fs/btrfs/transaction.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/fs.h>
  20 #include <linux/sched.h>
  21 #include <linux/writeback.h>
  22 #include <linux/pagemap.h>
  23 #include "ctree.h"
  24 #include "disk-io.h"
  25 #include "transaction.h"
  26 #include "locking.h"
  27 #include "ref-cache.h"
  28 #include "tree-log.h"
  29
  30 static int total_trans = 0;
  31 extern struct kmem_cache *btrfs_trans_handle_cachep;
  32 extern struct kmem_cache *btrfs_transaction_cachep;
  33
  34 #define BTRFS_ROOT_TRANS_TAG 0
  35
  36 static noinline void put_transaction(struct btrfs_transaction *transaction)
  37 {
  38         WARN_ON(transaction->use_count == 0);
  39         transaction->use_count--;
  40         if (transaction->use_count == 0) {
  41                 WARN_ON(total_trans == 0);
  42                 total_trans--;
  43                 list_del_init(&transaction->list);
  44                 memset(transaction, 0, sizeof(*transaction));
  45                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  46         }
  47 }
  48
  49 /*
  50  * either allocate a new transaction or hop into the existing one
  51  */
  52 static noinline int join_transaction(struct btrfs_root *root)
  53 {
  54         struct btrfs_transaction *cur_trans;
  55         cur_trans = root->fs_info->running_transaction;
  56         if (!cur_trans) {
  57                 cur_trans = kmem_cache_alloc(btrfs_transaction_cachep,
  58                                              GFP_NOFS);
  59                 total_trans++;
  60                 BUG_ON(!cur_trans);
  61                 root->fs_info->generation++;
  62                 root->fs_info->last_alloc = 0;
  63                 root->fs_info->last_data_alloc = 0;
  64                 root->fs_info->last_log_alloc = 0;
  65                 cur_trans->num_writers = 1;
  66                 cur_trans->num_joined = 0;
  67                 cur_trans->transid = root->fs_info->generation;
  68                 init_waitqueue_head(&cur_trans->writer_wait);
  69                 init_waitqueue_head(&cur_trans->commit_wait);
  70                 cur_trans->in_commit = 0;
  71                 cur_trans->blocked = 0;
  72                 cur_trans->use_count = 1;
  73                 cur_trans->commit_done = 0;
  74                 cur_trans->start_time = get_seconds();
  75                 INIT_LIST_HEAD(&cur_trans->pending_snapshots);
  76                 list_add_tail(&cur_trans->list, &root->fs_info->trans_list);
  77                 extent_io_tree_init(&cur_trans->dirty_pages,
  78                                      root->fs_info->btree_inode->i_mapping,
  79                                      GFP_NOFS);
  80                 spin_lock(&root->fs_info->new_trans_lock);
  81                 root->fs_info->running_transaction = cur_trans;
  82                 spin_unlock(&root->fs_info->new_trans_lock);
  83         } else {
  84                 cur_trans->num_writers++;
  85                 cur_trans->num_joined++;
  86         }
  87
  88         return 0;
  89 }
  90
  91 /*
  92  * this does all the record keeping required to make sure that a
  93  * reference counted root is properly recorded in a given transaction.
  94  * This is required to make sure the old root from before we joined the transaction
  95  * is deleted when the transaction commits
  96  */
  97 noinline int btrfs_record_root_in_trans(struct btrfs_root *root)
  98 {
  99         struct btrfs_dirty_root *dirty;
 100         u64 running_trans_id = root->fs_info->running_transaction->transid;
 101         if (root->ref_cows && root->last_trans < running_trans_id) {
 102                 WARN_ON(root == root->fs_info->extent_root);
 103                 if (root->root_item.refs != 0) {
 104                         radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 105                                    (unsigned long)root->root_key.objectid,
 106                                    BTRFS_ROOT_TRANS_TAG);
 107
 108                         dirty = kmalloc(sizeof(*dirty), GFP_NOFS);
 109                         BUG_ON(!dirty);
 110                         dirty->root = kmalloc(sizeof(*dirty->root), GFP_NOFS);
 111                         BUG_ON(!dirty->root);
 112                         dirty->latest_root = root;
 113                         INIT_LIST_HEAD(&dirty->list);
 114
 115                         root->commit_root = btrfs_root_node(root);
 116
 117                         memcpy(dirty->root, root, sizeof(*root));
 118                         spin_lock_init(&dirty->root->node_lock);
 119                         spin_lock_init(&dirty->root->list_lock);
 120                         mutex_init(&dirty->root->objectid_mutex);
 121                         mutex_init(&dirty->root->log_mutex);
 122                         INIT_LIST_HEAD(&dirty->root->dead_list);
 123                         dirty->root->node = root->commit_root;
 124                         dirty->root->commit_root = NULL;
 125
 126                         spin_lock(&root->list_lock);
 127                         list_add(&dirty->root->dead_list, &root->dead_list);
 128                         spin_unlock(&root->list_lock);
 129
 130                         root->dirty_root = dirty;
 131                 } else {
 132                         WARN_ON(1);
 133                 }
 134                 root->last_trans = running_trans_id;
 135         }
 136         return 0;
 137 }
 138
 139 /* wait for commit against the current transaction to become unblocked
 140  * when this is done, it is safe to start a new transaction, but the current
 141  * transaction might not be fully on disk.
