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                 root->fs_info->last_alloc = 0;
  57                 root->fs_info->last_data_alloc = 0;
  58                 cur_trans->num_writers = 1;
  59                 cur_trans->num_joined = 0;
  60                 cur_trans->transid = root->fs_info->generation;
  61                 init_waitqueue_head(&cur_trans->writer_wait);
  62                 init_waitqueue_head(&cur_trans->commit_wait);
  63                 cur_trans->in_commit = 0;
  64                 cur_trans->blocked = 0;
  65                 cur_trans->use_count = 1;
  66                 cur_trans->commit_done = 0;
  67                 cur_trans->start_time = get_seconds();
  68                 INIT_LIST_HEAD(&cur_trans->pending_snapshots);
  69                 list_add_tail(&cur_trans->list, &root->fs_info->trans_list);
  70                 extent_io_tree_init(&cur_trans->dirty_pages,
  71                                      root->fs_info->btree_inode->i_mapping,
  72                                      GFP_NOFS);
  73                 spin_lock(&root->fs_info->new_trans_lock);
  74                 root->fs_info->running_transaction = cur_trans;
  75                 spin_unlock(&root->fs_info->new_trans_lock);
  76         } else {
  77                 cur_trans->num_writers++;
  78                 cur_trans->num_joined++;
  79         }
  80
  81         return 0;
  82 }
  83
  84 /*
  85  * this does all the record keeping required to make sure that a reference
  86  * counted root is properly recorded in a given transaction.  This is required
  87  * to make sure the old root from before we joined the transaction is deleted
  88  * when the transaction commits
  89  */
  90 noinline int btrfs_record_root_in_trans(struct btrfs_root *root)
  91 {
  92         struct btrfs_dirty_root *dirty;
  93         u64 running_trans_id = root->fs_info->running_transaction->transid;
  94         if (root->ref_cows && root->last_trans < running_trans_id) {
  95                 WARN_ON(root == root->fs_info->extent_root);
  96                 if (root->root_item.refs != 0) {
  97                         radix_tree_tag_set(&root->fs_info->fs_roots_radix,
  98                                    (unsigned long)root->root_key.objectid,
  99                                    BTRFS_ROOT_TRANS_TAG);
 100
 101                         dirty = kmalloc(sizeof(*dirty), GFP_NOFS);
 102                         BUG_ON(!dirty);
 103                         dirty->root = kmalloc(sizeof(*dirty->root), GFP_NOFS);
 104                         BUG_ON(!dirty->root);
 105                         dirty->latest_root = root;
 106                         INIT_LIST_HEAD(&dirty->list);
 107
 108                         root->commit_root = btrfs_root_node(root);
 109
 110                         memcpy(dirty->root, root, sizeof(*root));
 111                         spin_lock_init(&dirty->root->node_lock);
 112                         spin_lock_init(&dirty->root->list_lock);
 113                         mutex_init(&dirty->root->objectid_mutex);
 114                         mutex_init(&dirty->root->log_mutex);
 115                         INIT_LIST_HEAD(&dirty->root->dead_list);
 116                         dirty->root->node = root->commit_root;
 117                         dirty->root->commit_root = NULL;
 118
 119                         spin_lock(&root->list_lock);
 120                         list_add(&dirty->root->dead_list, &root->dead_list);
 121                         spin_unlock(&root->list_lock);
 122
 123                         root->dirty_root = dirty;
 124                 } else {
 125                         WARN_ON(1);
 126                 }
 127                 root->last_trans = running_trans_id;
 128         }
 129         return 0;
 130 }
 131
 132 /* wait for commit against the current transaction to become unblocked
 133  * when this is done, it is safe to start a new transaction, but the current
 134  * transaction might not be fully on disk.
