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         TRANS_JOIN_NOLOCK,
 167 };
 168
 169 static int may_wait_transaction(struct btrfs_root *root, int type)
 170 {
 171         if (!root->fs_info->log_root_recovering &&
 172             ((type == TRANS_START && !root->fs_info->open_ioctl_trans) ||
 173              type == TRANS_USERSPACE))
 174                 return 1;
 175         return 0;
 176 }
 177
 178 static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
 179                                                     u64 num_items, int type)
 180 {
 181         struct btrfs_trans_handle *h;
 182         struct btrfs_transaction *cur_trans;
 183         int ret;
 184 again:
 185         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 186         if (!h)
 187                 return ERR_PTR(-ENOMEM);
 188
 189         if (type != TRANS_JOIN_NOLOCK)
 190                 mutex_lock(&root->fs_info->trans_mutex);
 191         if (may_wait_transaction(root, type))
 192                 wait_current_trans(root);
 193
 194         ret = join_transaction(root);
 195         BUG_ON(ret);
 196
 197         cur_trans = root->fs_info->running_transaction;
 198         cur_trans->use_count++;
 199         if (type != TRANS_JOIN_NOLOCK)
 200                 mutex_unlock(&root->fs_info->trans_mutex);
 201
 202         h->transid = cur_trans->transid;
 203         h->transaction = cur_trans;
 204         h->blocks_used = 0;
 205         h->block_group = 0;
 206         h->bytes_reserved = 0;
 207         h->delayed_ref_updates = 0;
 208         h->block_rsv = NULL;
 209
 210         smp_mb();
 211         if (cur_trans->blocked && may_wait_transaction(root, type)) {
 212                 btrfs_commit_transaction(h, root);
 213                 goto again;
 214         }
 215
 216         if (num_items > 0) {
 217                 ret = btrfs_trans_reserve_metadata(h, root, num_items);
 218                 if (ret == -EAGAIN) {
 219                         btrfs_commit_transaction(h, root);
 220                         goto again;
 221                 }
 222                 if (ret < 0) {
 223                         btrfs_end_transaction(h, root);
 224                         return ERR_PTR(ret);
 225                 }
 226         }
 227
 228         if (type != TRANS_JOIN_NOLOCK)
 229                 mutex_lock(&root->fs_info->trans_mutex);
 230         record_root_in_trans(h, root);
 231         if (type != TRANS_JOIN_NOLOCK)
 232                 mutex_unlock(&root->fs_info->trans_mutex);
 233
 234         if (!current->journal_info && type != TRANS_USERSPACE)
 235                 current->journal_info = h;
 236         return h;
 237 }
 238
 239 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 240                                                    int num_items)
 241 {
 242         return start_transaction(root, num_items, TRANS_START);
 243 }
 244 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root,
 245                                                    int num_blocks)
 246 {
 247         return start_transaction(root, 0, TRANS_JOIN);
 248 }
 249
 250 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root,
 251                                                           int num_blocks)
 252 {
 253         return start_transaction(root, 0, TRANS_JOIN_NOLOCK);
 254 }
 255
 256 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *r,
 257                                                          int num_blocks)
 258 {
 259         return start_transaction(r, 0, TRANS_USERSPACE);
 260 }
 261
 262 /* wait for a transaction commit to be fully complete */
 263 static noinline int wait_for_commit(struct btrfs_root *root,
 264                                     struct btrfs_transaction *commit)
 265 {
 266         DEFINE_WAIT(wait);
 267         mutex_lock(&root->fs_info->trans_mutex);
 268         while (!commit->commit_done) {
 269                 prepare_to_wait(&commit->commit_wait, &wait,
 270                                 TASK_UNINTERRUPTIBLE);
 271                 if (commit->commit_done)
 272                         break;
 273                 mutex_unlock(&root->fs_info->trans_mutex);
 274                 schedule();
 275                 mutex_lock(&root->fs_info->trans_mutex);
 276         }
 277         mutex_unlock(&root->fs_info->trans_mutex);
 278         finish_wait(&commit->commit_wait, &wait);
 279         return 0;
 280 }
 281
 282 #if 0
 283 /*
 284  * rate limit against the drop_snapshot code.  This helps to slow down new
 285  * operations if the drop_snapshot code isn't able to keep up.
