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(refcount_read(&transaction->use_count) == 0);
  64         if (refcount_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                         btrfs_err(transaction->fs_info,
  69                                   "pending csums is %llu",
  70                                   transaction->delayed_refs.pending_csums);
  71                 while (!list_empty(&transaction->pending_chunks)) {
  72                         struct extent_map *em;
  73
  74                         em = list_first_entry(&transaction->pending_chunks,
  75                                               struct extent_map, list);
  76                         list_del_init(&em->list);
  77                         free_extent_map(em);
  78                 }
  79                 /*
  80                  * If any block groups are found in ->deleted_bgs then it's
  81                  * because the transaction was aborted and a commit did not
  82                  * happen (things failed before writing the new superblock
  83                  * and calling btrfs_finish_extent_commit()), so we can not
  84                  * discard the physical locations of the block groups.
  85                  */
  86                 while (!list_empty(&transaction->deleted_bgs)) {
  87                         struct btrfs_block_group_cache *cache;
  88
  89                         cache = list_first_entry(&transaction->deleted_bgs,
  90                                                  struct btrfs_block_group_cache,
  91                                                  bg_list);
  92                         list_del_init(&cache->bg_list);
  93                         btrfs_put_block_group_trimming(cache);
  94                         btrfs_put_block_group(cache);
  95                 }
  96                 kfree(transaction);
  97         }
  98 }
  99
 100 static void clear_btree_io_tree(struct extent_io_tree *tree)
 101 {
 102         spin_lock(&tree->lock);
 103         /*
 104          * Do a single barrier for the waitqueue_active check here, the state
 105          * of the waitqueue should not change once clear_btree_io_tree is
 106          * called.
 107          */
 108         smp_mb();
 109         while (!RB_EMPTY_ROOT(&tree->state)) {
 110                 struct rb_node *node;
 111                 struct extent_state *state;
 112
 113                 node = rb_first(&tree->state);
 114                 state = rb_entry(node, struct extent_state, rb_node);
 115                 rb_erase(&state->rb_node, &tree->state);
 116                 RB_CLEAR_NODE(&state->rb_node);
 117                 /*
 118                  * btree io trees aren't supposed to have tasks waiting for
 119                  * changes in the flags of extent states ever.
 120                  */
 121                 ASSERT(!waitqueue_active(&state->wq));
 122                 free_extent_state(state);
 123
 124                 cond_resched_lock(&tree->lock);
 125         }
 126         spin_unlock(&tree->lock);
 127 }
 128
 129 static noinline void switch_commit_roots(struct btrfs_transaction *trans,
 130                                          struct btrfs_fs_info *fs_info)
 131 {
 132         struct btrfs_root *root, *tmp;
 133
 134         down_write(&fs_info->commit_root_sem);
 135         list_for_each_entry_safe(root, tmp, &trans->switch_commits,
 136                                  dirty_list) {
 137                 list_del_init(&root->dirty_list);
 138                 free_extent_buffer(root->commit_root);
 139                 root->commit_root = btrfs_root_node(root);
 140                 if (is_fstree(root->objectid))
 141                         btrfs_unpin_free_ino(root);
 142                 clear_btree_io_tree(&root->dirty_log_pages);
 143         }
 144
 145         /* We can free old roots now. */
 146         spin_lock(&trans->dropped_roots_lock);
 147         while (!list_empty(&trans->dropped_roots)) {
 148                 root = list_first_entry(&trans->dropped_roots,
 149                                         struct btrfs_root, root_list);
 150                 list_del_init(&root->root_list);
 151                 spin_unlock(&trans->dropped_roots_lock);
 152                 btrfs_drop_and_free_fs_root(fs_info, root);
 153                 spin_lock(&trans->dropped_roots_lock);
 154         }
 155         spin_unlock(&trans->dropped_roots_lock);
 156         up_write(&fs_info->commit_root_sem);
 157 }
 158
 159 static inline void extwriter_counter_inc(struct btrfs_transaction *trans,
 160                                          unsigned int type)
 161 {
 162         if (type & TRANS_EXTWRITERS)
 163                 atomic_inc(&trans->num_extwriters);
 164 }
 165
 166 static inline void extwriter_counter_dec(struct btrfs_transaction *trans,
 167                                          unsigned int type)
 168 {
 169         if (type & TRANS_EXTWRITERS)
 170                 atomic_dec(&trans->num_extwriters);
 171 }
 172
 173 static inline void extwriter_counter_init(struct btrfs_transaction *trans,
 174                                           unsigned int type)
 175 {
 176         atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0));
 177 }
 178
 179 static inline int extwriter_counter_read(struct btrfs_transaction *trans)
 180 {
 181         return atomic_read(&trans->num_extwriters);
 182 }
 183
 184 /*
 185  * either allocate a new transaction or hop into the existing one
 186  */
 187 static noinline int join_transaction(struct btrfs_fs_info *fs_info,
 188                                      unsigned int type)
 189 {
 190         struct btrfs_transaction *cur_trans;
 191
 192         spin_lock(&fs_info->trans_lock);
 193 loop:
 194         /* The file system has been taken offline. No new transactions. */
 195         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
 196                 spin_unlock(&fs_info->trans_lock);
 197                 return -EROFS;
 198         }
 199
 200         cur_trans = fs_info->running_transaction;
 201         if (cur_trans) {
 202                 if (cur_trans->aborted) {
 203                         spin_unlock(&fs_info->trans_lock);
 204                         return cur_trans->aborted;
 205                 }
 206                 if (btrfs_blocked_trans_types[cur_trans->state] & type) {
 207                         spin_unlock(&fs_info->trans_lock);
 208                         return -EBUSY;
 209                 }
 210                 refcount_inc(&cur_trans->use_count);
 211                 atomic_inc(&cur_trans->num_writers);
 212                 extwriter_counter_inc(cur_trans, type);
 213                 spin_unlock(&fs_info->trans_lock);
 214                 return 0;
 215         }
 216         spin_unlock(&fs_info->trans_lock);
 217
 218         /*
 219          * If we are ATTACH, we just want to catch the current transaction,
 220          * and commit it. If there is no transaction, just return ENOENT.
 221          */
 222         if (type == TRANS_ATTACH)
 223                 return -ENOENT;
 224
 225         /*
 226          * JOIN_NOLOCK only happens during the transaction commit, so
 227          * it is impossible that ->running_transaction is NULL
 228          */
 229         BUG_ON(type == TRANS_JOIN_NOLOCK);
 230
 231         cur_trans = kmalloc(sizeof(*cur_trans), GFP_NOFS);
 232         if (!cur_trans)
 233                 return -ENOMEM;
 234
 235         spin_lock(&fs_info->trans_lock);
 236         if (fs_info->running_transaction) {
 237                 /*
 238                  * someone started a transaction after we unlocked.  Make sure
 239                  * to redo the checks above
 240                  */
 241                 kfree(cur_trans);
 242                 goto loop;
 243         } else if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
 244                 spin_unlock(&fs_info->trans_lock);
 245                 kfree(cur_trans);
 246                 return -EROFS;
 247         }
 248
 249         cur_trans->fs_info = fs_info;
 250         atomic_set(&cur_trans->num_writers, 1);
 251         extwriter_counter_init(cur_trans, type);
 252         init_waitqueue_head(&cur_trans->writer_wait);
 253         init_waitqueue_head(&cur_trans->commit_wait);
 254         init_waitqueue_head(&cur_trans->pending_wait);
 255         cur_trans->state = TRANS_STATE_RUNNING;
 256         /*
 257          * One for this trans handle, one so it will live on until we
 258          * commit the transaction.
 259          */
 260         refcount_set(&cur_trans->use_count, 2);
 261         atomic_set(&cur_trans->pending_ordered, 0);
 262         cur_trans->flags = 0;
 263         cur_trans->start_time = get_seconds();
 264
 265         memset(&cur_trans->delayed_refs, 0, sizeof(cur_trans->delayed_refs));
 266
 267         cur_trans->delayed_refs.href_root = RB_ROOT;
 268         cur_trans->delayed_refs.dirty_extent_root = RB_ROOT;
 269         atomic_set(&cur_trans->delayed_refs.num_entries, 0);
 270
 271         /*
 272          * although the tree mod log is per file system and not per transaction,
 273          * the log must never go across transaction boundaries.
 274          */
 275         smp_mb();
 276         if (!list_empty(&fs_info->tree_mod_seq_list))
 277                 WARN(1, KERN_ERR "BTRFS: tree_mod_seq_list not empty when creating a fresh transaction\n");
 278         if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
 279                 WARN(1, KERN_ERR "BTRFS: tree_mod_log rb tree not empty when creating a fresh transaction\n");
 280         atomic64_set(&fs_info->tree_mod_seq, 0);
 281
 282         spin_lock_init(&cur_trans->delayed_refs.lock);
 283
 284         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 285         INIT_LIST_HEAD(&cur_trans->pending_chunks);
 286         INIT_LIST_HEAD(&cur_trans->switch_commits);
 287         INIT_LIST_HEAD(&cur_trans->dirty_bgs);
 288         INIT_LIST_HEAD(&cur_trans->io_bgs);
 289         INIT_LIST_HEAD(&cur_trans->dropped_roots);
 290         mutex_init(&cur_trans->cache_write_mutex);
 291         cur_trans->num_dirty_bgs = 0;
 292         spin_lock_init(&cur_trans->dirty_bgs_lock);
 293         INIT_LIST_HEAD(&cur_trans->deleted_bgs);
 294         spin_lock_init(&cur_trans->dropped_roots_lock);
 295         list_add_tail(&cur_trans->list, &fs_info->trans_list);
 296         extent_io_tree_init(&cur_trans->dirty_pages,
 297                              fs_info->btree_inode);
 298         fs_info->generation++;
 299         cur_trans->transid = fs_info->generation;
 300         fs_info->running_transaction = cur_trans;
 301         cur_trans->aborted = 0;
 302         spin_unlock(&fs_info->trans_lock);
 303
 304         return 0;
 305 }
 306
 307 /*
 308  * this does all the record keeping required to make sure that a reference
 309  * counted root is properly recorded in a given transaction.  This is required
 310  * to make sure the old root from before we joined the transaction is deleted
 311  * when the transaction commits
 312  */
 313 static int record_root_in_trans(struct btrfs_trans_handle *trans,
 314                                struct btrfs_root *root,
 315                                int force)
 316 {
 317         struct btrfs_fs_info *fs_info = root->fs_info;
 318
 319         if ((test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
 320             root->last_trans < trans->transid) || force) {
 321                 WARN_ON(root == fs_info->extent_root);
 322                 WARN_ON(root->commit_root != root->node);
 323
 324                 /*
 325                  * see below for IN_TRANS_SETUP usage rules
 326                  * we have the reloc mutex held now, so there
 327                  * is only one writer in this function
 328                  */
 329                 set_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
 330
 331                 /* make sure readers find IN_TRANS_SETUP before
 332                  * they find our root->last_trans update
 333                  */
 334                 smp_wmb();
 335
 336                 spin_lock(&fs_info->fs_roots_radix_lock);
 337                 if (root->last_trans == trans->transid && !force) {
 338                         spin_unlock(&fs_info->fs_roots_radix_lock);
 339                         return 0;
 340                 }
 341                 radix_tree_tag_set(&fs_info->fs_roots_radix,
 342                                    (unsigned long)root->root_key.objectid,
 343                                    BTRFS_ROOT_TRANS_TAG);
 344                 spin_unlock(&fs_info->fs_roots_radix_lock);
 345                 root->last_trans = trans->transid;
 346
 347                 /* this is pretty tricky.  We don't want to
 348                  * take the relocation lock in btrfs_record_root_in_trans
 349                  * unless we're really doing the first setup for this root in
 350                  * this transaction.
