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                 cur_trans->num_writers = 1;
  65                 cur_trans->num_joined = 0;
  66                 cur_trans->transid = root->fs_info->generation;
  67                 init_waitqueue_head(&cur_trans->writer_wait);
  68                 init_waitqueue_head(&cur_trans->commit_wait);
  69                 cur_trans->in_commit = 0;
  70                 cur_trans->blocked = 0;
  71                 cur_trans->use_count = 1;
  72                 cur_trans->commit_done = 0;
  73                 cur_trans->start_time = get_seconds();
  74                 INIT_LIST_HEAD(&cur_trans->pending_snapshots);
  75                 list_add_tail(&cur_trans->list, &root->fs_info->trans_list);
  76                 extent_io_tree_init(&cur_trans->dirty_pages,
  77                                      root->fs_info->btree_inode->i_mapping,
  78                                      GFP_NOFS);
  79                 spin_lock(&root->fs_info->new_trans_lock);
  80                 root->fs_info->running_transaction = cur_trans;
  81                 spin_unlock(&root->fs_info->new_trans_lock);
  82         } else {
  83                 cur_trans->num_writers++;
  84                 cur_trans->num_joined++;
  85         }
  86
  87         return 0;
  88 }
  89
  90 /*
  91  * this does all the record keeping required to make sure that a
  92  * reference counted root is properly recorded in a given transaction.
  93  * This is required to make sure the old root from before we joined the transaction
  94  * is deleted when the transaction commits
  95  */
  96 noinline int btrfs_record_root_in_trans(struct btrfs_root *root)
  97 {
  98         struct btrfs_dirty_root *dirty;
  99         u64 running_trans_id = root->fs_info->running_transaction->transid;
 100         if (root->ref_cows && root->last_trans < running_trans_id) {
 101                 WARN_ON(root == root->fs_info->extent_root);
 102                 if (root->root_item.refs != 0) {
 103                         radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 104                                    (unsigned long)root->root_key.objectid,
 105                                    BTRFS_ROOT_TRANS_TAG);
 106
 107                         dirty = kmalloc(sizeof(*dirty), GFP_NOFS);
 108                         BUG_ON(!dirty);
 109                         dirty->root = kmalloc(sizeof(*dirty->root), GFP_NOFS);
 110                         BUG_ON(!dirty->root);
 111                         dirty->latest_root = root;
 112                         INIT_LIST_HEAD(&dirty->list);
 113
 114                         root->commit_root = btrfs_root_node(root);
 115
 116                         memcpy(dirty->root, root, sizeof(*root));
 117                         spin_lock_init(&dirty->root->node_lock);
 118                         spin_lock_init(&dirty->root->list_lock);
 119                         mutex_init(&dirty->root->objectid_mutex);
 120                         mutex_init(&dirty->root->log_mutex);
 121                         INIT_LIST_HEAD(&dirty->root->dead_list);
 122                         dirty->root->node = root->commit_root;
 123                         dirty->root->commit_root = NULL;
 124
 125                         spin_lock(&root->list_lock);
 126                         list_add(&dirty->root->dead_list, &root->dead_list);
 127                         spin_unlock(&root->list_lock);
 128
 129                         root->dirty_root = dirty;
 130                 } else {
 131                         WARN_ON(1);
 132                 }
 133                 root->last_trans = running_trans_id;
 134         }
 135         return 0;
 136 }
 137
 138 /* wait for commit against the current transaction to become unblocked
 139  * when this is done, it is safe to start a new transaction, but the current
 140  * transaction might not be fully on disk.