 142  */
 143 static void wait_current_trans(struct btrfs_root *root)
 144 {
 145         struct btrfs_transaction *cur_trans;
 146
 147         cur_trans = root->fs_info->running_transaction;
 148         if (cur_trans && cur_trans->blocked) {
 149                 DEFINE_WAIT(wait);
 150                 cur_trans->use_count++;
 151                 while(1) {
 152                         prepare_to_wait(&root->fs_info->transaction_wait, &wait,
 153                                         TASK_UNINTERRUPTIBLE);
 154                         if (cur_trans->blocked) {
 155                                 mutex_unlock(&root->fs_info->trans_mutex);
 156                                 schedule();
 157                                 mutex_lock(&root->fs_info->trans_mutex);
 158                                 finish_wait(&root->fs_info->transaction_wait,
 159                                             &wait);
 160                         } else {
 161                                 finish_wait(&root->fs_info->transaction_wait,
 162                                             &wait);
 163                                 break;
 164                         }
 165                 }
 166                 put_transaction(cur_trans);
 167         }
 168 }
 169
 170 static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
 171                                              int num_blocks, int wait)
 172 {
 173         struct btrfs_trans_handle *h =
 174                 kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 175         int ret;
 176
 177         mutex_lock(&root->fs_info->trans_mutex);
 178         if (!root->fs_info->log_root_recovering &&
 179             ((wait == 1 && !root->fs_info->open_ioctl_trans) || wait == 2))
 180                 wait_current_trans(root);
 181         ret = join_transaction(root);
 182         BUG_ON(ret);
 183
 184         btrfs_record_root_in_trans(root);
 185         h->transid = root->fs_info->running_transaction->transid;
 186         h->transaction = root->fs_info->running_transaction;
 187         h->blocks_reserved = num_blocks;
 188         h->blocks_used = 0;
 189         h->block_group = NULL;
 190         h->alloc_exclude_nr = 0;
 191         h->alloc_exclude_start = 0;
 192         root->fs_info->running_transaction->use_count++;
 193         mutex_unlock(&root->fs_info->trans_mutex);
 194         return h;
 195 }
 196
 197 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 198                                                    int num_blocks)
 199 {
 200         return start_transaction(root, num_blocks, 1);
 201 }
 202 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root,
 203                                                    int num_blocks)
 204 {
 205         return start_transaction(root, num_blocks, 0);
 206 }
 207
 208 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *r,
 209                                                          int num_blocks)
 210 {
 211         return start_transaction(r, num_blocks, 2);
 212 }
 213
 214 /* wait for a transaction commit to be fully complete */
 215 static noinline int wait_for_commit(struct btrfs_root *root,
 216                                     struct btrfs_transaction *commit)
 217 {
 218         DEFINE_WAIT(wait);
 219         mutex_lock(&root->fs_info->trans_mutex);
 220         while(!commit->commit_done) {
 221                 prepare_to_wait(&commit->commit_wait, &wait,
 222                                 TASK_UNINTERRUPTIBLE);
 223                 if (commit->commit_done)
 224                         break;
 225                 mutex_unlock(&root->fs_info->trans_mutex);
 226                 schedule();
 227                 mutex_lock(&root->fs_info->trans_mutex);
 228         }
 229         mutex_unlock(&root->fs_info->trans_mutex);
 230         finish_wait(&commit->commit_wait, &wait);
 231         return 0;
 232 }
 233
 234 /*
 235  * rate limit against the drop_snapshot code.  This helps to slow down new operations
 236  * if the drop_snapshot code isn't able to keep up.