 351                  *
 352                  * Normally we'd use root->last_trans as a flag to decide
 353                  * if we want to take the expensive mutex.
 354                  *
 355                  * But, we have to set root->last_trans before we
 356                  * init the relocation root, otherwise, we trip over warnings
 357                  * in ctree.c.  The solution used here is to flag ourselves
 358                  * with root IN_TRANS_SETUP.  When this is 1, we're still
 359                  * fixing up the reloc trees and everyone must wait.
 360                  *
 361                  * When this is zero, they can trust root->last_trans and fly
 362                  * through btrfs_record_root_in_trans without having to take the
 363                  * lock.  smp_wmb() makes sure that all the writes above are
 364                  * done before we pop in the zero below
 365                  */
 366                 btrfs_init_reloc_root(trans, root);
 367                 smp_mb__before_atomic();
 368                 clear_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
 369         }
 370         return 0;
 371 }
 372
 373
 374 void btrfs_add_dropped_root(struct btrfs_trans_handle *trans,
 375                             struct btrfs_root *root)
 376 {
 377         struct btrfs_fs_info *fs_info = root->fs_info;
 378         struct btrfs_transaction *cur_trans = trans->transaction;
 379
 380         /* Add ourselves to the transaction dropped list */
 381         spin_lock(&cur_trans->dropped_roots_lock);
 382         list_add_tail(&root->root_list, &cur_trans->dropped_roots);
 383         spin_unlock(&cur_trans->dropped_roots_lock);
 384
 385         /* Make sure we don't try to update the root at commit time */
 386         spin_lock(&fs_info->fs_roots_radix_lock);
 387         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 388                              (unsigned long)root->root_key.objectid,
 389                              BTRFS_ROOT_TRANS_TAG);
 390         spin_unlock(&fs_info->fs_roots_radix_lock);
 391 }
 392
 393 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 394                                struct btrfs_root *root)
 395 {
 396         struct btrfs_fs_info *fs_info = root->fs_info;
 397
 398         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
 399                 return 0;
 400
 401         /*
 402          * see record_root_in_trans for comments about IN_TRANS_SETUP usage
 403          * and barriers
 404          */
 405         smp_rmb();
 406         if (root->last_trans == trans->transid &&
 407             !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state))
 408                 return 0;
 409
 410         mutex_lock(&fs_info->reloc_mutex);
 411         record_root_in_trans(trans, root, 0);
 412         mutex_unlock(&fs_info->reloc_mutex);
 413
 414         return 0;
 415 }
 416
 417 static inline int is_transaction_blocked(struct btrfs_transaction *trans)
 418 {
 419         return (trans->state >= TRANS_STATE_BLOCKED &&
 420                 trans->state < TRANS_STATE_UNBLOCKED &&
 421                 !trans->aborted);
 422 }
 423
 424 /* wait for commit against the current transaction to become unblocked
 425  * when this is done, it is safe to start a new transaction, but the current
 426  * transaction might not be fully on disk.
 427  */
 428 static void wait_current_trans(struct btrfs_fs_info *fs_info)
 429 {
 430         struct btrfs_transaction *cur_trans;
 431
 432         spin_lock(&fs_info->trans_lock);
 433         cur_trans = fs_info->running_transaction;
 434         if (cur_trans && is_transaction_blocked(cur_trans)) {
 435                 refcount_inc(&cur_trans->use_count);
 436                 spin_unlock(&fs_info->trans_lock);
 437
 438                 wait_event(fs_info->transaction_wait,
 439                            cur_trans->state >= TRANS_STATE_UNBLOCKED ||
 440                            cur_trans->aborted);
 441                 btrfs_put_transaction(cur_trans);
 442         } else {
 443                 spin_unlock(&fs_info->trans_lock);
 444         }
 445 }
 446
 447 static int may_wait_transaction(struct btrfs_fs_info *fs_info, int type)
 448 {
 449         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
 450                 return 0;
 451
 452         if (type == TRANS_USERSPACE)
 453                 return 1;
 454
 455         if (type == TRANS_START &&
 456             !atomic_read(&fs_info->open_ioctl_trans))
 457                 return 1;
 458
 459         return 0;
 460 }
 461
 462 static inline bool need_reserve_reloc_root(struct btrfs_root *root)
 463 {
 464         struct btrfs_fs_info *fs_info = root->fs_info;
 465
 466         if (!fs_info->reloc_ctl ||
 467             !test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
 468             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 469             root->reloc_root)
 470                 return false;
 471
 472         return true;
 473 }
 474
 475 static struct btrfs_trans_handle *
 476 start_transaction(struct btrfs_root *root, unsigned int num_items,
 477                   unsigned int type, enum btrfs_reserve_flush_enum flush,
 478                   bool enforce_qgroups)
 479 {
 480         struct btrfs_fs_info *fs_info = root->fs_info;
 481
 482         struct btrfs_trans_handle *h;
 483         struct btrfs_transaction *cur_trans;
 484         u64 num_bytes = 0;
 485         u64 qgroup_reserved = 0;
 486         bool reloc_reserved = false;
 487         int ret;
 488
 489         /* Send isn't supposed to start transactions. */
 490         ASSERT(current->journal_info != BTRFS_SEND_TRANS_STUB);
 491
 492         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
 493                 return ERR_PTR(-EROFS);
 494
 495         if (current->journal_info) {
 496                 WARN_ON(type & TRANS_EXTWRITERS);
 497                 h = current->journal_info;
 498                 refcount_inc(&h->use_count);
 499                 WARN_ON(refcount_read(&h->use_count) > 2);
 500                 h->orig_rsv = h->block_rsv;
 501                 h->block_rsv = NULL;
 502                 goto got_it;
 503         }
 504
 505         /*
 506          * Do the reservation before we join the transaction so we can do all
 507          * the appropriate flushing if need be.
 508          */
 509         if (num_items && root != fs_info->chunk_root) {
 510                 qgroup_reserved = num_items * fs_info->nodesize;
 511                 ret = btrfs_qgroup_reserve_meta(root, qgroup_reserved,
 512                                                 enforce_qgroups);
 513                 if (ret)
 514                         return ERR_PTR(ret);
 515
 516                 num_bytes = btrfs_calc_trans_metadata_size(fs_info, num_items);
 517                 /*
 518                  * Do the reservation for the relocation root creation
 519                  */
 520                 if (need_reserve_reloc_root(root)) {
 521                         num_bytes += fs_info->nodesize;
 522                         reloc_reserved = true;
 523                 }
 524
 525                 ret = btrfs_block_rsv_add(root, &fs_info->trans_block_rsv,
 526                                           num_bytes, flush);
 527                 if (ret)
 528                         goto reserve_fail;
 529         }
 530 again:
 531         h = kmem_cache_zalloc(btrfs_trans_handle_cachep, GFP_NOFS);
 532         if (!h) {
 533                 ret = -ENOMEM;
 534                 goto alloc_fail;
 535         }
 536
 537         /*
 538          * If we are JOIN_NOLOCK we're already committing a transaction and
 539          * waiting on this guy, so we don't need to do the sb_start_intwrite
 540          * because we're already holding a ref.  We need this because we could
 541          * have raced in and did an fsync() on a file which can kick a commit
 542          * and then we deadlock with somebody doing a freeze.
 543          *
 544          * If we are ATTACH, it means we just want to catch the current
 545          * transaction and commit it, so we needn't do sb_start_intwrite().
 546          */
 547         if (type & __TRANS_FREEZABLE)
 548                 sb_start_intwrite(fs_info->sb);
 549
 550         if (may_wait_transaction(fs_info, type))
 551                 wait_current_trans(fs_info);
 552
 553         do {
 554                 ret = join_transaction(fs_info, type);
 555                 if (ret == -EBUSY) {
 556                         wait_current_trans(fs_info);
 557                         if (unlikely(type == TRANS_ATTACH))
 558                                 ret = -ENOENT;
 559                 }
 560         } while (ret == -EBUSY);
 561
 562         if (ret < 0)
 563                 goto join_fail;
 564
 565         cur_trans = fs_info->running_transaction;
 566
 567         h->transid = cur_trans->transid;
 568         h->transaction = cur_trans;
 569         h->root = root;
 570         refcount_set(&h->use_count, 1);
 571         h->fs_info = root->fs_info;
 572
 573         h->type = type;
 574         h->can_flush_pending_bgs = true;
 575         INIT_LIST_HEAD(&h->new_bgs);
 576
 577         smp_mb();
 578         if (cur_trans->state >= TRANS_STATE_BLOCKED &&
 579             may_wait_transaction(fs_info, type)) {
 580                 current->journal_info = h;
 581                 btrfs_commit_transaction(h);
 582                 goto again;
 583         }
 584
 585         if (num_bytes) {
 586                 trace_btrfs_space_reservation(fs_info, "transaction",
 587                                               h->transid, num_bytes, 1);
 588                 h->block_rsv = &fs_info->trans_block_rsv;
 589                 h->bytes_reserved = num_bytes;
 590                 h->reloc_reserved = reloc_reserved;
 591         }
 592
 593 got_it:
 594         btrfs_record_root_in_trans(h, root);
 595
 596         if (!current->journal_info && type != TRANS_USERSPACE)
 597                 current->journal_info = h;
 598         return h;
 599
 600 join_fail:
 601         if (type & __TRANS_FREEZABLE)
 602                 sb_end_intwrite(fs_info->sb);
 603         kmem_cache_free(btrfs_trans_handle_cachep, h);
 604 alloc_fail:
 605         if (num_bytes)
 606                 btrfs_block_rsv_release(fs_info, &fs_info->trans_block_rsv,
 607                                         num_bytes);
 608 reserve_fail:
 609         btrfs_qgroup_free_meta(root, qgroup_reserved);
 610         return ERR_PTR(ret);
 611 }
 612
 613 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 614                                                    unsigned int num_items)
 615 {
 616         return start_transaction(root, num_items, TRANS_START,
 617                                  BTRFS_RESERVE_FLUSH_ALL, true);
 618 }
 619
 620 struct btrfs_trans_handle *btrfs_start_transaction_fallback_global_rsv(
 621                                         struct btrfs_root *root,
 622                                         unsigned int num_items,
 623                                         int min_factor)
 624 {
 625         struct btrfs_fs_info *fs_info = root->fs_info;
 626         struct btrfs_trans_handle *trans;
 627         u64 num_bytes;
 628         int ret;
 629
 630         /*
 631          * We have two callers: unlink and block group removal.  The
 632          * former should succeed even if we will temporarily exceed
 633          * quota and the latter operates on the extent root so
 634          * qgroup enforcement is ignored anyway.
