fs/btrfs/transaction.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/fs.h>
  20 #include <linux/sched.h>
  21 #include <linux/writeback.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/blkdev.h>
  24 #include "ctree.h"
  25 #include "disk-io.h"
  26 #include "transaction.h"
  27 #include "locking.h"
  28 #include "tree-log.h"
  29
  30 #define BTRFS_ROOT_TRANS_TAG 0
  31
  32 static noinline void put_transaction(struct btrfs_transaction *transaction)
  33 {
  34         WARN_ON(transaction->use_count == 0);
  35         transaction->use_count--;
  36         if (transaction->use_count == 0) {
  37                 list_del_init(&transaction->list);
  38                 memset(transaction, 0, sizeof(*transaction));
  39                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  40         }
  41 }
  42
  43 static noinline void switch_commit_root(struct btrfs_root *root)
  44 {
  45         free_extent_buffer(root->commit_root);
  46         root->commit_root = btrfs_root_node(root);
  47 }
  48
  49 /*
  50  * either allocate a new transaction or hop into the existing one
  51  */
  52 static noinline int join_transaction(struct btrfs_root *root)
  53 {
  54         struct btrfs_transaction *cur_trans;
  55         cur_trans = root->fs_info->running_transaction;
  56         if (!cur_trans) {
  57                 cur_trans = kmem_cache_alloc(btrfs_transaction_cachep,
  58                                              GFP_NOFS);
  59                 BUG_ON(!cur_trans);
  60                 root->fs_info->generation++;
  61                 cur_trans->num_writers = 1;
  62                 cur_trans->num_joined = 0;
  63                 cur_trans->transid = root->fs_info->generation;
  64                 init_waitqueue_head(&cur_trans->writer_wait);
  65                 init_waitqueue_head(&cur_trans->commit_wait);
  66                 cur_trans->in_commit = 0;
  67                 cur_trans->blocked = 0;
  68                 cur_trans->use_count = 1;
  69                 cur_trans->commit_done = 0;
  70                 cur_trans->start_time = get_seconds();
  71
  72                 cur_trans->delayed_refs.root.rb_node = NULL;
  73                 cur_trans->delayed_refs.num_entries = 0;
  74                 cur_trans->delayed_refs.num_heads_ready = 0;
  75                 cur_trans->delayed_refs.num_heads = 0;
  76                 cur_trans->delayed_refs.flushing = 0;
  77                 cur_trans->delayed_refs.run_delayed_start = 0;
  78                 spin_lock_init(&cur_trans->delayed_refs.lock);
  79
  80                 INIT_LIST_HEAD(&cur_trans->pending_snapshots);
  81                 list_add_tail(&cur_trans->list, &root->fs_info->trans_list);
  82                 extent_io_tree_init(&cur_trans->dirty_pages,
  83                                      root->fs_info->btree_inode->i_mapping,
  84                                      GFP_NOFS);
  85                 spin_lock(&root->fs_info->new_trans_lock);
  86                 root->fs_info->running_transaction = cur_trans;
  87                 spin_unlock(&root->fs_info->new_trans_lock);
  88         } else {
  89                 cur_trans->num_writers++;
  90                 cur_trans->num_joined++;
  91         }
  92
  93         return 0;
  94 }
  95
  96 /*
  97  * this does all the record keeping required to make sure that a reference
  98  * counted root is properly recorded in a given transaction.  This is required
  99  * to make sure the old root from before we joined the transaction is deleted
 100  * when the transaction commits
 101  */
 102 static noinline int record_root_in_trans(struct btrfs_trans_handle *trans,
 103                                          struct btrfs_root *root)
 104 {
 105         if (root->ref_cows && root->last_trans < trans->transid) {
 106                 WARN_ON(root == root->fs_info->extent_root);
 107                 WARN_ON(root->commit_root != root->node);
 108
 109                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 110                            (unsigned long)root->root_key.objectid,
 111                            BTRFS_ROOT_TRANS_TAG);
 112                 root->last_trans = trans->transid;
 113                 btrfs_init_reloc_root(trans, root);
 114         }
 115         return 0;
 116 }
 117
 118 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 119                                struct btrfs_root *root)
 120 {
 121         if (!root->ref_cows)
 122                 return 0;
 123
 124         mutex_lock(&root->fs_info->trans_mutex);
 125         if (root->last_trans == trans->transid) {
 126                 mutex_unlock(&root->fs_info->trans_mutex);
 127                 return 0;
 128         }
 129
 130         record_root_in_trans(trans, root);
 131         mutex_unlock(&root->fs_info->trans_mutex);
 132         return 0;
 133 }
 134
 135 /* wait for commit against the current transaction to become unblocked
 136  * when this is done, it is safe to start a new transaction, but the current
 137  * transaction might not be fully on disk.
