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                                int force)
 316 {
 317         if ((test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
 318             root->last_trans < trans->transid) || force) {
 319                 WARN_ON(root == root->fs_info->extent_root);
 320                 WARN_ON(root->commit_root != root->node);
 321
 322                 /*
 323                  * see below for IN_TRANS_SETUP usage rules
 324                  * we have the reloc mutex held now, so there
 325                  * is only one writer in this function
 326                  */
 327                 set_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
 328
 329                 /* make sure readers find IN_TRANS_SETUP before
 330                  * they find our root->last_trans update
 331                  */
 332                 smp_wmb();
 333
 334                 spin_lock(&root->fs_info->fs_roots_radix_lock);
 335                 if (root->last_trans == trans->transid && !force) {
 336                         spin_unlock(&root->fs_info->fs_roots_radix_lock);
 337                         return 0;
 338                 }
 339                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 340                            (unsigned long)root->root_key.objectid,
 341                            BTRFS_ROOT_TRANS_TAG);
 342                 spin_unlock(&root->fs_info->fs_roots_radix_lock);
 343                 root->last_trans = trans->transid;
 344
 345                 /* this is pretty tricky.  We don't want to
 346                  * take the relocation lock in btrfs_record_root_in_trans
 347                  * unless we're really doing the first setup for this root in
 348                  * this transaction.
 349                  *
 350                  * Normally we'd use root->last_trans as a flag to decide
 351                  * if we want to take the expensive mutex.
 352                  *
 353                  * But, we have to set root->last_trans before we
 354                  * init the relocation root, otherwise, we trip over warnings
 355                  * in ctree.c.  The solution used here is to flag ourselves
 356                  * with root IN_TRANS_SETUP.  When this is 1, we're still
 357                  * fixing up the reloc trees and everyone must wait.
 358                  *
 359                  * When this is zero, they can trust root->last_trans and fly
 360                  * through btrfs_record_root_in_trans without having to take the
 361                  * lock.  smp_wmb() makes sure that all the writes above are
 362                  * done before we pop in the zero below
 363                  */
 364                 btrfs_init_reloc_root(trans, root);
 365                 smp_mb__before_atomic();
 366                 clear_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
 367         }
 368         return 0;
 369 }
 370
 371
 372 void btrfs_add_dropped_root(struct btrfs_trans_handle *trans,
 373                             struct btrfs_root *root)
 374 {
 375         struct btrfs_transaction *cur_trans = trans->transaction;
 376
 377         /* Add ourselves to the transaction dropped list */
 378         spin_lock(&cur_trans->dropped_roots_lock);
 379         list_add_tail(&root->root_list, &cur_trans->dropped_roots);
 380         spin_unlock(&cur_trans->dropped_roots_lock);
 381
 382         /* Make sure we don't try to update the root at commit time */
 383         spin_lock(&root->fs_info->fs_roots_radix_lock);
 384         radix_tree_tag_clear(&root->fs_info->fs_roots_radix,
 385                              (unsigned long)root->root_key.objectid,
 386                              BTRFS_ROOT_TRANS_TAG);
 387         spin_unlock(&root->fs_info->fs_roots_radix_lock);
 388 }
 389
 390 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 391                                struct btrfs_root *root)
 392 {
 393         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
 394                 return 0;
 395
 396         /*
 397          * see record_root_in_trans for comments about IN_TRANS_SETUP usage
 398          * and barriers
 399          */
 400         smp_rmb();
 401         if (root->last_trans == trans->transid &&
 402             !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state))
 403                 return 0;
 404
 405         mutex_lock(&root->fs_info->reloc_mutex);
 406         record_root_in_trans(trans, root, 0);
 407         mutex_unlock(&root->fs_info->reloc_mutex);
 408
 409         return 0;
 410 }
 411
 412 static inline int is_transaction_blocked(struct btrfs_transaction *trans)
 413 {
 414         return (trans->state >= TRANS_STATE_BLOCKED &&
 415                 trans->state < TRANS_STATE_UNBLOCKED &&
 416                 !trans->aborted);
 417 }
 418
 419 /* wait for commit against the current transaction to become unblocked
 420  * when this is done, it is safe to start a new transaction, but the current
 421  * transaction might not be fully on disk.
 422  */
 423 static void wait_current_trans(struct btrfs_root *root)
 424 {
 425         struct btrfs_transaction *cur_trans;
 426
 427         spin_lock(&root->fs_info->trans_lock);
 428         cur_trans = root->fs_info->running_transaction;
 429         if (cur_trans && is_transaction_blocked(cur_trans)) {
 430                 atomic_inc(&cur_trans->use_count);
 431                 spin_unlock(&root->fs_info->trans_lock);
 432
 433                 wait_event(root->fs_info->transaction_wait,
 434                            cur_trans->state >= TRANS_STATE_UNBLOCKED ||
 435                            cur_trans->aborted);
 436                 btrfs_put_transaction(cur_trans);
 437         } else {
 438                 spin_unlock(&root->fs_info->trans_lock);
 439         }
 440 }
 441
 442 static int may_wait_transaction(struct btrfs_root *root, int type)
 443 {
 444         if (root->fs_info->log_root_recovering)
 445                 return 0;
 446
 447         if (type == TRANS_USERSPACE)
 448                 return 1;
 449
 450         if (type == TRANS_START &&
 451             !atomic_read(&root->fs_info->open_ioctl_trans))
 452                 return 1;
 453
 454         return 0;
 455 }
 456
 457 static inline bool need_reserve_reloc_root(struct btrfs_root *root)
 458 {
 459         if (!root->fs_info->reloc_ctl ||
 460             !test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
 461             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 462             root->reloc_root)
 463                 return false;
 464
 465         return true;
 466 }
 467
 468 static struct btrfs_trans_handle *
 469 start_transaction(struct btrfs_root *root, unsigned int num_items,
 470                   unsigned int type, enum btrfs_reserve_flush_enum flush)
 471 {
 472         struct btrfs_trans_handle *h;
 473         struct btrfs_transaction *cur_trans;
 474         u64 num_bytes = 0;
 475         u64 qgroup_reserved = 0;
 476         bool reloc_reserved = false;
 477         int ret;
 478
 479         /* Send isn't supposed to start transactions. */
 480         ASSERT(current->journal_info != BTRFS_SEND_TRANS_STUB);
 481
 482         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
 483                 return ERR_PTR(-EROFS);
 484
 485         if (current->journal_info) {
 486                 WARN_ON(type & TRANS_EXTWRITERS);
 487                 h = current->journal_info;
 488                 h->use_count++;
 489                 WARN_ON(h->use_count > 2);
 490                 h->orig_rsv = h->block_rsv;
 491                 h->block_rsv = NULL;
 492                 goto got_it;
 493         }
 494
 495         /*
 496          * Do the reservation before we join the transaction so we can do all
 497          * the appropriate flushing if need be.
 498          */
 499         if (num_items > 0 && root != root->fs_info->chunk_root) {
 500                 qgroup_reserved = num_items * root->nodesize;
 501                 ret = btrfs_qgroup_reserve_meta(root, qgroup_reserved);
 502                 if (ret)
 503                         return ERR_PTR(ret);
 504
 505                 num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
 506                 /*
 507                  * Do the reservation for the relocation root creation
 508                  */
 509                 if (need_reserve_reloc_root(root)) {
 510                         num_bytes += root->nodesize;
 511                         reloc_reserved = true;
 512                 }
 513
 514                 ret = btrfs_block_rsv_add(root,
 515                                           &root->fs_info->trans_block_rsv,
 516                                           num_bytes, flush);
 517                 if (ret)
 518                         goto reserve_fail;
 519         }
 520 again:
 521         h = kmem_cache_zalloc(btrfs_trans_handle_cachep, GFP_NOFS);
 522         if (!h) {
 523                 ret = -ENOMEM;
 524                 goto alloc_fail;
 525         }
 526
 527         /*
 528          * If we are JOIN_NOLOCK we're already committing a transaction and
 529          * waiting on this guy, so we don't need to do the sb_start_intwrite
 530          * because we're already holding a ref.  We need this because we could
 531          * have raced in and did an fsync() on a file which can kick a commit
 532          * and then we deadlock with somebody doing a freeze.
 533          *
 534          * If we are ATTACH, it means we just want to catch the current
 535          * transaction and commit it, so we needn't do sb_start_intwrite().