 635          */
 636         trans = start_transaction(root, num_items, TRANS_START,
 637                                   BTRFS_RESERVE_FLUSH_ALL, false);
 638         if (!IS_ERR(trans) || PTR_ERR(trans) != -ENOSPC)
 639                 return trans;
 640
 641         trans = btrfs_start_transaction(root, 0);
 642         if (IS_ERR(trans))
 643                 return trans;
 644
 645         num_bytes = btrfs_calc_trans_metadata_size(fs_info, num_items);
 646         ret = btrfs_cond_migrate_bytes(fs_info, &fs_info->trans_block_rsv,
 647                                        num_bytes, min_factor);
 648         if (ret) {
 649                 btrfs_end_transaction(trans);
 650                 return ERR_PTR(ret);
 651         }
 652
 653         trans->block_rsv = &fs_info->trans_block_rsv;
 654         trans->bytes_reserved = num_bytes;
 655         trace_btrfs_space_reservation(fs_info, "transaction",
 656                                       trans->transid, num_bytes, 1);
 657
 658         return trans;
 659 }
 660
 661 struct btrfs_trans_handle *btrfs_start_transaction_lflush(
 662                                         struct btrfs_root *root,
 663                                         unsigned int num_items)
 664 {
 665         return start_transaction(root, num_items, TRANS_START,
 666                                  BTRFS_RESERVE_FLUSH_LIMIT, true);
 667 }
 668
 669 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 670 {
 671         return start_transaction(root, 0, TRANS_JOIN, BTRFS_RESERVE_NO_FLUSH,
 672                                  true);
 673 }
 674
 675 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 676 {
 677         return start_transaction(root, 0, TRANS_JOIN_NOLOCK,
 678                                  BTRFS_RESERVE_NO_FLUSH, true);
 679 }
 680
 681 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 682 {
 683         return start_transaction(root, 0, TRANS_USERSPACE,
 684                                  BTRFS_RESERVE_NO_FLUSH, true);
 685 }
 686
 687 /*
 688  * btrfs_attach_transaction() - catch the running transaction
 689  *
 690  * It is used when we want to commit the current the transaction, but
 691  * don't want to start a new one.
 692  *
 693  * Note: If this function return -ENOENT, it just means there is no
 694  * running transaction. But it is possible that the inactive transaction
 695  * is still in the memory, not fully on disk. If you hope there is no
 696  * inactive transaction in the fs when -ENOENT is returned, you should
 697  * invoke
 698  *     btrfs_attach_transaction_barrier()
 699  */
 700 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
 701 {
 702         return start_transaction(root, 0, TRANS_ATTACH,
 703                                  BTRFS_RESERVE_NO_FLUSH, true);
 704 }
 705
 706 /*
 707  * btrfs_attach_transaction_barrier() - catch the running transaction
 708  *
 709  * It is similar to the above function, the differentia is this one
 710  * will wait for all the inactive transactions until they fully
 711  * complete.
 712  */
 713 struct btrfs_trans_handle *
 714 btrfs_attach_transaction_barrier(struct btrfs_root *root)
 715 {
 716         struct btrfs_trans_handle *trans;
 717
 718         trans = start_transaction(root, 0, TRANS_ATTACH,
 719                                   BTRFS_RESERVE_NO_FLUSH, true);
 720         if (IS_ERR(trans) && PTR_ERR(trans) == -ENOENT)
 721                 btrfs_wait_for_commit(root->fs_info, 0);
 722
 723         return trans;
 724 }
 725
 726 /* wait for a transaction commit to be fully complete */
 727 static noinline void wait_for_commit(struct btrfs_transaction *commit)
 728 {
 729         wait_event(commit->commit_wait, commit->state == TRANS_STATE_COMPLETED);
 730 }
 731
 732 int btrfs_wait_for_commit(struct btrfs_fs_info *fs_info, u64 transid)
 733 {
 734         struct btrfs_transaction *cur_trans = NULL, *t;
 735         int ret = 0;
 736
 737         if (transid) {
 738                 if (transid <= fs_info->last_trans_committed)
 739                         goto out;
 740
 741                 /* find specified transaction */
 742                 spin_lock(&fs_info->trans_lock);
 743                 list_for_each_entry(t, &fs_info->trans_list, list) {
 744                         if (t->transid == transid) {
 745                                 cur_trans = t;
 746                                 refcount_inc(&cur_trans->use_count);
 747                                 ret = 0;
 748                                 break;
 749                         }
 750                         if (t->transid > transid) {
 751                                 ret = 0;
 752                                 break;
 753                         }
 754                 }
 755                 spin_unlock(&fs_info->trans_lock);
 756
 757                 /*
 758                  * The specified transaction doesn't exist, or we
 759                  * raced with btrfs_commit_transaction
 760                  */
 761                 if (!cur_trans) {
 762                         if (transid > fs_info->last_trans_committed)
 763                                 ret = -EINVAL;
 764                         goto out;
 765                 }
 766         } else {
 767                 /* find newest transaction that is committing | committed */
 768                 spin_lock(&fs_info->trans_lock);
 769                 list_for_each_entry_reverse(t, &fs_info->trans_list,
 770                                             list) {
 771                         if (t->state >= TRANS_STATE_COMMIT_START) {
 772                                 if (t->state == TRANS_STATE_COMPLETED)
 773                                         break;
 774                                 cur_trans = t;
 775                                 refcount_inc(&cur_trans->use_count);
 776                                 break;
 777                         }
 778                 }
 779                 spin_unlock(&fs_info->trans_lock);
 780                 if (!cur_trans)
 781                         goto out;  /* nothing committing|committed */
 782         }
 783
 784         wait_for_commit(cur_trans);
 785         btrfs_put_transaction(cur_trans);
 786 out:
 787         return ret;
 788 }
 789
 790 void btrfs_throttle(struct btrfs_fs_info *fs_info)
 791 {
 792         if (!atomic_read(&fs_info->open_ioctl_trans))
 793                 wait_current_trans(fs_info);
 794 }
 795
 796 static int should_end_transaction(struct btrfs_trans_handle *trans)
 797 {
 798         struct btrfs_fs_info *fs_info = trans->fs_info;
 799
 800         if (btrfs_check_space_for_delayed_refs(trans, fs_info))
 801                 return 1;
 802
 803         return !!btrfs_block_rsv_check(&fs_info->global_block_rsv, 5);
 804 }
 805
 806 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans)
 807 {
 808         struct btrfs_transaction *cur_trans = trans->transaction;
 809         struct btrfs_fs_info *fs_info = trans->fs_info;
 810         int updates;
 811         int err;
 812
 813         smp_mb();
 814         if (cur_trans->state >= TRANS_STATE_BLOCKED ||
 815             cur_trans->delayed_refs.flushing)
 816                 return 1;
 817
 818         updates = trans->delayed_ref_updates;
 819         trans->delayed_ref_updates = 0;
 820         if (updates) {
 821                 err = btrfs_run_delayed_refs(trans, fs_info, updates * 2);
 822                 if (err) /* Error code will also eval true */
 823                         return err;
 824         }
 825
 826         return should_end_transaction(trans);
 827 }
 828
 829 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 830                                    int throttle)
 831 {
 832         struct btrfs_fs_info *info = trans->fs_info;
 833         struct btrfs_transaction *cur_trans = trans->transaction;
 834         u64 transid = trans->transid;
 835         unsigned long cur = trans->delayed_ref_updates;
 836         int lock = (trans->type != TRANS_JOIN_NOLOCK);
 837         int err = 0;
 838         int must_run_delayed_refs = 0;
 839
 840         if (refcount_read(&trans->use_count) > 1) {
 841                 refcount_dec(&trans->use_count);
 842                 trans->block_rsv = trans->orig_rsv;
 843                 return 0;
 844         }
 845
 846         btrfs_trans_release_metadata(trans, info);
 847         trans->block_rsv = NULL;
 848
 849         if (!list_empty(&trans->new_bgs))
 850                 btrfs_create_pending_block_groups(trans, info);
 851
 852         trans->delayed_ref_updates = 0;
 853         if (!trans->sync) {
 854                 must_run_delayed_refs =
 855                         btrfs_should_throttle_delayed_refs(trans, info);
 856                 cur = max_t(unsigned long, cur, 32);
 857
 858                 /*
 859                  * don't make the caller wait if they are from a NOLOCK
 860                  * or ATTACH transaction, it will deadlock with commit
 861                  */
 862                 if (must_run_delayed_refs == 1 &&
 863                     (trans->type & (__TRANS_JOIN_NOLOCK | __TRANS_ATTACH)))
 864                         must_run_delayed_refs = 2;
 865         }
 866
 867         btrfs_trans_release_metadata(trans, info);
 868         trans->block_rsv = NULL;
 869
 870         if (!list_empty(&trans->new_bgs))
 871                 btrfs_create_pending_block_groups(trans, info);
 872
 873         btrfs_trans_release_chunk_metadata(trans);
 874
 875         if (lock && !atomic_read(&info->open_ioctl_trans) &&
 876             should_end_transaction(trans) &&
 877             READ_ONCE(cur_trans->state) == TRANS_STATE_RUNNING) {
 878                 spin_lock(&info->trans_lock);
 879                 if (cur_trans->state == TRANS_STATE_RUNNING)
 880                         cur_trans->state = TRANS_STATE_BLOCKED;
 881                 spin_unlock(&info->trans_lock);
 882         }
 883
 884         if (lock && READ_ONCE(cur_trans->state) == TRANS_STATE_BLOCKED) {
 885                 if (throttle)
 886                         return btrfs_commit_transaction(trans);
 887                 else
 888                         wake_up_process(info->transaction_kthread);
 889         }
 890
 891         if (trans->type & __TRANS_FREEZABLE)
 892                 sb_end_intwrite(info->sb);
 893
 894         WARN_ON(cur_trans != info->running_transaction);
 895         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 896         atomic_dec(&cur_trans->num_writers);
 897         extwriter_counter_dec(cur_trans, trans->type);
 898
 899         /*
 900          * Make sure counter is updated before we wake up waiters.
