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_join_transaction(struct btrfs_root *root)
 662 {
 663         return start_transaction(root, 0, TRANS_JOIN, BTRFS_RESERVE_NO_FLUSH,
 664                                  true);
 665 }
 666
 667 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 668 {
 669         return start_transaction(root, 0, TRANS_JOIN_NOLOCK,
 670                                  BTRFS_RESERVE_NO_FLUSH, true);
 671 }
 672
 673 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 674 {
 675         return start_transaction(root, 0, TRANS_USERSPACE,
 676                                  BTRFS_RESERVE_NO_FLUSH, true);
 677 }
 678
 679 /*
 680  * btrfs_attach_transaction() - catch the running transaction
 681  *
 682  * It is used when we want to commit the current the transaction, but
 683  * don't want to start a new one.
 684  *
 685  * Note: If this function return -ENOENT, it just means there is no
 686  * running transaction. But it is possible that the inactive transaction
 687  * is still in the memory, not fully on disk. If you hope there is no
 688  * inactive transaction in the fs when -ENOENT is returned, you should
 689  * invoke
 690  *     btrfs_attach_transaction_barrier()
 691  */
 692 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
 693 {
 694         return start_transaction(root, 0, TRANS_ATTACH,
 695                                  BTRFS_RESERVE_NO_FLUSH, true);
 696 }
 697
 698 /*
 699  * btrfs_attach_transaction_barrier() - catch the running transaction
 700  *
 701  * It is similar to the above function, the differentia is this one
 702  * will wait for all the inactive transactions until they fully
 703  * complete.
 704  */
 705 struct btrfs_trans_handle *
 706 btrfs_attach_transaction_barrier(struct btrfs_root *root)
 707 {
 708         struct btrfs_trans_handle *trans;
 709
 710         trans = start_transaction(root, 0, TRANS_ATTACH,
 711                                   BTRFS_RESERVE_NO_FLUSH, true);
 712         if (IS_ERR(trans) && PTR_ERR(trans) == -ENOENT)
 713                 btrfs_wait_for_commit(root->fs_info, 0);
 714
 715         return trans;
 716 }
 717
 718 /* wait for a transaction commit to be fully complete */
 719 static noinline void wait_for_commit(struct btrfs_transaction *commit)
 720 {
 721         wait_event(commit->commit_wait, commit->state == TRANS_STATE_COMPLETED);
 722 }
 723
 724 int btrfs_wait_for_commit(struct btrfs_fs_info *fs_info, u64 transid)
 725 {
 726         struct btrfs_transaction *cur_trans = NULL, *t;
 727         int ret = 0;
 728
 729         if (transid) {
 730                 if (transid <= fs_info->last_trans_committed)
 731                         goto out;
 732
 733                 /* find specified transaction */
 734                 spin_lock(&fs_info->trans_lock);
 735                 list_for_each_entry(t, &fs_info->trans_list, list) {
 736                         if (t->transid == transid) {
 737                                 cur_trans = t;
 738                                 refcount_inc(&cur_trans->use_count);
 739                                 ret = 0;
 740                                 break;
 741                         }
 742                         if (t->transid > transid) {
 743                                 ret = 0;
 744                                 break;
 745                         }
 746                 }
 747                 spin_unlock(&fs_info->trans_lock);
 748
 749                 /*
 750                  * The specified transaction doesn't exist, or we
 751                  * raced with btrfs_commit_transaction
 752                  */
 753                 if (!cur_trans) {
 754                         if (transid > fs_info->last_trans_committed)
 755                                 ret = -EINVAL;
 756                         goto out;
 757                 }
 758         } else {
 759                 /* find newest transaction that is committing | committed */
 760                 spin_lock(&fs_info->trans_lock);
 761                 list_for_each_entry_reverse(t, &fs_info->trans_list,
 762                                             list) {
 763                         if (t->state >= TRANS_STATE_COMMIT_START) {
 764                                 if (t->state == TRANS_STATE_COMPLETED)
 765                                         break;
 766                                 cur_trans = t;
 767                                 refcount_inc(&cur_trans->use_count);
 768                                 break;
 769                         }
 770                 }
 771                 spin_unlock(&fs_info->trans_lock);
 772                 if (!cur_trans)
 773                         goto out;  /* nothing committing|committed */
 774         }
 775
 776         wait_for_commit(cur_trans);
 777         btrfs_put_transaction(cur_trans);
 778 out:
 779         return ret;
 780 }
 781
 782 void btrfs_throttle(struct btrfs_fs_info *fs_info)
 783 {
 784         if (!atomic_read(&fs_info->open_ioctl_trans))
 785                 wait_current_trans(fs_info);
 786 }
 787
 788 static int should_end_transaction(struct btrfs_trans_handle *trans)
 789 {
 790         struct btrfs_fs_info *fs_info = trans->fs_info;
 791
 792         if (btrfs_check_space_for_delayed_refs(trans, fs_info))
 793                 return 1;
 794
 795         return !!btrfs_block_rsv_check(&fs_info->global_block_rsv, 5);
 796 }
 797
 798 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans)
 799 {
 800         struct btrfs_transaction *cur_trans = trans->transaction;
 801         struct btrfs_fs_info *fs_info = trans->fs_info;
 802         int updates;
 803         int err;
 804
 805         smp_mb();
 806         if (cur_trans->state >= TRANS_STATE_BLOCKED ||
 807             cur_trans->delayed_refs.flushing)
 808                 return 1;
 809
 810         updates = trans->delayed_ref_updates;
 811         trans->delayed_ref_updates = 0;
 812         if (updates) {
 813                 err = btrfs_run_delayed_refs(trans, fs_info, updates * 2);
 814                 if (err) /* Error code will also eval true */
 815                         return err;
 816         }
 817
 818         return should_end_transaction(trans);
 819 }
 820
 821 static void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans)
 822
 823 {
 824         struct btrfs_fs_info *fs_info = trans->fs_info;
 825
 826         if (!trans->block_rsv) {
 827                 ASSERT(!trans->bytes_reserved);
 828                 return;
 829         }
 830
 831         if (!trans->bytes_reserved)
 832                 return;
 833
 834         ASSERT(trans->block_rsv == &fs_info->trans_block_rsv);
 835         trace_btrfs_space_reservation(fs_info, "transaction",
 836                                       trans->transid, trans->bytes_reserved, 0);
 837         btrfs_block_rsv_release(fs_info, trans->block_rsv,
 838                                 trans->bytes_reserved);
 839         trans->bytes_reserved = 0;
 840 }
 841
 842 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 843                                    int throttle)
 844 {
 845         struct btrfs_fs_info *info = trans->fs_info;
 846         struct btrfs_transaction *cur_trans = trans->transaction;
 847         u64 transid = trans->transid;
 848         unsigned long cur = trans->delayed_ref_updates;
 849         int lock = (trans->type != TRANS_JOIN_NOLOCK);
 850         int err = 0;
 851         int must_run_delayed_refs = 0;
 852
 853         if (refcount_read(&trans->use_count) > 1) {
 854                 refcount_dec(&trans->use_count);
 855                 trans->block_rsv = trans->orig_rsv;
 856                 return 0;
 857         }
 858
 859         btrfs_trans_release_metadata(trans);
 860         trans->block_rsv = NULL;
 861
 862         if (!list_empty(&trans->new_bgs))
 863                 btrfs_create_pending_block_groups(trans, info);
 864
 865         trans->delayed_ref_updates = 0;
 866         if (!trans->sync) {
 867                 must_run_delayed_refs =
 868                         btrfs_should_throttle_delayed_refs(trans, info);
 869                 cur = max_t(unsigned long, cur, 32);
 870
 871                 /*
 872                  * don't make the caller wait if they are from a NOLOCK
 873                  * or ATTACH transaction, it will deadlock with commit
 874                  */
 875                 if (must_run_delayed_refs == 1 &&
 876                     (trans->type & (__TRANS_JOIN_NOLOCK | __TRANS_ATTACH)))
 877                         must_run_delayed_refs = 2;
 878         }
 879
 880         btrfs_trans_release_metadata(trans);
 881         trans->block_rsv = NULL;
 882
 883         if (!list_empty(&trans->new_bgs))
 884                 btrfs_create_pending_block_groups(trans, info);
 885
 886         btrfs_trans_release_chunk_metadata(trans);
 887
 888         if (lock && !atomic_read(&info->open_ioctl_trans) &&
 889             should_end_transaction(trans) &&
 890             READ_ONCE(cur_trans->state) == TRANS_STATE_RUNNING) {
 891                 spin_lock(&info->trans_lock);
 892                 if (cur_trans->state == TRANS_STATE_RUNNING)
 893                         cur_trans->state = TRANS_STATE_BLOCKED;
 894                 spin_unlock(&info->trans_lock);
 895         }
 896
 897         if (lock && READ_ONCE(cur_trans->state) == TRANS_STATE_BLOCKED) {
 898                 if (throttle)
 899                         return btrfs_commit_transaction(trans);
 900                 else
 901                         wake_up_process(info->transaction_kthread);
 902         }
 903
 904         if (trans->type & __TRANS_FREEZABLE)
 905                 sb_end_intwrite(info->sb);
 906
 907         WARN_ON(cur_trans != info->running_transaction);
 908         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 909         atomic_dec(&cur_trans->num_writers);
 910         extwriter_counter_dec(cur_trans, trans->type);
 911
 912         /*
 913          * Make sure counter is updated before we wake up waiters.