 536          */
 537         if (type & __TRANS_FREEZABLE)
 538                 sb_start_intwrite(root->fs_info->sb);
 539
 540         if (may_wait_transaction(root, type))
 541                 wait_current_trans(root);
 542
 543         do {
 544                 ret = join_transaction(root, type);
 545                 if (ret == -EBUSY) {
 546                         wait_current_trans(root);
 547                         if (unlikely(type == TRANS_ATTACH))
 548                                 ret = -ENOENT;
 549                 }
 550         } while (ret == -EBUSY);
 551
 552         if (ret < 0) {
 553                 /* We must get the transaction if we are JOIN_NOLOCK. */
 554                 BUG_ON(type == TRANS_JOIN_NOLOCK);
 555                 goto join_fail;
 556         }
 557
 558         cur_trans = root->fs_info->running_transaction;
 559
 560         h->transid = cur_trans->transid;
 561         h->transaction = cur_trans;
 562         h->root = root;
 563         h->use_count = 1;
 564
 565         h->type = type;
 566         h->can_flush_pending_bgs = true;
 567         INIT_LIST_HEAD(&h->qgroup_ref_list);
 568         INIT_LIST_HEAD(&h->new_bgs);
 569
 570         smp_mb();
 571         if (cur_trans->state >= TRANS_STATE_BLOCKED &&
 572             may_wait_transaction(root, type)) {
 573                 current->journal_info = h;
 574                 btrfs_commit_transaction(h, root);
 575                 goto again;
 576         }
 577
 578         if (num_bytes) {
 579                 trace_btrfs_space_reservation(root->fs_info, "transaction",
 580                                               h->transid, num_bytes, 1);
 581                 h->block_rsv = &root->fs_info->trans_block_rsv;
 582                 h->bytes_reserved = num_bytes;
 583                 h->reloc_reserved = reloc_reserved;
 584         }
 585
 586 got_it:
 587         btrfs_record_root_in_trans(h, root);
 588
 589         if (!current->journal_info && type != TRANS_USERSPACE)
 590                 current->journal_info = h;
 591         return h;
 592
 593 join_fail:
 594         if (type & __TRANS_FREEZABLE)
 595                 sb_end_intwrite(root->fs_info->sb);
 596         kmem_cache_free(btrfs_trans_handle_cachep, h);
 597 alloc_fail:
 598         if (num_bytes)
 599                 btrfs_block_rsv_release(root, &root->fs_info->trans_block_rsv,
 600                                         num_bytes);
 601 reserve_fail:
 602         btrfs_qgroup_free_meta(root, qgroup_reserved);
 603         return ERR_PTR(ret);
 604 }
 605
 606 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 607                                                    unsigned int num_items)
 608 {
 609         return start_transaction(root, num_items, TRANS_START,
 610                                  BTRFS_RESERVE_FLUSH_ALL);
 611 }
 612 struct btrfs_trans_handle *btrfs_start_transaction_fallback_global_rsv(
 613                                         struct btrfs_root *root,
 614                                         unsigned int num_items,
 615                                         int min_factor)
 616 {
 617         struct btrfs_trans_handle *trans;
 618         u64 num_bytes;
 619         int ret;
 620
 621         trans = btrfs_start_transaction(root, num_items);
 622         if (!IS_ERR(trans) || PTR_ERR(trans) != -ENOSPC)
 623                 return trans;
 624
 625         trans = btrfs_start_transaction(root, 0);
 626         if (IS_ERR(trans))
 627                 return trans;
 628
 629         num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
 630         ret = btrfs_cond_migrate_bytes(root->fs_info,
 631                                        &root->fs_info->trans_block_rsv,
 632                                        num_bytes,
 633                                        min_factor);
 634         if (ret) {
 635                 btrfs_end_transaction(trans, root);
 636                 return ERR_PTR(ret);
 637         }
 638
 639         trans->block_rsv = &root->fs_info->trans_block_rsv;
 640         trans->bytes_reserved = num_bytes;
 641         trace_btrfs_space_reservation(root->fs_info, "transaction",
 642                                       trans->transid, num_bytes, 1);
 643
 644         return trans;
 645 }
 646
 647 struct btrfs_trans_handle *btrfs_start_transaction_lflush(
 648                                         struct btrfs_root *root,
 649                                         unsigned int num_items)
 650 {
 651         return start_transaction(root, num_items, TRANS_START,
 652                                  BTRFS_RESERVE_FLUSH_LIMIT);
 653 }
 654
 655 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 656 {
 657         return start_transaction(root, 0, TRANS_JOIN,
 658                                  BTRFS_RESERVE_NO_FLUSH);
 659 }
 660
 661 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 662 {
 663         return start_transaction(root, 0, TRANS_JOIN_NOLOCK,
 664                                  BTRFS_RESERVE_NO_FLUSH);
 665 }
 666
 667 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 668 {
 669         return start_transaction(root, 0, TRANS_USERSPACE,
 670                                  BTRFS_RESERVE_NO_FLUSH);
 671 }
 672
 673 /*
 674  * btrfs_attach_transaction() - catch the running transaction
 675  *
 676  * It is used when we want to commit the current the transaction, but
 677  * don't want to start a new one.
 678  *
 679  * Note: If this function return -ENOENT, it just means there is no
 680  * running transaction. But it is possible that the inactive transaction
 681  * is still in the memory, not fully on disk. If you hope there is no
 682  * inactive transaction in the fs when -ENOENT is returned, you should
 683  * invoke
 684  *     btrfs_attach_transaction_barrier()
 685  */
 686 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
 687 {
 688         return start_transaction(root, 0, TRANS_ATTACH,
 689                                  BTRFS_RESERVE_NO_FLUSH);
 690 }
 691
 692 /*
 693  * btrfs_attach_transaction_barrier() - catch the running transaction
 694  *
 695  * It is similar to the above function, the differentia is this one
 696  * will wait for all the inactive transactions until they fully
 697  * complete.
 698  */
 699 struct btrfs_trans_handle *
 700 btrfs_attach_transaction_barrier(struct btrfs_root *root)
 701 {
 702         struct btrfs_trans_handle *trans;
 703
 704         trans = start_transaction(root, 0, TRANS_ATTACH,
 705                                   BTRFS_RESERVE_NO_FLUSH);
 706         if (IS_ERR(trans) && PTR_ERR(trans) == -ENOENT)
 707                 btrfs_wait_for_commit(root, 0);
 708
 709         return trans;
 710 }
 711
 712 /* wait for a transaction commit to be fully complete */
 713 static noinline void wait_for_commit(struct btrfs_root *root,
 714                                     struct btrfs_transaction *commit)
 715 {
 716         wait_event(commit->commit_wait, commit->state == TRANS_STATE_COMPLETED);
 717 }
 718
 719 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 720 {
 721         struct btrfs_transaction *cur_trans = NULL, *t;
 722         int ret = 0;
 723
 724         if (transid) {
 725                 if (transid <= root->fs_info->last_trans_committed)
 726                         goto out;
 727
 728                 /* find specified transaction */
 729                 spin_lock(&root->fs_info->trans_lock);
 730                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 731                         if (t->transid == transid) {
 732                                 cur_trans = t;
 733                                 atomic_inc(&cur_trans->use_count);
 734                                 ret = 0;
 735                                 break;
 736                         }
 737                         if (t->transid > transid) {
 738                                 ret = 0;
 739                                 break;
 740                         }
 741                 }
 742                 spin_unlock(&root->fs_info->trans_lock);
 743
 744                 /*
 745                  * The specified transaction doesn't exist, or we
 746                  * raced with btrfs_commit_transaction
 747                  */
 748                 if (!cur_trans) {
 749                         if (transid > root->fs_info->last_trans_committed)
 750                                 ret = -EINVAL;
 751                         goto out;
 752                 }
 753         } else {
 754                 /* find newest transaction that is committing | committed */
 755                 spin_lock(&root->fs_info->trans_lock);
 756                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 757                                             list) {
 758                         if (t->state >= TRANS_STATE_COMMIT_START) {
 759                                 if (t->state == TRANS_STATE_COMPLETED)
 760                                         break;
 761                                 cur_trans = t;
 762                                 atomic_inc(&cur_trans->use_count);
 763                                 break;
 764                         }
 765                 }
 766                 spin_unlock(&root->fs_info->trans_lock);
 767                 if (!cur_trans)
 768                         goto out;  /* nothing committing|committed */
 769         }
 770
 771         wait_for_commit(root, cur_trans);
 772         btrfs_put_transaction(cur_trans);
 773 out:
 774         return ret;
 775 }
 776
 777 void btrfs_throttle(struct btrfs_root *root)
 778 {
 779         if (!atomic_read(&root->fs_info->open_ioctl_trans))
 780                 wait_current_trans(root);
 781 }
 782
 783 static int should_end_transaction(struct btrfs_trans_handle *trans,
 784                                   struct btrfs_root *root)
 785 {
 786         if (root->fs_info->global_block_rsv.space_info->full &&
 787             btrfs_check_space_for_delayed_refs(trans, root))
 788                 return 1;
 789
 790         return !!btrfs_block_rsv_check(root, &root->fs_info->global_block_rsv, 5);
 791 }
 792
 793 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 794                                  struct btrfs_root *root)
 795 {
 796         struct btrfs_transaction *cur_trans = trans->transaction;
 797         int updates;
 798         int err;
 799
 800         smp_mb();
 801         if (cur_trans->state >= TRANS_STATE_BLOCKED ||
 802             cur_trans->delayed_refs.flushing)
 803                 return 1;
 804
 805         updates = trans->delayed_ref_updates;
 806         trans->delayed_ref_updates = 0;
 807         if (updates) {
 808                 err = btrfs_run_delayed_refs(trans, root, updates * 2);
 809                 if (err) /* Error code will also eval true */
 810                         return err;
 811         }
 812
 813         return should_end_transaction(trans, root);
 814 }
 815
 816 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 817                           struct btrfs_root *root, int throttle)
 818 {
 819         struct btrfs_transaction *cur_trans = trans->transaction;
 820         struct btrfs_fs_info *info = root->fs_info;
 821         u64 transid = trans->transid;
 822         unsigned long cur = trans->delayed_ref_updates;
 823         int lock = (trans->type != TRANS_JOIN_NOLOCK);
 824         int err = 0;
 825         int must_run_delayed_refs = 0;
 826
 827         if (trans->use_count > 1) {
 828                 trans->use_count--;
 829                 trans->block_rsv = trans->orig_rsv;
 830                 return 0;
 831         }
 832
 833         btrfs_trans_release_metadata(trans, root);
 834         trans->block_rsv = NULL;
 835
 836         if (!list_empty(&trans->new_bgs))
 837                 btrfs_create_pending_block_groups(trans, root);
 838
 839         trans->delayed_ref_updates = 0;
 840         if (!trans->sync) {
 841                 must_run_delayed_refs =
 842                         btrfs_should_throttle_delayed_refs(trans, root);
 843                 cur = max_t(unsigned long, cur, 32);
 844
 845                 /*
 846                  * don't make the caller wait if they are from a NOLOCK
 847                  * or ATTACH transaction, it will deadlock with commit
 848                  */
 849                 if (must_run_delayed_refs == 1 &&
 850                     (trans->type & (__TRANS_JOIN_NOLOCK | __TRANS_ATTACH)))
 851                         must_run_delayed_refs = 2;
 852         }
 853
 854         btrfs_trans_release_metadata(trans, root);
 855         trans->block_rsv = NULL;
 856
 857         if (!list_empty(&trans->new_bgs))
 858                 btrfs_create_pending_block_groups(trans, root);
 859
 860         btrfs_trans_release_chunk_metadata(trans);
 861
 862         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
 863             should_end_transaction(trans, root) &&
 864             ACCESS_ONCE(cur_trans->state) == TRANS_STATE_RUNNING) {
 865                 spin_lock(&info->trans_lock);
 866                 if (cur_trans->state == TRANS_STATE_RUNNING)
 867                         cur_trans->state = TRANS_STATE_BLOCKED;
 868                 spin_unlock(&info->trans_lock);
 869         }
 870
 871         if (lock && ACCESS_ONCE(cur_trans->state) == TRANS_STATE_BLOCKED) {
 872                 if (throttle)
 873                         return btrfs_commit_transaction(trans, root);
 874                 else
 875                         wake_up_process(info->transaction_kthread);
 876         }
 877
 878         if (trans->type & __TRANS_FREEZABLE)
 879                 sb_end_intwrite(root->fs_info->sb);
 880
 881         WARN_ON(cur_trans != info->running_transaction);
 882         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 883         atomic_dec(&cur_trans->num_writers);
 884         extwriter_counter_dec(cur_trans, trans->type);
 885
 886         /*
 887          * Make sure counter is updated before we wake up waiters.
