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/slab.h>
  21 #include <linux/sched.h>
  22 #include <linux/writeback.h>
  23 #include <linux/pagemap.h>
  24 #include <linux/blkdev.h>
  25 #include "ctree.h"
  26 #include "disk-io.h"
  27 #include "transaction.h"
  28 #include "locking.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 static noinline void switch_commit_root(struct btrfs_root *root)
  45 {
  46         free_extent_buffer(root->commit_root);
  47         root->commit_root = btrfs_root_node(root);
  48 }
  49
  50 /*
  51  * either allocate a new transaction or hop into the existing one
  52  */
  53 static noinline int join_transaction(struct btrfs_root *root)
  54 {
  55         struct btrfs_transaction *cur_trans;
  56         cur_trans = root->fs_info->running_transaction;
  57         if (!cur_trans) {
  58                 cur_trans = kmem_cache_alloc(btrfs_transaction_cachep,
  59                                              GFP_NOFS);
  60                 BUG_ON(!cur_trans);
  61                 root->fs_info->generation++;
  62                 cur_trans->num_writers = 1;
  63                 cur_trans->num_joined = 0;
  64                 cur_trans->transid = root->fs_info->generation;
  65                 init_waitqueue_head(&cur_trans->writer_wait);
  66                 init_waitqueue_head(&cur_trans->commit_wait);
  67                 cur_trans->in_commit = 0;
  68                 cur_trans->blocked = 0;
  69                 cur_trans->use_count = 1;
  70                 cur_trans->commit_done = 0;
  71                 cur_trans->start_time = get_seconds();
  72
  73                 cur_trans->delayed_refs.root = RB_ROOT;
  74                 cur_trans->delayed_refs.num_entries = 0;
  75                 cur_trans->delayed_refs.num_heads_ready = 0;
  76                 cur_trans->delayed_refs.num_heads = 0;
  77                 cur_trans->delayed_refs.flushing = 0;
  78                 cur_trans->delayed_refs.run_delayed_start = 0;
  79                 spin_lock_init(&cur_trans->delayed_refs.lock);
  80
  81                 INIT_LIST_HEAD(&cur_trans->pending_snapshots);
  82                 list_add_tail(&cur_trans->list, &root->fs_info->trans_list);
  83                 extent_io_tree_init(&cur_trans->dirty_pages,
  84                                      root->fs_info->btree_inode->i_mapping,
  85                                      GFP_NOFS);
  86                 spin_lock(&root->fs_info->new_trans_lock);
  87                 root->fs_info->running_transaction = cur_trans;
  88                 spin_unlock(&root->fs_info->new_trans_lock);
  89         } else {
  90                 cur_trans->num_writers++;
  91                 cur_trans->num_joined++;
  92         }
  93
  94         return 0;
  95 }
  96
  97 /*
  98  * this does all the record keeping required to make sure that a reference
  99  * counted root is properly recorded in a given transaction.  This is required
 100  * to make sure the old root from before we joined the transaction is deleted
 101  * when the transaction commits
 102  */
 103 static noinline int record_root_in_trans(struct btrfs_trans_handle *trans,
 104                                          struct btrfs_root *root)
 105 {
 106         if (root->ref_cows && root->last_trans < trans->transid) {
 107                 WARN_ON(root == root->fs_info->extent_root);
 108                 WARN_ON(root->commit_root != root->node);
 109
 110                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 111                            (unsigned long)root->root_key.objectid,
 112                            BTRFS_ROOT_TRANS_TAG);
 113                 root->last_trans = trans->transid;
 114                 btrfs_init_reloc_root(trans, root);
 115         }
 116         return 0;
 117 }
 118
 119 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 120                                struct btrfs_root *root)
 121 {
 122         if (!root->ref_cows)
 123                 return 0;
 124
 125         mutex_lock(&root->fs_info->trans_mutex);
 126         if (root->last_trans == trans->transid) {
 127                 mutex_unlock(&root->fs_info->trans_mutex);
 128                 return 0;
 129         }
 130
 131         record_root_in_trans(trans, root);
 132         mutex_unlock(&root->fs_info->trans_mutex);
 133         return 0;
 134 }
 135
 136 /* wait for commit against the current transaction to become unblocked
 137  * when this is done, it is safe to start a new transaction, but the current
 138  * transaction might not be fully on disk.
