fs/btrfs/transaction.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/fs.h>
  20 #include <linux/slab.h>
  21 #include <linux/sched.h>
  22 #include <linux/writeback.h>
  23 #include <linux/pagemap.h>
  24 #include <linux/blkdev.h>
  25 #include <linux/uuid.h>
  26 #include "ctree.h"
  27 #include "disk-io.h"
  28 #include "transaction.h"
  29 #include "locking.h"
  30 #include "tree-log.h"
  31 #include "inode-map.h"
  32 #include "volumes.h"
  33 #include "dev-replace.h"
  34 #include "qgroup.h"
  35
  36 #define BTRFS_ROOT_TRANS_TAG 0
  37
  38 static const unsigned int btrfs_blocked_trans_types[TRANS_STATE_MAX] = {
  39         [TRANS_STATE_RUNNING]           = 0U,
  40         [TRANS_STATE_BLOCKED]           = (__TRANS_USERSPACE |
  41                                            __TRANS_START),
  42         [TRANS_STATE_COMMIT_START]      = (__TRANS_USERSPACE |
  43                                            __TRANS_START |
  44                                            __TRANS_ATTACH),
  45         [TRANS_STATE_COMMIT_DOING]      = (__TRANS_USERSPACE |
  46                                            __TRANS_START |
  47                                            __TRANS_ATTACH |
  48                                            __TRANS_JOIN),
  49         [TRANS_STATE_UNBLOCKED]         = (__TRANS_USERSPACE |
  50                                            __TRANS_START |
  51                                            __TRANS_ATTACH |
  52                                            __TRANS_JOIN |
  53                                            __TRANS_JOIN_NOLOCK),
  54         [TRANS_STATE_COMPLETED]         = (__TRANS_USERSPACE |
  55                                            __TRANS_START |
  56                                            __TRANS_ATTACH |
  57                                            __TRANS_JOIN |
  58                                            __TRANS_JOIN_NOLOCK),
  59 };
  60
  61 void btrfs_put_transaction(struct btrfs_transaction *transaction)
  62 {
  63         WARN_ON(atomic_read(&transaction->use_count) == 0);
  64         if (atomic_dec_and_test(&transaction->use_count)) {
  65                 BUG_ON(!list_empty(&transaction->list));
  66                 WARN_ON(!RB_EMPTY_ROOT(&transaction->delayed_refs.href_root));
  67                 if (transaction->delayed_refs.pending_csums)
  68                         printk(KERN_ERR "pending csums is %llu\n",
  69                                transaction->delayed_refs.pending_csums);
  70                 while (!list_empty(&transaction->pending_chunks)) {
  71                         struct extent_map *em;
  72
  73                         em = list_first_entry(&transaction->pending_chunks,
  74                                               struct extent_map, list);
  75                         list_del_init(&em->list);
  76                         free_extent_map(em);
  77                 }
  78                 /*
  79                  * If any block groups are found in ->deleted_bgs then it's
  80                  * because the transaction was aborted and a commit did not
  81                  * happen (things failed before writing the new superblock
  82                  * and calling btrfs_finish_extent_commit()), so we can not
  83                  * discard the physical locations of the block groups.
  84                  */
  85                 while (!list_empty(&transaction->deleted_bgs)) {
  86                         struct btrfs_block_group_cache *cache;
  87
  88                         cache = list_first_entry(&transaction->deleted_bgs,
  89                                                  struct btrfs_block_group_cache,
  90                                                  bg_list);
  91                         list_del_init(&cache->bg_list);
  92                         btrfs_put_block_group_trimming(cache);
  93                         btrfs_put_block_group(cache);
  94                 }
  95                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  96         }
  97 }
  98
  99 static void clear_btree_io_tree(struct extent_io_tree *tree)
 100 {
 101         spin_lock(&tree->lock);
 102         /*
 103          * Do a single barrier for the waitqueue_active check here, the state
 104          * of the waitqueue should not change once clear_btree_io_tree is
 105          * called.
 106          */
 107         smp_mb();
 108         while (!RB_EMPTY_ROOT(&tree->state)) {
 109                 struct rb_node *node;
 110                 struct extent_state *state;
 111
 112                 node = rb_first(&tree->state);
 113                 state = rb_entry(node, struct extent_state, rb_node);
 114                 rb_erase(&state->rb_node, &tree->state);
 115                 RB_CLEAR_NODE(&state->rb_node);
 116                 /*
 117                  * btree io trees aren't supposed to have tasks waiting for
 118                  * changes in the flags of extent states ever.
 119                  */
 120                 ASSERT(!waitqueue_active(&state->wq));
 121                 free_extent_state(state);
 122
 123                 cond_resched_lock(&tree->lock);
 124         }
 125         spin_unlock(&tree->lock);
 126 }
 127
 128 static noinline void switch_commit_roots(struct btrfs_transaction *trans,
 129                                          struct btrfs_fs_info *fs_info)
 130 {
 131         struct btrfs_root *root, *tmp;
 132
 133         down_write(&fs_info->commit_root_sem);
 134         list_for_each_entry_safe(root, tmp, &trans->switch_commits,
 135                                  dirty_list) {
 136                 list_del_init(&root->dirty_list);
 137                 free_extent_buffer(root->commit_root);
 138                 root->commit_root = btrfs_root_node(root);
 139                 if (is_fstree(root->objectid))
 140                         btrfs_unpin_free_ino(root);
 141                 clear_btree_io_tree(&root->dirty_log_pages);
 142         }
 143
 144         /* We can free old roots now. */
 145         spin_lock(&trans->dropped_roots_lock);
 146         while (!list_empty(&trans->dropped_roots)) {
 147                 root = list_first_entry(&trans->dropped_roots,
 148                                         struct btrfs_root, root_list);
 149                 list_del_init(&root->root_list);
 150                 spin_unlock(&trans->dropped_roots_lock);
 151                 btrfs_drop_and_free_fs_root(fs_info, root);
 152                 spin_lock(&trans->dropped_roots_lock);
 153         }
 154         spin_unlock(&trans->dropped_roots_lock);
 155         up_write(&fs_info->commit_root_sem);
 156 }
 157
 158 static inline void extwriter_counter_inc(struct btrfs_transaction *trans,
 159                                          unsigned int type)
 160 {
 161         if (type & TRANS_EXTWRITERS)
 162                 atomic_inc(&trans->num_extwriters);
 163 }
 164
 165 static inline void extwriter_counter_dec(struct btrfs_transaction *trans,
 166                                          unsigned int type)
 167 {
 168         if (type & TRANS_EXTWRITERS)
 169                 atomic_dec(&trans->num_extwriters);
 170 }
 171
 172 static inline void extwriter_counter_init(struct btrfs_transaction *trans,
 173                                           unsigned int type)
 174 {
 175         atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0));
 176 }
 177
 178 static inline int extwriter_counter_read(struct btrfs_transaction *trans)
 179 {
 180         return atomic_read(&trans->num_extwriters);
 181 }
 182
 183 /*
 184  * either allocate a new transaction or hop into the existing one
 185  */
 186 static noinline int join_transaction(struct btrfs_root *root, unsigned int type)
 187 {
 188         struct btrfs_transaction *cur_trans;
 189         struct btrfs_fs_info *fs_info = root->fs_info;
 190
 191         spin_lock(&fs_info->trans_lock);
 192 loop:
 193         /* The file system has been taken offline. No new transactions. */
 194         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
 195                 spin_unlock(&fs_info->trans_lock);
 196                 return -EROFS;
 197         }
 198
 199         cur_trans = fs_info->running_transaction;
 200         if (cur_trans) {
 201                 if (cur_trans->aborted) {
 202                         spin_unlock(&fs_info->trans_lock);
 203                         return cur_trans->aborted;
 204                 }
 205                 if (btrfs_blocked_trans_types[cur_trans->state] & type) {
 206                         spin_unlock(&fs_info->trans_lock);
 207                         return -EBUSY;
 208                 }
 209                 atomic_inc(&cur_trans->use_count);
 210                 atomic_inc(&cur_trans->num_writers);
 211                 extwriter_counter_inc(cur_trans, type);
 212                 spin_unlock(&fs_info->trans_lock);
 213                 return 0;
 214         }
 215         spin_unlock(&fs_info->trans_lock);
 216
 217         /*
 218          * If we are ATTACH, we just want to catch the current transaction,
 219          * and commit it. If there is no transaction, just return ENOENT.
 220          */
 221         if (type == TRANS_ATTACH)
 222                 return -ENOENT;
 223
 224         /*
 225          * JOIN_NOLOCK only happens during the transaction commit, so
 226          * it is impossible that ->running_transaction is NULL
 227          */
 228         BUG_ON(type == TRANS_JOIN_NOLOCK);
 229
 230         cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
 231         if (!cur_trans)
 232                 return -ENOMEM;
 233
 234         spin_lock(&fs_info->trans_lock);
 235         if (fs_info->running_transaction) {
 236                 /*
 237                  * someone started a transaction after we unlocked.  Make sure
 238                  * to redo the checks above
 239                  */
 240                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 241                 goto loop;
 242         } else if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
 243                 spin_unlock(&fs_info->trans_lock);
 244                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 245                 return -EROFS;
 246         }
 247
 248         atomic_set(&cur_trans->num_writers, 1);
 249         extwriter_counter_init(cur_trans, type);
 250         init_waitqueue_head(&cur_trans->writer_wait);
 251         init_waitqueue_head(&cur_trans->commit_wait);
 252         init_waitqueue_head(&cur_trans->pending_wait);
 253         cur_trans->state = TRANS_STATE_RUNNING;
 254         /*
 255          * One for this trans handle, one so it will live on until we
 256          * commit the transaction.
 257          */
 258         atomic_set(&cur_trans->use_count, 2);
 259         atomic_set(&cur_trans->pending_ordered, 0);
 260         cur_trans->flags = 0;
 261         cur_trans->start_time = get_seconds();
 262
 263         memset(&cur_trans->delayed_refs, 0, sizeof(cur_trans->delayed_refs));
 264
 265         cur_trans->delayed_refs.href_root = RB_ROOT;
 266         cur_trans->delayed_refs.dirty_extent_root = RB_ROOT;
 267         atomic_set(&cur_trans->delayed_refs.num_entries, 0);
 268
 269         /*
 270          * although the tree mod log is per file system and not per transaction,
 271          * the log must never go across transaction boundaries.