 141  */
 142 static void wait_current_trans(struct btrfs_root *root)
 143 {
 144         struct btrfs_transaction *cur_trans;
 145
 146         cur_trans = root->fs_info->running_transaction;
 147         if (cur_trans && cur_trans->blocked) {
 148                 DEFINE_WAIT(wait);
 149                 cur_trans->use_count++;
 150                 while(1) {
 151                         prepare_to_wait(&root->fs_info->transaction_wait, &wait,
 152                                         TASK_UNINTERRUPTIBLE);
 153                         if (cur_trans->blocked) {
 154                                 mutex_unlock(&root->fs_info->trans_mutex);
 155                                 schedule();
 156                                 mutex_lock(&root->fs_info->trans_mutex);
 157                                 finish_wait(&root->fs_info->transaction_wait,
 158                                             &wait);
 159                         } else {
 160                                 finish_wait(&root->fs_info->transaction_wait,
 161                                             &wait);
 162                                 break;
 163                         }
 164                 }
 165                 put_transaction(cur_trans);
 166         }
 167 }
 168
 169 static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
 170                                              int num_blocks, int wait)
 171 {
 172         struct btrfs_trans_handle *h =
 173                 kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 174         int ret;
 175
 176         mutex_lock(&root->fs_info->trans_mutex);
 177         if (!root->fs_info->log_root_recovering &&
 178             ((wait == 1 && !root->fs_info->open_ioctl_trans) || wait == 2))
 179                 wait_current_trans(root);
 180         ret = join_transaction(root);
 181         BUG_ON(ret);
 182
 183         btrfs_record_root_in_trans(root);
 184         h->transid = root->fs_info->running_transaction->transid;
 185         h->transaction = root->fs_info->running_transaction;
 186         h->blocks_reserved = num_blocks;
 187         h->blocks_used = 0;
 188         h->block_group = NULL;
 189         h->alloc_exclude_nr = 0;
 190         h->alloc_exclude_start = 0;
 191         root->fs_info->running_transaction->use_count++;
 192         mutex_unlock(&root->fs_info->trans_mutex);
 193         return h;
 194 }
 195
 196 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 197                                                    int num_blocks)
 198 {
 199         return start_transaction(root, num_blocks, 1);
 200 }
 201 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root,
 202                                                    int num_blocks)
 203 {
 204         return start_transaction(root, num_blocks, 0);
 205 }
 206
 207 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *r,
 208                                                          int num_blocks)
 209 {
 210         return start_transaction(r, num_blocks, 2);
 211 }
 212
 213 /* wait for a transaction commit to be fully complete */
 214 static noinline int wait_for_commit(struct btrfs_root *root,
 215                                     struct btrfs_transaction *commit)
 216 {
 217         DEFINE_WAIT(wait);
 218         mutex_lock(&root->fs_info->trans_mutex);
 219         while(!commit->commit_done) {
 220                 prepare_to_wait(&commit->commit_wait, &wait,
 221                                 TASK_UNINTERRUPTIBLE);
 222                 if (commit->commit_done)
 223                         break;
 224                 mutex_unlock(&root->fs_info->trans_mutex);
 225                 schedule();
 226                 mutex_lock(&root->fs_info->trans_mutex);
 227         }
 228         mutex_unlock(&root->fs_info->trans_mutex);
 229         finish_wait(&commit->commit_wait, &wait);
 230         return 0;
 231 }
 232
 233 /*
 234  * rate limit against the drop_snapshot code.  This helps to slow down new operations
 235  * if the drop_snapshot code isn't able to keep up.
 236  */
 237 static void throttle_on_drops(struct btrfs_root *root)
 238 {
 239         struct btrfs_fs_info *info = root->fs_info;
 240         int harder_count = 0;
 241
 242 harder:
 243         if (atomic_read(&info->throttles)) {
 244                 DEFINE_WAIT(wait);
 245                 int thr;
 246                 thr = atomic_read(&info->throttle_gen);
 247
 248                 do {
 249                         prepare_to_wait(&info->transaction_throttle,
 250                                         &wait, TASK_UNINTERRUPTIBLE);
 251                         if (!atomic_read(&info->throttles)) {
 252                                 finish_wait(&info->transaction_throttle, &wait);
 253                                 break;
 254                         }
 255                         schedule();
 256                         finish_wait(&info->transaction_throttle, &wait);
 257                 } while (thr == atomic_read(&info->throttle_gen));
 258                 harder_count++;
 259
 260                 if (root->fs_info->total_ref_cache_size > 1 * 1024 * 1024 &&
 261                     harder_count < 2)
 262                         goto harder;
 263
 264                 if (root->fs_info->total_ref_cache_size > 5 * 1024 * 1024 &&
 265                     harder_count < 10)
 266                         goto harder;
 267
 268                 if (root->fs_info->total_ref_cache_size > 10 * 1024 * 1024 &&
 269                     harder_count < 20)
 270                         goto harder;
 271         }
 272 }
 273
 274 void btrfs_throttle(struct btrfs_root *root)
 275 {
 276         mutex_lock(&root->fs_info->trans_mutex);
 277         if (!root->fs_info->open_ioctl_trans)
 278                 wait_current_trans(root);
 279         mutex_unlock(&root->fs_info->trans_mutex);
 280
 281         throttle_on_drops(root);
 282 }
 283