 914          */
 915         smp_mb();
 916         if (waitqueue_active(&cur_trans->writer_wait))
 917                 wake_up(&cur_trans->writer_wait);
 918         btrfs_put_transaction(cur_trans);
 919
 920         if (current->journal_info == trans)
 921                 current->journal_info = NULL;
 922
 923         if (throttle)
 924                 btrfs_run_delayed_iputs(info);
 925
 926         if (trans->aborted ||
 927             test_bit(BTRFS_FS_STATE_ERROR, &info->fs_state)) {
 928                 wake_up_process(info->transaction_kthread);
 929                 err = -EIO;
 930         }
 931
 932         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 933         if (must_run_delayed_refs) {
 934                 btrfs_async_run_delayed_refs(info, cur, transid,
 935                                              must_run_delayed_refs == 1);
 936         }
 937         return err;
 938 }
 939
 940 int btrfs_end_transaction(struct btrfs_trans_handle *trans)
 941 {
 942         return __btrfs_end_transaction(trans, 0);
 943 }
 944
 945 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans)
 946 {
 947         return __btrfs_end_transaction(trans, 1);
 948 }
 949
 950 /*
 951  * when btree blocks are allocated, they have some corresponding bits set for
 952  * them in one of two extent_io trees.  This is used to make sure all of
 953  * those extents are sent to disk but does not wait on them
 954  */
 955 int btrfs_write_marked_extents(struct btrfs_fs_info *fs_info,
 956                                struct extent_io_tree *dirty_pages, int mark)
 957 {
 958         int err = 0;
 959         int werr = 0;
 960         struct address_space *mapping = fs_info->btree_inode->i_mapping;
 961         struct extent_state *cached_state = NULL;
 962         u64 start = 0;
 963         u64 end;
 964
 965         atomic_inc(&BTRFS_I(fs_info->btree_inode)->sync_writers);
 966         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 967                                       mark, &cached_state)) {
 968                 bool wait_writeback = false;
 969
 970                 err = convert_extent_bit(dirty_pages, start, end,
 971                                          EXTENT_NEED_WAIT,
 972                                          mark, &cached_state);
 973                 /*
 974                  * convert_extent_bit can return -ENOMEM, which is most of the
 975                  * time a temporary error. So when it happens, ignore the error
 976                  * and wait for writeback of this range to finish - because we
 977                  * failed to set the bit EXTENT_NEED_WAIT for the range, a call
 978                  * to __btrfs_wait_marked_extents() would not know that
 979                  * writeback for this range started and therefore wouldn't
 980                  * wait for it to finish - we don't want to commit a
 981                  * superblock that points to btree nodes/leafs for which
 982                  * writeback hasn't finished yet (and without errors).
 983                  * We cleanup any entries left in the io tree when committing
 984                  * the transaction (through clear_btree_io_tree()).
 985                  */
 986                 if (err == -ENOMEM) {
 987                         err = 0;
 988                         wait_writeback = true;
 989                 }
 990                 if (!err)
 991                         err = filemap_fdatawrite_range(mapping, start, end);
 992                 if (err)
 993                         werr = err;
 994                 else if (wait_writeback)
 995                         werr = filemap_fdatawait_range(mapping, start, end);
 996                 free_extent_state(cached_state);
 997                 cached_state = NULL;
 998                 cond_resched();
 999                 start = end + 1;
1000         }
1001         atomic_dec(&BTRFS_I(fs_info->btree_inode)->sync_writers);
1002         return werr;
1003 }
1004
1005 /*
1006  * when btree blocks are allocated, they have some corresponding bits set for
1007  * them in one of two extent_io trees.  This is used to make sure all of
1008  * those extents are on disk for transaction or log commit.  We wait
1009  * on all the pages and clear them from the dirty pages state tree
1010  */
1011 static int __btrfs_wait_marked_extents(struct btrfs_fs_info *fs_info,
1012                                        struct extent_io_tree *dirty_pages)
1013 {
1014         int err = 0;
1015         int werr = 0;
1016         struct address_space *mapping = fs_info->btree_inode->i_mapping;
1017         struct extent_state *cached_state = NULL;
1018         u64 start = 0;
1019         u64 end;
1020
1021         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
1022                                       EXTENT_NEED_WAIT, &cached_state)) {
1023                 /*
1024                  * Ignore -ENOMEM errors returned by clear_extent_bit().
1025                  * When committing the transaction, we'll remove any entries
1026                  * left in the io tree. For a log commit, we don't remove them
1027                  * after committing the log because the tree can be accessed
1028                  * concurrently - we do it only at transaction commit time when
1029                  * it's safe to do it (through clear_btree_io_tree()).
1030                  */
1031                 err = clear_extent_bit(dirty_pages, start, end,
1032                                        EXTENT_NEED_WAIT, 0, 0, &cached_state);
1033                 if (err == -ENOMEM)
1034                         err = 0;
1035                 if (!err)
1036                         err = filemap_fdatawait_range(mapping, start, end);
1037                 if (err)
1038                         werr = err;
1039                 free_extent_state(cached_state);
1040                 cached_state = NULL;
1041                 cond_resched();
1042                 start = end + 1;
1043         }
1044         if (err)
1045                 werr = err;
1046         return werr;
1047 }
1048
1049 int btrfs_wait_extents(struct btrfs_fs_info *fs_info,
1050                        struct extent_io_tree *dirty_pages)
1051 {
1052         bool errors = false;
1053         int err;
1054
1055         err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1056         if (test_and_clear_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags))
1057                 errors = true;
1058
1059         if (errors && !err)
1060                 err = -EIO;
1061         return err;
1062 }
1063
1064 int btrfs_wait_tree_log_extents(struct btrfs_root *log_root, int mark)
1065 {
1066         struct btrfs_fs_info *fs_info = log_root->fs_info;
1067         struct extent_io_tree *dirty_pages = &log_root->dirty_log_pages;
1068         bool errors = false;
1069         int err;
1070
1071         ASSERT(log_root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
1072
1073         err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1074         if ((mark & EXTENT_DIRTY) &&
1075             test_and_clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags))
1076                 errors = true;
1077
1078         if ((mark & EXTENT_NEW) &&
1079             test_and_clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags))
1080                 errors = true;
1081
1082         if (errors && !err)
1083                 err = -EIO;
1084         return err;
1085 }
1086
1087 /*
1088  * When btree blocks are allocated the corresponding extents are marked dirty.
1089  * This function ensures such extents are persisted on disk for transaction or
1090  * log commit.
