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 retries = 0;
 184         int ret;
 185 again:
 186         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 187         if (!h)
 188                 return ERR_PTR(-ENOMEM);
 189
 190         if (type != TRANS_JOIN_NOLOCK)
 191                 mutex_lock(&root->fs_info->trans_mutex);
 192         if (may_wait_transaction(root, type))
 193                 wait_current_trans(root);
 194
 195         ret = join_transaction(root);
 196         BUG_ON(ret);
 197
 198         cur_trans = root->fs_info->running_transaction;
 199         cur_trans->use_count++;
 200         if (type != TRANS_JOIN_NOLOCK)
 201                 mutex_unlock(&root->fs_info->trans_mutex);
 202
 203         h->transid = cur_trans->transid;
 204         h->transaction = cur_trans;
 205         h->blocks_used = 0;
 206         h->block_group = 0;
 207         h->bytes_reserved = 0;
 208         h->delayed_ref_updates = 0;
 209         h->block_rsv = NULL;
 210
 211         smp_mb();
 212         if (cur_trans->blocked && may_wait_transaction(root, type)) {
 213                 btrfs_commit_transaction(h, root);
 214                 goto again;
 215         }
 216
 217         if (num_items > 0) {
 218                 ret = btrfs_trans_reserve_metadata(h, root, num_items,
 219                                                    &retries);
 220                 if (ret == -EAGAIN) {
 221                         btrfs_commit_transaction(h, root);
 222                         goto again;
 223                 }
 224                 if (ret < 0) {
 225                         btrfs_end_transaction(h, root);
 226                         return ERR_PTR(ret);
 227                 }
 228         }
 229
 230         if (type != TRANS_JOIN_NOLOCK)
 231                 mutex_lock(&root->fs_info->trans_mutex);
 232         record_root_in_trans(h, root);
 233         if (type != TRANS_JOIN_NOLOCK)
 234                 mutex_unlock(&root->fs_info->trans_mutex);
 235
 236         if (!current->journal_info && type != TRANS_USERSPACE)
 237                 current->journal_info = h;
 238         return h;
 239 }
 240
 241 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 242                                                    int num_items)
 243 {
 244         return start_transaction(root, num_items, TRANS_START);
 245 }
 246 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root,
 247                                                    int num_blocks)
 248 {
 249         return start_transaction(root, 0, TRANS_JOIN);
 250 }
 251
 252 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root,
 253                                                           int num_blocks)
 254 {
 255         return start_transaction(root, 0, TRANS_JOIN_NOLOCK);
 256 }
 257
 258 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *r,
 259                                                          int num_blocks)
 260 {
 261         return start_transaction(r, 0, TRANS_USERSPACE);
 262 }
 263
 264 /* wait for a transaction commit to be fully complete */
 265 static noinline int wait_for_commit(struct btrfs_root *root,
 266                                     struct btrfs_transaction *commit)
 267 {
 268         DEFINE_WAIT(wait);
 269         mutex_lock(&root->fs_info->trans_mutex);
 270         while (!commit->commit_done) {
 271                 prepare_to_wait(&commit->commit_wait, &wait,
 272                                 TASK_UNINTERRUPTIBLE);
 273                 if (commit->commit_done)
 274                         break;
 275                 mutex_unlock(&root->fs_info->trans_mutex);
 276                 schedule();
 277                 mutex_lock(&root->fs_info->trans_mutex);
 278         }
 279         mutex_unlock(&root->fs_info->trans_mutex);
 280         finish_wait(&commit->commit_wait, &wait);
 281         return 0;
 282 }
 283
 284 #if 0
 285 /*
 286  * rate limit against the drop_snapshot code.  This helps to slow down new
 287  * operations if the drop_snapshot code isn't able to keep up.