 139  */
 140 static void wait_current_trans(struct btrfs_root *root)
 141 {
 142         struct btrfs_transaction *cur_trans;
 143
 144         cur_trans = root->fs_info->running_transaction;
 145         if (cur_trans && cur_trans->blocked) {
 146                 DEFINE_WAIT(wait);
 147                 cur_trans->use_count++;
 148                 while (1) {
 149                         prepare_to_wait(&root->fs_info->transaction_wait, &wait,
 150                                         TASK_UNINTERRUPTIBLE);
 151                         if (!cur_trans->blocked)
 152                                 break;
 153                         mutex_unlock(&root->fs_info->trans_mutex);
 154                         schedule();
 155                         mutex_lock(&root->fs_info->trans_mutex);
 156                 }
 157                 finish_wait(&root->fs_info->transaction_wait, &wait);
 158                 put_transaction(cur_trans);
 159         }
 160 }
 161
 162 enum btrfs_trans_type {
 163         TRANS_START,
 164         TRANS_JOIN,
 165         TRANS_USERSPACE,
 166 };
 167
 168 static int may_wait_transaction(struct btrfs_root *root, int type)
 169 {
 170         if (!root->fs_info->log_root_recovering &&
 171             ((type == TRANS_START && !root->fs_info->open_ioctl_trans) ||
 172              type == TRANS_USERSPACE))
 173                 return 1;
 174         return 0;
 175 }
 176
 177 static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
 178                                                     u64 num_items, int type)
 179 {
 180         struct btrfs_trans_handle *h;
 181         struct btrfs_transaction *cur_trans;
 182         int retries = 0;
 183         int ret;
 184 again:
 185         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 186         if (!h)
 187                 return ERR_PTR(-ENOMEM);
 188
 189         mutex_lock(&root->fs_info->trans_mutex);
 190         if (may_wait_transaction(root, type))
 191                 wait_current_trans(root);
 192
 193         ret = join_transaction(root);
 194         BUG_ON(ret);
 195
 196         cur_trans = root->fs_info->running_transaction;
 197         cur_trans->use_count++;
 198         mutex_unlock(&root->fs_info->trans_mutex);
 199
 200         h->transid = cur_trans->transid;
 201         h->transaction = cur_trans;
 202         h->blocks_used = 0;
 203         h->block_group = 0;
 204         h->bytes_reserved = 0;
 205         h->delayed_ref_updates = 0;
 206         h->block_rsv = NULL;
 207
 208         smp_mb();
 209         if (cur_trans->blocked && may_wait_transaction(root, type)) {
 210                 btrfs_commit_transaction(h, root);
 211                 goto again;
 212         }
 213
 214         if (num_items > 0) {
 215                 ret = btrfs_trans_reserve_metadata(h, root, num_items,
 216                                                    &retries);
 217                 if (ret == -EAGAIN) {
 218                         btrfs_commit_transaction(h, root);
 219                         goto again;
 220                 }
 221                 if (ret < 0) {
 222                         btrfs_end_transaction(h, root);
 223                         return ERR_PTR(ret);
 224                 }
 225         }
 226
 227         mutex_lock(&root->fs_info->trans_mutex);
 228         record_root_in_trans(h, root);
 229         mutex_unlock(&root->fs_info->trans_mutex);
 230
 231         if (!current->journal_info && type != TRANS_USERSPACE)
 232                 current->journal_info = h;
 233         return h;
 234 }
 235
 236 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 237                                                    int num_items)
 238 {
 239         return start_transaction(root, num_items, TRANS_START);
 240 }
 241 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root,
 242                                                    int num_blocks)
 243 {
 244         return start_transaction(root, 0, TRANS_JOIN);
 245 }
 246
 247 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *r,
 248                                                          int num_blocks)
 249 {
 250         return start_transaction(r, 0, TRANS_USERSPACE);
 251 }
 252
 253 /* wait for a transaction commit to be fully complete */
 254 static noinline int wait_for_commit(struct btrfs_root *root,
 255                                     struct btrfs_transaction *commit)
 256 {
 257         DEFINE_WAIT(wait);
 258         mutex_lock(&root->fs_info->trans_mutex);
 259         while (!commit->commit_done) {
 260                 prepare_to_wait(&commit->commit_wait, &wait,
 261                                 TASK_UNINTERRUPTIBLE);
 262                 if (commit->commit_done)
 263                         break;
 264                 mutex_unlock(&root->fs_info->trans_mutex);
 265                 schedule();
 266                 mutex_lock(&root->fs_info->trans_mutex);
 267         }
 268         mutex_unlock(&root->fs_info->trans_mutex);
 269         finish_wait(&commit->commit_wait, &wait);
 270         return 0;
 271 }
 272
 273 #if 0
 274 /*
 275  * rate limit against the drop_snapshot code.  This helps to slow down new
 276  * operations if the drop_snapshot code isn't able to keep up.