 901          */
 902         smp_mb();
 903         if (waitqueue_active(&cur_trans->writer_wait))
 904                 wake_up(&cur_trans->writer_wait);
 905         btrfs_put_transaction(cur_trans);
 906
 907         if (current->journal_info == trans)
 908                 current->journal_info = NULL;
 909
 910         if (throttle)
 911                 btrfs_run_delayed_iputs(info);
 912
 913         if (trans->aborted ||
 914             test_bit(BTRFS_FS_STATE_ERROR, &info->fs_state)) {
 915                 wake_up_process(info->transaction_kthread);
 916                 err = -EIO;
 917         }
 918
 919         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 920         if (must_run_delayed_refs) {
 921                 btrfs_async_run_delayed_refs(info, cur, transid,
 922                                              must_run_delayed_refs == 1);
 923         }
 924         return err;
 925 }
 926
 927 int btrfs_end_transaction(struct btrfs_trans_handle *trans)
 928 {
 929         return __btrfs_end_transaction(trans, 0);
 930 }
 931
 932 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans)
 933 {
 934         return __btrfs_end_transaction(trans, 1);
 935 }
 936
 937 /*
 938  * when btree blocks are allocated, they have some corresponding bits set for
 939  * them in one of two extent_io trees.  This is used to make sure all of
 940  * those extents are sent to disk but does not wait on them
 941  */
 942 int btrfs_write_marked_extents(struct btrfs_fs_info *fs_info,
 943                                struct extent_io_tree *dirty_pages, int mark)
 944 {
 945         int err = 0;
 946         int werr = 0;
 947         struct address_space *mapping = fs_info->btree_inode->i_mapping;
 948         struct extent_state *cached_state = NULL;
 949         u64 start = 0;
 950         u64 end;
 951
 952         atomic_inc(&BTRFS_I(fs_info->btree_inode)->sync_writers);
 953         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 954                                       mark, &cached_state)) {
 955                 bool wait_writeback = false;
 956
 957                 err = convert_extent_bit(dirty_pages, start, end,
 958                                          EXTENT_NEED_WAIT,
 959                                          mark, &cached_state);
 960                 /*
 961                  * convert_extent_bit can return -ENOMEM, which is most of the
 962                  * time a temporary error. So when it happens, ignore the error
 963                  * and wait for writeback of this range to finish - because we
 964                  * failed to set the bit EXTENT_NEED_WAIT for the range, a call
 965                  * to __btrfs_wait_marked_extents() would not know that
 966                  * writeback for this range started and therefore wouldn't
 967                  * wait for it to finish - we don't want to commit a
 968                  * superblock that points to btree nodes/leafs for which
 969                  * writeback hasn't finished yet (and without errors).
 970                  * We cleanup any entries left in the io tree when committing
 971                  * the transaction (through clear_btree_io_tree()).
 972                  */
 973                 if (err == -ENOMEM) {
 974                         err = 0;
 975                         wait_writeback = true;
 976                 }
 977                 if (!err)
 978                         err = filemap_fdatawrite_range(mapping, start, end);
 979                 if (err)
 980                         werr = err;
 981                 else if (wait_writeback)
 982                         werr = filemap_fdatawait_range(mapping, start, end);
 983                 free_extent_state(cached_state);
 984                 cached_state = NULL;
 985                 cond_resched();
 986                 start = end + 1;
 987         }
 988         atomic_dec(&BTRFS_I(fs_info->btree_inode)->sync_writers);
 989         return werr;
 990 }
 991
 992 /*
 993  * when btree blocks are allocated, they have some corresponding bits set for
 994  * them in one of two extent_io trees.  This is used to make sure all of
 995  * those extents are on disk for transaction or log commit.  We wait
 996  * on all the pages and clear them from the dirty pages state tree
 997  */
 998 static int __btrfs_wait_marked_extents(struct btrfs_fs_info *fs_info,
 999                                        struct extent_io_tree *dirty_pages)
1000 {
1001         int err = 0;
1002         int werr = 0;
1003         struct address_space *mapping = fs_info->btree_inode->i_mapping;
1004         struct extent_state *cached_state = NULL;
1005         u64 start = 0;
1006         u64 end;
1007
1008         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
1009                                       EXTENT_NEED_WAIT, &cached_state)) {
1010                 /*
1011                  * Ignore -ENOMEM errors returned by clear_extent_bit().
1012                  * When committing the transaction, we'll remove any entries
1013                  * left in the io tree. For a log commit, we don't remove them
1014                  * after committing the log because the tree can be accessed
1015                  * concurrently - we do it only at transaction commit time when
1016                  * it's safe to do it (through clear_btree_io_tree()).
1017                  */
1018                 err = clear_extent_bit(dirty_pages, start, end,
1019                                        EXTENT_NEED_WAIT, 0, 0, &cached_state);
1020                 if (err == -ENOMEM)
1021                         err = 0;
1022                 if (!err)
1023                         err = filemap_fdatawait_range(mapping, start, end);
1024                 if (err)
1025                         werr = err;
1026                 free_extent_state(cached_state);
1027                 cached_state = NULL;
1028                 cond_resched();
1029                 start = end + 1;
1030         }
1031         if (err)
1032                 werr = err;
1033         return werr;
1034 }
1035
1036 int btrfs_wait_extents(struct btrfs_fs_info *fs_info,
1037                        struct extent_io_tree *dirty_pages)
1038 {
1039         bool errors = false;
1040         int err;
1041
1042         err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1043         if (test_and_clear_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags))
1044                 errors = true;
1045
1046         if (errors && !err)
1047                 err = -EIO;
1048         return err;
1049 }
1050
1051 int btrfs_wait_tree_log_extents(struct btrfs_root *log_root, int mark)
1052 {
1053         struct btrfs_fs_info *fs_info = log_root->fs_info;
1054         struct extent_io_tree *dirty_pages = &log_root->dirty_log_pages;
1055         bool errors = false;
1056         int err;
1057
1058         ASSERT(log_root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
1059
1060         err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1061         if ((mark & EXTENT_DIRTY) &&
1062             test_and_clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags))
1063                 errors = true;
1064
1065         if ((mark & EXTENT_NEW) &&
1066             test_and_clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags))
1067                 errors = true;
1068
1069         if (errors && !err)
1070                 err = -EIO;
1071         return err;
1072 }
1073
1074 /*
1075  * when btree blocks are allocated, they have some corresponding bits set for
1076  * them in one of two extent_io trees.  This is used to make sure all of
1077  * those extents are on disk for transaction or log commit
1078  */
1079 static int btrfs_write_and_wait_marked_extents(struct btrfs_fs_info *fs_info,
1080                                 struct extent_io_tree *dirty_pages, int mark)
1081 {
1082         int ret;
1083         int ret2;
1084         struct blk_plug plug;
1085
1086         blk_start_plug(&plug);
1087         ret = btrfs_write_marked_extents(fs_info, dirty_pages, mark);
1088         blk_finish_plug(&plug);
1089         ret2 = btrfs_wait_extents(fs_info, dirty_pages);
1090
1091         if (ret)
1092                 return ret;
1093         if (ret2)
1094                 return ret2;
1095         return 0;
1096 }
1097
1098 static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
1099                                             struct btrfs_fs_info *fs_info)
1100 {
1101         int ret;
1102
1103         ret = btrfs_write_and_wait_marked_extents(fs_info,
1104                                            &trans->transaction->dirty_pages,
1105                                            EXTENT_DIRTY);
1106         clear_btree_io_tree(&trans->transaction->dirty_pages);
1107
1108         return ret;
1109 }
1110
1111 /*
1112  * this is used to update the root pointer in the tree of tree roots.
1113  *
1114  * But, in the case of the extent allocation tree, updating the root
1115  * pointer may allocate blocks which may change the root of the extent
1116  * allocation tree.
1117  *
1118  * So, this loops and repeats and makes sure the cowonly root didn't
1119  * change while the root pointer was being updated in the metadata.
1120  */
1121 static int update_cowonly_root(struct btrfs_trans_handle *trans,
1122                                struct btrfs_root *root)
1123 {
1124         int ret;
1125         u64 old_root_bytenr;
1126         u64 old_root_used;
1127         struct btrfs_fs_info *fs_info = root->fs_info;
1128         struct btrfs_root *tree_root = fs_info->tree_root;
1129
1130         old_root_used = btrfs_root_used(&root->root_item);
1131
1132         while (1) {
1133                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1134                 if (old_root_bytenr == root->node->start &&
1135                     old_root_used == btrfs_root_used(&root->root_item))
1136                         break;
1137
1138                 btrfs_set_root_node(&root->root_item, root->node);
1139                 ret = btrfs_update_root(trans, tree_root,
1140                                         &root->root_key,
1141                                         &root->root_item);
1142                 if (ret)
1143                         return ret;
1144
1145                 old_root_used = btrfs_root_used(&root->root_item);
1146         }
1147
1148         return 0;
1149 }
1150
1151 /*
1152  * update all the cowonly tree roots on disk
1153  *
1154  * The error handling in this function may not be obvious. Any of the
1155  * failures will cause the file system to go offline. We still need
1156  * to clean up the delayed refs.