1091  *
1092  * @trans: transaction whose dirty pages we'd like to write
1093  */
1094 static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
1095                                             struct btrfs_fs_info *fs_info)
1096 {
1097         int ret;
1098         int ret2;
1099         struct extent_io_tree *dirty_pages = &trans->transaction->dirty_pages;
1100         struct blk_plug plug;
1101
1102         blk_start_plug(&plug);
1103         ret = btrfs_write_marked_extents(fs_info, dirty_pages, EXTENT_DIRTY);
1104         blk_finish_plug(&plug);
1105         ret2 = btrfs_wait_extents(fs_info, dirty_pages);
1106
1107         clear_btree_io_tree(&trans->transaction->dirty_pages);
1108
1109         if (ret)
1110                 return ret;
1111         else if (ret2)
1112                 return ret2;
1113         else
1114                 return 0;
1115 }
1116
1117 /*
1118  * this is used to update the root pointer in the tree of tree roots.
1119  *
1120  * But, in the case of the extent allocation tree, updating the root
1121  * pointer may allocate blocks which may change the root of the extent
1122  * allocation tree.
1123  *
1124  * So, this loops and repeats and makes sure the cowonly root didn't
1125  * change while the root pointer was being updated in the metadata.
1126  */
1127 static int update_cowonly_root(struct btrfs_trans_handle *trans,
1128                                struct btrfs_root *root)
1129 {
1130         int ret;
1131         u64 old_root_bytenr;
1132         u64 old_root_used;
1133         struct btrfs_fs_info *fs_info = root->fs_info;
1134         struct btrfs_root *tree_root = fs_info->tree_root;
1135
1136         old_root_used = btrfs_root_used(&root->root_item);
1137
1138         while (1) {
1139                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1140                 if (old_root_bytenr == root->node->start &&
1141                     old_root_used == btrfs_root_used(&root->root_item))
1142                         break;
1143
1144                 btrfs_set_root_node(&root->root_item, root->node);
1145                 ret = btrfs_update_root(trans, tree_root,
1146                                         &root->root_key,
1147                                         &root->root_item);
1148                 if (ret)
1149                         return ret;
1150
1151                 old_root_used = btrfs_root_used(&root->root_item);
1152         }
1153
1154         return 0;
1155 }
1156
1157 /*
1158  * update all the cowonly tree roots on disk
1159  *
1160  * The error handling in this function may not be obvious. Any of the
1161  * failures will cause the file system to go offline. We still need
1162  * to clean up the delayed refs.
1163  */
1164 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
1165                                          struct btrfs_fs_info *fs_info)
1166 {
1167         struct list_head *dirty_bgs = &trans->transaction->dirty_bgs;
1168         struct list_head *io_bgs = &trans->transaction->io_bgs;
1169         struct list_head *next;
1170         struct extent_buffer *eb;
1171         int ret;
1172
1173         eb = btrfs_lock_root_node(fs_info->tree_root);
1174         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
1175                               0, &eb);
1176         btrfs_tree_unlock(eb);
1177         free_extent_buffer(eb);
1178
1179         if (ret)
1180                 return ret;
1181
1182         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1183         if (ret)
1184                 return ret;
1185
1186         ret = btrfs_run_dev_stats(trans, fs_info);
1187         if (ret)
1188                 return ret;
1189         ret = btrfs_run_dev_replace(trans, fs_info);
1190         if (ret)
1191                 return ret;
1192         ret = btrfs_run_qgroups(trans, fs_info);
1193         if (ret)
1194                 return ret;
1195
1196         ret = btrfs_setup_space_cache(trans, fs_info);
1197         if (ret)
1198                 return ret;
1199
1200         /* run_qgroups might have added some more refs */
1201         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1202         if (ret)
1203                 return ret;
1204 again:
1205         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1206                 struct btrfs_root *root;
1207                 next = fs_info->dirty_cowonly_roots.next;
1208                 list_del_init(next);
1209                 root = list_entry(next, struct btrfs_root, dirty_list);
1210                 clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1211
1212                 if (root != fs_info->extent_root)
1213                         list_add_tail(&root->dirty_list,
1214                                       &trans->transaction->switch_commits);
1215                 ret = update_cowonly_root(trans, root);
1216                 if (ret)
1217                         return ret;
1218                 ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1219                 if (ret)
1220                         return ret;
1221         }
1222
1223         while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1224                 ret = btrfs_write_dirty_block_groups(trans, fs_info);
1225                 if (ret)
1226                         return ret;
1227                 ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1228                 if (ret)
1229                         return ret;
1230         }
1231
1232         if (!list_empty(&fs_info->dirty_cowonly_roots))
1233                 goto again;
1234
1235         list_add_tail(&fs_info->extent_root->dirty_list,
1236                       &trans->transaction->switch_commits);
1237         btrfs_after_dev_replace_commit(fs_info);
1238
1239         return 0;
1240 }
1241
1242 /*
1243  * dead roots are old snapshots that need to be deleted.  This allocates
1244  * a dirty root struct and adds it into the list of dead roots that need to
1245  * be deleted
1246  */
1247 void btrfs_add_dead_root(struct btrfs_root *root)
1248 {
1249         struct btrfs_fs_info *fs_info = root->fs_info;
1250
1251         spin_lock(&fs_info->trans_lock);
1252         if (list_empty(&root->root_list))
1253                 list_add_tail(&root->root_list, &fs_info->dead_roots);
1254         spin_unlock(&fs_info->trans_lock);
1255 }
1256
1257 /*
1258  * update all the cowonly tree roots on disk
1259  */
1260 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
1261                                     struct btrfs_fs_info *fs_info)
1262 {
1263         struct btrfs_root *gang[8];
1264         int i;
1265         int ret;
1266         int err = 0;
1267
1268         spin_lock(&fs_info->fs_roots_radix_lock);
1269         while (1) {
1270                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1271                                                  (void **)gang, 0,
1272                                                  ARRAY_SIZE(gang),
1273                                                  BTRFS_ROOT_TRANS_TAG);
1274                 if (ret == 0)
1275                         break;
1276                 for (i = 0; i < ret; i++) {
1277                         struct btrfs_root *root = gang[i];
1278                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
1279                                         (unsigned long)root->root_key.objectid,
1280                                         BTRFS_ROOT_TRANS_TAG);
1281                         spin_unlock(&fs_info->fs_roots_radix_lock);
1282
1283                         btrfs_free_log(trans, root);
1284                         btrfs_update_reloc_root(trans, root);
1285                         btrfs_orphan_commit_root(trans, root);
1286
1287                         btrfs_save_ino_cache(root, trans);
1288
1289                         /* see comments in should_cow_block() */
1290                         clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1291                         smp_mb__after_atomic();
1292
1293                         if (root->commit_root != root->node) {
1294                                 list_add_tail(&root->dirty_list,
1295                                         &trans->transaction->switch_commits);
1296                                 btrfs_set_root_node(&root->root_item,
1297                                                     root->node);
1298                         }
1299
1300                         err = btrfs_update_root(trans, fs_info->tree_root,
1301                                                 &root->root_key,
1302                                                 &root->root_item);
1303                         spin_lock(&fs_info->fs_roots_radix_lock);
1304                         if (err)
1305                                 break;
1306                         btrfs_qgroup_free_meta_all(root);
1307                 }
1308         }
1309         spin_unlock(&fs_info->fs_roots_radix_lock);
1310         return err;
1311 }
1312
1313 /*
1314  * defrag a given btree.
1315  * Every leaf in the btree is read and defragged.