 272          */
 273         smp_mb();
 274         if (!list_empty(&fs_info->tree_mod_seq_list))
 275                 WARN(1, KERN_ERR "BTRFS: tree_mod_seq_list not empty when "
 276                         "creating a fresh transaction\n");
 277         if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
 278                 WARN(1, KERN_ERR "BTRFS: tree_mod_log rb tree not empty when "
 279                         "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->i_mapping);
 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 {
 316         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
 317             root->last_trans < trans->transid) {
 318                 WARN_ON(root == root->fs_info->extent_root);
 319                 WARN_ON(root->commit_root != root->node);
 320
 321                 /*
 322                  * see below for IN_TRANS_SETUP usage rules
 323                  * we have the reloc mutex held now, so there
 324                  * is only one writer in this function
 325                  */
 326                 set_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
 327
 328                 /* make sure readers find IN_TRANS_SETUP before
 329                  * they find our root->last_trans update
 330                  */
 331                 smp_wmb();
 332
 333                 spin_lock(&root->fs_info->fs_roots_radix_lock);
 334                 if (root->last_trans == trans->transid) {
 335                         spin_unlock(&root->fs_info->fs_roots_radix_lock);
 336                         return 0;
 337                 }
 338                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 339                            (unsigned long)root->root_key.objectid,
 340                            BTRFS_ROOT_TRANS_TAG);
 341                 spin_unlock(&root->fs_info->fs_roots_radix_lock);
 342                 root->last_trans = trans->transid;
 343
 344                 /* this is pretty tricky.  We don't want to
 345                  * take the relocation lock in btrfs_record_root_in_trans
 346                  * unless we're really doing the first setup for this root in
 347                  * this transaction.
 348                  *
 349                  * Normally we'd use root->last_trans as a flag to decide
 350                  * if we want to take the expensive mutex.
 351                  *
 352                  * But, we have to set root->last_trans before we
 353                  * init the relocation root, otherwise, we trip over warnings
 354                  * in ctree.c.  The solution used here is to flag ourselves
 355                  * with root IN_TRANS_SETUP.  When this is 1, we're still
 356                  * fixing up the reloc trees and everyone must wait.
 357                  *
 358                  * When this is zero, they can trust root->last_trans and fly
 359                  * through btrfs_record_root_in_trans without having to take the
 360                  * lock.  smp_wmb() makes sure that all the writes above are
 361                  * done before we pop in the zero below
 362                  */
 363                 btrfs_init_reloc_root(trans, root);
 364                 smp_mb__before_atomic();
 365                 clear_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
 366         }
 367         return 0;
 368 }
 369
 370
 371 void btrfs_add_dropped_root(struct btrfs_trans_handle *trans,
 372                             struct btrfs_root *root)
 373 {
 374         struct btrfs_transaction *cur_trans = trans->transaction;
 375
 376         /* Add ourselves to the transaction dropped list */
 377         spin_lock(&cur_trans->dropped_roots_lock);
 378         list_add_tail(&root->root_list, &cur_trans->dropped_roots);
 379         spin_unlock(&cur_trans->dropped_roots_lock);
 380
 381         /* Make sure we don't try to update the root at commit time */
 382         spin_lock(&root->fs_info->fs_roots_radix_lock);
 383         radix_tree_tag_clear(&root->fs_info->fs_roots_radix,
 384                              (unsigned long)root->root_key.objectid,
 385                              BTRFS_ROOT_TRANS_TAG);
 386         spin_unlock(&root->fs_info->fs_roots_radix_lock);
 387 }
 388
 389 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 390                                struct btrfs_root *root)
 391 {
 392         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
 393                 return 0;
 394
 395         /*
 396          * see record_root_in_trans for comments about IN_TRANS_SETUP usage
 397          * and barriers
 398          */
 399         smp_rmb();
 400         if (root->last_trans == trans->transid &&
 401             !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state))
 402                 return 0;
 403
 404         mutex_lock(&root->fs_info->reloc_mutex);
 405         record_root_in_trans(trans, root);
 406         mutex_unlock(&root->fs_info->reloc_mutex);
 407
 408         return 0;
 409 }
 410
 411 static inline int is_transaction_blocked(struct btrfs_transaction *trans)
 412 {
 413         return (trans->state >= TRANS_STATE_BLOCKED &&
 414                 trans->state < TRANS_STATE_UNBLOCKED &&
 415                 !trans->aborted);
 416 }
 417
 418 /* wait for commit against the current transaction to become unblocked
 419  * when this is done, it is safe to start a new transaction, but the current
 420  * transaction might not be fully on disk.
 421  */
 422 static void wait_current_trans(struct btrfs_root *root)
 423 {
 424         struct btrfs_transaction *cur_trans;
 425
 426         spin_lock(&root->fs_info->trans_lock);
 427         cur_trans = root->fs_info->running_transaction;
 428         if (cur_trans && is_transaction_blocked(cur_trans)) {
 429                 atomic_inc(&cur_trans->use_count);
 430                 spin_unlock(&root->fs_info->trans_lock);
 431
 432                 wait_event(root->fs_info->transaction_wait,
 433                            cur_trans->state >= TRANS_STATE_UNBLOCKED ||
 434                            cur_trans->aborted);
 435                 btrfs_put_transaction(cur_trans);
 436         } else {
 437                 spin_unlock(&root->fs_info->trans_lock);
 438         }
 439 }
 440
 441 static int may_wait_transaction(struct btrfs_root *root, int type)
 442 {
 443         if (root->fs_info->log_root_recovering)
 444                 return 0;
 445
 446         if (type == TRANS_USERSPACE)
 447                 return 1;
 448
 449         if (type == TRANS_START &&
 450             !atomic_read(&root->fs_info->open_ioctl_trans))
 451                 return 1;
 452
 453         return 0;
 454 }
 455
 456 static inline bool need_reserve_reloc_root(struct btrfs_root *root)
 457 {
 458         if (!root->fs_info->reloc_ctl ||
 459             !test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
 460             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 461             root->reloc_root)
 462                 return false;
 463
 464         return true;
 465 }
 466
 467 static struct btrfs_trans_handle *
 468 start_transaction(struct btrfs_root *root, unsigned int num_items,
 469                   unsigned int type, enum btrfs_reserve_flush_enum flush)
 470 {
 471         struct btrfs_trans_handle *h;
 472         struct btrfs_transaction *cur_trans;
 473         u64 num_bytes = 0;
 474         u64 qgroup_reserved = 0;
 475         bool reloc_reserved = false;
 476         int ret;
 477
 478         /* Send isn't supposed to start transactions. */
 479         ASSERT(current->journal_info != BTRFS_SEND_TRANS_STUB);
 480
 481         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
 482                 return ERR_PTR(-EROFS);
 483
 484         if (current->journal_info) {
 485                 WARN_ON(type & TRANS_EXTWRITERS);
 486                 h = current->journal_info;
 487                 h->use_count++;
 488                 WARN_ON(h->use_count > 2);
 489                 h->orig_rsv = h->block_rsv;
 490                 h->block_rsv = NULL;
 491                 goto got_it;
 492         }
 493
 494         /*
 495          * Do the reservation before we join the transaction so we can do all
 496          * the appropriate flushing if need be.
 497          */
 498         if (num_items > 0 && root != root->fs_info->chunk_root) {
 499                 qgroup_reserved = num_items * root->nodesize;
 500                 ret = btrfs_qgroup_reserve_meta(root, qgroup_reserved);
 501                 if (ret)
 502                         return ERR_PTR(ret);
 503
 504                 num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
 505                 /*
 506                  * Do the reservation for the relocation root creation
 507                  */
 508                 if (need_reserve_reloc_root(root)) {
 509                         num_bytes += root->nodesize;
 510                         reloc_reserved = true;
 511                 }
 512
 513                 ret = btrfs_block_rsv_add(root,
 514                                           &root->fs_info->trans_block_rsv,
 515                                           num_bytes, flush);
 516                 if (ret)
 517                         goto reserve_fail;
 518         }
 519 again:
 520         h = kmem_cache_zalloc(btrfs_trans_handle_cachep, GFP_NOFS);
 521         if (!h) {
 522                 ret = -ENOMEM;
 523                 goto alloc_fail;
 524         }
 525
 526         /*
 527          * If we are JOIN_NOLOCK we're already committing a transaction and
 528          * waiting on this guy, so we don't need to do the sb_start_intwrite
 529          * because we're already holding a ref.  We need this because we could
 530          * have raced in and did an fsync() on a file which can kick a commit
 531          * and then we deadlock with somebody doing a freeze.
 532          *
 533          * If we are ATTACH, it means we just want to catch the current
 534          * transaction and commit it, so we needn't do sb_start_intwrite().