 135  */
 136 static void wait_current_trans(struct btrfs_root *root)
 137 {
 138         struct btrfs_transaction *cur_trans;
 139
 140         cur_trans = root->fs_info->running_transaction;
 141         if (cur_trans && cur_trans->blocked) {
 142                 DEFINE_WAIT(wait);
 143                 cur_trans->use_count++;
 144                 while (1) {
 145                         prepare_to_wait(&root->fs_info->transaction_wait, &wait,
 146                                         TASK_UNINTERRUPTIBLE);
 147                         if (cur_trans->blocked) {
 148                                 mutex_unlock(&root->fs_info->trans_mutex);
 149                                 schedule();
 150                                 mutex_lock(&root->fs_info->trans_mutex);
 151                                 finish_wait(&root->fs_info->transaction_wait,
 152                                             &wait);
 153                         } else {
 154                                 finish_wait(&root->fs_info->transaction_wait,
 155                                             &wait);
 156                                 break;
 157                         }
 158                 }
 159                 put_transaction(cur_trans);
 160         }
 161 }
 162
 163 static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
 164                                              int num_blocks, int wait)
 165 {
 166         struct btrfs_trans_handle *h =
 167                 kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 168         int ret;
 169
 170         mutex_lock(&root->fs_info->trans_mutex);
 171         if (!root->fs_info->log_root_recovering &&
 172             ((wait == 1 && !root->fs_info->open_ioctl_trans) || wait == 2))
 173                 wait_current_trans(root);
 174         ret = join_transaction(root);
 175         BUG_ON(ret);
 176
 177         btrfs_record_root_in_trans(root);
 178         h->transid = root->fs_info->running_transaction->transid;
 179         h->transaction = root->fs_info->running_transaction;
 180         h->blocks_reserved = num_blocks;
 181         h->blocks_used = 0;
 182         h->block_group = 0;
 183         h->alloc_exclude_nr = 0;
 184         h->alloc_exclude_start = 0;
 185         root->fs_info->running_transaction->use_count++;
 186         mutex_unlock(&root->fs_info->trans_mutex);
 187         return h;
 188 }
 189
 190 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 191                                                    int num_blocks)
 192 {
 193         return start_transaction(root, num_blocks, 1);
 194 }
 195 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root,
 196                                                    int num_blocks)
 197 {
 198         return start_transaction(root, num_blocks, 0);
 199 }
 200
 201 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *r,
 202                                                          int num_blocks)
 203 {
 204         return start_transaction(r, num_blocks, 2);
 205 }
 206
 207 /* wait for a transaction commit to be fully complete */
 208 static noinline int wait_for_commit(struct btrfs_root *root,
 209                                     struct btrfs_transaction *commit)
 210 {
 211         DEFINE_WAIT(wait);
 212         mutex_lock(&root->fs_info->trans_mutex);
 213         while (!commit->commit_done) {
 214                 prepare_to_wait(&commit->commit_wait, &wait,
 215                                 TASK_UNINTERRUPTIBLE);
 216                 if (commit->commit_done)
 217                         break;
 218                 mutex_unlock(&root->fs_info->trans_mutex);
 219                 schedule();
 220                 mutex_lock(&root->fs_info->trans_mutex);
 221         }
 222         mutex_unlock(&root->fs_info->trans_mutex);
 223         finish_wait(&commit->commit_wait, &wait);
 224         return 0;
 225 }
 226
 227 /*
 228  * rate limit against the drop_snapshot code.  This helps to slow down new
 229  * operations if the drop_snapshot code isn't able to keep up.
 230  */
 231 static void throttle_on_drops(struct btrfs_root *root)
 232 {
 233         struct btrfs_fs_info *info = root->fs_info;
 234         int harder_count = 0;
 235
 236 harder:
 237         if (atomic_read(&info->throttles)) {
 238                 DEFINE_WAIT(wait);
 239                 int thr;
 240                 thr = atomic_read(&info->throttle_gen);
 241
 242                 do {
 243                         prepare_to_wait(&info->transaction_throttle,
 244                                         &wait, TASK_UNINTERRUPTIBLE);
 245                         if (!atomic_read(&info->throttles)) {
 246                                 finish_wait(&info->transaction_throttle, &wait);
 247                                 break;
 248                         }
 249                         schedule();
 250                         finish_wait(&info->transaction_throttle, &wait);
 251                 } while (thr == atomic_read(&info->throttle_gen));
 252                 harder_count++;
 253
 254                 if (root->fs_info->total_ref_cache_size > 1 * 1024 * 1024 &&
 255                     harder_count < 2)
 256                         goto harder;
 257
 258                 if (root->fs_info->total_ref_cache_size > 5 * 1024 * 1024 &&
 259                     harder_count < 10)
 260                         goto harder;
 261
 262                 if (root->fs_info->total_ref_cache_size > 10 * 1024 * 1024 &&
 263                     harder_count < 20)
 264                         goto harder;
 265         }
 266 }
 267
 268 void btrfs_throttle(struct btrfs_root *root)
 269 {
 270         mutex_lock(&root->fs_info->trans_mutex);
 271         if (!root->fs_info->open_ioctl_trans)
 272                 wait_current_trans(root);
 273         mutex_unlock(&root->fs_info->trans_mutex);
 274
 275         throttle_on_drops(root);
 276 }
 277