1157  */
1158 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
1159                                          struct btrfs_fs_info *fs_info)
1160 {
1161         struct list_head *dirty_bgs = &trans->transaction->dirty_bgs;
1162         struct list_head *io_bgs = &trans->transaction->io_bgs;
1163         struct list_head *next;
1164         struct extent_buffer *eb;
1165         int ret;
1166
1167         eb = btrfs_lock_root_node(fs_info->tree_root);
1168         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
1169                               0, &eb);
1170         btrfs_tree_unlock(eb);
1171         free_extent_buffer(eb);
1172
1173         if (ret)
1174                 return ret;
1175
1176         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1177         if (ret)
1178                 return ret;
1179
1180         ret = btrfs_run_dev_stats(trans, fs_info);
1181         if (ret)
1182                 return ret;
1183         ret = btrfs_run_dev_replace(trans, fs_info);
1184         if (ret)
1185                 return ret;
1186         ret = btrfs_run_qgroups(trans, fs_info);
1187         if (ret)
1188                 return ret;
1189
1190         ret = btrfs_setup_space_cache(trans, fs_info);
1191         if (ret)
1192                 return ret;
1193
1194         /* run_qgroups might have added some more refs */
1195         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1196         if (ret)
1197                 return ret;
1198 again:
1199         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1200                 struct btrfs_root *root;
1201                 next = fs_info->dirty_cowonly_roots.next;
1202                 list_del_init(next);
1203                 root = list_entry(next, struct btrfs_root, dirty_list);
1204                 clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1205
1206                 if (root != fs_info->extent_root)
1207                         list_add_tail(&root->dirty_list,
1208                                       &trans->transaction->switch_commits);
1209                 ret = update_cowonly_root(trans, root);
1210                 if (ret)
1211                         return ret;
1212                 ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1213                 if (ret)
1214                         return ret;
1215         }
1216
1217         while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1218                 ret = btrfs_write_dirty_block_groups(trans, fs_info);
1219                 if (ret)
1220                         return ret;
1221                 ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1222                 if (ret)
1223                         return ret;
1224         }
1225
1226         if (!list_empty(&fs_info->dirty_cowonly_roots))
1227                 goto again;
1228
1229         list_add_tail(&fs_info->extent_root->dirty_list,
1230                       &trans->transaction->switch_commits);
1231         btrfs_after_dev_replace_commit(fs_info);
1232
1233         return 0;
1234 }
1235
1236 /*
1237  * dead roots are old snapshots that need to be deleted.  This allocates
1238  * a dirty root struct and adds it into the list of dead roots that need to
1239  * be deleted
1240  */
1241 void btrfs_add_dead_root(struct btrfs_root *root)
1242 {
1243         struct btrfs_fs_info *fs_info = root->fs_info;
1244
1245         spin_lock(&fs_info->trans_lock);
1246         if (list_empty(&root->root_list))
1247                 list_add_tail(&root->root_list, &fs_info->dead_roots);
1248         spin_unlock(&fs_info->trans_lock);
1249 }
1250
1251 /*
1252  * update all the cowonly tree roots on disk
1253  */
1254 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
1255                                     struct btrfs_fs_info *fs_info)
1256 {
1257         struct btrfs_root *gang[8];
1258         int i;
1259         int ret;
1260         int err = 0;
1261
1262         spin_lock(&fs_info->fs_roots_radix_lock);
1263         while (1) {
1264                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1265                                                  (void **)gang, 0,
1266                                                  ARRAY_SIZE(gang),
1267                                                  BTRFS_ROOT_TRANS_TAG);
1268                 if (ret == 0)
1269                         break;
1270                 for (i = 0; i < ret; i++) {
1271                         struct btrfs_root *root = gang[i];
1272                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
1273                                         (unsigned long)root->root_key.objectid,
1274                                         BTRFS_ROOT_TRANS_TAG);
1275                         spin_unlock(&fs_info->fs_roots_radix_lock);
1276
1277                         btrfs_free_log(trans, root);
1278                         btrfs_update_reloc_root(trans, root);
1279                         btrfs_orphan_commit_root(trans, root);
1280
1281                         btrfs_save_ino_cache(root, trans);
1282
1283                         /* see comments in should_cow_block() */
1284                         clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1285                         smp_mb__after_atomic();
1286
1287                         if (root->commit_root != root->node) {
1288                                 list_add_tail(&root->dirty_list,
1289                                         &trans->transaction->switch_commits);
1290                                 btrfs_set_root_node(&root->root_item,
1291                                                     root->node);
1292                         }
1293
1294                         err = btrfs_update_root(trans, fs_info->tree_root,
1295                                                 &root->root_key,
1296                                                 &root->root_item);
1297                         spin_lock(&fs_info->fs_roots_radix_lock);
1298                         if (err)
1299                                 break;
1300                         btrfs_qgroup_free_meta_all(root);
1301                 }
1302         }
1303         spin_unlock(&fs_info->fs_roots_radix_lock);
1304         return err;
1305 }
1306
1307 /*
1308  * defrag a given btree.
1309  * Every leaf in the btree is read and defragged.
1310  */
1311 int btrfs_defrag_root(struct btrfs_root *root)
1312 {
1313         struct btrfs_fs_info *info = root->fs_info;
1314         struct btrfs_trans_handle *trans;
1315         int ret;
1316
1317         if (test_and_set_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state))
1318                 return 0;
1319
1320         while (1) {
1321                 trans = btrfs_start_transaction(root, 0);
1322                 if (IS_ERR(trans))
1323                         return PTR_ERR(trans);
1324
1325                 ret = btrfs_defrag_leaves(trans, root);
1326
1327                 btrfs_end_transaction(trans);
1328                 btrfs_btree_balance_dirty(info);
1329                 cond_resched();
1330
1331                 if (btrfs_fs_closing(info) || ret != -EAGAIN)
1332                         break;
1333
1334                 if (btrfs_defrag_cancelled(info)) {
1335                         btrfs_debug(info, "defrag_root cancelled");
1336                         ret = -EAGAIN;
1337                         break;
1338                 }
1339         }
1340         clear_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state);
1341         return ret;
1342 }
1343
1344 /*
1345  * Do all special snapshot related qgroup dirty hack.
1346  *
1347  * Will do all needed qgroup inherit and dirty hack like switch commit
1348  * roots inside one transaction and write all btree into disk, to make
1349  * qgroup works.
1350  */
1351 static int qgroup_account_snapshot(struct btrfs_trans_handle *trans,
1352                                    struct btrfs_root *src,
1353                                    struct btrfs_root *parent,
1354                                    struct btrfs_qgroup_inherit *inherit,
1355                                    u64 dst_objectid)
1356 {
1357         struct btrfs_fs_info *fs_info = src->fs_info;
1358         int ret;
1359
1360         /*
1361          * Save some performance in the case that qgroups are not
1362          * enabled. If this check races with the ioctl, rescan will
1363          * kick in anyway.
1364          */
1365         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
1366                 return 0;
1367
1368         /*
1369          * We are going to commit transaction, see btrfs_commit_transaction()
1370          * comment for reason locking tree_log_mutex
1371          */
1372         mutex_lock(&fs_info->tree_log_mutex);
1373
1374         ret = commit_fs_roots(trans, fs_info);
1375         if (ret)
1376                 goto out;
1377         ret = btrfs_qgroup_account_extents(trans, fs_info);
1378         if (ret < 0)
1379                 goto out;
1380
1381         /* Now qgroup are all updated, we can inherit it to new qgroups */
1382         ret = btrfs_qgroup_inherit(trans, fs_info,
1383                                    src->root_key.objectid, dst_objectid,
1384                                    inherit);
1385         if (ret < 0)
1386                 goto out;
1387
1388         /*
1389          * Now we do a simplified commit transaction, which will:
1390          * 1) commit all subvolume and extent tree
1391          *    To ensure all subvolume and extent tree have a valid
1392          *    commit_root to accounting later insert_dir_item()
1393          * 2) write all btree blocks onto disk
1394          *    This is to make sure later btree modification will be cowed
1395          *    Or commit_root can be populated and cause wrong qgroup numbers
1396          * In this simplified commit, we don't really care about other trees
1397          * like chunk and root tree, as they won't affect qgroup.
1398          * And we don't write super to avoid half committed status.
1399          */
1400         ret = commit_cowonly_roots(trans, fs_info);
1401         if (ret)
1402                 goto out;
1403         switch_commit_roots(trans->transaction, fs_info);
1404         ret = btrfs_write_and_wait_transaction(trans, fs_info);
1405         if (ret)
1406                 btrfs_handle_fs_error(fs_info, ret,
1407                         "Error while writing out transaction for qgroup");
1408
1409 out:
1410         mutex_unlock(&fs_info->tree_log_mutex);
1411
1412         /*
1413          * Force parent root to be updated, as we recorded it before so its
1414          * last_trans == cur_transid.
1415          * Or it won't be committed again onto disk after later
1416          * insert_dir_item()
1417          */
1418         if (!ret)
1419                 record_root_in_trans(trans, parent, 1);
1420         return ret;
1421 }
1422
1423 /*
1424  * new snapshots need to be created at a very specific time in the
1425  * transaction commit.  This does the actual creation.
1426  *
1427  * Note:
1428  * If the error which may affect the commitment of the current transaction
1429  * happens, we should return the error number. If the error which just affect
1430  * the creation of the pending snapshots, just return 0.
1431  */
1432 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1433                                    struct btrfs_fs_info *fs_info,
1434                                    struct btrfs_pending_snapshot *pending)
1435 {
1436         struct btrfs_key key;
1437         struct btrfs_root_item *new_root_item;
1438         struct btrfs_root *tree_root = fs_info->tree_root;
1439         struct btrfs_root *root = pending->root;
1440         struct btrfs_root *parent_root;
1441         struct btrfs_block_rsv *rsv;
1442         struct inode *parent_inode;
1443         struct btrfs_path *path;
1444         struct btrfs_dir_item *dir_item;
1445         struct dentry *dentry;
1446         struct extent_buffer *tmp;
1447         struct extent_buffer *old;
1448         struct timespec cur_time;
1449         int ret = 0;
1450         u64 to_reserve = 0;
1451         u64 index = 0;
1452         u64 objectid;
1453         u64 root_flags;
1454         uuid_le new_uuid;
1455
1456         ASSERT(pending->path);
1457         path = pending->path;
1458
1459         ASSERT(pending->root_item);
1460         new_root_item = pending->root_item;
1461
1462         pending->error = btrfs_find_free_objectid(tree_root, &objectid);
1463         if (pending->error)
1464                 goto no_free_objectid;
1465
1466         /*
1467          * Make qgroup to skip current new snapshot's qgroupid, as it is
1468          * accounted by later btrfs_qgroup_inherit().
