git.oblomov.eu Git - linux-2.6/blob - 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 <linux/blkdev.h>
  24 #include "ctree.h"
  25 #include "disk-io.h"
  26 #include "transaction.h"
  27 #include "locking.h"
  28 #include "ref-cache.h"
  29 #include "tree-log.h"
  30
  31 #define BTRFS_ROOT_TRANS_TAG 0
  32
  33 static noinline void put_transaction(struct btrfs_transaction *transaction)
  34 {
  35         WARN_ON(transaction->use_count == 0);
  36         transaction->use_count--;
  37         if (transaction->use_count == 0) {
  38                 list_del_init(&transaction->list);
  39                 memset(transaction, 0, sizeof(*transaction));
  40                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  41         }
  42 }
  43
  44 /*
  45  * either allocate a new transaction or hop into the existing one
  46  */
  47 static noinline int join_transaction(struct btrfs_root *root)
  48 {
  49         struct btrfs_transaction *cur_trans;
  50         cur_trans = root->fs_info->running_transaction;
  51         if (!cur_trans) {
  52                 cur_trans = kmem_cache_alloc(btrfs_transaction_cachep,
  53                                              GFP_NOFS);
  54                 BUG_ON(!cur_trans);
  55                 root->fs_info->generation++;
  56                 cur_trans->num_writers = 1;
  57                 cur_trans->num_joined = 0;
  58                 cur_trans->transid = root->fs_info->generation;
  59                 init_waitqueue_head(&cur_trans->writer_wait);
  60                 init_waitqueue_head(&cur_trans->commit_wait);
  61                 cur_trans->in_commit = 0;
  62                 cur_trans->blocked = 0;
  63                 cur_trans->use_count = 1;
  64                 cur_trans->commit_done = 0;
  65                 cur_trans->start_time = get_seconds();
  66
  67                 cur_trans->delayed_refs.root.rb_node = NULL;
  68                 cur_trans->delayed_refs.num_entries = 0;
  69                 cur_trans->delayed_refs.num_heads_ready = 0;
  70                 cur_trans->delayed_refs.num_heads = 0;
  71                 cur_trans->delayed_refs.flushing = 0;
  72                 cur_trans->delayed_refs.run_delayed_start = 0;
  73                 spin_lock_init(&cur_trans->delayed_refs.lock);
  74
  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 reference
  93  * counted root is properly recorded in a given transaction.  This is required
  94  * to make sure the old root from before we joined the transaction is deleted
  95  * 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 = 0;
 190         h->alloc_exclude_nr = 0;
 191         h->alloc_exclude_start = 0;
 192         h->delayed_ref_updates = 0;
 193
 194         root->fs_info->running_transaction->use_count++;
 195         mutex_unlock(&root->fs_info->trans_mutex);
 196         return h;
 197 }
 198
 199 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 200                                                    int num_blocks)
 201 {
 202         return start_transaction(root, num_blocks, 1);
 203 }
 204 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root,
 205                                                    int num_blocks)
 206 {
 207         return start_transaction(root, num_blocks, 0);
 208 }
 209
 210 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *r,
 211                                                          int num_blocks)
 212 {
 213         return start_transaction(r, num_blocks, 2);
 214 }
 215
 216 /* wait for a transaction commit to be fully complete */
 217 static noinline int wait_for_commit(struct btrfs_root *root,
 218                                     struct btrfs_transaction *commit)
 219 {
 220         DEFINE_WAIT(wait);
 221         mutex_lock(&root->fs_info->trans_mutex);
 222         while (!commit->commit_done) {
 223                 prepare_to_wait(&commit->commit_wait, &wait,
 224                                 TASK_UNINTERRUPTIBLE);
 225                 if (commit->commit_done)
 226                         break;
 227                 mutex_unlock(&root->fs_info->trans_mutex);
 228                 schedule();
 229                 mutex_lock(&root->fs_info->trans_mutex);
 230         }
 231         mutex_unlock(&root->fs_info->trans_mutex);
 232         finish_wait(&commit->commit_wait, &wait);
 233         return 0;
 234 }
 235
 236 /*
 237  * rate limit against the drop_snapshot code.  This helps to slow down new
 238  * operations if the drop_snapshot code isn't able to keep up.