 888          */
 889         smp_mb();
 890         if (waitqueue_active(&cur_trans->writer_wait))
 891                 wake_up(&cur_trans->writer_wait);
 892         btrfs_put_transaction(cur_trans);
 893
 894         if (current->journal_info == trans)
 895                 current->journal_info = NULL;
 896
 897         if (throttle)
 898                 btrfs_run_delayed_iputs(root);
 899
 900         if (trans->aborted ||
 901             test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state)) {
 902                 wake_up_process(info->transaction_kthread);
 903                 err = -EIO;
 904         }
 905         assert_qgroups_uptodate(trans);
 906
 907         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 908         if (must_run_delayed_refs) {
 909                 btrfs_async_run_delayed_refs(root, cur, transid,
 910                                              must_run_delayed_refs == 1);
 911         }
 912         return err;
 913 }
 914
 915 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 916                           struct btrfs_root *root)
 917 {
 918         return __btrfs_end_transaction(trans, root, 0);
 919 }
 920
 921 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 922                                    struct btrfs_root *root)
 923 {
 924         return __btrfs_end_transaction(trans, root, 1);
 925 }
 926
 927 /*
 928  * when btree blocks are allocated, they have some corresponding bits set for
 929  * them in one of two extent_io trees.  This is used to make sure all of
 930  * those extents are sent to disk but does not wait on them
 931  */
 932 int btrfs_write_marked_extents(struct btrfs_root *root,
 933                                struct extent_io_tree *dirty_pages, int mark)
 934 {
 935         int err = 0;
 936         int werr = 0;
 937         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 938         struct extent_state *cached_state = NULL;
 939         u64 start = 0;
 940         u64 end;
 941
 942         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 943                                       mark, &cached_state)) {
 944                 bool wait_writeback = false;
 945
 946                 err = convert_extent_bit(dirty_pages, start, end,
 947                                          EXTENT_NEED_WAIT,
 948                                          mark, &cached_state);
 949                 /*
 950                  * convert_extent_bit can return -ENOMEM, which is most of the
 951                  * time a temporary error. So when it happens, ignore the error
 952                  * and wait for writeback of this range to finish - because we
 953                  * failed to set the bit EXTENT_NEED_WAIT for the range, a call
 954                  * to btrfs_wait_marked_extents() would not know that writeback
 955                  * for this range started and therefore wouldn't wait for it to
 956                  * finish - we don't want to commit a superblock that points to
 957                  * btree nodes/leafs for which writeback hasn't finished yet
 958                  * (and without errors).
 959                  * We cleanup any entries left in the io tree when committing
 960                  * the transaction (through clear_btree_io_tree()).
 961                  */
 962                 if (err == -ENOMEM) {
 963                         err = 0;
 964                         wait_writeback = true;
 965                 }
 966                 if (!err)
 967                         err = filemap_fdatawrite_range(mapping, start, end);
 968                 if (err)
 969                         werr = err;
 970                 else if (wait_writeback)
 971                         werr = filemap_fdatawait_range(mapping, start, end);
 972                 free_extent_state(cached_state);
 973                 cached_state = NULL;
 974                 cond_resched();
 975                 start = end + 1;
 976         }
 977         return werr;
 978 }
 979
 980 /*
 981  * when btree blocks are allocated, they have some corresponding bits set for
 982  * them in one of two extent_io trees.  This is used to make sure all of
 983  * those extents are on disk for transaction or log commit.  We wait
 984  * on all the pages and clear them from the dirty pages state tree
 985  */
 986 int btrfs_wait_marked_extents(struct btrfs_root *root,
 987                               struct extent_io_tree *dirty_pages, int mark)
 988 {
 989         int err = 0;
 990         int werr = 0;
 991         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 992         struct extent_state *cached_state = NULL;
 993         u64 start = 0;
 994         u64 end;
 995         struct btrfs_inode *btree_ino = BTRFS_I(root->fs_info->btree_inode);
 996         bool errors = false;
 997
 998         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 999                                       EXTENT_NEED_WAIT, &cached_state)) {
1000                 /*
1001                  * Ignore -ENOMEM errors returned by clear_extent_bit().
1002                  * When committing the transaction, we'll remove any entries
1003                  * left in the io tree. For a log commit, we don't remove them
1004                  * after committing the log because the tree can be accessed
1005                  * concurrently - we do it only at transaction commit time when
1006                  * it's safe to do it (through clear_btree_io_tree()).
1007                  */
1008                 err = clear_extent_bit(dirty_pages, start, end,
1009                                        EXTENT_NEED_WAIT,
1010                                        0, 0, &cached_state, GFP_NOFS);
1011                 if (err == -ENOMEM)
1012                         err = 0;
1013                 if (!err)
1014                         err = filemap_fdatawait_range(mapping, start, end);
1015                 if (err)
1016                         werr = err;
1017                 free_extent_state(cached_state);
1018                 cached_state = NULL;
1019                 cond_resched();
1020                 start = end + 1;
1021         }
1022         if (err)
1023                 werr = err;
1024
1025         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
1026                 if ((mark & EXTENT_DIRTY) &&
1027                     test_and_clear_bit(BTRFS_INODE_BTREE_LOG1_ERR,
1028                                        &btree_ino->runtime_flags))
1029                         errors = true;
1030
1031                 if ((mark & EXTENT_NEW) &&
1032                     test_and_clear_bit(BTRFS_INODE_BTREE_LOG2_ERR,
1033                                        &btree_ino->runtime_flags))
1034                         errors = true;
1035         } else {
1036                 if (test_and_clear_bit(BTRFS_INODE_BTREE_ERR,
1037                                        &btree_ino->runtime_flags))
1038                         errors = true;
1039         }
1040
1041         if (errors && !werr)
1042                 werr = -EIO;
1043
1044         return werr;
1045 }
1046
1047 /*
1048  * when btree blocks are allocated, they have some corresponding bits set for
1049  * them in one of two extent_io trees.  This is used to make sure all of
1050  * those extents are on disk for transaction or log commit
1051  */
1052 static int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
1053                                 struct extent_io_tree *dirty_pages, int mark)
1054 {
1055         int ret;
1056         int ret2;
1057         struct blk_plug plug;
1058
1059         blk_start_plug(&plug);
1060         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
1061         blk_finish_plug(&plug);
1062         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
1063
1064         if (ret)
1065                 return ret;
1066         if (ret2)
1067                 return ret2;
1068         return 0;
1069 }
1070
1071 static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
1072                                      struct btrfs_root *root)
1073 {
1074         int ret;
1075
1076         ret = btrfs_write_and_wait_marked_extents(root,
1077                                            &trans->transaction->dirty_pages,
1078                                            EXTENT_DIRTY);
1079         clear_btree_io_tree(&trans->transaction->dirty_pages);
1080
1081         return ret;
1082 }
1083
1084 /*
1085  * this is used to update the root pointer in the tree of tree roots.