 277  */
 278 static void throttle_on_drops(struct btrfs_root *root)
 279 {
 280         struct btrfs_fs_info *info = root->fs_info;
 281         int harder_count = 0;
 282
 283 harder:
 284         if (atomic_read(&info->throttles)) {
 285                 DEFINE_WAIT(wait);
 286                 int thr;
 287                 thr = atomic_read(&info->throttle_gen);
 288
 289                 do {
 290                         prepare_to_wait(&info->transaction_throttle,
 291                                         &wait, TASK_UNINTERRUPTIBLE);
 292                         if (!atomic_read(&info->throttles)) {
 293                                 finish_wait(&info->transaction_throttle, &wait);
 294                                 break;
 295                         }
 296                         schedule();
 297                         finish_wait(&info->transaction_throttle, &wait);
 298                 } while (thr == atomic_read(&info->throttle_gen));
 299                 harder_count++;
 300
 301                 if (root->fs_info->total_ref_cache_size > 1 * 1024 * 1024 &&
 302                     harder_count < 2)
 303                         goto harder;
 304
 305                 if (root->fs_info->total_ref_cache_size > 5 * 1024 * 1024 &&
 306                     harder_count < 10)
 307                         goto harder;
 308
 309                 if (root->fs_info->total_ref_cache_size > 10 * 1024 * 1024 &&
 310                     harder_count < 20)
 311                         goto harder;
 312         }
 313 }
 314 #endif
 315
 316 void btrfs_throttle(struct btrfs_root *root)
 317 {
 318         mutex_lock(&root->fs_info->trans_mutex);
 319         if (!root->fs_info->open_ioctl_trans)
 320                 wait_current_trans(root);
 321         mutex_unlock(&root->fs_info->trans_mutex);
 322 }
 323
 324 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 325                           struct btrfs_root *root, int throttle)
 326 {
 327         struct btrfs_transaction *cur_trans;
 328         struct btrfs_fs_info *info = root->fs_info;
 329         int count = 0;
 330
 331         while (count < 4) {
 332                 unsigned long cur = trans->delayed_ref_updates;
 333                 trans->delayed_ref_updates = 0;
 334                 if (cur &&
 335                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 336                         trans->delayed_ref_updates = 0;
 337
 338                         /*
 339                          * do a full flush if the transaction is trying
 340                          * to close
 341                          */
 342                         if (trans->transaction->delayed_refs.flushing)
 343                                 cur = 0;
 344                         btrfs_run_delayed_refs(trans, root, cur);
 345                 } else {
 346                         break;
 347                 }
 348                 count++;
 349         }
 350
 351         btrfs_trans_release_metadata(trans, root);
 352
 353         mutex_lock(&info->trans_mutex);
 354         cur_trans = info->running_transaction;
 355         WARN_ON(cur_trans != trans->transaction);
 356         WARN_ON(cur_trans->num_writers < 1);
 357         cur_trans->num_writers--;
 358
 359         if (waitqueue_active(&cur_trans->writer_wait))
 360                 wake_up(&cur_trans->writer_wait);
 361         put_transaction(cur_trans);
 362         mutex_unlock(&info->trans_mutex);
 363
 364         if (current->journal_info == trans)
 365                 current->journal_info = NULL;
 366         memset(trans, 0, sizeof(*trans));
 367         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 368
 369         if (throttle)
 370                 btrfs_run_delayed_iputs(root);
 371
 372         return 0;
 373 }
 374
 375 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 376                           struct btrfs_root *root)
 377 {
 378         return __btrfs_end_transaction(trans, root, 0);
 379 }
 380
 381 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 382                                    struct btrfs_root *root)
 383 {
 384         return __btrfs_end_transaction(trans, root, 1);
 385 }
 386
 387 /*
 388  * when btree blocks are allocated, they have some corresponding bits set for
 389  * them in one of two extent_io trees.  This is used to make sure all of
 390  * those extents are sent to disk but does not wait on them
 391  */
 392 int btrfs_write_marked_extents(struct btrfs_root *root,
 393                                struct extent_io_tree *dirty_pages, int mark)
 394 {
 395         int ret;
 396         int err = 0;
 397         int werr = 0;
 398         struct page *page;
 399         struct inode *btree_inode = root->fs_info->btree_inode;
 400         u64 start = 0;
 401         u64 end;
 402         unsigned long index;
 403
 404         while (1) {
 405                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 406                                             mark);
 407                 if (ret)
 408                         break;
 409                 while (start <= end) {
 410                         cond_resched();
 411
 412                         index = start >> PAGE_CACHE_SHIFT;
 413                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 414                         page = find_get_page(btree_inode->i_mapping, index);
 415                         if (!page)
 416                                 continue;
 417
 418                         btree_lock_page_hook(page);
 419                         if (!page->mapping) {
 420                                 unlock_page(page);
 421                                 page_cache_release(page);
 422                                 continue;
 423                         }
 424
 425                         if (PageWriteback(page)) {
 426                                 if (PageDirty(page))
 427                                         wait_on_page_writeback(page);
 428                                 else {