 239  */
 240 static void throttle_on_drops(struct btrfs_root *root)
 241 {
 242         struct btrfs_fs_info *info = root->fs_info;
 243         int harder_count = 0;
 244
 245 harder:
 246         if (atomic_read(&info->throttles)) {
 247                 DEFINE_WAIT(wait);
 248                 int thr;
 249                 thr = atomic_read(&info->throttle_gen);
 250
 251                 do {
 252                         prepare_to_wait(&info->transaction_throttle,
 253                                         &wait, TASK_UNINTERRUPTIBLE);
 254                         if (!atomic_read(&info->throttles)) {
 255                                 finish_wait(&info->transaction_throttle, &wait);
 256                                 break;
 257                         }
 258                         schedule();
 259                         finish_wait(&info->transaction_throttle, &wait);
 260                 } while (thr == atomic_read(&info->throttle_gen));
 261                 harder_count++;
 262
 263                 if (root->fs_info->total_ref_cache_size > 1 * 1024 * 1024 &&
 264                     harder_count < 2)
 265                         goto harder;
 266
 267                 if (root->fs_info->total_ref_cache_size > 5 * 1024 * 1024 &&
 268                     harder_count < 10)
 269                         goto harder;
 270
 271                 if (root->fs_info->total_ref_cache_size > 10 * 1024 * 1024 &&
 272                     harder_count < 20)
 273                         goto harder;
 274         }
 275 }
 276
 277 void btrfs_throttle(struct btrfs_root *root)
 278 {
 279         mutex_lock(&root->fs_info->trans_mutex);
 280         if (!root->fs_info->open_ioctl_trans)
 281                 wait_current_trans(root);
 282         mutex_unlock(&root->fs_info->trans_mutex);
 283         throttle_on_drops(root);
 284 }
 285
 286 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 287                           struct btrfs_root *root, int throttle)
 288 {
 289         struct btrfs_transaction *cur_trans;
 290         struct btrfs_fs_info *info = root->fs_info;
 291         int count = 0;
 292
 293         while (count < 4) {
 294                 unsigned long cur = trans->delayed_ref_updates;
 295                 trans->delayed_ref_updates = 0;
 296                 if (cur &&
 297                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 298                         trans->delayed_ref_updates = 0;
 299
 300                         /*
 301                          * do a full flush if the transaction is trying
 302                          * to close
 303                          */
 304                         if (trans->transaction->delayed_refs.flushing)
 305                                 cur = 0;
 306                         btrfs_run_delayed_refs(trans, root, cur);
 307                 } else {
 308                         break;
 309                 }
 310                 count++;
 311         }
 312
 313         mutex_lock(&info->trans_mutex);
 314         cur_trans = info->running_transaction;
 315         WARN_ON(cur_trans != trans->transaction);
 316         WARN_ON(cur_trans->num_writers < 1);
 317         cur_trans->num_writers--;
 318
 319         if (waitqueue_active(&cur_trans->writer_wait))
 320                 wake_up(&cur_trans->writer_wait);
 321         put_transaction(cur_trans);
 322         mutex_unlock(&info->trans_mutex);
 323         memset(trans, 0, sizeof(*trans));
 324         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 325
 326         if (throttle)
 327                 throttle_on_drops(root);
 328
 329         return 0;
 330 }
 331
 332 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 333                           struct btrfs_root *root)
 334 {
 335         return __btrfs_end_transaction(trans, root, 0);
 336 }
 337
 338 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 339                                    struct btrfs_root *root)
 340 {
 341         return __btrfs_end_transaction(trans, root, 1);
 342 }
 343
 344 /*
 345  * when btree blocks are allocated, they have some corresponding bits set for
 346  * them in one of two extent_io trees.  This is used to make sure all of
 347  * those extents are on disk for transaction or log commit
 348  */
 349 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 350                                         struct extent_io_tree *dirty_pages)
 351 {
 352         int ret;
 353         int err = 0;
 354         int werr = 0;
 355         struct page *page;
 356         struct inode *btree_inode = root->fs_info->btree_inode;
 357         u64 start = 0;
 358         u64 end;
 359         unsigned long index;
 360
 361         while (1) {
 362                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 363                                             EXTENT_DIRTY);
 364                 if (ret)
 365                         break;
 366                 while (start <= end) {
 367                         cond_resched();
 368
 369                         index = start >> PAGE_CACHE_SHIFT;
 370                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 371                         page = find_get_page(btree_inode->i_mapping, index);
 372                         if (!page)
 373                                 continue;
 374
 375                         btree_lock_page_hook(page);
 376                         if (!page->mapping) {
 377                                 unlock_page(page);
 378                                 page_cache_release(page);
 379                                 continue;
 380                         }
 381
 382                         if (PageWriteback(page)) {
 383                                 if (PageDirty(page))
 384                                         wait_on_page_writeback(page);
 385                                 else {
 386                                         unlock_page(page);
 387                                         page_cache_release(page);
 388                                         continue;
 389                                 }
 390                         }
 391                         err = write_one_page(page, 0);
 392                         if (err)
 393                                 werr = err;
 394                         page_cache_release(page);
 395                 }
 396         }
 397         while (1) {
 398                 ret = find_first_extent_bit(dirty_pages, 0, &start, &end,
 399                                             EXTENT_DIRTY);
 400                 if (ret)
 401                         break;
 402
 403                 clear_extent_dirty(dirty_pages, start, end, GFP_NOFS);
 404                 while (start <= end) {
 405                         index = start >> PAGE_CACHE_SHIFT;
 406                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 407                         page = find_get_page(btree_inode->i_mapping, index);
 408                         if (!page)
 409                                 continue;
 410                         if (PageDirty(page)) {
 411                                 btree_lock_page_hook(page);
 412                                 wait_on_page_writeback(page);
 413                                 err = write_one_page(page, 0);
 414                                 if (err)
 415                                         werr = err;
 416                         }
 417                         wait_on_page_writeback(page);
 418                         page_cache_release(page);
 419                         cond_resched();
 420                 }
 421         }
 422         if (err)
 423                 werr = err;
 424         return werr;
 425 }
 426
 427 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 428                                      struct btrfs_root *root)
 429 {
 430         if (!trans || !trans->transaction) {
 431                 struct inode *btree_inode;
 432                 btree_inode = root->fs_info->btree_inode;
 433                 return filemap_write_and_wait(btree_inode->i_mapping);
 434         }
 435         return btrfs_write_and_wait_marked_extents(root,
 436                                            &trans->transaction->dirty_pages);
 437 }
 438
 439 /*
 440  * this is used to update the root pointer in the tree of tree roots.