1086  *
1087  * But, in the case of the extent allocation tree, updating the root
1088  * pointer may allocate blocks which may change the root of the extent
1089  * allocation tree.
1090  *
1091  * So, this loops and repeats and makes sure the cowonly root didn't
1092  * change while the root pointer was being updated in the metadata.
1093  */
1094 static int update_cowonly_root(struct btrfs_trans_handle *trans,
1095                                struct btrfs_root *root)
1096 {
1097         int ret;
1098         u64 old_root_bytenr;
1099         u64 old_root_used;
1100         struct btrfs_root *tree_root = root->fs_info->tree_root;
1101
1102         old_root_used = btrfs_root_used(&root->root_item);
1103
1104         while (1) {
1105                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1106                 if (old_root_bytenr == root->node->start &&
1107                     old_root_used == btrfs_root_used(&root->root_item))
1108                         break;
1109
1110                 btrfs_set_root_node(&root->root_item, root->node);
1111                 ret = btrfs_update_root(trans, tree_root,
1112                                         &root->root_key,
1113                                         &root->root_item);
1114                 if (ret)
1115                         return ret;
1116
1117                 old_root_used = btrfs_root_used(&root->root_item);
1118         }
1119
1120         return 0;
1121 }
1122
1123 /*
1124  * update all the cowonly tree roots on disk
1125  *
1126  * The error handling in this function may not be obvious. Any of the
1127  * failures will cause the file system to go offline. We still need
1128  * to clean up the delayed refs.
1129  */
1130 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
1131                                          struct btrfs_root *root)
1132 {
1133         struct btrfs_fs_info *fs_info = root->fs_info;
1134         struct list_head *dirty_bgs = &trans->transaction->dirty_bgs;
1135         struct list_head *io_bgs = &trans->transaction->io_bgs;
1136         struct list_head *next;
1137         struct extent_buffer *eb;
1138         int ret;
1139
1140         eb = btrfs_lock_root_node(fs_info->tree_root);
1141         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
1142                               0, &eb);
1143         btrfs_tree_unlock(eb);
1144         free_extent_buffer(eb);
1145
1146         if (ret)
1147                 return ret;
1148
1149         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1150         if (ret)
1151                 return ret;
1152
1153         ret = btrfs_run_dev_stats(trans, root->fs_info);
1154         if (ret)
1155                 return ret;
1156         ret = btrfs_run_dev_replace(trans, root->fs_info);
1157         if (ret)
1158                 return ret;
1159         ret = btrfs_run_qgroups(trans, root->fs_info);
1160         if (ret)
1161                 return ret;
1162
1163         ret = btrfs_setup_space_cache(trans, root);
1164         if (ret)
1165                 return ret;
1166
1167         /* run_qgroups might have added some more refs */
1168         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1169         if (ret)
1170                 return ret;
1171 again:
1172         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1173                 next = fs_info->dirty_cowonly_roots.next;
1174                 list_del_init(next);
1175                 root = list_entry(next, struct btrfs_root, dirty_list);
1176                 clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1177
1178                 if (root != fs_info->extent_root)
1179                         list_add_tail(&root->dirty_list,
1180                                       &trans->transaction->switch_commits);
1181                 ret = update_cowonly_root(trans, root);
1182                 if (ret)
1183                         return ret;
1184                 ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1185                 if (ret)
1186                         return ret;
1187         }
1188
1189         while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1190                 ret = btrfs_write_dirty_block_groups(trans, root);
1191                 if (ret)
1192                         return ret;
1193                 ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1194                 if (ret)
1195                         return ret;
1196         }
1197
1198         if (!list_empty(&fs_info->dirty_cowonly_roots))
1199                 goto again;
1200
1201         list_add_tail(&fs_info->extent_root->dirty_list,
1202                       &trans->transaction->switch_commits);
1203         btrfs_after_dev_replace_commit(fs_info);
1204
1205         return 0;
1206 }
1207
1208 /*
1209  * dead roots are old snapshots that need to be deleted.  This allocates
1210  * a dirty root struct and adds it into the list of dead roots that need to
1211  * be deleted
1212  */
1213 void btrfs_add_dead_root(struct btrfs_root *root)
1214 {
1215         spin_lock(&root->fs_info->trans_lock);
1216         if (list_empty(&root->root_list))
1217                 list_add_tail(&root->root_list, &root->fs_info->dead_roots);
1218         spin_unlock(&root->fs_info->trans_lock);
1219 }
1220
1221 /*
1222  * update all the cowonly tree roots on disk
1223  */
1224 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
1225                                     struct btrfs_root *root)
1226 {
1227         struct btrfs_root *gang[8];
1228         struct btrfs_fs_info *fs_info = root->fs_info;
1229         int i;
1230         int ret;
1231         int err = 0;
1232
1233         spin_lock(&fs_info->fs_roots_radix_lock);
1234         while (1) {
1235                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1236                                                  (void **)gang, 0,
1237                                                  ARRAY_SIZE(gang),
1238                                                  BTRFS_ROOT_TRANS_TAG);
1239                 if (ret == 0)
1240                         break;
1241                 for (i = 0; i < ret; i++) {
1242                         root = gang[i];
1243                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
1244                                         (unsigned long)root->root_key.objectid,
1245                                         BTRFS_ROOT_TRANS_TAG);
1246                         spin_unlock(&fs_info->fs_roots_radix_lock);
1247
1248                         btrfs_free_log(trans, root);
1249                         btrfs_update_reloc_root(trans, root);
1250                         btrfs_orphan_commit_root(trans, root);
1251
1252                         btrfs_save_ino_cache(root, trans);
1253
1254                         /* see comments in should_cow_block() */
1255                         clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1256                         smp_mb__after_atomic();
1257
1258                         if (root->commit_root != root->node) {
1259                                 list_add_tail(&root->dirty_list,
1260                                         &trans->transaction->switch_commits);
1261                                 btrfs_set_root_node(&root->root_item,
1262                                                     root->node);
1263                         }
1264
1265                         err = btrfs_update_root(trans, fs_info->tree_root,
1266                                                 &root->root_key,
1267                                                 &root->root_item);
1268                         spin_lock(&fs_info->fs_roots_radix_lock);
1269                         if (err)
1270                                 break;
1271                         btrfs_qgroup_free_meta_all(root);
1272                 }
1273         }
1274         spin_unlock(&fs_info->fs_roots_radix_lock);
1275         return err;
1276 }
1277
1278 /*
1279  * defrag a given btree.
1280  * Every leaf in the btree is read and defragged.
1281  */
1282 int btrfs_defrag_root(struct btrfs_root *root)
1283 {
1284         struct btrfs_fs_info *info = root->fs_info;
1285         struct btrfs_trans_handle *trans;
1286         int ret;
1287
1288         if (test_and_set_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state))
1289                 return 0;
1290
1291         while (1) {
1292                 trans = btrfs_start_transaction(root, 0);
1293                 if (IS_ERR(trans))
1294                         return PTR_ERR(trans);
1295
1296                 ret = btrfs_defrag_leaves(trans, root);
1297
1298                 btrfs_end_transaction(trans, root);
1299                 btrfs_btree_balance_dirty(info->tree_root);
1300                 cond_resched();
1301
1302                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
1303                         break;
1304
1305                 if (btrfs_defrag_cancelled(root->fs_info)) {
1306                         pr_debug("BTRFS: defrag_root cancelled\n");
1307                         ret = -EAGAIN;
1308                         break;
1309                 }
1310         }
1311         clear_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state);
1312         return ret;
1313 }
1314
1315 /*
1316  * Do all special snapshot related qgroup dirty hack.
1317  *
1318  * Will do all needed qgroup inherit and dirty hack like switch commit
1319  * roots inside one transaction and write all btree into disk, to make
1320  * qgroup works.
1321  */
1322 static int qgroup_account_snapshot(struct btrfs_trans_handle *trans,
1323                                    struct btrfs_root *src,
1324                                    struct btrfs_root *parent,
1325                                    struct btrfs_qgroup_inherit *inherit,
1326                                    u64 dst_objectid)
1327 {
1328         struct btrfs_fs_info *fs_info = src->fs_info;
1329         int ret;
1330
1331         /*
1332          * Save some performance in the case that qgroups are not
1333          * enabled. If this check races with the ioctl, rescan will
1334          * kick in anyway.