 429                                         unlock_page(page);
 430                                         page_cache_release(page);
 431                                         continue;
 432                                 }
 433                         }
 434                         err = write_one_page(page, 0);
 435                         if (err)
 436                                 werr = err;
 437                         page_cache_release(page);
 438                 }
 439         }
 440         if (err)
 441                 werr = err;
 442         return werr;
 443 }
 444
 445 /*
 446  * when btree blocks are allocated, they have some corresponding bits set for
 447  * them in one of two extent_io trees.  This is used to make sure all of
 448  * those extents are on disk for transaction or log commit.  We wait
 449  * on all the pages and clear them from the dirty pages state tree
 450  */
 451 int btrfs_wait_marked_extents(struct btrfs_root *root,
 452                               struct extent_io_tree *dirty_pages, int mark)
 453 {
 454         int ret;
 455         int err = 0;
 456         int werr = 0;
 457         struct page *page;
 458         struct inode *btree_inode = root->fs_info->btree_inode;
 459         u64 start = 0;
 460         u64 end;
 461         unsigned long index;
 462
 463         while (1) {
 464                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 465                                             mark);
 466                 if (ret)
 467                         break;
 468
 469                 clear_extent_bits(dirty_pages, start, end, mark, GFP_NOFS);
 470                 while (start <= end) {
 471                         index = start >> PAGE_CACHE_SHIFT;
 472                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 473                         page = find_get_page(btree_inode->i_mapping, index);
 474                         if (!page)
 475                                 continue;
 476                         if (PageDirty(page)) {
 477                                 btree_lock_page_hook(page);
 478                                 wait_on_page_writeback(page);
 479                                 err = write_one_page(page, 0);
 480                                 if (err)
 481                                         werr = err;
 482                         }
 483                         wait_on_page_writeback(page);
 484                         page_cache_release(page);
 485                         cond_resched();
 486                 }
 487         }
 488         if (err)
 489                 werr = err;
 490         return werr;
 491 }
 492
 493 /*
 494  * when btree blocks are allocated, they have some corresponding bits set for
 495  * them in one of two extent_io trees.  This is used to make sure all of
 496  * those extents are on disk for transaction or log commit
 497  */
 498 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 499                                 struct extent_io_tree *dirty_pages, int mark)
 500 {
 501         int ret;
 502         int ret2;
 503
 504         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 505         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 506         return ret || ret2;
 507 }
 508
 509 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 510                                      struct btrfs_root *root)
 511 {
 512         if (!trans || !trans->transaction) {
 513                 struct inode *btree_inode;
 514                 btree_inode = root->fs_info->btree_inode;
 515                 return filemap_write_and_wait(btree_inode->i_mapping);
 516         }
 517         return btrfs_write_and_wait_marked_extents(root,
 518                                            &trans->transaction->dirty_pages,
 519                                            EXTENT_DIRTY);
 520 }
 521
 522 /*
 523  * this is used to update the root pointer in the tree of tree roots.
 524  *
 525  * But, in the case of the extent allocation tree, updating the root
 526  * pointer may allocate blocks which may change the root of the extent
 527  * allocation tree.
 528  *
 529  * So, this loops and repeats and makes sure the cowonly root didn't
 530  * change while the root pointer was being updated in the metadata.
 531  */
 532 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 533                                struct btrfs_root *root)
 534 {
 535         int ret;
 536         u64 old_root_bytenr;
 537         u64 old_root_used;
 538         struct btrfs_root *tree_root = root->fs_info->tree_root;
 539
 540         old_root_used = btrfs_root_used(&root->root_item);
 541         btrfs_write_dirty_block_groups(trans, root);
 542
 543         while (1) {
 544                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 545                 if (old_root_bytenr == root->node->start &&
 546                     old_root_used == btrfs_root_used(&root->root_item))
 547                         break;
 548
 549                 btrfs_set_root_node(&root->root_item, root->node);
 550                 ret = btrfs_update_root(trans, tree_root,
 551                                         &root->root_key,
 552                                         &root->root_item);
 553                 BUG_ON(ret);
 554
 555                 old_root_used = btrfs_root_used(&root->root_item);
 556                 ret = btrfs_write_dirty_block_groups(trans, root);
 557                 BUG_ON(ret);
 558         }
 559
 560         if (root != root->fs_info->extent_root)
 561                 switch_commit_root(root);
 562
 563         return 0;
 564 }
 565
 566 /*
 567  * update all the cowonly tree roots on disk
 568  */
 569 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 570                                          struct btrfs_root *root)
 571 {
 572         struct btrfs_fs_info *fs_info = root->fs_info;
 573         struct list_head *next;
 574         struct extent_buffer *eb;
 575         int ret;
 576
 577         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 578         BUG_ON(ret);