 237  */
 238 static void throttle_on_drops(struct btrfs_root *root)
 239 {
 240         struct btrfs_fs_info *info = root->fs_info;
 241         int harder_count = 0;
 242
 243 harder:
 244         if (atomic_read(&info->throttles)) {
 245                 DEFINE_WAIT(wait);
 246                 int thr;
 247                 thr = atomic_read(&info->throttle_gen);
 248
 249                 do {
 250                         prepare_to_wait(&info->transaction_throttle,
 251                                         &wait, TASK_UNINTERRUPTIBLE);
 252                         if (!atomic_read(&info->throttles)) {
 253                                 finish_wait(&info->transaction_throttle, &wait);
 254                                 break;
 255                         }
 256                         schedule();
 257                         finish_wait(&info->transaction_throttle, &wait);
 258                 } while (thr == atomic_read(&info->throttle_gen));
 259                 harder_count++;
 260
 261                 if (root->fs_info->total_ref_cache_size > 1 * 1024 * 1024 &&
 262                     harder_count < 2)
 263                         goto harder;
 264
 265                 if (root->fs_info->total_ref_cache_size > 5 * 1024 * 1024 &&
 266                     harder_count < 10)
 267                         goto harder;
 268
 269                 if (root->fs_info->total_ref_cache_size > 10 * 1024 * 1024 &&
 270                     harder_count < 20)
 271                         goto harder;
 272         }
 273 }
 274
 275 void btrfs_throttle(struct btrfs_root *root)
 276 {
 277         mutex_lock(&root->fs_info->trans_mutex);
 278         if (!root->fs_info->open_ioctl_trans)
 279                 wait_current_trans(root);
 280         mutex_unlock(&root->fs_info->trans_mutex);
 281
 282         throttle_on_drops(root);
 283 }
 284
 285 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 286                           struct btrfs_root *root, int throttle)
 287 {
 288         struct btrfs_transaction *cur_trans;
 289         struct btrfs_fs_info *info = root->fs_info;
 290
 291         mutex_lock(&info->trans_mutex);
 292         cur_trans = info->running_transaction;
 293         WARN_ON(cur_trans != trans->transaction);
 294         WARN_ON(cur_trans->num_writers < 1);
 295         cur_trans->num_writers--;
 296
 297         if (waitqueue_active(&cur_trans->writer_wait))
 298                 wake_up(&cur_trans->writer_wait);
 299         put_transaction(cur_trans);
 300         mutex_unlock(&info->trans_mutex);
 301         memset(trans, 0, sizeof(*trans));
 302         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 303
 304         if (throttle)
 305                 throttle_on_drops(root);
 306
 307         return 0;
 308 }
 309
 310 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 311                           struct btrfs_root *root)
 312 {
 313         return __btrfs_end_transaction(trans, root, 0);
 314 }
 315
 316 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 317                                    struct btrfs_root *root)
 318 {
 319         return __btrfs_end_transaction(trans, root, 1);
 320 }
 321
 322 /*
 323  * when btree blocks are allocated, they have some corresponding bits set for
 324  * them in one of two extent_io trees.  This is used to make sure all of
 325  * those extents are on disk for transaction or log commit
 326  */
 327 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 328                                         struct extent_io_tree *dirty_pages)
 329 {
 330         int ret;
 331         int err = 0;
 332         int werr = 0;
 333         struct page *page;
 334         struct inode *btree_inode = root->fs_info->btree_inode;
 335         u64 start = 0;
 336         u64 end;
 337         unsigned long index;
 338
 339         while(1) {
 340                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 341                                             EXTENT_DIRTY);
 342                 if (ret)
 343                         break;
 344                 while(start <= end) {
 345                         cond_resched();
 346
 347                         index = start >> PAGE_CACHE_SHIFT;
 348                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 349                         page = find_get_page(btree_inode->i_mapping, index);
 350                         if (!page)
 351                                 continue;
 352
 353                         btree_lock_page_hook(page);
 354                         if (!page->mapping) {
 355                                 unlock_page(page);
 356                                 page_cache_release(page);
 357                                 continue;
 358                         }
 359
 360                         if (PageWriteback(page)) {
 361                                 if (PageDirty(page))
 362                                         wait_on_page_writeback(page);
 363                                 else {
 364                                         unlock_page(page);
 365                                         page_cache_release(page);
 366                                         continue;
 367                                 }
 368                         }
 369                         err = write_one_page(page, 0);
 370                         if (err)
 371                                 werr = err;
 372                         page_cache_release(page);
 373                 }
 374         }
 375         while(1) {
 376                 ret = find_first_extent_bit(dirty_pages, 0, &start, &end,
 377                                             EXTENT_DIRTY);
 378                 if (ret)
 379                         break;
 380
 381                 clear_extent_dirty(dirty_pages, start, end, GFP_NOFS);
 382                 while(start <= end) {
 383                         index = start >> PAGE_CACHE_SHIFT;
 384                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 385                         page = find_get_page(btree_inode->i_mapping, index);
 386                         if (!page)
 387                                 continue;
 388                         if (PageDirty(page)) {
 389                                 btree_lock_page_hook(page);
 390                                 wait_on_page_writeback(page);
 391                                 err = write_one_page(page, 0);
 392                                 if (err)
 393                                         werr = err;
 394                         }
 395                         wait_on_page_writeback(page);
 396                         page_cache_release(page);
 397                         cond_resched();
 398                 }
 399         }
 400         if (err)
 401                 werr = err;
 402         return werr;
 403 }
 404
 405 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 406                                      struct btrfs_root *root)
 407 {
 408         if (!trans || !trans->transaction) {
 409                 struct inode *btree_inode;
 410                 btree_inode = root->fs_info->btree_inode;
 411                 return filemap_write_and_wait(btree_inode->i_mapping);
 412         }
 413         return btrfs_write_and_wait_marked_extents(root,
 414                                            &trans->transaction->dirty_pages);
 415 }
 416
 417 /*
 418  * this is used to update the root pointer in the tree of tree roots.