 138  */
 139 static void wait_current_trans(struct btrfs_root *root)
 140 {
 141         struct btrfs_transaction *cur_trans;
 142
 143         cur_trans = root->fs_info->running_transaction;
 144         if (cur_trans && cur_trans->blocked) {
 145                 DEFINE_WAIT(wait);
 146                 cur_trans->use_count++;
 147                 while (1) {
 148                         prepare_to_wait(&root->fs_info->transaction_wait, &wait,
 149                                         TASK_UNINTERRUPTIBLE);
 150                         if (cur_trans->blocked) {
 151                                 mutex_unlock(&root->fs_info->trans_mutex);
 152                                 schedule();
 153                                 mutex_lock(&root->fs_info->trans_mutex);
 154                                 finish_wait(&root->fs_info->transaction_wait,
 155                                             &wait);
 156                         } else {
 157                                 finish_wait(&root->fs_info->transaction_wait,
 158                                             &wait);
 159                                 break;
 160                         }
 161                 }
 162                 put_transaction(cur_trans);
 163         }
 164 }
 165
 166 static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
 167                                              int num_blocks, int wait)
 168 {
 169         struct btrfs_trans_handle *h =
 170                 kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 171         int ret;
 172
 173         mutex_lock(&root->fs_info->trans_mutex);
 174         if (!root->fs_info->log_root_recovering &&
 175             ((wait == 1 && !root->fs_info->open_ioctl_trans) || wait == 2))
 176                 wait_current_trans(root);
 177         ret = join_transaction(root);
 178         BUG_ON(ret);
 179
 180         h->transid = root->fs_info->running_transaction->transid;
 181         h->transaction = root->fs_info->running_transaction;
 182         h->blocks_reserved = num_blocks;
 183         h->blocks_used = 0;
 184         h->block_group = 0;
 185         h->alloc_exclude_nr = 0;
 186         h->alloc_exclude_start = 0;
 187         h->delayed_ref_updates = 0;
 188
 189         if (!current->journal_info)
 190                 current->journal_info = h;
 191
 192         root->fs_info->running_transaction->use_count++;
 193         record_root_in_trans(h, root);
 194         mutex_unlock(&root->fs_info->trans_mutex);
 195         return h;
 196 }
 197
 198 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 199                                                    int num_blocks)
 200 {
 201         return start_transaction(root, num_blocks, 1);
 202 }
 203 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root,
 204                                                    int num_blocks)
 205 {
 206         return start_transaction(root, num_blocks, 0);
 207 }
 208
 209 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *r,
 210                                                          int num_blocks)
 211 {
 212         return start_transaction(r, num_blocks, 2);
 213 }
 214
 215 /* wait for a transaction commit to be fully complete */
 216 static noinline int wait_for_commit(struct btrfs_root *root,
 217                                     struct btrfs_transaction *commit)
 218 {
 219         DEFINE_WAIT(wait);
 220         mutex_lock(&root->fs_info->trans_mutex);
 221         while (!commit->commit_done) {
 222                 prepare_to_wait(&commit->commit_wait, &wait,
 223                                 TASK_UNINTERRUPTIBLE);
 224                 if (commit->commit_done)
 225                         break;
 226                 mutex_unlock(&root->fs_info->trans_mutex);
 227                 schedule();
 228                 mutex_lock(&root->fs_info->trans_mutex);
 229         }
 230         mutex_unlock(&root->fs_info->trans_mutex);
 231         finish_wait(&commit->commit_wait, &wait);
 232         return 0;
 233 }
 234
 235 #if 0
 236 /*
 237  * rate limit against the drop_snapshot code.  This helps to slow down new
 238  * operations if the drop_snapshot code isn't able to keep up.