 284 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 285                           struct btrfs_root *root, int throttle)
 286 {
 287         struct btrfs_transaction *cur_trans;
 288         struct btrfs_fs_info *info = root->fs_info;
 289
 290         mutex_lock(&info->trans_mutex);
 291         cur_trans = info->running_transaction;
 292         WARN_ON(cur_trans != trans->transaction);
 293         WARN_ON(cur_trans->num_writers < 1);
 294         cur_trans->num_writers--;
 295
 296         if (waitqueue_active(&cur_trans->writer_wait))
 297                 wake_up(&cur_trans->writer_wait);
 298         put_transaction(cur_trans);
 299         mutex_unlock(&info->trans_mutex);
 300         memset(trans, 0, sizeof(*trans));
 301         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 302
 303         if (throttle)
 304                 throttle_on_drops(root);
 305
 306         return 0;
 307 }
 308
 309 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 310                           struct btrfs_root *root)
 311 {
 312         return __btrfs_end_transaction(trans, root, 0);
 313 }
 314
 315 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 316                                    struct btrfs_root *root)
 317 {
 318         return __btrfs_end_transaction(trans, root, 1);
 319 }
 320
 321 /*
 322  * when btree blocks are allocated, they have some corresponding bits set for
 323  * them in one of two extent_io trees.  This is used to make sure all of
 324  * those extents are on disk for transaction or log commit
 325  */
 326 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 327                                         struct extent_io_tree *dirty_pages)
 328 {
 329         int ret;
 330         int err = 0;
 331         int werr = 0;
 332         struct page *page;
 333         struct inode *btree_inode = root->fs_info->btree_inode;
 334         u64 start = 0;
 335         u64 end;
 336         unsigned long index;
 337
 338         while(1) {
 339                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 340                                             EXTENT_DIRTY);
 341                 if (ret)
 342                         break;
 343                 while(start <= end) {
 344                         cond_resched();
 345
 346                         index = start >> PAGE_CACHE_SHIFT;
 347                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 348                         page = find_get_page(btree_inode->i_mapping, index);
 349                         if (!page)
 350                                 continue;
 351
 352                         btree_lock_page_hook(page);
 353                         if (!page->mapping) {
 354                                 unlock_page(page);
 355                                 page_cache_release(page);
 356                                 continue;
 357                         }
 358
 359                         if (PageWriteback(page)) {
 360                                 if (PageDirty(page))
 361                                         wait_on_page_writeback(page);
 362                                 else {
 363                                         unlock_page(page);
 364                                         page_cache_release(page);
 365                                         continue;
 366                                 }
 367                         }
 368                         err = write_one_page(page, 0);
 369                         if (err)
 370                                 werr = err;
 371                         page_cache_release(page);
 372                 }
 373         }
 374         while(1) {
 375                 ret = find_first_extent_bit(dirty_pages, 0, &start, &end,
 376                                             EXTENT_DIRTY);
 377                 if (ret)
 378                         break;
 379
 380                 clear_extent_dirty(dirty_pages, start, end, GFP_NOFS);
 381                 while(start <= end) {
 382                         index = start >> PAGE_CACHE_SHIFT;
 383                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 384                         page = find_get_page(btree_inode->i_mapping, index);
 385                         if (!page)
 386                                 continue;
 387                         if (PageDirty(page)) {
 388                                 btree_lock_page_hook(page);
 389                                 wait_on_page_writeback(page);
 390                                 err = write_one_page(page, 0);
 391                                 if (err)
 392                                         werr = err;
 393                         }
 394                         wait_on_page_writeback(page);
 395                         page_cache_release(page);
 396                         cond_resched();
 397                 }
 398         }
 399         if (err)
 400                 werr = err;
 401         return werr;
 402 }
 403
 404 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 405                                      struct btrfs_root *root)
 406 {
 407         if (!trans || !trans->transaction) {
 408                 struct inode *btree_inode;
 409                 btree_inode = root->fs_info->btree_inode;
 410                 return filemap_write_and_wait(btree_inode->i_mapping);
 411         }
 412         return btrfs_write_and_wait_marked_extents(root,
 413                                            &trans->transaction->dirty_pages);
 414 }
 415
 416 /*
 417  * this is used to update the root pointer in the tree of tree roots.
 418  *
 419  * But, in the case of the extent allocation tree, updating the root
 420  * pointer may allocate blocks which may change the root of the extent
 421  * allocation tree.
 422  *
 423  * So, this loops and repeats and makes sure the cowonly root didn't
 424  * change while the root pointer was being updated in the metadata.