 278 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 279                           struct btrfs_root *root, int throttle)
 280 {
 281         struct btrfs_transaction *cur_trans;
 282         struct btrfs_fs_info *info = root->fs_info;
 283
 284         mutex_lock(&info->trans_mutex);
 285         cur_trans = info->running_transaction;
 286         WARN_ON(cur_trans != trans->transaction);
 287         WARN_ON(cur_trans->num_writers < 1);
 288         cur_trans->num_writers--;
 289
 290         if (waitqueue_active(&cur_trans->writer_wait))
 291                 wake_up(&cur_trans->writer_wait);
 292         put_transaction(cur_trans);
 293         mutex_unlock(&info->trans_mutex);
 294         memset(trans, 0, sizeof(*trans));
 295         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 296
 297         if (throttle)
 298                 throttle_on_drops(root);
 299
 300         return 0;
 301 }
 302
 303 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 304                           struct btrfs_root *root)
 305 {
 306         return __btrfs_end_transaction(trans, root, 0);
 307 }
 308
 309 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 310                                    struct btrfs_root *root)
 311 {
 312         return __btrfs_end_transaction(trans, root, 1);
 313 }
 314
 315 /*
 316  * when btree blocks are allocated, they have some corresponding bits set for
 317  * them in one of two extent_io trees.  This is used to make sure all of
 318  * those extents are on disk for transaction or log commit
 319  */
 320 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 321                                         struct extent_io_tree *dirty_pages)
 322 {
 323         int ret;
 324         int err = 0;
 325         int werr = 0;
 326         struct page *page;
 327         struct inode *btree_inode = root->fs_info->btree_inode;
 328         u64 start = 0;
 329         u64 end;
 330         unsigned long index;
 331
 332         while (1) {
 333                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 334                                             EXTENT_DIRTY);
 335                 if (ret)
 336                         break;
 337                 while (start <= end) {
 338                         cond_resched();
 339
 340                         index = start >> PAGE_CACHE_SHIFT;
 341                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 342                         page = find_get_page(btree_inode->i_mapping, index);
 343                         if (!page)
 344                                 continue;
 345
 346                         btree_lock_page_hook(page);
 347                         if (!page->mapping) {
 348                                 unlock_page(page);
 349                                 page_cache_release(page);
 350                                 continue;
 351                         }
 352
 353                         if (PageWriteback(page)) {
 354                                 if (PageDirty(page))
 355                                         wait_on_page_writeback(page);
 356                                 else {
 357                                         unlock_page(page);
 358                                         page_cache_release(page);
 359                                         continue;
 360                                 }
 361                         }
 362                         err = write_one_page(page, 0);
 363                         if (err)
 364                                 werr = err;
 365                         page_cache_release(page);
 366                 }
 367         }
 368         while (1) {
 369                 ret = find_first_extent_bit(dirty_pages, 0, &start, &end,
 370                                             EXTENT_DIRTY);
 371                 if (ret)
 372                         break;
 373
 374                 clear_extent_dirty(dirty_pages, start, end, GFP_NOFS);
 375                 while (start <= end) {
 376                         index = start >> PAGE_CACHE_SHIFT;
 377                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 378                         page = find_get_page(btree_inode->i_mapping, index);
 379                         if (!page)
 380                                 continue;
 381                         if (PageDirty(page)) {
 382                                 btree_lock_page_hook(page);
 383                                 wait_on_page_writeback(page);
 384                                 err = write_one_page(page, 0);
 385                                 if (err)
 386                                         werr = err;
 387                         }
 388                         wait_on_page_writeback(page);
 389                         page_cache_release(page);
 390                         cond_resched();
 391                 }
 392         }
 393         if (err)
 394                 werr = err;
 395         return werr;
 396 }
 397
 398 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 399                                      struct btrfs_root *root)
 400 {
 401         if (!trans || !trans->transaction) {
 402                 struct inode *btree_inode;
 403                 btree_inode = root->fs_info->btree_inode;
 404                 return filemap_write_and_wait(btree_inode->i_mapping);
 405         }
 406         return btrfs_write_and_wait_marked_extents(root,
 407                                            &trans->transaction->dirty_pages);
 408 }
 409
 410 /*
 411  * this is used to update the root pointer in the tree of tree roots.
 412  *
 413  * But, in the case of the extent allocation tree, updating the root
 414  * pointer may allocate blocks which may change the root of the extent
 415  * allocation tree.
 416  *
 417  * So, this loops and repeats and makes sure the cowonly root didn't
 418  * change while the root pointer was being updated in the metadata.