 239  */
 240 static void throttle_on_drops(struct btrfs_root *root)
 241 {
 242         struct btrfs_fs_info *info = root->fs_info;
 243         int harder_count = 0;
 244
 245 harder:
 246         if (atomic_read(&info->throttles)) {
 247                 DEFINE_WAIT(wait);
 248                 int thr;
 249                 thr = atomic_read(&info->throttle_gen);
 250
 251                 do {
 252                         prepare_to_wait(&info->transaction_throttle,
 253                                         &wait, TASK_UNINTERRUPTIBLE);
 254                         if (!atomic_read(&info->throttles)) {
 255                                 finish_wait(&info->transaction_throttle, &wait);
 256                                 break;
 257                         }
 258                         schedule();
 259                         finish_wait(&info->transaction_throttle, &wait);
 260                 } while (thr == atomic_read(&info->throttle_gen));
 261                 harder_count++;
 262
 263                 if (root->fs_info->total_ref_cache_size > 1 * 1024 * 1024 &&
 264                     harder_count < 2)
 265                         goto harder;
 266
 267                 if (root->fs_info->total_ref_cache_size > 5 * 1024 * 1024 &&
 268                     harder_count < 10)
 269                         goto harder;
 270
 271                 if (root->fs_info->total_ref_cache_size > 10 * 1024 * 1024 &&
 272                     harder_count < 20)
 273                         goto harder;
 274         }
 275 }
 276 #endif
 277
 278 void btrfs_throttle(struct btrfs_root *root)
 279 {
 280         mutex_lock(&root->fs_info->trans_mutex);
 281         if (!root->fs_info->open_ioctl_trans)
 282                 wait_current_trans(root);
 283         mutex_unlock(&root->fs_info->trans_mutex);
 284 }
 285
 286 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 287                           struct btrfs_root *root, int throttle)
 288 {
 289         struct btrfs_transaction *cur_trans;
 290         struct btrfs_fs_info *info = root->fs_info;
 291         int count = 0;
 292
 293         while (count < 4) {
 294                 unsigned long cur = trans->delayed_ref_updates;
 295                 trans->delayed_ref_updates = 0;
 296                 if (cur &&
 297                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 298                         trans->delayed_ref_updates = 0;
 299
 300                         /*
 301                          * do a full flush if the transaction is trying
 302                          * to close
 303                          */
 304                         if (trans->transaction->delayed_refs.flushing)
 305                                 cur = 0;
 306                         btrfs_run_delayed_refs(trans, root, cur);
 307                 } else {
 308                         break;
 309                 }
 310                 count++;
 311         }
 312
 313         mutex_lock(&info->trans_mutex);
 314         cur_trans = info->running_transaction;
 315         WARN_ON(cur_trans != trans->transaction);
 316         WARN_ON(cur_trans->num_writers < 1);
 317         cur_trans->num_writers--;
 318
 319         if (waitqueue_active(&cur_trans->writer_wait))
 320                 wake_up(&cur_trans->writer_wait);
 321         put_transaction(cur_trans);
 322         mutex_unlock(&info->trans_mutex);
 323
 324         if (current->journal_info == trans)
 325                 current->journal_info = NULL;
 326         memset(trans, 0, sizeof(*trans));
 327         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 328
 329         return 0;
 330 }
 331
 332 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 333                           struct btrfs_root *root)
 334 {
 335         return __btrfs_end_transaction(trans, root, 0);
 336 }
 337
 338 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 339                                    struct btrfs_root *root)
 340 {
 341         return __btrfs_end_transaction(trans, root, 1);
 342 }
 343
 344 /*
 345  * when btree blocks are allocated, they have some corresponding bits set for
 346  * them in one of two extent_io trees.  This is used to make sure all of
 347  * those extents are on disk for transaction or log commit
 348  */
 349 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 350                                         struct extent_io_tree *dirty_pages)
 351 {
 352         int ret;
 353         int err = 0;
 354         int werr = 0;
 355         struct page *page;
 356         struct inode *btree_inode = root->fs_info->btree_inode;
 357         u64 start = 0;
 358         u64 end;
 359         unsigned long index;
 360
 361         while (1) {
 362                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 363                                             EXTENT_DIRTY);
 364                 if (ret)
 365                         break;
 366                 while (start <= end) {
 367                         cond_resched();
 368
 369                         index = start >> PAGE_CACHE_SHIFT;
 370                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 371                         page = find_get_page(btree_inode->i_mapping, index);
 372                         if (!page)
 373                                 continue;
 374
 375                         btree_lock_page_hook(page);
 376                         if (!page->mapping) {
 377                                 unlock_page(page);
 378                                 page_cache_release(page);
 379                                 continue;
 380                         }
 381
 382                         if (PageWriteback(page)) {
 383                                 if (PageDirty(page))
 384                                         wait_on_page_writeback(page);
 385                                 else {
 386                                         unlock_page(page);
 387                                         page_cache_release(page);
 388                                         continue;
 389                                 }
 390                         }
 391                         err = write_one_page(page, 0);
 392                         if (err)
 393                                 werr = err;
 394                         page_cache_release(page);
 395                 }
 396         }
 397         while (1) {
 398                 ret = find_first_extent_bit(dirty_pages, 0, &start, &end,
 399                                             EXTENT_DIRTY);
 400                 if (ret)
 401                         break;
 402
 403                 clear_extent_dirty(dirty_pages, start, end, GFP_NOFS);
 404                 while (start <= end) {
 405                         index = start >> PAGE_CACHE_SHIFT;
 406                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 407                         page = find_get_page(btree_inode->i_mapping, index);
 408                         if (!page)
 409                                 continue;
 410                         if (PageDirty(page)) {
 411                                 btree_lock_page_hook(page);
 412                                 wait_on_page_writeback(page);
 413                                 err = write_one_page(page, 0);
 414                                 if (err)
 415                                         werr = err;
 416                         }
 417                         wait_on_page_writeback(page);
 418                         page_cache_release(page);
 419                         cond_resched();
 420                 }
 421         }
 422         if (err)
 423                 werr = err;
 424         return werr;
 425 }
 426
 427 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 428                                      struct btrfs_root *root)
 429 {
 430         if (!trans || !trans->transaction) {
 431                 struct inode *btree_inode;
 432                 btree_inode = root->fs_info->btree_inode;
 433                 return filemap_write_and_wait(btree_inode->i_mapping);
 434         }
 435         return btrfs_write_and_wait_marked_extents(root,
 436                                            &trans->transaction->dirty_pages);
 437 }
 438
 439 /*
 440  * this is used to update the root pointer in the tree of tree roots.