 286  */
 287 static void throttle_on_drops(struct btrfs_root *root)
 288 {
 289         struct btrfs_fs_info *info = root->fs_info;
 290         int harder_count = 0;
 291
 292 harder:
 293         if (atomic_read(&info->throttles)) {
 294                 DEFINE_WAIT(wait);
 295                 int thr;
 296                 thr = atomic_read(&info->throttle_gen);
 297
 298                 do {
 299                         prepare_to_wait(&info->transaction_throttle,
 300                                         &wait, TASK_UNINTERRUPTIBLE);
 301                         if (!atomic_read(&info->throttles)) {
 302                                 finish_wait(&info->transaction_throttle, &wait);
 303                                 break;
 304                         }
 305                         schedule();
 306                         finish_wait(&info->transaction_throttle, &wait);
 307                 } while (thr == atomic_read(&info->throttle_gen));
 308                 harder_count++;
 309
 310                 if (root->fs_info->total_ref_cache_size > 1 * 1024 * 1024 &&
 311                     harder_count < 2)
 312                         goto harder;
 313
 314                 if (root->fs_info->total_ref_cache_size > 5 * 1024 * 1024 &&
 315                     harder_count < 10)
 316                         goto harder;
 317
 318                 if (root->fs_info->total_ref_cache_size > 10 * 1024 * 1024 &&
 319                     harder_count < 20)
 320                         goto harder;
 321         }
 322 }
 323 #endif
 324
 325 void btrfs_throttle(struct btrfs_root *root)
 326 {
 327         mutex_lock(&root->fs_info->trans_mutex);
 328         if (!root->fs_info->open_ioctl_trans)
 329                 wait_current_trans(root);
 330         mutex_unlock(&root->fs_info->trans_mutex);
 331 }
 332
 333 static int should_end_transaction(struct btrfs_trans_handle *trans,
 334                                   struct btrfs_root *root)
 335 {
 336         int ret;
 337         ret = btrfs_block_rsv_check(trans, root,
 338                                     &root->fs_info->global_block_rsv, 0, 5);
 339         return ret ? 1 : 0;
 340 }
 341
 342 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 343                                  struct btrfs_root *root)
 344 {
 345         struct btrfs_transaction *cur_trans = trans->transaction;
 346         int updates;
 347
 348         if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
 349                 return 1;
 350
 351         updates = trans->delayed_ref_updates;
 352         trans->delayed_ref_updates = 0;
 353         if (updates)
 354                 btrfs_run_delayed_refs(trans, root, updates);
 355
 356         return should_end_transaction(trans, root);
 357 }
 358
 359 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 360                           struct btrfs_root *root, int throttle, int lock)
 361 {
 362         struct btrfs_transaction *cur_trans = trans->transaction;
 363         struct btrfs_fs_info *info = root->fs_info;
 364         int count = 0;
 365
 366         while (count < 4) {
 367                 unsigned long cur = trans->delayed_ref_updates;
 368                 trans->delayed_ref_updates = 0;
 369                 if (cur &&
 370                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 371                         trans->delayed_ref_updates = 0;
 372
 373                         /*
 374                          * do a full flush if the transaction is trying
 375                          * to close
 376                          */
 377                         if (trans->transaction->delayed_refs.flushing)
 378                                 cur = 0;
 379                         btrfs_run_delayed_refs(trans, root, cur);
 380                 } else {
 381                         break;
 382                 }
 383                 count++;
 384         }
 385
 386         btrfs_trans_release_metadata(trans, root);
 387
 388         if (lock && !root->fs_info->open_ioctl_trans &&
 389             should_end_transaction(trans, root))
 390                 trans->transaction->blocked = 1;
 391
 392         if (lock && cur_trans->blocked && !cur_trans->in_commit) {
 393                 if (throttle)
 394                         return btrfs_commit_transaction(trans, root);
 395                 else
 396                         wake_up_process(info->transaction_kthread);
 397         }
 398
 399         if (lock)
 400                 mutex_lock(&info->trans_mutex);
 401         WARN_ON(cur_trans != info->running_transaction);
 402         WARN_ON(cur_trans->num_writers < 1);
 403         cur_trans->num_writers--;
 404
 405         if (waitqueue_active(&cur_trans->writer_wait))
 406                 wake_up(&cur_trans->writer_wait);
 407         put_transaction(cur_trans);
 408         if (lock)
 409                 mutex_unlock(&info->trans_mutex);
 410
 411         if (current->journal_info == trans)
 412                 current->journal_info = NULL;
 413         memset(trans, 0, sizeof(*trans));
 414         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 415
 416         if (throttle)
 417                 btrfs_run_delayed_iputs(root);
 418
 419         return 0;
 420 }
 421
 422 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 423                           struct btrfs_root *root)
 424 {
 425         return __btrfs_end_transaction(trans, root, 0, 1);
 426 }
 427
 428 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 429                                    struct btrfs_root *root)
 430 {
 431         return __btrfs_end_transaction(trans, root, 1, 1);
 432 }
 433
 434 int btrfs_end_transaction_nolock(struct btrfs_trans_handle *trans,
 435                                  struct btrfs_root *root)
 436 {