 288  */
 289 static void throttle_on_drops(struct btrfs_root *root)
 290 {
 291         struct btrfs_fs_info *info = root->fs_info;
 292         int harder_count = 0;
 293
 294 harder:
 295         if (atomic_read(&info->throttles)) {
 296                 DEFINE_WAIT(wait);
 297                 int thr;
 298                 thr = atomic_read(&info->throttle_gen);
 299
 300                 do {
 301                         prepare_to_wait(&info->transaction_throttle,
 302                                         &wait, TASK_UNINTERRUPTIBLE);
 303                         if (!atomic_read(&info->throttles)) {
 304                                 finish_wait(&info->transaction_throttle, &wait);
 305                                 break;
 306                         }
 307                         schedule();
 308                         finish_wait(&info->transaction_throttle, &wait);
 309                 } while (thr == atomic_read(&info->throttle_gen));
 310                 harder_count++;
 311
 312                 if (root->fs_info->total_ref_cache_size > 1 * 1024 * 1024 &&
 313                     harder_count < 2)
 314                         goto harder;
 315
 316                 if (root->fs_info->total_ref_cache_size > 5 * 1024 * 1024 &&
 317                     harder_count < 10)
 318                         goto harder;
 319
 320                 if (root->fs_info->total_ref_cache_size > 10 * 1024 * 1024 &&
 321                     harder_count < 20)
 322                         goto harder;
 323         }
 324 }
 325 #endif
 326
 327 void btrfs_throttle(struct btrfs_root *root)
 328 {
 329         mutex_lock(&root->fs_info->trans_mutex);
 330         if (!root->fs_info->open_ioctl_trans)
 331                 wait_current_trans(root);
 332         mutex_unlock(&root->fs_info->trans_mutex);
 333 }
 334
 335 static int should_end_transaction(struct btrfs_trans_handle *trans,
 336                                   struct btrfs_root *root)
 337 {
 338         int ret;
 339         ret = btrfs_block_rsv_check(trans, root,
 340                                     &root->fs_info->global_block_rsv, 0, 5);
 341         return ret ? 1 : 0;
 342 }
 343
 344 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 345                                  struct btrfs_root *root)
 346 {
 347         struct btrfs_transaction *cur_trans = trans->transaction;
 348         int updates;
 349
 350         if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
 351                 return 1;
 352
 353         updates = trans->delayed_ref_updates;
 354         trans->delayed_ref_updates = 0;
 355         if (updates)
 356                 btrfs_run_delayed_refs(trans, root, updates);
 357
 358         return should_end_transaction(trans, root);
 359 }
 360
 361 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 362                           struct btrfs_root *root, int throttle, int lock)
 363 {
 364         struct btrfs_transaction *cur_trans = trans->transaction;
 365         struct btrfs_fs_info *info = root->fs_info;
 366         int count = 0;
 367
 368         while (count < 4) {
 369                 unsigned long cur = trans->delayed_ref_updates;
 370                 trans->delayed_ref_updates = 0;
 371                 if (cur &&
 372                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 373                         trans->delayed_ref_updates = 0;
 374
 375                         /*
 376                          * do a full flush if the transaction is trying
 377                          * to close
 378                          */
 379                         if (trans->transaction->delayed_refs.flushing)
 380                                 cur = 0;
 381                         btrfs_run_delayed_refs(trans, root, cur);
 382                 } else {
 383                         break;
 384                 }
 385                 count++;
 386         }
 387
 388         btrfs_trans_release_metadata(trans, root);
 389
 390         if (lock && !root->fs_info->open_ioctl_trans &&
 391             should_end_transaction(trans, root))
 392                 trans->transaction->blocked = 1;
 393
 394         if (lock && cur_trans->blocked && !cur_trans->in_commit) {
 395                 if (throttle)
 396                         return btrfs_commit_transaction(trans, root);
 397                 else
 398                         wake_up_process(info->transaction_kthread);
 399         }
 400
 401         if (lock)
 402                 mutex_lock(&info->trans_mutex);
 403         WARN_ON(cur_trans != info->running_transaction);
 404         WARN_ON(cur_trans->num_writers < 1);
 405         cur_trans->num_writers--;
 406
 407         if (waitqueue_active(&cur_trans->writer_wait))
 408                 wake_up(&cur_trans->writer_wait);
 409         put_transaction(cur_trans);
 410         if (lock)
 411                 mutex_unlock(&info->trans_mutex);
 412
 413         if (current->journal_info == trans)
 414                 current->journal_info = NULL;
 415         memset(trans, 0, sizeof(*trans));
 416         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 417
 418         if (throttle)
 419                 btrfs_run_delayed_iputs(root);
 420
 421         return 0;
 422 }
 423
 424 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 425                           struct btrfs_root *root)
 426 {
 427         return __btrfs_end_transaction(trans, root, 0, 1);
 428 }
 429
 430 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 431                                    struct btrfs_root *root)
 432 {
 433         return __btrfs_end_transaction(trans, root, 1, 1);
 434 }
 435
 436 int btrfs_end_transaction_nolock(struct btrfs_trans_handle *trans,
 437                                  struct btrfs_root *root)
 438 {
 439         return __btrfs_end_transaction(trans, root, 0, 0);
 440 }
 441
 442 /*
 443  * when btree blocks are allocated, they have some corresponding bits set for