 419  */
 420 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 421                                struct btrfs_root *root)
 422 {
 423         int ret;
 424         u64 old_root_bytenr;
 425         struct btrfs_root *tree_root = root->fs_info->tree_root;
 426
 427         btrfs_extent_post_op(trans, root);
 428         btrfs_write_dirty_block_groups(trans, root);
 429         btrfs_extent_post_op(trans, root);
 430
 431         while (1) {
 432                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 433                 if (old_root_bytenr == root->node->start)
 434                         break;
 435                 btrfs_set_root_bytenr(&root->root_item,
 436                                        root->node->start);
 437                 btrfs_set_root_level(&root->root_item,
 438                                      btrfs_header_level(root->node));
 439                 btrfs_set_root_generation(&root->root_item, trans->transid);
 440
 441                 btrfs_extent_post_op(trans, root);
 442
 443                 ret = btrfs_update_root(trans, tree_root,
 444                                         &root->root_key,
 445                                         &root->root_item);
 446                 BUG_ON(ret);
 447                 btrfs_write_dirty_block_groups(trans, root);
 448                 btrfs_extent_post_op(trans, root);
 449         }
 450         return 0;
 451 }
 452
 453 /*
 454  * update all the cowonly tree roots on disk
 455  */
 456 int btrfs_commit_tree_roots(struct btrfs_trans_handle *trans,
 457                             struct btrfs_root *root)
 458 {
 459         struct btrfs_fs_info *fs_info = root->fs_info;
 460         struct list_head *next;
 461         struct extent_buffer *eb;
 462
 463         btrfs_extent_post_op(trans, fs_info->tree_root);
 464
 465         eb = btrfs_lock_root_node(fs_info->tree_root);
 466         btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb, 0);
 467         btrfs_tree_unlock(eb);
 468         free_extent_buffer(eb);
 469
 470         btrfs_extent_post_op(trans, fs_info->tree_root);
 471
 472         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 473                 next = fs_info->dirty_cowonly_roots.next;
 474                 list_del_init(next);
 475                 root = list_entry(next, struct btrfs_root, dirty_list);
 476
 477                 update_cowonly_root(trans, root);
 478         }
 479         return 0;
 480 }
 481
 482 /*
 483  * dead roots are old snapshots that need to be deleted.  This allocates
 484  * a dirty root struct and adds it into the list of dead roots that need to
 485  * be deleted
 486  */
 487 int btrfs_add_dead_root(struct btrfs_root *root, struct btrfs_root *latest)
 488 {
 489         struct btrfs_dirty_root *dirty;
 490
 491         dirty = kmalloc(sizeof(*dirty), GFP_NOFS);
 492         if (!dirty)
 493                 return -ENOMEM;
 494         dirty->root = root;
 495         dirty->latest_root = latest;
 496
 497         mutex_lock(&root->fs_info->trans_mutex);
 498         list_add(&dirty->list, &latest->fs_info->dead_roots);
 499         mutex_unlock(&root->fs_info->trans_mutex);
 500         return 0;
 501 }
 502
 503 /*
 504  * at transaction commit time we need to schedule the old roots for
 505  * deletion via btrfs_drop_snapshot.  This runs through all the
 506  * reference counted roots that were modified in the current
 507  * transaction and puts them into the drop list
 508  */
 509 static noinline int add_dirty_roots(struct btrfs_trans_handle *trans,
 510                                     struct radix_tree_root *radix,
 511                                     struct list_head *list)
 512 {
 513         struct btrfs_dirty_root *dirty;
 514         struct btrfs_root *gang[8];
 515         struct btrfs_root *root;
 516         int i;
 517         int ret;
 518         int err = 0;
 519         u32 refs;
 520
 521         while (1) {
 522                 ret = radix_tree_gang_lookup_tag(radix, (void **)gang, 0,
 523                                                  ARRAY_SIZE(gang),
 524                                                  BTRFS_ROOT_TRANS_TAG);
 525                 if (ret == 0)
 526                         break;
 527                 for (i = 0; i < ret; i++) {
 528                         root = gang[i];
 529                         radix_tree_tag_clear(radix,
 530                                      (unsigned long)root->root_key.objectid,
 531                                      BTRFS_ROOT_TRANS_TAG);
 532
 533                         BUG_ON(!root->ref_tree);
 534                         dirty = root->dirty_root;
 535
 536                         btrfs_free_log(trans, root);
 537                         btrfs_free_reloc_root(trans, root);
 538
 539                         if (root->commit_root == root->node) {
 540                                 WARN_ON(root->node->start !=
 541                                         btrfs_root_bytenr(&root->root_item));
 542
 543                                 free_extent_buffer(root->commit_root);
 544                                 root->commit_root = NULL;
 545                                 root->dirty_root = NULL;
 546
 547                                 spin_lock(&root->list_lock);
 548                                 list_del_init(&dirty->root->dead_list);
 549                                 spin_unlock(&root->list_lock);
 550
 551                                 kfree(dirty->root);
 552                                 kfree(dirty);
 553
 554                                 /* make sure to update the root on disk
 555                                  * so we get any updates to the block used
 556                                  * counts
 557                                  */
 558                                 err = btrfs_update_root(trans,
 559                                                 root->fs_info->tree_root,
 560                                                 &root->root_key,
 561                                                 &root->root_item);
 562                                 continue;
 563                         }
 564
 565                         memset(&root->root_item.drop_progress, 0,
 566                                sizeof(struct btrfs_disk_key));
 567                         root->root_item.drop_level = 0;
 568                         root->commit_root = NULL;
 569                         root->dirty_root = NULL;
 570                         root->root_key.offset = root->fs_info->generation;
 571                         btrfs_set_root_bytenr(&root->root_item,
 572                                               root->node->start);
 573                         btrfs_set_root_level(&root->root_item,
 574                                              btrfs_header_level(root->node));
 575                         btrfs_set_root_generation(&root->root_item,
 576                                                   root->root_key.offset);
 577
 578                         err = btrfs_insert_root(trans, root->fs_info->tree_root,
 579                                                 &root->root_key,
 580                                                 &root->root_item);
 581                         if (err)
 582                                 break;
 583
 584                         refs = btrfs_root_refs(&dirty->root->root_item);
 585                         btrfs_set_root_refs(&dirty->root->root_item, refs - 1);
 586                         err = btrfs_update_root(trans, root->fs_info->tree_root,
 587                                                 &dirty->root->root_key,
 588                                                 &dirty->root->root_item);
 589
 590                         BUG_ON(err);
 591                         if (refs == 1) {
 592                                 list_add(&dirty->list, list);
 593                         } else {
 594                                 WARN_ON(1);
 595                                 free_extent_buffer(dirty->root->node);
 596                                 kfree(dirty->root);
 597                                 kfree(dirty);
 598                         }
 599                 }
 600         }
 601         return err;
 602 }
 603
 604 /*
 605  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 606  * otherwise every leaf in the btree is read and defragged.
