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 <linux/uuid.h>
  26 #include "ctree.h"
  27 #include "disk-io.h"
  28 #include "transaction.h"
  29 #include "locking.h"
  30 #include "tree-log.h"
  31 #include "inode-map.h"
  32 #include "volumes.h"
  33 #include "dev-replace.h"
  34 #include "qgroup.h"
  35
  36 #define BTRFS_ROOT_TRANS_TAG 0
  37
  38 static const unsigned int btrfs_blocked_trans_types[TRANS_STATE_MAX] = {
  39         [TRANS_STATE_RUNNING]           = 0U,
  40         [TRANS_STATE_BLOCKED]           = (__TRANS_USERSPACE |
  41                                            __TRANS_START),
  42         [TRANS_STATE_COMMIT_START]      = (__TRANS_USERSPACE |
  43                                            __TRANS_START |
  44                                            __TRANS_ATTACH),
  45         [TRANS_STATE_COMMIT_DOING]      = (__TRANS_USERSPACE |
  46                                            __TRANS_START |
  47                                            __TRANS_ATTACH |
  48                                            __TRANS_JOIN),
  49         [TRANS_STATE_UNBLOCKED]         = (__TRANS_USERSPACE |
  50                                            __TRANS_START |
  51                                            __TRANS_ATTACH |
  52                                            __TRANS_JOIN |
  53                                            __TRANS_JOIN_NOLOCK),
  54         [TRANS_STATE_COMPLETED]         = (__TRANS_USERSPACE |
  55                                            __TRANS_START |
  56                                            __TRANS_ATTACH |
  57                                            __TRANS_JOIN |
  58                                            __TRANS_JOIN_NOLOCK),
  59 };
  60
  61 void btrfs_put_transaction(struct btrfs_transaction *transaction)
  62 {
  63         WARN_ON(atomic_read(&transaction->use_count) == 0);
  64         if (atomic_dec_and_test(&transaction->use_count)) {
  65                 BUG_ON(!list_empty(&transaction->list));
  66                 WARN_ON(!RB_EMPTY_ROOT(&transaction->delayed_refs.href_root));
  67                 if (transaction->delayed_refs.pending_csums)
  68                         printk(KERN_ERR "pending csums is %llu\n",
  69                                transaction->delayed_refs.pending_csums);
  70                 while (!list_empty(&transaction->pending_chunks)) {
  71                         struct extent_map *em;
  72
  73                         em = list_first_entry(&transaction->pending_chunks,
  74                                               struct extent_map, list);
  75                         list_del_init(&em->list);
  76                         free_extent_map(em);
  77                 }
  78                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  79         }
  80 }
  81
  82 static void clear_btree_io_tree(struct extent_io_tree *tree)
  83 {
  84         spin_lock(&tree->lock);
  85         while (!RB_EMPTY_ROOT(&tree->state)) {
  86                 struct rb_node *node;
  87                 struct extent_state *state;
  88
  89                 node = rb_first(&tree->state);
  90                 state = rb_entry(node, struct extent_state, rb_node);
  91                 rb_erase(&state->rb_node, &tree->state);
  92                 RB_CLEAR_NODE(&state->rb_node);
  93                 /*
  94                  * btree io trees aren't supposed to have tasks waiting for
  95                  * changes in the flags of extent states ever.
  96                  */
  97                 ASSERT(!waitqueue_active(&state->wq));
  98                 free_extent_state(state);
  99
 100                 cond_resched_lock(&tree->lock);
 101         }
 102         spin_unlock(&tree->lock);
 103 }
 104
 105 static noinline void switch_commit_roots(struct btrfs_transaction *trans,
 106                                          struct btrfs_fs_info *fs_info)
 107 {
 108         struct btrfs_root *root, *tmp;
 109
 110         down_write(&fs_info->commit_root_sem);
 111         list_for_each_entry_safe(root, tmp, &trans->switch_commits,
 112                                  dirty_list) {
 113                 list_del_init(&root->dirty_list);
 114                 free_extent_buffer(root->commit_root);
 115                 root->commit_root = btrfs_root_node(root);
 116                 if (is_fstree(root->objectid))
 117                         btrfs_unpin_free_ino(root);
 118                 clear_btree_io_tree(&root->dirty_log_pages);
 119         }
 120         up_write(&fs_info->commit_root_sem);
 121 }
 122
 123 static inline void extwriter_counter_inc(struct btrfs_transaction *trans,
 124                                          unsigned int type)
 125 {
 126         if (type & TRANS_EXTWRITERS)
 127                 atomic_inc(&trans->num_extwriters);
 128 }
 129
 130 static inline void extwriter_counter_dec(struct btrfs_transaction *trans,
 131                                          unsigned int type)
 132 {
 133         if (type & TRANS_EXTWRITERS)
 134                 atomic_dec(&trans->num_extwriters);
 135 }
 136
 137 static inline void extwriter_counter_init(struct btrfs_transaction *trans,
 138                                           unsigned int type)
 139 {
 140         atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0));
 141 }
 142
 143 static inline int extwriter_counter_read(struct btrfs_transaction *trans)
 144 {
 145         return atomic_read(&trans->num_extwriters);
 146 }
 147
 148 /*
 149  * either allocate a new transaction or hop into the existing one
 150  */
 151 static noinline int join_transaction(struct btrfs_root *root, unsigned int type)
 152 {
 153         struct btrfs_transaction *cur_trans;
 154         struct btrfs_fs_info *fs_info = root->fs_info;
 155
 156         spin_lock(&fs_info->trans_lock);
 157 loop:
 158         /* The file system has been taken offline. No new transactions. */
 159         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
 160                 spin_unlock(&fs_info->trans_lock);
 161                 return -EROFS;
 162         }
 163
 164         cur_trans = fs_info->running_transaction;
 165         if (cur_trans) {
 166                 if (cur_trans->aborted) {
 167                         spin_unlock(&fs_info->trans_lock);
 168                         return cur_trans->aborted;
 169                 }
 170                 if (btrfs_blocked_trans_types[cur_trans->state] & type) {
 171                         spin_unlock(&fs_info->trans_lock);
 172                         return -EBUSY;
 173                 }
 174                 atomic_inc(&cur_trans->use_count);
 175                 atomic_inc(&cur_trans->num_writers);
 176                 extwriter_counter_inc(cur_trans, type);
 177                 spin_unlock(&fs_info->trans_lock);
 178                 return 0;
 179         }
 180         spin_unlock(&fs_info->trans_lock);
 181
 182         /*
 183          * If we are ATTACH, we just want to catch the current transaction,
 184          * and commit it. If there is no transaction, just return ENOENT.
 185          */
 186         if (type == TRANS_ATTACH)
 187                 return -ENOENT;
 188
 189         /*
 190          * JOIN_NOLOCK only happens during the transaction commit, so
 191          * it is impossible that ->running_transaction is NULL
 192          */
 193         BUG_ON(type == TRANS_JOIN_NOLOCK);
 194
 195         cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
 196         if (!cur_trans)
 197                 return -ENOMEM;
 198
 199         spin_lock(&fs_info->trans_lock);
 200         if (fs_info->running_transaction) {
 201                 /*
 202                  * someone started a transaction after we unlocked.  Make sure
 203                  * to redo the checks above
 204                  */
 205                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 206                 goto loop;
 207         } else if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
 208                 spin_unlock(&fs_info->trans_lock);
 209                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 210                 return -EROFS;
 211         }
 212
 213         atomic_set(&cur_trans->num_writers, 1);
 214         extwriter_counter_init(cur_trans, type);
 215         init_waitqueue_head(&cur_trans->writer_wait);
 216         init_waitqueue_head(&cur_trans->commit_wait);
 217         cur_trans->state = TRANS_STATE_RUNNING;
 218         /*
 219          * One for this trans handle, one so it will live on until we
 220          * commit the transaction.
 221          */
 222         atomic_set(&cur_trans->use_count, 2);
 223         cur_trans->have_free_bgs = 0;
 224         cur_trans->start_time = get_seconds();
 225         cur_trans->dirty_bg_run = 0;
 226
 227         cur_trans->delayed_refs.href_root = RB_ROOT;
 228         cur_trans->delayed_refs.dirty_extent_root = RB_ROOT;
 229         atomic_set(&cur_trans->delayed_refs.num_entries, 0);
 230         cur_trans->delayed_refs.num_heads_ready = 0;
 231         cur_trans->delayed_refs.pending_csums = 0;
 232         cur_trans->delayed_refs.num_heads = 0;
 233         cur_trans->delayed_refs.flushing = 0;
 234         cur_trans->delayed_refs.run_delayed_start = 0;
 235
 236         /*
 237          * although the tree mod log is per file system and not per transaction,
 238          * the log must never go across transaction boundaries.