 444  * them in one of two extent_io trees.  This is used to make sure all of
 445  * those extents are sent to disk but does not wait on them
 446  */
 447 int btrfs_write_marked_extents(struct btrfs_root *root,
 448                                struct extent_io_tree *dirty_pages, int mark)
 449 {
 450         int ret;
 451         int err = 0;
 452         int werr = 0;
 453         struct page *page;
 454         struct inode *btree_inode = root->fs_info->btree_inode;
 455         u64 start = 0;
 456         u64 end;
 457         unsigned long index;
 458
 459         while (1) {
 460                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 461                                             mark);
 462                 if (ret)
 463                         break;
 464                 while (start <= end) {
 465                         cond_resched();
 466
 467                         index = start >> PAGE_CACHE_SHIFT;
 468                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 469                         page = find_get_page(btree_inode->i_mapping, index);
 470                         if (!page)
 471                                 continue;
 472
 473                         btree_lock_page_hook(page);
 474                         if (!page->mapping) {
 475                                 unlock_page(page);
 476                                 page_cache_release(page);
 477                                 continue;
 478                         }
 479
 480                         if (PageWriteback(page)) {
 481                                 if (PageDirty(page))
 482                                         wait_on_page_writeback(page);
 483                                 else {
 484                                         unlock_page(page);
 485                                         page_cache_release(page);
 486                                         continue;
 487                                 }
 488                         }
 489                         err = write_one_page(page, 0);
 490                         if (err)
 491                                 werr = err;
 492                         page_cache_release(page);
 493                 }
 494         }
 495         if (err)
 496                 werr = err;
 497         return werr;
 498 }
 499
 500 /*
 501  * when btree blocks are allocated, they have some corresponding bits set for
 502  * them in one of two extent_io trees.  This is used to make sure all of
 503  * those extents are on disk for transaction or log commit.  We wait
 504  * on all the pages and clear them from the dirty pages state tree
 505  */
 506 int btrfs_wait_marked_extents(struct btrfs_root *root,
 507                               struct extent_io_tree *dirty_pages, int mark)
 508 {
 509         int ret;
 510         int err = 0;
 511         int werr = 0;
 512         struct page *page;
 513         struct inode *btree_inode = root->fs_info->btree_inode;
 514         u64 start = 0;
 515         u64 end;
 516         unsigned long index;
 517
 518         while (1) {
 519                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 520                                             mark);
 521                 if (ret)
 522                         break;
 523
 524                 clear_extent_bits(dirty_pages, start, end, mark, GFP_NOFS);
 525                 while (start <= end) {
 526                         index = start >> PAGE_CACHE_SHIFT;
 527                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 528                         page = find_get_page(btree_inode->i_mapping, index);
 529                         if (!page)
 530                                 continue;
 531                         if (PageDirty(page)) {
 532                                 btree_lock_page_hook(page);
 533                                 wait_on_page_writeback(page);
 534                                 err = write_one_page(page, 0);
 535                                 if (err)
 536                                         werr = err;
 537                         }
 538                         wait_on_page_writeback(page);
 539                         page_cache_release(page);
 540                         cond_resched();
 541                 }
 542         }
 543         if (err)
 544                 werr = err;
 545         return werr;
 546 }
 547
 548 /*
 549  * when btree blocks are allocated, they have some corresponding bits set for
 550  * them in one of two extent_io trees.  This is used to make sure all of
 551  * those extents are on disk for transaction or log commit
 552  */
 553 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 554                                 struct extent_io_tree *dirty_pages, int mark)
 555 {
 556         int ret;
 557         int ret2;
 558
 559         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 560         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 561         return ret || ret2;
 562 }
 563
 564 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 565                                      struct btrfs_root *root)
 566 {
 567         if (!trans || !trans->transaction) {
 568                 struct inode *btree_inode;
 569                 btree_inode = root->fs_info->btree_inode;
 570                 return filemap_write_and_wait(btree_inode->i_mapping);
 571         }
 572         return btrfs_write_and_wait_marked_extents(root,
 573                                            &trans->transaction->dirty_pages,
 574                                            EXTENT_DIRTY);
 575 }
 576
 577 /*
 578  * this is used to update the root pointer in the tree of tree roots.
 579  *
 580  * But, in the case of the extent allocation tree, updating the root
 581  * pointer may allocate blocks which may change the root of the extent
 582  * allocation tree.
 583  *
 584  * So, this loops and repeats and makes sure the cowonly root didn't
 585  * change while the root pointer was being updated in the metadata.