 607  */
 608 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 609 {
 610         struct btrfs_fs_info *info = root->fs_info;
 611         int ret;
 612         struct btrfs_trans_handle *trans;
 613         unsigned long nr;
 614
 615         smp_mb();
 616         if (root->defrag_running)
 617                 return 0;
 618         trans = btrfs_start_transaction(root, 1);
 619         while (1) {
 620                 root->defrag_running = 1;
 621                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 622                 nr = trans->blocks_used;
 623                 btrfs_end_transaction(trans, root);
 624                 btrfs_btree_balance_dirty(info->tree_root, nr);
 625                 cond_resched();
 626
 627                 trans = btrfs_start_transaction(root, 1);
 628                 if (root->fs_info->closing || ret != -EAGAIN)
 629                         break;
 630         }
 631         root->defrag_running = 0;
 632         smp_mb();
 633         btrfs_end_transaction(trans, root);
 634         return 0;
 635 }
 636
 637 /*
 638  * Given a list of roots that need to be deleted, call btrfs_drop_snapshot on
 639  * all of them
 640  */
 641 static noinline int drop_dirty_roots(struct btrfs_root *tree_root,
 642                                      struct list_head *list)
 643 {
 644         struct btrfs_dirty_root *dirty;
 645         struct btrfs_trans_handle *trans;
 646         unsigned long nr;
 647         u64 num_bytes;
 648         u64 bytes_used;
 649         u64 max_useless;
 650         int ret = 0;
 651         int err;
 652
 653         while (!list_empty(list)) {
 654                 struct btrfs_root *root;
 655
 656                 dirty = list_entry(list->prev, struct btrfs_dirty_root, list);
 657                 list_del_init(&dirty->list);
 658
 659                 num_bytes = btrfs_root_used(&dirty->root->root_item);
 660                 root = dirty->latest_root;
 661                 atomic_inc(&root->fs_info->throttles);
 662
 663                 while (1) {
 664                         trans = btrfs_start_transaction(tree_root, 1);
 665                         mutex_lock(&root->fs_info->drop_mutex);
 666                         ret = btrfs_drop_snapshot(trans, dirty->root);
 667                         if (ret != -EAGAIN)
 668                                 break;
 669                         mutex_unlock(&root->fs_info->drop_mutex);
 670
 671                         err = btrfs_update_root(trans,
 672                                         tree_root,
 673                                         &dirty->root->root_key,
 674                                         &dirty->root->root_item);
 675                         if (err)
 676                                 ret = err;
 677                         nr = trans->blocks_used;
 678                         ret = btrfs_end_transaction(trans, tree_root);
 679                         BUG_ON(ret);
 680
 681                         btrfs_btree_balance_dirty(tree_root, nr);
 682                         cond_resched();
 683                 }
 684                 BUG_ON(ret);
 685                 atomic_dec(&root->fs_info->throttles);
 686                 wake_up(&root->fs_info->transaction_throttle);
 687
 688                 num_bytes -= btrfs_root_used(&dirty->root->root_item);
 689                 bytes_used = btrfs_root_used(&root->root_item);
 690                 if (num_bytes) {
 691                         btrfs_record_root_in_trans(root);
 692                         btrfs_set_root_used(&root->root_item,
 693                                             bytes_used - num_bytes);
 694                 }
 695
 696                 ret = btrfs_del_root(trans, tree_root, &dirty->root->root_key);
 697                 if (ret) {
 698                         BUG();
 699                         break;
 700                 }
 701                 mutex_unlock(&root->fs_info->drop_mutex);
 702
 703                 spin_lock(&root->list_lock);
 704                 list_del_init(&dirty->root->dead_list);
 705                 if (!list_empty(&root->dead_list)) {
 706                         struct btrfs_root *oldest;
 707                         oldest = list_entry(root->dead_list.prev,
 708                                             struct btrfs_root, dead_list);
 709                         max_useless = oldest->root_key.offset - 1;
 710                 } else {
 711                         max_useless = root->root_key.offset - 1;
 712                 }
 713                 spin_unlock(&root->list_lock);
 714
 715                 nr = trans->blocks_used;
 716                 ret = btrfs_end_transaction(trans, tree_root);
 717                 BUG_ON(ret);
 718
 719                 ret = btrfs_remove_leaf_refs(root, max_useless, 0);
 720                 BUG_ON(ret);
 721
 722                 free_extent_buffer(dirty->root->node);
 723                 kfree(dirty->root);
 724                 kfree(dirty);
 725
 726                 btrfs_btree_balance_dirty(tree_root, nr);
 727                 cond_resched();