 239          */
 240         smp_mb();
 241         if (!list_empty(&fs_info->tree_mod_seq_list))
 242                 WARN(1, KERN_ERR "BTRFS: tree_mod_seq_list not empty when "
 243                         "creating a fresh transaction\n");
 244         if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
 245                 WARN(1, KERN_ERR "BTRFS: tree_mod_log rb tree not empty when "
 246                         "creating a fresh transaction\n");
 247         atomic64_set(&fs_info->tree_mod_seq, 0);
 248
 249         spin_lock_init(&cur_trans->delayed_refs.lock);
 250
 251         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 252         INIT_LIST_HEAD(&cur_trans->pending_chunks);
 253         INIT_LIST_HEAD(&cur_trans->switch_commits);
 254         INIT_LIST_HEAD(&cur_trans->pending_ordered);
 255         INIT_LIST_HEAD(&cur_trans->dirty_bgs);
 256         INIT_LIST_HEAD(&cur_trans->io_bgs);
 257         mutex_init(&cur_trans->cache_write_mutex);
 258         cur_trans->num_dirty_bgs = 0;
 259         spin_lock_init(&cur_trans->dirty_bgs_lock);
 260         list_add_tail(&cur_trans->list, &fs_info->trans_list);
 261         extent_io_tree_init(&cur_trans->dirty_pages,
 262                              fs_info->btree_inode->i_mapping);
 263         fs_info->generation++;
 264         cur_trans->transid = fs_info->generation;
 265         fs_info->running_transaction = cur_trans;
 266         cur_trans->aborted = 0;
 267         spin_unlock(&fs_info->trans_lock);
 268
 269         return 0;
 270 }
 271
 272 /*
 273  * this does all the record keeping required to make sure that a reference
 274  * counted root is properly recorded in a given transaction.  This is required
 275  * to make sure the old root from before we joined the transaction is deleted
 276  * when the transaction commits
 277  */
 278 static int record_root_in_trans(struct btrfs_trans_handle *trans,
 279                                struct btrfs_root *root)
 280 {
 281         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
 282             root->last_trans < trans->transid) {
 283                 WARN_ON(root == root->fs_info->extent_root);
 284                 WARN_ON(root->commit_root != root->node);
 285
 286                 /*
 287                  * see below for IN_TRANS_SETUP usage rules
 288                  * we have the reloc mutex held now, so there
 289                  * is only one writer in this function
 290                  */
 291                 set_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
 292
 293                 /* make sure readers find IN_TRANS_SETUP before
 294                  * they find our root->last_trans update
 295                  */
 296                 smp_wmb();
 297
 298                 spin_lock(&root->fs_info->fs_roots_radix_lock);
 299                 if (root->last_trans == trans->transid) {
 300                         spin_unlock(&root->fs_info->fs_roots_radix_lock);
 301                         return 0;
 302                 }
 303                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 304                            (unsigned long)root->root_key.objectid,
 305                            BTRFS_ROOT_TRANS_TAG);
 306                 spin_unlock(&root->fs_info->fs_roots_radix_lock);
 307                 root->last_trans = trans->transid;
 308
 309                 /* this is pretty tricky.  We don't want to
 310                  * take the relocation lock in btrfs_record_root_in_trans
 311                  * unless we're really doing the first setup for this root in
 312                  * this transaction.
 313                  *
 314                  * Normally we'd use root->last_trans as a flag to decide
 315                  * if we want to take the expensive mutex.
 316                  *
 317                  * But, we have to set root->last_trans before we
 318                  * init the relocation root, otherwise, we trip over warnings
 319                  * in ctree.c.  The solution used here is to flag ourselves
 320                  * with root IN_TRANS_SETUP.  When this is 1, we're still
 321                  * fixing up the reloc trees and everyone must wait.
 322                  *
 323                  * When this is zero, they can trust root->last_trans and fly
 324                  * through btrfs_record_root_in_trans without having to take the
 325                  * lock.  smp_wmb() makes sure that all the writes above are
 326                  * done before we pop in the zero below
 327                  */
 328                 btrfs_init_reloc_root(trans, root);
 329                 smp_mb__before_atomic();
 330                 clear_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
 331         }
 332         return 0;
 333 }
 334
 335
 336 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 337                                struct btrfs_root *root)
 338 {
 339         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
 340                 return 0;
 341
 342         /*
 343          * see record_root_in_trans for comments about IN_TRANS_SETUP usage
 344          * and barriers
 345          */
 346         smp_rmb();
 347         if (root->last_trans == trans->transid &&
 348             !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state))
 349                 return 0;
 350
 351         mutex_lock(&root->fs_info->reloc_mutex);
 352         record_root_in_trans(trans, root);
 353         mutex_unlock(&root->fs_info->reloc_mutex);
 354
 355         return 0;
 356 }
 357
 358 static inline int is_transaction_blocked(struct btrfs_transaction *trans)
 359 {
 360         return (trans->state >= TRANS_STATE_BLOCKED &&
 361                 trans->state < TRANS_STATE_UNBLOCKED &&
 362                 !trans->aborted);
 363 }
 364
 365 /* wait for commit against the current transaction to become unblocked
 366  * when this is done, it is safe to start a new transaction, but the current
 367  * transaction might not be fully on disk.
 368  */
 369 static void wait_current_trans(struct btrfs_root *root)
 370 {
 371         struct btrfs_transaction *cur_trans;
 372
 373         spin_lock(&root->fs_info->trans_lock);
 374         cur_trans = root->fs_info->running_transaction;
 375         if (cur_trans && is_transaction_blocked(cur_trans)) {
 376                 atomic_inc(&cur_trans->use_count);
 377                 spin_unlock(&root->fs_info->trans_lock);
 378
 379                 wait_event(root->fs_info->transaction_wait,
 380                            cur_trans->state >= TRANS_STATE_UNBLOCKED ||
 381                            cur_trans->aborted);
 382                 btrfs_put_transaction(cur_trans);
 383         } else {
 384                 spin_unlock(&root->fs_info->trans_lock);
 385         }
 386 }
 387
 388 static int may_wait_transaction(struct btrfs_root *root, int type)
 389 {
 390         if (root->fs_info->log_root_recovering)
 391                 return 0;
 392
 393         if (type == TRANS_USERSPACE)
 394                 return 1;
 395
 396         if (type == TRANS_START &&
 397             !atomic_read(&root->fs_info->open_ioctl_trans))
 398                 return 1;
 399
 400         return 0;
 401 }
 402
 403 static inline bool need_reserve_reloc_root(struct btrfs_root *root)
 404 {
 405         if (!root->fs_info->reloc_ctl ||
 406             !test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
 407             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 408             root->reloc_root)
 409                 return false;
 410
 411         return true;
 412 }
 413
 414 static struct btrfs_trans_handle *
 415 start_transaction(struct btrfs_root *root, u64 num_items, unsigned int type,
 416                   enum btrfs_reserve_flush_enum flush)
 417 {
 418         struct btrfs_trans_handle *h;
 419         struct btrfs_transaction *cur_trans;
 420         u64 num_bytes = 0;
 421         u64 qgroup_reserved = 0;
 422         bool reloc_reserved = false;
 423         int ret;
 424
 425         /* Send isn't supposed to start transactions. */
 426         ASSERT(current->journal_info != BTRFS_SEND_TRANS_STUB);
 427
 428         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
 429                 return ERR_PTR(-EROFS);
 430
 431         if (current->journal_info) {
 432                 WARN_ON(type & TRANS_EXTWRITERS);
 433                 h = current->journal_info;
 434                 h->use_count++;
 435                 WARN_ON(h->use_count > 2);
 436                 h->orig_rsv = h->block_rsv;
 437                 h->block_rsv = NULL;
 438                 goto got_it;
 439         }
 440
 441         /*
 442          * Do the reservation before we join the transaction so we can do all
 443          * the appropriate flushing if need be.
 444          */
 445         if (num_items > 0 && root != root->fs_info->chunk_root) {
 446                 if (root->fs_info->quota_enabled &&
 447                     is_fstree(root->root_key.objectid)) {
 448                         qgroup_reserved = num_items * root->nodesize;
 449                         ret = btrfs_qgroup_reserve(root, qgroup_reserved);
 450                         if (ret)
 451                                 return ERR_PTR(ret);
 452                 }
 453
 454                 num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
 455                 /*
 456                  * Do the reservation for the relocation root creation
 457                  */
 458                 if (need_reserve_reloc_root(root)) {
 459                         num_bytes += root->nodesize;
 460                         reloc_reserved = true;
 461                 }
 462
 463                 ret = btrfs_block_rsv_add(root,
 464                                           &root->fs_info->trans_block_rsv,
 465                                           num_bytes, flush);
 466                 if (ret)
 467                         goto reserve_fail;
 468         }
 469 again:
 470         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 471         if (!h) {
 472                 ret = -ENOMEM;
 473                 goto alloc_fail;
 474         }
 475
 476         /*
 477          * If we are JOIN_NOLOCK we're already committing a transaction and
 478          * waiting on this guy, so we don't need to do the sb_start_intwrite
 479          * because we're already holding a ref.  We need this because we could
 480          * have raced in and did an fsync() on a file which can kick a commit
 481          * and then we deadlock with somebody doing a freeze.
 482          *
 483          * If we are ATTACH, it means we just want to catch the current
 484          * transaction and commit it, so we needn't do sb_start_intwrite().