 441  *
 442  * But, in the case of the extent allocation tree, updating the root
 443  * pointer may allocate blocks which may change the root of the extent
 444  * allocation tree.
 445  *
 446  * So, this loops and repeats and makes sure the cowonly root didn't
 447  * change while the root pointer was being updated in the metadata.
 448  */
 449 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 450                                struct btrfs_root *root)
 451 {
 452         int ret;
 453         u64 old_root_bytenr;
 454         struct btrfs_root *tree_root = root->fs_info->tree_root;
 455
 456         btrfs_write_dirty_block_groups(trans, root);
 457
 458         while (1) {
 459                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 460                 if (old_root_bytenr == root->node->start)
 461                         break;
 462
 463                 btrfs_set_root_node(&root->root_item, root->node);
 464                 ret = btrfs_update_root(trans, tree_root,
 465                                         &root->root_key,
 466                                         &root->root_item);
 467                 BUG_ON(ret);
 468
 469                 ret = btrfs_write_dirty_block_groups(trans, root);
 470                 BUG_ON(ret);
 471         }
 472
 473         if (root != root->fs_info->extent_root)
 474                 switch_commit_root(root);
 475
 476         return 0;
 477 }
 478
 479 /*
 480  * update all the cowonly tree roots on disk
 481  */
 482 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 483                                          struct btrfs_root *root)
 484 {
 485         struct btrfs_fs_info *fs_info = root->fs_info;
 486         struct list_head *next;
 487         struct extent_buffer *eb;
 488         int ret;
 489
 490         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 491         BUG_ON(ret);
 492
 493         eb = btrfs_lock_root_node(fs_info->tree_root);
 494         btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
 495         btrfs_tree_unlock(eb);
 496         free_extent_buffer(eb);
 497
 498         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 499         BUG_ON(ret);
 500
 501         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 502                 next = fs_info->dirty_cowonly_roots.next;
 503                 list_del_init(next);
 504                 root = list_entry(next, struct btrfs_root, dirty_list);
 505
 506                 update_cowonly_root(trans, root);
 507         }
 508
 509         down_write(&fs_info->extent_commit_sem);
 510         switch_commit_root(fs_info->extent_root);
 511         up_write(&fs_info->extent_commit_sem);
 512
 513         return 0;
 514 }
 515
 516 /*
 517  * dead roots are old snapshots that need to be deleted.  This allocates
 518  * a dirty root struct and adds it into the list of dead roots that need to
 519  * be deleted
 520  */
 521 int btrfs_add_dead_root(struct btrfs_root *root)
 522 {
 523         mutex_lock(&root->fs_info->trans_mutex);
 524         list_add(&root->root_list, &root->fs_info->dead_roots);
 525         mutex_unlock(&root->fs_info->trans_mutex);
 526         return 0;
 527 }
 528
 529 /*
 530  * update all the cowonly tree roots on disk
 531  */
 532 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 533                                     struct btrfs_root *root)
 534 {
 535         struct btrfs_root *gang[8];
 536         struct btrfs_fs_info *fs_info = root->fs_info;
 537         int i;
 538         int ret;
 539         int err = 0;
 540
 541         while (1) {
 542                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 543                                                  (void **)gang, 0,
 544                                                  ARRAY_SIZE(gang),
 545                                                  BTRFS_ROOT_TRANS_TAG);
 546                 if (ret == 0)
 547                         break;
 548                 for (i = 0; i < ret; i++) {
 549                         root = gang[i];
 550                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 551                                         (unsigned long)root->root_key.objectid,
 552                                         BTRFS_ROOT_TRANS_TAG);
 553
 554                         btrfs_free_log(trans, root);
 555                         btrfs_update_reloc_root(trans, root);
 556
 557                         if (root->commit_root != root->node) {
 558                                 switch_commit_root(root);
 559                                 btrfs_set_root_node(&root->root_item,
 560                                                     root->node);
 561                         }
 562
 563                         err = btrfs_update_root(trans, fs_info->tree_root,
 564                                                 &root->root_key,
 565                                                 &root->root_item);
 566                         if (err)
 567                                 break;
 568                 }
 569         }
 570         return err;
 571 }
 572
 573 /*
 574  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 575  * otherwise every leaf in the btree is read and defragged.
