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