 419  *
 420  * But, in the case of the extent allocation tree, updating the root
 421  * pointer may allocate blocks which may change the root of the extent
 422  * allocation tree.
 423  *
 424  * So, this loops and repeats and makes sure the cowonly root didn't
 425  * change while the root pointer was being updated in the metadata.
 426  */
 427 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 428                                struct btrfs_root *root)
 429 {
 430         int ret;
 431         u64 old_root_bytenr;
 432         struct btrfs_root *tree_root = root->fs_info->tree_root;
 433
 434         btrfs_write_dirty_block_groups(trans, root);
 435         while(1) {
 436                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 437                 if (old_root_bytenr == root->node->start)
 438                         break;
 439                 btrfs_set_root_bytenr(&root->root_item,
 440                                        root->node->start);
 441                 btrfs_set_root_level(&root->root_item,
 442                                      btrfs_header_level(root->node));
 443                 ret = btrfs_update_root(trans, tree_root,
 444                                         &root->root_key,
 445                                         &root->root_item);
 446                 BUG_ON(ret);
 447                 btrfs_write_dirty_block_groups(trans, root);
 448         }
 449         return 0;
 450 }
 451
 452 /*
 453  * update all the cowonly tree roots on disk
 454  */
 455 int btrfs_commit_tree_roots(struct btrfs_trans_handle *trans,
 456                             struct btrfs_root *root)
 457 {
 458         struct btrfs_fs_info *fs_info = root->fs_info;
 459         struct list_head *next;
 460
 461         while(!list_empty(&fs_info->dirty_cowonly_roots)) {
 462                 next = fs_info->dirty_cowonly_roots.next;
 463                 list_del_init(next);
 464                 root = list_entry(next, struct btrfs_root, dirty_list);
 465                 update_cowonly_root(trans, root);
 466         }
 467         return 0;
 468 }
 469
 470 /*
 471  * dead roots are old snapshots that need to be deleted.  This allocates
 472  * a dirty root struct and adds it into the list of dead roots that need to
 473  * be deleted
 474  */
 475 int btrfs_add_dead_root(struct btrfs_root *root, struct btrfs_root *latest)
 476 {
 477         struct btrfs_dirty_root *dirty;
 478
 479         dirty = kmalloc(sizeof(*dirty), GFP_NOFS);
 480         if (!dirty)
 481                 return -ENOMEM;
 482         dirty->root = root;
 483         dirty->latest_root = latest;
 484
 485         mutex_lock(&root->fs_info->trans_mutex);
 486         list_add(&dirty->list, &latest->fs_info->dead_roots);
 487         mutex_unlock(&root->fs_info->trans_mutex);
 488         return 0;
 489 }
 490
 491 /*
 492  * at transaction commit time we need to schedule the old roots for
 493  * deletion via btrfs_drop_snapshot.  This runs through all the
 494  * reference counted roots that were modified in the current
 495  * transaction and puts them into the drop list
 496  */
 497 static noinline int add_dirty_roots(struct btrfs_trans_handle *trans,
 498                                     struct radix_tree_root *radix,
 499                                     struct list_head *list)
 500 {
 501         struct btrfs_dirty_root *dirty;
 502         struct btrfs_root *gang[8];
 503         struct btrfs_root *root;
 504         int i;
 505         int ret;
 506         int err = 0;
 507         u32 refs;
 508
 509         while(1) {
 510                 ret = radix_tree_gang_lookup_tag(radix, (void **)gang, 0,
 511                                                  ARRAY_SIZE(gang),
 512                                                  BTRFS_ROOT_TRANS_TAG);
 513                 if (ret == 0)
 514                         break;
 515                 for (i = 0; i < ret; i++) {
 516                         root = gang[i];
 517                         radix_tree_tag_clear(radix,
 518                                      (unsigned long)root->root_key.objectid,
 519                                      BTRFS_ROOT_TRANS_TAG);
 520
 521                         BUG_ON(!root->ref_tree);
 522                         dirty = root->dirty_root;
 523
 524                         btrfs_free_log(trans, root);
 525                         btrfs_free_reloc_root(root);
 526
 527                         if (root->commit_root == root->node) {
 528                                 WARN_ON(root->node->start !=
 529                                         btrfs_root_bytenr(&root->root_item));
 530
 531                                 free_extent_buffer(root->commit_root);
 532                                 root->commit_root = NULL;
 533                                 root->dirty_root = NULL;
 534
 535                                 spin_lock(&root->list_lock);
 536                                 list_del_init(&dirty->root->dead_list);
 537                                 spin_unlock(&root->list_lock);
 538
 539                                 kfree(dirty->root);
 540                                 kfree(dirty);
 541
 542                                 /* make sure to update the root on disk