 579
 580         eb = btrfs_lock_root_node(fs_info->tree_root);
 581         btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
 582         btrfs_tree_unlock(eb);
 583         free_extent_buffer(eb);
 584
 585         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 586         BUG_ON(ret);
 587
 588         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 589                 next = fs_info->dirty_cowonly_roots.next;
 590                 list_del_init(next);
 591                 root = list_entry(next, struct btrfs_root, dirty_list);
 592
 593                 update_cowonly_root(trans, root);
 594         }
 595
 596         down_write(&fs_info->extent_commit_sem);
 597         switch_commit_root(fs_info->extent_root);
 598         up_write(&fs_info->extent_commit_sem);
 599
 600         return 0;
 601 }
 602
 603 /*
 604  * dead roots are old snapshots that need to be deleted.  This allocates
 605  * a dirty root struct and adds it into the list of dead roots that need to
 606  * be deleted
 607  */
 608 int btrfs_add_dead_root(struct btrfs_root *root)
 609 {
 610         mutex_lock(&root->fs_info->trans_mutex);
 611         list_add(&root->root_list, &root->fs_info->dead_roots);
 612         mutex_unlock(&root->fs_info->trans_mutex);
 613         return 0;
 614 }
 615
 616 /*
 617  * update all the cowonly tree roots on disk
 618  */
 619 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 620                                     struct btrfs_root *root)
 621 {
 622         struct btrfs_root *gang[8];
 623         struct btrfs_fs_info *fs_info = root->fs_info;
 624         int i;
 625         int ret;
 626         int err = 0;
 627
 628         while (1) {
 629                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 630                                                  (void **)gang, 0,
 631                                                  ARRAY_SIZE(gang),
 632                                                  BTRFS_ROOT_TRANS_TAG);
 633                 if (ret == 0)
 634                         break;
 635                 for (i = 0; i < ret; i++) {
 636                         root = gang[i];
 637                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 638                                         (unsigned long)root->root_key.objectid,
 639                                         BTRFS_ROOT_TRANS_TAG);
 640
 641                         btrfs_free_log(trans, root);
 642                         btrfs_update_reloc_root(trans, root);
 643
 644                         if (root->commit_root != root->node) {
 645                                 switch_commit_root(root);
 646                                 btrfs_set_root_node(&root->root_item,
 647                                                     root->node);
 648                         }
 649
 650                         err = btrfs_update_root(trans, fs_info->tree_root,
 651                                                 &root->root_key,
 652                                                 &root->root_item);
 653                         if (err)
 654                                 break;
 655                 }
 656         }
 657         return err;
 658 }
 659
 660 /*
 661  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 662  * otherwise every leaf in the btree is read and defragged.
 663  */
 664 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 665 {
 666         struct btrfs_fs_info *info = root->fs_info;
 667         int ret;
 668         struct btrfs_trans_handle *trans;
 669         unsigned long nr;
 670
 671         smp_mb();
 672         if (root->defrag_running)
 673                 return 0;
 674         trans = btrfs_start_transaction(root, 1);
 675         while (1) {
 676                 root->defrag_running = 1;
 677                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 678                 nr = trans->blocks_used;
 679                 btrfs_end_transaction(trans, root);
 680                 btrfs_btree_balance_dirty(info->tree_root, nr);
 681                 cond_resched();
 682
 683                 trans = btrfs_start_transaction(root, 1);
 684                 if (root->fs_info->closing || ret != -EAGAIN)
 685                         break;
 686         }
 687         root->defrag_running = 0;
 688         smp_mb();
 689         btrfs_end_transaction(trans, root);
 690         return 0;
 691 }
 692
 693 #if 0
 694 /*
 695  * when dropping snapshots, we generate a ton of delayed refs, and it makes
 696  * sense not to join the transaction while it is trying to flush the current
 697  * queue of delayed refs out.
 698  *
 699  * This is used by the drop snapshot code only
 700  */
 701 static noinline int wait_transaction_pre_flush(struct btrfs_fs_info *info)
 702 {
 703         DEFINE_WAIT(wait);
 704
 705         mutex_lock(&info->trans_mutex);
 706         while (info->running_transaction &&
 707                info->running_transaction->delayed_refs.flushing) {
 708                 prepare_to_wait(&info->transaction_wait, &wait,
 709                                 TASK_UNINTERRUPTIBLE);
 710                 mutex_unlock(&info->trans_mutex);
 711
 712                 schedule();
 713
 714                 mutex_lock(&info->trans_mutex);
 715                 finish_wait(&info->transaction_wait, &wait);
 716         }
 717         mutex_unlock(&info->trans_mutex);
 718         return 0;
 719 }
 720
 721 /*
 722  * Given a list of roots that need to be deleted, call btrfs_drop_snapshot on
 723  * all of them
 724  */
 725 int btrfs_drop_dead_root(struct btrfs_root *root)
 726 {
 727         struct btrfs_trans_handle *trans;
 728         struct btrfs_root *tree_root = root->fs_info->tree_root;
 729         unsigned long nr;
 730         int ret;
 731
 732         while (1) {
 733                 /*
 734                  * we don't want to jump in and create a bunch of
 735                  * delayed refs if the transaction is starting to close
 736                  */
