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                                 mutex_unlock(&root->fs_info->trans_mutex);
 153                                 schedule();
 154                                 mutex_lock(&root->fs_info->trans_mutex);
 155                                 finish_wait(&root->fs_info->transaction_wait,
 156                                             &wait);
 157                         } else {
 158                                 finish_wait(&root->fs_info->transaction_wait,
 159                                             &wait);
 160                                 break;
 161                         }
 162                 }
 163                 put_transaction(cur_trans);
 164         }
 165 }
 166
 167 enum btrfs_trans_type {
 168         TRANS_START,
 169         TRANS_JOIN,
 170         TRANS_USERSPACE,
 171 };
 172
 173 static struct btrfs_trans_handle *start_transaction(struct btrfs_root *root,
 174                                              int num_blocks, int type)
 175 {
 176         struct btrfs_trans_handle *h =
 177                 kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 178         int ret;
 179
 180         mutex_lock(&root->fs_info->trans_mutex);
 181         if (!root->fs_info->log_root_recovering &&
 182             ((type == TRANS_START && !root->fs_info->open_ioctl_trans) ||
 183              type == TRANS_USERSPACE))
 184                 wait_current_trans(root);
 185         ret = join_transaction(root);
 186         BUG_ON(ret);
 187
 188         h->transid = root->fs_info->running_transaction->transid;
 189         h->transaction = root->fs_info->running_transaction;
 190         h->blocks_reserved = num_blocks;
 191         h->blocks_used = 0;
 192         h->block_group = 0;
 193         h->alloc_exclude_nr = 0;
 194         h->alloc_exclude_start = 0;
 195         h->delayed_ref_updates = 0;
 196
 197         if (!current->journal_info && type != TRANS_USERSPACE)
 198                 current->journal_info = h;
 199
 200         root->fs_info->running_transaction->use_count++;
 201         record_root_in_trans(h, root);
 202         mutex_unlock(&root->fs_info->trans_mutex);
 203         return h;
 204 }
 205
 206 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 207                                                    int num_blocks)
 208 {
 209         return start_transaction(root, num_blocks, TRANS_START);
 210 }
 211 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root,
 212                                                    int num_blocks)
 213 {
 214         return start_transaction(root, num_blocks, TRANS_JOIN);
 215 }
 216
 217 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *r,
 218                                                          int num_blocks)
 219 {
 220         return start_transaction(r, num_blocks, TRANS_USERSPACE);
 221 }
 222
 223 /* wait for a transaction commit to be fully complete */
 224 static noinline int wait_for_commit(struct btrfs_root *root,
 225                                     struct btrfs_transaction *commit)
 226 {
 227         DEFINE_WAIT(wait);
 228         mutex_lock(&root->fs_info->trans_mutex);
 229         while (!commit->commit_done) {
 230                 prepare_to_wait(&commit->commit_wait, &wait,
 231                                 TASK_UNINTERRUPTIBLE);
 232                 if (commit->commit_done)
 233                         break;
 234                 mutex_unlock(&root->fs_info->trans_mutex);
 235                 schedule();
 236                 mutex_lock(&root->fs_info->trans_mutex);
 237         }
 238         mutex_unlock(&root->fs_info->trans_mutex);
 239         finish_wait(&commit->commit_wait, &wait);
 240         return 0;
 241 }
 242
 243 #if 0
 244 /*
 245  * rate limit against the drop_snapshot code.  This helps to slow down new
 246  * operations if the drop_snapshot code isn't able to keep up.