 441  *
 442  * But, in the case of the extent allocation tree, updating the root
 443  * pointer may allocate blocks which may change the root of the extent
 444  * allocation tree.
 445  *
 446  * So, this loops and repeats and makes sure the cowonly root didn't
 447  * change while the root pointer was being updated in the metadata.
 448  */
 449 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 450                                struct btrfs_root *root)
 451 {
 452         int ret;
 453         u64 old_root_bytenr;
 454         struct btrfs_root *tree_root = root->fs_info->tree_root;
 455
 456         btrfs_write_dirty_block_groups(trans, root);
 457
 458         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 459         BUG_ON(ret);
 460
 461         while (1) {
 462                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 463                 if (old_root_bytenr == root->node->start)
 464                         break;
 465                 btrfs_set_root_bytenr(&root->root_item,
 466                                        root->node->start);
 467                 btrfs_set_root_level(&root->root_item,
 468                                      btrfs_header_level(root->node));
 469                 btrfs_set_root_generation(&root->root_item, trans->transid);
 470
 471                 ret = btrfs_update_root(trans, tree_root,
 472                                         &root->root_key,
 473                                         &root->root_item);
 474                 BUG_ON(ret);
 475                 btrfs_write_dirty_block_groups(trans, root);
 476
 477                 ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 478                 BUG_ON(ret);
 479         }
 480         return 0;
 481 }
 482
 483 /*
 484  * update all the cowonly tree roots on disk
 485  */
 486 int btrfs_commit_tree_roots(struct btrfs_trans_handle *trans,
 487                             struct btrfs_root *root)
 488 {
 489         struct btrfs_fs_info *fs_info = root->fs_info;
 490         struct list_head *next;
 491         struct extent_buffer *eb;
 492         int ret;
 493
 494         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 495         BUG_ON(ret);
 496
 497         eb = btrfs_lock_root_node(fs_info->tree_root);
 498         btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
 499         btrfs_tree_unlock(eb);
 500         free_extent_buffer(eb);
 501
 502         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 503         BUG_ON(ret);
 504
 505         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 506                 next = fs_info->dirty_cowonly_roots.next;
 507                 list_del_init(next);
 508                 root = list_entry(next, struct btrfs_root, dirty_list);
 509
 510                 update_cowonly_root(trans, root);
 511
 512                 ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 513                 BUG_ON(ret);
 514         }
 515         return 0;
 516 }
 517
 518 /*
 519  * dead roots are old snapshots that need to be deleted.  This allocates
 520  * a dirty root struct and adds it into the list of dead roots that need to
 521  * be deleted
 522  */
 523 int btrfs_add_dead_root(struct btrfs_root *root, struct btrfs_root *latest)
 524 {
 525         struct btrfs_dirty_root *dirty;
 526
 527         dirty = kmalloc(sizeof(*dirty), GFP_NOFS);
 528         if (!dirty)
 529                 return -ENOMEM;
 530         dirty->root = root;
 531         dirty->latest_root = latest;
 532
 533         mutex_lock(&root->fs_info->trans_mutex);
 534         list_add(&dirty->list, &latest->fs_info->dead_roots);
 535         mutex_unlock(&root->fs_info->trans_mutex);
 536         return 0;
 537 }
 538
 539 /*
 540  * at transaction commit time we need to schedule the old roots for
 541  * deletion via btrfs_drop_snapshot.  This runs through all the
 542  * reference counted roots that were modified in the current
 543  * transaction and puts them into the drop list
 544  */
 545 static noinline int add_dirty_roots(struct btrfs_trans_handle *trans,
 546                                     struct radix_tree_root *radix,
 547                                     struct list_head *list)
 548 {
 549         struct btrfs_dirty_root *dirty;
 550         struct btrfs_root *gang[8];
 551         struct btrfs_root *root;
 552         int i;
 553         int ret;
 554         int err = 0;
 555         u32 refs;
 556
 557         while (1) {
 558                 ret = radix_tree_gang_lookup_tag(radix, (void **)gang, 0,
 559                                                  ARRAY_SIZE(gang),
 560                                                  BTRFS_ROOT_TRANS_TAG);
 561                 if (ret == 0)
 562                         break;
 563                 for (i = 0; i < ret; i++) {
 564                         root = gang[i];
 565                         radix_tree_tag_clear(radix,
 566                                      (unsigned long)root->root_key.objectid,
 567                                      BTRFS_ROOT_TRANS_TAG);
 568