 425  */
 426 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 427                                struct btrfs_root *root)
 428 {
 429         int ret;
 430         u64 old_root_bytenr;
 431         struct btrfs_root *tree_root = root->fs_info->tree_root;
 432
 433         btrfs_write_dirty_block_groups(trans, root);
 434         while(1) {
 435                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 436                 if (old_root_bytenr == root->node->start)
 437                         break;
 438                 btrfs_set_root_bytenr(&root->root_item,
 439                                        root->node->start);
 440                 btrfs_set_root_level(&root->root_item,
 441                                      btrfs_header_level(root->node));
 442                 btrfs_set_root_generation(&root->root_item, trans->transid);
 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         struct extent_buffer *eb;
 461
 462         eb = btrfs_lock_root_node(fs_info->tree_root);
 463         btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb, 0);
 464         btrfs_tree_unlock(eb);
 465         free_extent_buffer(eb);
 466
 467         while(!list_empty(&fs_info->dirty_cowonly_roots)) {
 468                 next = fs_info->dirty_cowonly_roots.next;
 469                 list_del_init(next);
 470                 root = list_entry(next, struct btrfs_root, dirty_list);
 471                 update_cowonly_root(trans, root);
 472         }
 473         return 0;
 474 }
 475
 476 /*
 477  * dead roots are old snapshots that need to be deleted.  This allocates
 478  * a dirty root struct and adds it into the list of dead roots that need to
 479  * be deleted
 480  */
 481 int btrfs_add_dead_root(struct btrfs_root *root, struct btrfs_root *latest)
 482 {
 483         struct btrfs_dirty_root *dirty;
 484
 485         dirty = kmalloc(sizeof(*dirty), GFP_NOFS);
 486         if (!dirty)
 487                 return -ENOMEM;
 488         dirty->root = root;
 489         dirty->latest_root = latest;
 490
 491         mutex_lock(&root->fs_info->trans_mutex);
 492         list_add(&dirty->list, &latest->fs_info->dead_roots);
 493         mutex_unlock(&root->fs_info->trans_mutex);
 494         return 0;
 495 }
 496
 497 /*
 498  * at transaction commit time we need to schedule the old roots for
 499  * deletion via btrfs_drop_snapshot.  This runs through all the
 500  * reference counted roots that were modified in the current
 501  * transaction and puts them into the drop list
 502  */
 503 static noinline int add_dirty_roots(struct btrfs_trans_handle *trans,
 504                                     struct radix_tree_root *radix,
 505                                     struct list_head *list)
 506 {
 507         struct btrfs_dirty_root *dirty;
 508         struct btrfs_root *gang[8];
 509         struct btrfs_root *root;
 510         int i;
 511         int ret;
 512         int err = 0;
 513         u32 refs;
 514
 515         while(1) {
 516                 ret = radix_tree_gang_lookup_tag(radix, (void **)gang, 0,
 517                                                  ARRAY_SIZE(gang),
 518                                                  BTRFS_ROOT_TRANS_TAG);
 519                 if (ret == 0)
 520                         break;
 521                 for (i = 0; i < ret; i++) {
 522                         root = gang[i];
 523                         radix_tree_tag_clear(radix,
 524                                      (unsigned long)root->root_key.objectid,
 525                                      BTRFS_ROOT_TRANS_TAG);
 526
 527                         BUG_ON(!root->ref_tree);
 528                         dirty = root->dirty_root;
 529
 530                         btrfs_free_log(trans, root);
 531                         btrfs_free_reloc_root(trans, root);
 532
 533                         if (root->commit_root == root->node) {
 534                                 WARN_ON(root->node->start !=
 535                                         btrfs_root_bytenr(&root->root_item));
 536
 537                                 free_extent_buffer(root->commit_root);
 538                                 root->commit_root = NULL;
 539                                 root->dirty_root = NULL;
 540
 541                                 spin_lock(&root->list_lock);
 542                                 list_del_init(&dirty->root->dead_list);
 543                                 spin_unlock(&root->list_lock);
 544
 545                                 kfree(dirty->root);
 546                                 kfree(dirty);
 547
 548                                 /* make sure to update the root on disk
 549                                  * so we get any updates to the block used
 550                                  * counts
 551                                  */
 552                                 err = btrfs_update_root(trans,
 553                                                 root->fs_info->tree_root,
 554                                                 &root->root_key,
 555                                                 &root->root_item);
 556                                 continue;
 557                         }
 558
 559                         memset(&root->root_item.drop_progress, 0,
 560                                sizeof(struct btrfs_disk_key));
 561                         root->root_item.drop_level = 0;
 562                         root->commit_root = NULL;
 563                         root->dirty_root = NULL;
 564                         root->root_key.offset = root->fs_info->generation;
 565                         btrfs_set_root_bytenr(&root->root_item,
 566                                               root->node->start);
 567                         btrfs_set_root_level(&root->root_item,
 568                                              btrfs_header_level(root->node));
 569                         btrfs_set_root_generation(&root->root_item,
 570                                                   root->root_key.offset);
 571
 572                         err = btrfs_insert_root(trans, root->fs_info->tree_root,
 573                                                 &root->root_key,
 574                                                 &root->root_item);
 575                         if (err)
 576                                 break;
 577
 578                         refs = btrfs_root_refs(&dirty->root->root_item);
 579                         btrfs_set_root_refs(&dirty->root->root_item, refs - 1);
 580                         err = btrfs_update_root(trans, root->fs_info->tree_root,
 581                                                 &dirty->root->root_key,
 582                                                 &dirty->root->root_item);
 583
 584                         BUG_ON(err);
 585                         if (refs == 1) {
 586                                 list_add(&dirty->list, list);
 587                         } else {
 588                                 WARN_ON(1);
 589                                 free_extent_buffer(dirty->root->node);
 590                                 kfree(dirty->root);
 591                                 kfree(dirty);
 592                         }
 593                 }
 594         }
 595         return err;
 596 }
 597
 598 /*
 599  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 600  * otherwise every leaf in the btree is read and defragged.