 485          */
 486         if (type & __TRANS_FREEZABLE)
 487                 sb_start_intwrite(root->fs_info->sb);
 488
 489         if (may_wait_transaction(root, type))
 490                 wait_current_trans(root);
 491
 492         do {
 493                 ret = join_transaction(root, type);
 494                 if (ret == -EBUSY) {
 495                         wait_current_trans(root);
 496                         if (unlikely(type == TRANS_ATTACH))
 497                                 ret = -ENOENT;
 498                 }
 499         } while (ret == -EBUSY);
 500
 501         if (ret < 0) {
 502                 /* We must get the transaction if we are JOIN_NOLOCK. */
 503                 BUG_ON(type == TRANS_JOIN_NOLOCK);
 504                 goto join_fail;
 505         }
 506
 507         cur_trans = root->fs_info->running_transaction;
 508
 509         h->transid = cur_trans->transid;
 510         h->transaction = cur_trans;
 511         h->blocks_used = 0;
 512         h->bytes_reserved = 0;
 513         h->chunk_bytes_reserved = 0;
 514         h->root = root;
 515         h->delayed_ref_updates = 0;
 516         h->use_count = 1;
 517         h->adding_csums = 0;
 518         h->block_rsv = NULL;
 519         h->orig_rsv = NULL;
 520         h->aborted = 0;
 521         h->qgroup_reserved = 0;
 522         h->delayed_ref_elem.seq = 0;
 523         h->type = type;
 524         h->allocating_chunk = false;
 525         h->reloc_reserved = false;
 526         h->sync = false;
 527         INIT_LIST_HEAD(&h->qgroup_ref_list);
 528         INIT_LIST_HEAD(&h->new_bgs);
 529         INIT_LIST_HEAD(&h->ordered);
 530
 531         smp_mb();
 532         if (cur_trans->state >= TRANS_STATE_BLOCKED &&
 533             may_wait_transaction(root, type)) {
 534                 current->journal_info = h;
 535                 btrfs_commit_transaction(h, root);
 536                 goto again;
 537         }
 538
 539         if (num_bytes) {
 540                 trace_btrfs_space_reservation(root->fs_info, "transaction",
 541                                               h->transid, num_bytes, 1);
 542                 h->block_rsv = &root->fs_info->trans_block_rsv;
 543                 h->bytes_reserved = num_bytes;
 544                 h->reloc_reserved = reloc_reserved;
 545         }
 546         h->qgroup_reserved = qgroup_reserved;
 547
 548 got_it:
 549         btrfs_record_root_in_trans(h, root);
 550
 551         if (!current->journal_info && type != TRANS_USERSPACE)
 552                 current->journal_info = h;
 553         return h;
 554
 555 join_fail:
 556         if (type & __TRANS_FREEZABLE)
 557                 sb_end_intwrite(root->fs_info->sb);
 558         kmem_cache_free(btrfs_trans_handle_cachep, h);
 559 alloc_fail:
 560         if (num_bytes)
 561                 btrfs_block_rsv_release(root, &root->fs_info->trans_block_rsv,
 562                                         num_bytes);
 563 reserve_fail:
 564         if (qgroup_reserved)
 565                 btrfs_qgroup_free(root, qgroup_reserved);
 566         return ERR_PTR(ret);
 567 }
 568
 569 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 570                                                    int num_items)
 571 {
 572         return start_transaction(root, num_items, TRANS_START,
 573                                  BTRFS_RESERVE_FLUSH_ALL);
 574 }
 575
 576 struct btrfs_trans_handle *btrfs_start_transaction_lflush(
 577                                         struct btrfs_root *root, int num_items)
 578 {
 579         return start_transaction(root, num_items, TRANS_START,
 580                                  BTRFS_RESERVE_FLUSH_LIMIT);
 581 }
 582
 583 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 584 {
 585         return start_transaction(root, 0, TRANS_JOIN, 0);
 586 }
 587
 588 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 589 {
 590         return start_transaction(root, 0, TRANS_JOIN_NOLOCK, 0);
 591 }
 592
 593 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 594 {
 595         return start_transaction(root, 0, TRANS_USERSPACE, 0);
 596 }
 597
 598 /*
 599  * btrfs_attach_transaction() - catch the running transaction
 600  *
 601  * It is used when we want to commit the current the transaction, but
 602  * don't want to start a new one.
 603  *
 604  * Note: If this function return -ENOENT, it just means there is no
 605  * running transaction. But it is possible that the inactive transaction
 606  * is still in the memory, not fully on disk. If you hope there is no
 607  * inactive transaction in the fs when -ENOENT is returned, you should
 608  * invoke
 609  *     btrfs_attach_transaction_barrier()
 610  */
 611 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
 612 {
 613         return start_transaction(root, 0, TRANS_ATTACH, 0);
 614 }
 615
 616 /*
 617  * btrfs_attach_transaction_barrier() - catch the running transaction
 618  *
 619  * It is similar to the above function, the differentia is this one
 620  * will wait for all the inactive transactions until they fully
 621  * complete.
 622  */
 623 struct btrfs_trans_handle *
 624 btrfs_attach_transaction_barrier(struct btrfs_root *root)
 625 {
 626         struct btrfs_trans_handle *trans;
 627
 628         trans = start_transaction(root, 0, TRANS_ATTACH, 0);
 629         if (IS_ERR(trans) && PTR_ERR(trans) == -ENOENT)
 630                 btrfs_wait_for_commit(root, 0);
 631
 632         return trans;
 633 }
 634
 635 /* wait for a transaction commit to be fully complete */
 636 static noinline void wait_for_commit(struct btrfs_root *root,
 637                                     struct btrfs_transaction *commit)
 638 {
 639         wait_event(commit->commit_wait, commit->state == TRANS_STATE_COMPLETED);
 640 }
 641
 642 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 643 {
 644         struct btrfs_transaction *cur_trans = NULL, *t;
 645         int ret = 0;
 646
 647         if (transid) {
 648                 if (transid <= root->fs_info->last_trans_committed)
 649                         goto out;
 650
 651                 /* find specified transaction */
 652                 spin_lock(&root->fs_info->trans_lock);
 653                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 654                         if (t->transid == transid) {
 655                                 cur_trans = t;
 656                                 atomic_inc(&cur_trans->use_count);
 657                                 ret = 0;
 658                                 break;
 659                         }
 660                         if (t->transid > transid) {
 661                                 ret = 0;
 662                                 break;
 663                         }
 664                 }
 665                 spin_unlock(&root->fs_info->trans_lock);
 666
 667                 /*
 668                  * The specified transaction doesn't exist, or we
 669                  * raced with btrfs_commit_transaction
 670                  */
 671                 if (!cur_trans) {
 672                         if (transid > root->fs_info->last_trans_committed)
 673                                 ret = -EINVAL;
 674                         goto out;
 675                 }
 676         } else {
 677                 /* find newest transaction that is committing | committed */
 678                 spin_lock(&root->fs_info->trans_lock);
 679                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 680                                             list) {
 681                         if (t->state >= TRANS_STATE_COMMIT_START) {
 682                                 if (t->state == TRANS_STATE_COMPLETED)
 683                                         break;
 684                                 cur_trans = t;
 685                                 atomic_inc(&cur_trans->use_count);
 686                                 break;
 687                         }
 688                 }
 689                 spin_unlock(&root->fs_info->trans_lock);
 690                 if (!cur_trans)
 691                         goto out;  /* nothing committing|committed */
 692         }
 693
 694         wait_for_commit(root, cur_trans);
 695         btrfs_put_transaction(cur_trans);
 696 out:
 697         return ret;
 698 }
 699
 700 void btrfs_throttle(struct btrfs_root *root)
 701 {
 702         if (!atomic_read(&root->fs_info->open_ioctl_trans))
 703                 wait_current_trans(root);
 704 }
 705
 706 static int should_end_transaction(struct btrfs_trans_handle *trans,
 707                                   struct btrfs_root *root)
 708 {
 709         if (root->fs_info->global_block_rsv.space_info->full &&
 710             btrfs_check_space_for_delayed_refs(trans, root))
 711                 return 1;
 712
 713         return !!btrfs_block_rsv_check(root, &root->fs_info->global_block_rsv, 5);
 714 }
 715
 716 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 717                                  struct btrfs_root *root)
 718 {
 719         struct btrfs_transaction *cur_trans = trans->transaction;
 720         int updates;
 721         int err;
 722
 723         smp_mb();
 724         if (cur_trans->state >= TRANS_STATE_BLOCKED ||
 725             cur_trans->delayed_refs.flushing)
 726                 return 1;
 727
 728         updates = trans->delayed_ref_updates;
 729         trans->delayed_ref_updates = 0;
 730         if (updates) {
 731                 err = btrfs_run_delayed_refs(trans, root, updates * 2);
 732                 if (err) /* Error code will also eval true */
 733                         return err;
 734         }
 735
 736         return should_end_transaction(trans, root);
 737 }
 738
 739 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 740                           struct btrfs_root *root, int throttle)
 741 {
 742         struct btrfs_transaction *cur_trans = trans->transaction;
 743         struct btrfs_fs_info *info = root->fs_info;
 744         unsigned long cur = trans->delayed_ref_updates;
 745         int lock = (trans->type != TRANS_JOIN_NOLOCK);
 746         int err = 0;
 747         int must_run_delayed_refs = 0;
 748
 749         if (trans->use_count > 1) {
 750                 trans->use_count--;
 751                 trans->block_rsv = trans->orig_rsv;
 752                 return 0;
 753         }
 754
 755         btrfs_trans_release_metadata(trans, root);
 756         trans->block_rsv = NULL;
 757
 758         if (!list_empty(&trans->new_bgs))
 759                 btrfs_create_pending_block_groups(trans, root);
 760
 761         if (!list_empty(&trans->ordered)) {
 762                 spin_lock(&info->trans_lock);
 763                 list_splice(&trans->ordered, &cur_trans->pending_ordered);
 764                 spin_unlock(&info->trans_lock);
 765         }
 766
 767         trans->delayed_ref_updates = 0;
 768         if (!trans->sync) {
 769                 must_run_delayed_refs =
 770                         btrfs_should_throttle_delayed_refs(trans, root);
 771                 cur = max_t(unsigned long, cur, 32);
 772
 773                 /*
 774                  * don't make the caller wait if they are from a NOLOCK
 775                  * or ATTACH transaction, it will deadlock with commit
 776                  */
 777                 if (must_run_delayed_refs == 1 &&
 778                     (trans->type & (__TRANS_JOIN_NOLOCK | __TRANS_ATTACH)))
 779                         must_run_delayed_refs = 2;
 780         }
 781
 782         if (trans->qgroup_reserved) {
 783                 /*
 784                  * the same root has to be passed here between start_transaction
 785                  * and end_transaction. Subvolume quota depends on this.