 569                         BUG_ON(!root->ref_tree);
 570                         dirty = root->dirty_root;
 571
 572                         btrfs_free_log(trans, root);
 573                         btrfs_free_reloc_root(trans, root);
 574
 575                         if (root->commit_root == root->node) {
 576                                 WARN_ON(root->node->start !=
 577                                         btrfs_root_bytenr(&root->root_item));
 578
 579                                 free_extent_buffer(root->commit_root);
 580                                 root->commit_root = NULL;
 581                                 root->dirty_root = NULL;
 582
 583                                 spin_lock(&root->list_lock);
 584                                 list_del_init(&dirty->root->dead_list);
 585                                 spin_unlock(&root->list_lock);
 586
 587                                 kfree(dirty->root);
 588                                 kfree(dirty);
 589
 590                                 /* make sure to update the root on disk
 591                                  * so we get any updates to the block used
 592                                  * counts
 593                                  */
 594                                 err = btrfs_update_root(trans,
 595                                                 root->fs_info->tree_root,
 596                                                 &root->root_key,
 597                                                 &root->root_item);
 598                                 continue;
 599                         }
 600
 601                         memset(&root->root_item.drop_progress, 0,
 602                                sizeof(struct btrfs_disk_key));
 603                         root->root_item.drop_level = 0;
 604                         root->commit_root = NULL;
 605                         root->dirty_root = NULL;
 606                         root->root_key.offset = root->fs_info->generation;
 607                         btrfs_set_root_bytenr(&root->root_item,
 608                                               root->node->start);
 609                         btrfs_set_root_level(&root->root_item,
 610                                              btrfs_header_level(root->node));
 611                         btrfs_set_root_generation(&root->root_item,
 612                                                   root->root_key.offset);
 613
 614                         err = btrfs_insert_root(trans, root->fs_info->tree_root,
 615                                                 &root->root_key,
 616                                                 &root->root_item);
 617                         if (err)
 618                                 break;
 619
 620                         refs = btrfs_root_refs(&dirty->root->root_item);
 621                         btrfs_set_root_refs(&dirty->root->root_item, refs - 1);
 622                         err = btrfs_update_root(trans, root->fs_info->tree_root,
 623                                                 &dirty->root->root_key,
 624                                                 &dirty->root->root_item);
 625
 626                         BUG_ON(err);
 627                         if (refs == 1) {
 628                                 list_add(&dirty->list, list);
 629                         } else {
 630                                 WARN_ON(1);
 631                                 free_extent_buffer(dirty->root->node);
 632                                 kfree(dirty->root);
 633                                 kfree(dirty);
 634                         }
 635                 }
 636         }
 637         return err;
 638 }
 639
 640 /*
 641  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 642  * otherwise every leaf in the btree is read and defragged.
 643  */
 644 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 645 {
 646         struct btrfs_fs_info *info = root->fs_info;
 647         int ret;
 648         struct btrfs_trans_handle *trans;
 649         unsigned long nr;
 650
 651         smp_mb();
 652         if (root->defrag_running)
 653                 return 0;
 654         trans = btrfs_start_transaction(root, 1);
 655         while (1) {
 656                 root->defrag_running = 1;
 657                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 658                 nr = trans->blocks_used;
 659                 btrfs_end_transaction(trans, root);
 660                 btrfs_btree_balance_dirty(info->tree_root, nr);
 661                 cond_resched();
 662
 663                 trans = btrfs_start_transaction(root, 1);
 664                 if (root->fs_info->closing || ret != -EAGAIN)
 665                         break;
 666         }
 667         root->defrag_running = 0;
 668         smp_mb();
 669         btrfs_end_transaction(trans, root);
 670         return 0;
 671 }
 672
 673 /*
 674  * when dropping snapshots, we generate a ton of delayed refs, and it makes
 675  * sense not to join the transaction while it is trying to flush the current
 676  * queue of delayed refs out.