 737                 wait_transaction_pre_flush(tree_root->fs_info);
 738                 trans = btrfs_start_transaction(tree_root, 1);
 739
 740                 /*
 741                  * we've joined a transaction, make sure it isn't
 742                  * closing right now
 743                  */
 744                 if (trans->transaction->delayed_refs.flushing) {
 745                         btrfs_end_transaction(trans, tree_root);
 746                         continue;
 747                 }
 748
 749                 ret = btrfs_drop_snapshot(trans, root);
 750                 if (ret != -EAGAIN)
 751                         break;
 752
 753                 ret = btrfs_update_root(trans, tree_root,
 754                                         &root->root_key,
 755                                         &root->root_item);
 756                 if (ret)
 757                         break;
 758
 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
 768         ret = btrfs_del_root(trans, tree_root, &root->root_key);
 769         BUG_ON(ret);
 770
 771         nr = trans->blocks_used;
 772         ret = btrfs_end_transaction(trans, tree_root);
 773         BUG_ON(ret);
 774
 775         free_extent_buffer(root->node);
 776         free_extent_buffer(root->commit_root);
 777         kfree(root);
 778
 779         btrfs_btree_balance_dirty(tree_root, nr);
 780         return ret;
 781 }
 782 #endif
 783
 784 /*
 785  * new snapshots need to be created at a very specific time in the
 786  * transaction commit.  This does the actual creation
 787  */
 788 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 789                                    struct btrfs_fs_info *fs_info,
 790                                    struct btrfs_pending_snapshot *pending)
 791 {
 792         struct btrfs_key key;
 793         struct btrfs_root_item *new_root_item;
 794         struct btrfs_root *tree_root = fs_info->tree_root;
 795         struct btrfs_root *root = pending->root;
 796         struct btrfs_root *parent_root;
 797         struct inode *parent_inode;
 798         struct dentry *dentry;
 799         struct extent_buffer *tmp;
 800         struct extent_buffer *old;
 801         int ret;
 802         u64 index = 0;
 803         u64 objectid;
 804
 805         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 806         if (!new_root_item) {
 807                 pending->error = -ENOMEM;
 808                 goto fail;
 809         }
 810
 811         ret = btrfs_find_free_objectid(trans, tree_root, 0, &objectid);
 812         if (ret) {
 813                 pending->error = ret;
 814                 goto fail;
 815         }
 816
 817         key.objectid = objectid;
 818         key.offset = (u64)-1;
 819         key.type = BTRFS_ROOT_ITEM_KEY;
 820
 821         trans->block_rsv = &pending->block_rsv;
 822
 823         dentry = pending->dentry;
 824         parent_inode = dentry->d_parent->d_inode;
 825         parent_root = BTRFS_I(parent_inode)->root;
 826         record_root_in_trans(trans, parent_root);
 827
 828         /*
 829          * insert the directory item
 830          */
 831         ret = btrfs_set_inode_index(parent_inode, &index);
 832         BUG_ON(ret);
 833         ret = btrfs_insert_dir_item(trans, parent_root,
 834                                 dentry->d_name.name, dentry->d_name.len,
 835                                 parent_inode->i_ino, &key,
 836                                 BTRFS_FT_DIR, index);
 837         BUG_ON(ret);
 838
 839         btrfs_i_size_write(parent_inode, parent_inode->i_size +
 840                                          dentry->d_name.len * 2);
 841         ret = btrfs_update_inode(trans, parent_root, parent_inode);
 842         BUG_ON(ret);
 843
 844         record_root_in_trans(trans, root);
 845         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
 846         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
 847
 848         old = btrfs_lock_root_node(root);
 849         btrfs_cow_block(trans, root, old, NULL, 0, &old);
 850         btrfs_set_lock_blocking(old);
 851
 852         btrfs_copy_root(trans, root, old, &tmp, objectid);
 853         btrfs_tree_unlock(old);
 854         free_extent_buffer(old);
 855
 856         btrfs_set_root_node(new_root_item, tmp);
 857         /* record when the snapshot was created in key.offset */
 858         key.offset = trans->transid;
 859         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
 860         btrfs_tree_unlock(tmp);
 861         free_extent_buffer(tmp);
 862         BUG_ON(ret);
 863
 864         /*
 865          * insert root back/forward references
 866          */
 867         ret = btrfs_add_root_ref(trans, tree_root, objectid,
 868                                  parent_root->root_key.objectid,
 869                                  parent_inode->i_ino, index,
 870                                  dentry->d_name.name, dentry->d_name.len);
 871         BUG_ON(ret);
 872
 873         key.offset = (u64)-1;
 874         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
 875         BUG_ON(IS_ERR(pending->snap));
 876 fail:
 877         kfree(new_root_item);
 878         btrfs_block_rsv_release(root, &pending->block_rsv, (u64)-1);
 879         return 0;
 880 }
 881
 882 /*
 883  * create all the snapshots we've scheduled for creation
 884  */
 885 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
 886                                              struct btrfs_fs_info *fs_info)
 887 {
 888         struct btrfs_pending_snapshot *pending;
 889         struct list_head *head = &trans->transaction->pending_snapshots;
 890         int ret;
 891
 892         list_for_each_entry(pending, head, list) {
 893                 ret = create_pending_snapshot(trans, fs_info, pending);
 894                 BUG_ON(ret);
 895         }
 896         return 0;
 897 }
 898
 899 static void update_super_roots(struct btrfs_root *root)
 900 {
 901         struct btrfs_root_item *root_item;
 902         struct btrfs_super_block *super;
 903