 247  */
 248 static void throttle_on_drops(struct btrfs_root *root)
 249 {
 250         struct btrfs_fs_info *info = root->fs_info;
 251         int harder_count = 0;
 252
 253 harder:
 254         if (atomic_read(&info->throttles)) {
 255                 DEFINE_WAIT(wait);
 256                 int thr;
 257                 thr = atomic_read(&info->throttle_gen);
 258
 259                 do {
 260                         prepare_to_wait(&info->transaction_throttle,
 261                                         &wait, TASK_UNINTERRUPTIBLE);
 262                         if (!atomic_read(&info->throttles)) {
 263                                 finish_wait(&info->transaction_throttle, &wait);
 264                                 break;
 265                         }
 266                         schedule();
 267                         finish_wait(&info->transaction_throttle, &wait);
 268                 } while (thr == atomic_read(&info->throttle_gen));
 269                 harder_count++;
 270
 271                 if (root->fs_info->total_ref_cache_size > 1 * 1024 * 1024 &&
 272                     harder_count < 2)
 273                         goto harder;
 274
 275                 if (root->fs_info->total_ref_cache_size > 5 * 1024 * 1024 &&
 276                     harder_count < 10)
 277                         goto harder;
 278
 279                 if (root->fs_info->total_ref_cache_size > 10 * 1024 * 1024 &&
 280                     harder_count < 20)
 281                         goto harder;
 282         }
 283 }
 284 #endif
 285
 286 void btrfs_throttle(struct btrfs_root *root)
 287 {
 288         mutex_lock(&root->fs_info->trans_mutex);
 289         if (!root->fs_info->open_ioctl_trans)
 290                 wait_current_trans(root);
 291         mutex_unlock(&root->fs_info->trans_mutex);
 292 }
 293
 294 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 295                           struct btrfs_root *root, int throttle)
 296 {
 297         struct btrfs_transaction *cur_trans;
 298         struct btrfs_fs_info *info = root->fs_info;
 299         int count = 0;
 300
 301         while (count < 4) {
 302                 unsigned long cur = trans->delayed_ref_updates;
 303                 trans->delayed_ref_updates = 0;
 304                 if (cur &&
 305                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 306                         trans->delayed_ref_updates = 0;
 307
 308                         /*
 309                          * do a full flush if the transaction is trying
 310                          * to close
 311                          */
 312                         if (trans->transaction->delayed_refs.flushing)
 313                                 cur = 0;
 314                         btrfs_run_delayed_refs(trans, root, cur);
 315                 } else {
 316                         break;
 317                 }
 318                 count++;
 319         }
 320
 321         mutex_lock(&info->trans_mutex);
 322         cur_trans = info->running_transaction;
 323         WARN_ON(cur_trans != trans->transaction);
 324         WARN_ON(cur_trans->num_writers < 1);
 325         cur_trans->num_writers--;
 326
 327         if (waitqueue_active(&cur_trans->writer_wait))
 328                 wake_up(&cur_trans->writer_wait);
 329         put_transaction(cur_trans);
 330         mutex_unlock(&info->trans_mutex);
 331
 332         if (current->journal_info == trans)
 333                 current->journal_info = NULL;
 334         memset(trans, 0, sizeof(*trans));
 335         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 336
 337         if (throttle)
 338                 btrfs_run_delayed_iputs(root);
 339
 340         return 0;
 341 }
 342
 343 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 344                           struct btrfs_root *root)
 345 {
 346         return __btrfs_end_transaction(trans, root, 0);
 347 }
 348
 349 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 350                                    struct btrfs_root *root)
 351 {
 352         return __btrfs_end_transaction(trans, root, 1);
 353 }
 354
 355 /*
 356  * when btree blocks are allocated, they have some corresponding bits set for
 357  * them in one of two extent_io trees.  This is used to make sure all of
 358  * those extents are sent to disk but does not wait on them
 359  */
 360 int btrfs_write_marked_extents(struct btrfs_root *root,
 361                                struct extent_io_tree *dirty_pages, int mark)
 362 {
 363         int ret;
 364         int err = 0;
 365         int werr = 0;
 366         struct page *page;
 367         struct inode *btree_inode = root->fs_info->btree_inode;
 368         u64 start = 0;
 369         u64 end;
 370         unsigned long index;
 371
 372         while (1) {
 373                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 374                                             mark);
 375                 if (ret)
 376                         break;
 377                 while (start <= end) {
 378                         cond_resched();
 379
 380                         index = start >> PAGE_CACHE_SHIFT;
 381                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 382                         page = find_get_page(btree_inode->i_mapping, index);
 383                         if (!page)
 384                                 continue;
 385
 386                         btree_lock_page_hook(page);
 387                         if (!page->mapping) {
 388                                 unlock_page(page);