 437         return __btrfs_end_transaction(trans, root, 0, 0);
 438 }
 439
 440 /*
 441  * when btree blocks are allocated, they have some corresponding bits set for
 442  * them in one of two extent_io trees.  This is used to make sure all of
 443  * those extents are sent to disk but does not wait on them
 444  */
 445 int btrfs_write_marked_extents(struct btrfs_root *root,
 446                                struct extent_io_tree *dirty_pages, int mark)
 447 {
 448         int ret;
 449         int err = 0;
 450         int werr = 0;
 451         struct page *page;
 452         struct inode *btree_inode = root->fs_info->btree_inode;
 453         u64 start = 0;
 454         u64 end;
 455         unsigned long index;
 456
 457         while (1) {
 458                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 459                                             mark);
 460                 if (ret)
 461                         break;
 462                 while (start <= end) {
 463                         cond_resched();
 464
 465                         index = start >> PAGE_CACHE_SHIFT;
 466                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 467                         page = find_get_page(btree_inode->i_mapping, index);
 468                         if (!page)
 469                                 continue;
 470
 471                         btree_lock_page_hook(page);
 472                         if (!page->mapping) {
 473                                 unlock_page(page);
 474                                 page_cache_release(page);
 475                                 continue;
 476                         }
 477
 478                         if (PageWriteback(page)) {
 479                                 if (PageDirty(page))
 480                                         wait_on_page_writeback(page);
 481                                 else {
 482                                         unlock_page(page);
 483                                         page_cache_release(page);
 484                                         continue;
 485                                 }
 486                         }
 487                         err = write_one_page(page, 0);
 488                         if (err)
 489                                 werr = err;
 490                         page_cache_release(page);
 491                 }
 492         }
 493         if (err)
 494                 werr = err;
 495         return werr;
 496 }
 497
 498 /*
 499  * when btree blocks are allocated, they have some corresponding bits set for
 500  * them in one of two extent_io trees.  This is used to make sure all of
 501  * those extents are on disk for transaction or log commit.  We wait
 502  * on all the pages and clear them from the dirty pages state tree
 503  */
 504 int btrfs_wait_marked_extents(struct btrfs_root *root,
 505                               struct extent_io_tree *dirty_pages, int mark)
 506 {
 507         int ret;
 508         int err = 0;
 509         int werr = 0;
 510         struct page *page;
 511         struct inode *btree_inode = root->fs_info->btree_inode;
 512         u64 start = 0;
 513         u64 end;
 514         unsigned long index;
 515
 516         while (1) {
 517                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 518                                             mark);
 519                 if (ret)
 520                         break;
 521
 522                 clear_extent_bits(dirty_pages, start, end, mark, GFP_NOFS);
 523                 while (start <= end) {
 524                         index = start >> PAGE_CACHE_SHIFT;
 525                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 526                         page = find_get_page(btree_inode->i_mapping, index);
 527                         if (!page)
 528                                 continue;
 529                         if (PageDirty(page)) {
 530                                 btree_lock_page_hook(page);
 531                                 wait_on_page_writeback(page);
 532                                 err = write_one_page(page, 0);
 533                                 if (err)
 534                                         werr = err;
 535                         }
 536                         wait_on_page_writeback(page);
 537                         page_cache_release(page);
 538                         cond_resched();
 539                 }
 540         }
 541         if (err)
 542                 werr = err;
 543         return werr;
 544 }
 545
 546 /*
 547  * when btree blocks are allocated, they have some corresponding bits set for
 548  * them in one of two extent_io trees.  This is used to make sure all of
 549  * those extents are on disk for transaction or log commit
 550  */
 551 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 552                                 struct extent_io_tree *dirty_pages, int mark)
 553 {
 554         int ret;
 555         int ret2;
 556
 557         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 558         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 559         return ret || ret2;
 560 }
 561
 562 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 563                                      struct btrfs_root *root)
 564 {
 565         if (!trans || !trans->transaction) {
 566                 struct inode *btree_inode;
 567                 btree_inode = root->fs_info->btree_inode;
 568                 return filemap_write_and_wait(btree_inode->i_mapping);
 569         }
 570         return btrfs_write_and_wait_marked_extents(root,
 571                                            &trans->transaction->dirty_pages,
 572                                            EXTENT_DIRTY);
 573 }
 574
 575 /*
 576  * this is used to update the root pointer in the tree of tree roots.