 601  */
 602 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 603 {
 604         struct btrfs_fs_info *info = root->fs_info;
 605         int ret;
 606         struct btrfs_trans_handle *trans;
 607         unsigned long nr;
 608
 609         smp_mb();
 610         if (root->defrag_running)
 611                 return 0;
 612         trans = btrfs_start_transaction(root, 1);
 613         while (1) {
 614                 root->defrag_running = 1;
 615                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 616                 nr = trans->blocks_used;
 617                 btrfs_end_transaction(trans, root);
 618                 btrfs_btree_balance_dirty(info->tree_root, nr);
 619                 cond_resched();
 620
 621                 trans = btrfs_start_transaction(root, 1);
 622                 if (root->fs_info->closing || ret != -EAGAIN)
 623                         break;
 624         }
 625         root->defrag_running = 0;
 626         smp_mb();
 627         btrfs_end_transaction(trans, root);
 628         return 0;
 629 }
 630
 631 /*
 632  * Given a list of roots that need to be deleted, call btrfs_drop_snapshot on
 633  * all of them
 634  */
 635 static noinline int drop_dirty_roots(struct btrfs_root *tree_root,
 636                                      struct list_head *list)
 637 {
 638         struct btrfs_dirty_root *dirty;
 639         struct btrfs_trans_handle *trans;
 640         unsigned long nr;
 641         u64 num_bytes;
 642         u64 bytes_used;
 643         u64 max_useless;
 644         int ret = 0;
 645         int err;
 646
 647         while(!list_empty(list)) {
 648                 struct btrfs_root *root;
 649
 650                 dirty = list_entry(list->prev, struct btrfs_dirty_root, list);
 651                 list_del_init(&dirty->list);
 652
 653                 num_bytes = btrfs_root_used(&dirty->root->root_item);
 654                 root = dirty->latest_root;
 655                 atomic_inc(&root->fs_info->throttles);
 656
 657                 while(1) {
 658                         trans = btrfs_start_transaction(tree_root, 1);
 659                         mutex_lock(&root->fs_info->drop_mutex);
 660                         ret = btrfs_drop_snapshot(trans, dirty->root);
 661                         if (ret != -EAGAIN) {
 662                                 break;
 663                         }
 664                         mutex_unlock(&root->fs_info->drop_mutex);
 665
 666                         err = btrfs_update_root(trans,
 667                                         tree_root,
 668                                         &dirty->root->root_key,
 669                                         &dirty->root->root_item);
 670                         if (err)
 671                                 ret = err;
 672                         nr = trans->blocks_used;
 673                         ret = btrfs_end_transaction(trans, tree_root);
 674                         BUG_ON(ret);
 675
 676                         btrfs_btree_balance_dirty(tree_root, nr);
 677                         cond_resched();
 678                 }
 679                 BUG_ON(ret);
 680                 atomic_dec(&root->fs_info->throttles);
 681                 wake_up(&root->fs_info->transaction_throttle);
 682
 683                 num_bytes -= btrfs_root_used(&dirty->root->root_item);
 684                 bytes_used = btrfs_root_used(&root->root_item);
 685                 if (num_bytes) {
 686                         btrfs_record_root_in_trans(root);
 687                         btrfs_set_root_used(&root->root_item,
 688                                             bytes_used - num_bytes);
 689                 }
 690
 691                 ret = btrfs_del_root(trans, tree_root, &dirty->root->root_key);
 692                 if (ret) {
 693                         BUG();
 694                         break;
 695                 }
 696                 mutex_unlock(&root->fs_info->drop_mutex);
 697
 698                 spin_lock(&root->list_lock);
 699                 list_del_init(&dirty->root->dead_list);
 700                 if (!list_empty(&root->dead_list)) {
 701                         struct btrfs_root *oldest;
 702                         oldest = list_entry(root->dead_list.prev,
 703                                             struct btrfs_root, dead_list);
 704                         max_useless = oldest->root_key.offset - 1;
 705                 } else {