 389                                 page_cache_release(page);
 390                                 continue;
 391                         }
 392
 393                         if (PageWriteback(page)) {
 394                                 if (PageDirty(page))
 395                                         wait_on_page_writeback(page);
 396                                 else {
 397                                         unlock_page(page);
 398                                         page_cache_release(page);
 399                                         continue;
 400                                 }
 401                         }
 402                         err = write_one_page(page, 0);
 403                         if (err)
 404                                 werr = err;
 405                         page_cache_release(page);
 406                 }
 407         }
 408         if (err)
 409                 werr = err;
 410         return werr;
 411 }
 412
 413 /*
 414  * when btree blocks are allocated, they have some corresponding bits set for
 415  * them in one of two extent_io trees.  This is used to make sure all of
 416  * those extents are on disk for transaction or log commit.  We wait
 417  * on all the pages and clear them from the dirty pages state tree
 418  */
 419 int btrfs_wait_marked_extents(struct btrfs_root *root,
 420                               struct extent_io_tree *dirty_pages, int mark)
 421 {
 422         int ret;
 423         int err = 0;
 424         int werr = 0;
 425         struct page *page;
 426         struct inode *btree_inode = root->fs_info->btree_inode;
 427         u64 start = 0;
 428         u64 end;
 429         unsigned long index;
 430
 431         while (1) {
 432                 ret = find_first_extent_bit(dirty_pages, start, &start, &end,
 433                                             mark);
 434                 if (ret)
 435                         break;
 436
 437                 clear_extent_bits(dirty_pages, start, end, mark, GFP_NOFS);
 438                 while (start <= end) {
 439                         index = start >> PAGE_CACHE_SHIFT;
 440                         start = (u64)(index + 1) << PAGE_CACHE_SHIFT;
 441                         page = find_get_page(btree_inode->i_mapping, index);
 442                         if (!page)
 443                                 continue;
 444                         if (PageDirty(page)) {
 445                                 btree_lock_page_hook(page);
 446                                 wait_on_page_writeback(page);
 447                                 err = write_one_page(page, 0);
 448                                 if (err)
 449                                         werr = err;
 450                         }
 451                         wait_on_page_writeback(page);
 452                         page_cache_release(page);
 453                         cond_resched();
 454                 }
 455         }
 456         if (err)
 457                 werr = err;
 458         return werr;
 459 }
 460
 461 /*
 462  * when btree blocks are allocated, they have some corresponding bits set for
 463  * them in one of two extent_io trees.  This is used to make sure all of
 464  * those extents are on disk for transaction or log commit
 465  */
 466 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 467                                 struct extent_io_tree *dirty_pages, int mark)
 468 {
 469         int ret;
 470         int ret2;
 471
 472         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 473         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 474         return ret || ret2;
 475 }
 476
 477 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 478                                      struct btrfs_root *root)
 479 {
 480         if (!trans || !trans->transaction) {
 481                 struct inode *btree_inode;
 482                 btree_inode = root->fs_info->btree_inode;
 483                 return filemap_write_and_wait(btree_inode->i_mapping);
 484         }
 485         return btrfs_write_and_wait_marked_extents(root,
 486                                            &trans->transaction->dirty_pages,
 487                                            EXTENT_DIRTY);
 488 }
 489
 490 /*
 491  * this is used to update the root pointer in the tree of tree roots.
 492  *
 493  * But, in the case of the extent allocation tree, updating the root
 494  * pointer may allocate blocks which may change the root of the extent
 495  * allocation tree.
 496  *
 497  * So, this loops and repeats and makes sure the cowonly root didn't
 498  * change while the root pointer was being updated in the metadata.
 499  */
 500 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 501                                struct btrfs_root *root)
 502 {
 503         int ret;
 504         u64 old_root_bytenr;
 505         u64 old_root_used;
 506         struct btrfs_root *tree_root = root->fs_info->tree_root;
 507
 508         old_root_used = btrfs_root_used(&root->root_item);
 509         btrfs_write_dirty_block_groups(trans, root);
 510
 511         while (1) {
 512                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 513                 if (old_root_bytenr == root->node->start &&
 514                     old_root_used == btrfs_root_used(&root->root_item))
 515                         break;
 516
 517                 btrfs_set_root_node(&root->root_item, root->node);
 518                 ret = btrfs_update_root(trans, tree_root,
 519                                         &root->root_key,