 543                                  * so we get any updates to the block used
 544                                  * counts
 545                                  */
 546                                 err = btrfs_update_root(trans,
 547                                                 root->fs_info->tree_root,
 548                                                 &root->root_key,
 549                                                 &root->root_item);
 550                                 continue;
 551                         }
 552
 553                         memset(&root->root_item.drop_progress, 0,
 554                                sizeof(struct btrfs_disk_key));
 555                         root->root_item.drop_level = 0;
 556                         root->commit_root = NULL;
 557                         root->dirty_root = NULL;
 558                         root->root_key.offset = root->fs_info->generation;
 559                         btrfs_set_root_bytenr(&root->root_item,
 560                                               root->node->start);
 561                         btrfs_set_root_level(&root->root_item,
 562                                              btrfs_header_level(root->node));
 563                         err = btrfs_insert_root(trans, root->fs_info->tree_root,
 564                                                 &root->root_key,
 565                                                 &root->root_item);
 566                         if (err)
 567                                 break;
 568
 569                         refs = btrfs_root_refs(&dirty->root->root_item);
 570                         btrfs_set_root_refs(&dirty->root->root_item, refs - 1);
 571                         err = btrfs_update_root(trans, root->fs_info->tree_root,
 572                                                 &dirty->root->root_key,
 573                                                 &dirty->root->root_item);
 574
 575                         BUG_ON(err);
 576                         if (refs == 1) {
 577                                 list_add(&dirty->list, list);
 578                         } else {
 579                                 WARN_ON(1);
 580                                 free_extent_buffer(dirty->root->node);
 581                                 kfree(dirty->root);
 582                                 kfree(dirty);
 583                         }
 584                 }
 585         }
 586         return err;
 587 }
 588
 589 /*
 590  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 591  * otherwise every leaf in the btree is read and defragged.
 592  */
 593 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 594 {
 595         struct btrfs_fs_info *info = root->fs_info;
 596         int ret;
 597         struct btrfs_trans_handle *trans;
 598         unsigned long nr;
 599
 600         smp_mb();
 601         if (root->defrag_running)
 602                 return 0;
 603         trans = btrfs_start_transaction(root, 1);
 604         while (1) {
 605                 root->defrag_running = 1;
 606                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 607                 nr = trans->blocks_used;
 608                 btrfs_end_transaction(trans, root);
 609                 btrfs_btree_balance_dirty(info->tree_root, nr);
 610                 cond_resched();
 611
 612                 trans = btrfs_start_transaction(root, 1);
 613                 if (root->fs_info->closing || ret != -EAGAIN)
 614                         break;
 615         }
 616         root->defrag_running = 0;
 617         smp_mb();
 618         btrfs_end_transaction(trans, root);
 619         return 0;
 620 }
 621
 622 /*
 623  * Given a list of roots that need to be deleted, call btrfs_drop_snapshot on
 624  * all of them
 625  */
 626 static noinline int drop_dirty_roots(struct btrfs_root *tree_root,
 627                                      struct list_head *list)
 628 {
 629         struct btrfs_dirty_root *dirty;
 630         struct btrfs_trans_handle *trans;
 631         unsigned long nr;
 632         u64 num_bytes;
 633         u64 bytes_used;
 634         u64 max_useless;
 635         int ret = 0;
 636         int err;
 637
 638         while(!list_empty(list)) {
 639                 struct btrfs_root *root;
 640
 641                 dirty = list_entry(list->prev, struct btrfs_dirty_root, list);
 642                 list_del_init(&dirty->list);
 643
 644                 num_bytes = btrfs_root_used(&dirty->root->root_item);
 645                 root = dirty->latest_root;
 646                 atomic_inc(&root->fs_info->throttles);
 647
 648                 while(1) {
 649                         trans = btrfs_start_transaction(tree_root, 1);
 650                         mutex_lock(&root->fs_info->drop_mutex);
 651                         ret = btrfs_drop_snapshot(trans, dirty->root);
 652                         if (ret != -EAGAIN) {
 653                                 break;
 654                         }
 655                         mutex_unlock(&root->fs_info->drop_mutex);
 656
 657                         err = btrfs_update_root(trans,
 658                                         tree_root,