 677  *
 678  * This is used by the drop snapshot code only
 679  */
 680 static noinline int wait_transaction_pre_flush(struct btrfs_fs_info *info)
 681 {
 682         DEFINE_WAIT(wait);
 683
 684         mutex_lock(&info->trans_mutex);
 685         while (info->running_transaction &&
 686                info->running_transaction->delayed_refs.flushing) {
 687                 prepare_to_wait(&info->transaction_wait, &wait,
 688                                 TASK_UNINTERRUPTIBLE);
 689                 mutex_unlock(&info->trans_mutex);
 690
 691                 atomic_dec(&info->throttles);
 692                 wake_up(&info->transaction_throttle);
 693
 694                 schedule();
 695
 696                 atomic_inc(&info->throttles);
 697                 mutex_lock(&info->trans_mutex);
 698                 finish_wait(&info->transaction_wait, &wait);
 699         }
 700         mutex_unlock(&info->trans_mutex);
 701         return 0;
 702 }
 703
 704 /*
 705  * Given a list of roots that need to be deleted, call btrfs_drop_snapshot on
 706  * all of them
 707  */
 708 static noinline int drop_dirty_roots(struct btrfs_root *tree_root,
 709                                      struct list_head *list)
 710 {
 711         struct btrfs_dirty_root *dirty;
 712         struct btrfs_trans_handle *trans;
 713         unsigned long nr;
 714         u64 num_bytes;
 715         u64 bytes_used;
 716         u64 max_useless;
 717         int ret = 0;
 718         int err;
 719
 720         while (!list_empty(list)) {
 721                 struct btrfs_root *root;
 722
 723                 dirty = list_entry(list->prev, struct btrfs_dirty_root, list);
 724                 list_del_init(&dirty->list);
 725
 726                 num_bytes = btrfs_root_used(&dirty->root->root_item);
 727                 root = dirty->latest_root;
 728                 atomic_inc(&root->fs_info->throttles);
 729
 730                 while (1) {
 731                         /*
 732                          * we don't want to jump in and create a bunch of
 733                          * delayed refs if the transaction is starting to close
 734                          */
 735                         wait_transaction_pre_flush(tree_root->fs_info);
 736                         trans = btrfs_start_transaction(tree_root, 1);
 737
 738                         /*
 739                          * we've joined a transaction, make sure it isn't
 740                          * closing right now
 741                          */
 742                         if (trans->transaction->delayed_refs.flushing) {
 743                                 btrfs_end_transaction(trans, tree_root);
 744                                 continue;
 745                         }
 746
 747                         mutex_lock(&root->fs_info->drop_mutex);
 748                         ret = btrfs_drop_snapshot(trans, dirty->root);
 749                         if (ret != -EAGAIN)
 750                                 break;
 751                         mutex_unlock(&root->fs_info->drop_mutex);
 752
 753                         err = btrfs_update_root(trans,
 754                                         tree_root,
 755                                         &dirty->root->root_key,
 756                                         &dirty->root->root_item);
 757                         if (err)
 758                                 ret = err;
 759                         nr = trans->blocks_used;
 760                         ret = btrfs_end_transaction(trans, tree_root);
 761                         BUG_ON(ret);
 762
 763                         btrfs_btree_balance_dirty(tree_root, nr);
 764                         cond_resched();
 765                 }
 766                 BUG_ON(ret);
 767                 atomic_dec(&root->fs_info->throttles);
 768                 wake_up(&root->fs_info->transaction_throttle);
 769
 770                 num_bytes -= btrfs_root_used(&dirty->root->root_item);
 771                 bytes_used = btrfs_root_used(&root->root_item);
 772                 if (num_bytes) {
 773                         mutex_lock(&root->fs_info->trans_mutex);
 774                         btrfs_record_root_in_trans(root);
 775                         mutex_unlock(&root->fs_info->trans_mutex);
 776                         btrfs_set_root_used(&root->root_item,
 777                                             bytes_used - num_bytes);
 778                 }
 779
 780                 ret = btrfs_del_root(trans, tree_root, &dirty->root->root_key);
 781                 if (ret) {
 782                         BUG();
 783                         break;
 784                 }
 785                 mutex_unlock(&root->fs_info->drop_mutex);
 786
 787                 spin_lock(&root->list_lock);
 788                 list_del_init(&dirty->root->dead_list);
 789                 if (!list_empty(&root->dead_list)) {
 790                         struct btrfs_root *oldest;
 791                         oldest = list_entry(root->dead_list.prev,
 792                                             struct btrfs_root, dead_list);
 793                         max_useless = oldest->root_key.offset - 1;
 794                 } else {
 795                         max_useless = root->root_key.offset - 1;
 796                 }
 797                 spin_unlock(&root->list_lock);
 798
 799                 nr = trans->blocks_used;
 800                 ret = btrfs_end_transaction(trans, tree_root);
 801                 BUG_ON(ret);
 802
 803                 ret = btrfs_remove_leaf_refs(root, max_useless, 0);
 804                 BUG_ON(ret);
 805
 806                 free_extent_buffer(dirty->root->node);
 807                 kfree(dirty->root);
 808                 kfree(dirty);
 809
 810                 btrfs_btree_balance_dirty(tree_root, nr);
 811                 cond_resched();
 812         }
 813         return ret;
 814 }
 815
 816 /*
 817  * new snapshots need to be created at a very specific time in the
 818  * transaction commit.  This does the actual creation
 819  */
 820 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 821                                    struct btrfs_fs_info *fs_info,
 822                                    struct btrfs_pending_snapshot *pending)
 823 {
 824         struct btrfs_key key;
 825         struct btrfs_root_item *new_root_item;
 826         struct btrfs_root *tree_root = fs_info->tree_root;
 827         struct btrfs_root *root = pending->root;
 828         struct extent_buffer *tmp;
 829         struct extent_buffer *old;
 830         int ret;