 786                  */
 787                 btrfs_qgroup_free(trans->root, trans->qgroup_reserved);
 788                 trans->qgroup_reserved = 0;
 789         }
 790
 791         btrfs_trans_release_metadata(trans, root);
 792         trans->block_rsv = NULL;
 793
 794         if (!list_empty(&trans->new_bgs))
 795                 btrfs_create_pending_block_groups(trans, root);
 796
 797         btrfs_trans_release_chunk_metadata(trans);
 798
 799         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
 800             should_end_transaction(trans, root) &&
 801             ACCESS_ONCE(cur_trans->state) == TRANS_STATE_RUNNING) {
 802                 spin_lock(&info->trans_lock);
 803                 if (cur_trans->state == TRANS_STATE_RUNNING)
 804                         cur_trans->state = TRANS_STATE_BLOCKED;
 805                 spin_unlock(&info->trans_lock);
 806         }
 807
 808         if (lock && ACCESS_ONCE(cur_trans->state) == TRANS_STATE_BLOCKED) {
 809                 if (throttle)
 810                         return btrfs_commit_transaction(trans, root);
 811                 else
 812                         wake_up_process(info->transaction_kthread);
 813         }
 814
 815         if (trans->type & __TRANS_FREEZABLE)
 816                 sb_end_intwrite(root->fs_info->sb);
 817
 818         WARN_ON(cur_trans != info->running_transaction);
 819         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 820         atomic_dec(&cur_trans->num_writers);
 821         extwriter_counter_dec(cur_trans, trans->type);
 822
 823         smp_mb();
 824         if (waitqueue_active(&cur_trans->writer_wait))
 825                 wake_up(&cur_trans->writer_wait);
 826         btrfs_put_transaction(cur_trans);
 827
 828         if (current->journal_info == trans)
 829                 current->journal_info = NULL;
 830
 831         if (throttle)
 832                 btrfs_run_delayed_iputs(root);
 833
 834         if (trans->aborted ||
 835             test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state)) {
 836                 wake_up_process(info->transaction_kthread);
 837                 err = -EIO;
 838         }
 839         assert_qgroups_uptodate(trans);
 840
 841         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 842         if (must_run_delayed_refs) {
 843                 btrfs_async_run_delayed_refs(root, cur,
 844                                              must_run_delayed_refs == 1);
 845         }
 846         return err;
 847 }
 848
 849 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 850                           struct btrfs_root *root)
 851 {
 852         return __btrfs_end_transaction(trans, root, 0);
 853 }
 854
 855 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 856                                    struct btrfs_root *root)
 857 {
 858         return __btrfs_end_transaction(trans, root, 1);
 859 }
 860
 861 /*
 862  * when btree blocks are allocated, they have some corresponding bits set for
 863  * them in one of two extent_io trees.  This is used to make sure all of
 864  * those extents are sent to disk but does not wait on them
 865  */
 866 int btrfs_write_marked_extents(struct btrfs_root *root,
 867                                struct extent_io_tree *dirty_pages, int mark)
 868 {
 869         int err = 0;
 870         int werr = 0;
 871         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 872         struct extent_state *cached_state = NULL;
 873         u64 start = 0;
 874         u64 end;
 875
 876         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 877                                       mark, &cached_state)) {
 878                 bool wait_writeback = false;
 879
 880                 err = convert_extent_bit(dirty_pages, start, end,
 881                                          EXTENT_NEED_WAIT,
 882                                          mark, &cached_state, GFP_NOFS);
 883                 /*
 884                  * convert_extent_bit can return -ENOMEM, which is most of the
 885                  * time a temporary error. So when it happens, ignore the error
 886                  * and wait for writeback of this range to finish - because we
 887                  * failed to set the bit EXTENT_NEED_WAIT for the range, a call
 888                  * to btrfs_wait_marked_extents() would not know that writeback
 889                  * for this range started and therefore wouldn't wait for it to
 890                  * finish - we don't want to commit a superblock that points to
 891                  * btree nodes/leafs for which writeback hasn't finished yet
 892                  * (and without errors).
 893                  * We cleanup any entries left in the io tree when committing
 894                  * the transaction (through clear_btree_io_tree()).
 895                  */
 896                 if (err == -ENOMEM) {
 897                         err = 0;
 898                         wait_writeback = true;
 899                 }
 900                 if (!err)
 901                         err = filemap_fdatawrite_range(mapping, start, end);
 902                 if (err)
 903                         werr = err;
 904                 else if (wait_writeback)
 905                         werr = filemap_fdatawait_range(mapping, start, end);
 906                 free_extent_state(cached_state);
 907                 cached_state = NULL;
 908                 cond_resched();
 909                 start = end + 1;
 910         }
 911         return werr;
 912 }
 913
 914 /*
 915  * when btree blocks are allocated, they have some corresponding bits set for
 916  * them in one of two extent_io trees.  This is used to make sure all of
 917  * those extents are on disk for transaction or log commit.  We wait
 918  * on all the pages and clear them from the dirty pages state tree
 919  */
 920 int btrfs_wait_marked_extents(struct btrfs_root *root,
 921                               struct extent_io_tree *dirty_pages, int mark)
 922 {
 923         int err = 0;
 924         int werr = 0;
 925         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 926         struct extent_state *cached_state = NULL;
 927         u64 start = 0;
 928         u64 end;
 929         struct btrfs_inode *btree_ino = BTRFS_I(root->fs_info->btree_inode);
 930         bool errors = false;
 931
 932         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 933                                       EXTENT_NEED_WAIT, &cached_state)) {
 934                 /*
 935                  * Ignore -ENOMEM errors returned by clear_extent_bit().
 936                  * When committing the transaction, we'll remove any entries
 937                  * left in the io tree. For a log commit, we don't remove them
 938                  * after committing the log because the tree can be accessed
 939                  * concurrently - we do it only at transaction commit time when
 940                  * it's safe to do it (through clear_btree_io_tree()).
 941                  */
 942                 err = clear_extent_bit(dirty_pages, start, end,
 943                                        EXTENT_NEED_WAIT,
 944                                        0, 0, &cached_state, GFP_NOFS);
 945                 if (err == -ENOMEM)
 946                         err = 0;
 947                 if (!err)
 948                         err = filemap_fdatawait_range(mapping, start, end);
 949                 if (err)
 950                         werr = err;
 951                 free_extent_state(cached_state);
 952                 cached_state = NULL;
 953                 cond_resched();
 954                 start = end + 1;
 955         }
 956         if (err)
 957                 werr = err;
 958
 959         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
 960                 if ((mark & EXTENT_DIRTY) &&
 961                     test_and_clear_bit(BTRFS_INODE_BTREE_LOG1_ERR,
 962                                        &btree_ino->runtime_flags))
 963                         errors = true;
 964
 965                 if ((mark & EXTENT_NEW) &&
 966                     test_and_clear_bit(BTRFS_INODE_BTREE_LOG2_ERR,
 967                                        &btree_ino->runtime_flags))
 968                         errors = true;
 969         } else {
 970                 if (test_and_clear_bit(BTRFS_INODE_BTREE_ERR,
 971                                        &btree_ino->runtime_flags))
 972                         errors = true;
 973         }
 974
 975         if (errors && !werr)
 976                 werr = -EIO;
 977
 978         return werr;
 979 }
 980
 981 /*
 982  * when btree blocks are allocated, they have some corresponding bits set for
 983  * them in one of two extent_io trees.  This is used to make sure all of
 984  * those extents are on disk for transaction or log commit
 985  */
 986 static int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 987                                 struct extent_io_tree *dirty_pages, int mark)
 988 {
 989         int ret;
 990         int ret2;
 991         struct blk_plug plug;
 992
 993         blk_start_plug(&plug);
 994         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 995         blk_finish_plug(&plug);
 996         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 997
 998         if (ret)
 999                 return ret;
1000         if (ret2)
1001                 return ret2;
1002         return 0;
1003 }
1004
1005 static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
1006                                      struct btrfs_root *root)
1007 {
1008         int ret;
1009
1010         ret = btrfs_write_and_wait_marked_extents(root,
1011                                            &trans->transaction->dirty_pages,
1012                                            EXTENT_DIRTY);
1013         clear_btree_io_tree(&trans->transaction->dirty_pages);
1014
1015         return ret;
1016 }
1017
1018 /*
1019  * this is used to update the root pointer in the tree of tree roots.
1020  *
1021  * But, in the case of the extent allocation tree, updating the root
1022  * pointer may allocate blocks which may change the root of the extent
1023  * allocation tree.
1024  *
1025  * So, this loops and repeats and makes sure the cowonly root didn't
1026  * change while the root pointer was being updated in the metadata.
1027  */
1028 static int update_cowonly_root(struct btrfs_trans_handle *trans,
1029                                struct btrfs_root *root)
1030 {
1031         int ret;
1032         u64 old_root_bytenr;
1033         u64 old_root_used;
1034         struct btrfs_root *tree_root = root->fs_info->tree_root;
1035
1036         old_root_used = btrfs_root_used(&root->root_item);
1037
1038         while (1) {
1039                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1040                 if (old_root_bytenr == root->node->start &&
1041                     old_root_used == btrfs_root_used(&root->root_item))
1042                         break;
1043
1044                 btrfs_set_root_node(&root->root_item, root->node);
1045                 ret = btrfs_update_root(trans, tree_root,
1046                                         &root->root_key,
1047                                         &root->root_item);
1048                 if (ret)
1049                         return ret;
1050
1051                 old_root_used = btrfs_root_used(&root->root_item);
1052         }
1053
1054         return 0;
1055 }
1056
1057 /*
1058  * update all the cowonly tree roots on disk
1059  *
1060  * The error handling in this function may not be obvious. Any of the
1061  * failures will cause the file system to go offline. We still need
1062  * to clean up the delayed refs.