1316  */
1317 int btrfs_defrag_root(struct btrfs_root *root)
1318 {
1319         struct btrfs_fs_info *info = root->fs_info;
1320         struct btrfs_trans_handle *trans;
1321         int ret;
1322
1323         if (test_and_set_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state))
1324                 return 0;
1325
1326         while (1) {
1327                 trans = btrfs_start_transaction(root, 0);
1328                 if (IS_ERR(trans))
1329                         return PTR_ERR(trans);
1330
1331                 ret = btrfs_defrag_leaves(trans, root);
1332
1333                 btrfs_end_transaction(trans);
1334                 btrfs_btree_balance_dirty(info);
1335                 cond_resched();
1336
1337                 if (btrfs_fs_closing(info) || ret != -EAGAIN)
1338                         break;
1339
1340                 if (btrfs_defrag_cancelled(info)) {
1341                         btrfs_debug(info, "defrag_root cancelled");
1342                         ret = -EAGAIN;
1343                         break;
1344                 }
1345         }
1346         clear_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state);
1347         return ret;
1348 }
1349
1350 /*
1351  * Do all special snapshot related qgroup dirty hack.
1352  *
1353  * Will do all needed qgroup inherit and dirty hack like switch commit
1354  * roots inside one transaction and write all btree into disk, to make
1355  * qgroup works.
1356  */
1357 static int qgroup_account_snapshot(struct btrfs_trans_handle *trans,
1358                                    struct btrfs_root *src,
1359                                    struct btrfs_root *parent,
1360                                    struct btrfs_qgroup_inherit *inherit,
1361                                    u64 dst_objectid)
1362 {
1363         struct btrfs_fs_info *fs_info = src->fs_info;
1364         int ret;
1365
1366         /*
1367          * Save some performance in the case that qgroups are not
1368          * enabled. If this check races with the ioctl, rescan will
1369          * kick in anyway.
1370          */
1371         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
1372                 return 0;
1373
1374         /*
1375          * We are going to commit transaction, see btrfs_commit_transaction()
1376          * comment for reason locking tree_log_mutex
1377          */
1378         mutex_lock(&fs_info->tree_log_mutex);
1379
1380         ret = commit_fs_roots(trans, fs_info);
1381         if (ret)
1382                 goto out;
1383         ret = btrfs_qgroup_account_extents(trans, fs_info);
1384         if (ret < 0)
1385                 goto out;
1386
1387         /* Now qgroup are all updated, we can inherit it to new qgroups */
1388         ret = btrfs_qgroup_inherit(trans, fs_info,
1389                                    src->root_key.objectid, dst_objectid,
1390                                    inherit);
1391         if (ret < 0)
1392                 goto out;
1393
1394         /*
1395          * Now we do a simplified commit transaction, which will:
1396          * 1) commit all subvolume and extent tree
1397          *    To ensure all subvolume and extent tree have a valid
1398          *    commit_root to accounting later insert_dir_item()
1399          * 2) write all btree blocks onto disk
1400          *    This is to make sure later btree modification will be cowed
1401          *    Or commit_root can be populated and cause wrong qgroup numbers
1402          * In this simplified commit, we don't really care about other trees
1403          * like chunk and root tree, as they won't affect qgroup.
1404          * And we don't write super to avoid half committed status.
1405          */
1406         ret = commit_cowonly_roots(trans, fs_info);
1407         if (ret)
1408                 goto out;
1409         switch_commit_roots(trans->transaction, fs_info);
1410         ret = btrfs_write_and_wait_transaction(trans, fs_info);
1411         if (ret)
1412                 btrfs_handle_fs_error(fs_info, ret,
1413                         "Error while writing out transaction for qgroup");
1414
1415 out:
1416         mutex_unlock(&fs_info->tree_log_mutex);
1417
1418         /*
1419          * Force parent root to be updated, as we recorded it before so its
1420          * last_trans == cur_transid.
1421          * Or it won't be committed again onto disk after later
1422          * insert_dir_item()
1423          */
1424         if (!ret)
1425                 record_root_in_trans(trans, parent, 1);
1426         return ret;
1427 }
1428
1429 /*
1430  * new snapshots need to be created at a very specific time in the
1431  * transaction commit.  This does the actual creation.
1432  *
1433  * Note:
1434  * If the error which may affect the commitment of the current transaction
1435  * happens, we should return the error number. If the error which just affect
1436  * the creation of the pending snapshots, just return 0.
1437  */
1438 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1439                                    struct btrfs_fs_info *fs_info,
1440                                    struct btrfs_pending_snapshot *pending)
1441 {
1442         struct btrfs_key key;
1443         struct btrfs_root_item *new_root_item;
1444         struct btrfs_root *tree_root = fs_info->tree_root;
1445         struct btrfs_root *root = pending->root;
1446         struct btrfs_root *parent_root;
1447         struct btrfs_block_rsv *rsv;
1448         struct inode *parent_inode;
1449         struct btrfs_path *path;
1450         struct btrfs_dir_item *dir_item;
1451         struct dentry *dentry;
1452         struct extent_buffer *tmp;
1453         struct extent_buffer *old;
1454         struct timespec cur_time;
1455         int ret = 0;
1456         u64 to_reserve = 0;
1457         u64 index = 0;
1458         u64 objectid;
1459         u64 root_flags;
1460         uuid_le new_uuid;
1461
1462         ASSERT(pending->path);
1463         path = pending->path;
1464
1465         ASSERT(pending->root_item);
1466         new_root_item = pending->root_item;
1467
1468         pending->error = btrfs_find_free_objectid(tree_root, &objectid);
1469         if (pending->error)
1470                 goto no_free_objectid;
1471
1472         /*
1473          * Make qgroup to skip current new snapshot's qgroupid, as it is
1474          * accounted by later btrfs_qgroup_inherit().
1475          */
1476         btrfs_set_skip_qgroup(trans, objectid);
1477
1478         btrfs_reloc_pre_snapshot(pending, &to_reserve);
1479
1480         if (to_reserve > 0) {
1481                 pending->error = btrfs_block_rsv_add(root,
1482                                                      &pending->block_rsv,
1483                                                      to_reserve,
1484                                                      BTRFS_RESERVE_NO_FLUSH);
1485                 if (pending->error)
1486                         goto clear_skip_qgroup;
1487         }
1488
1489         key.objectid = objectid;
1490         key.offset = (u64)-1;
1491         key.type = BTRFS_ROOT_ITEM_KEY;
1492
1493         rsv = trans->block_rsv;
1494         trans->block_rsv = &pending->block_rsv;
1495         trans->bytes_reserved = trans->block_rsv->reserved;
1496         trace_btrfs_space_reservation(fs_info, "transaction",
1497                                       trans->transid,
1498                                       trans->bytes_reserved, 1);
1499         dentry = pending->dentry;
1500         parent_inode = pending->dir;
1501         parent_root = BTRFS_I(parent_inode)->root;
1502         record_root_in_trans(trans, parent_root, 0);
1503
1504         cur_time = current_time(parent_inode);
1505
1506         /*
1507          * insert the directory item
1508          */
1509         ret = btrfs_set_inode_index(BTRFS_I(parent_inode), &index);
1510         BUG_ON(ret); /* -ENOMEM */
1511
1512         /* check if there is a file/dir which has the same name. */
1513         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1514                                          btrfs_ino(BTRFS_I(parent_inode)),
1515                                          dentry->d_name.name,
1516                                          dentry->d_name.len, 0);
1517         if (dir_item != NULL && !IS_ERR(dir_item)) {
1518                 pending->error = -EEXIST;
1519                 goto dir_item_existed;
1520         } else if (IS_ERR(dir_item)) {
1521                 ret = PTR_ERR(dir_item);
1522                 btrfs_abort_transaction(trans, ret);
1523                 goto fail;
1524         }
1525         btrfs_release_path(path);
1526
1527         /*
1528          * pull in the delayed directory update
1529          * and the delayed inode item
1530          * otherwise we corrupt the FS during
1531          * snapshot
1532          */
1533         ret = btrfs_run_delayed_items(trans, fs_info);
1534         if (ret) {      /* Transaction aborted */
1535                 btrfs_abort_transaction(trans, ret);
1536                 goto fail;
1537         }
1538
1539         record_root_in_trans(trans, root, 0);
1540         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1541         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1542         btrfs_check_and_init_root_item(new_root_item);
1543
1544         root_flags = btrfs_root_flags(new_root_item);
1545         if (pending->readonly)
1546                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1547         else
