fs_info->new_trans_lock);
1026         root->fs_info->running_transaction = NULL;
1027         spin_unlock(&root->fs_info->new_trans_lock);
1028
1029         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1030                             root->fs_info->tree_root->node);
1031         switch_commit_root(root->fs_info->tree_root);
1032
1033         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1034                             root->fs_info->chunk_root->node);
1035         switch_commit_root(root->fs_info->chunk_root);
1036
1037         update_super_roots(root);
1038
1039         if (!root->fs_info->log_root_recovering) {
1040                 btrfs_set_super_log_root(&root->fs_info->super_copy, 0);
1041                 btrfs_set_super_log_root_level(&root->fs_info->super_copy, 0);
1042         }
1043
1044         memcpy(&root->fs_info->super_for_commit, &root->fs_info->super_copy,
1045                sizeof(root->fs_info->super_copy));
1046
1047         trans->transaction->blocked = 0;
1048
1049         wake_up(&root->fs_info->transaction_wait);
1050
1051         mutex_unlock(&root->fs_info->trans_mutex);
1052         ret = btrfs_write_and_wait_transaction(trans, root);
1053         BUG_ON(ret);
1054         write_ctree_super(trans, root, 0);
1055
1056         /*
1057          * the super is written, we can safely allow the tree-loggers
1058          * to go about their business
1059          */
1060         mutex_unlock(&root->fs_info->tree_log_mutex);
1061
1062         btrfs_finish_extent_commit(trans, root);
1063
1064         mutex_lock(&root->fs_info->trans_mutex);
1065
1066         cur_trans->commit_done = 1;
1067
1068         root->fs_info->last_trans_committed = cur_trans->transid;
1069
1070         wake_up(&cur_trans->commit_wait);
1071
1072         put_transaction(cur_trans);
1073         put_transaction(cur_trans);
1074
1075         mutex_unlock(&root->fs_info->trans_mutex);
1076
1077         if (current->journal_info == trans)
1078                 current->journal_info = NULL;
1079
1080         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1081
1082         if (current != root->fs_info->transaction_kthread)
1083                 btrfs_run_delayed_iputs(root);
1084
1085         return ret;
1086 }
1087
1088 /*
1089  * interface function to delete all the snapshots we have scheduled for deletion
1090  */
1091 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1092 {
1093         LIST_HEAD(list);
1094         struct btrfs_fs_info *fs_info = root->fs_info;
1095
1096         mutex_lock(&fs_info->trans_mutex);
1097         list_splice_init(&fs_info->dead_roots, &list);
1098         mutex_unlock(&fs_info->trans_mutex);
1099
1100         while (!list_empty(&list)) {
1101                 root = list_entry(list.next, struct btrfs_root, root_list);
1102                 list_del(&root->root_list);
1103
1104                 if (btrfs_header_backref_rev(root->node) <
1105                     BTRFS_MIXED_BACKREF_REV)
1106                         btrfs_drop_snapshot(root, 0);
1107                 else
1108                         btrfs_drop_snapshot(root, 1);
1109         }
1110         return 0;
1111 }