 577  *
 578  * But, in the case of the extent allocation tree, updating the root
 579  * pointer may allocate blocks which may change the root of the extent
 580  * allocation tree.
 581  *
 582  * So, this loops and repeats and makes sure the cowonly root didn't
 583  * change while the root pointer was being updated in the metadata.
 584  */
 585 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 586                                struct btrfs_root *root)
 587 {
 588         int ret;
 589         u64 old_root_bytenr;
 590         u64 old_root_used;
 591         struct btrfs_root *tree_root = root->fs_info->tree_root;
 592
 593         old_root_used = btrfs_root_used(&root->root_item);
 594         btrfs_write_dirty_block_groups(trans, root);
 595
 596         while (1) {
 597                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 598                 if (old_root_bytenr == root->node->start &&
 599                     old_root_used == btrfs_root_used(&root->root_item))
 600                         break;
 601
 602                 btrfs_set_root_node(&root->root_item, root->node);
 603                 ret = btrfs_update_root(trans, tree_root,
 604                                         &root->root_key,
 605                                         &root->root_item);
 606                 BUG_ON(ret);
 607
 608                 old_root_used = btrfs_root_used(&root->root_item);
 609                 ret = btrfs_write_dirty_block_groups(trans, root);
 610                 BUG_ON(ret);
 611         }
 612
 613         if (root != root->fs_info->extent_root)
 614                 switch_commit_root(root);
 615
 616         return 0;
 617 }
 618
 619 /*
 620  * update all the cowonly tree roots on disk
 621  */
 622 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 623                                          struct btrfs_root *root)
 624 {
 625         struct btrfs_fs_info *fs_info = root->fs_info;
 626         struct list_head *next;
 627         struct extent_buffer *eb;
 628         int ret;
 629
 630         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 631         BUG_ON(ret);
 632
 633         eb = btrfs_lock_root_node(fs_info->tree_root);
 634         btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
 635         btrfs_tree_unlock(eb);
 636         free_extent_buffer(eb);
 637
 638         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 639         BUG_ON(ret);
 640
 641         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 642                 next = fs_info->dirty_cowonly_roots.next;
 643                 list_del_init(next);
 644                 root = list_entry(next, struct btrfs_root, dirty_list);
 645
 646                 update_cowonly_root(trans, root);
 647         }
 648
 649         down_write(&fs_info->extent_commit_sem);
 650         switch_commit_root(fs_info->extent_root);
 651         up_write(&fs_info->extent_commit_sem);
 652
 653         return 0;
 654 }
 655
 656 /*
 657  * dead roots are old snapshots that need to be deleted.  This allocates
 658  * a dirty root struct and adds it into the list of dead roots that need to
 659  * be deleted
 660  */
 661 int btrfs_add_dead_root(struct btrfs_root *root)
 662 {
 663         mutex_lock(&root->fs_info->trans_mutex);
 664         list_add(&root->root_list, &root->fs_info->dead_roots);
 665         mutex_unlock(&root->fs_info->trans_mutex);
 666         return 0;
 667 }
 668
 669 /*
 670  * update all the cowonly tree roots on disk
 671  */
 672 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 673                                     struct btrfs_root *root)
 674 {
 675         struct btrfs_root *gang[8];
 676         struct btrfs_fs_info *fs_info = root->fs_info;
 677         int i;
 678         int ret;
 679         int err = 0;
 680
 681         while (1) {
 682                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 683                                                  (void **)gang, 0,
 684                                                  ARRAY_SIZE(gang),
 685                                                  BTRFS_ROOT_TRANS_TAG);
 686                 if (ret == 0)
 687                         break;
 688                 for (i = 0; i < ret; i++) {
 689                         root = gang[i];
 690                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 691                                         (unsigned long)root->root_key.objectid,
 692                                         BTRFS_ROOT_TRANS_TAG);
 693
 694                         btrfs_free_log(trans, root);
 695                         btrfs_update_reloc_root(trans, root);
 696                         btrfs_orphan_commit_root(trans, root);
 697
 698                         if (root->commit_root != root->node) {
 699                                 switch_commit_root(root);
 700                                 btrfs_set_root_node(&root->root_item,
 701                                                     root->node);
 702                         }
 703
 704                         err = btrfs_update_root(trans, fs_info->tree_root,
 705                                                 &root->root_key,
 706                                                 &root->root_item);
 707                         if (err)
 708                                 break;
 709                 }
 710         }
 711         return err;
 712 }
 713
 714 /*
 715  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 716  * otherwise every leaf in the btree is read and defragged.