 535          */
 536         if (type & __TRANS_FREEZABLE)
 537                 sb_start_intwrite(root->fs_info->sb);
 538
 539         if (may_wait_transaction(root, type))
 540                 wait_current_trans(root);
 541
 542         do {
 543                 ret = join_transaction(root, type);
 544                 if (ret == -EBUSY) {
 545                         wait_current_trans(root);
 546                         if (unlikely(type == TRANS_ATTACH))
 547                                 ret = -ENOENT;
 548                 }
 549         } while (ret == -EBUSY);
 550
 551         if (ret < 0) {
 552                 /* We must get the transaction if we are JOIN_NOLOCK. */
 553                 BUG_ON(type == TRANS_JOIN_NOLOCK);
 554                 goto join_fail;
 555         }
 556
 557         cur_trans = root->fs_info->running_transaction;
 558
 559         h->transid = cur_trans->transid;
 560         h->transaction = cur_trans;
 561         h->root = root;
 562         h->use_count = 1;
 563
 564         h->type = type;
 565         h->can_flush_pending_bgs = true;
 566         INIT_LIST_HEAD(&h->qgroup_ref_list);
 567         INIT_LIST_HEAD(&h->new_bgs);
 568
 569         smp_mb();
 570         if (cur_trans->state >= TRANS_STATE_BLOCKED &&
 571             may_wait_transaction(root, type)) {
 572                 current->journal_info = h;
 573                 btrfs_commit_transaction(h, root);
 574                 goto again;
 575         }
 576
 577         if (num_bytes) {
 578                 trace_btrfs_space_reservation(root->fs_info, "transaction",
 579                                               h->transid, num_bytes, 1);
 580                 h->block_rsv = &root->fs_info->trans_block_rsv;
 581                 h->bytes_reserved = num_bytes;
 582                 h->reloc_reserved = reloc_reserved;
 583         }
 584
 585 got_it:
 586         btrfs_record_root_in_trans(h, root);
 587
 588         if (!current->journal_info && type != TRANS_USERSPACE)
 589                 current->journal_info = h;
 590         return h;
 591
 592 join_fail:
 593         if (type & __TRANS_FREEZABLE)
 594                 sb_end_intwrite(root->fs_info->sb);
 595         kmem_cache_free(btrfs_trans_handle_cachep, h);
 596 alloc_fail:
 597         if (num_bytes)
 598                 btrfs_block_rsv_release(root, &root->fs_info->trans_block_rsv,
 599                                         num_bytes);
 600 reserve_fail:
 601         btrfs_qgroup_free_meta(root, qgroup_reserved);
 602         return ERR_PTR(ret);
 603 }
 604
 605 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 606                                                    unsigned int num_items)
 607 {
 608         return start_transaction(root, num_items, TRANS_START,
 609                                  BTRFS_RESERVE_FLUSH_ALL);
 610 }
 611 struct btrfs_trans_handle *btrfs_start_transaction_fallback_global_rsv(
 612                                         struct btrfs_root *root,
 613                                         unsigned int num_items,
 614                                         int min_factor)
 615 {
 616         struct btrfs_trans_handle *trans;
 617         u64 num_bytes;
 618         int ret;
 619
 620         trans = btrfs_start_transaction(root, num_items);
 621         if (!IS_ERR(trans) || PTR_ERR(trans) != -ENOSPC)
 622                 return trans;
 623
 624         trans = btrfs_start_transaction(root, 0);
 625         if (IS_ERR(trans))
 626                 return trans;
 627
 628         num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
 629         ret = btrfs_cond_migrate_bytes(root->fs_info,
 630                                        &root->fs_info->trans_block_rsv,
 631                                        num_bytes,
 632                                        min_factor);
 633         if (ret) {
 634                 btrfs_end_transaction(trans, root);
 635                 return ERR_PTR(ret);
 636         }
 637
 638         trans->block_rsv = &root->fs_info->trans_block_rsv;
 639         trans->bytes_reserved = num_bytes;
 640
 641         return trans;
 642 }
 643
 644 struct btrfs_trans_handle *btrfs_start_transaction_lflush(
 645                                         struct btrfs_root *root,
 646                                         unsigned int num_items)
 647 {
 648         return start_transaction(root, num_items, TRANS_START,
 649                                  BTRFS_RESERVE_FLUSH_LIMIT);
 650 }
 651
 652 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 653 {
 654         return start_transaction(root, 0, TRANS_JOIN,
 655                                  BTRFS_RESERVE_NO_FLUSH);
 656 }
 657
 658 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 659 {
 660         return start_transaction(root, 0, TRANS_JOIN_NOLOCK,
 661                                  BTRFS_RESERVE_NO_FLUSH);
 662 }
 663
 664 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 665 {
 666         return start_transaction(root, 0, TRANS_USERSPACE,
 667                                  BTRFS_RESERVE_NO_FLUSH);
 668 }
 669
 670 /*
 671  * btrfs_attach_transaction() - catch the running transaction
 672  *
 673  * It is used when we want to commit the current the transaction, but
 674  * don't want to start a new one.
 675  *
 676  * Note: If this function return -ENOENT, it just means there is no
 677  * running transaction. But it is possible that the inactive transaction
 678  * is still in the memory, not fully on disk. If you hope there is no
 679  * inactive transaction in the fs when -ENOENT is returned, you should
 680  * invoke
 681  *     btrfs_attach_transaction_barrier()
 682  */
 683 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
 684 {
 685         return start_transaction(root, 0, TRANS_ATTACH,
 686                                  BTRFS_RESERVE_NO_FLUSH);
 687 }
 688
 689 /*
 690  * btrfs_attach_transaction_barrier() - catch the running transaction
 691  *
 692  * It is similar to the above function, the differentia is this one
 693  * will wait for all the inactive transactions until they fully
 694  * complete.
 695  */
 696 struct btrfs_trans_handle *
 697 btrfs_attach_transaction_barrier(struct btrfs_root *root)
 698 {
 699         struct btrfs_trans_handle *trans;
 700
 701         trans = start_transaction(root, 0, TRANS_ATTACH,
 702                                   BTRFS_RESERVE_NO_FLUSH);
 703         if (IS_ERR(trans) && PTR_ERR(trans) == -ENOENT)
 704                 btrfs_wait_for_commit(root, 0);
 705
 706         return trans;
 707 }
 708
 709 /* wait for a transaction commit to be fully complete */
 710 static noinline void wait_for_commit(struct btrfs_root *root,
 711                                     struct btrfs_transaction *commit)
 712 {
 713         wait_event(commit->commit_wait, commit->state == TRANS_STATE_COMPLETED);
 714 }
 715
 716 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 717 {
 718         struct btrfs_transaction *cur_trans = NULL, *t;
 719         int ret = 0;
 720
 721         if (transid) {
 722                 if (transid <= root->fs_info->last_trans_committed)
 723                         goto out;
 724
 725                 /* find specified transaction */
 726                 spin_lock(&root->fs_info->trans_lock);
 727                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 728                         if (t->transid == transid) {
 729                                 cur_trans = t;
 730                                 atomic_inc(&cur_trans->use_count);
 731                                 ret = 0;
 732                                 break;
 733                         }
 734                         if (t->transid > transid) {
 735                                 ret = 0;
 736                                 break;
 737                         }
 738                 }
 739                 spin_unlock(&root->fs_info->trans_lock);
 740
 741                 /*
 742                  * The specified transaction doesn't exist, or we
 743                  * raced with btrfs_commit_transaction
 744                  */
 745                 if (!cur_trans) {
 746                         if (transid > root->fs_info->last_trans_committed)
 747                                 ret = -EINVAL;
 748                         goto out;
 749                 }
 750         } else {
 751                 /* find newest transaction that is committing | committed */
 752                 spin_lock(&root->fs_info->trans_lock);
 753                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 754                                             list) {
 755                         if (t->state >= TRANS_STATE_COMMIT_START) {
 756                                 if (t->state == TRANS_STATE_COMPLETED)
 757                                         break;
 758                                 cur_trans = t;
 759                                 atomic_inc(&cur_trans->use_count);
 760                                 break;
 761                         }
 762                 }
 763                 spin_unlock(&root->fs_info->trans_lock);
 764                 if (!cur_trans)
 765                         goto out;  /* nothing committing|committed */
 766         }
 767
 768         wait_for_commit(root, cur_trans);
 769         btrfs_put_transaction(cur_trans);
 770 out:
 771         return ret;
 772 }
 773
 774 void btrfs_throttle(struct btrfs_root *root)
 775 {
 776         if (!atomic_read(&root->fs_info->open_ioctl_trans))
 777                 wait_current_trans(root);
 778 }
 779
 780 static int should_end_transaction(struct btrfs_trans_handle *trans,
 781                                   struct btrfs_root *root)
 782 {
 783         if (root->fs_info->global_block_rsv.space_info->full &&
 784             btrfs_check_space_for_delayed_refs(trans, root))
 785                 return 1;
 786
 787         return !!btrfs_block_rsv_check(root, &root->fs_info->global_block_rsv, 5);
 788 }
 789
 790 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 791                                  struct btrfs_root *root)
 792 {
 793         struct btrfs_transaction *cur_trans = trans->transaction;
 794         int updates;
 795         int err;
 796
 797         smp_mb();
 798         if (cur_trans->state >= TRANS_STATE_BLOCKED ||
 799             cur_trans->delayed_refs.flushing)
 800                 return 1;
 801
 802         updates = trans->delayed_ref_updates;
 803         trans->delayed_ref_updates = 0;
 804         if (updates) {
 805                 err = btrfs_run_delayed_refs(trans, root, updates * 2);
 806                 if (err) /* Error code will also eval true */
 807                         return err;
 808         }
 809
 810         return should_end_transaction(trans, root);
 811 }
 812
 813 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 814                           struct btrfs_root *root, int throttle)
 815 {
 816         struct btrfs_transaction *cur_trans = trans->transaction;
 817         struct btrfs_fs_info *info = root->fs_info;
 818         unsigned long cur = trans->delayed_ref_updates;
 819         int lock = (trans->type != TRANS_JOIN_NOLOCK);
 820         int err = 0;
 821         int must_run_delayed_refs = 0;
 822
 823         if (trans->use_count > 1) {
 824                 trans->use_count--;
 825                 trans->block_rsv = trans->orig_rsv;
 826                 return 0;
 827         }
 828
 829         btrfs_trans_release_metadata(trans, root);
 830         trans->block_rsv = NULL;
 831
 832         if (!list_empty(&trans->new_bgs))
 833                 btrfs_create_pending_block_groups(trans, root);
 834
 835         trans->delayed_ref_updates = 0;
 836         if (!trans->sync) {
 837                 must_run_delayed_refs =
 838                         btrfs_should_throttle_delayed_refs(trans, root);
 839                 cur = max_t(unsigned long, cur, 32);
 840
 841                 /*
 842                  * don't make the caller wait if they are from a NOLOCK
 843                  * or ATTACH transaction, it will deadlock with commit
 844                  */
 845                 if (must_run_delayed_refs == 1 &&
 846                     (trans->type & (__TRANS_JOIN_NOLOCK | __TRANS_ATTACH)))
 847                         must_run_delayed_refs = 2;
 848         }
 849
 850         btrfs_trans_release_metadata(trans, root);
 851         trans->block_rsv = NULL;
 852
 853         if (!list_empty(&trans->new_bgs))
 854                 btrfs_create_pending_block_groups(trans, root);
 855
 856         btrfs_trans_release_chunk_metadata(trans);
 857
 858         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
 859             should_end_transaction(trans, root) &&
 860             ACCESS_ONCE(cur_trans->state) == TRANS_STATE_RUNNING) {
 861                 spin_lock(&info->trans_lock);
 862                 if (cur_trans->state == TRANS_STATE_RUNNING)
 863                         cur_trans->state = TRANS_STATE_BLOCKED;
 864                 spin_unlock(&info->trans_lock);
 865         }
 866
 867         if (lock && ACCESS_ONCE(cur_trans->state) == TRANS_STATE_BLOCKED) {
 868                 if (throttle)
 869                         return btrfs_commit_transaction(trans, root);
 870                 else
 871                         wake_up_process(info->transaction_kthread);
 872         }
 873
 874         if (trans->type & __TRANS_FREEZABLE)
 875                 sb_end_intwrite(root->fs_info->sb);
 876
 877         WARN_ON(cur_trans != info->running_transaction);
 878         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 879         atomic_dec(&cur_trans->num_writers);
 880         extwriter_counter_dec(cur_trans, trans->type);
 881
 882         /*
 883          * Make sure counter is updated before we wake up waiters.