 576  */
 577 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 578 {
 579         struct btrfs_fs_info *info = root->fs_info;
 580         int ret;
 581         struct btrfs_trans_handle *trans;
 582         unsigned long nr;
 583
 584         smp_mb();
 585         if (root->defrag_running)
 586                 return 0;
 587         trans = btrfs_start_transaction(root, 1);
 588         while (1) {
 589                 root->defrag_running = 1;
 590                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 591                 nr = trans->blocks_used;
 592                 btrfs_end_transaction(trans, root);
 593                 btrfs_btree_balance_dirty(info->tree_root, nr);
 594                 cond_resched();
 595
 596                 trans = btrfs_start_transaction(root, 1);
 597                 if (root->fs_info->closing || ret != -EAGAIN)
 598                         break;
 599         }
 600         root->defrag_running = 0;
 601         smp_mb();
 602         btrfs_end_transaction(trans, root);
 603         return 0;
 604 }
 605
 606 #if 0
 607 /*
 608  * when dropping snapshots, we generate a ton of delayed refs, and it makes
 609  * sense not to join the transaction while it is trying to flush the current
 610  * queue of delayed refs out.
 611  *
 612  * This is used by the drop snapshot code only
 613  */
 614 static noinline int wait_transaction_pre_flush(struct btrfs_fs_info *info)
 615 {
 616         DEFINE_WAIT(wait);
 617
 618         mutex_lock(&info->trans_mutex);
 619         while (info->running_transaction &&
 620                info->running_transaction->delayed_refs.flushing) {
 621                 prepare_to_wait(&info->transaction_wait, &wait,
 622                                 TASK_UNINTERRUPTIBLE);
 623                 mutex_unlock(&info->trans_mutex);
 624
 625                 schedule();
 626
 627                 mutex_lock(&info->trans_mutex);
 628                 finish_wait(&info->transaction_wait, &wait);
 629         }
 630         mutex_unlock(&info->trans_mutex);
 631         return 0;
 632 }
 633
 634 /*
 635  * Given a list of roots that need to be deleted, call btrfs_drop_snapshot on
 636  * all of them
 637  */
 638 int btrfs_drop_dead_root(struct btrfs_root *root)
 639 {
 640         struct btrfs_trans_handle *trans;
 641         struct btrfs_root *tree_root = root->fs_info->tree_root;
 642         unsigned long nr;
 643         int ret;
 644
 645         while (1) {
 646                 /*
 647                  * we don't want to jump in and create a bunch of
 648                  * delayed refs if the transaction is starting to close
 649                  */
 650                 wait_transaction_pre_flush(tree_root->fs_info);
 651                 trans = btrfs_start_transaction(tree_root, 1);
 652
 653                 /*
 654                  * we've joined a transaction, make sure it isn't
 655                  * closing right now
 656                  */
 657                 if (trans->transaction->delayed_refs.flushing) {
 658                         btrfs_end_transaction(trans, tree_root);
 659                         continue;
 660                 }
 661
 662                 ret = btrfs_drop_snapshot(trans, root);
 663                 if (ret != -EAGAIN)
 664                         break;
 665
 666                 ret = btrfs_update_root(trans, tree_root,
 667                                         &root->root_key,
 668                                         &root->root_item);
 669                 if (ret)
 670                         break;
 671
 672                 nr = trans->blocks_used;
 673                 ret = btrfs_end_transaction(trans, tree_root);
 674                 BUG_ON(ret);
 675
 676                 btrfs_btree_balance_dirty(tree_root, nr);
 677                 cond_resched();
 678         }
 679         BUG_ON(ret);
 680
 681         ret = btrfs_del_root(trans, tree_root, &root->root_key);
 682         BUG_ON(ret);
 683
 684         nr = trans->blocks_used;
 685         ret = btrfs_end_transaction(trans, tree_root);
 686         BUG_ON(ret);
 687
 688         free_extent_buffer(root->node);
 689         free_extent_buffer(root->commit_root);
 690         kfree(root);
 691
 692         btrfs_btree_balance_dirty(tree_root, nr);
 693         return ret;
 694 }
 695 #endif
 696
 697 /*
 698  * new snapshots need to be created at a very specific time in the
 699  * transaction commit.  This does the actual creation
 700  */
 701 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 702                                    struct btrfs_fs_info *fs_info,
 703                                    struct btrfs_pending_snapshot *pending)
 704 {
 705         struct btrfs_key key;
 706         struct btrfs_root_item *new_root_item;
 707         struct btrfs_root *tree_root = fs_info->tree_root;
 708         struct btrfs_root *root = pending->root;
 709         struct extent_buffer *tmp;
 710         struct extent_buffer *old;
 711         int ret;
 712         u64 objectid;
 713
 714         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 715         if (!new_root_item) {
 716                 ret = -ENOMEM;
 717                 goto fail;
 718         }
 719         ret = btrfs_find_free_objectid(trans, tree_root, 0, &objectid);
 720         if (ret)
 721                 goto fail;
 722
 723         record_root_in_trans(trans, root);