 586  */
 587 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 588                                struct btrfs_root *root)
 589 {
 590         int ret;
 591         u64 old_root_bytenr;
 592         u64 old_root_used;
 593         struct btrfs_root *tree_root = root->fs_info->tree_root;
 594
 595         old_root_used = btrfs_root_used(&root->root_item);
 596         btrfs_write_dirty_block_groups(trans, root);
 597
 598         while (1) {
 599                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 600                 if (old_root_bytenr == root->node->start &&
 601                     old_root_used == btrfs_root_used(&root->root_item))
 602                         break;
 603
 604                 btrfs_set_root_node(&root->root_item, root->node);
 605                 ret = btrfs_update_root(trans, tree_root,
 606                                         &root->root_key,
 607                                         &root->root_item);
 608                 BUG_ON(ret);
 609
 610                 old_root_used = btrfs_root_used(&root->root_item);
 611                 ret = btrfs_write_dirty_block_groups(trans, root);
 612                 BUG_ON(ret);
 613         }
 614
 615         if (root != root->fs_info->extent_root)
 616                 switch_commit_root(root);
 617
 618         return 0;
 619 }
 620
 621 /*
 622  * update all the cowonly tree roots on disk
 623  */
 624 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 625                                          struct btrfs_root *root)
 626 {
 627         struct btrfs_fs_info *fs_info = root->fs_info;
 628         struct list_head *next;
 629         struct extent_buffer *eb;
 630         int ret;
 631
 632         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 633         BUG_ON(ret);
 634
 635         eb = btrfs_lock_root_node(fs_info->tree_root);
 636         btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
 637         btrfs_tree_unlock(eb);
 638         free_extent_buffer(eb);
 639
 640         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 641         BUG_ON(ret);
 642
 643         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 644                 next = fs_info->dirty_cowonly_roots.next;
 645                 list_del_init(next);
 646                 root = list_entry(next, struct btrfs_root, dirty_list);
 647
 648                 update_cowonly_root(trans, root);
 649         }
 650
 651         down_write(&fs_info->extent_commit_sem);
 652         switch_commit_root(fs_info->extent_root);
 653         up_write(&fs_info->extent_commit_sem);
 654
 655         return 0;
 656 }
 657
 658 /*
 659  * dead roots are old snapshots that need to be deleted.  This allocates
 660  * a dirty root struct and adds it into the list of dead roots that need to
 661  * be deleted
 662  */
 663 int btrfs_add_dead_root(struct btrfs_root *root)
 664 {
 665         mutex_lock(&root->fs_info->trans_mutex);
 666         list_add(&root->root_list, &root->fs_info->dead_roots);
 667         mutex_unlock(&root->fs_info->trans_mutex);
 668         return 0;
 669 }
 670
 671 /*
 672  * update all the cowonly tree roots on disk
 673  */
 674 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 675                                     struct btrfs_root *root)
 676 {
 677         struct btrfs_root *gang[8];
 678         struct btrfs_fs_info *fs_info = root->fs_info;
 679         int i;
 680         int ret;
 681         int err = 0;
 682
 683         while (1) {
 684                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 685                                                  (void **)gang, 0,
 686                                                  ARRAY_SIZE(gang),
 687                                                  BTRFS_ROOT_TRANS_TAG);
 688                 if (ret == 0)
 689                         break;
 690                 for (i = 0; i < ret; i++) {
 691                         root = gang[i];
 692                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 693                                         (unsigned long)root->root_key.objectid,
 694                                         BTRFS_ROOT_TRANS_TAG);
 695
 696                         btrfs_free_log(trans, root);
 697                         btrfs_update_reloc_root(trans, root);
 698                         btrfs_orphan_commit_root(trans, root);
 699
 700                         if (root->commit_root != root->node) {
 701                                 switch_commit_root(root);
 702                                 btrfs_set_root_node(&root->root_item,
 703                                                     root->node);
 704                         }
 705
 706                         err = btrfs_update_root(trans, fs_info->tree_root,
 707                                                 &root->root_key,
 708                                                 &root->root_item);
 709                         if (err)
 710                                 break;
 711                 }
 712         }
 713         return err;
 714 }
 715
 716 /*
 717  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 718  * otherwise every leaf in the btree is read and defragged.