 520                                         &root->root_item);
 521                 BUG_ON(ret);
 522
 523                 old_root_used = btrfs_root_used(&root->root_item);
 524                 ret = btrfs_write_dirty_block_groups(trans, root);
 525                 BUG_ON(ret);
 526         }
 527
 528         if (root != root->fs_info->extent_root)
 529                 switch_commit_root(root);
 530
 531         return 0;
 532 }
 533
 534 /*
 535  * update all the cowonly tree roots on disk
 536  */
 537 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 538                                          struct btrfs_root *root)
 539 {
 540         struct btrfs_fs_info *fs_info = root->fs_info;
 541         struct list_head *next;
 542         struct extent_buffer *eb;
 543         int ret;
 544
 545         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 546         BUG_ON(ret);
 547
 548         eb = btrfs_lock_root_node(fs_info->tree_root);
 549         btrfs_cow_block(trans, fs_info->tree_root, eb, NULL, 0, &eb);
 550         btrfs_tree_unlock(eb);
 551         free_extent_buffer(eb);
 552
 553         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 554         BUG_ON(ret);
 555
 556         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 557                 next = fs_info->dirty_cowonly_roots.next;
 558                 list_del_init(next);
 559                 root = list_entry(next, struct btrfs_root, dirty_list);
 560
 561                 update_cowonly_root(trans, root);
 562         }
 563
 564         down_write(&fs_info->extent_commit_sem);
 565         switch_commit_root(fs_info->extent_root);
 566         up_write(&fs_info->extent_commit_sem);
 567
 568         return 0;
 569 }
 570
 571 /*
 572  * dead roots are old snapshots that need to be deleted.  This allocates
 573  * a dirty root struct and adds it into the list of dead roots that need to
 574  * be deleted
 575  */
 576 int btrfs_add_dead_root(struct btrfs_root *root)
 577 {
 578         mutex_lock(&root->fs_info->trans_mutex);
 579         list_add(&root->root_list, &root->fs_info->dead_roots);
 580         mutex_unlock(&root->fs_info->trans_mutex);
 581         return 0;
 582 }
 583
 584 /*
 585  * update all the cowonly tree roots on disk
 586  */
 587 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 588                                     struct btrfs_root *root)
 589 {
 590         struct btrfs_root *gang[8];
 591         struct btrfs_fs_info *fs_info = root->fs_info;
 592         int i;
 593         int ret;
 594         int err = 0;
 595
 596         while (1) {
 597                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 598                                                  (void **)gang, 0,
 599                                                  ARRAY_SIZE(gang),
 600                                                  BTRFS_ROOT_TRANS_TAG);
 601                 if (ret == 0)
 602                         break;
 603                 for (i = 0; i < ret; i++) {
 604                         root = gang[i];
 605                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 606                                         (unsigned long)root->root_key.objectid,
 607                                         BTRFS_ROOT_TRANS_TAG);
 608
 609                         btrfs_free_log(trans, root);
 610                         btrfs_update_reloc_root(trans, root);
 611
 612                         if (root->commit_root != root->node) {
 613                                 switch_commit_root(root);
 614                                 btrfs_set_root_node(&root->root_item,
 615                                                     root->node);
 616                         }
 617
 618                         err = btrfs_update_root(trans, fs_info->tree_root,
 619                                                 &root->root_key,
 620                                                 &root->root_item);
 621                         if (err)
 622                                 break;
 623                 }
 624         }
 625         return err;
 626 }
 627
 628 /*
 629  * defrag a given btree.  If cacheonly == 1, this won't read from the disk,
 630  * otherwise every leaf in the btree is read and defragged.
 631  */
 632 int btrfs_defrag_root(struct btrfs_root *root, int cacheonly)
 633 {
 634         struct btrfs_fs_info *info = root->fs_info;
 635         int ret;
 636         struct btrfs_trans_handle *trans;
 637         unsigned long nr;
 638
 639         smp_mb();
 640         if (root->defrag_running)
 641                 return 0;
 642         trans = btrfs_start_transaction(root, 1);
 643         while (1) {
 644                 root->defrag_running = 1;
 645                 ret = btrfs_defrag_leaves(trans, root, cacheonly);
 646                 nr = trans->blocks_used;
 647                 btrfs_end_transaction(trans, root);
 648                 btrfs_btree_balance_dirty(info->tree_root, nr);
 649                 cond_resched();
 650
 651                 trans = btrfs_start_transaction(root, 1);
 652                 if (root->fs_info->closing || ret != -EAGAIN)
 653                         break;
 654         }
 655         root->defrag_running = 0;
 656         smp_mb();
 657         btrfs_end_transaction(trans, root);
 658         return 0;
 659 }
 660
 661 #if 0
 662 /*
 663  * when dropping snapshots, we generate a ton of delayed refs, and it makes
 664  * sense not to join the transaction while it is trying to flush the current
 665  * queue of delayed refs out.
