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         unsigned long cur = trans->delayed_ref_updates;
 822         int lock = (trans->type != TRANS_JOIN_NOLOCK);
 823         int err = 0;
 824         int must_run_delayed_refs = 0;
 825
 826         if (trans->use_count > 1) {
 827                 trans->use_count--;
 828                 trans->block_rsv = trans->orig_rsv;
 829                 return 0;
 830         }
 831
 832         btrfs_trans_release_metadata(trans, root);
 833         trans->block_rsv = NULL;
 834
 835         if (!list_empty(&trans->new_bgs))
 836                 btrfs_create_pending_block_groups(trans, root);
 837
 838         trans->delayed_ref_updates = 0;
 839         if (!trans->sync) {
 840                 must_run_delayed_refs =
 841                         btrfs_should_throttle_delayed_refs(trans, root);
 842                 cur = max_t(unsigned long, cur, 32);
 843
 844                 /*
 845                  * don't make the caller wait if they are from a NOLOCK
 846                  * or ATTACH transaction, it will deadlock with commit
 847                  */
 848                 if (must_run_delayed_refs == 1 &&
 849                     (trans->type & (__TRANS_JOIN_NOLOCK | __TRANS_ATTACH)))
 850                         must_run_delayed_refs = 2;
 851         }
 852
 853         btrfs_trans_release_metadata(trans, root);
 854         trans->block_rsv = NULL;
 855
 856         if (!list_empty(&trans->new_bgs))
 857                 btrfs_create_pending_block_groups(trans, root);
 858
 859         btrfs_trans_release_chunk_metadata(trans);
 860
 861         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
 862             should_end_transaction(trans, root) &&
 863             ACCESS_ONCE(cur_trans->state) == TRANS_STATE_RUNNING) {
 864                 spin_lock(&info->trans_lock);
 865                 if (cur_trans->state == TRANS_STATE_RUNNING)
 866                         cur_trans->state = TRANS_STATE_BLOCKED;
 867                 spin_unlock(&info->trans_lock);
 868         }
 869
 870         if (lock && ACCESS_ONCE(cur_trans->state) == TRANS_STATE_BLOCKED) {
 871                 if (throttle)
 872                         return btrfs_commit_transaction(trans, root);
 873                 else
 874                         wake_up_process(info->transaction_kthread);
 875         }
 876
 877         if (trans->type & __TRANS_FREEZABLE)
 878                 sb_end_intwrite(root->fs_info->sb);
 879
 880         WARN_ON(cur_trans != info->running_transaction);
 881         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 882         atomic_dec(&cur_trans->num_writers);
 883         extwriter_counter_dec(cur_trans, trans->type);
 884
 885         /*
 886          * Make sure counter is updated before we wake up waiters.
 887          */
 888         smp_mb();
 889         if (waitqueue_active(&cur_trans->writer_wait))
 890                 wake_up(&cur_trans->writer_wait);
 891         btrfs_put_transaction(cur_trans);
 892
 893         if (current->journal_info == trans)
 894                 current->journal_info = NULL;
 895
 896         if (throttle)
 897                 btrfs_run_delayed_iputs(root);
 898
 899         if (trans->aborted ||
 900             test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state)) {
 901                 wake_up_process(info->transaction_kthread);
 902                 err = -EIO;
 903         }
 904         assert_qgroups_uptodate(trans);
 905
 906         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 907         if (must_run_delayed_refs) {
 908                 btrfs_async_run_delayed_refs(root, cur,
 909                                              must_run_delayed_refs == 1);
 910         }
 911         return err;
 912 }
 913
 914 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 915                           struct btrfs_root *root)
 916 {
 917         return __btrfs_end_transaction(trans, root, 0);
 918 }
 919
 920 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 921                                    struct btrfs_root *root)
 922 {
 923         return __btrfs_end_transaction(trans, root, 1);
 924 }
 925
 926 /*
 927  * when btree blocks are allocated, they have some corresponding bits set for
 928  * them in one of two extent_io trees.  This is used to make sure all of
 929  * those extents are sent to disk but does not wait on them
 930  */
 931 int btrfs_write_marked_extents(struct btrfs_root *root,
 932                                struct extent_io_tree *dirty_pages, int mark)
 933 {
 934         int err = 0;
 935         int werr = 0;
 936         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 937         struct extent_state *cached_state = NULL;
 938         u64 start = 0;
 939         u64 end;
 940
 941         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 942                                       mark, &cached_state)) {
 943                 bool wait_writeback = false;
 944
 945                 err = convert_extent_bit(dirty_pages, start, end,
 946                                          EXTENT_NEED_WAIT,
 947                                          mark, &cached_state);
 948                 /*
 949                  * convert_extent_bit can return -ENOMEM, which is most of the
 950                  * time a temporary error. So when it happens, ignore the error
 951                  * and wait for writeback of this range to finish - because we
 952                  * failed to set the bit EXTENT_NEED_WAIT for the range, a call
 953                  * to btrfs_wait_marked_extents() would not know that writeback
 954                  * for this range started and therefore wouldn't wait for it to
 955                  * finish - we don't want to commit a superblock that points to
 956                  * btree nodes/leafs for which writeback hasn't finished yet
 957                  * (and without errors).
 958                  * We cleanup any entries left in the io tree when committing
 959                  * the transaction (through clear_btree_io_tree()).