 728         }
 729         return ret;
 730 }
 731
 732 /*
 733  * new snapshots need to be created at a very specific time in the
 734  * transaction commit.  This does the actual creation
 735  */
 736 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 737                                    struct btrfs_fs_info *fs_info,
 738                                    struct btrfs_pending_snapshot *pending)
 739 {
 740         struct btrfs_key key;
 741         struct btrfs_root_item *new_root_item;
 742         struct btrfs_root *tree_root = fs_info->tree_root;
 743         struct btrfs_root *root = pending->root;
 744         struct extent_buffer *tmp;
 745         struct extent_buffer *old;
 746         int ret;
 747         u64 objectid;
 748
 749         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 750         if (!new_root_item) {
 751                 ret = -ENOMEM;
 752                 goto fail;
 753         }
 754         ret = btrfs_find_free_objectid(trans, tree_root, 0, &objectid);
 755         if (ret)
 756                 goto fail;
 757
 758         btrfs_record_root_in_trans(root);
 759         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
 760         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
 761
 762         key.objectid = objectid;
 763         key.offset = trans->transid;
 764         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
 765
 766         old = btrfs_lock_root_node(root);
 767         btrfs_cow_block(trans, root, old, NULL, 0, &old, 0);
 768
 769         btrfs_copy_root(trans, root, old, &tmp, objectid);
 770         btrfs_tree_unlock(old);
 771         free_extent_buffer(old);
 772
 773         btrfs_set_root_bytenr(new_root_item, tmp->start);
 774         btrfs_set_root_level(new_root_item, btrfs_header_level(tmp));
 775         btrfs_set_root_generation(new_root_item, trans->transid);
 776         ret = btrfs_insert_root(trans, root->fs_info->tree_root, &key,
 777                                 new_root_item);
 778         btrfs_tree_unlock(tmp);
 779         free_extent_buffer(tmp);
 780         if (ret)
 781                 goto fail;
 782
 783         key.offset = (u64)-1;
 784         memcpy(&pending->root_key, &key, sizeof(key));
 785 fail:
 786         kfree(new_root_item);
 787         return ret;
 788 }
 789
 790 static noinline int finish_pending_snapshot(struct btrfs_fs_info *fs_info,
 791                                    struct btrfs_pending_snapshot *pending)
 792 {
 793         int ret;
 794         int namelen;
 795         u64 index = 0;
 796         struct btrfs_trans_handle *trans;
 797         struct inode *parent_inode;
 798         struct inode *inode;
 799         struct btrfs_root *parent_root;
 800
 801         parent_inode = pending->dentry->d_parent->d_inode;
 802         parent_root = BTRFS_I(parent_inode)->root;
 803         trans = btrfs_join_transaction(parent_root, 1);
 804
 805         /*
 806          * insert the directory item
 807          */
 808         namelen = strlen(pending->name);
 809         ret = btrfs_set_inode_index(parent_inode, &index);
 810         ret = btrfs_insert_dir_item(trans, parent_root,
 811                             pending->name, namelen,
 812                             parent_inode->i_ino,
 813                             &pending->root_key, BTRFS_FT_DIR, index);
 814
 815         if (ret)
 816                 goto fail;
 817
 818         btrfs_i_size_write(parent_inode, parent_inode->i_size + namelen * 2);
 819         ret = btrfs_update_inode(trans, parent_root, parent_inode);
 820         BUG_ON(ret);
 821
 822         /* add the backref first */
 823         ret = btrfs_add_root_ref(trans, parent_root->fs_info->tree_root,
 824                                  pending->root_key.objectid,
 825                                  BTRFS_ROOT_BACKREF_KEY,
 826                                  parent_root->root_key.objectid,
 827                                  parent_inode->i_ino, index, pending->name,
 828                                  namelen);
 829
 830         BUG_ON(ret);
 831
 832         /* now add the forward ref */
 833         ret = btrfs_add_root_ref(trans, parent_root->fs_info->tree_root,
 834                                  parent_root->root_key.objectid,
 835                                  BTRFS_ROOT_REF_KEY,
 836                                  pending->root_key.objectid,
 837                                  parent_inode->i_ino, index, pending->name,
 838                                  namelen);
 839
 840         inode = btrfs_lookup_dentry(parent_inode, pending->dentry);
 841         d_instantiate(pending->dentry, inode);
 842 fail:
 843         btrfs_end_transaction(trans, fs_info->fs_root);
 844         return ret;
 845 }
 846
 847 /*
 848  * create all the snapshots we've scheduled for creation
 849  */
 850 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
 851                                              struct btrfs_fs_info *fs_info)
 852 {
 853         struct btrfs_pending_snapshot *pending;
 854         struct list_head *head = &trans->transaction->pending_snapshots;