 884          */
 885         smp_mb();
 886         if (waitqueue_active(&cur_trans->writer_wait))
 887                 wake_up(&cur_trans->writer_wait);
 888         btrfs_put_transaction(cur_trans);
 889
 890         if (current->journal_info == trans)
 891                 current->journal_info = NULL;
 892
 893         if (throttle)
 894                 btrfs_run_delayed_iputs(root);
 895
 896         if (trans->aborted ||
 897             test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state)) {
 898                 wake_up_process(info->transaction_kthread);
 899                 err = -EIO;
 900         }
 901         assert_qgroups_uptodate(trans);
 902
 903         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 904         if (must_run_delayed_refs) {
 905                 btrfs_async_run_delayed_refs(root, cur,
 906                                              must_run_delayed_refs == 1);
 907         }
 908         return err;
 909 }
 910
 911 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 912                           struct btrfs_root *root)
 913 {
 914         return __btrfs_end_transaction(trans, root, 0);
 915 }
 916
 917 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 918                                    struct btrfs_root *root)
 919 {
 920         return __btrfs_end_transaction(trans, root, 1);
 921 }
 922
 923 /*
 924  * when btree blocks are allocated, they have some corresponding bits set for
 925  * them in one of two extent_io trees.  This is used to make sure all of
 926  * those extents are sent to disk but does not wait on them
 927  */
 928 int btrfs_write_marked_extents(struct btrfs_root *root,
 929                                struct extent_io_tree *dirty_pages, int mark)
 930 {
 931         int err = 0;
 932         int werr = 0;
 933         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 934         struct extent_state *cached_state = NULL;
 935         u64 start = 0;
 936         u64 end;
 937
 938         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 939                                       mark, &cached_state)) {
 940                 bool wait_writeback = false;
 941
 942                 err = convert_extent_bit(dirty_pages, start, end,
 943                                          EXTENT_NEED_WAIT,
 944                                          mark, &cached_state, GFP_NOFS);
 945                 /*
 946                  * convert_extent_bit can return -ENOMEM, which is most of the
 947                  * time a temporary error. So when it happens, ignore the error
 948                  * and wait for writeback of this range to finish - because we
 949                  * failed to set the bit EXTENT_NEED_WAIT for the range, a call
 950                  * to btrfs_wait_marked_extents() would not know that writeback
 951                  * for this range started and therefore wouldn't wait for it to
 952                  * finish - we don't want to commit a superblock that points to
 953                  * btree nodes/leafs for which writeback hasn't finished yet
 954                  * (and without errors).
 955                  * We cleanup any entries left in the io tree when committing
 956                  * the transaction (through clear_btree_io_tree()).
 957                  */
 958                 if (err == -ENOMEM) {
 959                         err = 0;
 960                         wait_writeback = true;
 961                 }
 962                 if (!err)
 963                         err = filemap_fdatawrite_range(mapping, start, end);
 964                 if (err)
 965                         werr = err;
 966                 else if (wait_writeback)
 967                         werr = filemap_fdatawait_range(mapping, start, end);
 968                 free_extent_state(cached_state);
 969                 cached_state = NULL;
 970                 cond_resched();
 971                 start = end + 1;
 972         }
 973         return werr;
 974 }
 975
 976 /*
 977  * when btree blocks are allocated, they have some corresponding bits set for
 978  * them in one of two extent_io trees.  This is used to make sure all of
 979  * those extents are on disk for transaction or log commit.  We wait
 980  * on all the pages and clear them from the dirty pages state tree
 981  */
 982 int btrfs_wait_marked_extents(struct btrfs_root *root,
 983                               struct extent_io_tree *dirty_pages, int mark)
 984 {
 985         int err = 0;
 986         int werr = 0;
 987         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 988         struct extent_state *cached_state = NULL;
 989         u64 start = 0;
 990         u64 end;
 991         struct btrfs_inode *btree_ino = BTRFS_I(root->fs_info->btree_inode);
 992         bool errors = false;
 993
 994         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 995                                       EXTENT_NEED_WAIT, &cached_state)) {
 996                 /*
 997                  * Ignore -ENOMEM errors returned by clear_extent_bit().
 998                  * When committing the transaction, we'll remove any entries
 999                  * left in the io tree. For a log commit, we don't remove them
1000                  * after committing the log because the tree can be accessed
1001                  * concurrently - we do it only at transaction commit time when
1002                  * it's safe to do it (through clear_btree_io_tree()).
1003                  */
1004                 err = clear_extent_bit(dirty_pages, start, end,
1005                                        EXTENT_NEED_WAIT,
1006                                        0, 0, &cached_state, GFP_NOFS);
1007                 if (err == -ENOMEM)
1008                         err = 0;
1009                 if (!err)
1010                         err = filemap_fdatawait_range(mapping, start, end);
1011                 if (err)
1012                         werr = err;
1013                 free_extent_state(cached_state);
1014                 cached_state = NULL;
1015                 cond_resched();
1016                 start = end + 1;
1017         }
1018         if (err)
1019                 werr = err;
1020
1021         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
1022                 if ((mark & EXTENT_DIRTY) &&
1023                     test_and_clear_bit(BTRFS_INODE_BTREE_LOG1_ERR,
1024                                        &btree_ino->runtime_flags))
1025                         errors = true;
1026
1027                 if ((mark & EXTENT_NEW) &&
1028                     test_and_clear_bit(BTRFS_INODE_BTREE_LOG2_ERR,
1029                                        &btree_ino->runtime_flags))
1030                         errors = true;
1031         } else {
1032                 if (test_and_clear_bit(BTRFS_INODE_BTREE_ERR,
1033                                        &btree_ino->runtime_flags))
1034                         errors = true;
1035         }
1036
1037         if (errors && !werr)
1038                 werr = -EIO;
1039
1040         return werr;
1041 }
1042
1043 /*
1044  * when btree blocks are allocated, they have some corresponding bits set for
1045  * them in one of two extent_io trees.  This is used to make sure all of
1046  * those extents are on disk for transaction or log commit
1047  */
1048 static int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
1049                                 struct extent_io_tree *dirty_pages, int mark)
1050 {
1051         int ret;
1052         int ret2;
1053         struct blk_plug plug;
1054
1055         blk_start_plug(&plug);
1056         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
1057         blk_finish_plug(&plug);
1058         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
1059
1060         if (ret)
1061                 return ret;
1062         if (ret2)
1063                 return ret2;
1064         return 0;
1065 }
1066
1067 static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
1068                                      struct btrfs_root *root)
1069 {
1070         int ret;
1071
1072         ret = btrfs_write_and_wait_marked_extents(root,
1073                                            &trans->transaction->dirty_pages,
1074                                            EXTENT_DIRTY);
1075         clear_btree_io_tree(&trans->transaction->dirty_pages);
1076
1077         return ret;
1078 }
1079
1080 /*
1081  * this is used to update the root pointer in the tree of tree roots.
1082  *
1083  * But, in the case of the extent allocation tree, updating the root
1084  * pointer may allocate blocks which may change the root of the extent
1085  * allocation tree.
1086  *
1087  * So, this loops and repeats and makes sure the cowonly root didn't
1088  * change while the root pointer was being updated in the metadata.
1089  */
1090 static int update_cowonly_root(struct btrfs_trans_handle *trans,
1091                                struct btrfs_root *root)
1092 {
1093         int ret;
1094         u64 old_root_bytenr;
1095         u64 old_root_used;
1096         struct btrfs_root *tree_root = root->fs_info->tree_root;
1097
1098         old_root_used = btrfs_root_used(&root->root_item);
1099
1100         while (1) {
1101                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1102                 if (old_root_bytenr == root->node->start &&
1103                     old_root_used == btrfs_root_used(&root->root_item))
1104                         break;
1105
1106                 btrfs_set_root_node(&root->root_item, root->node);
1107                 ret = btrfs_update_root(trans, tree_root,
1108                                         &root->root_key,
1109                                         &root->root_item);
1110                 if (ret)
1111                         return ret;
1112
1113                 old_root_used = btrfs_root_used(&root->root_item);
1114         }
1115
1116         return 0;
1117 }
1118
1119 /*
1120  * update all the cowonly tree roots on disk
1121  *
1122  * The error handling in this function may not be obvious. Any of the
1123  * failures will cause the file system to go offline. We still need
1124  * to clean up the delayed refs.