 659                                         &dirty->root->root_key,
 660                                         &dirty->root->root_item);
 661                         if (err)
 662                                 ret = err;
 663                         nr = trans->blocks_used;
 664                         ret = btrfs_end_transaction(trans, tree_root);
 665                         BUG_ON(ret);
 666
 667                         btrfs_btree_balance_dirty(tree_root, nr);
 668                         cond_resched();
 669                 }
 670                 BUG_ON(ret);
 671                 atomic_dec(&root->fs_info->throttles);
 672                 wake_up(&root->fs_info->transaction_throttle);
 673
 674                 mutex_lock(&root->fs_info->alloc_mutex);
 675                 num_bytes -= btrfs_root_used(&dirty->root->root_item);
 676                 bytes_used = btrfs_root_used(&root->root_item);
 677                 if (num_bytes) {
 678                         btrfs_record_root_in_trans(root);
 679                         btrfs_set_root_used(&root->root_item,
 680                                             bytes_used - num_bytes);
 681                 }
 682                 mutex_unlock(&root->fs_info->alloc_mutex);
 683
 684                 ret = btrfs_del_root(trans, tree_root, &dirty->root->root_key);
 685                 if (ret) {
 686                         BUG();
 687                         break;
 688                 }
 689                 mutex_unlock(&root->fs_info->drop_mutex);
 690
 691                 spin_lock(&root->list_lock);
 692                 list_del_init(&dirty->root->dead_list);
 693                 if (!list_empty(&root->dead_list)) {
 694                         struct btrfs_root *oldest;
 695                         oldest = list_entry(root->dead_list.prev,
 696                                             struct btrfs_root, dead_list);
 697                         max_useless = oldest->root_key.offset - 1;
 698                 } else {
 699                         max_useless = root->root_key.offset - 1;
 700                 }
 701                 spin_unlock(&root->list_lock);
 702
 703                 nr = trans->blocks_used;
 704                 ret = btrfs_end_transaction(trans, tree_root);
 705                 BUG_ON(ret);
 706
 707                 ret = btrfs_remove_leaf_refs(root, max_useless, 0);
 708                 BUG_ON(ret);
 709
 710                 free_extent_buffer(dirty->root->node);
 711                 kfree(dirty->root);
 712                 kfree(dirty);
 713
 714                 btrfs_btree_balance_dirty(tree_root, nr);
 715                 cond_resched();
 716         }
 717         return ret;
 718 }
 719
 720 /*
 721  * new snapshots need to be created at a very specific time in the
 722  * transaction commit.  This does the actual creation
 723  */
 724 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 725                                    struct btrfs_fs_info *fs_info,
 726                                    struct btrfs_pending_snapshot *pending)
 727 {
 728         struct btrfs_key key;
 729         struct btrfs_root_item *new_root_item;
 730         struct btrfs_root *tree_root = fs_info->tree_root;
 731         struct btrfs_root *root = pending->root;
 732         struct extent_buffer *tmp;
 733         struct extent_buffer *old;
 734         int ret;
 735         int namelen;
 736         u64 objectid;
 737
 738         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 739         if (!new_root_item) {
 740                 ret = -ENOMEM;
 741                 goto fail;
 742         }
 743         ret = btrfs_find_free_objectid(trans, tree_root, 0, &objectid);
 744         if (ret)
 745                 goto fail;
 746
 747         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
 748
 749         key.objectid = objectid;
 750         key.offset = trans->transid;
 751         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
 752
 753         old = btrfs_lock_root_node(root);
 754         btrfs_cow_block(trans, root, old, NULL, 0, &old, 0);
 755
 756         btrfs_copy_root(trans, root, old, &tmp, objectid);
 757         btrfs_tree_unlock(old);
 758         free_extent_buffer(old);
 759
 760         btrfs_set_root_bytenr(new_root_item, tmp->start);
 761         btrfs_set_root_level(new_root_item, btrfs_header_level(tmp));
 762         ret = btrfs_insert_root(trans, root->fs_info->tree_root, &key,
 763                                 new_root_item);
 764         btrfs_tree_unlock(tmp);
 765         free_extent_buffer(tmp);
 766         if (ret)
 767                 goto fail;
 768
 769         /*
 770          * insert the directory item
 771          */
 772         key.offset = (u64)-1;
 773         namelen = strlen(pending->name);
 774         ret = btrfs_insert_dir_item(trans, root->fs_info->tree_root,
 775                                     pending->name, namelen,
 776                                     root->fs_info->sb->s_root->d_inode->i_ino,
 777                                     &key, BTRFS_FT_DIR, 0);
 778
 779         if (ret)
 780                 goto fail;
 781
 782         ret = btrfs_insert_inode_ref(trans, root->fs_info->tree_root,
 783                              pending->name, strlen(pending->name), objectid,