 960                  */
 961                 if (err == -ENOMEM) {
 962                         err = 0;
 963                         wait_writeback = true;
 964                 }
 965                 if (!err)
 966                         err = filemap_fdatawrite_range(mapping, start, end);
 967                 if (err)
 968                         werr = err;
 969                 else if (wait_writeback)
 970                         werr = filemap_fdatawait_range(mapping, start, end);
 971                 free_extent_state(cached_state);
 972                 cached_state = NULL;
 973                 cond_resched();
 974                 start = end + 1;
 975         }
 976         return werr;
 977 }
 978
 979 /*
 980  * when btree blocks are allocated, they have some corresponding bits set for
 981  * them in one of two extent_io trees.  This is used to make sure all of
 982  * those extents are on disk for transaction or log commit.  We wait
 983  * on all the pages and clear them from the dirty pages state tree
 984  */
 985 int btrfs_wait_marked_extents(struct btrfs_root *root,
 986                               struct extent_io_tree *dirty_pages, int mark)
 987 {
 988         int err = 0;
 989         int werr = 0;
 990         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 991         struct extent_state *cached_state = NULL;
 992         u64 start = 0;
 993         u64 end;
 994         struct btrfs_inode *btree_ino = BTRFS_I(root->fs_info->btree_inode);
 995         bool errors = false;
 996
 997         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 998                                       EXTENT_NEED_WAIT, &cached_state)) {
 999                 /*
1000                  * Ignore -ENOMEM errors returned by clear_extent_bit().
1001                  * When committing the transaction, we'll remove any entries
1002                  * left in the io tree. For a log commit, we don't remove them
1003                  * after committing the log because the tree can be accessed
1004                  * concurrently - we do it only at transaction commit time when
1005                  * it's safe to do it (through clear_btree_io_tree()).
1006                  */
1007                 err = clear_extent_bit(dirty_pages, start, end,
1008                                        EXTENT_NEED_WAIT,
1009                                        0, 0, &cached_state, GFP_NOFS);
1010                 if (err == -ENOMEM)
1011                         err = 0;
1012                 if (!err)
1013                         err = filemap_fdatawait_range(mapping, start, end);
1014                 if (err)
1015                         werr = err;
1016                 free_extent_state(cached_state);
1017                 cached_state = NULL;
1018                 cond_resched();
1019                 start = end + 1;
1020         }
1021         if (err)
1022                 werr = err;
1023
1024         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
1025                 if ((mark & EXTENT_DIRTY) &&
1026                     test_and_clear_bit(BTRFS_INODE_BTREE_LOG1_ERR,
1027                                        &btree_ino->runtime_flags))
1028                         errors = true;
1029
1030                 if ((mark & EXTENT_NEW) &&
1031                     test_and_clear_bit(BTRFS_INODE_BTREE_LOG2_ERR,
1032                                        &btree_ino->runtime_flags))
1033                         errors = true;
1034         } else {
1035                 if (test_and_clear_bit(BTRFS_INODE_BTREE_ERR,
1036                                        &btree_ino->runtime_flags))
1037                         errors = true;
1038         }
1039
1040         if (errors && !werr)
1041                 werr = -EIO;
1042
1043         return werr;
1044 }
1045
1046 /*
1047  * when btree blocks are allocated, they have some corresponding bits set for
1048  * them in one of two extent_io trees.  This is used to make sure all of
1049  * those extents are on disk for transaction or log commit
1050  */
1051 static int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
1052                                 struct extent_io_tree *dirty_pages, int mark)
1053 {
1054         int ret;
1055         int ret2;
1056         struct blk_plug plug;
1057
1058         blk_start_plug(&plug);
1059         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
1060         blk_finish_plug(&plug);
1061         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
1062
1063         if (ret)
1064                 return ret;
1065         if (ret2)
1066                 return ret2;
1067         return 0;
1068 }
1069
1070 static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
1071                                      struct btrfs_root *root)
1072 {
1073         int ret;
1074
1075         ret = btrfs_write_and_wait_marked_extents(root,
1076                                            &trans->transaction->dirty_pages,
1077                                            EXTENT_DIRTY);
1078         clear_btree_io_tree(&trans->transaction->dirty_pages);
1079
1080         return ret;
1081 }
1082
1083 /*
1084  * this is used to update the root pointer in the tree of tree roots.
1085  *
1086  * But, in the case of the extent allocation tree, updating the root
1087  * pointer may allocate blocks which may change the root of the extent
1088  * allocation tree.
1089  *
1090  * So, this loops and repeats and makes sure the cowonly root didn't
1091  * change while the root pointer was being updated in the metadata.
1092  */
1093 static int update_cowonly_root(struct btrfs_trans_handle *trans,
1094                                struct btrfs_root *root)
1095 {
1096         int ret;
1097         u64 old_root_bytenr;
1098         u64 old_root_used;
1099         struct btrfs_root *tree_root = root->fs_info->tree_root;
1100
1101         old_root_used = btrfs_root_used(&root->root_item);
1102
1103         while (1) {
1104                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1105                 if (old_root_bytenr == root->node->start &&
1106                     old_root_used == btrfs_root_used(&root->root_item))
1107                         break;
1108
1109                 btrfs_set_root_node(&root->root_item, root->node);
1110                 ret = btrfs_update_root(trans, tree_root,
1111                                         &root->root_key,
1112                                         &root->root_item);
1113                 if (ret)
1114                         return ret;
1115
1116                 old_root_used = btrfs_root_used(&root->root_item);
1117         }
1118
1119         return 0;
1120 }
1121
1122 /*
1123  * update all the cowonly tree roots on disk
1124  *
1125  * The error handling in this function may not be obvious. Any of the
1126  * failures will cause the file system to go offline. We still need
1127  * to clean up the delayed refs.