1063  */
1064 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
1065                                          struct btrfs_root *root)
1066 {
1067         struct btrfs_fs_info *fs_info = root->fs_info;
1068         struct list_head *dirty_bgs = &trans->transaction->dirty_bgs;
1069         struct list_head *io_bgs = &trans->transaction->io_bgs;
1070         struct list_head *next;
1071         struct extent_buffer *eb;
1072         int ret;
1073
1074         eb = btrfs_lock_root_node(fs_info->tree_root);
1075         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
1076                               0, &eb);
1077         btrfs_tree_unlock(eb);
1078         free_extent_buffer(eb);
1079
1080         if (ret)
1081                 return ret;
1082
1083         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1084         if (ret)
1085                 return ret;
1086
1087         ret = btrfs_run_dev_stats(trans, root->fs_info);
1088         if (ret)
1089                 return ret;
1090         ret = btrfs_run_dev_replace(trans, root->fs_info);
1091         if (ret)
1092                 return ret;
1093         ret = btrfs_run_qgroups(trans, root->fs_info);
1094         if (ret)
1095                 return ret;
1096
1097         ret = btrfs_setup_space_cache(trans, root);
1098         if (ret)
1099                 return ret;
1100
1101         /* run_qgroups might have added some more refs */
1102         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1103         if (ret)
1104                 return ret;
1105 again:
1106         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1107                 next = fs_info->dirty_cowonly_roots.next;
1108                 list_del_init(next);
1109                 root = list_entry(next, struct btrfs_root, dirty_list);
1110                 clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1111
1112                 if (root != fs_info->extent_root)
1113                         list_add_tail(&root->dirty_list,
1114                                       &trans->transaction->switch_commits);
1115                 ret = update_cowonly_root(trans, root);
1116                 if (ret)
1117                         return ret;
1118                 ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1119                 if (ret)
1120                         return ret;
1121         }
1122
1123         while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1124                 ret = btrfs_write_dirty_block_groups(trans, root);
1125                 if (ret)
1126                         return ret;
1127                 ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1128                 if (ret)
1129                         return ret;
1130         }
1131
1132         if (!list_empty(&fs_info->dirty_cowonly_roots))
1133                 goto again;
1134
1135         list_add_tail(&fs_info->extent_root->dirty_list,
1136                       &trans->transaction->switch_commits);
1137         btrfs_after_dev_replace_commit(fs_info);
1138
1139         return 0;
1140 }
1141
1142 /*
1143  * dead roots are old snapshots that need to be deleted.  This allocates
1144  * a dirty root struct and adds it into the list of dead roots that need to
1145  * be deleted
1146  */
1147 void btrfs_add_dead_root(struct btrfs_root *root)
1148 {
1149         spin_lock(&root->fs_info->trans_lock);
1150         if (list_empty(&root->root_list))
1151                 list_add_tail(&root->root_list, &root->fs_info->dead_roots);
1152         spin_unlock(&root->fs_info->trans_lock);
1153 }
1154
1155 /*
1156  * update all the cowonly tree roots on disk
1157  */
1158 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
1159                                     struct btrfs_root *root)
1160 {
1161         struct btrfs_root *gang[8];
1162         struct btrfs_fs_info *fs_info = root->fs_info;
1163         int i;
1164         int ret;
1165         int err = 0;
1166
1167         spin_lock(&fs_info->fs_roots_radix_lock);
1168         while (1) {
1169                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1170                                                  (void **)gang, 0,
1171                                                  ARRAY_SIZE(gang),
1172                                                  BTRFS_ROOT_TRANS_TAG);
1173                 if (ret == 0)
1174                         break;
1175                 for (i = 0; i < ret; i++) {
1176                         root = gang[i];
1177                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
1178                                         (unsigned long)root->root_key.objectid,
1179                                         BTRFS_ROOT_TRANS_TAG);
1180                         spin_unlock(&fs_info->fs_roots_radix_lock);
1181
1182                         btrfs_free_log(trans, root);
1183                         btrfs_update_reloc_root(trans, root);
1184                         btrfs_orphan_commit_root(trans, root);
1185
1186                         btrfs_save_ino_cache(root, trans);
1187
1188                         /* see comments in should_cow_block() */
1189                         clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1190                         smp_mb__after_atomic();
1191
1192                         if (root->commit_root != root->node) {
1193                                 list_add_tail(&root->dirty_list,
1194                                         &trans->transaction->switch_commits);
1195                                 btrfs_set_root_node(&root->root_item,
1196                                                     root->node);
1197                         }
1198
1199                         err = btrfs_update_root(trans, fs_info->tree_root,
1200                                                 &root->root_key,
1201                                                 &root->root_item);
1202                         spin_lock(&fs_info->fs_roots_radix_lock);
1203                         if (err)
1204                                 break;
1205                 }
1206         }
1207         spin_unlock(&fs_info->fs_roots_radix_lock);
1208         return err;
1209 }
1210
1211 /*
1212  * defrag a given btree.
1213  * Every leaf in the btree is read and defragged.
1214  */
1215 int btrfs_defrag_root(struct btrfs_root *root)
1216 {
1217         struct btrfs_fs_info *info = root->fs_info;
1218         struct btrfs_trans_handle *trans;
1219         int ret;
1220
1221         if (test_and_set_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state))
1222                 return 0;
1223
1224         while (1) {
1225                 trans = btrfs_start_transaction(root, 0);
1226                 if (IS_ERR(trans))
1227                         return PTR_ERR(trans);
1228
1229                 ret = btrfs_defrag_leaves(trans, root);
1230
1231                 btrfs_end_transaction(trans, root);
1232                 btrfs_btree_balance_dirty(info->tree_root);
1233                 cond_resched();
1234
1235                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
1236                         break;
1237
1238                 if (btrfs_defrag_cancelled(root->fs_info)) {
1239                         pr_debug("BTRFS: defrag_root cancelled\n");
1240                         ret = -EAGAIN;
1241                         break;
1242                 }
1243         }
1244         clear_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state);
1245         return ret;
1246 }
1247
1248 /*
1249  * new snapshots need to be created at a very specific time in the
1250  * transaction commit.  This does the actual creation.
1251  *
1252  * Note:
1253  * If the error which may affect the commitment of the current transaction
1254  * happens, we should return the error number. If the error which just affect
1255  * the creation of the pending snapshots, just return 0.
1256  */
1257 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1258                                    struct btrfs_fs_info *fs_info,
1259                                    struct btrfs_pending_snapshot *pending)
1260 {
1261         struct btrfs_key key;
1262         struct btrfs_root_item *new_root_item;
1263         struct btrfs_root *tree_root = fs_info->tree_root;
1264         struct btrfs_root *root = pending->root;
1265         struct btrfs_root *parent_root;
1266         struct btrfs_block_rsv *rsv;
1267         struct inode *parent_inode;
1268         struct btrfs_path *path;
1269         struct btrfs_dir_item *dir_item;
1270         struct dentry *dentry;
1271         struct extent_buffer *tmp;
1272         struct extent_buffer *old;
1273         struct timespec cur_time = CURRENT_TIME;
1274         int ret = 0;
1275         u64 to_reserve = 0;
1276         u64 index = 0;
1277         u64 objectid;
1278         u64 root_flags;
1279         uuid_le new_uuid;
1280
1281         path = btrfs_alloc_path();
1282         if (!path) {
1283                 pending->error = -ENOMEM;
1284                 return 0;
1285         }
1286
1287         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
1288         if (!new_root_item) {
1289                 pending->error = -ENOMEM;
1290                 goto root_item_alloc_fail;
1291         }
1292
1293         pending->error = btrfs_find_free_objectid(tree_root, &objectid);
1294         if (pending->error)
1295                 goto no_free_objectid;
1296
1297         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
1298
1299         if (to_reserve > 0) {
1300                 pending->error = btrfs_block_rsv_add(root,
1301                                                      &pending->block_rsv,
1302                                                      to_reserve,
1303                                                      BTRFS_RESERVE_NO_FLUSH);
1304                 if (pending->error)
1305                         goto no_free_objectid;
1306         }
1307
1308         key.objectid = objectid;
1309         key.offset = (u64)-1;
1310         key.type = BTRFS_ROOT_ITEM_KEY;
1311
1312         rsv = trans->block_rsv;
1313         trans->block_rsv = &pending->block_rsv;
1314         trans->bytes_reserved = trans->block_rsv->reserved;
1315
1316         dentry = pending->dentry;
1317         parent_inode = pending->dir;
1318         parent_root = BTRFS_I(parent_inode)->root;
1319         record_root_in_trans(trans, parent_root);
1320
1321         /*
1322          * insert the directory item
1323          */
1324         ret = btrfs_set_inode_index(parent_inode, &index);
1325         BUG_ON(ret); /* -ENOMEM */
1326
1327         /* check if there is a file/dir which has the same name. */
1328         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1329                                          btrfs_ino(parent_inode),
1330                                          dentry->d_name.name,
1331                                          dentry->d_name.len, 0);
1332         if (dir_item != NULL && !IS_ERR(dir_item)) {
1333                 pending->error = -EEXIST;
1334                 goto dir_item_existed;
1335         } else if (IS_ERR(dir_item)) {
1336                 ret = PTR_ERR(dir_item);
1337                 btrfs_abort_transaction(trans, root, ret);
1338                 goto fail;
1339         }
1340         btrfs_release_path(path);
1341
1342         /*
1343          * pull in the delayed directory update
1344          * and the delayed inode item
1345          * otherwise we corrupt the FS during
1346          * snapshot
1347          */
1348         ret = btrfs_run_delayed_items(trans, root);
1349         if (ret) {      /* Transaction aborted */
1350                 btrfs_abort_transaction(trans, root, ret);
1351                 goto fail;
1352         }
1353
1354         record_root_in_trans(trans, root);
1355         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1356         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1357         btrfs_check_and_init_root_item(new_root_item);
1358
1359         root_flags = btrfs_root_flags(new_root_item);
1360         if (pending->readonly)
1361                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1362         else
1363                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1364         btrfs_set_root_flags(new_root_item, root_flags);
1365
1366         btrfs_set_root_generation_v2(new_root_item,
1367                         trans->transid);
1368         uuid_le_gen(&new_uuid);
1369         memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1370         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1371                         BTRFS_UUID_SIZE);
1372         if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1373                 memset(new_root_item->received_uuid, 0,
1374                        sizeof(new_root_item->received_uuid));
1375                 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1376                 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1377                 btrfs_set_root_stransid(new_root_item, 0);
1378                 btrfs_set_root_rtransid(new_root_item, 0);
1379         }
1380         btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1381         btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1382         btrfs_set_root_otransid(new_root_item, trans->transid);
1383
1384         old = btrfs_lock_root_node(root);
1385         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1386         if (ret) {
1387                 btrfs_tree_unlock(old);
1388                 free_extent_buffer(old);
1389                 btrfs_abort_transaction(trans, root, ret);
1390                 goto fail;
1391         }
1392
1393         btrfs_set_lock_blocking(old);
1394
1395         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1396         /* clean up in any case */
1397         btrfs_tree_unlock(old);
1398         free_extent_buffer(old);
1399         if (ret) {
1400                 btrfs_abort_transaction(trans, root, ret);
1401                 goto fail;
1402         }
1403
1404         /*
1405          * We need to flush delayed refs in order to make sure all of our quota
1406          * operations have been done before we call btrfs_qgroup_inherit.