 719  */
 720 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 721 {
 722         struct btrfs_fs_info *info = root->fs_info;
 723         struct btrfs_trans_handle *trans;
 724         int ret;
 725         unsigned long nr;
 726
 727         if (xchg(&root->defrag_running, 1))
 728                 return 0;
 729
 730         while (1) {
 731                 trans = btrfs_start_transaction(root, 0);
 732                 if (IS_ERR(trans))
 733                         return PTR_ERR(trans);
 734
 735                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 736
 737                 nr = trans->blocks_used;
 738                 btrfs_end_transaction(trans, root);
 739                 btrfs_btree_balance_dirty(info->tree_root, nr);
 740                 cond_resched();
 741
 742                 if (root->fs_info->closing || ret != -EAGAIN)
 743                         break;
 744         }
 745         root->defrag_running = 0;
 746         return ret;
 747 }
 748
 749 #if 0
 750 /*
 751  * when dropping snapshots, we generate a ton of delayed refs, and it makes
 752  * sense not to join the transaction while it is trying to flush the current
 753  * queue of delayed refs out.
 754  *
 755  * This is used by the drop snapshot code only
 756  */
 757 static noinline int wait_transaction_pre_flush(struct btrfs_fs_info *info)
 758 {
 759         DEFINE_WAIT(wait);
 760
 761         mutex_lock(&info->trans_mutex);
 762         while (info->running_transaction &&
 763                info->running_transaction->delayed_refs.flushing) {
 764                 prepare_to_wait(&info->transaction_wait, &wait,
 765                                 TASK_UNINTERRUPTIBLE);
 766                 mutex_unlock(&info->trans_mutex);
 767
 768                 schedule();
 769
 770                 mutex_lock(&info->trans_mutex);
 771                 finish_wait(&info->transaction_wait, &wait);
 772         }
 773         mutex_unlock(&info->trans_mutex);
 774         return 0;
 775 }
 776
 777 /*
 778  * Given a list of roots that need to be deleted, call btrfs_drop_snapshot on
 779  * all of them
 780  */
 781 int btrfs_drop_dead_root(struct btrfs_root *root)
 782 {
 783         struct btrfs_trans_handle *trans;
 784         struct btrfs_root *tree_root = root->fs_info->tree_root;
 785         unsigned long nr;
 786         int ret;
 787
 788         while (1) {
 789                 /*
 790                  * we don't want to jump in and create a bunch of
 791                  * delayed refs if the transaction is starting to close
 792                  */
 793                 wait_transaction_pre_flush(tree_root->fs_info);
 794                 trans = btrfs_start_transaction(tree_root, 1);
 795
 796                 /*
 797                  * we've joined a transaction, make sure it isn't
 798                  * closing right now
 799                  */
 800                 if (trans->transaction->delayed_refs.flushing) {
 801                         btrfs_end_transaction(trans, tree_root);
 802                         continue;
 803                 }
 804
 805                 ret = btrfs_drop_snapshot(trans, root);
 806                 if (ret != -EAGAIN)
 807                         break;
 808
 809                 ret = btrfs_update_root(trans, tree_root,
 810                                         &root->root_key,
 811                                         &root->root_item);
 812                 if (ret)
 813                         break;
 814
 815                 nr = trans->blocks_used;
 816                 ret = btrfs_end_transaction(trans, tree_root);
 817                 BUG_ON(ret);
 818
 819                 btrfs_btree_balance_dirty(tree_root, nr);
 820                 cond_resched();
 821         }
 822         BUG_ON(ret);
 823
 824         ret = btrfs_del_root(trans, tree_root, &root->root_key);
 825         BUG_ON(ret);
 826
 827         nr = trans->blocks_used;
 828         ret = btrfs_end_transaction(trans, tree_root);
 829         BUG_ON(ret);
 830
 831         free_extent_buffer(root->node);
 832         free_extent_buffer(root->commit_root);
 833         kfree(root);
 834
 835         btrfs_btree_balance_dirty(tree_root, nr);
 836         return ret;
 837 }
 838 #endif
 839
 840 /*
 841  * new snapshots need to be created at a very specific time in the
 842  * transaction commit.  This does the actual creation
 843  */
 844 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 845                                    struct btrfs_fs_info *fs_info,
 846                                    struct btrfs_pending_snapshot *pending)
 847 {
 848         struct btrfs_key key;
 849         struct btrfs_root_item *new_root_item;
 850         struct btrfs_root *tree_root = fs_info->tree_root;
 851         struct btrfs_root *root = pending->root;
 852         struct btrfs_root *parent_root;
 853         struct inode *parent_inode;
 854         struct dentry *dentry;
 855         struct extent_buffer *tmp;
 856         struct extent_buffer *old;
 857         int ret;
 858         int retries = 0;
 859         u64 to_reserve = 0;
 860         u64 index = 0;
 861         u64 objectid;
 862
 863         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 864         if (!new_root_item) {
 865                 pending->error = -ENOMEM;
 866                 goto fail;
 867         }