1128  */
1129 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
1130                                          struct btrfs_root *root)
1131 {
1132         struct btrfs_fs_info *fs_info = root->fs_info;
1133         struct list_head *dirty_bgs = &trans->transaction->dirty_bgs;
1134         struct list_head *io_bgs = &trans->transaction->io_bgs;
1135         struct list_head *next;
1136         struct extent_buffer *eb;
1137         int ret;
1138
1139         eb = btrfs_lock_root_node(fs_info->tree_root);
1140         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
1141                               0, &eb);
1142         btrfs_tree_unlock(eb);
1143         free_extent_buffer(eb);
1144
1145         if (ret)
1146                 return ret;
1147
1148         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1149         if (ret)
1150                 return ret;
1151
1152         ret = btrfs_run_dev_stats(trans, root->fs_info);
1153         if (ret)
1154                 return ret;
1155         ret = btrfs_run_dev_replace(trans, root->fs_info);
1156         if (ret)
1157                 return ret;
1158         ret = btrfs_run_qgroups(trans, root->fs_info);
1159         if (ret)
1160                 return ret;
1161
1162         ret = btrfs_setup_space_cache(trans, root);
1163         if (ret)
1164                 return ret;
1165
1166         /* run_qgroups might have added some more refs */
1167         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1168         if (ret)
1169                 return ret;
1170 again:
1171         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1172                 next = fs_info->dirty_cowonly_roots.next;
1173                 list_del_init(next);
1174                 root = list_entry(next, struct btrfs_root, dirty_list);
1175                 clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1176
1177                 if (root != fs_info->extent_root)
1178                         list_add_tail(&root->dirty_list,
1179                                       &trans->transaction->switch_commits);
1180                 ret = update_cowonly_root(trans, root);
1181                 if (ret)
1182                         return ret;
1183                 ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1184                 if (ret)
1185                         return ret;
1186         }
1187
1188         while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1189                 ret = btrfs_write_dirty_block_groups(trans, root);
1190                 if (ret)
1191                         return ret;
1192                 ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1193                 if (ret)
1194                         return ret;
1195         }
1196
1197         if (!list_empty(&fs_info->dirty_cowonly_roots))
1198                 goto again;
1199
1200         list_add_tail(&fs_info->extent_root->dirty_list,
1201                       &trans->transaction->switch_commits);
1202         btrfs_after_dev_replace_commit(fs_info);
1203
1204         return 0;
1205 }
1206
1207 /*
1208  * dead roots are old snapshots that need to be deleted.  This allocates
1209  * a dirty root struct and adds it into the list of dead roots that need to
1210  * be deleted
1211  */
1212 void btrfs_add_dead_root(struct btrfs_root *root)
1213 {
1214         spin_lock(&root->fs_info->trans_lock);
1215         if (list_empty(&root->root_list))
1216                 list_add_tail(&root->root_list, &root->fs_info->dead_roots);
1217         spin_unlock(&root->fs_info->trans_lock);
1218 }
1219
1220 /*
1221  * update all the cowonly tree roots on disk
1222  */
1223 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
1224                                     struct btrfs_root *root)
1225 {
1226         struct btrfs_root *gang[8];
1227         struct btrfs_fs_info *fs_info = root->fs_info;
1228         int i;
1229         int ret;
1230         int err = 0;
1231
1232         spin_lock(&fs_info->fs_roots_radix_lock);
1233         while (1) {
1234                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1235                                                  (void **)gang, 0,
1236                                                  ARRAY_SIZE(gang),
1237                                                  BTRFS_ROOT_TRANS_TAG);
1238                 if (ret == 0)
1239                         break;
1240                 for (i = 0; i < ret; i++) {
1241                         root = gang[i];
1242                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
1243                                         (unsigned long)root->root_key.objectid,
1244                                         BTRFS_ROOT_TRANS_TAG);
1245                         spin_unlock(&fs_info->fs_roots_radix_lock);
1246
1247                         btrfs_free_log(trans, root);
1248                         btrfs_update_reloc_root(trans, root);
1249                         btrfs_orphan_commit_root(trans, root);
1250
1251                         btrfs_save_ino_cache(root, trans);
1252
1253                         /* see comments in should_cow_block() */
1254                         clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1255                         smp_mb__after_atomic();
1256
1257                         if (root->commit_root != root->node) {
1258                                 list_add_tail(&root->dirty_list,
1259                                         &trans->transaction->switch_commits);
1260                                 btrfs_set_root_node(&root->root_item,
1261                                                     root->node);
1262                         }
1263
1264                         err = btrfs_update_root(trans, fs_info->tree_root,
1265                                                 &root->root_key,
1266                                                 &root->root_item);
1267                         spin_lock(&fs_info->fs_roots_radix_lock);
1268                         if (err)
1269                                 break;
1270                         btrfs_qgroup_free_meta_all(root);
1271                 }
1272         }
1273         spin_unlock(&fs_info->fs_roots_radix_lock);
1274         return err;
1275 }
1276
1277 /*
1278  * defrag a given btree.
1279  * Every leaf in the btree is read and defragged.
1280  */
1281 int btrfs_defrag_root(struct btrfs_root *root)
1282 {
1283         struct btrfs_fs_info *info = root->fs_info;
1284         struct btrfs_trans_handle *trans;
1285         int ret;
1286
1287         if (test_and_set_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state))
1288                 return 0;
1289
1290         while (1) {
1291                 trans = btrfs_start_transaction(root, 0);
1292                 if (IS_ERR(trans))
1293                         return PTR_ERR(trans);
1294
1295                 ret = btrfs_defrag_leaves(trans, root);
1296
1297                 btrfs_end_transaction(trans, root);
1298                 btrfs_btree_balance_dirty(info->tree_root);
1299                 cond_resched();
1300
1301                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
1302                         break;
1303
1304                 if (btrfs_defrag_cancelled(root->fs_info)) {
1305                         pr_debug("BTRFS: defrag_root cancelled\n");
1306                         ret = -EAGAIN;
1307                         break;
1308                 }
1309         }
1310         clear_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state);
1311         return ret;
1312 }
1313
1314 /* Bisesctability fixup, remove in 4.8 */
1315 #ifndef btrfs_std_error
1316 #define btrfs_std_error btrfs_handle_fs_error
1317 #endif
1318
1319 /*
1320  * Do all special snapshot related qgroup dirty hack.
1321  *
1322  * Will do all needed qgroup inherit and dirty hack like switch commit
1323  * roots inside one transaction and write all btree into disk, to make
1324  * qgroup works.
1325  */
1326 static int qgroup_account_snapshot(struct btrfs_trans_handle *trans,
1327                                    struct btrfs_root *src,
1328                                    struct btrfs_root *parent,
1329                                    struct btrfs_qgroup_inherit *inherit,
1330                                    u64 dst_objectid)
1331 {
1332         struct btrfs_fs_info *fs_info = src->fs_info;
1333         int ret;
1334
1335         /*
1336          * Save some performance in the case that qgroups are not
1337          * enabled. If this check races with the ioctl, rescan will
1338          * kick in anyway.