 706                         max_useless = root->root_key.offset - 1;
 707                 }
 708                 spin_unlock(&root->list_lock);
 709
 710                 nr = trans->blocks_used;
 711                 ret = btrfs_end_transaction(trans, tree_root);
 712                 BUG_ON(ret);
 713
 714                 ret = btrfs_remove_leaf_refs(root, max_useless, 0);
 715                 BUG_ON(ret);
 716
 717                 free_extent_buffer(dirty->root->node);
 718                 kfree(dirty->root);
 719                 kfree(dirty);
 720
 721                 btrfs_btree_balance_dirty(tree_root, nr);
 722                 cond_resched();
 723         }
 724         return ret;
 725 }
 726
 727 /*
 728  * new snapshots need to be created at a very specific time in the
 729  * transaction commit.  This does the actual creation
 730  */
 731 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 732                                    struct btrfs_fs_info *fs_info,
 733                                    struct btrfs_pending_snapshot *pending)
 734 {
 735         struct btrfs_key key;
 736         struct btrfs_root_item *new_root_item;
 737         struct btrfs_root *tree_root = fs_info->tree_root;
 738         struct btrfs_root *root = pending->root;
 739         struct extent_buffer *tmp;
 740         struct extent_buffer *old;
 741         int ret;
 742         int namelen;
 743         u64 objectid;
 744
 745         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 746         if (!new_root_item) {
 747                 ret = -ENOMEM;
 748                 goto fail;
 749         }
 750         ret = btrfs_find_free_objectid(trans, tree_root, 0, &objectid);
 751         if (ret)
 752                 goto fail;
 753
 754         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
 755
 756         key.objectid = objectid;
 757         key.offset = trans->transid;
 758         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
 759
 760         old = btrfs_lock_root_node(root);
 761         btrfs_cow_block(trans, root, old, NULL, 0, &old, 0);
 762
 763         btrfs_copy_root(trans, root, old, &tmp, objectid);
 764         btrfs_tree_unlock(old);
 765         free_extent_buffer(old);
 766
 767         btrfs_set_root_bytenr(new_root_item, tmp->start);
 768         btrfs_set_root_level(new_root_item, btrfs_header_level(tmp));
 769         btrfs_set_root_generation(new_root_item, trans->transid);
 770         ret = btrfs_insert_root(trans, root->fs_info->tree_root, &key,
 771                                 new_root_item);
 772         btrfs_tree_unlock(tmp);
 773         free_extent_buffer(tmp);
 774         if (ret)
 775                 goto fail;
 776
 777         /*
 778          * insert the directory item
 779          */
 780         key.offset = (u64)-1;
 781         namelen = strlen(pending->name);
 782         ret = btrfs_insert_dir_item(trans, root->fs_info->tree_root,
 783                                     pending->name, namelen,
 784                                     root->fs_info->sb->s_root->d_inode->i_ino,
 785                                     &key, BTRFS_FT_DIR, 0);
 786
 787         if (ret)
 788                 goto fail;
 789
 790         ret = btrfs_insert_inode_ref(trans, root->fs_info->tree_root,
 791                              pending->name, strlen(pending->name), objectid,
 792                              root->fs_info->sb->s_root->d_inode->i_ino, 0);
 793
 794         /* Invalidate existing dcache entry for new snapshot. */
 795         btrfs_invalidate_dcache_root(root, pending->name, namelen);
 796
 797 fail:
 798         kfree(new_root_item);
 799         return ret;
 800 }
 801
 802 /*
 803  * create all the snapshots we've scheduled for creation
 804  */
 805 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
 806                                              struct btrfs_fs_info *fs_info)
 807 {
 808         struct btrfs_pending_snapshot *pending;
 809         struct list_head *head = &trans->transaction->pending_snapshots;
 810         int ret;
 811
 812         while(!list_empty(head)) {
 813                 pending = list_entry(head->next,
 814                                      struct btrfs_pending_snapshot, list);
 815                 ret = create_pending_snapshot(trans, fs_info, pending);
 816                 BUG_ON(ret);
 817                 list_del(&pending->list);
 818                 kfree(pending->name);