 904         super = &root->fs_info->super_copy;
 905
 906         root_item = &root->fs_info->chunk_root->root_item;
 907         super->chunk_root = root_item->bytenr;
 908         super->chunk_root_generation = root_item->generation;
 909         super->chunk_root_level = root_item->level;
 910
 911         root_item = &root->fs_info->tree_root->root_item;
 912         super->root = root_item->bytenr;
 913         super->generation = root_item->generation;
 914         super->root_level = root_item->level;
 915 }
 916
 917 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
 918 {
 919         int ret = 0;
 920         spin_lock(&info->new_trans_lock);
 921         if (info->running_transaction)
 922                 ret = info->running_transaction->in_commit;
 923         spin_unlock(&info->new_trans_lock);
 924         return ret;
 925 }
 926
 927 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
 928                              struct btrfs_root *root)
 929 {
 930         unsigned long joined = 0;
 931         unsigned long timeout = 1;
 932         struct btrfs_transaction *cur_trans;
 933         struct btrfs_transaction *prev_trans = NULL;
 934         DEFINE_WAIT(wait);
 935         int ret;
 936         int should_grow = 0;
 937         unsigned long now = get_seconds();
 938         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
 939
 940         btrfs_run_ordered_operations(root, 0);
 941
 942         /* make a pass through all the delayed refs we have so far
 943          * any runnings procs may add more while we are here
 944          */
 945         ret = btrfs_run_delayed_refs(trans, root, 0);
 946         BUG_ON(ret);
 947
 948         btrfs_trans_release_metadata(trans, root);
 949
 950         cur_trans = trans->transaction;
 951         /*
 952          * set the flushing flag so procs in this transaction have to
 953          * start sending their work down.
 954          */
 955         cur_trans->delayed_refs.flushing = 1;
 956
 957         ret = btrfs_run_delayed_refs(trans, root, 0);
 958         BUG_ON(ret);
 959
 960         mutex_lock(&root->fs_info->trans_mutex);
 961         if (cur_trans->in_commit) {
 962                 cur_trans->use_count++;
 963                 mutex_unlock(&root->fs_info->trans_mutex);
 964                 btrfs_end_transaction(trans, root);
 965
 966                 ret = wait_for_commit(root, cur_trans);
 967                 BUG_ON(ret);
 968
 969                 mutex_lock(&root->fs_info->trans_mutex);
 970                 put_transaction(cur_trans);
 971                 mutex_unlock(&root->fs_info->trans_mutex);
 972
 973                 return 0;
 974         }
 975
 976         trans->transaction->in_commit = 1;
 977         trans->transaction->blocked = 1;
 978         if (cur_trans->list.prev != &root->fs_info->trans_list) {
 979                 prev_trans = list_entry(cur_trans->list.prev,
 980                                         struct btrfs_transaction, list);
 981                 if (!prev_trans->commit_done) {
 982                         prev_trans->use_count++;
 983                         mutex_unlock(&root->fs_info->trans_mutex);
 984
 985                         wait_for_commit(root, prev_trans);
 986
 987                         mutex_lock(&root->fs_info->trans_mutex);
 988                         put_transaction(prev_trans);
 989                 }
 990         }
 991
 992         if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
 993                 should_grow = 1;
 994
 995         do {
 996                 int snap_pending = 0;
 997                 joined = cur_trans->num_joined;
 998                 if (!list_empty(&trans->transaction->pending_snapshots))
 999                         snap_pending = 1;
1000
1001                 WARN_ON(cur_trans != trans->transaction);
1002                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1003                                 TASK_UNINTERRUPTIBLE);
1004
1005                 if (cur_trans->num_writers > 1)
1006                         timeout = MAX_SCHEDULE_TIMEOUT;
1007                 else if (should_grow)
1008                         timeout = 1;
1009
1010                 mutex_unlock(&root->fs_info->trans_mutex);
1011
1012                 if (flush_on_commit || snap_pending) {
1013                         btrfs_start_delalloc_inodes(root, 1);
1014                         ret = btrfs_wait_ordered_extents(root, 0, 1);
1015                         BUG_ON(ret);
1016                 }
1017
1018                 /*
1019                  * rename don't use btrfs_join_transaction, so, once we
1020                  * set the transaction to blocked above, we aren't going
1021                  * to get any new ordered operations.  We can safely run
1022                  * it here and no for sure that nothing new will be added
1023                  * to the list
1024                  */
1025                 btrfs_run_ordered_operations(root, 1);
1026
1027                 smp_mb();
1028                 if (cur_trans->num_writers > 1 || should_grow)
1029                         schedule_timeout(timeout);
1030
1031                 mutex_lock(&root->fs_info->trans_mutex);
1032                 finish_wait(&cur_trans->writer_wait, &wait);
1033         } while (cur_trans->num_writers > 1 ||
1034                  (should_grow && cur_trans->num_joined != joined));
1035
1036         ret = create_pending_snapshots(trans, root->fs_info);
1037         BUG_ON(ret);
1038
1039         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1040         BUG_ON(ret);
1041
1042         WARN_ON(cur_trans != trans->transaction);
1043
1044         /* btrfs_commit_tree_roots is responsible for getting the
1045          * various roots consistent with each other.  Every pointer
1046          * in the tree of tree roots has to point to the most up to date
1047          * root for every subvolume and other tree.  So, we have to keep
1048          * the tree logging code from jumping in and changing any
1049          * of the trees.