1125  */
1126 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
1127                                          struct btrfs_root *root)
1128 {
1129         struct btrfs_fs_info *fs_info = root->fs_info;
1130         struct list_head *dirty_bgs = &trans->transaction->dirty_bgs;
1131         struct list_head *io_bgs = &trans->transaction->io_bgs;
1132         struct list_head *next;
1133         struct extent_buffer *eb;
1134         int ret;
1135
1136         eb = btrfs_lock_root_node(fs_info->tree_root);
1137         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
1138                               0, &eb);
1139         btrfs_tree_unlock(eb);
1140         free_extent_buffer(eb);
1141
1142         if (ret)
1143                 return ret;
1144
1145         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1146         if (ret)
1147                 return ret;
1148
1149         ret = btrfs_run_dev_stats(trans, root->fs_info);
1150         if (ret)
1151                 return ret;
1152         ret = btrfs_run_dev_replace(trans, root->fs_info);
1153         if (ret)
1154                 return ret;
1155         ret = btrfs_run_qgroups(trans, root->fs_info);
1156         if (ret)
1157                 return ret;
1158
1159         ret = btrfs_setup_space_cache(trans, root);
1160         if (ret)
1161                 return ret;
1162
1163         /* run_qgroups might have added some more refs */
1164         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1165         if (ret)
1166                 return ret;
1167 again:
1168         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1169                 next = fs_info->dirty_cowonly_roots.next;
1170                 list_del_init(next);
1171                 root = list_entry(next, struct btrfs_root, dirty_list);
1172                 clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1173
1174                 if (root != fs_info->extent_root)
1175                         list_add_tail(&root->dirty_list,
1176                                       &trans->transaction->switch_commits);
1177                 ret = update_cowonly_root(trans, root);
1178                 if (ret)
1179                         return ret;
1180                 ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1181                 if (ret)
1182                         return ret;
1183         }
1184
1185         while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1186                 ret = btrfs_write_dirty_block_groups(trans, root);
1187                 if (ret)
1188                         return ret;
1189                 ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1190                 if (ret)
1191                         return ret;
1192         }
1193
1194         if (!list_empty(&fs_info->dirty_cowonly_roots))
1195                 goto again;
1196
1197         list_add_tail(&fs_info->extent_root->dirty_list,
1198                       &trans->transaction->switch_commits);
1199         btrfs_after_dev_replace_commit(fs_info);
1200
1201         return 0;
1202 }
1203
1204 /*
1205  * dead roots are old snapshots that need to be deleted.  This allocates
1206  * a dirty root struct and adds it into the list of dead roots that need to
1207  * be deleted
1208  */
1209 void btrfs_add_dead_root(struct btrfs_root *root)
1210 {
1211         spin_lock(&root->fs_info->trans_lock);
1212         if (list_empty(&root->root_list))
1213                 list_add_tail(&root->root_list, &root->fs_info->dead_roots);
1214         spin_unlock(&root->fs_info->trans_lock);
1215 }
1216
1217 /*
1218  * update all the cowonly tree roots on disk
1219  */
1220 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
1221                                     struct btrfs_root *root)
1222 {
1223         struct btrfs_root *gang[8];
1224         struct btrfs_fs_info *fs_info = root->fs_info;
1225         int i;
1226         int ret;
1227         int err = 0;
1228
1229         spin_lock(&fs_info->fs_roots_radix_lock);
1230         while (1) {
1231                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1232                                                  (void **)gang, 0,
1233                                                  ARRAY_SIZE(gang),
1234                                                  BTRFS_ROOT_TRANS_TAG);
1235                 if (ret == 0)
1236                         break;
1237                 for (i = 0; i < ret; i++) {
1238                         root = gang[i];
1239                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
1240                                         (unsigned long)root->root_key.objectid,
1241                                         BTRFS_ROOT_TRANS_TAG);
1242                         spin_unlock(&fs_info->fs_roots_radix_lock);
1243
1244                         btrfs_free_log(trans, root);
1245                         btrfs_update_reloc_root(trans, root);
1246                         btrfs_orphan_commit_root(trans, root);
1247
1248                         btrfs_save_ino_cache(root, trans);
1249
1250                         /* see comments in should_cow_block() */
1251                         clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1252                         smp_mb__after_atomic();
1253
1254                         if (root->commit_root != root->node) {
1255                                 list_add_tail(&root->dirty_list,
1256                                         &trans->transaction->switch_commits);
1257                                 btrfs_set_root_node(&root->root_item,
1258                                                     root->node);
1259                         }
1260
1261                         err = btrfs_update_root(trans, fs_info->tree_root,
1262                                                 &root->root_key,
1263                                                 &root->root_item);
1264                         spin_lock(&fs_info->fs_roots_radix_lock);
1265                         if (err)
1266                                 break;
1267                         btrfs_qgroup_free_meta_all(root);
1268                 }
1269         }
1270         spin_unlock(&fs_info->fs_roots_radix_lock);
1271         return err;
1272 }
1273
1274 /*
1275  * defrag a given btree.
1276  * Every leaf in the btree is read and defragged.
1277  */
1278 int btrfs_defrag_root(struct btrfs_root *root)
1279 {
1280         struct btrfs_fs_info *info = root->fs_info;
1281         struct btrfs_trans_handle *trans;
1282         int ret;
1283
1284         if (test_and_set_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state))
1285                 return 0;
1286
1287         while (1) {
1288                 trans = btrfs_start_transaction(root, 0);
1289                 if (IS_ERR(trans))
1290                         return PTR_ERR(trans);
1291
1292                 ret = btrfs_defrag_leaves(trans, root);
1293
1294                 btrfs_end_transaction(trans, root);
1295                 btrfs_btree_balance_dirty(info->tree_root);
1296                 cond_resched();
1297
1298                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
1299                         break;
1300
1301                 if (btrfs_defrag_cancelled(root->fs_info)) {
1302                         pr_debug("BTRFS: defrag_root cancelled\n");
1303                         ret = -EAGAIN;
1304                         break;
1305                 }
1306         }
1307         clear_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state);
1308         return ret;
1309 }
1310
1311 /*
1312  * new snapshots need to be created at a very specific time in the
1313  * transaction commit.  This does the actual creation.
1314  *
1315  * Note:
1316  * If the error which may affect the commitment of the current transaction
1317  * happens, we should return the error number. If the error which just affect
1318  * the creation of the pending snapshots, just return 0.
1319  */
1320 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1321                                    struct btrfs_fs_info *fs_info,
1322                                    struct btrfs_pending_snapshot *pending)
1323 {
1324         struct btrfs_key key;
1325         struct btrfs_root_item *new_root_item;
1326         struct btrfs_root *tree_root = fs_info->tree_root;
1327         struct btrfs_root *root = pending->root;
1328         struct btrfs_root *parent_root;
1329         struct btrfs_block_rsv *rsv;
1330         struct inode *parent_inode;
1331         struct btrfs_path *path;
1332         struct btrfs_dir_item *dir_item;
1333         struct dentry *dentry;
1334         struct extent_buffer *tmp;
1335         struct extent_buffer *old;
1336         struct timespec cur_time = CURRENT_TIME;
1337         int ret = 0;
1338         u64 to_reserve = 0;
1339         u64 index = 0;
1340         u64 objectid;
1341         u64 root_flags;
1342         uuid_le new_uuid;
1343
1344         ASSERT(pending->path);
1345         path = pending->path;
1346
1347         ASSERT(pending->root_item);
1348         new_root_item = pending->root_item;
1349
1350         pending->error = btrfs_find_free_objectid(tree_root, &objectid);
1351         if (pending->error)
1352                 goto no_free_objectid;
1353
1354         /*
1355          * Make qgroup to skip current new snapshot's qgroupid, as it is
1356          * accounted by later btrfs_qgroup_inherit().