1548                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1549         btrfs_set_root_flags(new_root_item, root_flags);
1550
1551         btrfs_set_root_generation_v2(new_root_item,
1552                         trans->transid);
1553         uuid_le_gen(&new_uuid);
1554         memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1555         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1556                         BTRFS_UUID_SIZE);
1557         if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1558                 memset(new_root_item->received_uuid, 0,
1559                        sizeof(new_root_item->received_uuid));
1560                 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1561                 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1562                 btrfs_set_root_stransid(new_root_item, 0);
1563                 btrfs_set_root_rtransid(new_root_item, 0);
1564         }
1565         btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1566         btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1567         btrfs_set_root_otransid(new_root_item, trans->transid);
1568
1569         old = btrfs_lock_root_node(root);
1570         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1571         if (ret) {
1572                 btrfs_tree_unlock(old);
1573                 free_extent_buffer(old);
1574                 btrfs_abort_transaction(trans, ret);
1575                 goto fail;
1576         }
1577
1578         btrfs_set_lock_blocking(old);
1579
1580         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1581         /* clean up in any case */
1582         btrfs_tree_unlock(old);
1583         free_extent_buffer(old);
1584         if (ret) {
1585                 btrfs_abort_transaction(trans, ret);
1586                 goto fail;
1587         }
1588         /* see comments in should_cow_block() */
1589         set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1590         smp_wmb();
1591
1592         btrfs_set_root_node(new_root_item, tmp);
1593         /* record when the snapshot was created in key.offset */
1594         key.offset = trans->transid;
1595         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1596         btrfs_tree_unlock(tmp);
1597         free_extent_buffer(tmp);
1598         if (ret) {
1599                 btrfs_abort_transaction(trans, ret);
1600                 goto fail;
1601         }
1602
1603         /*
1604          * insert root back/forward references
1605          */
1606         ret = btrfs_add_root_ref(trans, fs_info, objectid,
1607                                  parent_root->root_key.objectid,
1608                                  btrfs_ino(BTRFS_I(parent_inode)), index,
1609                                  dentry->d_name.name, dentry->d_name.len);
1610         if (ret) {
1611                 btrfs_abort_transaction(trans, ret);
1612                 goto fail;
1613         }
1614
1615         key.offset = (u64)-1;
1616         pending->snap = btrfs_read_fs_root_no_name(fs_info, &key);
1617         if (IS_ERR(pending->snap)) {
1618                 ret = PTR_ERR(pending->snap);
1619                 btrfs_abort_transaction(trans, ret);
1620                 goto fail;
1621         }
1622
1623         ret = btrfs_reloc_post_snapshot(trans, pending);
1624         if (ret) {
1625                 btrfs_abort_transaction(trans, ret);
1626                 goto fail;
1627         }
1628
1629         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1630         if (ret) {
1631                 btrfs_abort_transaction(trans, ret);
1632                 goto fail;
1633         }
1634
1635         /*
1636          * Do special qgroup accounting for snapshot, as we do some qgroup
1637          * snapshot hack to do fast snapshot.
1638          * To co-operate with that hack, we do hack again.
1639          * Or snapshot will be greatly slowed down by a subtree qgroup rescan
1640          */
1641         ret = qgroup_account_snapshot(trans, root, parent_root,
1642                                       pending->inherit, objectid);
1643         if (ret < 0)
1644                 goto fail;
1645
1646         ret = btrfs_insert_dir_item(trans, parent_root,
1647                                     dentry->d_name.name, dentry->d_name.len,
1648                                     BTRFS_I(parent_inode), &key,
1649                                     BTRFS_FT_DIR, index);
1650         /* We have check then name at the beginning, so it is impossible. */
1651         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1652         if (ret) {
1653                 btrfs_abort_transaction(trans, ret);
1654                 goto fail;
1655         }
1656
1657         btrfs_i_size_write(BTRFS_I(parent_inode), parent_inode->i_size +
1658                                          dentry->d_name.len * 2);
1659         parent_inode->i_mtime = parent_inode->i_ctime =
1660                 current_time(parent_inode);
1661         ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1662         if (ret) {
1663                 btrfs_abort_transaction(trans, ret);
1664                 goto fail;
1665         }
1666         ret = btrfs_uuid_tree_add(trans, fs_info, new_uuid.b,
1667                                   BTRFS_UUID_KEY_SUBVOL, objectid);
1668         if (ret) {
1669                 btrfs_abort_transaction(trans, ret);
1670                 goto fail;
1671         }
1672         if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1673                 ret = btrfs_uuid_tree_add(trans, fs_info,
1674                                           new_root_item->received_uuid,
1675                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1676                                           objectid);
1677                 if (ret && ret != -EEXIST) {
1678                         btrfs_abort_transaction(trans, ret);
1679                         goto fail;
1680                 }
1681         }
1682
1683         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
1684         if (ret) {
1685                 btrfs_abort_transaction(trans, ret);
1686                 goto fail;
1687         }
1688
1689 fail:
1690         pending->error = ret;
1691 dir_item_existed:
1692         trans->block_rsv = rsv;
1693         trans->bytes_reserved = 0;
1694 clear_skip_qgroup:
1695         btrfs_clear_skip_qgroup(trans);
1696 no_free_objectid:
1697         kfree(new_root_item);
1698         pending->root_item = NULL;
1699         btrfs_free_path(path);
1700         pending->path = NULL;
1701
1702         return ret;
1703 }
1704
1705 /*
1706  * create all the snapshots we've scheduled for creation
1707  */
1708 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1709                                              struct btrfs_fs_info *fs_info)
1710 {
1711         struct btrfs_pending_snapshot *pending, *next;
1712         struct list_head *head = &trans->transaction->pending_snapshots;
1713         int ret = 0;
1714
1715         list_for_each_entry_safe(pending, next, head, list) {
1716                 list_del(&pending->list);
1717                 ret = create_pending_snapshot(trans, fs_info, pending);
1718                 if (ret)
1719                         break;
1720         }
1721         return ret;
1722 }
1723
1724 static void update_super_roots(struct btrfs_fs_info *fs_info)
1725 {
1726         struct btrfs_root_item *root_item;
1727         struct btrfs_super_block *super;
1728
1729         super = fs_info->super_copy;
1730
1731         root_item = &fs_info->chunk_root->root_item;
1732         super->chunk_root = root_item->bytenr;
1733         super->chunk_root_generation = root_item->generation;
1734         super->chunk_root_level = root_item->level;
1735
1736         root_item = &fs_info->tree_root->root_item;
1737         super->root = root_item->bytenr;
1738         super->generation = root_item->generation;
1739         super->root_level = root_item->level;
1740         if (btrfs_test_opt(fs_info, SPACE_CACHE))
1741                 super->cache_generation = root_item->generation;
1742         if (test_bit(BTRFS_FS_UPDATE_UUID_TREE_GEN, &fs_info->flags))
1743                 super->uuid_tree_generation = root_item->generation;
1744 }
1745
1746 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1747 {
1748         struct btrfs_transaction *trans;
1749         int ret = 0;
1750
1751         spin_lock(&info->trans_lock);
1752         trans = info->running_transaction;
1753         if (trans)
1754                 ret = (trans->state >= TRANS_STATE_COMMIT_START);
1755         spin_unlock(&info->trans_lock);
1756         return ret;
1757 }
1758
1759 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1760 {
1761         struct btrfs_transaction *trans;
1762         int ret = 0;
1763
1764         spin_lock(&info->trans_lock);
1765         trans = info->running_transaction;
1766         if (trans)
1767                 ret = is_transaction_blocked(trans);
1768         spin_unlock(&info->trans_lock);
1769         return ret;
1770 }
1771
1772 /*
1773  * wait for the current transaction commit to start and block subsequent
1774  * transaction joins
1775  */
1776 static void wait_current_trans_commit_start(struct btrfs_fs_info *fs_info,
1777                                             struct btrfs_transaction *trans)
1778 {
1779         wait_event(fs_info->transaction_blocked_wait,
1780                    trans->state >= TRANS_STATE_COMMIT_START || trans->aborted);
1781 }
1782
1783 /*
1784  * wait for the current transaction to start and then become unblocked.