 868
 869         ret = btrfs_find_free_objectid(trans, tree_root, 0, &objectid);
 870         if (ret) {
 871                 pending->error = ret;
 872                 goto fail;
 873         }
 874
 875         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
 876         btrfs_orphan_pre_snapshot(trans, pending, &to_reserve);
 877
 878         if (to_reserve > 0) {
 879                 ret = btrfs_block_rsv_add(trans, root, &pending->block_rsv,
 880                                           to_reserve, &retries);
 881                 if (ret) {
 882                         pending->error = ret;
 883                         goto fail;
 884                 }
 885         }
 886
 887         key.objectid = objectid;
 888         key.offset = (u64)-1;
 889         key.type = BTRFS_ROOT_ITEM_KEY;
 890
 891         trans->block_rsv = &pending->block_rsv;
 892
 893         dentry = pending->dentry;
 894         parent_inode = dentry->d_parent->d_inode;
 895         parent_root = BTRFS_I(parent_inode)->root;
 896         record_root_in_trans(trans, parent_root);
 897
 898         /*
 899          * insert the directory item
 900          */
 901         ret = btrfs_set_inode_index(parent_inode, &index);
 902         BUG_ON(ret);
 903         ret = btrfs_insert_dir_item(trans, parent_root,
 904                                 dentry->d_name.name, dentry->d_name.len,
 905                                 parent_inode->i_ino, &key,
 906                                 BTRFS_FT_DIR, index);
 907         BUG_ON(ret);
 908
 909         btrfs_i_size_write(parent_inode, parent_inode->i_size +
 910                                          dentry->d_name.len * 2);
 911         ret = btrfs_update_inode(trans, parent_root, parent_inode);
 912         BUG_ON(ret);
 913
 914         record_root_in_trans(trans, root);
 915         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
 916         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
 917
 918         old = btrfs_lock_root_node(root);
 919         btrfs_cow_block(trans, root, old, NULL, 0, &old);
 920         btrfs_set_lock_blocking(old);
 921
 922         btrfs_copy_root(trans, root, old, &tmp, objectid);
 923         btrfs_tree_unlock(old);
 924         free_extent_buffer(old);
 925
 926         btrfs_set_root_node(new_root_item, tmp);
 927         /* record when the snapshot was created in key.offset */
 928         key.offset = trans->transid;
 929         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
 930         btrfs_tree_unlock(tmp);
 931         free_extent_buffer(tmp);
 932         BUG_ON(ret);
 933
 934         /*
 935          * insert root back/forward references
 936          */
 937         ret = btrfs_add_root_ref(trans, tree_root, objectid,
 938                                  parent_root->root_key.objectid,
 939                                  parent_inode->i_ino, index,
 940                                  dentry->d_name.name, dentry->d_name.len);
 941         BUG_ON(ret);
 942
 943         key.offset = (u64)-1;
 944         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
 945         BUG_ON(IS_ERR(pending->snap));
 946
 947         btrfs_reloc_post_snapshot(trans, pending);
 948         btrfs_orphan_post_snapshot(trans, pending);
 949 fail:
 950         kfree(new_root_item);
 951         btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
 952         return 0;
 953 }
 954
 955 /*
 956  * create all the snapshots we've scheduled for creation
 957  */
 958 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
 959                                              struct btrfs_fs_info *fs_info)
 960 {
 961         struct btrfs_pending_snapshot *pending;
 962         struct list_head *head = &trans->transaction->pending_snapshots;
 963         int ret;
 964
 965         list_for_each_entry(pending, head, list) {
 966                 ret = create_pending_snapshot(trans, fs_info, pending);
 967                 BUG_ON(ret);
 968         }
 969         return 0;
 970 }
 971
 972 static void update_super_roots(struct btrfs_root *root)
 973 {
 974         struct btrfs_root_item *root_item;
 975         struct btrfs_super_block *super;
 976
 977         super = &root->fs_info->super_copy;
 978
 979         root_item = &root->fs_info->chunk_root->root_item;
 980         super->chunk_root = root_item->bytenr;
 981         super->chunk_root_generation = root_item->generation;
 982         super->chunk_root_level = root_item->level;
 983
 984         root_item = &root->fs_info->tree_root->root_item;
 985         super->root = root_item->bytenr;
 986         super->generation = root_item->generation;
 987         super->root_level = root_item->level;
 988         if (super->cache_generation != 0 || btrfs_test_opt(root, SPACE_CACHE))
 989                 super->cache_generation = root_item->generation;
 990 }
 991
 992 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
 993 {
 994         int ret = 0;
 995         spin_lock(&info->new_trans_lock);
 996         if (info->running_transaction)
 997                 ret = info->running_transaction->in_commit;
 998         spin_unlock(&info->new_trans_lock);
 999         return ret;
1000 }
1001
1002 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1003 {
1004         int ret = 0;
1005         spin_lock(&info->new_trans_lock);