1407          */
1408         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1409         if (ret) {
1410                 btrfs_abort_transaction(trans, root, ret);
1411                 goto fail;
1412         }
1413
1414         ret = btrfs_qgroup_inherit(trans, fs_info,
1415                                    root->root_key.objectid,
1416                                    objectid, pending->inherit);
1417         if (ret) {
1418                 btrfs_abort_transaction(trans, root, ret);
1419                 goto fail;
1420         }
1421
1422         /* see comments in should_cow_block() */
1423         set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1424         smp_wmb();
1425
1426         btrfs_set_root_node(new_root_item, tmp);
1427         /* record when the snapshot was created in key.offset */
1428         key.offset = trans->transid;
1429         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1430         btrfs_tree_unlock(tmp);
1431         free_extent_buffer(tmp);
1432         if (ret) {
1433                 btrfs_abort_transaction(trans, root, ret);
1434                 goto fail;
1435         }
1436
1437         /*
1438          * insert root back/forward references
1439          */
1440         ret = btrfs_add_root_ref(trans, tree_root, objectid,
1441                                  parent_root->root_key.objectid,
1442                                  btrfs_ino(parent_inode), index,
1443                                  dentry->d_name.name, dentry->d_name.len);
1444         if (ret) {
1445                 btrfs_abort_transaction(trans, root, ret);
1446                 goto fail;
1447         }
1448
1449         key.offset = (u64)-1;
1450         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1451         if (IS_ERR(pending->snap)) {
1452                 ret = PTR_ERR(pending->snap);
1453                 btrfs_abort_transaction(trans, root, ret);
1454                 goto fail;
1455         }
1456
1457         ret = btrfs_reloc_post_snapshot(trans, pending);
1458         if (ret) {
1459                 btrfs_abort_transaction(trans, root, ret);
1460                 goto fail;
1461         }
1462
1463         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1464         if (ret) {
1465                 btrfs_abort_transaction(trans, root, ret);
1466                 goto fail;
1467         }
1468
1469         ret = btrfs_insert_dir_item(trans, parent_root,
1470                                     dentry->d_name.name, dentry->d_name.len,
1471                                     parent_inode, &key,
1472                                     BTRFS_FT_DIR, index);
1473         /* We have check then name at the beginning, so it is impossible. */
1474         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1475         if (ret) {
1476                 btrfs_abort_transaction(trans, root, ret);
1477                 goto fail;
1478         }
1479
1480         btrfs_i_size_write(parent_inode, parent_inode->i_size +
1481                                          dentry->d_name.len * 2);
1482         parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
1483         ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1484         if (ret) {
1485                 btrfs_abort_transaction(trans, root, ret);
1486                 goto fail;
1487         }
1488         ret = btrfs_uuid_tree_add(trans, fs_info->uuid_root, new_uuid.b,
1489                                   BTRFS_UUID_KEY_SUBVOL, objectid);
1490         if (ret) {
1491                 btrfs_abort_transaction(trans, root, ret);
1492                 goto fail;
1493         }
1494         if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1495                 ret = btrfs_uuid_tree_add(trans, fs_info->uuid_root,
1496                                           new_root_item->received_uuid,
1497                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1498                                           objectid);
1499                 if (ret && ret != -EEXIST) {
1500                         btrfs_abort_transaction(trans, root, ret);
1501                         goto fail;
1502                 }
1503         }
1504 fail:
1505         pending->error = ret;
1506 dir_item_existed:
1507         trans->block_rsv = rsv;
1508         trans->bytes_reserved = 0;
1509 no_free_objectid:
1510         kfree(new_root_item);
1511 root_item_alloc_fail:
1512         btrfs_free_path(path);
1513         return ret;
1514 }
1515
1516 /*
1517  * create all the snapshots we've scheduled for creation
1518  */
1519 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1520                                              struct btrfs_fs_info *fs_info)
1521 {
1522         struct btrfs_pending_snapshot *pending, *next;
1523         struct list_head *head = &trans->transaction->pending_snapshots;
1524         int ret = 0;
1525
1526         list_for_each_entry_safe(pending, next, head, list) {
1527                 list_del(&pending->list);
1528                 ret = create_pending_snapshot(trans, fs_info, pending);
1529                 if (ret)
1530                         break;
1531         }
1532         return ret;
1533 }
1534
1535 static void update_super_roots(struct btrfs_root *root)
1536 {
1537         struct btrfs_root_item *root_item;
1538         struct btrfs_super_block *super;
1539
1540         super = root->fs_info->super_copy;
1541
1542         root_item = &root->fs_info->chunk_root->root_item;
1543         super->chunk_root = root_item->bytenr;
1544         super->chunk_root_generation = root_item->generation;
1545         super->chunk_root_level = root_item->level;
1546
1547         root_item = &root->fs_info->tree_root->root_item;
1548         super->root = root_item->bytenr;
1549         super->generation = root_item->generation;
1550         super->root_level = root_item->level;
1551         if (btrfs_test_opt(root, SPACE_CACHE))
1552                 super->cache_generation = root_item->generation;
1553         if (root->fs_info->update_uuid_tree_gen)
1554                 super->uuid_tree_generation = root_item->generation;
1555 }
1556
1557 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1558 {
1559         struct btrfs_transaction *trans;
1560         int ret = 0;
1561
1562         spin_lock(&info->trans_lock);
1563         trans = info->running_transaction;
1564         if (trans)
1565                 ret = (trans->state >= TRANS_STATE_COMMIT_START);
1566         spin_unlock(&info->trans_lock);
1567         return ret;
1568 }
1569
1570 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1571 {
1572         struct btrfs_transaction *trans;
1573         int ret = 0;
1574
1575         spin_lock(&info->trans_lock);
1576         trans = info->running_transaction;
1577         if (trans)
1578                 ret = is_transaction_blocked(trans);
1579         spin_unlock(&info->trans_lock);
1580         return ret;
1581 }
1582
1583 /*
1584  * wait for the current transaction commit to start and block subsequent
1585  * transaction joins
1586  */
1587 static void wait_current_trans_commit_start(struct btrfs_root *root,
1588                                             struct btrfs_transaction *trans)
1589 {
1590         wait_event(root->fs_info->transaction_blocked_wait,
1591                    trans->state >= TRANS_STATE_COMMIT_START ||
1592                    trans->aborted);
1593 }
1594
1595 /*
1596  * wait for the current transaction to start and then become unblocked.
1597  * caller holds ref.
1598  */
1599 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1600                                          struct btrfs_transaction *trans)
1601 {
1602         wait_event(root->fs_info->transaction_wait,
1603                    trans->state >= TRANS_STATE_UNBLOCKED ||
1604                    trans->aborted);
1605 }
1606
1607 /*
1608  * commit transactions asynchronously. once btrfs_commit_transaction_async
1609  * returns, any subsequent transaction will not be allowed to join.
1610  */
1611 struct btrfs_async_commit {
1612         struct btrfs_trans_handle *newtrans;
1613         struct btrfs_root *root;
1614         struct work_struct work;
1615 };
1616
1617 static void do_async_commit(struct work_struct *work)
1618 {
1619         struct btrfs_async_commit *ac =
1620                 container_of(work, struct btrfs_async_commit, work);
1621
1622         /*
1623          * We've got freeze protection passed with the transaction.
1624          * Tell lockdep about it.
1625          */
1626         if (ac->newtrans->type & __TRANS_FREEZABLE)
1627                 rwsem_acquire_read(
1628                      &ac->root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
1629                      0, 1, _THIS_IP_);
1630
1631         current->journal_info = ac->newtrans;
1632
1633         btrfs_commit_transaction(ac->newtrans, ac->root);
1634         kfree(ac);
1635 }
1636
1637 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1638                                    struct btrfs_root *root,
1639                                    int wait_for_unblock)
1640 {
1641         struct btrfs_async_commit *ac;
1642         struct btrfs_transaction *cur_trans;
1643
1644         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1645         if (!ac)
1646                 return -ENOMEM;
1647
1648         INIT_WORK(&ac->work, do_async_commit);
1649         ac->root = root;
1650         ac->newtrans = btrfs_join_transaction(root);
1651         if (IS_ERR(ac->newtrans)) {
1652                 int err = PTR_ERR(ac->newtrans);
1653                 kfree(ac);
1654                 return err;
1655         }
1656
1657         /* take transaction reference */
1658         cur_trans = trans->transaction;
1659         atomic_inc(&cur_trans->use_count);
1660
1661         btrfs_end_transaction(trans, root);
1662
1663         /*
1664          * Tell lockdep we've released the freeze rwsem, since the
1665          * async commit thread will be the one to unlock it.
1666          */
1667         if (ac->newtrans->type & __TRANS_FREEZABLE)
1668                 rwsem_release(
1669                         &root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
1670                         1, _THIS_IP_);
1671
1672         schedule_work(&ac->work);
1673
1674         /* wait for transaction to start and unblock */
1675         if (wait_for_unblock)
1676                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1677         else
1678                 wait_current_trans_commit_start(root, cur_trans);
1679
1680         if (current->journal_info == trans)
1681                 current->journal_info = NULL;
1682
1683         btrfs_put_transaction(cur_trans);
1684         return 0;
1685 }
1686
1687
1688 static void cleanup_transaction(struct btrfs_trans_handle *trans,
1689                                 struct btrfs_root *root, int err)
1690 {
1691         struct btrfs_transaction *cur_trans = trans->transaction;
1692         DEFINE_WAIT(wait);
1693
1694         WARN_ON(trans->use_count > 1);
1695
1696         btrfs_abort_transaction(trans, root, err);
1697
1698         spin_lock(&root->fs_info->trans_lock);
1699
1700         /*
1701          * If the transaction is removed from the list, it means this
1702          * transaction has been committed successfully, so it is impossible
1703          * to call the cleanup function.