 831         u64 objectid;
 832
 833         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 834         if (!new_root_item) {
 835                 ret = -ENOMEM;
 836                 goto fail;
 837         }
 838         ret = btrfs_find_free_objectid(trans, tree_root, 0, &objectid);
 839         if (ret)
 840                 goto fail;
 841
 842         btrfs_record_root_in_trans(root);
 843         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
 844         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
 845
 846         key.objectid = objectid;
 847         key.offset = trans->transid;
 848         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
 849
 850         old = btrfs_lock_root_node(root);
 851         btrfs_cow_block(trans, root, old, NULL, 0, &old);
 852
 853         btrfs_copy_root(trans, root, old, &tmp, objectid);
 854         btrfs_tree_unlock(old);
 855         free_extent_buffer(old);
 856
 857         btrfs_set_root_bytenr(new_root_item, tmp->start);
 858         btrfs_set_root_level(new_root_item, btrfs_header_level(tmp));
 859         btrfs_set_root_generation(new_root_item, trans->transid);
 860         ret = btrfs_insert_root(trans, root->fs_info->tree_root, &key,
 861                                 new_root_item);
 862         btrfs_tree_unlock(tmp);
 863         free_extent_buffer(tmp);
 864         if (ret)
 865                 goto fail;
 866
 867         key.offset = (u64)-1;
 868         memcpy(&pending->root_key, &key, sizeof(key));
 869 fail:
 870         kfree(new_root_item);
 871         return ret;
 872 }
 873
 874 static noinline int finish_pending_snapshot(struct btrfs_fs_info *fs_info,
 875                                    struct btrfs_pending_snapshot *pending)
 876 {
 877         int ret;
 878         int namelen;
 879         u64 index = 0;
 880         struct btrfs_trans_handle *trans;
 881         struct inode *parent_inode;
 882         struct inode *inode;
 883         struct btrfs_root *parent_root;
 884
 885         parent_inode = pending->dentry->d_parent->d_inode;
 886         parent_root = BTRFS_I(parent_inode)->root;
 887         trans = btrfs_join_transaction(parent_root, 1);
 888
 889         /*
 890          * insert the directory item
 891          */
 892         namelen = strlen(pending->name);
 893         ret = btrfs_set_inode_index(parent_inode, &index);
 894         ret = btrfs_insert_dir_item(trans, parent_root,
 895                             pending->name, namelen,
 896                             parent_inode->i_ino,
 897                             &pending->root_key, BTRFS_FT_DIR, index);
 898
 899         if (ret)
 900                 goto fail;
 901
 902         btrfs_i_size_write(parent_inode, parent_inode->i_size + namelen * 2);
 903         ret = btrfs_update_inode(trans, parent_root, parent_inode);
 904         BUG_ON(ret);
 905
 906         /* add the backref first */
 907         ret = btrfs_add_root_ref(trans, parent_root->fs_info->tree_root,
 908                                  pending->root_key.objectid,
 909                                  BTRFS_ROOT_BACKREF_KEY,
 910                                  parent_root->root_key.objectid,
 911                                  parent_inode->i_ino, index, pending->name,
 912                                  namelen);
 913
 914         BUG_ON(ret);
 915
 916         /* now add the forward ref */
 917         ret = btrfs_add_root_ref(trans, parent_root->fs_info->tree_root,
 918                                  parent_root->root_key.objectid,
 919                                  BTRFS_ROOT_REF_KEY,
 920                                  pending->root_key.objectid,
 921                                  parent_inode->i_ino, index, pending->name,
 922                                  namelen);
 923
 924         inode = btrfs_lookup_dentry(parent_inode, pending->dentry);
 925         d_instantiate(pending->dentry, inode);
 926 fail:
 927         btrfs_end_transaction(trans, fs_info->fs_root);
 928         return ret;
 929 }
 930
 931 /*
 932  * create all the snapshots we've scheduled for creation
 933  */
 934 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
 935                                              struct btrfs_fs_info *fs_info)
 936 {
 937         struct btrfs_pending_snapshot *pending;
 938         struct list_head *head = &trans->transaction->pending_snapshots;
 939         int ret;
 940
 941         list_for_each_entry(pending, head, list) {
 942                 ret = create_pending_snapshot(trans, fs_info, pending);
 943                 BUG_ON(ret);
 944         }
 945         return 0;
 946 }
 947
 948 static noinline int finish_pending_snapshots(struct btrfs_trans_handle *trans,
 949                                              struct btrfs_fs_info *fs_info)
 950 {
 951         struct btrfs_pending_snapshot *pending;
 952         struct list_head *head = &trans->transaction->pending_snapshots;
 953         int ret;
 954
 955         while (!list_empty(head)) {
 956                 pending = list_entry(head->next,
 957                                      struct btrfs_pending_snapshot, list);
 958                 ret = finish_pending_snapshot(fs_info, pending);
 959                 BUG_ON(ret);
 960                 list_del(&pending->list);
 961                 kfree(pending->name);
 962                 kfree(pending);
 963         }
 964         return 0;
 965 }
 966
 967 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
 968                              struct btrfs_root *root)
 969 {
 970         unsigned long joined = 0;
 971         unsigned long timeout = 1;
 972         struct btrfs_transaction *cur_trans;
 973         struct btrfs_transaction *prev_trans = NULL;
 974         struct btrfs_root *chunk_root = root->fs_info->chunk_root;
 975         struct list_head dirty_fs_roots;
 976         struct extent_io_tree *pinned_copy;
 977         DEFINE_WAIT(wait);
 978         int ret;
 979         int should_grow = 0;
 980         unsigned long now = get_seconds();
 981         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
 982
 983         btrfs_run_ordered_operations(root, 0);
 984
 985         /* make a pass through all the delayed refs we have so far
 986          * any runnings procs may add more while we are here
 987          */
 988         ret = btrfs_run_delayed_refs(trans, root, 0);
 989         BUG_ON(ret);
 990
 991         cur_trans = trans->transaction;
 992         /*
 993          * set the flushing flag so procs in this transaction have to
 994          * start sending their work down.
