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