1704          */
1705         BUG_ON(list_empty(&cur_trans->list));
1706
1707         list_del_init(&cur_trans->list);
1708         if (cur_trans == root->fs_info->running_transaction) {
1709                 cur_trans->state = TRANS_STATE_COMMIT_DOING;
1710                 spin_unlock(&root->fs_info->trans_lock);
1711                 wait_event(cur_trans->writer_wait,
1712                            atomic_read(&cur_trans->num_writers) == 1);
1713
1714                 spin_lock(&root->fs_info->trans_lock);
1715         }
1716         spin_unlock(&root->fs_info->trans_lock);
1717
1718         btrfs_cleanup_one_transaction(trans->transaction, root);
1719
1720         spin_lock(&root->fs_info->trans_lock);
1721         if (cur_trans == root->fs_info->running_transaction)
1722                 root->fs_info->running_transaction = NULL;
1723         spin_unlock(&root->fs_info->trans_lock);
1724
1725         if (trans->type & __TRANS_FREEZABLE)
1726                 sb_end_intwrite(root->fs_info->sb);
1727         btrfs_put_transaction(cur_trans);
1728         btrfs_put_transaction(cur_trans);
1729
1730         trace_btrfs_transaction_commit(root);
1731
1732         if (current->journal_info == trans)
1733                 current->journal_info = NULL;
1734         btrfs_scrub_cancel(root->fs_info);
1735
1736         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1737 }
1738
1739 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
1740 {
1741         if (btrfs_test_opt(fs_info->tree_root, FLUSHONCOMMIT))
1742                 return btrfs_start_delalloc_roots(fs_info, 1, -1);
1743         return 0;
1744 }
1745
1746 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
1747 {
1748         if (btrfs_test_opt(fs_info->tree_root, FLUSHONCOMMIT))
1749                 btrfs_wait_ordered_roots(fs_info, -1);
1750 }
1751
1752 static inline void
1753 btrfs_wait_pending_ordered(struct btrfs_transaction *cur_trans,
1754                            struct btrfs_fs_info *fs_info)
1755 {
1756         struct btrfs_ordered_extent *ordered;
1757
1758         spin_lock(&fs_info->trans_lock);
1759         while (!list_empty(&cur_trans->pending_ordered)) {
1760                 ordered = list_first_entry(&cur_trans->pending_ordered,
1761                                            struct btrfs_ordered_extent,
1762                                            trans_list);
1763                 list_del_init(&ordered->trans_list);
1764                 spin_unlock(&fs_info->trans_lock);
1765
1766                 wait_event(ordered->wait, test_bit(BTRFS_ORDERED_COMPLETE,
1767                                                    &ordered->flags));
1768                 btrfs_put_ordered_extent(ordered);
1769                 spin_lock(&fs_info->trans_lock);
1770         }
1771         spin_unlock(&fs_info->trans_lock);
1772 }
1773
1774 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1775                              struct btrfs_root *root)
1776 {
1777         struct btrfs_transaction *cur_trans = trans->transaction;
1778         struct btrfs_transaction *prev_trans = NULL;
1779         struct btrfs_inode *btree_ino = BTRFS_I(root->fs_info->btree_inode);
1780         int ret;
1781
1782         /* Stop the commit early if ->aborted is set */
1783         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1784                 ret = cur_trans->aborted;
1785                 btrfs_end_transaction(trans, root);
1786                 return ret;
1787         }
1788
1789         /* make a pass through all the delayed refs we have so far
1790          * any runnings procs may add more while we are here
1791          */
1792         ret = btrfs_run_delayed_refs(trans, root, 0);
1793         if (ret) {
1794                 btrfs_end_transaction(trans, root);
1795                 return ret;
1796         }
1797
1798         btrfs_trans_release_metadata(trans, root);
1799         trans->block_rsv = NULL;
1800         if (trans->qgroup_reserved) {
1801                 btrfs_qgroup_free(root, trans->qgroup_reserved);
1802                 trans->qgroup_reserved = 0;
1803         }
1804
1805         cur_trans = trans->transaction;
1806
1807         /*
1808          * set the flushing flag so procs in this transaction have to
1809          * start sending their work down.
1810          */
1811         cur_trans->delayed_refs.flushing = 1;
1812         smp_wmb();
1813
1814         if (!list_empty(&trans->new_bgs))
1815                 btrfs_create_pending_block_groups(trans, root);
1816
1817         ret = btrfs_run_delayed_refs(trans, root, 0);
1818         if (ret) {
1819                 btrfs_end_transaction(trans, root);
1820                 return ret;
1821         }
1822
1823         if (!cur_trans->dirty_bg_run) {
1824                 int run_it = 0;
1825
1826                 /* this mutex is also taken before trying to set
1827                  * block groups readonly.  We need to make sure
1828                  * that nobody has set a block group readonly
1829                  * after a extents from that block group have been
1830                  * allocated for cache files.  btrfs_set_block_group_ro
1831                  * will wait for the transaction to commit if it
1832                  * finds dirty_bg_run = 1
1833                  *
1834                  * The dirty_bg_run flag is also used to make sure only
1835                  * one process starts all the block group IO.  It wouldn't
1836                  * hurt to have more than one go through, but there's no
1837                  * real advantage to it either.
1838                  */
1839                 mutex_lock(&root->fs_info->ro_block_group_mutex);
1840                 if (!cur_trans->dirty_bg_run) {
1841                         run_it = 1;
1842                         cur_trans->dirty_bg_run = 1;
1843                 }
1844                 mutex_unlock(&root->fs_info->ro_block_group_mutex);
1845
1846                 if (run_it)
1847                         ret = btrfs_start_dirty_block_groups(trans, root);
1848         }
1849         if (ret) {
1850                 btrfs_end_transaction(trans, root);
1851                 return ret;
1852         }
1853
1854         spin_lock(&root->fs_info->trans_lock);
1855         list_splice(&trans->ordered, &cur_trans->pending_ordered);
1856         if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
1857                 spin_unlock(&root->fs_info->trans_lock);
1858                 atomic_inc(&cur_trans->use_count);
1859                 ret = btrfs_end_transaction(trans, root);
1860
1861                 wait_for_commit(root, cur_trans);
1862
1863                 if (unlikely(cur_trans->aborted))
1864                         ret = cur_trans->aborted;
1865
1866                 btrfs_put_transaction(cur_trans);
1867
1868                 return ret;
1869         }
1870
1871         cur_trans->state = TRANS_STATE_COMMIT_START;
1872         wake_up(&root->fs_info->transaction_blocked_wait);
1873
1874         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1875                 prev_trans = list_entry(cur_trans->list.prev,
1876                                         struct btrfs_transaction, list);
1877                 if (prev_trans->state != TRANS_STATE_COMPLETED) {
1878                         atomic_inc(&prev_trans->use_count);
1879                         spin_unlock(&root->fs_info->trans_lock);
1880
1881                         wait_for_commit(root, prev_trans);
1882
1883                         btrfs_put_transaction(prev_trans);
1884                 } else {
1885                         spin_unlock(&root->fs_info->trans_lock);
1886                 }
1887         } else {
1888                 spin_unlock(&root->fs_info->trans_lock);
1889         }
1890
1891         extwriter_counter_dec(cur_trans, trans->type);
1892
1893         ret = btrfs_start_delalloc_flush(root->fs_info);
1894         if (ret)
1895                 goto cleanup_transaction;
1896
1897         ret = btrfs_run_delayed_items(trans, root);
1898         if (ret)
1899                 goto cleanup_transaction;
1900
1901         wait_event(cur_trans->writer_wait,
1902                    extwriter_counter_read(cur_trans) == 0);
1903
1904         /* some pending stuffs might be added after the previous flush. */
1905         ret = btrfs_run_delayed_items(trans, root);
1906         if (ret)
1907                 goto cleanup_transaction;
1908
1909         btrfs_wait_delalloc_flush(root->fs_info);
1910
1911         btrfs_wait_pending_ordered(cur_trans, root->fs_info);
1912
1913         btrfs_scrub_pause(root);
1914         /*
1915          * Ok now we need to make sure to block out any other joins while we
1916          * commit the transaction.  We could have started a join before setting
1917          * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
1918          */
1919         spin_lock(&root->fs_info->trans_lock);
1920         cur_trans->state = TRANS_STATE_COMMIT_DOING;
1921         spin_unlock(&root->fs_info->trans_lock);
1922         wait_event(cur_trans->writer_wait,
1923                    atomic_read(&cur_trans->num_writers) == 1);
1924
1925         /* ->aborted might be set after the previous check, so check it */
1926         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1927                 ret = cur_trans->aborted;
1928                 goto scrub_continue;
1929         }
1930         /*
1931          * the reloc mutex makes sure that we stop
1932          * the balancing code from coming in and moving
1933          * extents around in the middle of the commit
1934          */
1935         mutex_lock(&root->fs_info->reloc_mutex);
1936
1937         /*
1938          * We needn't worry about the delayed items because we will
1939          * deal with them in create_pending_snapshot(), which is the
1940          * core function of the snapshot creation.
1941          */
1942         ret = create_pending_snapshots(trans, root->fs_info);
1943         if (ret) {
1944                 mutex_unlock(&root->fs_info->reloc_mutex);
1945                 goto scrub_continue;
1946         }
1947
1948         /*
1949          * We insert the dir indexes of the snapshots and update the inode
1950          * of the snapshots' parents after the snapshot creation, so there
1951          * are some delayed items which are not dealt with. Now deal with
1952          * them.
1953          *
1954          * We needn't worry that this operation will corrupt the snapshots,
1955          * because all the tree which are snapshoted will be forced to COW
1956          * the nodes and leaves.