1357          */
1358         btrfs_set_skip_qgroup(trans, objectid);
1359
1360         btrfs_reloc_pre_snapshot(pending, &to_reserve);
1361
1362         if (to_reserve > 0) {
1363                 pending->error = btrfs_block_rsv_add(root,
1364                                                      &pending->block_rsv,
1365                                                      to_reserve,
1366                                                      BTRFS_RESERVE_NO_FLUSH);
1367                 if (pending->error)
1368                         goto clear_skip_qgroup;
1369         }
1370
1371         key.objectid = objectid;
1372         key.offset = (u64)-1;
1373         key.type = BTRFS_ROOT_ITEM_KEY;
1374
1375         rsv = trans->block_rsv;
1376         trans->block_rsv = &pending->block_rsv;
1377         trans->bytes_reserved = trans->block_rsv->reserved;
1378
1379         dentry = pending->dentry;
1380         parent_inode = pending->dir;
1381         parent_root = BTRFS_I(parent_inode)->root;
1382         record_root_in_trans(trans, parent_root);
1383
1384         /*
1385          * insert the directory item
1386          */
1387         ret = btrfs_set_inode_index(parent_inode, &index);
1388         BUG_ON(ret); /* -ENOMEM */
1389
1390         /* check if there is a file/dir which has the same name. */
1391         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1392                                          btrfs_ino(parent_inode),
1393                                          dentry->d_name.name,
1394                                          dentry->d_name.len, 0);
1395         if (dir_item != NULL && !IS_ERR(dir_item)) {
1396                 pending->error = -EEXIST;
1397                 goto dir_item_existed;
1398         } else if (IS_ERR(dir_item)) {
1399                 ret = PTR_ERR(dir_item);
1400                 btrfs_abort_transaction(trans, root, ret);
1401                 goto fail;
1402         }
1403         btrfs_release_path(path);
1404
1405         /*
1406          * pull in the delayed directory update
1407          * and the delayed inode item
1408          * otherwise we corrupt the FS during
1409          * snapshot
1410          */
1411         ret = btrfs_run_delayed_items(trans, root);
1412         if (ret) {      /* Transaction aborted */
1413                 btrfs_abort_transaction(trans, root, ret);
1414                 goto fail;
1415         }
1416
1417         record_root_in_trans(trans, root);
1418         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1419         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1420         btrfs_check_and_init_root_item(new_root_item);
1421
1422         root_flags = btrfs_root_flags(new_root_item);
1423         if (pending->readonly)
1424                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1425         else
1426                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1427         btrfs_set_root_flags(new_root_item, root_flags);
1428
1429         btrfs_set_root_generation_v2(new_root_item,
1430                         trans->transid);
1431         uuid_le_gen(&new_uuid);
1432         memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1433         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1434                         BTRFS_UUID_SIZE);
1435         if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1436                 memset(new_root_item->received_uuid, 0,
1437                        sizeof(new_root_item->received_uuid));
1438                 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1439                 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1440                 btrfs_set_root_stransid(new_root_item, 0);
1441                 btrfs_set_root_rtransid(new_root_item, 0);
1442         }
1443         btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1444         btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1445         btrfs_set_root_otransid(new_root_item, trans->transid);
1446
1447         old = btrfs_lock_root_node(root);
1448         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1449         if (ret) {
1450                 btrfs_tree_unlock(old);
1451                 free_extent_buffer(old);
1452                 btrfs_abort_transaction(trans, root, ret);
1453                 goto fail;
1454         }
1455
1456         btrfs_set_lock_blocking(old);
1457
1458         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1459         /* clean up in any case */
1460         btrfs_tree_unlock(old);
1461         free_extent_buffer(old);
1462         if (ret) {
1463                 btrfs_abort_transaction(trans, root, ret);
1464                 goto fail;
1465         }
1466         /* see comments in should_cow_block() */
1467         set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1468         smp_wmb();
1469
1470         btrfs_set_root_node(new_root_item, tmp);
1471         /* record when the snapshot was created in key.offset */
1472         key.offset = trans->transid;
1473         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1474         btrfs_tree_unlock(tmp);
1475         free_extent_buffer(tmp);
1476         if (ret) {
1477                 btrfs_abort_transaction(trans, root, ret);
1478                 goto fail;
1479         }
1480
1481         /*
1482          * insert root back/forward references
1483          */
1484         ret = btrfs_add_root_ref(trans, tree_root, objectid,
1485                                  parent_root->root_key.objectid,
1486                                  btrfs_ino(parent_inode), index,
1487                                  dentry->d_name.name, dentry->d_name.len);
1488         if (ret) {
1489                 btrfs_abort_transaction(trans, root, ret);
1490                 goto fail;
1491         }
1492
1493         key.offset = (u64)-1;
1494         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1495         if (IS_ERR(pending->snap)) {
1496                 ret = PTR_ERR(pending->snap);
1497                 btrfs_abort_transaction(trans, root, ret);
1498                 goto fail;
1499         }
1500
1501         ret = btrfs_reloc_post_snapshot(trans, pending);
1502         if (ret) {
1503                 btrfs_abort_transaction(trans, root, ret);
1504                 goto fail;
1505         }
1506
1507         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1508         if (ret) {
1509                 btrfs_abort_transaction(trans, root, ret);
1510                 goto fail;
1511         }
1512
1513         ret = btrfs_insert_dir_item(trans, parent_root,
1514                                     dentry->d_name.name, dentry->d_name.len,
1515                                     parent_inode, &key,
1516                                     BTRFS_FT_DIR, index);
1517         /* We have check then name at the beginning, so it is impossible. */
1518         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1519         if (ret) {
1520                 btrfs_abort_transaction(trans, root, ret);
1521                 goto fail;
1522         }
1523
1524         btrfs_i_size_write(parent_inode, parent_inode->i_size +
1525                                          dentry->d_name.len * 2);
1526         parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
1527         ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1528         if (ret) {
1529                 btrfs_abort_transaction(trans, root, ret);
1530                 goto fail;
1531         }
1532         ret = btrfs_uuid_tree_add(trans, fs_info->uuid_root, new_uuid.b,
1533                                   BTRFS_UUID_KEY_SUBVOL, objectid);
1534         if (ret) {
1535                 btrfs_abort_transaction(trans, root, ret);
1536                 goto fail;
1537         }
1538         if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1539                 ret = btrfs_uuid_tree_add(trans, fs_info->uuid_root,
1540                                           new_root_item->received_uuid,
1541                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1542                                           objectid);
1543                 if (ret && ret != -EEXIST) {
1544                         btrfs_abort_transaction(trans, root, ret);
1545                         goto fail;
1546                 }
1547         }
1548
1549         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1550         if (ret) {
1551                 btrfs_abort_transaction(trans, root, ret);
1552                 goto fail;
1553         }
1554
1555         /*
1556          * account qgroup counters before qgroup_inherit()
1557          */
1558         ret = btrfs_qgroup_prepare_account_extents(trans, fs_info);
1559         if (ret)
1560                 goto fail;
1561         ret = btrfs_qgroup_account_extents(trans, fs_info);
1562         if (ret)
1563                 goto fail;
1564         ret = btrfs_qgroup_inherit(trans, fs_info,
1565                                    root->root_key.objectid,
1566                                    objectid, pending->inherit);
1567         if (ret) {
1568                 btrfs_abort_transaction(trans, root, ret);
1569                 goto fail;
1570         }
1571
1572 fail:
1573         pending->error = ret;
1574 dir_item_existed:
1575         trans->block_rsv = rsv;
1576         trans->bytes_reserved = 0;
1577 clear_skip_qgroup:
1578         btrfs_clear_skip_qgroup(trans);
1579 no_free_objectid:
1580         kfree(new_root_item);
1581         pending->root_item = NULL;
1582         btrfs_free_path(path);
1583         pending->path = NULL;
1584
1585         return ret;
1586 }
1587
1588 /*
1589  * create all the snapshots we've scheduled for creation
1590  */
1591 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1592                                              struct btrfs_fs_info *fs_info)
1593 {
1594         struct btrfs_pending_snapshot *pending, *next;
1595         struct list_head *head = &trans->transaction->pending_snapshots;
1596         int ret = 0;
1597
1598         list_for_each_entry_safe(pending, next, head, list) {
1599                 list_del(&pending->list);
1600                 ret = create_pending_snapshot(trans, fs_info, pending);
1601                 if (ret)
1602                         break;
1603         }
1604         return ret;
1605 }
1606
1607 static void update_super_roots(struct btrfs_root *root)
1608 {
1609         struct btrfs_root_item *root_item;
1610         struct btrfs_super_block *super;
1611
1612         super = root->fs_info->super_copy;
1613
1614         root_item = &root->fs_info->chunk_root->root_item;
1615         super->chunk_root = root_item->bytenr;
1616         super->chunk_root_generation = root_item->generation;
1617         super->chunk_root_level = root_item->level;
1618
1619         root_item = &root->fs_info->tree_root->root_item;
1620         super->root = root_item->bytenr;
1621         super->generation = root_item->generation;
1622         super->root_level = root_item->level;
1623         if (btrfs_test_opt(root, SPACE_CACHE))
1624                 super->cache_generation = root_item->generation;
1625         if (root->fs_info->update_uuid_tree_gen)
1626                 super->uuid_tree_generation = root_item->generation;
1627 }
1628
1629 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1630 {
1631         struct btrfs_transaction *trans;
1632         int ret = 0;
1633
1634         spin_lock(&info->trans_lock);
1635         trans = info->running_transaction;
1636         if (trans)
1637                 ret = (trans->state >= TRANS_STATE_COMMIT_START);
1638         spin_unlock(&info->trans_lock);
1639         return ret;
1640 }
1641
1642 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1643 {
1644         struct btrfs_transaction *trans;
1645         int ret = 0;
1646
1647         spin_lock(&info->trans_lock);
1648         trans = info->running_transaction;
1649         if (trans)
1650                 ret = is_transaction_blocked(trans);
1651         spin_unlock(&info->trans_lock);
1652         return ret;
1653 }
1654
1655 /*
1656  * wait for the current transaction commit to start and block subsequent
1657  * transaction joins
1658  */
1659 static void wait_current_trans_commit_start(struct btrfs_root *root,
1660                                             struct btrfs_transaction *trans)
1661 {
1662         wait_event(root->fs_info->transaction_blocked_wait,
1663                    trans->state >= TRANS_STATE_COMMIT_START ||
1664                    trans->aborted);
1665 }
1666
1667 /*
1668  * wait for the current transaction to start and then become unblocked.
1669  * caller holds ref.
1670  */
1671 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1672                                          struct btrfs_transaction *trans)
1673 {
1674         wait_event(root->fs_info->transaction_wait,
1675                    trans->state >= TRANS_STATE_UNBLOCKED ||
1676                    trans->aborted);
1677 }
1678
1679 /*
1680  * commit transactions asynchronously. once btrfs_commit_transaction_async
1681  * returns, any subsequent transaction will not be allowed to join.
1682  */
1683 struct btrfs_async_commit {
1684         struct btrfs_trans_handle *newtrans;
1685         struct btrfs_root *root;
1686         struct work_struct work;
1687 };
1688
1689 static void do_async_commit(struct work_struct *work)
1690 {
1691         struct btrfs_async_commit *ac =
1692                 container_of(work, struct btrfs_async_commit, work);
1693
1694         /*
1695          * We've got freeze protection passed with the transaction.
1696          * Tell lockdep about it.
1697          */
1698         if (ac->newtrans->type & __TRANS_FREEZABLE)
1699                 __sb_writers_acquired(ac->root->fs_info->sb, SB_FREEZE_FS);
1700
1701         current->journal_info = ac->newtrans;
1702
1703         btrfs_commit_transaction(ac->newtrans, ac->root);
1704         kfree(ac);
1705 }
1706
1707 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1708                                    struct btrfs_root *root,
1709                                    int wait_for_unblock)
1710 {
1711         struct btrfs_async_commit *ac;
1712         struct btrfs_transaction *cur_trans;
1713
1714         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1715         if (!ac)
1716                 return -ENOMEM;
1717
1718         INIT_WORK(&ac->work, do_async_commit);
1719         ac->root = root;
1720         ac->newtrans = btrfs_join_transaction(root);
1721         if (IS_ERR(ac->newtrans)) {
1722                 int err = PTR_ERR(ac->newtrans);
1723                 kfree(ac);
1724                 return err;
1725         }
1726
1727         /* take transaction reference */
1728         cur_trans = trans->transaction;
1729         atomic_inc(&cur_trans->use_count);
1730
1731         btrfs_end_transaction(trans, root);
1732
1733         /*
1734          * Tell lockdep we've released the freeze rwsem, since the
1735          * async commit thread will be the one to unlock it.