 724         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
 725         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
 726
 727         key.objectid = objectid;
 728         /* record when the snapshot was created in key.offset */
 729         key.offset = trans->transid;
 730         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
 731
 732         old = btrfs_lock_root_node(root);
 733         btrfs_cow_block(trans, root, old, NULL, 0, &old);
 734         btrfs_set_lock_blocking(old);
 735
 736         btrfs_copy_root(trans, root, old, &tmp, objectid);
 737         btrfs_tree_unlock(old);
 738         free_extent_buffer(old);
 739
 740         btrfs_set_root_node(new_root_item, tmp);
 741         ret = btrfs_insert_root(trans, root->fs_info->tree_root, &key,
 742                                 new_root_item);
 743         btrfs_tree_unlock(tmp);
 744         free_extent_buffer(tmp);
 745         if (ret)
 746                 goto fail;
 747
 748         key.offset = (u64)-1;
 749         memcpy(&pending->root_key, &key, sizeof(key));
 750 fail:
 751         kfree(new_root_item);
 752         btrfs_unreserve_metadata_space(root, 6);
 753         return ret;
 754 }
 755
 756 static noinline int finish_pending_snapshot(struct btrfs_fs_info *fs_info,
 757                                    struct btrfs_pending_snapshot *pending)
 758 {
 759         int ret;
 760         int namelen;
 761         u64 index = 0;
 762         struct btrfs_trans_handle *trans;
 763         struct inode *parent_inode;
 764         struct inode *inode;
 765         struct btrfs_root *parent_root;
 766
 767         parent_inode = pending->dentry->d_parent->d_inode;
 768         parent_root = BTRFS_I(parent_inode)->root;
 769         trans = btrfs_join_transaction(parent_root, 1);
 770
 771         /*
 772          * insert the directory item
 773          */
 774         namelen = strlen(pending->name);
 775         ret = btrfs_set_inode_index(parent_inode, &index);
 776         ret = btrfs_insert_dir_item(trans, parent_root,
 777                             pending->name, namelen,
 778                             parent_inode->i_ino,
 779                             &pending->root_key, BTRFS_FT_DIR, index);
 780
 781         if (ret)
 782                 goto fail;
 783
 784         btrfs_i_size_write(parent_inode, parent_inode->i_size + namelen * 2);
 785         ret = btrfs_update_inode(trans, parent_root, parent_inode);
 786         BUG_ON(ret);
 787
 788         ret = btrfs_add_root_ref(trans, parent_root->fs_info->tree_root,
 789                                  pending->root_key.objectid,
 790                                  parent_root->root_key.objectid,
 791                                  parent_inode->i_ino, index, pending->name,
 792                                  namelen);
 793
 794         BUG_ON(ret);
 795
 796         inode = btrfs_lookup_dentry(parent_inode, pending->dentry);
 797         d_instantiate(pending->dentry, inode);
 798 fail:
 799         btrfs_end_transaction(trans, fs_info->fs_root);
 800         return ret;
 801 }
 802
 803 /*
 804  * create all the snapshots we've scheduled for creation
 805  */
 806 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
 807                                              struct btrfs_fs_info *fs_info)
 808 {
 809         struct btrfs_pending_snapshot *pending;
 810         struct list_head *head = &trans->transaction->pending_snapshots;
 811         int ret;
 812
 813         list_for_each_entry(pending, head, list) {
 814                 ret = create_pending_snapshot(trans, fs_info, pending);
 815                 BUG_ON(ret);
 816         }
 817         return 0;
 818 }
 819
 820 static noinline int finish_pending_snapshots(struct btrfs_trans_handle *trans,
 821                                              struct btrfs_fs_info *fs_info)
 822 {
 823         struct btrfs_pending_snapshot *pending;
 824         struct list_head *head = &trans->transaction->pending_snapshots;
 825         int ret;
 826
 827         while (!list_empty(head)) {
 828                 pending = list_entry(head->next,
 829                                      struct btrfs_pending_snapshot, list);
 830                 ret = finish_pending_snapshot(fs_info, pending);
 831                 BUG_ON(ret);
 832                 list_del(&pending->list);
 833                 kfree(pending->name);
 834                 kfree(pending);
 835         }
 836         return 0;
 837 }
 838
 839 static void update_super_roots(struct btrfs_root *root)
 840 {
 841         struct btrfs_root_item *root_item;
 842         struct btrfs_super_block *super;
 843
 844         super = &root->fs_info->super_copy;
 845
 846         root_item = &root->fs_info->chunk_root->root_item;
 847         super->chunk_root = root_item->bytenr;
 848         super->chunk_root_generation = root_item->generation;
 849         super->chunk_root_level = root_item->level;
 850
 851         root_item = &root->fs_info->tree_root->root_item;
 852         super->root = root_item->bytenr;
 853         super->generation = root_item->generation;
 854         super->root_level = root_item->level;
 855 }
 856
 857 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
 858 {
 859         int ret = 0;
 860         spin_lock(&info->new_trans_lock);
 861         if (info->running_transaction)