1469          */
1470         btrfs_set_skip_qgroup(trans, objectid);
1471
1472         btrfs_reloc_pre_snapshot(pending, &to_reserve);
1473
1474         if (to_reserve > 0) {
1475                 pending->error = btrfs_block_rsv_add(root,
1476                                                      &pending->block_rsv,
1477                                                      to_reserve,
1478                                                      BTRFS_RESERVE_NO_FLUSH);
1479                 if (pending->error)
1480                         goto clear_skip_qgroup;
1481         }
1482
1483         key.objectid = objectid;
1484         key.offset = (u64)-1;
1485         key.type = BTRFS_ROOT_ITEM_KEY;
1486
1487         rsv = trans->block_rsv;
1488         trans->block_rsv = &pending->block_rsv;
1489         trans->bytes_reserved = trans->block_rsv->reserved;
1490         trace_btrfs_space_reservation(fs_info, "transaction",
1491                                       trans->transid,
1492                                       trans->bytes_reserved, 1);
1493         dentry = pending->dentry;
1494         parent_inode = pending->dir;
1495         parent_root = BTRFS_I(parent_inode)->root;
1496         record_root_in_trans(trans, parent_root, 0);
1497
1498         cur_time = current_time(parent_inode);
1499
1500         /*
1501          * insert the directory item
1502          */
1503         ret = btrfs_set_inode_index(BTRFS_I(parent_inode), &index);
1504         BUG_ON(ret); /* -ENOMEM */
1505
1506         /* check if there is a file/dir which has the same name. */
1507         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1508                                          btrfs_ino(BTRFS_I(parent_inode)),
1509                                          dentry->d_name.name,
1510                                          dentry->d_name.len, 0);
1511         if (dir_item != NULL && !IS_ERR(dir_item)) {
1512                 pending->error = -EEXIST;
1513                 goto dir_item_existed;
1514         } else if (IS_ERR(dir_item)) {
1515                 ret = PTR_ERR(dir_item);
1516                 btrfs_abort_transaction(trans, ret);
1517                 goto fail;
1518         }
1519         btrfs_release_path(path);
1520
1521         /*
1522          * pull in the delayed directory update
1523          * and the delayed inode item
1524          * otherwise we corrupt the FS during
1525          * snapshot
1526          */
1527         ret = btrfs_run_delayed_items(trans, fs_info);
1528         if (ret) {      /* Transaction aborted */
1529                 btrfs_abort_transaction(trans, ret);
1530                 goto fail;
1531         }
1532
1533         record_root_in_trans(trans, root, 0);
1534         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1535         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1536         btrfs_check_and_init_root_item(new_root_item);
1537
1538         root_flags = btrfs_root_flags(new_root_item);
1539         if (pending->readonly)
1540                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1541         else
1542                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1543         btrfs_set_root_flags(new_root_item, root_flags);
1544
1545         btrfs_set_root_generation_v2(new_root_item,
1546                         trans->transid);
1547         uuid_le_gen(&new_uuid);
1548         memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1549         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1550                         BTRFS_UUID_SIZE);
1551         if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1552                 memset(new_root_item->received_uuid, 0,
1553                        sizeof(new_root_item->received_uuid));
1554                 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1555                 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1556                 btrfs_set_root_stransid(new_root_item, 0);
1557                 btrfs_set_root_rtransid(new_root_item, 0);
1558         }
1559         btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1560         btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1561         btrfs_set_root_otransid(new_root_item, trans->transid);
1562
1563         old = btrfs_lock_root_node(root);
1564         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1565         if (ret) {
1566                 btrfs_tree_unlock(old);
1567                 free_extent_buffer(old);
1568                 btrfs_abort_transaction(trans, ret);
1569                 goto fail;
1570         }
1571
1572         btrfs_set_lock_blocking(old);
1573
1574         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1575         /* clean up in any case */
1576         btrfs_tree_unlock(old);
1577         free_extent_buffer(old);
1578         if (ret) {
1579                 btrfs_abort_transaction(trans, ret);
1580                 goto fail;
1581         }
1582         /* see comments in should_cow_block() */
1583         set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1584         smp_wmb();
1585
1586         btrfs_set_root_node(new_root_item, tmp);
1587         /* record when the snapshot was created in key.offset */
1588         key.offset = trans->transid;
1589         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1590         btrfs_tree_unlock(tmp);
1591         free_extent_buffer(tmp);
1592         if (ret) {
1593                 btrfs_abort_transaction(trans, ret);
1594                 goto fail;
1595         }
1596
1597         /*
1598          * insert root back/forward references
1599          */
1600         ret = btrfs_add_root_ref(trans, fs_info, objectid,
1601                                  parent_root->root_key.objectid,
1602                                  btrfs_ino(BTRFS_I(parent_inode)), index,
1603                                  dentry->d_name.name, dentry->d_name.len);
1604         if (ret) {
1605                 btrfs_abort_transaction(trans, ret);
1606                 goto fail;
1607         }
1608
1609         key.offset = (u64)-1;
1610         pending->snap = btrfs_read_fs_root_no_name(fs_info, &key);
1611         if (IS_ERR(pending->snap)) {
1612                 ret = PTR_ERR(pending->snap);
1613                 btrfs_abort_transaction(trans, ret);
1614                 goto fail;
1615         }
1616
1617         ret = btrfs_reloc_post_snapshot(trans, pending);
1618         if (ret) {
1619                 btrfs_abort_transaction(trans, ret);
1620                 goto fail;
1621         }
1622
1623         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1624         if (ret) {
1625                 btrfs_abort_transaction(trans, ret);
1626                 goto fail;
1627         }
1628
1629         /*
1630          * Do special qgroup accounting for snapshot, as we do some qgroup
1631          * snapshot hack to do fast snapshot.
1632          * To co-operate with that hack, we do hack again.
1633          * Or snapshot will be greatly slowed down by a subtree qgroup rescan
1634          */
1635         ret = qgroup_account_snapshot(trans, root, parent_root,
1636                                       pending->inherit, objectid);
1637         if (ret < 0)
1638                 goto fail;
1639
1640         ret = btrfs_insert_dir_item(trans, parent_root,
1641                                     dentry->d_name.name, dentry->d_name.len,
1642                                     BTRFS_I(parent_inode), &key,
1643                                     BTRFS_FT_DIR, index);
1644         /* We have check then name at the beginning, so it is impossible. */
1645         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1646         if (ret) {
1647                 btrfs_abort_transaction(trans, ret);
1648                 goto fail;
1649         }
1650
1651         btrfs_i_size_write(BTRFS_I(parent_inode), parent_inode->i_size +
1652                                          dentry->d_name.len * 2);
1653         parent_inode->i_mtime = parent_inode->i_ctime =
1654                 current_time(parent_inode);
1655         ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1656         if (ret) {
1657                 btrfs_abort_transaction(trans, ret);
1658                 goto fail;
1659         }
1660         ret = btrfs_uuid_tree_add(trans, fs_info, new_uuid.b,
1661                                   BTRFS_UUID_KEY_SUBVOL, objectid);
1662         if (ret) {
1663                 btrfs_abort_transaction(trans, ret);
1664                 goto fail;
1665         }
1666         if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1667                 ret = btrfs_uuid_tree_add(trans, fs_info,
1668                                           new_root_item->received_uuid,
1669                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1670                                           objectid);
1671                 if (ret && ret != -EEXIST) {
1672                         btrfs_abort_transaction(trans, ret);
1673                         goto fail;
1674                 }
1675         }
1676
1677         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1678         if (ret) {
1679                 btrfs_abort_transaction(trans, ret);
1680                 goto fail;
1681         }
1682
1683 fail:
1684         pending->error = ret;
1685 dir_item_existed:
1686         trans->block_rsv = rsv;
1687         trans->bytes_reserved = 0;
1688 clear_skip_qgroup:
1689         btrfs_clear_skip_qgroup(trans);
1690 no_free_objectid:
1691         kfree(new_root_item);
1692         pending->root_item = NULL;
1693         btrfs_free_path(path);
1694         pending->path = NULL;
1695
1696         return ret;
1697 }
1698
1699 /*
1700  * create all the snapshots we've scheduled for creation
1701  */
1702 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1703                                              struct btrfs_fs_info *fs_info)
1704 {
1705         struct btrfs_pending_snapshot *pending, *next;
1706         struct list_head *head = &trans->transaction->pending_snapshots;
1707         int ret = 0;
1708
1709         list_for_each_entry_safe(pending, next, head, list) {
1710                 list_del(&pending->list);
1711                 ret = create_pending_snapshot(trans, fs_info, pending);
1712                 if (ret)
1713                         break;
1714         }
1715         return ret;
1716 }
1717
1718 static void update_super_roots(struct btrfs_fs_info *fs_info)
1719 {
1720         struct btrfs_root_item *root_item;
1721         struct btrfs_super_block *super;
1722
1723         super = fs_info->super_copy;
1724
1725         root_item = &fs_info->chunk_root->root_item;
1726         super->chunk_root = root_item->bytenr;
1727         super->chunk_root_generation = root_item->generation;
1728         super->chunk_root_level = root_item->level;
1729
1730         root_item = &fs_info->tree_root->root_item;
1731         super->root = root_item->bytenr;
1732         super->generation = root_item->generation;
1733         super->root_level = root_item->level;
1734         if (btrfs_test_opt(fs_info, SPACE_CACHE))
1735                 super->cache_generation = root_item->generation;
1736         if (test_bit(BTRFS_FS_UPDATE_UUID_TREE_GEN, &fs_info->flags))
1737                 super->uuid_tree_generation = root_item->generation;
1738 }
1739
1740 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1741 {
1742         struct btrfs_transaction *trans;
1743         int ret = 0;
1744
1745         spin_lock(&info->trans_lock);
1746         trans = info->running_transaction;
1747         if (trans)
1748                 ret = (trans->state >= TRANS_STATE_COMMIT_START);
1749         spin_unlock(&info->trans_lock);
1750         return ret;
1751 }
1752
1753 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1754 {
1755         struct btrfs_transaction *trans;
1756         int ret = 0;
1757
1758         spin_lock(&info->trans_lock);
1759         trans = info->running_transaction;
1760         if (trans)
1761                 ret = is_transaction_blocked(trans);
1762         spin_unlock(&info->trans_lock);
1763         return ret;
1764 }
1765
1766 /*
1767  * wait for the current transaction commit to start and block subsequent
1768  * transaction joins
1769  */
1770 static void wait_current_trans_commit_start(struct btrfs_fs_info *fs_info,
1771                                             struct btrfs_transaction *trans)
1772 {
1773         wait_event(fs_info->transaction_blocked_wait,
1774                    trans->state >= TRANS_STATE_COMMIT_START || trans->aborted);
1775 }
1776
1777 /*
1778  * wait for the current transaction to start and then become unblocked.
1779  * caller holds ref.
1780  */
1781 static void wait_current_trans_commit_start_and_unblock(
1782                                         struct btrfs_fs_info *fs_info,
1783                                         struct btrfs_transaction *trans)
1784 {
1785         wait_event(fs_info->transaction_wait,
1786                    trans->state >= TRANS_STATE_UNBLOCKED || trans->aborted);
1787 }
1788
1789 /*
1790  * commit transactions asynchronously. once btrfs_commit_transaction_async
1791  * returns, any subsequent transaction will not be allowed to join.
1792  */
1793 struct btrfs_async_commit {
1794         struct btrfs_trans_handle *newtrans;
1795         struct work_struct work;
1796 };
1797
1798 static void do_async_commit(struct work_struct *work)
1799 {
1800         struct btrfs_async_commit *ac =
1801                 container_of(work, struct btrfs_async_commit, work);
1802
1803         /*
1804          * We've got freeze protection passed with the transaction.
1805          * Tell lockdep about it.
1806          */
1807         if (ac->newtrans->type & __TRANS_FREEZABLE)
1808                 __sb_writers_acquired(ac->newtrans->fs_info->sb, SB_FREEZE_FS);
1809
1810         current->journal_info = ac->newtrans;
1811
1812         btrfs_commit_transaction(ac->newtrans);
1813         kfree(ac);
1814 }
1815
1816 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1817                                    int wait_for_unblock)
1818 {
1819         struct btrfs_fs_info *fs_info = trans->fs_info;
1820         struct btrfs_async_commit *ac;
1821         struct btrfs_transaction *cur_trans;
1822
1823         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1824         if (!ac)
1825                 return -ENOMEM;
1826
1827         INIT_WORK(&ac->work, do_async_commit);
1828         ac->newtrans = btrfs_join_transaction(trans->root);
1829         if (IS_ERR(ac->newtrans)) {
1830                 int err = PTR_ERR(ac->newtrans);
1831                 kfree(ac);
1832                 return err;
1833         }
1834
1835         /* take transaction reference */
1836         cur_trans = trans->transaction;
1837         refcount_inc(&cur_trans->use_count);
1838
1839         btrfs_end_transaction(trans);
1840
1841         /*
1842          * Tell lockdep we've released the freeze rwsem, since the
1843          * async commit thread will be the one to unlock it.