1050          *
1051          * At this point in the commit, there can't be any tree-log
1052          * writers, but a little lower down we drop the trans mutex
1053          * and let new people in.  By holding the tree_log_mutex
1054          * from now until after the super is written, we avoid races
1055          * with the tree-log code.
1056          */
1057         mutex_lock(&root->fs_info->tree_log_mutex);
1058
1059         ret = commit_fs_roots(trans, root);
1060         BUG_ON(ret);
1061
1062         /* commit_fs_roots gets rid of all the tree log roots, it is now
1063          * safe to free the root of tree log roots
1064          */
1065         btrfs_free_log_root_tree(trans, root->fs_info);
1066
1067         ret = commit_cowonly_roots(trans, root);
1068         BUG_ON(ret);
1069
1070         btrfs_prepare_extent_commit(trans, root);
1071
1072         cur_trans = root->fs_info->running_transaction;
1073         spin_lock(&root->fs_info->new_trans_lock);
1074         root->fs_info->running_transaction = NULL;
1075         spin_unlock(&root->fs_info->new_trans_lock);
1076
1077         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1078                             root->fs_info->tree_root->node);
1079         switch_commit_root(root->fs_info->tree_root);
1080
1081         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1082                             root->fs_info->chunk_root->node);
1083         switch_commit_root(root->fs_info->chunk_root);
1084
1085         update_super_roots(root);
1086
1087         if (!root->fs_info->log_root_recovering) {
1088                 btrfs_set_super_log_root(&root->fs_info->super_copy, 0);
1089                 btrfs_set_super_log_root_level(&root->fs_info->super_copy, 0);
1090         }
1091
1092         memcpy(&root->fs_info->super_for_commit, &root->fs_info->super_copy,
1093                sizeof(root->fs_info->super_copy));
1094
1095         trans->transaction->blocked = 0;
1096
1097         wake_up(&root->fs_info->transaction_wait);
1098
1099         mutex_unlock(&root->fs_info->trans_mutex);
1100         ret = btrfs_write_and_wait_transaction(trans, root);
1101         BUG_ON(ret);
1102         write_ctree_super(trans, root, 0);
1103
1104         /*
1105          * the super is written, we can safely allow the tree-loggers
1106          * to go about their business
1107          */
1108         mutex_unlock(&root->fs_info->tree_log_mutex);
1109
1110         btrfs_finish_extent_commit(trans, root);
1111
1112         mutex_lock(&root->fs_info->trans_mutex);
1113
1114         cur_trans->commit_done = 1;
1115
1116         root->fs_info->last_trans_committed = cur_trans->transid;
1117
1118         wake_up(&cur_trans->commit_wait);
1119
1120         put_transaction(cur_trans);
1121         put_transaction(cur_trans);
1122
1123         mutex_unlock(&root->fs_info->trans_mutex);
1124
1125         if (current->journal_info == trans)
1126                 current->journal_info = NULL;
1127
1128         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1129
1130         if (current != root->fs_info->transaction_kthread)
1131                 btrfs_run_delayed_iputs(root);
1132
1133         return ret;
1134 }
1135
1136 /*
1137  * interface function to delete all the snapshots we have scheduled for deletion
1138  */
1139 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1140 {
1141         LIST_HEAD(list);
1142         struct btrfs_fs_info *fs_info = root->fs_info;
1143
1144         mutex_lock(&fs_info->trans_mutex);
1145         list_splice_init(&fs_info->dead_roots, &list);
1146         mutex_unlock(&fs_info->trans_mutex);
1147
1148         while (!list_empty(&list)) {
1149                 root = list_entry(list.next, struct btrfs_root, root_list);
1150                 list_del(&root->root_list);
1151
1152                 if (btrfs_header_backref_rev(root->node) <
1153                     BTRFS_MIXED_BACKREF_REV)
1154                         btrfs_drop_snapshot(root, 0);
1155                 else
1156                         btrfs_drop_snapshot(root, 1);
1157         }
1158         return 0;
1159 }