 819                 kfree(pending);
 820         }
 821         return 0;
 822 }
 823
 824 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
 825                              struct btrfs_root *root)
 826 {
 827         unsigned long joined = 0;
 828         unsigned long timeout = 1;
 829         struct btrfs_transaction *cur_trans;
 830         struct btrfs_transaction *prev_trans = NULL;
 831         struct btrfs_root *chunk_root = root->fs_info->chunk_root;
 832         struct list_head dirty_fs_roots;
 833         struct extent_io_tree *pinned_copy;
 834         DEFINE_WAIT(wait);
 835         int ret;
 836
 837         INIT_LIST_HEAD(&dirty_fs_roots);
 838         mutex_lock(&root->fs_info->trans_mutex);
 839         if (trans->transaction->in_commit) {
 840                 cur_trans = trans->transaction;
 841                 trans->transaction->use_count++;
 842                 mutex_unlock(&root->fs_info->trans_mutex);
 843                 btrfs_end_transaction(trans, root);
 844
 845                 ret = wait_for_commit(root, cur_trans);
 846                 BUG_ON(ret);
 847
 848                 mutex_lock(&root->fs_info->trans_mutex);
 849                 put_transaction(cur_trans);
 850                 mutex_unlock(&root->fs_info->trans_mutex);
 851
 852                 return 0;
 853         }
 854
 855         pinned_copy = kmalloc(sizeof(*pinned_copy), GFP_NOFS);
 856         if (!pinned_copy)
 857                 return -ENOMEM;
 858
 859         extent_io_tree_init(pinned_copy,
 860                              root->fs_info->btree_inode->i_mapping, GFP_NOFS);
 861
 862         trans->transaction->in_commit = 1;
 863         trans->transaction->blocked = 1;
 864         cur_trans = trans->transaction;
 865         if (cur_trans->list.prev != &root->fs_info->trans_list) {
 866                 prev_trans = list_entry(cur_trans->list.prev,
 867                                         struct btrfs_transaction, list);
 868                 if (!prev_trans->commit_done) {
 869                         prev_trans->use_count++;
 870                         mutex_unlock(&root->fs_info->trans_mutex);
 871
 872                         wait_for_commit(root, prev_trans);
 873
 874                         mutex_lock(&root->fs_info->trans_mutex);
 875                         put_transaction(prev_trans);
 876                 }
 877         }
 878
 879         do {
 880                 int snap_pending = 0;
 881                 joined = cur_trans->num_joined;
 882                 if (!list_empty(&trans->transaction->pending_snapshots))
 883                         snap_pending = 1;
 884
 885                 WARN_ON(cur_trans != trans->transaction);
 886                 prepare_to_wait(&cur_trans->writer_wait, &wait,
 887                                 TASK_UNINTERRUPTIBLE);
 888
 889                 if (cur_trans->num_writers > 1)
 890                         timeout = MAX_SCHEDULE_TIMEOUT;
 891                 else
 892                         timeout = 1;
 893
 894                 mutex_unlock(&root->fs_info->trans_mutex);
 895
 896                 if (snap_pending) {
 897                         ret = btrfs_wait_ordered_extents(root, 1);
 898                         BUG_ON(ret);
 899                 }
 900
 901                 schedule_timeout(timeout);
 902
 903                 mutex_lock(&root->fs_info->trans_mutex);
 904                 finish_wait(&cur_trans->writer_wait, &wait);
 905         } while (cur_trans->num_writers > 1 ||
 906                  (cur_trans->num_joined != joined));
 907
 908         ret = create_pending_snapshots(trans, root->fs_info);
 909         BUG_ON(ret);
 910
 911         WARN_ON(cur_trans != trans->transaction);
 912
 913         /* btrfs_commit_tree_roots is responsible for getting the
 914          * various roots consistent with each other.  Every pointer
 915          * in the tree of tree roots has to point to the most up to date
 916          * root for every subvolume and other tree.  So, we have to keep
 917          * the tree logging code from jumping in and changing any
 918          * of the trees.
 919          *
 920          * At this point in the commit, there can't be any tree-log
 921          * writers, but a little lower down we drop the trans mutex
 922          * and let new people in.  By holding the tree_log_mutex
 923          * from now until after the super is written, we avoid races
 924          * with the tree-log code.