1339          */
1340         mutex_lock(&fs_info->qgroup_ioctl_lock);
1341         if (!fs_info->quota_enabled) {
1342                 mutex_unlock(&fs_info->qgroup_ioctl_lock);
1343                 return 0;
1344         }
1345         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1346
1347         /*
1348          * We are going to commit transaction, see btrfs_commit_transaction()
1349          * comment for reason locking tree_log_mutex
1350          */
1351         mutex_lock(&fs_info->tree_log_mutex);
1352
1353         ret = commit_fs_roots(trans, src);
1354         if (ret)
1355                 goto out;
1356         ret = btrfs_qgroup_prepare_account_extents(trans, fs_info);
1357         if (ret < 0)
1358                 goto out;
1359         ret = btrfs_qgroup_account_extents(trans, fs_info);
1360         if (ret < 0)
1361                 goto out;
1362
1363         /* Now qgroup are all updated, we can inherit it to new qgroups */
1364         ret = btrfs_qgroup_inherit(trans, fs_info,
1365                                    src->root_key.objectid, dst_objectid,
1366                                    inherit);
1367         if (ret < 0)
1368                 goto out;
1369
1370         /*
1371          * Now we do a simplified commit transaction, which will:
1372          * 1) commit all subvolume and extent tree
1373          *    To ensure all subvolume and extent tree have a valid
1374          *    commit_root to accounting later insert_dir_item()
1375          * 2) write all btree blocks onto disk
1376          *    This is to make sure later btree modification will be cowed
1377          *    Or commit_root can be populated and cause wrong qgroup numbers
1378          * In this simplified commit, we don't really care about other trees
1379          * like chunk and root tree, as they won't affect qgroup.
1380          * And we don't write super to avoid half committed status.
1381          */
1382         ret = commit_cowonly_roots(trans, src);
1383         if (ret)
1384                 goto out;
1385         switch_commit_roots(trans->transaction, fs_info);
1386         ret = btrfs_write_and_wait_transaction(trans, src);
1387         if (ret)
1388                 btrfs_std_error(fs_info, ret,
1389                         "Error while writing out transaction for qgroup");
1390
1391 out:
1392         mutex_unlock(&fs_info->tree_log_mutex);
1393
1394         /*
1395          * Force parent root to be updated, as we recorded it before so its
1396          * last_trans == cur_transid.
1397          * Or it won't be committed again onto disk after later
1398          * insert_dir_item()
1399          */
1400         if (!ret)
1401                 record_root_in_trans(trans, parent, 1);
1402         return ret;
1403 }
1404
1405 /*
1406  * new snapshots need to be created at a very specific time in the
1407  * transaction commit.  This does the actual creation.
1408  *
1409  * Note:
1410  * If the error which may affect the commitment of the current transaction
1411  * happens, we should return the error number. If the error which just affect
1412  * the creation of the pending snapshots, just return 0.
1413  */
1414 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1415                                    struct btrfs_fs_info *fs_info,
1416                                    struct btrfs_pending_snapshot *pending)
1417 {
1418         struct btrfs_key key;
1419         struct btrfs_root_item *new_root_item;
1420         struct btrfs_root *tree_root = fs_info->tree_root;
1421         struct btrfs_root *root = pending->root;
1422         struct btrfs_root *parent_root;
1423         struct btrfs_block_rsv *rsv;
1424         struct inode *parent_inode;
1425         struct btrfs_path *path;
1426         struct btrfs_dir_item *dir_item;
1427         struct dentry *dentry;
1428         struct extent_buffer *tmp;
1429         struct extent_buffer *old;
1430         struct timespec cur_time;
1431         int ret = 0;
1432         u64 to_reserve = 0;
1433         u64 index = 0;
1434         u64 objectid;
1435         u64 root_flags;
1436         uuid_le new_uuid;
1437
1438         ASSERT(pending->path);
1439         path = pending->path;
1440
1441         ASSERT(pending->root_item);
1442         new_root_item = pending->root_item;
1443
1444         pending->error = btrfs_find_free_objectid(tree_root, &objectid);
1445         if (pending->error)
1446                 goto no_free_objectid;
1447
1448         /*
1449          * Make qgroup to skip current new snapshot's qgroupid, as it is
1450          * accounted by later btrfs_qgroup_inherit().