 717  */
 718 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 719 {
 720         struct btrfs_fs_info *info = root->fs_info;
 721         struct btrfs_trans_handle *trans;
 722         int ret;
 723         unsigned long nr;
 724
 725         if (xchg(&root->defrag_running, 1))
 726                 return 0;
 727
 728         while (1) {
 729                 trans = btrfs_start_transaction(root, 0);
 730                 if (IS_ERR(trans))
 731                         return PTR_ERR(trans);
 732
 733                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 734
 735                 nr = trans->blocks_used;
 736                 btrfs_end_transaction(trans, root);
 737                 btrfs_btree_balance_dirty(info->tree_root, nr);
 738                 cond_resched();
 739
 740                 if (root->fs_info->closing || ret != -EAGAIN)
 741                         break;
 742         }
 743         root->defrag_running = 0;
 744         return ret;
 745 }
 746
 747 #if 0
 748 /*
 749  * when dropping snapshots, we generate a ton of delayed refs, and it makes
 750  * sense not to join the transaction while it is trying to flush the current
 751  * queue of delayed refs out.
 752  *
 753  * This is used by the drop snapshot code only
 754  */
 755 static noinline int wait_transaction_pre_flush(struct btrfs_fs_info *info)
 756 {
 757         DEFINE_WAIT(wait);
 758
 759         mutex_lock(&info->trans_mutex);
 760         while (info->running_transaction &&
 761                info->running_transaction->delayed_refs.flushing) {
 762                 prepare_to_wait(&info->transaction_wait, &wait,
 763                                 TASK_UNINTERRUPTIBLE);
 764                 mutex_unlock(&info->trans_mutex);
 765
 766                 schedule();
 767
 768                 mutex_lock(&info->trans_mutex);
 769                 finish_wait(&info->transaction_wait, &wait);
 770         }
 771         mutex_unlock(&info->trans_mutex);
 772         return 0;
 773 }
 774
 775 /*
 776  * Given a list of roots that need to be deleted, call btrfs_drop_snapshot on
 777  * all of them
 778  */
 779 int btrfs_drop_dead_root(struct btrfs_root *root)
 780 {
 781         struct btrfs_trans_handle *trans;
 782         struct btrfs_root *tree_root = root->fs_info->tree_root;
 783         unsigned long nr;
 784         int ret;
 785
 786         while (1) {
 787                 /*
 788                  * we don't want to jump in and create a bunch of
 789                  * delayed refs if the transaction is starting to close
 790                  */
 791                 wait_transaction_pre_flush(tree_root->fs_info);
 792                 trans = btrfs_start_transaction(tree_root, 1);
 793
 794                 /*
 795                  * we've joined a transaction, make sure it isn't
 796                  * closing right now
 797                  */
 798                 if (trans->transaction->delayed_refs.flushing) {
 799                         btrfs_end_transaction(trans, tree_root);
 800                         continue;
 801                 }
 802
 803                 ret = btrfs_drop_snapshot(trans, root);
 804                 if (ret != -EAGAIN)
 805                         break;
 806
 807                 ret = btrfs_update_root(trans, tree_root,
 808                                         &root->root_key,
 809                                         &root->root_item);
 810                 if (ret)
 811                         break;
 812
 813                 nr = trans->blocks_used;
 814                 ret = btrfs_end_transaction(trans, tree_root);
 815                 BUG_ON(ret);
 816
 817                 btrfs_btree_balance_dirty(tree_root, nr);
 818                 cond_resched();
 819         }
 820         BUG_ON(ret);
 821
 822         ret = btrfs_del_root(trans, tree_root, &root->root_key);
 823         BUG_ON(ret);
 824
 825         nr = trans->blocks_used;
 826         ret = btrfs_end_transaction(trans, tree_root);
 827         BUG_ON(ret);
 828
 829         free_extent_buffer(root->node);
 830         free_extent_buffer(root->commit_root);
 831         kfree(root);
 832
 833         btrfs_btree_balance_dirty(tree_root, nr);
 834         return ret;
 835 }
 836 #endif
 837
 838 /*
 839  * new snapshots need to be created at a very specific time in the
 840  * transaction commit.  This does the actual creation
 841  */
 842 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 843                                    struct btrfs_fs_info *fs_info,
 844                                    struct btrfs_pending_snapshot *pending)
 845 {
 846         struct btrfs_key key;
 847         struct btrfs_root_item *new_root_item;
 848         struct btrfs_root *tree_root = fs_info->tree_root;
 849         struct btrfs_root *root = pending->root;
 850         struct btrfs_root *parent_root;
 851         struct inode *parent_inode;
 852         struct dentry *dentry;
 853         struct extent_buffer *tmp;
 854         struct extent_buffer *old;
 855         int ret;
 856         u64 to_reserve = 0;
 857         u64 index = 0;
 858         u64 objectid;
 859
 860         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 861         if (!new_root_item) {
 862                 pending->error = -ENOMEM;
 863                 goto fail;
 864         }
 865