 925          */
 926         mutex_lock(&root->fs_info->tree_log_mutex);
 927         /*
 928          * keep tree reloc code from adding new reloc trees
 929          */
 930         mutex_lock(&root->fs_info->tree_reloc_mutex);
 931
 932
 933         ret = add_dirty_roots(trans, &root->fs_info->fs_roots_radix,
 934                               &dirty_fs_roots);
 935         BUG_ON(ret);
 936
 937         /* add_dirty_roots gets rid of all the tree log roots, it is now
 938          * safe to free the root of tree log roots
 939          */
 940         btrfs_free_log_root_tree(trans, root->fs_info);
 941
 942         ret = btrfs_commit_tree_roots(trans, root);
 943         BUG_ON(ret);
 944
 945         cur_trans = root->fs_info->running_transaction;
 946         spin_lock(&root->fs_info->new_trans_lock);
 947         root->fs_info->running_transaction = NULL;
 948         spin_unlock(&root->fs_info->new_trans_lock);
 949         btrfs_set_super_generation(&root->fs_info->super_copy,
 950                                    cur_trans->transid);
 951         btrfs_set_super_root(&root->fs_info->super_copy,
 952                              root->fs_info->tree_root->node->start);
 953         btrfs_set_super_root_level(&root->fs_info->super_copy,
 954                            btrfs_header_level(root->fs_info->tree_root->node));
 955
 956         btrfs_set_super_chunk_root(&root->fs_info->super_copy,
 957                                    chunk_root->node->start);
 958         btrfs_set_super_chunk_root_level(&root->fs_info->super_copy,
 959                                          btrfs_header_level(chunk_root->node));
 960         btrfs_set_super_chunk_root_generation(&root->fs_info->super_copy,
 961                                 btrfs_header_generation(chunk_root->node));
 962
 963         if (!root->fs_info->log_root_recovering) {
 964                 btrfs_set_super_log_root(&root->fs_info->super_copy, 0);
 965                 btrfs_set_super_log_root_level(&root->fs_info->super_copy, 0);
 966         }
 967
 968         memcpy(&root->fs_info->super_for_commit, &root->fs_info->super_copy,
 969                sizeof(root->fs_info->super_copy));
 970
 971         btrfs_copy_pinned(root, pinned_copy);
 972
 973         trans->transaction->blocked = 0;
 974         wake_up(&root->fs_info->transaction_throttle);
 975         wake_up(&root->fs_info->transaction_wait);
 976
 977         mutex_unlock(&root->fs_info->trans_mutex);
 978         ret = btrfs_write_and_wait_transaction(trans, root);
 979         BUG_ON(ret);
 980         write_ctree_super(trans, root);
 981
 982         /*
 983          * the super is written, we can safely allow the tree-loggers
 984          * to go about their business
 985          */
 986         mutex_unlock(&root->fs_info->tree_log_mutex);
 987
 988         btrfs_finish_extent_commit(trans, root, pinned_copy);
 989         kfree(pinned_copy);
 990
 991         btrfs_drop_dead_reloc_roots(root);
 992         mutex_unlock(&root->fs_info->tree_reloc_mutex);
 993
 994         mutex_lock(&root->fs_info->trans_mutex);
 995
 996         cur_trans->commit_done = 1;
 997         root->fs_info->last_trans_committed = cur_trans->transid;
 998         wake_up(&cur_trans->commit_wait);
 999         put_transaction(cur_trans);
1000         put_transaction(cur_trans);
1001
1002         list_splice_init(&dirty_fs_roots, &root->fs_info->dead_roots);
1003         if (root->fs_info->closing)
1004                 list_splice_init(&root->fs_info->dead_roots, &dirty_fs_roots);
1005
1006         mutex_unlock(&root->fs_info->trans_mutex);
1007         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1008
1009         if (root->fs_info->closing) {
1010                 drop_dirty_roots(root->fs_info->tree_root, &dirty_fs_roots);
1011         }
1012         return ret;
1013 }
1014
1015 /*
1016  * interface function to delete all the snapshots we have scheduled for deletion
1017  */
1018 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1019 {
1020         struct list_head dirty_roots;
1021         INIT_LIST_HEAD(&dirty_roots);
1022 again:
1023         mutex_lock(&root->fs_info->trans_mutex);
1024         list_splice_init(&root->fs_info->dead_roots, &dirty_roots);
1025         mutex_unlock(&root->fs_info->trans_mutex);
1026
1027         if (!list_empty(&dirty_roots)) {
1028                 drop_dirty_roots(root, &dirty_roots);
1029                 goto again;
1030         }
1031         return 0;
1032 }