1335          */
1336         mutex_lock(&fs_info->qgroup_ioctl_lock);
1337         if (!fs_info->quota_enabled) {
1338                 mutex_unlock(&fs_info->qgroup_ioctl_lock);
1339                 return 0;
1340         }
1341         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1342
1343         /*
1344          * We are going to commit transaction, see btrfs_commit_transaction()
1345          * comment for reason locking tree_log_mutex
1346          */
1347         mutex_lock(&fs_info->tree_log_mutex);
1348
1349         ret = commit_fs_roots(trans, src);
1350         if (ret)
1351                 goto out;
1352         ret = btrfs_qgroup_prepare_account_extents(trans, fs_info);
1353         if (ret < 0)
1354                 goto out;
1355         ret = btrfs_qgroup_account_extents(trans, fs_info);
1356         if (ret < 0)
1357                 goto out;
1358
1359         /* Now qgroup are all updated, we can inherit it to new qgroups */
1360         ret = btrfs_qgroup_inherit(trans, fs_info,
1361                                    src->root_key.objectid, dst_objectid,
1362                                    inherit);
1363         if (ret < 0)
1364                 goto out;
1365
1366         /*
1367          * Now we do a simplified commit transaction, which will:
1368          * 1) commit all subvolume and extent tree
1369          *    To ensure all subvolume and extent tree have a valid
1370          *    commit_root to accounting later insert_dir_item()
1371          * 2) write all btree blocks onto disk
1372          *    This is to make sure later btree modification will be cowed
1373          *    Or commit_root can be populated and cause wrong qgroup numbers
1374          * In this simplified commit, we don't really care about other trees
1375          * like chunk and root tree, as they won't affect qgroup.
1376          * And we don't write super to avoid half committed status.
1377          */
1378         ret = commit_cowonly_roots(trans, src);
1379         if (ret)
1380                 goto out;
1381         switch_commit_roots(trans->transaction, fs_info);
1382         ret = btrfs_write_and_wait_transaction(trans, src);
1383         if (ret)
1384                 btrfs_handle_fs_error(fs_info, ret,
1385                         "Error while writing out transaction for qgroup");
1386
1387 out:
1388         mutex_unlock(&fs_info->tree_log_mutex);
1389
1390         /*
1391          * Force parent root to be updated, as we recorded it before so its
1392          * last_trans == cur_transid.
1393          * Or it won't be committed again onto disk after later
1394          * insert_dir_item()
1395          */
1396         if (!ret)
1397                 record_root_in_trans(trans, parent, 1);
1398         return ret;
1399 }
1400
1401 /*
1402  * new snapshots need to be created at a very specific time in the
1403  * transaction commit.  This does the actual creation.
1404  *
1405  * Note:
1406  * If the error which may affect the commitment of the current transaction
1407  * happens, we should return the error number. If the error which just affect
1408  * the creation of the pending snapshots, just return 0.
1409  */
1410 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1411                                    struct btrfs_fs_info *fs_info,
1412                                    struct btrfs_pending_snapshot *pending)
1413 {
1414         struct btrfs_key key;
1415         struct btrfs_root_item *new_root_item;
1416         struct btrfs_root *tree_root = fs_info->tree_root;
1417         struct btrfs_root *root = pending->root;
1418         struct btrfs_root *parent_root;
1419         struct btrfs_block_rsv *rsv;
1420         struct inode *parent_inode;
1421         struct btrfs_path *path;
1422         struct btrfs_dir_item *dir_item;
1423         struct dentry *dentry;
1424         struct extent_buffer *tmp;
1425         struct extent_buffer *old;
1426         struct timespec cur_time;
1427         int ret = 0;
1428         u64 to_reserve = 0;
1429         u64 index = 0;
1430         u64 objectid;
1431         u64 root_flags;
1432         uuid_le new_uuid;
1433
1434         ASSERT(pending->path);
1435         path = pending->path;
1436
1437         ASSERT(pending->root_item);
1438         new_root_item = pending->root_item;
1439
1440         pending->error = btrfs_find_free_objectid(tree_root, &objectid);
1441         if (pending->error)
1442                 goto no_free_objectid;
1443
1444         /*
1445          * Make qgroup to skip current new snapshot's qgroupid, as it is
1446          * accounted by later btrfs_qgroup_inherit().
1447          */
1448         btrfs_set_skip_qgroup(trans, objectid);
1449
1450         btrfs_reloc_pre_snapshot(pending, &to_reserve);
1451
1452         if (to_reserve > 0) {
1453                 pending->error = btrfs_block_rsv_add(root,
1454                                                      &pending->block_rsv,
1455                                                      to_reserve,
1456                                                      BTRFS_RESERVE_NO_FLUSH);
1457                 if (pending->error)
1458                         goto clear_skip_qgroup;
1459         }
1460
1461         key.objectid = objectid;
1462         key.offset = (u64)-1;
1463         key.type = BTRFS_ROOT_ITEM_KEY;
1464
1465         rsv = trans->block_rsv;
1466         trans->block_rsv = &pending->block_rsv;
1467         trans->bytes_reserved = trans->block_rsv->reserved;
1468         trace_btrfs_space_reservation(root->fs_info, "transaction",
1469                                       trans->transid,
1470                                       trans->bytes_reserved, 1);
1471         dentry = pending->dentry;
1472         parent_inode = pending->dir;
1473         parent_root = BTRFS_I(parent_inode)->root;
1474         record_root_in_trans(trans, parent_root, 0);
1475
1476         cur_time = current_fs_time(parent_inode->i_sb);
1477
1478         /*
1479          * insert the directory item
1480          */
1481         ret = btrfs_set_inode_index(parent_inode, &index);
1482         BUG_ON(ret); /* -ENOMEM */
1483
1484         /* check if there is a file/dir which has the same name. */
1485         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1486                                          btrfs_ino(parent_inode),
1487                                          dentry->d_name.name,
1488                                          dentry->d_name.len, 0);
1489         if (dir_item != NULL && !IS_ERR(dir_item)) {
1490                 pending->error = -EEXIST;
1491                 goto dir_item_existed;
1492         } else if (IS_ERR(dir_item)) {
1493                 ret = PTR_ERR(dir_item);
1494                 btrfs_abort_transaction(trans, root, ret);
1495                 goto fail;
1496         }
1497         btrfs_release_path(path);
1498
1499         /*
1500          * pull in the delayed directory update
1501          * and the delayed inode item
1502          * otherwise we corrupt the FS during
1503          * snapshot
1504          */
1505         ret = btrfs_run_delayed_items(trans, root);
1506         if (ret) {      /* Transaction aborted */
1507                 btrfs_abort_transaction(trans, root, ret);
1508                 goto fail;
1509         }
1510
1511         record_root_in_trans(trans, root, 0);
1512         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1513         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1514         btrfs_check_and_init_root_item(new_root_item);
1515
1516         root_flags = btrfs_root_flags(new_root_item);
1517         if (pending->readonly)
1518                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1519         else
1520                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1521         btrfs_set_root_flags(new_root_item, root_flags);
1522
1523         btrfs_set_root_generation_v2(new_root_item,
1524                         trans->transid);
1525         uuid_le_gen(&new_uuid);
1526         memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1527         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1528                         BTRFS_UUID_SIZE);
1529         if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1530                 memset(new_root_item->received_uuid, 0,
1531                        sizeof(new_root_item->received_uuid));
1532                 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1533                 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1534                 btrfs_set_root_stransid(new_root_item, 0);
1535                 btrfs_set_root_rtransid(new_root_item, 0);
1536         }
1537         btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1538         btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1539         btrfs_set_root_otransid(new_root_item, trans->transid);
1540
1541         old = btrfs_lock_root_node(root);
1542         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1543         if (ret) {
1544                 btrfs_tree_unlock(old);
1545                 free_extent_buffer(old);
1546                 btrfs_abort_transaction(trans, root, ret);
1547                 goto fail;
1548         }
1549
1550         btrfs_set_lock_blocking(old);
1551
1552         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1553         /* clean up in any case */
1554         btrfs_tree_unlock(old);
1555         free_extent_buffer(old);
1556         if (ret) {
1557                 btrfs_abort_transaction(trans, root, ret);
1558                 goto fail;
1559         }
1560         /* see comments in should_cow_block() */
1561         set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1562         smp_wmb();
1563
1564         btrfs_set_root_node(new_root_item, tmp);
1565         /* record when the snapshot was created in key.offset */
1566         key.offset = trans->transid;
1567         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1568         btrfs_tree_unlock(tmp);
1569         free_extent_buffer(tmp);
1570         if (ret) {
1571                 btrfs_abort_transaction(trans, root, ret);
1572                 goto fail;
1573         }
1574
1575         /*
1576          * insert root back/forward references
1577          */
1578         ret = btrfs_add_root_ref(trans, tree_root, objectid,
1579                                  parent_root->root_key.objectid,
1580                                  btrfs_ino(parent_inode), index,
1581                                  dentry->d_name.name, dentry->d_name.len);
1582         if (ret) {
1583                 btrfs_abort_transaction(trans, root, ret);
1584                 goto fail;
1585         }
1586
1587         key.offset = (u64)-1;
1588         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1589         if (IS_ERR(pending->snap)) {
1590                 ret = PTR_ERR(pending->snap);
1591                 btrfs_abort_transaction(trans, root, ret);
1592                 goto fail;
1593         }
1594
1595         ret = btrfs_reloc_post_snapshot(trans, pending);
1596         if (ret) {
1597                 btrfs_abort_transaction(trans, root, ret);
1598                 goto fail;
1599         }
1600
1601         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1602         if (ret) {
1603                 btrfs_abort_transaction(trans, root, ret);
1604                 goto fail;
1605         }
1606
1607         /*
1608          * Do special qgroup accounting for snapshot, as we do some qgroup
1609          * snapshot hack to do fast snapshot.
1610          * To co-operate with that hack, we do hack again.