 784                              root->fs_info->sb->s_root->d_inode->i_ino, 0);
 785
 786         /* Invalidate existing dcache entry for new snapshot. */
 787         btrfs_invalidate_dcache_root(root, pending->name, namelen);
 788
 789 fail:
 790         kfree(new_root_item);
 791         return ret;
 792 }
 793
 794 /*
 795  * create all the snapshots we've scheduled for creation
 796  */
 797 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
 798                                              struct btrfs_fs_info *fs_info)
 799 {
 800         struct btrfs_pending_snapshot *pending;
 801         struct list_head *head = &trans->transaction->pending_snapshots;
 802         int ret;
 803
 804         while(!list_empty(head)) {
 805                 pending = list_entry(head->next,
 806                                      struct btrfs_pending_snapshot, list);
 807                 ret = create_pending_snapshot(trans, fs_info, pending);
 808                 BUG_ON(ret);
 809                 list_del(&pending->list);
 810                 kfree(pending->name);
 811                 kfree(pending);
 812         }
 813         return 0;
 814 }
 815
 816 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
 817                              struct btrfs_root *root)
 818 {
 819         unsigned long joined = 0;
 820         unsigned long timeout = 1;
 821         struct btrfs_transaction *cur_trans;
 822         struct btrfs_transaction *prev_trans = NULL;
 823         struct btrfs_root *chunk_root = root->fs_info->chunk_root;
 824         struct list_head dirty_fs_roots;
 825         struct extent_io_tree *pinned_copy;
 826         DEFINE_WAIT(wait);
 827         int ret;
 828
 829         INIT_LIST_HEAD(&dirty_fs_roots);
 830         mutex_lock(&root->fs_info->trans_mutex);
 831         if (trans->transaction->in_commit) {
 832                 cur_trans = trans->transaction;
 833                 trans->transaction->use_count++;
 834                 mutex_unlock(&root->fs_info->trans_mutex);
 835                 btrfs_end_transaction(trans, root);
 836
 837                 ret = wait_for_commit(root, cur_trans);
 838                 BUG_ON(ret);
 839
 840                 mutex_lock(&root->fs_info->trans_mutex);
 841                 put_transaction(cur_trans);
 842                 mutex_unlock(&root->fs_info->trans_mutex);
 843
 844                 return 0;
 845         }
 846
 847         pinned_copy = kmalloc(sizeof(*pinned_copy), GFP_NOFS);
 848         if (!pinned_copy)
 849                 return -ENOMEM;
 850
 851         extent_io_tree_init(pinned_copy,
 852                              root->fs_info->btree_inode->i_mapping, GFP_NOFS);
 853
 854         trans->transaction->in_commit = 1;
 855         trans->transaction->blocked = 1;
 856         cur_trans = trans->transaction;
 857         if (cur_trans->list.prev != &root->fs_info->trans_list) {
 858                 prev_trans = list_entry(cur_trans->list.prev,
 859                                         struct btrfs_transaction, list);
 860                 if (!prev_trans->commit_done) {
 861                         prev_trans->use_count++;
 862                         mutex_unlock(&root->fs_info->trans_mutex);
 863
 864                         wait_for_commit(root, prev_trans);
 865
 866                         mutex_lock(&root->fs_info->trans_mutex);
 867                         put_transaction(prev_trans);
 868                 }
 869         }
 870
 871         do {
 872                 int snap_pending = 0;
 873                 joined = cur_trans->num_joined;
 874                 if (!list_empty(&trans->transaction->pending_snapshots))
 875                         snap_pending = 1;
 876
 877                 WARN_ON(cur_trans != trans->transaction);
 878                 prepare_to_wait(&cur_trans->writer_wait, &wait,
 879                                 TASK_UNINTERRUPTIBLE);
 880
 881                 if (cur_trans->num_writers > 1)
 882                         timeout = MAX_SCHEDULE_TIMEOUT;
 883                 else
 884                         timeout = 1;
 885
 886                 mutex_unlock(&root->fs_info->trans_mutex);
 887
 888                 if (snap_pending) {
 889                         ret = btrfs_wait_ordered_extents(root, 1);
 890                         BUG_ON(ret);
 891                 }
 892
 893                 schedule_timeout(timeout);
 894
 895                 mutex_lock(&root->fs_info->trans_mutex);
 896                 finish_wait(&cur_trans->writer_wait, &wait);
 897         } while (cur_trans->num_writers > 1 ||
 898                  (cur_trans->num_joined != joined));
 899
 900         ret = create_pending_snapshots(trans, root->fs_info);
 901         BUG_ON(ret);
 902
 903         WARN_ON(cur_trans != trans->transaction);
 904
 905         /* btrfs_commit_tree_roots is responsible for getting the
 906          * various roots consistent with each other.  Every pointer
 907          * in the tree of tree roots has to point to the most up to date
 908          * root for every subvolume and other tree.  So, we have to keep
 909          * the tree logging code from jumping in and changing any
 910          * of the trees.