1006         if (info->running_transaction)
1007                 ret = info->running_transaction->blocked;
1008         spin_unlock(&info->new_trans_lock);
1009         return ret;
1010 }
1011
1012 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1013                              struct btrfs_root *root)
1014 {
1015         unsigned long joined = 0;
1016         unsigned long timeout = 1;
1017         struct btrfs_transaction *cur_trans;
1018         struct btrfs_transaction *prev_trans = NULL;
1019         DEFINE_WAIT(wait);
1020         int ret;
1021         int should_grow = 0;
1022         unsigned long now = get_seconds();
1023         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1024
1025         btrfs_run_ordered_operations(root, 0);
1026
1027         /* make a pass through all the delayed refs we have so far
1028          * any runnings procs may add more while we are here
1029          */
1030         ret = btrfs_run_delayed_refs(trans, root, 0);
1031         BUG_ON(ret);
1032
1033         btrfs_trans_release_metadata(trans, root);
1034
1035         cur_trans = trans->transaction;
1036         /*
1037          * set the flushing flag so procs in this transaction have to
1038          * start sending their work down.
1039          */
1040         cur_trans->delayed_refs.flushing = 1;
1041
1042         ret = btrfs_run_delayed_refs(trans, root, 0);
1043         BUG_ON(ret);
1044
1045         mutex_lock(&root->fs_info->trans_mutex);
1046         if (cur_trans->in_commit) {
1047                 cur_trans->use_count++;
1048                 mutex_unlock(&root->fs_info->trans_mutex);
1049                 btrfs_end_transaction(trans, root);
1050
1051                 ret = wait_for_commit(root, cur_trans);
1052                 BUG_ON(ret);
1053
1054                 mutex_lock(&root->fs_info->trans_mutex);
1055                 put_transaction(cur_trans);
1056                 mutex_unlock(&root->fs_info->trans_mutex);
1057
1058                 return 0;
1059         }
1060
1061         trans->transaction->in_commit = 1;
1062         trans->transaction->blocked = 1;
1063         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1064                 prev_trans = list_entry(cur_trans->list.prev,
1065                                         struct btrfs_transaction, list);
1066                 if (!prev_trans->commit_done) {
1067                         prev_trans->use_count++;
1068                         mutex_unlock(&root->fs_info->trans_mutex);
1069
1070                         wait_for_commit(root, prev_trans);
1071
1072                         mutex_lock(&root->fs_info->trans_mutex);
1073                         put_transaction(prev_trans);
1074                 }
1075         }
1076
1077         if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
1078                 should_grow = 1;
1079
1080         do {
1081                 int snap_pending = 0;
1082                 joined = cur_trans->num_joined;
1083                 if (!list_empty(&trans->transaction->pending_snapshots))
1084                         snap_pending = 1;
1085
1086                 WARN_ON(cur_trans != trans->transaction);
1087                 if (cur_trans->num_writers > 1)
1088                         timeout = MAX_SCHEDULE_TIMEOUT;
1089                 else if (should_grow)
1090                         timeout = 1;
1091
1092                 mutex_unlock(&root->fs_info->trans_mutex);
1093
1094                 if (flush_on_commit || snap_pending) {
1095                         btrfs_start_delalloc_inodes(root, 1);
1096                         ret = btrfs_wait_ordered_extents(root, 0, 1);
1097                         BUG_ON(ret);
1098                 }
1099
1100                 /*
1101                  * rename don't use btrfs_join_transaction, so, once we
1102                  * set the transaction to blocked above, we aren't going
1103                  * to get any new ordered operations.  We can safely run
1104                  * it here and no for sure that nothing new will be added
1105                  * to the list
1106                  */
1107                 btrfs_run_ordered_operations(root, 1);
1108
1109                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1110                                 TASK_UNINTERRUPTIBLE);
1111
1112                 smp_mb();
1113                 if (cur_trans->num_writers > 1 || should_grow)
1114                         schedule_timeout(timeout);
1115
1116                 mutex_lock(&root->fs_info->trans_mutex);
1117                 finish_wait(&cur_trans->writer_wait, &wait);
1118         } while (cur_trans->num_writers > 1 ||
1119                  (should_grow && cur_trans->num_joined != joined));
1120
1121         ret = create_pending_snapshots(trans, root->fs_info);
1122         BUG_ON(ret);
1123
1124         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1125         BUG_ON(ret);
1126
1127         WARN_ON(cur_trans != trans->transaction);
1128
1129         /* btrfs_commit_tree_roots is responsible for getting the
1130          * various roots consistent with each other.  Every pointer
1131          * in the tree of tree roots has to point to the most up to date
1132          * root for every subvolume and other tree.  So, we have to keep
1133          * the tree logging code from jumping in and changing any
1134          * of the trees.