1611          * Or snapshot will be greatly slowed down by a subtree qgroup rescan
1612          */
1613         ret = qgroup_account_snapshot(trans, root, parent_root,
1614                                       pending->inherit, objectid);
1615         if (ret < 0)
1616                 goto fail;
1617
1618         ret = btrfs_insert_dir_item(trans, parent_root,
1619                                     dentry->d_name.name, dentry->d_name.len,
1620                                     parent_inode, &key,
1621                                     BTRFS_FT_DIR, index);
1622         /* We have check then name at the beginning, so it is impossible. */
1623         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1624         if (ret) {
1625                 btrfs_abort_transaction(trans, root, ret);
1626                 goto fail;
1627         }
1628
1629         btrfs_i_size_write(parent_inode, parent_inode->i_size +
1630                                          dentry->d_name.len * 2);
1631         parent_inode->i_mtime = parent_inode->i_ctime =
1632                 current_fs_time(parent_inode->i_sb);
1633         ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1634         if (ret) {
1635                 btrfs_abort_transaction(trans, root, ret);
1636                 goto fail;
1637         }
1638         ret = btrfs_uuid_tree_add(trans, fs_info->uuid_root, new_uuid.b,
1639                                   BTRFS_UUID_KEY_SUBVOL, objectid);
1640         if (ret) {
1641                 btrfs_abort_transaction(trans, root, ret);
1642                 goto fail;
1643         }
1644         if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1645                 ret = btrfs_uuid_tree_add(trans, fs_info->uuid_root,
1646                                           new_root_item->received_uuid,
1647                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1648                                           objectid);
1649                 if (ret && ret != -EEXIST) {
1650                         btrfs_abort_transaction(trans, root, ret);
1651                         goto fail;
1652                 }
1653         }
1654
1655         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1656         if (ret) {
1657                 btrfs_abort_transaction(trans, root, ret);
1658                 goto fail;
1659         }
1660
1661 fail:
1662         pending->error = ret;
1663 dir_item_existed:
1664         trans->block_rsv = rsv;
1665         trans->bytes_reserved = 0;
1666 clear_skip_qgroup:
1667         btrfs_clear_skip_qgroup(trans);
1668 no_free_objectid:
1669         kfree(new_root_item);
1670         pending->root_item = NULL;
1671         btrfs_free_path(path);
1672         pending->path = NULL;
1673
1674         return ret;
1675 }
1676
1677 /*
1678  * create all the snapshots we've scheduled for creation
1679  */
1680 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1681                                              struct btrfs_fs_info *fs_info)
1682 {
1683         struct btrfs_pending_snapshot *pending, *next;
1684         struct list_head *head = &trans->transaction->pending_snapshots;
1685         int ret = 0;
1686
1687         list_for_each_entry_safe(pending, next, head, list) {
1688                 list_del(&pending->list);
1689                 ret = create_pending_snapshot(trans, fs_info, pending);
1690                 if (ret)
1691                         break;
1692         }
1693         return ret;
1694 }
1695
1696 static void update_super_roots(struct btrfs_root *root)
1697 {
1698         struct btrfs_root_item *root_item;
1699         struct btrfs_super_block *super;
1700
1701         super = root->fs_info->super_copy;
1702
1703         root_item = &root->fs_info->chunk_root->root_item;
1704         super->chunk_root = root_item->bytenr;
1705         super->chunk_root_generation = root_item->generation;
1706         super->chunk_root_level = root_item->level;
1707
1708         root_item = &root->fs_info->tree_root->root_item;
1709         super->root = root_item->bytenr;
1710         super->generation = root_item->generation;
1711         super->root_level = root_item->level;
1712         if (btrfs_test_opt(root, SPACE_CACHE))
1713                 super->cache_generation = root_item->generation;
1714         if (root->fs_info->update_uuid_tree_gen)
1715                 super->uuid_tree_generation = root_item->generation;
1716 }
1717
1718 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1719 {
1720         struct btrfs_transaction *trans;
1721         int ret = 0;
1722
1723         spin_lock(&info->trans_lock);
1724         trans = info->running_transaction;
1725         if (trans)
1726                 ret = (trans->state >= TRANS_STATE_COMMIT_START);
1727         spin_unlock(&info->trans_lock);
1728         return ret;
1729 }
1730
1731 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1732 {
1733         struct btrfs_transaction *trans;
1734         int ret = 0;
1735
1736         spin_lock(&info->trans_lock);
1737         trans = info->running_transaction;
1738         if (trans)
1739                 ret = is_transaction_blocked(trans);
1740         spin_unlock(&info->trans_lock);
1741         return ret;
1742 }
1743
1744 /*
1745  * wait for the current transaction commit to start and block subsequent
1746  * transaction joins
1747  */
1748 static void wait_current_trans_commit_start(struct btrfs_root *root,
1749                                             struct btrfs_transaction *trans)
1750 {
1751         wait_event(root->fs_info->transaction_blocked_wait,
1752                    trans->state >= TRANS_STATE_COMMIT_START ||
1753                    trans->aborted);
1754 }
1755
1756 /*
1757  * wait for the current transaction to start and then become unblocked.
1758  * caller holds ref.
1759  */
1760 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1761                                          struct btrfs_transaction *trans)
1762 {
1763         wait_event(root->fs_info->transaction_wait,
1764                    trans->state >= TRANS_STATE_UNBLOCKED ||
1765                    trans->aborted);
1766 }
1767
1768 /*
1769  * commit transactions asynchronously. once btrfs_commit_transaction_async
1770  * returns, any subsequent transaction will not be allowed to join.
1771  */
1772 struct btrfs_async_commit {
1773         struct btrfs_trans_handle *newtrans;
1774         struct btrfs_root *root;
1775         struct work_struct work;
1776 };
1777
1778 static void do_async_commit(struct work_struct *work)
1779 {
1780         struct btrfs_async_commit *ac =
1781                 container_of(work, struct btrfs_async_commit, work);
1782
1783         /*
1784          * We've got freeze protection passed with the transaction.
1785          * Tell lockdep about it.
1786          */
1787         if (ac->newtrans->type & __TRANS_FREEZABLE)
1788                 __sb_writers_acquired(ac->root->fs_info->sb, SB_FREEZE_FS);
1789
1790         current->journal_info = ac->newtrans;
1791
1792         btrfs_commit_transaction(ac->newtrans, ac->root);
1793         kfree(ac);
1794 }
1795
1796 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1797                                    struct btrfs_root *root,
1798                                    int wait_for_unblock)
1799 {
1800         struct btrfs_async_commit *ac;
1801         struct btrfs_transaction *cur_trans;
1802
1803         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1804         if (!ac)
1805                 return -ENOMEM;
1806
1807         INIT_WORK(&ac->work, do_async_commit);
1808         ac->root = root;
1809         ac->newtrans = btrfs_join_transaction(root);
1810         if (IS_ERR(ac->newtrans)) {
1811                 int err = PTR_ERR(ac->newtrans);
1812                 kfree(ac);
1813                 return err;
1814         }
1815
1816         /* take transaction reference */
1817         cur_trans = trans->transaction;
1818         atomic_inc(&cur_trans->use_count);
1819
1820         btrfs_end_transaction(trans, root);
1821
1822         /*
1823          * Tell lockdep we've released the freeze rwsem, since the
1824          * async commit thread will be the one to unlock it.
1825          */
1826         if (ac->newtrans->type & __TRANS_FREEZABLE)
1827                 __sb_writers_release(root->fs_info->sb, SB_FREEZE_FS);
1828
1829         schedule_work(&ac->work);
1830
1831         /* wait for transaction to start and unblock */
1832         if (wait_for_unblock)
1833                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1834         else
1835                 wait_current_trans_commit_start(root, cur_trans);
1836
1837         if (current->journal_info == trans)
1838                 current->journal_info = NULL;
1839
1840         btrfs_put_transaction(cur_trans);
1841         return 0;
1842 }
1843
1844
1845 static void cleanup_transaction(struct btrfs_trans_handle *trans,
1846                                 struct btrfs_root *root, int err)
1847 {
1848         struct btrfs_transaction *cur_trans = trans->transaction;
1849         DEFINE_WAIT(wait);
1850
1851         WARN_ON(trans->use_count > 1);
1852
1853         btrfs_abort_transaction(trans, root, err);
1854
1855         spin_lock(&root->fs_info->trans_lock);
1856
1857         /*
1858          * If the transaction is removed from the list, it means this
1859          * transaction has been committed successfully, so it is impossible
1860          * to call the cleanup function.