 855         int ret;
 856
 857         list_for_each_entry(pending, head, list) {
 858                 ret = create_pending_snapshot(trans, fs_info, pending);
 859                 BUG_ON(ret);
 860         }
 861         return 0;
 862 }
 863
 864 static noinline int finish_pending_snapshots(struct btrfs_trans_handle *trans,
 865                                              struct btrfs_fs_info *fs_info)
 866 {
 867         struct btrfs_pending_snapshot *pending;
 868         struct list_head *head = &trans->transaction->pending_snapshots;
 869         int ret;
 870
 871         while (!list_empty(head)) {
 872                 pending = list_entry(head->next,
 873                                      struct btrfs_pending_snapshot, list);
 874                 ret = finish_pending_snapshot(fs_info, pending);
 875                 BUG_ON(ret);
 876                 list_del(&pending->list);
 877                 kfree(pending->name);
 878                 kfree(pending);
 879         }
 880         return 0;
 881 }
 882
 883 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
 884                              struct btrfs_root *root)
 885 {
 886         unsigned long joined = 0;
 887         unsigned long timeout = 1;
 888         struct btrfs_transaction *cur_trans;
 889         struct btrfs_transaction *prev_trans = NULL;
 890         struct btrfs_root *chunk_root = root->fs_info->chunk_root;
 891         struct list_head dirty_fs_roots;
 892         struct extent_io_tree *pinned_copy;
 893         DEFINE_WAIT(wait);
 894         int ret;
 895
 896         INIT_LIST_HEAD(&dirty_fs_roots);
 897         mutex_lock(&root->fs_info->trans_mutex);
 898         if (trans->transaction->in_commit) {
 899                 cur_trans = trans->transaction;
 900                 trans->transaction->use_count++;
 901                 mutex_unlock(&root->fs_info->trans_mutex);
 902                 btrfs_end_transaction(trans, root);
 903
 904                 ret = wait_for_commit(root, cur_trans);
 905                 BUG_ON(ret);
 906
 907                 mutex_lock(&root->fs_info->trans_mutex);
 908                 put_transaction(cur_trans);
 909                 mutex_unlock(&root->fs_info->trans_mutex);
 910
 911                 return 0;
 912         }
 913
 914         pinned_copy = kmalloc(sizeof(*pinned_copy), GFP_NOFS);
 915         if (!pinned_copy)
 916                 return -ENOMEM;
 917
 918         extent_io_tree_init(pinned_copy,
 919                              root->fs_info->btree_inode->i_mapping, GFP_NOFS);
 920
 921         trans->transaction->in_commit = 1;
 922         trans->transaction->blocked = 1;
 923         cur_trans = trans->transaction;
 924         if (cur_trans->list.prev != &root->fs_info->trans_list) {
 925                 prev_trans = list_entry(cur_trans->list.prev,
 926                                         struct btrfs_transaction, list);
 927                 if (!prev_trans->commit_done) {
 928                         prev_trans->use_count++;
 929                         mutex_unlock(&root->fs_info->trans_mutex);
 930
 931                         wait_for_commit(root, prev_trans);
 932
 933                         mutex_lock(&root->fs_info->trans_mutex);
 934                         put_transaction(prev_trans);
 935                 }
 936         }
 937
 938         do {
 939                 int snap_pending = 0;
 940                 joined = cur_trans->num_joined;
 941                 if (!list_empty(&trans->transaction->pending_snapshots))
 942                         snap_pending = 1;
 943
 944                 WARN_ON(cur_trans != trans->transaction);
 945                 prepare_to_wait(&cur_trans->writer_wait, &wait,
 946                                 TASK_UNINTERRUPTIBLE);
 947
 948                 if (cur_trans->num_writers > 1)
 949                         timeout = MAX_SCHEDULE_TIMEOUT;
 950                 else
 951                         timeout = 1;
 952
 953                 mutex_unlock(&root->fs_info->trans_mutex);
 954
 955                 if (snap_pending) {
 956                         ret = btrfs_wait_ordered_extents(root, 1);
 957                         BUG_ON(ret);
 958                 }
 959
 960                 schedule_timeout(timeout);
 961
 962                 mutex_lock(&root->fs_info->trans_mutex);
 963                 finish_wait(&cur_trans->writer_wait, &wait);
 964         } while (cur_trans->num_writers > 1 ||
 965                  (cur_trans->num_joined != joined));
 966
 967         ret = create_pending_snapshots(trans, root->fs_info);
 968         BUG_ON(ret);
 969
 970         WARN_ON(cur_trans != trans->transaction);
 971
 972         /* btrfs_commit_tree_roots is responsible for getting the
 973          * various roots consistent with each other.  Every pointer
 974          * in the tree of tree roots has to point to the most up to date
 975          * root for every subvolume and other tree.  So, we have to keep
 976          * the tree logging code from jumping in and changing any
 977          * of the trees.