 862                 ret = info->running_transaction->in_commit;
 863         spin_unlock(&info->new_trans_lock);
 864         return ret;
 865 }
 866
 867 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
 868                              struct btrfs_root *root)
 869 {
 870         unsigned long joined = 0;
 871         unsigned long timeout = 1;
 872         struct btrfs_transaction *cur_trans;
 873         struct btrfs_transaction *prev_trans = NULL;
 874         DEFINE_WAIT(wait);
 875         int ret;
 876         int should_grow = 0;
 877         unsigned long now = get_seconds();
 878         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
 879
 880         btrfs_run_ordered_operations(root, 0);
 881
 882         /* make a pass through all the delayed refs we have so far
 883          * any runnings procs may add more while we are here
 884          */
 885         ret = btrfs_run_delayed_refs(trans, root, 0);
 886         BUG_ON(ret);
 887
 888         cur_trans = trans->transaction;
 889         /*
 890          * set the flushing flag so procs in this transaction have to
 891          * start sending their work down.
 892          */
 893         cur_trans->delayed_refs.flushing = 1;
 894
 895         ret = btrfs_run_delayed_refs(trans, root, 0);
 896         BUG_ON(ret);
 897
 898         mutex_lock(&root->fs_info->trans_mutex);
 899         if (cur_trans->in_commit) {
 900                 cur_trans->use_count++;
 901                 mutex_unlock(&root->fs_info->trans_mutex);
 902                 btrfs_end_transaction(trans, root);
 903
 904                 ret = wait_for_commit(root, cur_trans);
 905                 BUG_ON(ret);
 906
 907                 mutex_lock(&root->fs_info->trans_mutex);
 908                 put_transaction(cur_trans);
 909                 mutex_unlock(&root->fs_info->trans_mutex);
 910
 911                 return 0;
 912         }
 913
 914         trans->transaction->in_commit = 1;
 915         trans->transaction->blocked = 1;
 916         if (cur_trans->list.prev != &root->fs_info->trans_list) {
 917                 prev_trans = list_entry(cur_trans->list.prev,
 918                                         struct btrfs_transaction, list);
 919                 if (!prev_trans->commit_done) {
 920                         prev_trans->use_count++;
 921                         mutex_unlock(&root->fs_info->trans_mutex);
 922
 923                         wait_for_commit(root, prev_trans);
 924
 925                         mutex_lock(&root->fs_info->trans_mutex);
 926                         put_transaction(prev_trans);
 927                 }
 928         }
 929
 930         if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
 931                 should_grow = 1;
 932
 933         do {
 934                 int snap_pending = 0;
 935                 joined = cur_trans->num_joined;
 936                 if (!list_empty(&trans->transaction->pending_snapshots))
 937                         snap_pending = 1;
 938
 939                 WARN_ON(cur_trans != trans->transaction);
 940                 prepare_to_wait(&cur_trans->writer_wait, &wait,
 941                                 TASK_UNINTERRUPTIBLE);
 942
 943                 if (cur_trans->num_writers > 1)
 944                         timeout = MAX_SCHEDULE_TIMEOUT;
 945                 else if (should_grow)
 946                         timeout = 1;
 947
 948                 mutex_unlock(&root->fs_info->trans_mutex);
 949
 950                 if (flush_on_commit) {
 951                         btrfs_start_delalloc_inodes(root);
 952                         ret = btrfs_wait_ordered_extents(root, 0);
 953                         BUG_ON(ret);
 954                 } else if (snap_pending) {
 955                         ret = btrfs_wait_ordered_extents(root, 1);
 956                         BUG_ON(ret);
 957                 }
 958
 959                 /*
 960                  * rename don't use btrfs_join_transaction, so, once we
 961                  * set the transaction to blocked above, we aren't going
 962                  * to get any new ordered operations.  We can safely run
 963                  * it here and no for sure that nothing new will be added
 964                  * to the list
 965                  */
 966                 btrfs_run_ordered_operations(root, 1);
 967
 968                 smp_mb();
 969                 if (cur_trans->num_writers > 1 || should_grow)
 970                         schedule_timeout(timeout);
 971
 972                 mutex_lock(&root->fs_info->trans_mutex);
 973                 finish_wait(&cur_trans->writer_wait, &wait);
 974         } while (cur_trans->num_writers > 1 ||
 975                  (should_grow && cur_trans->num_joined != joined));
 976
 977         ret = create_pending_snapshots(trans, root->fs_info);
 978         BUG_ON(ret);
 979
 980         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 981         BUG_ON(ret);
 982
 983         WARN_ON(cur_trans != trans->transaction);
 984
 985         /* btrfs_commit_tree_roots is responsible for getting the
 986          * various roots consistent with each other.  Every pointer
 987          * in the tree of tree roots has to point to the most up to date
 988          * root for every subvolume and other tree.  So, we have to keep
 989          * the tree logging code from jumping in and changing any
 990          * of the trees.