1785  * caller holds ref.
1786  */
1787 static void wait_current_trans_commit_start_and_unblock(
1788                                         struct btrfs_fs_info *fs_info,
1789                                         struct btrfs_transaction *trans)
1790 {
1791         wait_event(fs_info->transaction_wait,
1792                    trans->state >= TRANS_STATE_UNBLOCKED || trans->aborted);
1793 }
1794
1795 /*
1796  * commit transactions asynchronously. once btrfs_commit_transaction_async
1797  * returns, any subsequent transaction will not be allowed to join.
1798  */
1799 struct btrfs_async_commit {
1800         struct btrfs_trans_handle *newtrans;
1801         struct work_struct work;
1802 };
1803
1804 static void do_async_commit(struct work_struct *work)
1805 {
1806         struct btrfs_async_commit *ac =
1807                 container_of(work, struct btrfs_async_commit, work);
1808
1809         /*
1810          * We've got freeze protection passed with the transaction.
1811          * Tell lockdep about it.
1812          */
1813         if (ac->newtrans->type & __TRANS_FREEZABLE)
1814                 __sb_writers_acquired(ac->newtrans->fs_info->sb, SB_FREEZE_FS);
1815
1816         current->journal_info = ac->newtrans;
1817
1818         btrfs_commit_transaction(ac->newtrans);
1819         kfree(ac);
1820 }
1821
1822 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1823                                    int wait_for_unblock)
1824 {
1825         struct btrfs_fs_info *fs_info = trans->fs_info;
1826         struct btrfs_async_commit *ac;
1827         struct btrfs_transaction *cur_trans;
1828
1829         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1830         if (!ac)
1831                 return -ENOMEM;
1832
1833         INIT_WORK(&ac->work, do_async_commit);
1834         ac->newtrans = btrfs_join_transaction(trans->root);
1835         if (IS_ERR(ac->newtrans)) {
1836                 int err = PTR_ERR(ac->newtrans);
1837                 kfree(ac);
1838                 return err;
1839         }
1840
1841         /* take transaction reference */
1842         cur_trans = trans->transaction;
1843         refcount_inc(&cur_trans->use_count);
1844
1845         btrfs_end_transaction(trans);
1846
1847         /*
1848          * Tell lockdep we've released the freeze rwsem, since the
1849          * async commit thread will be the one to unlock it.
1850          */
1851         if (ac->newtrans->type & __TRANS_FREEZABLE)
1852                 __sb_writers_release(fs_info->sb, SB_FREEZE_FS);
1853
1854         schedule_work(&ac->work);
1855
1856         /* wait for transaction to start and unblock */
1857         if (wait_for_unblock)
1858                 wait_current_trans_commit_start_and_unblock(fs_info, cur_trans);
1859         else
1860                 wait_current_trans_commit_start(fs_info, cur_trans);
1861
1862         if (current->journal_info == trans)
1863                 current->journal_info = NULL;
1864
1865         btrfs_put_transaction(cur_trans);
1866         return 0;
1867 }
1868
1869
1870 static void cleanup_transaction(struct btrfs_trans_handle *trans,
1871                                 struct btrfs_root *root, int err)
1872 {
1873         struct btrfs_fs_info *fs_info = root->fs_info;
1874         struct btrfs_transaction *cur_trans = trans->transaction;
1875         DEFINE_WAIT(wait);
1876
1877         WARN_ON(refcount_read(&trans->use_count) > 1);
1878
1879         btrfs_abort_transaction(trans, err);
1880
1881         spin_lock(&fs_info->trans_lock);
1882
1883         /*
1884          * If the transaction is removed from the list, it means this
1885          * transaction has been committed successfully, so it is impossible
1886          * to call the cleanup function.
1887          */
1888         BUG_ON(list_empty(&cur_trans->list));
1889
1890         list_del_init(&cur_trans->list);
1891         if (cur_trans == fs_info->running_transaction) {
1892                 cur_trans->state = TRANS_STATE_COMMIT_DOING;
1893                 spin_unlock(&fs_info->trans_lock);
1894                 wait_event(cur_trans->writer_wait,
1895                            atomic_read(&cur_trans->num_writers) == 1);
1896
1897                 spin_lock(&fs_info->trans_lock);
1898         }
1899         spin_unlock(&fs_info->trans_lock);
1900
1901         btrfs_cleanup_one_transaction(trans->transaction, fs_info);
1902
1903         spin_lock(&fs_info->trans_lock);
1904         if (cur_trans == fs_info->running_transaction)
1905                 fs_info->running_transaction = NULL;
1906         spin_unlock(&fs_info->trans_lock);
1907
1908         if (trans->type & __TRANS_FREEZABLE)
1909                 sb_end_intwrite(fs_info->sb);
1910         btrfs_put_transaction(cur_trans);
1911         btrfs_put_transaction(cur_trans);
1912
1913         trace_btrfs_transaction_commit(root);
1914
1915         if (current->journal_info == trans)
1916                 current->journal_info = NULL;
1917         btrfs_scrub_cancel(fs_info);
1918
1919         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1920 }
1921
1922 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
1923 {
1924         /*
1925          * We use writeback_inodes_sb here because if we used
1926          * btrfs_start_delalloc_roots we would deadlock with fs freeze.
1927          * Currently are holding the fs freeze lock, if we do an async flush
1928          * we'll do btrfs_join_transaction() and deadlock because we need to
1929          * wait for the fs freeze lock.  Using the direct flushing we benefit
1930          * from already being in a transaction and our join_transaction doesn't
1931          * have to re-take the fs freeze lock.
1932          */
1933         if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
1934                 writeback_inodes_sb(fs_info->sb, WB_REASON_SYNC);
1935         return 0;
1936 }
1937
1938 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
1939 {
1940         if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
1941                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0, (u64)-1);
1942 }
1943
1944 static inline void
1945 btrfs_wait_pending_ordered(struct btrfs_transaction *cur_trans)
1946 {
1947         wait_event(cur_trans->pending_wait,
1948                    atomic_read(&cur_trans->pending_ordered) == 0);
1949 }
1950
1951 int btrfs_commit_transaction(struct btrfs_trans_handle *trans)
1952 {
1953         struct btrfs_fs_info *fs_info = trans->fs_info;
1954         struct btrfs_transaction *cur_trans = trans->transaction;
1955         struct btrfs_transaction *prev_trans = NULL;
1956         int ret;
1957
1958         /* Stop the commit early if ->aborted is set */
1959         if (unlikely(READ_ONCE(cur_trans->aborted))) {
1960                 ret = cur_trans->aborted;
1961                 btrfs_end_transaction(trans);
1962                 return ret;
1963         }
1964
1965         /* make a pass through all the delayed refs we have so far
1966          * any runnings procs may add more while we are here
1967          */
1968         ret = btrfs_run_delayed_refs(trans, fs_info, 0);
1969         if (ret) {
1970                 btrfs_end_transaction(trans);
1971                 return ret;
1972         }
1973
1974         btrfs_trans_release_metadata(trans);
1975         trans->block_rsv = NULL;
1976
1977         cur_trans = trans->transaction;
1978
1979         /*
1980          * set the flushing flag so procs in this transaction have to
1981          * start sending their work down.
1982          */
1983         cur_trans->delayed_refs.flushing = 1;
1984         smp_wmb();
1985
1986         if (!list_empty(&trans->new_bgs))
1987                 btrfs_create_pending_block_groups(trans, fs_info);
1988
1989         ret = btrfs_run_delayed_refs(trans, fs_info, 0);
1990         if (ret) {
1991                 btrfs_end_transaction(trans);
1992                 return ret;
1993         }
1994
1995         if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) {
1996                 int run_it = 0;
1997
1998                 /* this mutex is also taken before trying to set
1999                  * block groups readonly.  We need to make sure
2000                  * that nobody has set a block group readonly
2001                  * after a extents from that block group have been
2002                  * allocated for cache files.  btrfs_set_block_group_ro
2003                  * will wait for the transaction to commit if it
2004                  * finds BTRFS_TRANS_DIRTY_BG_RUN set.