 911          *
 912          * At this point in the commit, there can't be any tree-log
 913          * writers, but a little lower down we drop the trans mutex
 914          * and let new people in.  By holding the tree_log_mutex
 915          * from now until after the super is written, we avoid races
 916          * with the tree-log code.
 917          */
 918         mutex_lock(&root->fs_info->tree_log_mutex);
 919         /*
 920          * keep tree reloc code from adding new reloc trees
 921          */
 922         mutex_lock(&root->fs_info->tree_reloc_mutex);
 923
 924
 925         ret = add_dirty_roots(trans, &root->fs_info->fs_roots_radix,
 926                               &dirty_fs_roots);
 927         BUG_ON(ret);
 928
 929         /* add_dirty_roots gets rid of all the tree log roots, it is now
 930          * safe to free the root of tree log roots
 931          */
 932         btrfs_free_log_root_tree(trans, root->fs_info);
 933
 934         btrfs_free_reloc_mappings(root);
 935
 936         ret = btrfs_commit_tree_roots(trans, root);
 937         BUG_ON(ret);
 938
 939         cur_trans = root->fs_info->running_transaction;
 940         spin_lock(&root->fs_info->new_trans_lock);
 941         root->fs_info->running_transaction = NULL;
 942         spin_unlock(&root->fs_info->new_trans_lock);
 943         btrfs_set_super_generation(&root->fs_info->super_copy,
 944                                    cur_trans->transid);
 945         btrfs_set_super_root(&root->fs_info->super_copy,
 946                              root->fs_info->tree_root->node->start);
 947         btrfs_set_super_root_level(&root->fs_info->super_copy,
 948                            btrfs_header_level(root->fs_info->tree_root->node));
 949
 950         btrfs_set_super_chunk_root(&root->fs_info->super_copy,
 951                                    chunk_root->node->start);
 952         btrfs_set_super_chunk_root_level(&root->fs_info->super_copy,
 953                                          btrfs_header_level(chunk_root->node));
 954
 955         if (!root->fs_info->log_root_recovering) {
 956                 btrfs_set_super_log_root(&root->fs_info->super_copy, 0);
 957                 btrfs_set_super_log_root_level(&root->fs_info->super_copy, 0);
 958         }
 959
 960         memcpy(&root->fs_info->super_for_commit, &root->fs_info->super_copy,
 961                sizeof(root->fs_info->super_copy));
 962
 963         btrfs_copy_pinned(root, pinned_copy);
 964
 965         trans->transaction->blocked = 0;
 966         wake_up(&root->fs_info->transaction_throttle);
 967         wake_up(&root->fs_info->transaction_wait);
 968
 969         mutex_unlock(&root->fs_info->trans_mutex);
 970         ret = btrfs_write_and_wait_transaction(trans, root);
 971         BUG_ON(ret);
 972         write_ctree_super(trans, root);
 973
 974         /*
 975          * the super is written, we can safely allow the tree-loggers
 976          * to go about their business
 977          */
 978         mutex_unlock(&root->fs_info->tree_log_mutex);
 979
 980         btrfs_finish_extent_commit(trans, root, pinned_copy);
 981         kfree(pinned_copy);
 982
 983         btrfs_drop_dead_reloc_roots(root);
 984         mutex_unlock(&root->fs_info->tree_reloc_mutex);
 985
 986         mutex_lock(&root->fs_info->trans_mutex);
 987
 988         cur_trans->commit_done = 1;
 989         root->fs_info->last_trans_committed = cur_trans->transid;
 990         wake_up(&cur_trans->commit_wait);
 991         put_transaction(cur_trans);
 992         put_transaction(cur_trans);
 993
 994         list_splice_init(&dirty_fs_roots, &root->fs_info->dead_roots);
 995         if (root->fs_info->closing)
 996                 list_splice_init(&root->fs_info->dead_roots, &dirty_fs_roots);
 997
 998         mutex_unlock(&root->fs_info->trans_mutex);
 999         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1000
1001         if (root->fs_info->closing) {
1002                 drop_dirty_roots(root->fs_info->tree_root, &dirty_fs_roots);
1003         }
1004         return ret;
1005 }
1006
1007 /*
1008  * interface function to delete all the snapshots we have scheduled for deletion
1009  */
1010 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1011 {
1012         struct list_head dirty_roots;
1013         INIT_LIST_HEAD(&dirty_roots);
1014 again:
1015         mutex_lock(&root->fs_info->trans_mutex);
1016         list_splice_init(&root->fs_info->dead_roots, &dirty_roots);
1017         mutex_unlock(&root->fs_info->trans_mutex);
1018
1019         if (!list_empty(&dirty_roots)) {
1020                 drop_dirty_roots(root, &dirty_roots);
1021                 goto again;
1022         }
1023         return 0;
1024 }