1861          */
1862         BUG_ON(list_empty(&cur_trans->list));
1863
1864         list_del_init(&cur_trans->list);
1865         if (cur_trans == root->fs_info->running_transaction) {
1866                 cur_trans->state = TRANS_STATE_COMMIT_DOING;
1867                 spin_unlock(&root->fs_info->trans_lock);
1868                 wait_event(cur_trans->writer_wait,
1869                            atomic_read(&cur_trans->num_writers) == 1);
1870
1871                 spin_lock(&root->fs_info->trans_lock);
1872         }
1873         spin_unlock(&root->fs_info->trans_lock);
1874
1875         btrfs_cleanup_one_transaction(trans->transaction, root);
1876
1877         spin_lock(&root->fs_info->trans_lock);
1878         if (cur_trans == root->fs_info->running_transaction)
1879                 root->fs_info->running_transaction = NULL;
1880         spin_unlock(&root->fs_info->trans_lock);
1881
1882         if (trans->type & __TRANS_FREEZABLE)
1883                 sb_end_intwrite(root->fs_info->sb);
1884         btrfs_put_transaction(cur_trans);
1885         btrfs_put_transaction(cur_trans);
1886
1887         trace_btrfs_transaction_commit(root);
1888
1889         if (current->journal_info == trans)
1890                 current->journal_info = NULL;
1891         btrfs_scrub_cancel(root->fs_info);
1892
1893         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1894 }
1895
1896 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
1897 {
1898         if (btrfs_test_opt(fs_info->tree_root, FLUSHONCOMMIT))
1899                 return btrfs_start_delalloc_roots(fs_info, 1, -1);
1900         return 0;
1901 }
1902
1903 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
1904 {
1905         if (btrfs_test_opt(fs_info->tree_root, FLUSHONCOMMIT))
1906                 btrfs_wait_ordered_roots(fs_info, -1, 0, (u64)-1);
1907 }
1908
1909 static inline void
1910 btrfs_wait_pending_ordered(struct btrfs_transaction *cur_trans)
1911 {
1912         wait_event(cur_trans->pending_wait,
1913                    atomic_read(&cur_trans->pending_ordered) == 0);
1914 }
1915
1916 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1917                              struct btrfs_root *root)
1918 {
1919         struct btrfs_transaction *cur_trans = trans->transaction;
1920         struct btrfs_transaction *prev_trans = NULL;
1921         struct btrfs_inode *btree_ino = BTRFS_I(root->fs_info->btree_inode);
1922         int ret;
1923
1924         /* Stop the commit early if ->aborted is set */
1925         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1926                 ret = cur_trans->aborted;
1927                 btrfs_end_transaction(trans, root);
1928                 return ret;
1929         }
1930
1931         /* make a pass through all the delayed refs we have so far
1932          * any runnings procs may add more while we are here
1933          */
1934         ret = btrfs_run_delayed_refs(trans, root, 0);
1935         if (ret) {
1936                 btrfs_end_transaction(trans, root);
1937                 return ret;
1938         }
1939
1940         btrfs_trans_release_metadata(trans, root);
1941         trans->block_rsv = NULL;
1942
1943         cur_trans = trans->transaction;
1944
1945         /*
1946          * set the flushing flag so procs in this transaction have to
1947          * start sending their work down.
1948          */
1949         cur_trans->delayed_refs.flushing = 1;
1950         smp_wmb();
1951
1952         if (!list_empty(&trans->new_bgs))
1953                 btrfs_create_pending_block_groups(trans, root);
1954
1955         ret = btrfs_run_delayed_refs(trans, root, 0);
1956         if (ret) {
1957                 btrfs_end_transaction(trans, root);
1958                 return ret;
1959         }
1960
1961         if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) {
1962                 int run_it = 0;
1963
1964                 /* this mutex is also taken before trying to set
1965                  * block groups readonly.  We need to make sure
1966                  * that nobody has set a block group readonly
1967                  * after a extents from that block group have been
1968                  * allocated for cache files.  btrfs_set_block_group_ro
1969                  * will wait for the transaction to commit if it
1970                  * finds BTRFS_TRANS_DIRTY_BG_RUN set.
1971                  *
1972                  * The BTRFS_TRANS_DIRTY_BG_RUN flag is also used to make sure
1973                  * only one process starts all the block group IO.  It wouldn't
1974                  * hurt to have more than one go through, but there's no
1975                  * real advantage to it either.
1976                  */
1977                 mutex_lock(&root->fs_info->ro_block_group_mutex);
1978                 if (!test_and_set_bit(BTRFS_TRANS_DIRTY_BG_RUN,
1979                                       &cur_trans->flags))
1980                         run_it = 1;
1981                 mutex_unlock(&root->fs_info->ro_block_group_mutex);
1982
1983                 if (run_it)
1984                         ret = btrfs_start_dirty_block_groups(trans, root);
1985         }
1986         if (ret) {
1987                 btrfs_end_transaction(trans, root);
1988                 return ret;
1989         }
1990
1991         spin_lock(&root->fs_info->trans_lock);
1992         if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
1993                 spin_unlock(&root->fs_info->trans_lock);
1994                 atomic_inc(&cur_trans->use_count);
1995                 ret = btrfs_end_transaction(trans, root);
1996
1997                 wait_for_commit(root, cur_trans);
1998
1999                 if (unlikely(cur_trans->aborted))
2000                         ret = cur_trans->aborted;
2001
2002                 btrfs_put_transaction(cur_trans);
2003
2004                 return ret;
2005         }
2006
2007         cur_trans->state = TRANS_STATE_COMMIT_START;
2008         wake_up(&root->fs_info->transaction_blocked_wait);
2009
2010         if (cur_trans->list.prev != &root->fs_info->trans_list) {
2011                 prev_trans = list_entry(cur_trans->list.prev,
2012                                         struct btrfs_transaction, list);
2013                 if (prev_trans->state != TRANS_STATE_COMPLETED) {
2014                         atomic_inc(&prev_trans->use_count);
2015                         spin_unlock(&root->fs_info->trans_lock);
2016
2017                         wait_for_commit(root, prev_trans);
2018                         ret = prev_trans->aborted;
2019
2020                         btrfs_put_transaction(prev_trans);
2021                         if (ret)
2022                                 goto cleanup_transaction;
2023                 } else {
2024                         spin_unlock(&root->fs_info->trans_lock);
2025                 }
2026         } else {
2027                 spin_unlock(&root->fs_info->trans_lock);
2028         }
2029
2030         extwriter_counter_dec(cur_trans, trans->type);
2031
2032         ret = btrfs_start_delalloc_flush(root->fs_info);
2033         if (ret)
2034                 goto cleanup_transaction;
2035
2036         ret = btrfs_run_delayed_items(trans, root);
2037         if (ret)
2038                 goto cleanup_transaction;
2039
2040         wait_event(cur_trans->writer_wait,
2041                    extwriter_counter_read(cur_trans) == 0);
2042
2043         /* some pending stuffs might be added after the previous flush. */
2044         ret = btrfs_run_delayed_items(trans, root);
2045         if (ret)
2046                 goto cleanup_transaction;
2047
2048         btrfs_wait_delalloc_flush(root->fs_info);
2049
2050         btrfs_wait_pending_ordered(cur_trans);
2051
2052         btrfs_scrub_pause(root);
2053         /*
2054          * Ok now we need to make sure to block out any other joins while we
2055          * commit the transaction.  We could have started a join before setting
2056          * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
2057          */
2058         spin_lock(&root->fs_info->trans_lock);
2059         cur_trans->state = TRANS_STATE_COMMIT_DOING;
2060         spin_unlock(&root->fs_info->trans_lock);
2061         wait_event(cur_trans->writer_wait,
2062                    atomic_read(&cur_trans->num_writers) == 1);
2063
2064         /* ->aborted might be set after the previous check, so check it */
2065         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
2066                 ret = cur_trans->aborted;
2067                 goto scrub_continue;
2068         }
2069         /*
2070          * the reloc mutex makes sure that we stop
2071          * the balancing code from coming in and moving
2072          * extents around in the middle of the commit
2073          */
2074         mutex_lock(&root->fs_info->reloc_mutex);
2075
2076         /*
2077          * We needn't worry about the delayed items because we will
2078          * deal with them in create_pending_snapshot(), which is the
2079          * core function of the snapshot creation.
2080          */
2081         ret = create_pending_snapshots(trans, root->fs_info);
2082         if (ret) {
2083                 mutex_unlock(&root->fs_info->reloc_mutex);
2084                 goto scrub_continue;
2085         }
2086
2087         /*
2088          * We insert the dir indexes of the snapshots and update the inode
2089          * of the snapshots' parents after the snapshot creation, so there
2090          * are some delayed items which are not dealt with. Now deal with
2091          * them.
2092          *
2093          * We needn't worry that this operation will corrupt the snapshots,
2094          * because all the tree which are snapshoted will be forced to COW
2095          * the nodes and leaves.
2096          */
2097         ret = btrfs_run_delayed_items(trans, root);
2098         if (ret) {
2099                 mutex_unlock(&root->fs_info->reloc_mutex);
2100                 goto scrub_continue;
2101         }
2102
2103         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
2104         if (ret) {
2105                 mutex_unlock(&root->fs_info->reloc_mutex);
2106                 goto scrub_continue;
2107         }
2108
2109         /* Reocrd old roots for later qgroup accounting */
2110         ret = btrfs_qgroup_prepare_account_extents(trans, root->fs_info);
2111         if (ret) {
2112                 mutex_unlock(&root->fs_info->reloc_mutex);
2113                 goto scrub_continue;
2114         }
2115
2116         /*
2117          * make sure none of the code above managed to slip in a
2118          * delayed item
2119          */
2120         btrfs_assert_delayed_root_empty(root);
2121
2122         WARN_ON(cur_trans != trans->transaction);
2123
2124         /* btrfs_commit_tree_roots is responsible for getting the
2125          * various roots consistent with each other.  Every pointer
2126          * in the tree of tree roots has to point to the most up to date
2127          * root for every subvolume and other tree.  So, we have to keep
2128          * the tree logging code from jumping in and changing any
2129          * of the trees.