1451          */
1452         btrfs_set_skip_qgroup(trans, objectid);
1453
1454         btrfs_reloc_pre_snapshot(pending, &to_reserve);
1455
1456         if (to_reserve > 0) {
1457                 pending->error = btrfs_block_rsv_add(root,
1458                                                      &pending->block_rsv,
1459                                                      to_reserve,
1460                                                      BTRFS_RESERVE_NO_FLUSH);
1461                 if (pending->error)
1462                         goto clear_skip_qgroup;
1463         }
1464
1465         key.objectid = objectid;
1466         key.offset = (u64)-1;
1467         key.type = BTRFS_ROOT_ITEM_KEY;
1468
1469         rsv = trans->block_rsv;
1470         trans->block_rsv = &pending->block_rsv;
1471         trans->bytes_reserved = trans->block_rsv->reserved;
1472         trace_btrfs_space_reservation(root->fs_info, "transaction",
1473                                       trans->transid,
1474                                       trans->bytes_reserved, 1);
1475         dentry = pending->dentry;
1476         parent_inode = pending->dir;
1477         parent_root = BTRFS_I(parent_inode)->root;
1478         record_root_in_trans(trans, parent_root, 0);
1479
1480         cur_time = current_fs_time(parent_inode->i_sb);
1481
1482         /*
1483          * insert the directory item
1484          */
1485         ret = btrfs_set_inode_index(parent_inode, &index);
1486         BUG_ON(ret); /* -ENOMEM */
1487
1488         /* check if there is a file/dir which has the same name. */
1489         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1490                                          btrfs_ino(parent_inode),
1491                                          dentry->d_name.name,
1492                                          dentry->d_name.len, 0);
1493         if (dir_item != NULL && !IS_ERR(dir_item)) {
1494                 pending->error = -EEXIST;
1495                 goto dir_item_existed;
1496         } else if (IS_ERR(dir_item)) {
1497                 ret = PTR_ERR(dir_item);
1498                 btrfs_abort_transaction(trans, root, ret);
1499                 goto fail;
1500         }
1501         btrfs_release_path(path);
1502
1503         /*
1504          * pull in the delayed directory update
1505          * and the delayed inode item
1506          * otherwise we corrupt the FS during
1507          * snapshot
1508          */
1509         ret = btrfs_run_delayed_items(trans, root);
1510         if (ret) {      /* Transaction aborted */
1511                 btrfs_abort_transaction(trans, root, ret);
1512                 goto fail;
1513         }
1514
1515         record_root_in_trans(trans, root, 0);
1516         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1517         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1518         btrfs_check_and_init_root_item(new_root_item);
1519
1520         root_flags = btrfs_root_flags(new_root_item);
1521         if (pending->readonly)
1522                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1523         else
1524                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1525         btrfs_set_root_flags(new_root_item, root_flags);
1526
1527         btrfs_set_root_generation_v2(new_root_item,
1528                         trans->transid);
1529         uuid_le_gen(&new_uuid);
1530         memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1531         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1532                         BTRFS_UUID_SIZE);
1533         if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1534                 memset(new_root_item->received_uuid, 0,
1535                        sizeof(new_root_item->received_uuid));
1536                 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1537                 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1538                 btrfs_set_root_stransid(new_root_item, 0);
1539                 btrfs_set_root_rtransid(new_root_item, 0);
1540         }
1541         btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1542         btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1543         btrfs_set_root_otransid(new_root_item, trans->transid);
1544
1545         old = btrfs_lock_root_node(root);
1546         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1547         if (ret) {
1548                 btrfs_tree_unlock(old);
1549                 free_extent_buffer(old);
1550                 btrfs_abort_transaction(trans, root, ret);
1551                 goto fail;
1552         }
1553
1554         btrfs_set_lock_blocking(old);
1555
1556         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1557         /* clean up in any case */
1558         btrfs_tree_unlock(old);
1559         free_extent_buffer(old);
1560         if (ret) {
1561                 btrfs_abort_transaction(trans, root, ret);
1562                 goto fail;
1563         }
1564         /* see comments in should_cow_block() */
1565         set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1566         smp_wmb();
1567
1568         btrfs_set_root_node(new_root_item, tmp);
1569         /* record when the snapshot was created in key.offset */
1570         key.offset = trans->transid;
1571         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1572         btrfs_tree_unlock(tmp);
1573         free_extent_buffer(tmp);
1574         if (ret) {
1575                 btrfs_abort_transaction(trans, root, ret);
1576                 goto fail;
1577         }
1578
1579         /*
1580          * insert root back/forward references
1581          */
1582         ret = btrfs_add_root_ref(trans, tree_root, objectid,
1583                                  parent_root->root_key.objectid,
1584                                  btrfs_ino(parent_inode), index,
1585                                  dentry->d_name.name, dentry->d_name.len);
1586         if (ret) {
1587                 btrfs_abort_transaction(trans, root, ret);
1588                 goto fail;
1589         }
1590
1591         key.offset = (u64)-1;
1592         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1593         if (IS_ERR(pending->snap)) {
1594                 ret = PTR_ERR(pending->snap);
1595                 btrfs_abort_transaction(trans, root, ret);
1596                 goto fail;
1597         }
1598
1599         ret = btrfs_reloc_post_snapshot(trans, pending);
1600         if (ret) {
1601                 btrfs_abort_transaction(trans, root, ret);
1602                 goto fail;
1603         }
1604
1605         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1606         if (ret) {
1607                 btrfs_abort_transaction(trans, root, ret);
1608                 goto fail;
1609         }
1610
1611         /*
1612          * Do special qgroup accounting for snapshot, as we do some qgroup
1613          * snapshot hack to do fast snapshot.
1614          * To co-operate with that hack, we do hack again.
1615          * Or snapshot will be greatly slowed down by a subtree qgroup rescan
1616          */
1617         ret = qgroup_account_snapshot(trans, root, parent_root,
1618                                       pending->inherit, objectid);
1619         if (ret < 0)
1620                 goto fail;
1621
1622         ret = btrfs_insert_dir_item(trans, parent_root,
1623                                     dentry->d_name.name, dentry->d_name.len,
1624                                     parent_inode, &key,
1625                                     BTRFS_FT_DIR, index);
1626         /* We have check then name at the beginning, so it is impossible. */
1627         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1628         if (ret) {
1629                 btrfs_abort_transaction(trans, root, ret);
1630                 goto fail;
1631         }
1632
1633         btrfs_i_size_write(parent_inode, parent_inode->i_size +
1634                                          dentry->d_name.len * 2);
1635         parent_inode->i_mtime = parent_inode->i_ctime =
1636                 current_fs_time(parent_inode->i_sb);
1637         ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1638         if (ret) {
1639                 btrfs_abort_transaction(trans, root, ret);