2005                  *
2006                  * The BTRFS_TRANS_DIRTY_BG_RUN flag is also used to make sure
2007                  * only one process starts all the block group IO.  It wouldn't
2008                  * hurt to have more than one go through, but there's no
2009                  * real advantage to it either.
2010                  */
2011                 mutex_lock(&fs_info->ro_block_group_mutex);
2012                 if (!test_and_set_bit(BTRFS_TRANS_DIRTY_BG_RUN,
2013                                       &cur_trans->flags))
2014                         run_it = 1;
2015                 mutex_unlock(&fs_info->ro_block_group_mutex);
2016
2017                 if (run_it)
2018                         ret = btrfs_start_dirty_block_groups(trans, fs_info);
2019         }
2020         if (ret) {
2021                 btrfs_end_transaction(trans);
2022                 return ret;
2023         }
2024
2025         spin_lock(&fs_info->trans_lock);
2026         if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
2027                 spin_unlock(&fs_info->trans_lock);
2028                 refcount_inc(&cur_trans->use_count);
2029                 ret = btrfs_end_transaction(trans);
2030
2031                 wait_for_commit(cur_trans);
2032
2033                 if (unlikely(cur_trans->aborted))
2034                         ret = cur_trans->aborted;
2035
2036                 btrfs_put_transaction(cur_trans);
2037
2038                 return ret;
2039         }
2040
2041         cur_trans->state = TRANS_STATE_COMMIT_START;
2042         wake_up(&fs_info->transaction_blocked_wait);
2043
2044         if (cur_trans->list.prev != &fs_info->trans_list) {
2045                 prev_trans = list_entry(cur_trans->list.prev,
2046                                         struct btrfs_transaction, list);
2047                 if (prev_trans->state != TRANS_STATE_COMPLETED) {
2048                         refcount_inc(&prev_trans->use_count);
2049                         spin_unlock(&fs_info->trans_lock);
2050
2051                         wait_for_commit(prev_trans);
2052                         ret = prev_trans->aborted;
2053
2054                         btrfs_put_transaction(prev_trans);
2055                         if (ret)
2056                                 goto cleanup_transaction;
2057                 } else {
2058                         spin_unlock(&fs_info->trans_lock);
2059                 }
2060         } else {
2061                 spin_unlock(&fs_info->trans_lock);
2062         }
2063
2064         extwriter_counter_dec(cur_trans, trans->type);
2065
2066         ret = btrfs_start_delalloc_flush(fs_info);
2067         if (ret)
2068                 goto cleanup_transaction;
2069
2070         ret = btrfs_run_delayed_items(trans, fs_info);
2071         if (ret)
2072                 goto cleanup_transaction;
2073
2074         wait_event(cur_trans->writer_wait,
2075                    extwriter_counter_read(cur_trans) == 0);
2076
2077         /* some pending stuffs might be added after the previous flush. */
2078         ret = btrfs_run_delayed_items(trans, fs_info);
2079         if (ret)
2080                 goto cleanup_transaction;
2081
2082         btrfs_wait_delalloc_flush(fs_info);
2083
2084         btrfs_wait_pending_ordered(cur_trans);
2085
2086         btrfs_scrub_pause(fs_info);
2087         /*
2088          * Ok now we need to make sure to block out any other joins while we
2089          * commit the transaction.  We could have started a join before setting
2090          * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
2091          */
2092         spin_lock(&fs_info->trans_lock);
2093         cur_trans->state = TRANS_STATE_COMMIT_DOING;
2094         spin_unlock(&fs_info->trans_lock);
2095         wait_event(cur_trans->writer_wait,
2096                    atomic_read(&cur_trans->num_writers) == 1);
2097
2098         /* ->aborted might be set after the previous check, so check it */
2099         if (unlikely(READ_ONCE(cur_trans->aborted))) {
2100                 ret = cur_trans->aborted;
2101                 goto scrub_continue;
2102         }
2103         /*
2104          * the reloc mutex makes sure that we stop
2105          * the balancing code from coming in and moving
2106          * extents around in the middle of the commit
2107          */
2108         mutex_lock(&fs_info->reloc_mutex);
2109
2110         /*
2111          * We needn't worry about the delayed items because we will
2112          * deal with them in create_pending_snapshot(), which is the
2113          * core function of the snapshot creation.
2114          */
2115         ret = create_pending_snapshots(trans, fs_info);
2116         if (ret) {
2117                 mutex_unlock(&fs_info->reloc_mutex);
2118                 goto scrub_continue;
2119         }
2120
2121         /*
2122          * We insert the dir indexes of the snapshots and update the inode
2123          * of the snapshots' parents after the snapshot creation, so there
2124          * are some delayed items which are not dealt with. Now deal with
2125          * them.
2126          *
2127          * We needn't worry that this operation will corrupt the snapshots,
2128          * because all the tree which are snapshoted will be forced to COW
2129          * the nodes and leaves.
2130          */
2131         ret = btrfs_run_delayed_items(trans, fs_info);
2132         if (ret) {
2133                 mutex_unlock(&fs_info->reloc_mutex);
2134                 goto scrub_continue;
2135         }
2136
2137         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
2138         if (ret) {
2139                 mutex_unlock(&fs_info->reloc_mutex);
2140                 goto scrub_continue;
2141         }
2142
2143         /*
2144          * make sure none of the code above managed to slip in a
2145          * delayed item
2146          */
2147         btrfs_assert_delayed_root_empty(fs_info);
2148
2149         WARN_ON(cur_trans != trans->transaction);
2150
2151         /* btrfs_commit_tree_roots is responsible for getting the
2152          * various roots consistent with each other.  Every pointer
2153          * in the tree of tree roots has to point to the most up to date
2154          * root for every subvolume and other tree.  So, we have to keep
2155          * the tree logging code from jumping in and changing any
2156          * of the trees.
2157          *
2158          * At this point in the commit, there can't be any tree-log
2159          * writers, but a little lower down we drop the trans mutex
2160          * and let new people in.  By holding the tree_log_mutex
2161          * from now until after the super is written, we avoid races
2162          * with the tree-log code.
2163          */
2164         mutex_lock(&fs_info->tree_log_mutex);
2165
2166         ret = commit_fs_roots(trans, fs_info);
2167         if (ret) {
2168                 mutex_unlock(&fs_info->tree_log_mutex);
2169                 mutex_unlock(&fs_info->reloc_mutex);
2170                 goto scrub_continue;
2171         }
2172
2173         /*
2174          * Since the transaction is done, we can apply the pending changes
2175          * before the next transaction.
2176          */
2177         btrfs_apply_pending_changes(fs_info);
2178
2179         /* commit_fs_roots gets rid of all the tree log roots, it is now
2180          * safe to free the root of tree log roots
2181          */
2182         btrfs_free_log_root_tree(trans, fs_info);
2183
2184         /*
2185          * commit_fs_roots() can call btrfs_save_ino_cache(), which generates
2186          * new delayed refs. Must handle them or qgroup can be wrong.
2187          */
2188         ret = btrfs_run_delayed_refs(trans, fs_info, (unsigned long)-1);
2189         if (ret) {
2190                 mutex_unlock(&fs_info->tree_log_mutex);
2191                 mutex_unlock(&fs_info->reloc_mutex);
2192                 goto scrub_continue;
2193         }
2194
2195         /*
2196          * Since fs roots are all committed, we can get a quite accurate
2197          * new_roots. So let's do quota accounting.