1957          */
1958         ret = btrfs_run_delayed_items(trans, root);
1959         if (ret) {
1960                 mutex_unlock(&root->fs_info->reloc_mutex);
1961                 goto scrub_continue;
1962         }
1963
1964         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1965         if (ret) {
1966                 mutex_unlock(&root->fs_info->reloc_mutex);
1967                 goto scrub_continue;
1968         }
1969
1970         /* Reocrd old roots for later qgroup accounting */
1971         ret = btrfs_qgroup_prepare_account_extents(trans, root->fs_info);
1972         if (ret) {
1973                 mutex_unlock(&root->fs_info->reloc_mutex);
1974                 goto scrub_continue;
1975         }
1976
1977         /*
1978          * make sure none of the code above managed to slip in a
1979          * delayed item
1980          */
1981         btrfs_assert_delayed_root_empty(root);
1982
1983         WARN_ON(cur_trans != trans->transaction);
1984
1985         /* btrfs_commit_tree_roots is responsible for getting the
1986          * various roots consistent with each other.  Every pointer
1987          * in the tree of tree roots has to point to the most up to date
1988          * root for every subvolume and other tree.  So, we have to keep
1989          * the tree logging code from jumping in and changing any
1990          * of the trees.
1991          *
1992          * At this point in the commit, there can't be any tree-log
1993          * writers, but a little lower down we drop the trans mutex
1994          * and let new people in.  By holding the tree_log_mutex
1995          * from now until after the super is written, we avoid races
1996          * with the tree-log code.
1997          */
1998         mutex_lock(&root->fs_info->tree_log_mutex);
1999
2000         ret = commit_fs_roots(trans, root);
2001         if (ret) {
2002                 mutex_unlock(&root->fs_info->tree_log_mutex);
2003                 mutex_unlock(&root->fs_info->reloc_mutex);
2004                 goto scrub_continue;
2005         }
2006
2007         /*
2008          * Since the transaction is done, we can apply the pending changes
2009          * before the next transaction.
2010          */
2011         btrfs_apply_pending_changes(root->fs_info);
2012
2013         /* commit_fs_roots gets rid of all the tree log roots, it is now
2014          * safe to free the root of tree log roots
2015          */
2016         btrfs_free_log_root_tree(trans, root->fs_info);
2017
2018         /*
2019          * Since fs roots are all committed, we can get a quite accurate
2020          * new_roots. So let's do quota accounting.
2021          */
2022         ret = btrfs_qgroup_account_extents(trans, root->fs_info);
2023         if (ret < 0) {
2024                 mutex_unlock(&root->fs_info->tree_log_mutex);
2025                 mutex_unlock(&root->fs_info->reloc_mutex);
2026                 goto scrub_continue;
2027         }
2028
2029         ret = commit_cowonly_roots(trans, root);
2030         if (ret) {
2031                 mutex_unlock(&root->fs_info->tree_log_mutex);
2032                 mutex_unlock(&root->fs_info->reloc_mutex);
2033                 goto scrub_continue;
2034         }
2035
2036         /*
2037          * The tasks which save the space cache and inode cache may also
2038          * update ->aborted, check it.
2039          */
2040         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
2041                 ret = cur_trans->aborted;
2042                 mutex_unlock(&root->fs_info->tree_log_mutex);
2043                 mutex_unlock(&root->fs_info->reloc_mutex);
2044                 goto scrub_continue;
2045         }
2046
2047         btrfs_prepare_extent_commit(trans, root);
2048
2049         cur_trans = root->fs_info->running_transaction;
2050
2051         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
2052                             root->fs_info->tree_root->node);
2053         list_add_tail(&root->fs_info->tree_root->dirty_list,
2054                       &cur_trans->switch_commits);
2055
2056         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
2057                             root->fs_info->chunk_root->node);
2058         list_add_tail(&root->fs_info->chunk_root->dirty_list,
2059                       &cur_trans->switch_commits);
2060
2061         switch_commit_roots(cur_trans, root->fs_info);
2062
2063         assert_qgroups_uptodate(trans);
2064         ASSERT(list_empty(&cur_trans->dirty_bgs));
2065         ASSERT(list_empty(&cur_trans->io_bgs));
2066         update_super_roots(root);
2067
2068         btrfs_set_super_log_root(root->fs_info->super_copy, 0);
2069         btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
2070         memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
2071                sizeof(*root->fs_info->super_copy));
2072
2073         btrfs_update_commit_device_size(root->fs_info);
2074         btrfs_update_commit_device_bytes_used(root, cur_trans);
2075
2076         clear_bit(BTRFS_INODE_BTREE_LOG1_ERR, &btree_ino->runtime_flags);
2077         clear_bit(BTRFS_INODE_BTREE_LOG2_ERR, &btree_ino->runtime_flags);
2078
2079         btrfs_trans_release_chunk_metadata(trans);
2080
2081         spin_lock(&root->fs_info->trans_lock);
2082         cur_trans->state = TRANS_STATE_UNBLOCKED;
2083         root->fs_info->running_transaction = NULL;
2084         spin_unlock(&root->fs_info->trans_lock);
2085         mutex_unlock(&root->fs_info->reloc_mutex);
2086
2087         wake_up(&root->fs_info->transaction_wait);
2088
2089         ret = btrfs_write_and_wait_transaction(trans, root);
2090         if (ret) {
2091                 btrfs_error(root->fs_info, ret,
2092                             "Error while writing out transaction");
2093                 mutex_unlock(&root->fs_info->tree_log_mutex);
2094                 goto scrub_continue;
2095         }
2096
2097         ret = write_ctree_super(trans, root, 0);
2098         if (ret) {
2099                 mutex_unlock(&root->fs_info->tree_log_mutex);
2100                 goto scrub_continue;
2101         }
2102
2103         /*
2104          * the super is written, we can safely allow the tree-loggers
2105          * to go about their business
2106          */
2107         mutex_unlock(&root->fs_info->tree_log_mutex);
2108
2109         btrfs_finish_extent_commit(trans, root);
2110
2111         if (cur_trans->have_free_bgs)
2112                 btrfs_clear_space_info_full(root->fs_info);
2113
2114         root->fs_info->last_trans_committed = cur_trans->transid;
2115         /*
2116          * We needn't acquire the lock here because there is no other task
2117          * which can change it.
2118          */
2119         cur_trans->state = TRANS_STATE_COMPLETED;
2120         wake_up(&cur_trans->commit_wait);
2121
2122         spin_lock(&root->fs_info->trans_lock);
2123         list_del_init(&cur_trans->list);
2124         spin_unlock(&root->fs_info->trans_lock);
2125
2126         btrfs_put_transaction(cur_trans);
2127         btrfs_put_transaction(cur_trans);
2128
2129         if (trans->type & __TRANS_FREEZABLE)
2130                 sb_end_intwrite(root->fs_info->sb);
2131
2132         trace_btrfs_transaction_commit(root);
2133
2134         btrfs_scrub_continue(root);
2135
2136         if (current->journal_info == trans)
2137                 current->journal_info = NULL;
2138
2139         kmem_cache_free(btrfs_trans_handle_cachep, trans);
2140
2141         if (current != root->fs_info->transaction_kthread)
2142                 btrfs_run_delayed_iputs(root);
2143
2144         return ret;
2145
2146 scrub_continue:
2147         btrfs_scrub_continue(root);
2148 cleanup_transaction:
2149         btrfs_trans_release_metadata(trans, root);
2150         btrfs_trans_release_chunk_metadata(trans);
2151         trans->block_rsv = NULL;
2152         if (trans->qgroup_reserved) {
2153                 btrfs_qgroup_free(root, trans->qgroup_reserved);
2154                 trans->qgroup_reserved = 0;
2155         }
2156         btrfs_warn(root->fs_info, "Skipping commit of aborted transaction.");
2157         if (current->journal_info == trans)
2158                 current->journal_info = NULL;
2159         cleanup_transaction(trans, root, ret);
2160
2161         return ret;
2162 }
2163
2164 /*
2165  * return < 0 if error
2166  * 0 if there are no more dead_roots at the time of call
2167  * 1 there are more to be processed, call me again
2168  *
2169  * The return value indicates there are certainly more snapshots to delete, but
2170  * if there comes a new one during processing, it may return 0. We don't mind,
2171  * because btrfs_commit_super will poke cleaner thread and it will process it a
2172  * few seconds later.
2173  */
2174 int btrfs_clean_one_deleted_snapshot(struct btrfs_root *root)
2175 {
2176         int ret;
2177         struct btrfs_fs_info *fs_info = root->fs_info;
2178
2179         spin_lock(&fs_info->trans_lock);
2180         if (list_empty(&fs_info->dead_roots)) {
2181                 spin_unlock(&fs_info->trans_lock);
2182                 return 0;
2183         }
2184         root = list_first_entry(&fs_info->dead_roots,
2185                         struct btrfs_root, root_list);
2186         list_del_init(&root->root_list);
2187         spin_unlock(&fs_info->trans_lock);
2188
2189         pr_debug("BTRFS: cleaner removing %llu\n", root->objectid);
2190
2191         btrfs_kill_all_delayed_nodes(root);
2192
2193         if (btrfs_header_backref_rev(root->node) <
2194                         BTRFS_MIXED_BACKREF_REV)
2195                 ret = btrfs_drop_snapshot(root, NULL, 0, 0);
2196         else
2197                 ret = btrfs_drop_snapshot(root, NULL, 1, 0);
2198
2199         return (ret < 0) ? 0 : 1;
2200 }
2201
2202 void btrfs_apply_pending_changes(struct btrfs_fs_info *fs_info)
2203 {
2204         unsigned long prev;
2205         unsigned long bit;
2206
2207         prev = xchg(&fs_info->pending_changes, 0);
2208         if (!prev)
2209                 return;
2210
2211         bit = 1 << BTRFS_PENDING_SET_INODE_MAP_CACHE;
2212         if (prev & bit)
2213                 btrfs_set_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2214         prev &= ~bit;
2215
2216         bit = 1 << BTRFS_PENDING_CLEAR_INODE_MAP_CACHE;
2217         if (prev & bit)
2218                 btrfs_clear_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2219         prev &= ~bit;
2220
2221         bit = 1 << BTRFS_PENDING_COMMIT;
2222         if (prev & bit)
2223                 btrfs_debug(fs_info, "pending commit done");
2224         prev &= ~bit;
2225
2226         if (prev)
2227                 btrfs_warn(fs_info,
2228                         "unknown pending changes left 0x%lx, ignoring", prev);
2229 }