 666  *
 667  * This is used by the drop snapshot code only
 668  */
 669 static noinline int wait_transaction_pre_flush(struct btrfs_fs_info *info)
 670 {
 671         DEFINE_WAIT(wait);
 672
 673         mutex_lock(&info->trans_mutex);
 674         while (info->running_transaction &&
 675                info->running_transaction->delayed_refs.flushing) {
 676                 prepare_to_wait(&info->transaction_wait, &wait,
 677                                 TASK_UNINTERRUPTIBLE);
 678                 mutex_unlock(&info->trans_mutex);
 679
 680                 schedule();
 681
 682                 mutex_lock(&info->trans_mutex);
 683                 finish_wait(&info->transaction_wait, &wait);
 684         }
 685         mutex_unlock(&info->trans_mutex);
 686         return 0;
 687 }
 688
 689 /*
 690  * Given a list of roots that need to be deleted, call btrfs_drop_snapshot on
 691  * all of them
 692  */
 693 int btrfs_drop_dead_root(struct btrfs_root *root)
 694 {
 695         struct btrfs_trans_handle *trans;
 696         struct btrfs_root *tree_root = root->fs_info->tree_root;
 697         unsigned long nr;
 698         int ret;
 699
 700         while (1) {
 701                 /*
 702                  * we don't want to jump in and create a bunch of
 703                  * delayed refs if the transaction is starting to close
 704                  */
 705                 wait_transaction_pre_flush(tree_root->fs_info);
 706                 trans = btrfs_start_transaction(tree_root, 1);
 707
 708                 /*
 709                  * we've joined a transaction, make sure it isn't
 710                  * closing right now
 711                  */
 712                 if (trans->transaction->delayed_refs.flushing) {
 713                         btrfs_end_transaction(trans, tree_root);
 714                         continue;
 715                 }
 716
 717                 ret = btrfs_drop_snapshot(trans, root);
 718                 if (ret != -EAGAIN)
 719                         break;
 720
 721                 ret = btrfs_update_root(trans, tree_root,
 722                                         &root->root_key,
 723                                         &root->root_item);
 724                 if (ret)
 725                         break;
 726
 727                 nr = trans->blocks_used;
 728                 ret = btrfs_end_transaction(trans, tree_root);
 729                 BUG_ON(ret);
 730
 731                 btrfs_btree_balance_dirty(tree_root, nr);
 732                 cond_resched();
 733         }
 734         BUG_ON(ret);
 735
 736         ret = btrfs_del_root(trans, tree_root, &root->root_key);
 737         BUG_ON(ret);
 738
 739         nr = trans->blocks_used;
 740         ret = btrfs_end_transaction(trans, tree_root);
 741         BUG_ON(ret);
 742
 743         free_extent_buffer(root->node);
 744         free_extent_buffer(root->commit_root);
 745         kfree(root);
 746
 747         btrfs_btree_balance_dirty(tree_root, nr);
 748         return ret;
 749 }
 750 #endif
 751
 752 /*
 753  * new snapshots need to be created at a very specific time in the
 754  * transaction commit.  This does the actual creation
 755  */
 756 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
 757                                    struct btrfs_fs_info *fs_info,
 758                                    struct btrfs_pending_snapshot *pending)
 759 {
 760         struct btrfs_key key;
 761         struct btrfs_root_item *new_root_item;
 762         struct btrfs_root *tree_root = fs_info->tree_root;
 763         struct btrfs_root *root = pending->root;
 764         struct extent_buffer *tmp;
 765         struct extent_buffer *old;
 766         int ret;
 767         u64 objectid;
 768
 769         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
 770         if (!new_root_item) {
 771                 ret = -ENOMEM;
 772                 goto fail;
 773         }
 774         ret = btrfs_find_free_objectid(trans, tree_root, 0, &objectid);
 775         if (ret)
 776                 goto fail;
 777
 778         record_root_in_trans(trans, root);
 779         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
 780         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
 781
 782         key.objectid = objectid;
 783         /* record when the snapshot was created in key.offset */
 784         key.offset = trans->transid;
 785         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
 786
 787         old = btrfs_lock_root_node(root);
 788         btrfs_cow_block(trans, root, old, NULL, 0, &old);
 789         btrfs_set_lock_blocking(old);
 790
 791         btrfs_copy_root(trans, root, old, &tmp, objectid);
 792         btrfs_tree_unlock(old);
 793         free_extent_buffer(old);
 794
 795         btrfs_set_root_node(new_root_item, tmp);
 796         ret = btrfs_insert_root(trans, root->fs_info->tree_root, &key,
 797                                 new_root_item);
 798         btrfs_tree_unlock(tmp);
 799         free_extent_buffer(tmp);
 800         if (ret)
 801                 goto fail;
 802
 803         key.offset = (u64)-1;
 804         memcpy(&pending->root_key, &key, sizeof(key));
 805 fail:
 806         kfree(new_root_item);
 807         return ret;
 808 }
 809
 810 static noinline int finish_pending_snapshot(struct btrfs_fs_info *fs_info,
 811                                    struct btrfs_pending_snapshot *pending)
 812 {
 813         int ret;
 814         int namelen;
 815         u64 index = 0;
 816         struct btrfs_trans_handle *trans;
 817         struct inode *parent_inode;
 818         struct btrfs_root *parent_root;
 819
 820         parent_inode = pending->dentry->d_parent->d_inode;
 821         parent_root = BTRFS_I(parent_inode)->root;
 822         trans = btrfs_join_transaction(parent_root, 1);
 823
 824         /*
 825          * insert the directory item
 826          */
 827         namelen = strlen(pending->name);
 828         ret = btrfs_set_inode_index(parent_inode, &index);
 829         ret = btrfs_insert_dir_item(trans, parent_root,
 830                             pending->name, namelen,
 831                             parent_inode->i_ino,
 832                             &pending->root_key, BTRFS_FT_DIR, index);
 833
 834         if (ret)
 835                 goto fail;
 836
 837         btrfs_i_size_write(parent_inode, parent_inode->i_size + namelen * 2);
 838         ret = btrfs_update_inode(trans, parent_root, parent_inode);
 839         BUG_ON(ret);
 840
 841         ret = btrfs_add_root_ref(trans, parent_root->fs_info->tree_root,
 842                                  pending->root_key.objectid,
 843                                  parent_root->root_key.objectid,
 844                                  parent_inode->i_ino, index, pending->name,
 845                                  namelen);
 846
 847         BUG_ON(ret);
 848
 849 fail:
 850         btrfs_end_transaction(trans, fs_info->fs_root);
 851         return ret;
 852 }
 853
 854 /*
 855  * create all the snapshots we've scheduled for creation
 856  */
 857 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
 858                                              struct btrfs_fs_info *fs_info)
 859 {
 860         struct btrfs_pending_snapshot *pending;
 861         struct list_head *head = &trans->transaction->pending_snapshots;
 862         int ret;
 863
 864         list_for_each_entry(pending, head, list) {
 865                 ret = create_pending_snapshot(trans, fs_info, pending);
 866                 BUG_ON(ret);
 867         }
 868         return 0;
 869 }
 870
 871 static noinline int finish_pending_snapshots(struct btrfs_trans_handle *trans,
 872                                              struct btrfs_fs_info *fs_info)
 873 {
 874         struct btrfs_pending_snapshot *pending;
 875         struct list_head *head = &trans->transaction->pending_snapshots;
 876         int ret;
 877
 878         while (!list_empty(head)) {
 879                 pending = list_entry(head->next,
 880                                      struct btrfs_pending_snapshot, list);
 881                 ret = finish_pending_snapshot(fs_info, pending);
 882                 BUG_ON(ret);
 883                 list_del(&pending->list);
 884                 kfree(pending->name);
 885                 kfree(pending);
 886         }
 887         return 0;
 888 }
 889
 890 static void update_super_roots(struct btrfs_root *root)
 891 {
 892         struct btrfs_root_item *root_item;
 893         struct btrfs_super_block *super;
 894
 895         super = &root->fs_info->super_copy;
 896
 897         root_item = &root->fs_info->chunk_root->root_item;
 898         super->chunk_root = root_item->bytenr;
 899         super->chunk_root_generation = root_item->generation;
 900         super->chunk_root_level = root_item->level;
 901
 902         root_item = &root->fs_info->tree_root->root_item;
 903         super->root = root_item->bytenr;
 904         super->generation = root_item->generation;
 905         super->root_level = root_item->level;
 906 }
 907
 908 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
 909 {
 910         int ret = 0;
 911         spin_lock(&info->new_trans_lock);
 912         if (info->running_transaction)
 913                 ret = info->running_transaction->in_commit;
 914         spin_unlock(&info->new_trans_lock);
 915         return ret;
 916 }
 917
 918 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
 919                              struct btrfs_root *root)
 920 {
 921         unsigned long joined = 0;
 922         unsigned long timeout = 1;
 923         struct btrfs_transaction *cur_trans;
 924         struct btrfs_transaction *prev_trans = NULL;
 925         DEFINE_WAIT(wait);
 926         int ret;
 927         int should_grow = 0;
 928         unsigned long now = get_seconds();
 929         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
 930
 931         btrfs_run_ordered_operations(root, 0);
 932
 933         /* make a pass through all the delayed refs we have so far
 934          * any runnings procs may add more while we are here
 935          */
 936         ret = btrfs_run_delayed_refs(trans, root, 0);
 937         BUG_ON(ret);
 938
 939         cur_trans = trans->transaction;
 940         /*
 941          * set the flushing flag so procs in this transaction have to
 942          * start sending their work down.