 991          *
 992          * At this point in the commit, there can't be any tree-log
 993          * writers, but a little lower down we drop the trans mutex
 994          * and let new people in.  By holding the tree_log_mutex
 995          * from now until after the super is written, we avoid races
 996          * with the tree-log code.
 997          */
 998         mutex_lock(&root->fs_info->tree_log_mutex);
 999
1000         ret = commit_fs_roots(trans, root);
1001         BUG_ON(ret);
1002
1003         /* commit_fs_roots gets rid of all the tree log roots, it is now
1004          * safe to free the root of tree log roots
1005          */
1006         btrfs_free_log_root_tree(trans, root->fs_info);
1007
1008         ret = commit_cowonly_roots(trans, root);
1009         BUG_ON(ret);
1010
1011         btrfs_prepare_extent_commit(trans, root);
1012
1013         cur_trans = root->fs_info->running_transaction;
1014         spin_lock(&root->fs_info->new_trans_lock);
1015         root->fs_info->running_transaction = NULL;
1016         spin_unlock(&root->fs_info->new_trans_lock);
1017
1018         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1019                             root->fs_info->tree_root->node);
1020         switch_commit_root(root->fs_info->tree_root);
1021
1022         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1023                             root->fs_info->chunk_root->node);
1024         switch_commit_root(root->fs_info->chunk_root);
1025
1026         update_super_roots(root);
1027
1028         if (!root->fs_info->log_root_recovering) {
1029                 btrfs_set_super_log_root(&root->fs_info->super_copy, 0);
1030                 btrfs_set_super_log_root_level(&root->fs_info->super_copy, 0);
1031         }
1032
1033         memcpy(&root->fs_info->super_for_commit, &root->fs_info->super_copy,
1034                sizeof(root->fs_info->super_copy));
1035
1036         trans->transaction->blocked = 0;
1037
1038         wake_up(&root->fs_info->transaction_wait);
1039
1040         mutex_unlock(&root->fs_info->trans_mutex);
1041         ret = btrfs_write_and_wait_transaction(trans, root);
1042         BUG_ON(ret);
1043         write_ctree_super(trans, root, 0);
1044
1045         /*
1046          * the super is written, we can safely allow the tree-loggers
1047          * to go about their business
1048          */
1049         mutex_unlock(&root->fs_info->tree_log_mutex);
1050
1051         btrfs_finish_extent_commit(trans, root);
1052
1053         /* do the directory inserts of any pending snapshot creations */
1054         finish_pending_snapshots(trans, root->fs_info);
1055
1056         mutex_lock(&root->fs_info->trans_mutex);
1057
1058         cur_trans->commit_done = 1;
1059
1060         root->fs_info->last_trans_committed = cur_trans->transid;
1061
1062         wake_up(&cur_trans->commit_wait);
1063
1064         put_transaction(cur_trans);
1065         put_transaction(cur_trans);
1066
1067         mutex_unlock(&root->fs_info->trans_mutex);
1068
1069         if (current->journal_info == trans)
1070                 current->journal_info = NULL;
1071
1072         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1073         return ret;
1074 }
1075
1076 /*
1077  * interface function to delete all the snapshots we have scheduled for deletion
1078  */
1079 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1080 {
1081         LIST_HEAD(list);
1082         struct btrfs_fs_info *fs_info = root->fs_info;
1083
1084         mutex_lock(&fs_info->trans_mutex);
1085         list_splice_init(&fs_info->dead_roots, &list);
1086         mutex_unlock(&fs_info->trans_mutex);
1087
1088         while (!list_empty(&list)) {
1089                 root = list_entry(list.next, struct btrfs_root, root_list);
1090                 list_del(&root->root_list);
1091
1092                 if (btrfs_header_backref_rev(root->node) <
1093                     BTRFS_MIXED_BACKREF_REV)
1094                         btrfs_drop_snapshot(root, 0);
1095                 else
1096                         btrfs_drop_snapshot(root, 1);
1097         }
1098         return 0;
1099 }