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