1844          */
1845         if (ac->newtrans->type & __TRANS_FREEZABLE)
1846                 __sb_writers_release(fs_info->sb, SB_FREEZE_FS);
1847
1848         schedule_work(&ac->work);
1849
1850         /* wait for transaction to start and unblock */
1851         if (wait_for_unblock)
1852                 wait_current_trans_commit_start_and_unblock(fs_info, cur_trans);
1853         else
1854                 wait_current_trans_commit_start(fs_info, cur_trans);
1855
1856         if (current->journal_info == trans)
1857                 current->journal_info = NULL;
1858
1859         btrfs_put_transaction(cur_trans);
1860         return 0;
1861 }
1862
1863
1864 static void cleanup_transaction(struct btrfs_trans_handle *trans,
1865                                 struct btrfs_root *root, int err)
1866 {
1867         struct btrfs_fs_info *fs_info = root->fs_info;
1868         struct btrfs_transaction *cur_trans = trans->transaction;
1869         DEFINE_WAIT(wait);
1870
1871         WARN_ON(refcount_read(&trans->use_count) > 1);
1872
1873         btrfs_abort_transaction(trans, err);
1874
1875         spin_lock(&fs_info->trans_lock);
1876
1877         /*
1878          * If the transaction is removed from the list, it means this
1879          * transaction has been committed successfully, so it is impossible
1880          * to call the cleanup function.
1881          */
1882         BUG_ON(list_empty(&cur_trans->list));
1883
1884         list_del_init(&cur_trans->list);
1885         if (cur_trans == fs_info->running_transaction) {
1886                 cur_trans->state = TRANS_STATE_COMMIT_DOING;
1887                 spin_unlock(&fs_info->trans_lock);
1888                 wait_event(cur_trans->writer_wait,
1889                            atomic_read(&cur_trans->num_writers) == 1);
1890
1891                 spin_lock(&fs_info->trans_lock);
1892         }
1893         spin_unlock(&fs_info->trans_lock);
1894
1895         btrfs_cleanup_one_transaction(trans->transaction, fs_info);
1896
1897         spin_lock(&fs_info->trans_lock);
1898         if (cur_trans == fs_info->running_transaction)
1899                 fs_info->running_transaction = NULL;
1900         spin_unlock(&fs_info->trans_lock);
1901
1902         if (trans->type & __TRANS_FREEZABLE)
1903                 sb_end_intwrite(fs_info->sb);
1904         btrfs_put_transaction(cur_trans);
1905         btrfs_put_transaction(cur_trans);
1906
1907         trace_btrfs_transaction_commit(root);
1908
1909         if (current->journal_info == trans)
1910                 current->journal_info = NULL;
1911         btrfs_scrub_cancel(fs_info);
1912
1913         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1914 }
1915
1916 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
1917 {
1918         /*
1919          * We use writeback_inodes_sb here because if we used
1920          * btrfs_start_delalloc_roots we would deadlock with fs freeze.
1921          * Currently are holding the fs freeze lock, if we do an async flush
1922          * we'll do btrfs_join_transaction() and deadlock because we need to
1923          * wait for the fs freeze lock.  Using the direct flushing we benefit
1924          * from already being in a transaction and our join_transaction doesn't
1925          * have to re-take the fs freeze lock.
1926          */
1927         if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
1928                 writeback_inodes_sb(fs_info->sb, WB_REASON_SYNC);
1929         return 0;
1930 }
1931
1932 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
1933 {
1934         if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
1935                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0, (u64)-1);
1936 }
1937
1938 static inline void
1939 btrfs_wait_pending_ordered(struct btrfs_transaction *cur_trans)
1940 {
1941         wait_event(cur_trans->pending_wait,
1942                    atomic_read(&cur_trans->pending_ordered) == 0);
1943 }
1944
1945 int btrfs_commit_transaction(struct btrfs_trans_handle *trans)
1946 {
1947         struct btrfs_fs_info *fs_info = trans->fs_info;
1948         struct btrfs_transaction *cur_trans = trans->transaction;
1949         struct btrfs_transaction *prev_trans = NULL;
1950         int ret;
1951
1952         /* Stop the commit early if ->aborted is set */
1953         if (unlikely(READ_ONCE(cur_trans->aborted))) {
1954                 ret = cur_trans->aborted;
1955                 btrfs_end_transaction(trans);
1956                 return ret;
1957         }
1958
1959         /* make a pass through all the delayed refs we have so far
1960          * any runnings procs may add more while we are here
1961          */
1962         ret = btrfs_run_delayed_refs(trans, fs_info, 0);
1963         if (ret) {
1964                 btrfs_end_transaction(trans);
1965                 return ret;
1966         }
1967
1968         btrfs_trans_release_metadata(trans, fs_info);
1969         trans->block_rsv = NULL;
1970
1971         cur_trans = trans->transaction;
1972
1973         /*
1974          * set the flushing flag so procs in this transaction have to
1975          * start sending their work down.
1976          */
1977         cur_trans->delayed_refs.flushing = 1;
1978         smp_wmb();
1979
1980         if (!list_empty(&trans->new_bgs))
1981                 btrfs_create_pending_block_groups(trans, fs_info);
1982
1983         ret = btrfs_run_delayed_refs(trans, fs_info, 0);
1984         if (ret) {
1985                 btrfs_end_transaction(trans);
1986                 return ret;
1987         }
1988
1989         if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) {
1990                 int run_it = 0;
1991
1992                 /* this mutex is also taken before trying to set
1993                  * block groups readonly.  We need to make sure
1994                  * that nobody has set a block group readonly
1995                  * after a extents from that block group have been
1996                  * allocated for cache files.  btrfs_set_block_group_ro
1997                  * will wait for the transaction to commit if it
1998                  * finds BTRFS_TRANS_DIRTY_BG_RUN set.
1999                  *
2000                  * The BTRFS_TRANS_DIRTY_BG_RUN flag is also used to make sure
2001                  * only one process starts all the block group IO.  It wouldn't
2002                  * hurt to have more than one go through, but there's no
2003                  * real advantage to it either.
2004                  */
2005                 mutex_lock(&fs_info->ro_block_group_mutex);
2006                 if (!test_and_set_bit(BTRFS_TRANS_DIRTY_BG_RUN,
2007                                       &cur_trans->flags))
2008                         run_it = 1;
2009                 mutex_unlock(&fs_info->ro_block_group_mutex);
2010
2011                 if (run_it)
2012                         ret = btrfs_start_dirty_block_groups(trans, fs_info);
2013         }
2014         if (ret) {
2015                 btrfs_end_transaction(trans);
2016                 return ret;
2017         }
2018
2019         spin_lock(&fs_info->trans_lock);
2020         if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
2021                 spin_unlock(&fs_info->trans_lock);
2022                 refcount_inc(&cur_trans->use_count);
2023                 ret = btrfs_end_transaction(trans);
2024
2025                 wait_for_commit(cur_trans);
2026
2027                 if (unlikely(cur_trans->aborted))
2028                         ret = cur_trans->aborted;
2029
2030                 btrfs_put_transaction(cur_trans);
2031
2032                 return ret;
2033         }
2034
2035         cur_trans->state = TRANS_STATE_COMMIT_START;
2036         wake_up(&fs_info->transaction_blocked_wait);
2037
2038         if (cur_trans->list.prev != &fs_info->trans_list) {
2039                 prev_trans = list_entry(cur_trans->list.prev,
2040                                         struct btrfs_transaction, list);
2041                 if (prev_trans->state != TRANS_STATE_COMPLETED) {
2042                         refcount_inc(&prev_trans->use_count);
2043                         spin_unlock(&fs_info->trans_lock);
2044
2045                         wait_for_commit(prev_trans);
2046                         ret = prev_trans->aborted;
2047
2048                         btrfs_put_transaction(prev_trans);
2049                         if (ret)
2050                                 goto cleanup_transaction;
2051                 } else {
2052                         spin_unlock(&fs_info->trans_lock);
2053                 }
2054         } else {
2055                 spin_unlock(&fs_info->trans_lock);
2056         }
2057
2058         extwriter_counter_dec(cur_trans, trans->type);
2059
2060         ret = btrfs_start_delalloc_flush(fs_info);
2061         if (ret)
2062                 goto cleanup_transaction;
2063
2064         ret = btrfs_run_delayed_items(trans, fs_info);
2065         if (ret)
2066                 goto cleanup_transaction;
2067
2068         wait_event(cur_trans->writer_wait,
2069                    extwriter_counter_read(cur_trans) == 0);
2070
2071         /* some pending stuffs might be added after the previous flush. */
2072         ret = btrfs_run_delayed_items(trans, fs_info);
2073         if (ret)
2074                 goto cleanup_transaction;
2075
2076         btrfs_wait_delalloc_flush(fs_info);
2077
2078         btrfs_wait_pending_ordered(cur_trans);
2079
2080         btrfs_scrub_pause(fs_info);
2081         /*
2082          * Ok now we need to make sure to block out any other joins while we
2083          * commit the transaction.  We could have started a join before setting
2084          * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
2085          */
2086         spin_lock(&fs_info->trans_lock);
2087         cur_trans->state = TRANS_STATE_COMMIT_DOING;
2088         spin_unlock(&fs_info->trans_lock);
2089         wait_event(cur_trans->writer_wait,
2090                    atomic_read(&cur_trans->num_writers) == 1);
2091
2092         /* ->aborted might be set after the previous check, so check it */
2093         if (unlikely(READ_ONCE(cur_trans->aborted))) {
2094                 ret = cur_trans->aborted;
2095                 goto scrub_continue;
2096         }
2097         /*
2098          * the reloc mutex makes sure that we stop
2099          * the balancing code from coming in and moving
2100          * extents around in the middle of the commit
2101          */
2102         mutex_lock(&fs_info->reloc_mutex);
2103
2104         /*
2105          * We needn't worry about the delayed items because we will
2106          * deal with them in create_pending_snapshot(), which is the
2107          * core function of the snapshot creation.
2108          */
2109         ret = create_pending_snapshots(trans, fs_info);
2110         if (ret) {
2111                 mutex_unlock(&fs_info->reloc_mutex);
2112                 goto scrub_continue;
2113         }
2114
2115         /*
2116          * We insert the dir indexes of the snapshots and update the inode
2117          * of the snapshots' parents after the snapshot creation, so there
2118          * are some delayed items which are not dealt with. Now deal with
2119          * them.