 866         ret = btrfs_find_free_objectid(trans, tree_root, 0, &objectid);
 867         if (ret) {
 868                 pending->error = ret;
 869                 goto fail;
 870         }
 871
 872         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
 873         btrfs_orphan_pre_snapshot(trans, pending, &to_reserve);
 874
 875         if (to_reserve > 0) {
 876                 ret = btrfs_block_rsv_add(trans, root, &pending->block_rsv,
 877                                           to_reserve);
 878                 if (ret) {
 879                         pending->error = ret;
 880                         goto fail;
 881                 }
 882         }
 883
 884         key.objectid = objectid;
 885         key.offset = (u64)-1;
 886         key.type = BTRFS_ROOT_ITEM_KEY;
 887
 888         trans->block_rsv = &pending->block_rsv;
 889
 890         dentry = pending->dentry;
 891         parent_inode = dentry->d_parent->d_inode;
 892         parent_root = BTRFS_I(parent_inode)->root;
 893         record_root_in_trans(trans, parent_root);
 894
 895         /*
 896          * insert the directory item
 897          */
 898         ret = btrfs_set_inode_index(parent_inode, &index);
 899         BUG_ON(ret);
 900         ret = btrfs_insert_dir_item(trans, parent_root,
 901                                 dentry->d_name.name, dentry->d_name.len,
 902                                 parent_inode->i_ino, &key,
 903                                 BTRFS_FT_DIR, index);
 904         BUG_ON(ret);
 905
 906         btrfs_i_size_write(parent_inode, parent_inode->i_size +
 907                                          dentry->d_name.len * 2);
 908         ret = btrfs_update_inode(trans, parent_root, parent_inode);
 909         BUG_ON(ret);
 910
 911         record_root_in_trans(trans, root);
 912         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
 913         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
 914
 915         old = btrfs_lock_root_node(root);
 916         btrfs_cow_block(trans, root, old, NULL, 0, &old);
 917         btrfs_set_lock_blocking(old);
 918
 919         btrfs_copy_root(trans, root, old, &tmp, objectid);
 920         btrfs_tree_unlock(old);
 921         free_extent_buffer(old);
 922
 923         btrfs_set_root_node(new_root_item, tmp);
 924         /* record when the snapshot was created in key.offset */
 925         key.offset = trans->transid;
 926         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
 927         btrfs_tree_unlock(tmp);
 928         free_extent_buffer(tmp);
 929         BUG_ON(ret);
 930
 931         /*
 932          * insert root back/forward references
 933          */
 934         ret = btrfs_add_root_ref(trans, tree_root, objectid,
 935                                  parent_root->root_key.objectid,
 936                                  parent_inode->i_ino, index,
 937                                  dentry->d_name.name, dentry->d_name.len);
 938         BUG_ON(ret);
 939
 940         key.offset = (u64)-1;
 941         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
 942         BUG_ON(IS_ERR(pending->snap));
 943
 944         btrfs_reloc_post_snapshot(trans, pending);
 945         btrfs_orphan_post_snapshot(trans, pending);
 946 fail:
 947         kfree(new_root_item);
 948         btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
 949         return 0;
 950 }
 951
 952 /*
 953  * create all the snapshots we've scheduled for creation
 954  */
 955 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
 956                                              struct btrfs_fs_info *fs_info)
 957 {
 958         struct btrfs_pending_snapshot *pending;
 959         struct list_head *head = &trans->transaction->pending_snapshots;
 960         int ret;
 961
 962         list_for_each_entry(pending, head, list) {
 963                 ret = create_pending_snapshot(trans, fs_info, pending);
 964                 BUG_ON(ret);
 965         }
 966         return 0;
 967 }
 968
 969 static void update_super_roots(struct btrfs_root *root)
 970 {
 971         struct btrfs_root_item *root_item;
 972         struct btrfs_super_block *super;
 973
 974         super = &root->fs_info->super_copy;
 975
 976         root_item = &root->fs_info->chunk_root->root_item;
 977         super->chunk_root = root_item->bytenr;
 978         super->chunk_root_generation = root_item->generation;
 979         super->chunk_root_level = root_item->level;
 980
 981         root_item = &root->fs_info->tree_root->root_item;
 982         super->root = root_item->bytenr;
 983         super->generation = root_item->generation;
 984         super->root_level = root_item->level;
 985         if (super->cache_generation != 0 || btrfs_test_opt(root, SPACE_CACHE))
 986                 super->cache_generation = root_item->generation;
 987 }
 988
 989 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
 990 {
 991         int ret = 0;
 992         spin_lock(&info->new_trans_lock);
 993         if (info->running_transaction)
 994                 ret = info->running_transaction->in_commit;
 995         spin_unlock(&info->new_trans_lock);
 996         return ret;
 997 }
 998
 999 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1000 {
1001         int ret = 0;
1002         spin_lock(&info->new_trans_lock);
1003         if (info->running_transaction)