 995          */
 996         cur_trans->delayed_refs.flushing = 1;
 997
 998         ret = btrfs_run_delayed_refs(trans, root, 0);
 999         BUG_ON(ret);
1000
1001         mutex_lock(&root->fs_info->trans_mutex);
1002         INIT_LIST_HEAD(&dirty_fs_roots);
1003         if (cur_trans->in_commit) {
1004                 cur_trans->use_count++;
1005                 mutex_unlock(&root->fs_info->trans_mutex);
1006                 btrfs_end_transaction(trans, root);
1007
1008                 ret = wait_for_commit(root, cur_trans);
1009                 BUG_ON(ret);
1010
1011                 mutex_lock(&root->fs_info->trans_mutex);
1012                 put_transaction(cur_trans);
1013                 mutex_unlock(&root->fs_info->trans_mutex);
1014
1015                 return 0;
1016         }
1017
1018         pinned_copy = kmalloc(sizeof(*pinned_copy), GFP_NOFS);
1019         if (!pinned_copy)
1020                 return -ENOMEM;
1021
1022         extent_io_tree_init(pinned_copy,
1023                              root->fs_info->btree_inode->i_mapping, GFP_NOFS);
1024
1025         trans->transaction->in_commit = 1;
1026         trans->transaction->blocked = 1;
1027         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1028                 prev_trans = list_entry(cur_trans->list.prev,
1029                                         struct btrfs_transaction, list);
1030                 if (!prev_trans->commit_done) {
1031                         prev_trans->use_count++;
1032                         mutex_unlock(&root->fs_info->trans_mutex);
1033
1034                         wait_for_commit(root, prev_trans);
1035
1036                         mutex_lock(&root->fs_info->trans_mutex);
1037                         put_transaction(prev_trans);
1038                 }
1039         }
1040
1041         if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
1042                 should_grow = 1;
1043
1044         do {
1045                 int snap_pending = 0;
1046                 joined = cur_trans->num_joined;
1047                 if (!list_empty(&trans->transaction->pending_snapshots))
1048                         snap_pending = 1;
1049
1050                 WARN_ON(cur_trans != trans->transaction);
1051                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1052                                 TASK_UNINTERRUPTIBLE);
1053
1054                 if (cur_trans->num_writers > 1)
1055                         timeout = MAX_SCHEDULE_TIMEOUT;
1056                 else if (should_grow)
1057                         timeout = 1;
1058
1059                 mutex_unlock(&root->fs_info->trans_mutex);
1060
1061                 if (flush_on_commit || snap_pending) {
1062                         if (flush_on_commit)
1063                                 btrfs_start_delalloc_inodes(root);
1064                         ret = btrfs_wait_ordered_extents(root, 1);
1065                         BUG_ON(ret);
1066                 }
1067
1068                 /*
1069                  * rename don't use btrfs_join_transaction, so, once we
1070                  * set the transaction to blocked above, we aren't going
1071                  * to get any new ordered operations.  We can safely run
1072                  * it here and no for sure that nothing new will be added
1073                  * to the list
1074                  */
1075                 btrfs_run_ordered_operations(root, 1);
1076
1077                 smp_mb();
1078                 if (cur_trans->num_writers > 1 || should_grow)
1079                         schedule_timeout(timeout);
1080
1081                 mutex_lock(&root->fs_info->trans_mutex);
1082                 finish_wait(&cur_trans->writer_wait, &wait);
1083         } while (cur_trans->num_writers > 1 ||
1084                  (should_grow && cur_trans->num_joined != joined));
1085
1086         ret = create_pending_snapshots(trans, root->fs_info);
1087         BUG_ON(ret);
1088
1089         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1090         BUG_ON(ret);
1091
1092         WARN_ON(cur_trans != trans->transaction);
1093
1094         /* btrfs_commit_tree_roots is responsible for getting the
1095          * various roots consistent with each other.  Every pointer
1096          * in the tree of tree roots has to point to the most up to date
1097          * root for every subvolume and other tree.  So, we have to keep
1098          * the tree logging code from jumping in and changing any
1099          * of the trees.