1640                 goto fail;
1641         }
1642         ret = btrfs_uuid_tree_add(trans, fs_info->uuid_root, new_uuid.b,
1643                                   BTRFS_UUID_KEY_SUBVOL, objectid);
1644         if (ret) {
1645                 btrfs_abort_transaction(trans, root, ret);
1646                 goto fail;
1647         }
1648         if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1649                 ret = btrfs_uuid_tree_add(trans, fs_info->uuid_root,
1650                                           new_root_item->received_uuid,
1651                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1652                                           objectid);
1653                 if (ret && ret != -EEXIST) {
1654                         btrfs_abort_transaction(trans, root, ret);
1655                         goto fail;
1656                 }
1657         }
1658
1659         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1660         if (ret) {
1661                 btrfs_abort_transaction(trans, root, ret);
1662                 goto fail;
1663         }
1664
1665 fail:
1666         pending->error = ret;
1667 dir_item_existed:
1668         trans->block_rsv = rsv;
1669         trans->bytes_reserved = 0;
1670 clear_skip_qgroup:
1671         btrfs_clear_skip_qgroup(trans);
1672 no_free_objectid:
1673         kfree(new_root_item);
1674         pending->root_item = NULL;
1675         btrfs_free_path(path);
1676         pending->path = NULL;
1677
1678         return ret;
1679 }
1680
1681 /*
1682  * create all the snapshots we've scheduled for creation
1683  */
1684 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1685                                              struct btrfs_fs_info *fs_info)
1686 {
1687         struct btrfs_pending_snapshot *pending, *next;
1688         struct list_head *head = &trans->transaction->pending_snapshots;
1689         int ret = 0;
1690
1691         list_for_each_entry_safe(pending, next, head, list) {
1692                 list_del(&pending->list);
1693                 ret = create_pending_snapshot(trans, fs_info, pending);
1694                 if (ret)
1695                         break;
1696         }
1697         return ret;
1698 }
1699
1700 static void update_super_roots(struct btrfs_root *root)
1701 {
1702         struct btrfs_root_item *root_item;
1703         struct btrfs_super_block *super;
1704
1705         super = root->fs_info->super_copy;
1706
1707         root_item = &root->fs_info->chunk_root->root_item;
1708         super->chunk_root = root_item->bytenr;
1709         super->chunk_root_generation = root_item->generation;
1710         super->chunk_root_level = root_item->level;
1711
1712         root_item = &root->fs_info->tree_root->root_item;
1713         super->root = root_item->bytenr;
1714         super->generation = root_item->generation;
1715         super->root_level = root_item->level;
1716         if (btrfs_test_opt(root, SPACE_CACHE))
1717                 super->cache_generation = root_item->generation;
1718         if (root->fs_info->update_uuid_tree_gen)
1719                 super->uuid_tree_generation = root_item->generation;
1720 }
1721
1722 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1723 {
1724         struct btrfs_transaction *trans;
1725         int ret = 0;
1726
1727         spin_lock(&info->trans_lock);
1728         trans = info->running_transaction;
1729         if (trans)
1730                 ret = (trans->state >= TRANS_STATE_COMMIT_START);
1731         spin_unlock(&info->trans_lock);
1732         return ret;
1733 }
1734
1735 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1736 {
1737         struct btrfs_transaction *trans;
1738         int ret = 0;
1739
1740         spin_lock(&info->trans_lock);
1741         trans = info->running_transaction;
1742         if (trans)
1743                 ret = is_transaction_blocked(trans);
1744         spin_unlock(&info->trans_lock);
1745         return ret;
1746 }
1747
1748 /*
1749  * wait for the current transaction commit to start and block subsequent
1750  * transaction joins
1751  */
1752 static void wait_current_trans_commit_start(struct btrfs_root *root,
1753                                             struct btrfs_transaction *trans)
1754 {
1755         wait_event(root->fs_info->transaction_blocked_wait,
1756                    trans->state >= TRANS_STATE_COMMIT_START ||
1757                    trans->aborted);
1758 }
1759
1760 /*
1761  * wait for the current transaction to start and then become unblocked.
1762  * caller holds ref.
1763  */
1764 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1765                                          struct btrfs_transaction *trans)
1766 {
1767         wait_event(root->fs_info->transaction_wait,
1768                    trans->state >= TRANS_STATE_UNBLOCKED ||
1769                    trans->aborted);
1770 }
1771
1772 /*
1773  * commit transactions asynchronously. once btrfs_commit_transaction_async
1774  * returns, any subsequent transaction will not be allowed to join.
1775  */
1776 struct btrfs_async_commit {
1777         struct btrfs_trans_handle *newtrans;
1778         struct btrfs_root *root;
1779         struct work_struct work;
1780 };
1781
1782 static void do_async_commit(struct work_struct *work)
1783 {
1784         struct btrfs_async_commit *ac =
1785                 container_of(work, struct btrfs_async_commit, work);
1786
1787         /*
1788          * We've got freeze protection passed with the transaction.
1789          * Tell lockdep about it.
1790          */
1791         if (ac->newtrans->type & __TRANS_FREEZABLE)
1792                 __sb_writers_acquired(ac->root->fs_info->sb, SB_FREEZE_FS);
1793
1794         current->journal_info = ac->newtrans;
1795
1796         btrfs_commit_transaction(ac->newtrans, ac->root);
1797         kfree(ac);
1798 }
1799
1800 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1801                                    struct btrfs_root *root,
1802                                    int wait_for_unblock)
1803 {
1804         struct btrfs_async_commit *ac;
1805         struct btrfs_transaction *cur_trans;
1806
1807         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1808         if (!ac)
1809                 return -ENOMEM;
1810
1811         INIT_WORK(&ac->work, do_async_commit);
1812         ac->root = root;
1813         ac->newtrans = btrfs_join_transaction(root);
1814         if (IS_ERR(ac->newtrans)) {
1815                 int err = PTR_ERR(ac->newtrans);
1816                 kfree(ac);
1817                 return err;
1818         }
1819
1820         /* take transaction reference */
1821         cur_trans = trans->transaction;
1822         atomic_inc(&cur_trans->use_count);
1823
1824         btrfs_end_transaction(trans, root);
1825
1826         /*
1827          * Tell lockdep we've released the freeze rwsem, since the
1828          * async commit thread will be the one to unlock it.
1829          */
1830         if (ac->newtrans->type & __TRANS_FREEZABLE)
1831                 __sb_writers_release(root->fs_info->sb, SB_FREEZE_FS);
1832
1833         schedule_work(&ac->work);
1834
1835         /* wait for transaction to start and unblock */
1836         if (wait_for_unblock)
1837                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1838         else
1839                 wait_current_trans_commit_start(root, cur_trans);
1840
1841         if (current->journal_info == trans)
1842                 current->journal_info = NULL;
1843
1844         btrfs_put_transaction(cur_trans);
1845         return 0;
1846 }
1847
1848
1849 static void cleanup_transaction(struct btrfs_trans_handle *trans,
1850                                 struct btrfs_root *root, int err)
1851 {
1852         struct btrfs_transaction *cur_trans = trans->transaction;
1853         DEFINE_WAIT(wait);
1854
1855         WARN_ON(trans->use_count > 1);
1856
1857         btrfs_abort_transaction(trans, root, err);
1858
1859         spin_lock(&root->fs_info->trans_lock);
1860
1861         /*
1862          * If the transaction is removed from the list, it means this
1863          * transaction has been committed successfully, so it is impossible
1864          * to call the cleanup function.