2120          *
2121          * We needn't worry that this operation will corrupt the snapshots,
2122          * because all the tree which are snapshoted will be forced to COW
2123          * the nodes and leaves.
2124          */
2125         ret = btrfs_run_delayed_items(trans, fs_info);
2126         if (ret) {
2127                 mutex_unlock(&fs_info->reloc_mutex);
2128                 goto scrub_continue;
2129         }
2130
2131         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
2132         if (ret) {
2133                 mutex_unlock(&fs_info->reloc_mutex);
2134                 goto scrub_continue;
2135         }
2136
2137         /*
2138          * make sure none of the code above managed to slip in a
2139          * delayed item
2140          */
2141         btrfs_assert_delayed_root_empty(fs_info);
2142
2143         WARN_ON(cur_trans != trans->transaction);
2144
2145         /* btrfs_commit_tree_roots is responsible for getting the
2146          * various roots consistent with each other.  Every pointer
2147          * in the tree of tree roots has to point to the most up to date
2148          * root for every subvolume and other tree.  So, we have to keep
2149          * the tree logging code from jumping in and changing any
2150          * of the trees.
2151          *
2152          * At this point in the commit, there can't be any tree-log
2153          * writers, but a little lower down we drop the trans mutex
2154          * and let new people in.  By holding the tree_log_mutex
2155          * from now until after the super is written, we avoid races
2156          * with the tree-log code.
2157          */
2158         mutex_lock(&fs_info->tree_log_mutex);
2159
2160         ret = commit_fs_roots(trans, fs_info);
2161         if (ret) {
2162                 mutex_unlock(&fs_info->tree_log_mutex);
2163                 mutex_unlock(&fs_info->reloc_mutex);
2164                 goto scrub_continue;
2165         }
2166
2167         /*
2168          * Since the transaction is done, we can apply the pending changes
2169          * before the next transaction.
2170          */
2171         btrfs_apply_pending_changes(fs_info);
2172
2173         /* commit_fs_roots gets rid of all the tree log roots, it is now
2174          * safe to free the root of tree log roots
2175          */
2176         btrfs_free_log_root_tree(trans, fs_info);
2177
2178         /*
2179          * commit_fs_roots() can call btrfs_save_ino_cache(), which generates
2180          * new delayed refs. Must handle them or qgroup can be wrong.
2181          */
2182         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
2183         if (ret) {
2184                 mutex_unlock(&fs_info->tree_log_mutex);
2185                 mutex_unlock(&fs_info->reloc_mutex);
2186                 goto scrub_continue;
2187         }
2188
2189         /*
2190          * Since fs roots are all committed, we can get a quite accurate
2191          * new_roots. So let's do quota accounting.
2192          */
2193         ret = btrfs_qgroup_account_extents(trans, fs_info);
2194         if (ret < 0) {
2195                 mutex_unlock(&fs_info->tree_log_mutex);
2196                 mutex_unlock(&fs_info->reloc_mutex);
2197                 goto scrub_continue;
2198         }
2199
2200         ret = commit_cowonly_roots(trans, fs_info);
2201         if (ret) {
2202                 mutex_unlock(&fs_info->tree_log_mutex);
2203                 mutex_unlock(&fs_info->reloc_mutex);
2204                 goto scrub_continue;
2205         }
2206
2207         /*
2208          * The tasks which save the space cache and inode cache may also
2209          * update ->aborted, check it.
2210          */
2211         if (unlikely(READ_ONCE(cur_trans->aborted))) {
2212                 ret = cur_trans->aborted;
2213                 mutex_unlock(&fs_info->tree_log_mutex);
2214                 mutex_unlock(&fs_info->reloc_mutex);
2215                 goto scrub_continue;
2216         }
2217
2218         btrfs_prepare_extent_commit(fs_info);
2219
2220         cur_trans = fs_info->running_transaction;
2221
2222         btrfs_set_root_node(&fs_info->tree_root->root_item,
2223                             fs_info->tree_root->node);
2224         list_add_tail(&fs_info->tree_root->dirty_list,
2225                       &cur_trans->switch_commits);
2226
2227         btrfs_set_root_node(&fs_info->chunk_root->root_item,
2228                             fs_info->chunk_root->node);
2229         list_add_tail(&fs_info->chunk_root->dirty_list,
2230                       &cur_trans->switch_commits);
2231
2232         switch_commit_roots(cur_trans, fs_info);
2233
2234         ASSERT(list_empty(&cur_trans->dirty_bgs));
2235         ASSERT(list_empty(&cur_trans->io_bgs));
2236         update_super_roots(fs_info);
2237
2238         btrfs_set_super_log_root(fs_info->super_copy, 0);
2239         btrfs_set_super_log_root_level(fs_info->super_copy, 0);
2240         memcpy(fs_info->super_for_commit, fs_info->super_copy,
2241                sizeof(*fs_info->super_copy));
2242
2243         btrfs_update_commit_device_size(fs_info);
2244         btrfs_update_commit_device_bytes_used(fs_info, cur_trans);
2245
2246         clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
2247         clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
2248
2249         btrfs_trans_release_chunk_metadata(trans);
2250
2251         spin_lock(&fs_info->trans_lock);
2252         cur_trans->state = TRANS_STATE_UNBLOCKED;
2253         fs_info->running_transaction = NULL;
2254         spin_unlock(&fs_info->trans_lock);
2255         mutex_unlock(&fs_info->reloc_mutex);
2256
2257         wake_up(&fs_info->transaction_wait);
2258
2259         ret = btrfs_write_and_wait_transaction(trans, fs_info);
2260         if (ret) {
2261                 btrfs_handle_fs_error(fs_info, ret,
2262                                       "Error while writing out transaction");
2263                 mutex_unlock(&fs_info->tree_log_mutex);
2264                 goto scrub_continue;
2265         }
2266
2267         ret = write_all_supers(fs_info, 0);
2268         /*
2269          * the super is written, we can safely allow the tree-loggers
2270          * to go about their business
2271          */
2272         mutex_unlock(&fs_info->tree_log_mutex);
2273         if (ret)
2274                 goto scrub_continue;
2275
2276         btrfs_finish_extent_commit(trans, fs_info);
2277
2278         if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &cur_trans->flags))
2279                 btrfs_clear_space_info_full(fs_info);
2280
2281         fs_info->last_trans_committed = cur_trans->transid;
2282         /*
2283          * We needn't acquire the lock here because there is no other task
2284          * which can change it.
2285          */
2286         cur_trans->state = TRANS_STATE_COMPLETED;
2287         wake_up(&cur_trans->commit_wait);
2288
2289         spin_lock(&fs_info->trans_lock);
2290         list_del_init(&cur_trans->list);
2291         spin_unlock(&fs_info->trans_lock);
2292
2293         btrfs_put_transaction(cur_trans);
2294         btrfs_put_transaction(cur_trans);
2295
2296         if (trans->type & __TRANS_FREEZABLE)
2297                 sb_end_intwrite(fs_info->sb);
2298
2299         trace_btrfs_transaction_commit(trans->root);
2300
2301         btrfs_scrub_continue(fs_info);
2302
2303         if (current->journal_info == trans)
2304                 current->journal_info = NULL;
2305
2306         kmem_cache_free(btrfs_trans_handle_cachep, trans);
2307
2308         /*
2309          * If fs has been frozen, we can not handle delayed iputs, otherwise
2310          * it'll result in deadlock about SB_FREEZE_FS.
2311          */
2312         if (current != fs_info->transaction_kthread &&
2313             current != fs_info->cleaner_kthread &&
2314             !test_bit(BTRFS_FS_FROZEN, &fs_info->flags))
2315                 btrfs_run_delayed_iputs(fs_info);
2316
2317         return ret;
2318
2319 scrub_continue:
2320         btrfs_scrub_continue(fs_info);
2321 cleanup_transaction:
2322         btrfs_trans_release_metadata(trans, fs_info);
2323         btrfs_trans_release_chunk_metadata(trans);
2324         trans->block_rsv = NULL;
2325         btrfs_warn(fs_info, "Skipping commit of aborted transaction.");
2326         if (current->journal_info == trans)
2327                 current->journal_info = NULL;
2328         cleanup_transaction(trans, trans->root, ret);
2329
2330         return ret;
2331 }
2332
2333 /*
2334  * return < 0 if error
2335  * 0 if there are no more dead_roots at the time of call
2336  * 1 there are more to be processed, call me again
2337  *
2338  * The return value indicates there are certainly more snapshots to delete, but
2339  * if there comes a new one during processing, it may return 0. We don't mind,
2340  * because btrfs_commit_super will poke cleaner thread and it will process it a
2341  * few seconds later.
2342  */
2343 int btrfs_clean_one_deleted_snapshot(struct btrfs_root *root)
2344 {
2345         int ret;
2346         struct btrfs_fs_info *fs_info = root->fs_info;
2347
2348         spin_lock(&fs_info->trans_lock);
2349         if (list_empty(&fs_info->dead_roots)) {
2350                 spin_unlock(&fs_info->trans_lock);
2351                 return 0;
2352         }
2353         root = list_first_entry(&fs_info->dead_roots,
2354                         struct btrfs_root, root_list);
2355         list_del_init(&root->root_list);
2356         spin_unlock(&fs_info->trans_lock);
2357
2358         btrfs_debug(fs_info, "cleaner removing %llu", root->objectid);
2359
2360         btrfs_kill_all_delayed_nodes(root);
2361
2362         if (btrfs_header_backref_rev(root->node) <
2363                         BTRFS_MIXED_BACKREF_REV)
2364                 ret = btrfs_drop_snapshot(root, NULL, 0, 0);
2365         else
2366                 ret = btrfs_drop_snapshot(root, NULL, 1, 0);
2367
2368         return (ret < 0) ? 0 : 1;
2369 }
2370
2371 void btrfs_apply_pending_changes(struct btrfs_fs_info *fs_info)
2372 {
2373         unsigned long prev;
2374         unsigned long bit;
2375
2376         prev = xchg(&fs_info->pending_changes, 0);
2377         if (!prev)
2378                 return;
2379
2380         bit = 1 << BTRFS_PENDING_SET_INODE_MAP_CACHE;
2381         if (prev & bit)
2382                 btrfs_set_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2383         prev &= ~bit;
2384
2385         bit = 1 << BTRFS_PENDING_CLEAR_INODE_MAP_CACHE;
2386         if (prev & bit)
2387                 btrfs_clear_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2388         prev &= ~bit;
2389
2390         bit = 1 << BTRFS_PENDING_COMMIT;
2391         if (prev & bit)
2392                 btrfs_debug(fs_info, "pending commit done");
2393         prev &= ~bit;
2394
2395         if (prev)
2396                 btrfs_warn(fs_info,
2397                         "unknown pending changes left 0x%lx, ignoring", prev);
2398 }