1100          *
1101          * At this point in the commit, there can't be any tree-log
1102          * writers, but a little lower down we drop the trans mutex
1103          * and let new people in.  By holding the tree_log_mutex
1104          * from now until after the super is written, we avoid races
1105          * with the tree-log code.
1106          */
1107         mutex_lock(&root->fs_info->tree_log_mutex);
1108         /*
1109          * keep tree reloc code from adding new reloc trees
1110          */
1111         mutex_lock(&root->fs_info->tree_reloc_mutex);
1112
1113
1114         ret = add_dirty_roots(trans, &root->fs_info->fs_roots_radix,
1115                               &dirty_fs_roots);
1116         BUG_ON(ret);
1117
1118         /* add_dirty_roots gets rid of all the tree log roots, it is now
1119          * safe to free the root of tree log roots
1120          */
1121         btrfs_free_log_root_tree(trans, root->fs_info);
1122
1123         ret = btrfs_commit_tree_roots(trans, root);
1124         BUG_ON(ret);
1125
1126         cur_trans = root->fs_info->running_transaction;
1127         spin_lock(&root->fs_info->new_trans_lock);
1128         root->fs_info->running_transaction = NULL;
1129         spin_unlock(&root->fs_info->new_trans_lock);
1130         btrfs_set_super_generation(&root->fs_info->super_copy,
1131                                    cur_trans->transid);
1132         btrfs_set_super_root(&root->fs_info->super_copy,
1133                              root->fs_info->tree_root->node->start);
1134         btrfs_set_super_root_level(&root->fs_info->super_copy,
1135                            btrfs_header_level(root->fs_info->tree_root->node));
1136
1137         btrfs_set_super_chunk_root(&root->fs_info->super_copy,
1138                                    chunk_root->node->start);
1139         btrfs_set_super_chunk_root_level(&root->fs_info->super_copy,
1140                                          btrfs_header_level(chunk_root->node));
1141         btrfs_set_super_chunk_root_generation(&root->fs_info->super_copy,
1142                                 btrfs_header_generation(chunk_root->node));
1143
1144         if (!root->fs_info->log_root_recovering) {
1145                 btrfs_set_super_log_root(&root->fs_info->super_copy, 0);
1146                 btrfs_set_super_log_root_level(&root->fs_info->super_copy, 0);
1147         }
1148
1149         memcpy(&root->fs_info->super_for_commit, &root->fs_info->super_copy,
1150                sizeof(root->fs_info->super_copy));
1151
1152         btrfs_copy_pinned(root, pinned_copy);
1153
1154         trans->transaction->blocked = 0;
1155
1156         wake_up(&root->fs_info->transaction_throttle);
1157         wake_up(&root->fs_info->transaction_wait);
1158
1159         mutex_unlock(&root->fs_info->trans_mutex);
1160         ret = btrfs_write_and_wait_transaction(trans, root);
1161         BUG_ON(ret);
1162         write_ctree_super(trans, root, 0);
1163
1164         /*
1165          * the super is written, we can safely allow the tree-loggers
1166          * to go about their business
1167          */
1168         mutex_unlock(&root->fs_info->tree_log_mutex);
1169
1170         btrfs_finish_extent_commit(trans, root, pinned_copy);
1171         kfree(pinned_copy);
1172
1173         btrfs_drop_dead_reloc_roots(root);
1174         mutex_unlock(&root->fs_info->tree_reloc_mutex);
1175
1176         /* do the directory inserts of any pending snapshot creations */
1177         finish_pending_snapshots(trans, root->fs_info);
1178
1179         mutex_lock(&root->fs_info->trans_mutex);
1180
1181         cur_trans->commit_done = 1;
1182
1183         root->fs_info->last_trans_committed = cur_trans->transid;
1184         wake_up(&cur_trans->commit_wait);
1185
1186         put_transaction(cur_trans);
1187         put_transaction(cur_trans);
1188
1189         list_splice_init(&dirty_fs_roots, &root->fs_info->dead_roots);
1190         if (root->fs_info->closing)
1191                 list_splice_init(&root->fs_info->dead_roots, &dirty_fs_roots);
1192
1193         mutex_unlock(&root->fs_info->trans_mutex);
1194
1195         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1196
1197         if (root->fs_info->closing)
1198                 drop_dirty_roots(root->fs_info->tree_root, &dirty_fs_roots);
1199         return ret;
1200 }
1201
1202 /*
1203  * interface function to delete all the snapshots we have scheduled for deletion
1204  */
1205 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1206 {
1207         struct list_head dirty_roots;
1208         INIT_LIST_HEAD(&dirty_roots);
1209 again:
1210         mutex_lock(&root->fs_info->trans_mutex);
1211         list_splice_init(&root->fs_info->dead_roots, &dirty_roots);
1212         mutex_unlock(&root->fs_info->trans_mutex);
1213
1214         if (!list_empty(&dirty_roots)) {
1215                 drop_dirty_roots(root, &dirty_roots);
1216                 goto again;
1217         }
1218         return 0;
1219 }