 943          */
 944         cur_trans->delayed_refs.flushing = 1;
 945
 946         ret = btrfs_run_delayed_refs(trans, root, 0);
 947         BUG_ON(ret);
 948
 949         mutex_lock(&root->fs_info->trans_mutex);
 950         if (cur_trans->in_commit) {
 951                 cur_trans->use_count++;
 952                 mutex_unlock(&root->fs_info->trans_mutex);
 953                 btrfs_end_transaction(trans, root);
 954
 955                 ret = wait_for_commit(root, cur_trans);
 956                 BUG_ON(ret);
 957
 958                 mutex_lock(&root->fs_info->trans_mutex);
 959                 put_transaction(cur_trans);
 960                 mutex_unlock(&root->fs_info->trans_mutex);
 961
 962                 return 0;
 963         }
 964
 965         trans->transaction->in_commit = 1;
 966         trans->transaction->blocked = 1;
 967         if (cur_trans->list.prev != &root->fs_info->trans_list) {
 968                 prev_trans = list_entry(cur_trans->list.prev,
 969                                         struct btrfs_transaction, list);
 970                 if (!prev_trans->commit_done) {
 971                         prev_trans->use_count++;
 972                         mutex_unlock(&root->fs_info->trans_mutex);
 973
 974                         wait_for_commit(root, prev_trans);
 975
 976                         mutex_lock(&root->fs_info->trans_mutex);
 977                         put_transaction(prev_trans);
 978                 }
 979         }
 980
 981         if (now < cur_trans->start_time || now - cur_trans->start_time < 1)
 982                 should_grow = 1;
 983
 984         do {
 985                 int snap_pending = 0;
 986                 joined = cur_trans->num_joined;
 987                 if (!list_empty(&trans->transaction->pending_snapshots))
 988                         snap_pending = 1;
 989
 990                 WARN_ON(cur_trans != trans->transaction);
 991                 prepare_to_wait(&cur_trans->writer_wait, &wait,
 992                                 TASK_UNINTERRUPTIBLE);
 993
 994                 if (cur_trans->num_writers > 1)
 995                         timeout = MAX_SCHEDULE_TIMEOUT;
 996                 else if (should_grow)
 997                         timeout = 1;
 998
 999                 mutex_unlock(&root->fs_info->trans_mutex);
1000
1001                 if (flush_on_commit || snap_pending) {
1002                         btrfs_start_delalloc_inodes(root, 1);
1003                         ret = btrfs_wait_ordered_extents(root, 0, 1);
1004                         BUG_ON(ret);
1005                 }
1006
1007                 /*
1008                  * rename don't use btrfs_join_transaction, so, once we
1009                  * set the transaction to blocked above, we aren't going
1010                  * to get any new ordered operations.  We can safely run
1011                  * it here and no for sure that nothing new will be added
1012                  * to the list
1013                  */
1014                 btrfs_run_ordered_operations(root, 1);
1015
1016                 smp_mb();
1017                 if (cur_trans->num_writers > 1 || should_grow)
1018                         schedule_timeout(timeout);
1019
1020                 mutex_lock(&root->fs_info->trans_mutex);
1021                 finish_wait(&cur_trans->writer_wait, &wait);
1022         } while (cur_trans->num_writers > 1 ||
1023                  (should_grow && cur_trans->num_joined != joined));
1024
1025         ret = create_pending_snapshots(trans, root->fs_info);
1026         BUG_ON(ret);
1027
1028         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1029         BUG_ON(ret);
1030
1031         WARN_ON(cur_trans != trans->transaction);
1032
1033         /* btrfs_commit_tree_roots is responsible for getting the
1034          * various roots consistent with each other.  Every pointer
1035          * in the tree of tree roots has to point to the most up to date
1036          * root for every subvolume and other tree.  So, we have to keep
1037          * the tree logging code from jumping in and changing any
1038          * of the trees.