1865          */
1866         BUG_ON(list_empty(&cur_trans->list));
1867
1868         list_del_init(&cur_trans->list);
1869         if (cur_trans == root->fs_info->running_transaction) {
1870                 cur_trans->state = TRANS_STATE_COMMIT_DOING;
1871                 spin_unlock(&root->fs_info->trans_lock);
1872                 wait_event(cur_trans->writer_wait,
1873                            atomic_read(&cur_trans->num_writers) == 1);
1874
1875                 spin_lock(&root->fs_info->trans_lock);
1876         }
1877         spin_unlock(&root->fs_info->trans_lock);
1878
1879         btrfs_cleanup_one_transaction(trans->transaction, root);
1880
1881         spin_lock(&root->fs_info->trans_lock);
1882         if (cur_trans == root->fs_info->running_transaction)
1883                 root->fs_info->running_transaction = NULL;
1884         spin_unlock(&root->fs_info->trans_lock);
1885
1886         if (trans->type & __TRANS_FREEZABLE)
1887                 sb_end_intwrite(root->fs_info->sb);
1888         btrfs_put_transaction(cur_trans);
1889         btrfs_put_transaction(cur_trans);
1890
1891         trace_btrfs_transaction_commit(root);
1892
1893         if (current->journal_info == trans)
1894                 current->journal_info = NULL;
1895         btrfs_scrub_cancel(root->fs_info);
1896
1897         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1898 }
1899
1900 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
1901 {
1902         if (btrfs_test_opt(fs_info->tree_root, FLUSHONCOMMIT))
1903                 return btrfs_start_delalloc_roots(fs_info, 1, -1);
1904         return 0;
1905 }
1906
1907 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
1908 {
1909         if (btrfs_test_opt(fs_info->tree_root, FLUSHONCOMMIT))
1910                 btrfs_wait_ordered_roots(fs_info, -1, 0, (u64)-1);
1911 }
1912
1913 static inline void
1914 btrfs_wait_pending_ordered(struct btrfs_transaction *cur_trans)
1915 {
1916         wait_event(cur_trans->pending_wait,
1917                    atomic_read(&cur_trans->pending_ordered) == 0);
1918 }
1919
1920 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1921                              struct btrfs_root *root)
1922 {
1923         struct btrfs_transaction *cur_trans = trans->transaction;
1924         struct btrfs_transaction *prev_trans = NULL;
1925         struct btrfs_inode *btree_ino = BTRFS_I(root->fs_info->btree_inode);
1926         int ret;
1927
1928         /* Stop the commit early if ->aborted is set */
1929         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1930                 ret = cur_trans->aborted;
1931                 btrfs_end_transaction(trans, root);
1932                 return ret;
1933         }
1934
1935         /* make a pass through all the delayed refs we have so far
1936          * any runnings procs may add more while we are here
1937          */
1938         ret = btrfs_run_delayed_refs(trans, root, 0);
1939         if (ret) {
1940                 btrfs_end_transaction(trans, root);
1941                 return ret;
1942         }
1943
1944         btrfs_trans_release_metadata(trans, root);
1945         trans->block_rsv = NULL;
1946
1947         cur_trans = trans->transaction;
1948
1949         /*
1950          * set the flushing flag so procs in this transaction have to
1951          * start sending their work down.
1952          */
1953         cur_trans->delayed_refs.flushing = 1;
1954         smp_wmb();
1955
1956         if (!list_empty(&trans->new_bgs))
1957                 btrfs_create_pending_block_groups(trans, root);
1958
1959         ret = btrfs_run_delayed_refs(trans, root, 0);
1960         if (ret) {
1961                 btrfs_end_transaction(trans, root);
1962                 return ret;
1963         }
1964
1965         if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) {
1966                 int run_it = 0;
1967
1968                 /* this mutex is also taken before trying to set
1969                  * block groups readonly.  We need to make sure
1970                  * that nobody has set a block group readonly
1971                  * after a extents from that block group have been
1972                  * allocated for cache files.  btrfs_set_block_group_ro
1973                  * will wait for the transaction to commit if it
1974                  * finds BTRFS_TRANS_DIRTY_BG_RUN set.
1975                  *
1976                  * The BTRFS_TRANS_DIRTY_BG_RUN flag is also used to make sure
1977                  * only one process starts all the block group IO.  It wouldn't
1978                  * hurt to have more than one go through, but there's no
1979                  * real advantage to it either.