1004                 ret = info->running_transaction->blocked;
1005         spin_unlock(&info->new_trans_lock);
1006         return ret;
1007 }
1008
1009 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1010                              struct btrfs_root *root)
1011 {
1012         unsigned long joined = 0;
1013         unsigned long timeout = 1;
1014         struct btrfs_transaction *cur_trans;
1015         struct btrfs_transaction *prev_trans = NULL;
1016         DEFINE_WAIT(wait);
1017         int ret;
1018         int should_grow = 0;
1019         unsigned long now = get_seconds();
1020         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1021
1022         btrfs_run_ordered_operations(root, 0);
1023
1024         /* make a pass through all the delayed refs we have so far
1025          * any runnings procs may add more while we are here
1026          */
1027         ret = btrfs_run_delayed_refs(trans, root, 0);
1028         BUG_ON(ret);
1029
1030         btrfs_trans_release_metadata(trans, root);
1031
1032         cur_trans = trans->transaction;
1033         /*
1034          * set the flushing flag so procs in this transaction have to
1035          * start sending their work down.
1036          */
1037         cur_trans->delayed_refs.flushing = 1;
1038
1039         ret = btrfs_run_delayed_refs(trans, root, 0);
1040         BUG_ON(ret);
1041
1042         mutex_lock(&root->fs_info->trans_mutex);
1043         if (cur_trans->in_commit) {
1044                 cur_trans->use_count++;
1045                 mutex_unlock(&root->fs_info->trans_mutex);
1046                 btrfs_end_transaction(trans, root);
1047
1048                 ret = wait_for_commit(root, cur_trans);
1049                 BUG_ON(ret);
1050
1051                 mutex_lock(&root->fs_info->trans_mutex);
1052                 put_transaction(cur_trans);
1053                 mutex_unlock(&root->fs_info->trans_mutex);
1054
1055                 return 0;
1056         }
1057
1058         trans->transaction->in_commit = 1;
1059         trans->transaction->blocked = 1;
1060         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1061                 prev_trans = list_entry(cur_trans->list.prev,
1062                                         struct btrfs_transaction, list);
1063                 if (!prev_trans->commit_done) {
1064                         prev_trans->use_count++;
1065                         mutex_unlock(&root->fs_info->trans_mutex);
1066
1067                         wait_for_commit(root, prev_trans);
1068
1069                         mutex_lock(&root->fs_info->trans_mutex);
1070                         put_transaction(prev_trans);
1071                 }
1072         }
1073
1074         if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
1075                 should_grow = 1;
1076
1077         do {
1078                 int snap_pending = 0;
1079                 joined = cur_trans->num_joined;
1080                 if (!list_empty(&trans->transaction->pending_snapshots))
1081                         snap_pending = 1;
1082
1083                 WARN_ON(cur_trans != trans->transaction);
1084                 if (cur_trans->num_writers > 1)
1085                         timeout = MAX_SCHEDULE_TIMEOUT;
1086                 else if (should_grow)
1087                         timeout = 1;
1088
1089                 mutex_unlock(&root->fs_info->trans_mutex);
1090
1091                 if (flush_on_commit || snap_pending) {
1092                         btrfs_start_delalloc_inodes(root, 1);
1093                         ret = btrfs_wait_ordered_extents(root, 0, 1);
1094                         BUG_ON(ret);
1095                 }
1096
1097                 /*
1098                  * rename don't use btrfs_join_transaction, so, once we
1099                  * set the transaction to blocked above, we aren't going
1100                  * to get any new ordered operations.  We can safely run
1101                  * it here and no for sure that nothing new will be added
1102                  * to the list
1103                  */
1104                 btrfs_run_ordered_operations(root, 1);
1105
1106                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1107                                 TASK_UNINTERRUPTIBLE);
1108
1109                 smp_mb();
1110                 if (cur_trans->num_writers > 1 || should_grow)
1111                         schedule_timeout(timeout);
1112
1113                 mutex_lock(&root->fs_info->trans_mutex);
1114                 finish_wait(&cur_trans->writer_wait, &wait);
1115         } while (cur_trans->num_writers > 1 ||
1116                  (should_grow && cur_trans->num_joined != joined));
1117
1118         ret = create_pending_snapshots(trans, root->fs_info);
1119         BUG_ON(ret);
1120
1121         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1122         BUG_ON(ret);
1123
1124         WARN_ON(cur_trans != trans->transaction);
1125
1126         /* btrfs_commit_tree_roots is responsible for getting the
1127          * various roots consistent with each other.  Every pointer
1128          * in the tree of tree roots has to point to the most up to date
1129          * root for every subvolume and other tree.  So, we have to keep
1130          * the tree logging code from jumping in and changing any
1131          * of the trees.