 978          *
 979          * At this point in the commit, there can't be any tree-log
 980          * writers, but a little lower down we drop the trans mutex
 981          * and let new people in.  By holding the tree_log_mutex
 982          * from now until after the super is written, we avoid races
 983          * with the tree-log code.
 984          */
 985         mutex_lock(&root->fs_info->tree_log_mutex);
 986         /*
 987          * keep tree reloc code from adding new reloc trees
 988          */
 989         mutex_lock(&root->fs_info->tree_reloc_mutex);
 990
 991
 992         ret = add_dirty_roots(trans, &root->fs_info->fs_roots_radix,
 993                               &dirty_fs_roots);
 994         BUG_ON(ret);
 995
 996         /* add_dirty_roots gets rid of all the tree log roots, it is now
 997          * safe to free the root of tree log roots
 998          */
 999         btrfs_free_log_root_tree(trans, root->fs_info);
1000
1001         ret = btrfs_commit_tree_roots(trans, root);
1002         BUG_ON(ret);
1003
1004         cur_trans = root->fs_info->running_transaction;
1005         spin_lock(&root->fs_info->new_trans_lock);
1006         root->fs_info->running_transaction = NULL;
1007         spin_unlock(&root->fs_info->new_trans_lock);
1008         btrfs_set_super_generation(&root->fs_info->super_copy,
1009                                    cur_trans->transid);
1010         btrfs_set_super_root(&root->fs_info->super_copy,
1011                              root->fs_info->tree_root->node->start);
1012         btrfs_set_super_root_level(&root->fs_info->super_copy,
1013                            btrfs_header_level(root->fs_info->tree_root->node));
1014
1015         btrfs_set_super_chunk_root(&root->fs_info->super_copy,
1016                                    chunk_root->node->start);
1017         btrfs_set_super_chunk_root_level(&root->fs_info->super_copy,
1018                                          btrfs_header_level(chunk_root->node));
1019         btrfs_set_super_chunk_root_generation(&root->fs_info->super_copy,
1020                                 btrfs_header_generation(chunk_root->node));
1021
1022         if (!root->fs_info->log_root_recovering) {
1023                 btrfs_set_super_log_root(&root->fs_info->super_copy, 0);
1024                 btrfs_set_super_log_root_level(&root->fs_info->super_copy, 0);
1025         }
1026
1027         memcpy(&root->fs_info->super_for_commit, &root->fs_info->super_copy,
1028                sizeof(root->fs_info->super_copy));
1029
1030         btrfs_copy_pinned(root, pinned_copy);
1031
1032         trans->transaction->blocked = 0;
1033         wake_up(&root->fs_info->transaction_throttle);
1034         wake_up(&root->fs_info->transaction_wait);
1035
1036         mutex_unlock(&root->fs_info->trans_mutex);
1037         ret = btrfs_write_and_wait_transaction(trans, root);
1038         BUG_ON(ret);
1039         write_ctree_super(trans, root, 0);
1040
1041         /*
1042          * the super is written, we can safely allow the tree-loggers
1043          * to go about their business
1044          */
1045         mutex_unlock(&root->fs_info->tree_log_mutex);
1046
1047         btrfs_finish_extent_commit(trans, root, pinned_copy);
1048         kfree(pinned_copy);
1049
1050         btrfs_drop_dead_reloc_roots(root);
1051         mutex_unlock(&root->fs_info->tree_reloc_mutex);
1052
1053         /* do the directory inserts of any pending snapshot creations */
1054         finish_pending_snapshots(trans, root->fs_info);
1055
1056         mutex_lock(&root->fs_info->trans_mutex);
1057
1058         cur_trans->commit_done = 1;
1059         root->fs_info->last_trans_committed = cur_trans->transid;
1060         wake_up(&cur_trans->commit_wait);
1061
1062         put_transaction(cur_trans);
1063         put_transaction(cur_trans);
1064
1065         list_splice_init(&dirty_fs_roots, &root->fs_info->dead_roots);
1066         if (root->fs_info->closing)
1067                 list_splice_init(&root->fs_info->dead_roots, &dirty_fs_roots);
1068
1069         mutex_unlock(&root->fs_info->trans_mutex);
1070
1071         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1072
1073         if (root->fs_info->closing)
1074                 drop_dirty_roots(root->fs_info->tree_root, &dirty_fs_roots);
1075         return ret;
1076 }
1077
1078 /*
1079  * interface function to delete all the snapshots we have scheduled for deletion
1080  */
1081 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1082 {
1083         struct list_head dirty_roots;
1084         INIT_LIST_HEAD(&dirty_roots);
1085 again:
1086         mutex_lock(&root->fs_info->trans_mutex);
1087         list_splice_init(&root->fs_info->dead_roots, &dirty_roots);
1088         mutex_unlock(&root->fs_info->trans_mutex);
1089
1090         if (!list_empty(&dirty_roots)) {
1091                 drop_dirty_roots(root, &dirty_roots);
1092                 goto again;
1093         }
1094         return 0;
1095 }