1135          *
1136          * At this point in the commit, there can't be any tree-log
1137          * writers, but a little lower down we drop the trans mutex
1138          * and let new people in.  By holding the tree_log_mutex
1139          * from now until after the super is written, we avoid races
1140          * with the tree-log code.
1141          */
1142         mutex_lock(&root->fs_info->tree_log_mutex);
1143
1144         ret = commit_fs_roots(trans, root);
1145         BUG_ON(ret);
1146
1147         /* commit_fs_roots gets rid of all the tree log roots, it is now
1148          * safe to free the root of tree log roots
1149          */
1150         btrfs_free_log_root_tree(trans, root->fs_info);
1151
1152         ret = commit_cowonly_roots(trans, root);
1153         BUG_ON(ret);
1154
1155         btrfs_prepare_extent_commit(trans, root);
1156
1157         cur_trans = root->fs_info->running_transaction;
1158         spin_lock(&root->fs_info->new_trans_lock);
1159         root->fs_info->running_transaction = NULL;
1160         spin_unlock(&root->fs_info->new_trans_lock);
1161
1162         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1163                             root->fs_info->tree_root->node);
1164         switch_commit_root(root->fs_info->tree_root);
1165
1166         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1167                             root->fs_info->chunk_root->node);
1168         switch_commit_root(root->fs_info->chunk_root);
1169
1170         update_super_roots(root);
1171
1172         if (!root->fs_info->log_root_recovering) {
1173                 btrfs_set_super_log_root(&root->fs_info->super_copy, 0);
1174                 btrfs_set_super_log_root_level(&root->fs_info->super_copy, 0);
1175         }
1176
1177         memcpy(&root->fs_info->super_for_commit, &root->fs_info->super_copy,
1178                sizeof(root->fs_info->super_copy));
1179
1180         trans->transaction->blocked = 0;
1181
1182         wake_up(&root->fs_info->transaction_wait);
1183
1184         mutex_unlock(&root->fs_info->trans_mutex);
1185         ret = btrfs_write_and_wait_transaction(trans, root);
1186         BUG_ON(ret);
1187         write_ctree_super(trans, root, 0);
1188
1189         /*
1190          * the super is written, we can safely allow the tree-loggers
1191          * to go about their business
1192          */
1193         mutex_unlock(&root->fs_info->tree_log_mutex);
1194
1195         btrfs_finish_extent_commit(trans, root);
1196
1197         mutex_lock(&root->fs_info->trans_mutex);
1198
1199         cur_trans->commit_done = 1;
1200
1201         root->fs_info->last_trans_committed = cur_trans->transid;
1202
1203         wake_up(&cur_trans->commit_wait);
1204
1205         put_transaction(cur_trans);
1206         put_transaction(cur_trans);
1207
1208         mutex_unlock(&root->fs_info->trans_mutex);
1209
1210         if (current->journal_info == trans)
1211                 current->journal_info = NULL;
1212
1213         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1214
1215         if (current != root->fs_info->transaction_kthread)
1216                 btrfs_run_delayed_iputs(root);
1217
1218         return ret;
1219 }
1220
1221 /*
1222  * interface function to delete all the snapshots we have scheduled for deletion
1223  */
1224 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1225 {
1226         LIST_HEAD(list);
1227         struct btrfs_fs_info *fs_info = root->fs_info;
1228
1229         mutex_lock(&fs_info->trans_mutex);
1230         list_splice_init(&fs_info->dead_roots, &list);
1231         mutex_unlock(&fs_info->trans_mutex);
1232
1233         while (!list_empty(&list)) {
1234                 root = list_entry(list.next, struct btrfs_root, root_list);
1235                 list_del(&root->root_list);
1236
1237                 if (btrfs_header_backref_rev(root->node) <
1238                     BTRFS_MIXED_BACKREF_REV)
1239                         btrfs_drop_snapshot(root, NULL, 0);
1240                 else
1241                         btrfs_drop_snapshot(root, NULL, 1);
1242         }
1243         return 0;
1244 }