1132          *
1133          * At this point in the commit, there can't be any tree-log
1134          * writers, but a little lower down we drop the trans mutex
1135          * and let new people in.  By holding the tree_log_mutex
1136          * from now until after the super is written, we avoid races
1137          * with the tree-log code.
1138          */
1139         mutex_lock(&root->fs_info->tree_log_mutex);
1140
1141         ret = commit_fs_roots(trans, root);
1142         BUG_ON(ret);
1143
1144         /* commit_fs_roots gets rid of all the tree log roots, it is now
1145          * safe to free the root of tree log roots
1146          */
1147         btrfs_free_log_root_tree(trans, root->fs_info);
1148
1149         ret = commit_cowonly_roots(trans, root);
1150         BUG_ON(ret);
1151
1152         btrfs_prepare_extent_commit(trans, root);
1153
1154         cur_trans = root->fs_info->running_transaction;
1155         spin_lock(&root->fs_info->new_trans_lock);
1156         root->fs_info->running_transaction = NULL;
1157         spin_unlock(&root->fs_info->new_trans_lock);
1158
1159         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1160                             root->fs_info->tree_root->node);
1161         switch_commit_root(root->fs_info->tree_root);
1162
1163         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1164                             root->fs_info->chunk_root->node);
1165         switch_commit_root(root->fs_info->chunk_root);
1166
1167         update_super_roots(root);
1168
1169         if (!root->fs_info->log_root_recovering) {
1170                 btrfs_set_super_log_root(&root->fs_info->super_copy, 0);
1171                 btrfs_set_super_log_root_level(&root->fs_info->super_copy, 0);
1172         }
1173
1174         memcpy(&root->fs_info->super_for_commit, &root->fs_info->super_copy,
1175                sizeof(root->fs_info->super_copy));
1176
1177         trans->transaction->blocked = 0;
1178
1179         wake_up(&root->fs_info->transaction_wait);
1180
1181         mutex_unlock(&root->fs_info->trans_mutex);
1182         ret = btrfs_write_and_wait_transaction(trans, root);
1183         BUG_ON(ret);
1184         write_ctree_super(trans, root, 0);
1185
1186         /*
1187          * the super is written, we can safely allow the tree-loggers
1188          * to go about their business
1189          */
1190         mutex_unlock(&root->fs_info->tree_log_mutex);
1191
1192         btrfs_finish_extent_commit(trans, root);
1193
1194         mutex_lock(&root->fs_info->trans_mutex);
1195
1196         cur_trans->commit_done = 1;
1197
1198         root->fs_info->last_trans_committed = cur_trans->transid;
1199
1200         wake_up(&cur_trans->commit_wait);
1201
1202         put_transaction(cur_trans);
1203         put_transaction(cur_trans);
1204
1205         mutex_unlock(&root->fs_info->trans_mutex);
1206
1207         if (current->journal_info == trans)
1208                 current->journal_info = NULL;
1209
1210         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1211
1212         if (current != root->fs_info->transaction_kthread)
1213                 btrfs_run_delayed_iputs(root);
1214
1215         return ret;
1216 }
1217
1218 /*
1219  * interface function to delete all the snapshots we have scheduled for deletion
1220  */
1221 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1222 {
1223         LIST_HEAD(list);
1224         struct btrfs_fs_info *fs_info = root->fs_info;
1225
1226         mutex_lock(&fs_info->trans_mutex);
1227         list_splice_init(&fs_info->dead_roots, &list);
1228         mutex_unlock(&fs_info->trans_mutex);
1229
1230         while (!list_empty(&list)) {
1231                 root = list_entry(list.next, struct btrfs_root, root_list);
1232                 list_del(&root->root_list);
1233
1234                 if (btrfs_header_backref_rev(root->node) <
1235                     BTRFS_MIXED_BACKREF_REV)
1236                         btrfs_drop_snapshot(root, NULL, 0);
1237                 else
1238                         btrfs_drop_snapshot(root, NULL, 1);
1239         }
1240         return 0;
1241 }