1736          */
1737         if (ac->newtrans->type & __TRANS_FREEZABLE)
1738                 __sb_writers_release(root->fs_info->sb, SB_FREEZE_FS);
1739
1740         schedule_work(&ac->work);
1741
1742         /* wait for transaction to start and unblock */
1743         if (wait_for_unblock)
1744                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1745         else
1746                 wait_current_trans_commit_start(root, cur_trans);
1747
1748         if (current->journal_info == trans)
1749                 current->journal_info = NULL;
1750
1751         btrfs_put_transaction(cur_trans);
1752         return 0;
1753 }
1754
1755
1756 static void cleanup_transaction(struct btrfs_trans_handle *trans,
1757                                 struct btrfs_root *root, int err)
1758 {
1759         struct btrfs_transaction *cur_trans = trans->transaction;
1760         DEFINE_WAIT(wait);
1761
1762         WARN_ON(trans->use_count > 1);
1763
1764         btrfs_abort_transaction(trans, root, err);
1765
1766         spin_lock(&root->fs_info->trans_lock);
1767
1768         /*
1769          * If the transaction is removed from the list, it means this
1770          * transaction has been committed successfully, so it is impossible
1771          * to call the cleanup function.
1772          */
1773         BUG_ON(list_empty(&cur_trans->list));
1774
1775         list_del_init(&cur_trans->list);
1776         if (cur_trans == root->fs_info->running_transaction) {
1777                 cur_trans->state = TRANS_STATE_COMMIT_DOING;
1778                 spin_unlock(&root->fs_info->trans_lock);
1779                 wait_event(cur_trans->writer_wait,
1780                            atomic_read(&cur_trans->num_writers) == 1);
1781
1782                 spin_lock(&root->fs_info->trans_lock);
1783         }
1784         spin_unlock(&root->fs_info->trans_lock);
1785
1786         btrfs_cleanup_one_transaction(trans->transaction, root);
1787
1788         spin_lock(&root->fs_info->trans_lock);
1789         if (cur_trans == root->fs_info->running_transaction)
1790                 root->fs_info->running_transaction = NULL;
1791         spin_unlock(&root->fs_info->trans_lock);
1792
1793         if (trans->type & __TRANS_FREEZABLE)
1794                 sb_end_intwrite(root->fs_info->sb);
1795         btrfs_put_transaction(cur_trans);
1796         btrfs_put_transaction(cur_trans);
1797
1798         trace_btrfs_transaction_commit(root);
1799
1800         if (current->journal_info == trans)
1801                 current->journal_info = NULL;
1802         btrfs_scrub_cancel(root->fs_info);
1803
1804         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1805 }
1806
1807 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
1808 {
1809         if (btrfs_test_opt(fs_info->tree_root, FLUSHONCOMMIT))
1810                 return btrfs_start_delalloc_roots(fs_info, 1, -1);
1811         return 0;
1812 }
1813
1814 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
1815 {
1816         if (btrfs_test_opt(fs_info->tree_root, FLUSHONCOMMIT))
1817                 btrfs_wait_ordered_roots(fs_info, -1);
1818 }
1819
1820 static inline void
1821 btrfs_wait_pending_ordered(struct btrfs_transaction *cur_trans)
1822 {
1823         wait_event(cur_trans->pending_wait,
1824                    atomic_read(&cur_trans->pending_ordered) == 0);
1825 }
1826
1827 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1828                              struct btrfs_root *root)
1829 {
1830         struct btrfs_transaction *cur_trans = trans->transaction;
1831         struct btrfs_transaction *prev_trans = NULL;
1832         struct btrfs_inode *btree_ino = BTRFS_I(root->fs_info->btree_inode);
1833         int ret;
1834
1835         /* Stop the commit early if ->aborted is set */
1836         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1837                 ret = cur_trans->aborted;
1838                 btrfs_end_transaction(trans, root);
1839                 return ret;
1840         }
1841
1842         /* make a pass through all the delayed refs we have so far
1843          * any runnings procs may add more while we are here
1844          */
1845         ret = btrfs_run_delayed_refs(trans, root, 0);
1846         if (ret) {
1847                 btrfs_end_transaction(trans, root);
1848                 return ret;
1849         }
1850
1851         btrfs_trans_release_metadata(trans, root);
1852         trans->block_rsv = NULL;
1853
1854         cur_trans = trans->transaction;
1855
1856         /*
1857          * set the flushing flag so procs in this transaction have to
1858          * start sending their work down.
1859          */
1860         cur_trans->delayed_refs.flushing = 1;
1861         smp_wmb();
1862
1863         if (!list_empty(&trans->new_bgs))
1864                 btrfs_create_pending_block_groups(trans, root);
1865
1866         ret = btrfs_run_delayed_refs(trans, root, 0);
1867         if (ret) {
1868                 btrfs_end_transaction(trans, root);
1869                 return ret;
1870         }
1871
1872         if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) {
1873                 int run_it = 0;
1874
1875                 /* this mutex is also taken before trying to set
1876                  * block groups readonly.  We need to make sure
1877                  * that nobody has set a block group readonly
1878                  * after a extents from that block group have been
1879                  * allocated for cache files.  btrfs_set_block_group_ro
1880                  * will wait for the transaction to commit if it
1881                  * finds BTRFS_TRANS_DIRTY_BG_RUN set.
1882                  *
1883                  * The BTRFS_TRANS_DIRTY_BG_RUN flag is also used to make sure
1884                  * only one process starts all the block group IO.  It wouldn't
1885                  * hurt to have more than one go through, but there's no
1886                  * real advantage to it either.
1887                  */
1888                 mutex_lock(&root->fs_info->ro_block_group_mutex);
1889                 if (!test_and_set_bit(BTRFS_TRANS_DIRTY_BG_RUN,
1890                                       &cur_trans->flags))
1891                         run_it = 1;
1892                 mutex_unlock(&root->fs_info->ro_block_group_mutex);
1893
1894                 if (run_it)
1895                         ret = btrfs_start_dirty_block_groups(trans, root);
1896         }
1897         if (ret) {
1898                 btrfs_end_transaction(trans, root);
1899                 return ret;
1900         }
1901
1902         spin_lock(&root->fs_info->trans_lock);
1903         if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
1904                 spin_unlock(&root->fs_info->trans_lock);
1905                 atomic_inc(&cur_trans->use_count);
1906                 ret = btrfs_end_transaction(trans, root);
1907
1908                 wait_for_commit(root, cur_trans);
1909
1910                 if (unlikely(cur_trans->aborted))
1911                         ret = cur_trans->aborted;
1912
1913                 btrfs_put_transaction(cur_trans);
1914
1915                 return ret;
1916         }
1917
1918         cur_trans->state = TRANS_STATE_COMMIT_START;
1919         wake_up(&root->fs_info->transaction_blocked_wait);
1920
1921         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1922                 prev_trans = list_entry(cur_trans->list.prev,
1923                                         struct btrfs_transaction, list);
1924                 if (prev_trans->state != TRANS_STATE_COMPLETED) {
1925                         atomic_inc(&prev_trans->use_count);
1926                         spin_unlock(&root->fs_info->trans_lock);
1927
1928                         wait_for_commit(root, prev_trans);
1929                         ret = prev_trans->aborted;
1930
1931                         btrfs_put_transaction(prev_trans);
1932                         if (ret)
1933                                 goto cleanup_transaction;
1934                 } else {
1935                         spin_unlock(&root->fs_info->trans_lock);
1936                 }
1937         } else {
1938                 spin_unlock(&root->fs_info->trans_lock);
1939         }
1940
1941         extwriter_counter_dec(cur_trans, trans->type);
1942
1943         ret = btrfs_start_delalloc_flush(root->fs_info);
1944         if (ret)
1945                 goto cleanup_transaction;
1946
1947         ret = btrfs_run_delayed_items(trans, root);
1948         if (ret)
1949                 goto cleanup_transaction;
1950
1951         wait_event(cur_trans->writer_wait,
1952                    extwriter_counter_read(cur_trans) == 0);
1953
1954         /* some pending stuffs might be added after the previous flush. */
1955         ret = btrfs_run_delayed_items(trans, root);
1956         if (ret)
1957                 goto cleanup_transaction;
1958
1959         btrfs_wait_delalloc_flush(root->fs_info);
1960
1961         btrfs_wait_pending_ordered(cur_trans);
1962
1963         btrfs_scrub_pause(root);
1964         /*
1965          * Ok now we need to make sure to block out any other joins while we
1966          * commit the transaction.  We could have started a join before setting
1967          * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
1968          */
1969         spin_lock(&root->fs_info->trans_lock);
1970         cur_trans->state = TRANS_STATE_COMMIT_DOING;
1971         spin_unlock(&root->fs_info->trans_lock);
1972         wait_event(cur_trans->writer_wait,
1973                    atomic_read(&cur_trans->num_writers) == 1);
1974
1975         /* ->aborted might be set after the previous check, so check it */
1976         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1977                 ret = cur_trans->aborted;
1978                 goto scrub_continue;
1979         }
1980         /*
1981          * the reloc mutex makes sure that we stop
1982          * the balancing code from coming in and moving
1983          * extents around in the middle of the commit
1984          */
1985         mutex_lock(&root->fs_info->reloc_mutex);
1986
1987         /*
1988          * We needn't worry about the delayed items because we will
1989          * deal with them in create_pending_snapshot(), which is the
1990          * core function of the snapshot creation.
1991          */
1992         ret = create_pending_snapshots(trans, root->fs_info);
1993         if (ret) {
1994                 mutex_unlock(&root->fs_info->reloc_mutex);
1995                 goto scrub_continue;
1996         }
1997
1998         /*
1999          * We insert the dir indexes of the snapshots and update the inode
2000          * of the snapshots' parents after the snapshot creation, so there
2001          * are some delayed items which are not dealt with. Now deal with
2002          * them.
2003          *
2004          * We needn't worry that this operation will corrupt the snapshots,
2005          * because all the tree which are snapshoted will be forced to COW
2006          * the nodes and leaves.