2198          */
2199         ret = btrfs_qgroup_account_extents(trans, fs_info);
2200         if (ret < 0) {
2201                 mutex_unlock(&fs_info->tree_log_mutex);
2202                 mutex_unlock(&fs_info->reloc_mutex);
2203                 goto scrub_continue;
2204         }
2205
2206         ret = commit_cowonly_roots(trans, fs_info);
2207         if (ret) {
2208                 mutex_unlock(&fs_info->tree_log_mutex);
2209                 mutex_unlock(&fs_info->reloc_mutex);
2210                 goto scrub_continue;
2211         }
2212
2213         /*
2214          * The tasks which save the space cache and inode cache may also
2215          * update ->aborted, check it.
2216          */
2217         if (unlikely(READ_ONCE(cur_trans->aborted))) {
2218                 ret = cur_trans->aborted;
2219                 mutex_unlock(&fs_info->tree_log_mutex);
2220                 mutex_unlock(&fs_info->reloc_mutex);
2221                 goto scrub_continue;
2222         }
2223
2224         btrfs_prepare_extent_commit(fs_info);
2225
2226         cur_trans = fs_info->running_transaction;
2227
2228         btrfs_set_root_node(&fs_info->tree_root->root_item,
2229                             fs_info->tree_root->node);
2230         list_add_tail(&fs_info->tree_root->dirty_list,
2231                       &cur_trans->switch_commits);
2232
2233         btrfs_set_root_node(&fs_info->chunk_root->root_item,
2234                             fs_info->chunk_root->node);
2235         list_add_tail(&fs_info->chunk_root->dirty_list,
2236                       &cur_trans->switch_commits);
2237
2238         switch_commit_roots(cur_trans, fs_info);
2239
2240         ASSERT(list_empty(&cur_trans->dirty_bgs));
2241         ASSERT(list_empty(&cur_trans->io_bgs));
2242         update_super_roots(fs_info);
2243
2244         btrfs_set_super_log_root(fs_info->super_copy, 0);
2245         btrfs_set_super_log_root_level(fs_info->super_copy, 0);
2246         memcpy(fs_info->super_for_commit, fs_info->super_copy,
2247                sizeof(*fs_info->super_copy));
2248
2249         btrfs_update_commit_device_size(fs_info);
2250         btrfs_update_commit_device_bytes_used(fs_info, cur_trans);
2251
2252         clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
2253         clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
2254
2255         btrfs_trans_release_chunk_metadata(trans);
2256
2257         spin_lock(&fs_info->trans_lock);
2258         cur_trans->state = TRANS_STATE_UNBLOCKED;
2259         fs_info->running_transaction = NULL;
2260         spin_unlock(&fs_info->trans_lock);
2261         mutex_unlock(&fs_info->reloc_mutex);
2262
2263         wake_up(&fs_info->transaction_wait);
2264
2265         ret = btrfs_write_and_wait_transaction(trans, fs_info);
2266         if (ret) {
2267                 btrfs_handle_fs_error(fs_info, ret,
2268                                       "Error while writing out transaction");
2269                 mutex_unlock(&fs_info->tree_log_mutex);
2270                 goto scrub_continue;
2271         }
2272
2273         ret = write_all_supers(fs_info, 0);
2274         /*
2275          * the super is written, we can safely allow the tree-loggers
2276          * to go about their business
2277          */
2278         mutex_unlock(&fs_info->tree_log_mutex);
2279         if (ret)
2280                 goto scrub_continue;
2281
2282         btrfs_finish_extent_commit(trans, fs_info);
2283
2284         if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &cur_trans->flags))
2285                 btrfs_clear_space_info_full(fs_info);
2286
2287         fs_info->last_trans_committed = cur_trans->transid;
2288         /*
2289          * We needn't acquire the lock here because there is no other task
2290          * which can change it.
2291          */
2292         cur_trans->state = TRANS_STATE_COMPLETED;
2293         wake_up(&cur_trans->commit_wait);
2294
2295         spin_lock(&fs_info->trans_lock);
2296         list_del_init(&cur_trans->list);
2297         spin_unlock(&fs_info->trans_lock);
2298
2299         btrfs_put_transaction(cur_trans);
2300         btrfs_put_transaction(cur_trans);
2301
2302         if (trans->type & __TRANS_FREEZABLE)
2303                 sb_end_intwrite(fs_info->sb);
2304
2305         trace_btrfs_transaction_commit(trans->root);
2306
2307         btrfs_scrub_continue(fs_info);
2308
2309         if (current->journal_info == trans)
2310                 current->journal_info = NULL;
2311
2312         kmem_cache_free(btrfs_trans_handle_cachep, trans);
2313
2314         /*
2315          * If fs has been frozen, we can not handle delayed iputs, otherwise
2316          * it'll result in deadlock about SB_FREEZE_FS.
2317          */
2318         if (current != fs_info->transaction_kthread &&
2319             current != fs_info->cleaner_kthread &&
2320             !test_bit(BTRFS_FS_FROZEN, &fs_info->flags))
2321                 btrfs_run_delayed_iputs(fs_info);
2322
2323         return ret;
2324
2325 scrub_continue:
2326         btrfs_scrub_continue(fs_info);
2327 cleanup_transaction:
2328         btrfs_trans_release_metadata(trans);
2329         btrfs_trans_release_chunk_metadata(trans);
2330         trans->block_rsv = NULL;
2331         btrfs_warn(fs_info, "Skipping commit of aborted transaction.");
2332         if (current->journal_info == trans)
2333                 current->journal_info = NULL;
2334         cleanup_transaction(trans, trans->root, ret);
2335
2336         return ret;
2337 }
2338
2339 /*
2340  * return < 0 if error
2341  * 0 if there are no more dead_roots at the time of call
2342  * 1 there are more to be processed, call me again
2343  *
2344  * The return value indicates there are certainly more snapshots to delete, but
2345  * if there comes a new one during processing, it may return 0. We don't mind,
2346  * because btrfs_commit_super will poke cleaner thread and it will process it a
2347  * few seconds later.
2348  */
2349 int btrfs_clean_one_deleted_snapshot(struct btrfs_root *root)
2350 {
2351         int ret;
2352         struct btrfs_fs_info *fs_info = root->fs_info;
2353
2354         spin_lock(&fs_info->trans_lock);
2355         if (list_empty(&fs_info->dead_roots)) {
2356                 spin_unlock(&fs_info->trans_lock);
2357                 return 0;
2358         }
2359         root = list_first_entry(&fs_info->dead_roots,
2360                         struct btrfs_root, root_list);
2361         list_del_init(&root->root_list);
2362         spin_unlock(&fs_info->trans_lock);
2363
2364         btrfs_debug(fs_info, "cleaner removing %llu", root->objectid);
2365
2366         btrfs_kill_all_delayed_nodes(root);
2367
2368         if (btrfs_header_backref_rev(root->node) <
2369                         BTRFS_MIXED_BACKREF_REV)
2370                 ret = btrfs_drop_snapshot(root, NULL, 0, 0);
2371         else
2372                 ret = btrfs_drop_snapshot(root, NULL, 1, 0);
2373
2374         return (ret < 0) ? 0 : 1;
2375 }
2376
2377 void btrfs_apply_pending_changes(struct btrfs_fs_info *fs_info)
2378 {
2379         unsigned long prev;
2380         unsigned long bit;
2381
2382         prev = xchg(&fs_info->pending_changes, 0);
2383         if (!prev)
2384                 return;
2385
2386         bit = 1 << BTRFS_PENDING_SET_INODE_MAP_CACHE;
2387         if (prev & bit)
2388                 btrfs_set_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2389         prev &= ~bit;
2390
2391         bit = 1 << BTRFS_PENDING_CLEAR_INODE_MAP_CACHE;
2392         if (prev & bit)
2393                 btrfs_clear_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2394         prev &= ~bit;
2395
2396         bit = 1 << BTRFS_PENDING_COMMIT;
2397         if (prev & bit)
2398                 btrfs_debug(fs_info, "pending commit done");
2399         prev &= ~bit;
2400
2401         if (prev)
2402                 btrfs_warn(fs_info,
2403                         "unknown pending changes left 0x%lx, ignoring", prev);
2404 }