1039          *
1040          * At this point in the commit, there can't be any tree-log
1041          * writers, but a little lower down we drop the trans mutex
1042          * and let new people in.  By holding the tree_log_mutex
1043          * from now until after the super is written, we avoid races
1044          * with the tree-log code.
1045          */
1046         mutex_lock(&root->fs_info->tree_log_mutex);
1047
1048         ret = commit_fs_roots(trans, root);
1049         BUG_ON(ret);
1050
1051         /* commit_fs_roots gets rid of all the tree log roots, it is now
1052          * safe to free the root of tree log roots
1053          */
1054         btrfs_free_log_root_tree(trans, root->fs_info);
1055
1056         ret = commit_cowonly_roots(trans, root);
1057         BUG_ON(ret);
1058
1059         btrfs_prepare_extent_commit(trans, root);
1060
1061         cur_trans = root->fs_info->running_transaction;
1062         spin_lock(&root->fs_info->new_trans_lock);
1063         root->fs_info->running_transaction = NULL;
1064         spin_unlock(&root->fs_info->new_trans_lock);
1065
1066         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1067                             root->fs_info->tree_root->node);
1068         switch_commit_root(root->fs_info->tree_root);
1069
1070         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1071                             root->fs_info->chunk_root->node);
1072         switch_commit_root(root->fs_info->chunk_root);
1073
1074         update_super_roots(root);
1075
1076         if (!root->fs_info->log_root_recovering) {
1077                 btrfs_set_super_log_root(&root->fs_info->super_copy, 0);
1078                 btrfs_set_super_log_root_level(&root->fs_info->super_copy, 0);
1079         }
1080
1081         memcpy(&root->fs_info->super_for_commit, &root->fs_info->super_copy,
1082                sizeof(root->fs_info->super_copy));
1083
1084         trans->transaction->blocked = 0;
1085
1086         wake_up(&root->fs_info->transaction_wait);
1087
1088         mutex_unlock(&root->fs_info->trans_mutex);
1089         ret = btrfs_write_and_wait_transaction(trans, root);
1090         BUG_ON(ret);
1091         write_ctree_super(trans, root, 0);
1092
1093         /*
1094          * the super is written, we can safely allow the tree-loggers
1095          * to go about their business
1096          */
1097         mutex_unlock(&root->fs_info->tree_log_mutex);
1098
1099         btrfs_finish_extent_commit(trans, root);
1100
1101         /* do the directory inserts of any pending snapshot creations */
1102         finish_pending_snapshots(trans, root->fs_info);
1103
1104         mutex_lock(&root->fs_info->trans_mutex);
1105
1106         cur_trans->commit_done = 1;
1107
1108         root->fs_info->last_trans_committed = cur_trans->transid;
1109
1110         wake_up(&cur_trans->commit_wait);
1111
1112         put_transaction(cur_trans);
1113         put_transaction(cur_trans);
1114
1115         mutex_unlock(&root->fs_info->trans_mutex);
1116
1117         if (current->journal_info == trans)
1118                 current->journal_info = NULL;
1119
1120         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1121
1122         if (current != root->fs_info->transaction_kthread)
1123                 btrfs_run_delayed_iputs(root);
1124
1125         return ret;
1126 }
1127
1128 /*
1129  * interface function to delete all the snapshots we have scheduled for deletion
1130  */
1131 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1132 {
1133         LIST_HEAD(list);
1134         struct btrfs_fs_info *fs_info = root->fs_info;
1135
1136         mutex_lock(&fs_info->trans_mutex);
1137         list_splice_init(&fs_info->dead_roots, &list);
1138         mutex_unlock(&fs_info->trans_mutex);
1139
1140         while (!list_empty(&list)) {
1141                 root = list_entry(list.next, struct btrfs_root, root_list);
1142                 list_del(&root->root_list);
1143
1144                 if (btrfs_header_backref_rev(root->node) <
1145                     BTRFS_MIXED_BACKREF_REV)
1146                         btrfs_drop_snapshot(root, 0);
1147                 else
1148                         btrfs_drop_snapshot(root, 1);
1149         }
1150         return 0;
1151 }