2007          */
2008         ret = btrfs_run_delayed_items(trans, root);
2009         if (ret) {
2010                 mutex_unlock(&root->fs_info->reloc_mutex);
2011                 goto scrub_continue;
2012         }
2013
2014         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
2015         if (ret) {
2016                 mutex_unlock(&root->fs_info->reloc_mutex);
2017                 goto scrub_continue;
2018         }
2019
2020         /* Reocrd old roots for later qgroup accounting */
2021         ret = btrfs_qgroup_prepare_account_extents(trans, root->fs_info);
2022         if (ret) {
2023                 mutex_unlock(&root->fs_info->reloc_mutex);
2024                 goto scrub_continue;
2025         }
2026
2027         /*
2028          * make sure none of the code above managed to slip in a
2029          * delayed item
2030          */
2031         btrfs_assert_delayed_root_empty(root);
2032
2033         WARN_ON(cur_trans != trans->transaction);
2034
2035         /* btrfs_commit_tree_roots is responsible for getting the
2036          * various roots consistent with each other.  Every pointer
2037          * in the tree of tree roots has to point to the most up to date
2038          * root for every subvolume and other tree.  So, we have to keep
2039          * the tree logging code from jumping in and changing any
2040          * of the trees.
2041          *
2042          * At this point in the commit, there can't be any tree-log
2043          * writers, but a little lower down we drop the trans mutex
2044          * and let new people in.  By holding the tree_log_mutex
2045          * from now until after the super is written, we avoid races
2046          * with the tree-log code.
2047          */
2048         mutex_lock(&root->fs_info->tree_log_mutex);
2049
2050         ret = commit_fs_roots(trans, root);
2051         if (ret) {
2052                 mutex_unlock(&root->fs_info->tree_log_mutex);
2053                 mutex_unlock(&root->fs_info->reloc_mutex);
2054                 goto scrub_continue;
2055         }
2056
2057         /*
2058          * Since the transaction is done, we can apply the pending changes
2059          * before the next transaction.
2060          */
2061         btrfs_apply_pending_changes(root->fs_info);
2062
2063         /* commit_fs_roots gets rid of all the tree log roots, it is now
2064          * safe to free the root of tree log roots
2065          */
2066         btrfs_free_log_root_tree(trans, root->fs_info);
2067
2068         /*
2069          * Since fs roots are all committed, we can get a quite accurate
2070          * new_roots. So let's do quota accounting.
2071          */
2072         ret = btrfs_qgroup_account_extents(trans, root->fs_info);
2073         if (ret < 0) {
2074                 mutex_unlock(&root->fs_info->tree_log_mutex);
2075                 mutex_unlock(&root->fs_info->reloc_mutex);
2076                 goto scrub_continue;
2077         }
2078
2079         ret = commit_cowonly_roots(trans, root);
2080         if (ret) {
2081                 mutex_unlock(&root->fs_info->tree_log_mutex);
2082                 mutex_unlock(&root->fs_info->reloc_mutex);
2083                 goto scrub_continue;
2084         }
2085
2086         /*
2087          * The tasks which save the space cache and inode cache may also
2088          * update ->aborted, check it.
2089          */
2090         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
2091                 ret = cur_trans->aborted;
2092                 mutex_unlock(&root->fs_info->tree_log_mutex);
2093                 mutex_unlock(&root->fs_info->reloc_mutex);
2094                 goto scrub_continue;
2095         }
2096
2097         btrfs_prepare_extent_commit(trans, root);
2098
2099         cur_trans = root->fs_info->running_transaction;
2100
2101         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
2102                             root->fs_info->tree_root->node);
2103         list_add_tail(&root->fs_info->tree_root->dirty_list,
2104                       &cur_trans->switch_commits);
2105
2106         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
2107                             root->fs_info->chunk_root->node);
2108         list_add_tail(&root->fs_info->chunk_root->dirty_list,
2109                       &cur_trans->switch_commits);
2110
2111         switch_commit_roots(cur_trans, root->fs_info);
2112
2113         assert_qgroups_uptodate(trans);
2114         ASSERT(list_empty(&cur_trans->dirty_bgs));
2115         ASSERT(list_empty(&cur_trans->io_bgs));
2116         update_super_roots(root);
2117
2118         btrfs_set_super_log_root(root->fs_info->super_copy, 0);
2119         btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
2120         memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
2121                sizeof(*root->fs_info->super_copy));
2122
2123         btrfs_update_commit_device_size(root->fs_info);
2124         btrfs_update_commit_device_bytes_used(root, cur_trans);
2125
2126         clear_bit(BTRFS_INODE_BTREE_LOG1_ERR, &btree_ino->runtime_flags);
2127         clear_bit(BTRFS_INODE_BTREE_LOG2_ERR, &btree_ino->runtime_flags);
2128
2129         btrfs_trans_release_chunk_metadata(trans);
2130
2131         spin_lock(&root->fs_info->trans_lock);
2132         cur_trans->state = TRANS_STATE_UNBLOCKED;
2133         root->fs_info->running_transaction = NULL;
2134         spin_unlock(&root->fs_info->trans_lock);
2135         mutex_unlock(&root->fs_info->reloc_mutex);
2136
2137         wake_up(&root->fs_info->transaction_wait);
2138
2139         ret = btrfs_write_and_wait_transaction(trans, root);
2140         if (ret) {
2141                 btrfs_std_error(root->fs_info, ret,
2142                             "Error while writing out transaction");
2143                 mutex_unlock(&root->fs_info->tree_log_mutex);
2144                 goto scrub_continue;
2145         }
2146
2147         ret = write_ctree_super(trans, root, 0);
2148         if (ret) {
2149                 mutex_unlock(&root->fs_info->tree_log_mutex);
2150                 goto scrub_continue;
2151         }
2152
2153         /*
2154          * the super is written, we can safely allow the tree-loggers
2155          * to go about their business
2156          */
2157         mutex_unlock(&root->fs_info->tree_log_mutex);
2158
2159         btrfs_finish_extent_commit(trans, root);
2160
2161         if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &cur_trans->flags))
2162                 btrfs_clear_space_info_full(root->fs_info);
2163
2164         root->fs_info->last_trans_committed = cur_trans->transid;
2165         /*
2166          * We needn't acquire the lock here because there is no other task
2167          * which can change it.
2168          */
2169         cur_trans->state = TRANS_STATE_COMPLETED;
2170         wake_up(&cur_trans->commit_wait);
2171
2172         spin_lock(&root->fs_info->trans_lock);
2173         list_del_init(&cur_trans->list);
2174         spin_unlock(&root->fs_info->trans_lock);
2175
2176         btrfs_put_transaction(cur_trans);
2177         btrfs_put_transaction(cur_trans);
2178
2179         if (trans->type & __TRANS_FREEZABLE)
2180                 sb_end_intwrite(root->fs_info->sb);
2181
2182         trace_btrfs_transaction_commit(root);
2183
2184         btrfs_scrub_continue(root);
2185
2186         if (current->journal_info == trans)
2187                 current->journal_info = NULL;
2188
2189         kmem_cache_free(btrfs_trans_handle_cachep, trans);
2190
2191         if (current != root->fs_info->transaction_kthread &&
2192             current != root->fs_info->cleaner_kthread)
2193                 btrfs_run_delayed_iputs(root);
2194
2195         return ret;
2196
2197 scrub_continue:
2198         btrfs_scrub_continue(root);
2199 cleanup_transaction:
2200         btrfs_trans_release_metadata(trans, root);
2201         btrfs_trans_release_chunk_metadata(trans);
2202         trans->block_rsv = NULL;
2203         btrfs_warn(root->fs_info, "Skipping commit of aborted transaction.");
2204         if (current->journal_info == trans)
2205                 current->journal_info = NULL;
2206         cleanup_transaction(trans, root, ret);
2207
2208         return ret;
2209 }
2210
2211 /*
2212  * return < 0 if error
2213  * 0 if there are no more dead_roots at the time of call
2214  * 1 there are more to be processed, call me again
2215  *
2216  * The return value indicates there are certainly more snapshots to delete, but
2217  * if there comes a new one during processing, it may return 0. We don't mind,
2218  * because btrfs_commit_super will poke cleaner thread and it will process it a
2219  * few seconds later.
2220  */
2221 int btrfs_clean_one_deleted_snapshot(struct btrfs_root *root)
2222 {
2223         int ret;
2224         struct btrfs_fs_info *fs_info = root->fs_info;
2225
2226         spin_lock(&fs_info->trans_lock);
2227         if (list_empty(&fs_info->dead_roots)) {
2228                 spin_unlock(&fs_info->trans_lock);
2229                 return 0;
2230         }
2231         root = list_first_entry(&fs_info->dead_roots,
2232                         struct btrfs_root, root_list);
2233         list_del_init(&root->root_list);
2234         spin_unlock(&fs_info->trans_lock);
2235
2236         pr_debug("BTRFS: cleaner removing %llu\n", root->objectid);
2237
2238         btrfs_kill_all_delayed_nodes(root);
2239
2240         if (btrfs_header_backref_rev(root->node) <
2241                         BTRFS_MIXED_BACKREF_REV)
2242                 ret = btrfs_drop_snapshot(root, NULL, 0, 0);
2243         else
2244                 ret = btrfs_drop_snapshot(root, NULL, 1, 0);
2245
2246         return (ret < 0) ? 0 : 1;
2247 }
2248
2249 void btrfs_apply_pending_changes(struct btrfs_fs_info *fs_info)
2250 {
2251         unsigned long prev;
2252         unsigned long bit;
2253
2254         prev = xchg(&fs_info->pending_changes, 0);
2255         if (!prev)
2256                 return;
2257
2258         bit = 1 << BTRFS_PENDING_SET_INODE_MAP_CACHE;
2259         if (prev & bit)
2260                 btrfs_set_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2261         prev &= ~bit;
2262
2263         bit = 1 << BTRFS_PENDING_CLEAR_INODE_MAP_CACHE;
2264         if (prev & bit)
2265                 btrfs_clear_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2266         prev &= ~bit;
2267
2268         bit = 1 << BTRFS_PENDING_COMMIT;
2269         if (prev & bit)
2270                 btrfs_debug(fs_info, "pending commit done");
2271         prev &= ~bit;
2272
2273         if (prev)
2274                 btrfs_warn(fs_info,
2275                         "unknown pending changes left 0x%lx, ignoring", prev);
2276 }