2130          *
2131          * At this point in the commit, there can't be any tree-log
2132          * writers, but a little lower down we drop the trans mutex
2133          * and let new people in.  By holding the tree_log_mutex
2134          * from now until after the super is written, we avoid races
2135          * with the tree-log code.
2136          */
2137         mutex_lock(&root->fs_info->tree_log_mutex);
2138
2139         ret = commit_fs_roots(trans, root);
2140         if (ret) {
2141                 mutex_unlock(&root->fs_info->tree_log_mutex);
2142                 mutex_unlock(&root->fs_info->reloc_mutex);
2143                 goto scrub_continue;
2144         }
2145
2146         /*
2147          * Since the transaction is done, we can apply the pending changes
2148          * before the next transaction.
2149          */
2150         btrfs_apply_pending_changes(root->fs_info);
2151
2152         /* commit_fs_roots gets rid of all the tree log roots, it is now
2153          * safe to free the root of tree log roots
2154          */
2155         btrfs_free_log_root_tree(trans, root->fs_info);
2156
2157         /*
2158          * Since fs roots are all committed, we can get a quite accurate
2159          * new_roots. So let's do quota accounting.
2160          */
2161         ret = btrfs_qgroup_account_extents(trans, root->fs_info);
2162         if (ret < 0) {
2163                 mutex_unlock(&root->fs_info->tree_log_mutex);
2164                 mutex_unlock(&root->fs_info->reloc_mutex);
2165                 goto scrub_continue;
2166         }
2167
2168         ret = commit_cowonly_roots(trans, root);
2169         if (ret) {
2170                 mutex_unlock(&root->fs_info->tree_log_mutex);
2171                 mutex_unlock(&root->fs_info->reloc_mutex);
2172                 goto scrub_continue;
2173         }
2174
2175         /*
2176          * The tasks which save the space cache and inode cache may also
2177          * update ->aborted, check it.
2178          */
2179         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
2180                 ret = cur_trans->aborted;
2181                 mutex_unlock(&root->fs_info->tree_log_mutex);
2182                 mutex_unlock(&root->fs_info->reloc_mutex);
2183                 goto scrub_continue;
2184         }
2185
2186         btrfs_prepare_extent_commit(trans, root);
2187
2188         cur_trans = root->fs_info->running_transaction;
2189
2190         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
2191                             root->fs_info->tree_root->node);
2192         list_add_tail(&root->fs_info->tree_root->dirty_list,
2193                       &cur_trans->switch_commits);
2194
2195         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
2196                             root->fs_info->chunk_root->node);
2197         list_add_tail(&root->fs_info->chunk_root->dirty_list,
2198                       &cur_trans->switch_commits);
2199
2200         switch_commit_roots(cur_trans, root->fs_info);
2201
2202         assert_qgroups_uptodate(trans);
2203         ASSERT(list_empty(&cur_trans->dirty_bgs));
2204         ASSERT(list_empty(&cur_trans->io_bgs));
2205         update_super_roots(root);
2206
2207         btrfs_set_super_log_root(root->fs_info->super_copy, 0);
2208         btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
2209         memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
2210                sizeof(*root->fs_info->super_copy));
2211
2212         btrfs_update_commit_device_size(root->fs_info);
2213         btrfs_update_commit_device_bytes_used(root, cur_trans);
2214
2215         clear_bit(BTRFS_INODE_BTREE_LOG1_ERR, &btree_ino->runtime_flags);
2216         clear_bit(BTRFS_INODE_BTREE_LOG2_ERR, &btree_ino->runtime_flags);
2217
2218         btrfs_trans_release_chunk_metadata(trans);
2219
2220         spin_lock(&root->fs_info->trans_lock);
2221         cur_trans->state = TRANS_STATE_UNBLOCKED;
2222         root->fs_info->running_transaction = NULL;
2223         spin_unlock(&root->fs_info->trans_lock);
2224         mutex_unlock(&root->fs_info->reloc_mutex);
2225
2226         wake_up(&root->fs_info->transaction_wait);
2227
2228         ret = btrfs_write_and_wait_transaction(trans, root);
2229         if (ret) {
2230                 btrfs_handle_fs_error(root->fs_info, ret,
2231                             "Error while writing out transaction");
2232                 mutex_unlock(&root->fs_info->tree_log_mutex);
2233                 goto scrub_continue;
2234         }
2235
2236         ret = write_ctree_super(trans, root, 0);
2237         if (ret) {
2238                 mutex_unlock(&root->fs_info->tree_log_mutex);
2239                 goto scrub_continue;
2240         }
2241
2242         /*
2243          * the super is written, we can safely allow the tree-loggers
2244          * to go about their business
2245          */
2246         mutex_unlock(&root->fs_info->tree_log_mutex);
2247
2248         btrfs_finish_extent_commit(trans, root);
2249
2250         if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &cur_trans->flags))
2251                 btrfs_clear_space_info_full(root->fs_info);
2252
2253         root->fs_info->last_trans_committed = cur_trans->transid;
2254         /*
2255          * We needn't acquire the lock here because there is no other task
2256          * which can change it.
2257          */
2258         cur_trans->state = TRANS_STATE_COMPLETED;
2259         wake_up(&cur_trans->commit_wait);
2260
2261         spin_lock(&root->fs_info->trans_lock);
2262         list_del_init(&cur_trans->list);
2263         spin_unlock(&root->fs_info->trans_lock);
2264
2265         btrfs_put_transaction(cur_trans);
2266         btrfs_put_transaction(cur_trans);
2267
2268         if (trans->type & __TRANS_FREEZABLE)
2269                 sb_end_intwrite(root->fs_info->sb);
2270
2271         trace_btrfs_transaction_commit(root);
2272
2273         btrfs_scrub_continue(root);
2274
2275         if (current->journal_info == trans)
2276                 current->journal_info = NULL;
2277
2278         kmem_cache_free(btrfs_trans_handle_cachep, trans);
2279
2280         if (current != root->fs_info->transaction_kthread &&
2281             current != root->fs_info->cleaner_kthread)
2282                 btrfs_run_delayed_iputs(root);
2283
2284         return ret;
2285
2286 scrub_continue:
2287         btrfs_scrub_continue(root);
2288 cleanup_transaction:
2289         btrfs_trans_release_metadata(trans, root);
2290         btrfs_trans_release_chunk_metadata(trans);
2291         trans->block_rsv = NULL;
2292         btrfs_warn(root->fs_info, "Skipping commit of aborted transaction.");
2293         if (current->journal_info == trans)
2294                 current->journal_info = NULL;
2295         cleanup_transaction(trans, root, ret);
2296
2297         return ret;
2298 }
2299
2300 /*
2301  * return < 0 if error
2302  * 0 if there are no more dead_roots at the time of call
2303  * 1 there are more to be processed, call me again
2304  *
2305  * The return value indicates there are certainly more snapshots to delete, but
2306  * if there comes a new one during processing, it may return 0. We don't mind,
2307  * because btrfs_commit_super will poke cleaner thread and it will process it a
2308  * few seconds later.
2309  */
2310 int btrfs_clean_one_deleted_snapshot(struct btrfs_root *root)
2311 {
2312         int ret;
2313         struct btrfs_fs_info *fs_info = root->fs_info;
2314
2315         spin_lock(&fs_info->trans_lock);
2316         if (list_empty(&fs_info->dead_roots)) {
2317                 spin_unlock(&fs_info->trans_lock);
2318                 return 0;
2319         }
2320         root = list_first_entry(&fs_info->dead_roots,
2321                         struct btrfs_root, root_list);
2322         list_del_init(&root->root_list);
2323         spin_unlock(&fs_info->trans_lock);
2324
2325         pr_debug("BTRFS: cleaner removing %llu\n", root->objectid);
2326
2327         btrfs_kill_all_delayed_nodes(root);
2328
2329         if (btrfs_header_backref_rev(root->node) <
2330                         BTRFS_MIXED_BACKREF_REV)
2331                 ret = btrfs_drop_snapshot(root, NULL, 0, 0);
2332         else
2333                 ret = btrfs_drop_snapshot(root, NULL, 1, 0);
2334
2335         return (ret < 0) ? 0 : 1;
2336 }
2337
2338 void btrfs_apply_pending_changes(struct btrfs_fs_info *fs_info)
2339 {
2340         unsigned long prev;
2341         unsigned long bit;
2342
2343         prev = xchg(&fs_info->pending_changes, 0);
2344         if (!prev)
2345                 return;
2346
2347         bit = 1 << BTRFS_PENDING_SET_INODE_MAP_CACHE;
2348         if (prev & bit)
2349                 btrfs_set_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2350         prev &= ~bit;
2351
2352         bit = 1 << BTRFS_PENDING_CLEAR_INODE_MAP_CACHE;
2353         if (prev & bit)
2354                 btrfs_clear_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2355         prev &= ~bit;
2356
2357         bit = 1 << BTRFS_PENDING_COMMIT;
2358         if (prev & bit)
2359                 btrfs_debug(fs_info, "pending commit done");
2360         prev &= ~bit;
2361
2362         if (prev)
2363                 btrfs_warn(fs_info,
2364                         "unknown pending changes left 0x%lx, ignoring", prev);
2365 }