1980                  */
1981                 mutex_lock(&root->fs_info->ro_block_group_mutex);
1982                 if (!test_and_set_bit(BTRFS_TRANS_DIRTY_BG_RUN,
1983                                       &cur_trans->flags))
1984                         run_it = 1;
1985                 mutex_unlock(&root->fs_info->ro_block_group_mutex);
1986
1987                 if (run_it)
1988                         ret = btrfs_start_dirty_block_groups(trans, root);
1989         }
1990         if (ret) {
1991                 btrfs_end_transaction(trans, root);
1992                 return ret;
1993         }
1994
1995         spin_lock(&root->fs_info->trans_lock);
1996         if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
1997                 spin_unlock(&root->fs_info->trans_lock);
1998                 atomic_inc(&cur_trans->use_count);
1999                 ret = btrfs_end_transaction(trans, root);
2000
2001                 wait_for_commit(root, cur_trans);
2002
2003                 if (unlikely(cur_trans->aborted))
2004                         ret = cur_trans->aborted;
2005
2006                 btrfs_put_transaction(cur_trans);
2007
2008                 return ret;
2009         }
2010
2011         cur_trans->state = TRANS_STATE_COMMIT_START;
2012         wake_up(&root->fs_info->transaction_blocked_wait);
2013
2014         if (cur_trans->list.prev != &root->fs_info->trans_list) {
2015                 prev_trans = list_entry(cur_trans->list.prev,
2016                                         struct btrfs_transaction, list);
2017                 if (prev_trans->state != TRANS_STATE_COMPLETED) {
2018                         atomic_inc(&prev_trans->use_count);
2019                         spin_unlock(&root->fs_info->trans_lock);
2020
2021                         wait_for_commit(root, prev_trans);
2022                         ret = prev_trans->aborted;
2023
2024                         btrfs_put_transaction(prev_trans);
2025                         if (ret)
2026                                 goto cleanup_transaction;
2027                 } else {
2028                         spin_unlock(&root->fs_info->trans_lock);
2029                 }
2030         } else {
2031                 spin_unlock(&root->fs_info->trans_lock);
2032         }
2033
2034         extwriter_counter_dec(cur_trans, trans->type);
2035
2036         ret = btrfs_start_delalloc_flush(root->fs_info);
2037         if (ret)
2038                 goto cleanup_transaction;
2039
2040         ret = btrfs_run_delayed_items(trans, root);
2041         if (ret)
2042                 goto cleanup_transaction;
2043
2044         wait_event(cur_trans->writer_wait,
2045                    extwriter_counter_read(cur_trans) == 0);
2046
2047         /* some pending stuffs might be added after the previous flush. */
2048         ret = btrfs_run_delayed_items(trans, root);
2049         if (ret)
2050                 goto cleanup_transaction;
2051
2052         btrfs_wait_delalloc_flush(root->fs_info);
2053
2054         btrfs_wait_pending_ordered(cur_trans);
2055
2056         btrfs_scrub_pause(root);
2057         /*
2058          * Ok now we need to make sure to block out any other joins while we
2059          * commit the transaction.  We could have started a join before setting
2060          * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
2061          */
2062         spin_lock(&root->fs_info->trans_lock);
2063         cur_trans->state = TRANS_STATE_COMMIT_DOING;
2064         spin_unlock(&root->fs_info->trans_lock);
2065         wait_event(cur_trans->writer_wait,
2066                    atomic_read(&cur_trans->num_writers) == 1);
2067
2068         /* ->aborted might be set after the previous check, so check it */
2069         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
2070                 ret = cur_trans->aborted;
2071                 goto scrub_continue;
2072         }
2073         /*
2074          * the reloc mutex makes sure that we stop
2075          * the balancing code from coming in and moving
2076          * extents around in the middle of the commit
2077          */
2078         mutex_lock(&root->fs_info->reloc_mutex);
2079
2080         /*
2081          * We needn't worry about the delayed items because we will
2082          * deal with them in create_pending_snapshot(), which is the
2083          * core function of the snapshot creation.
2084          */
2085         ret = create_pending_snapshots(trans, root->fs_info);
2086         if (ret) {
2087                 mutex_unlock(&root->fs_info->reloc_mutex);
2088                 goto scrub_continue;
2089         }
2090
2091         /*
2092          * We insert the dir indexes of the snapshots and update the inode
2093          * of the snapshots' parents after the snapshot creation, so there
2094          * are some delayed items which are not dealt with. Now deal with
2095          * them.
2096          *
2097          * We needn't worry that this operation will corrupt the snapshots,
2098          * because all the tree which are snapshoted will be forced to COW
2099          * the nodes and leaves.
2100          */
2101         ret = btrfs_run_delayed_items(trans, root);
2102         if (ret) {
2103                 mutex_unlock(&root->fs_info->reloc_mutex);
2104                 goto scrub_continue;
2105         }
2106
2107         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
2108         if (ret) {
2109                 mutex_unlock(&root->fs_info->reloc_mutex);
2110                 goto scrub_continue;
2111         }
2112
2113         /* Reocrd old roots for later qgroup accounting */
2114         ret = btrfs_qgroup_prepare_account_extents(trans, root->fs_info);
2115         if (ret) {
2116                 mutex_unlock(&root->fs_info->reloc_mutex);
2117                 goto scrub_continue;
2118         }
2119
2120         /*
2121          * make sure none of the code above managed to slip in a
2122          * delayed item
2123          */
2124         btrfs_assert_delayed_root_empty(root);
2125
2126         WARN_ON(cur_trans != trans->transaction);
2127
2128         /* btrfs_commit_tree_roots is responsible for getting the
2129          * various roots consistent with each other.  Every pointer
2130          * in the tree of tree roots has to point to the most up to date
2131          * root for every subvolume and other tree.  So, we have to keep
2132          * the tree logging code from jumping in and changing any
2133          * of the trees.
2134          *
2135          * At this point in the commit, there can't be any tree-log
2136          * writers, but a little lower down we drop the trans mutex
2137          * and let new people in.  By holding the tree_log_mutex
2138          * from now until after the super is written, we avoid races
2139          * with the tree-log code.
2140          */
2141         mutex_lock(&root->fs_info->tree_log_mutex);
2142
2143         ret = commit_fs_roots(trans, root);
2144         if (ret) {
2145                 mutex_unlock(&root->fs_info->tree_log_mutex);
2146                 mutex_unlock(&root->fs_info->reloc_mutex);
2147                 goto scrub_continue;
2148         }
2149
2150         /*
2151          * Since the transaction is done, we can apply the pending changes
2152          * before the next transaction.
2153          */
2154         btrfs_apply_pending_changes(root->fs_info);
2155
2156         /* commit_fs_roots gets rid of all the tree log roots, it is now
2157          * safe to free the root of tree log roots
2158          */
2159         btrfs_free_log_root_tree(trans, root->fs_info);
2160
2161         /*
2162          * Since fs roots are all committed, we can get a quite accurate
2163          * new_roots. So let's do quota accounting.
2164          */
2165         ret = btrfs_qgroup_account_extents(trans, root->fs_info);
2166         if (ret < 0) {
2167                 mutex_unlock(&root->fs_info->tree_log_mutex);
2168                 mutex_unlock(&root->fs_info->reloc_mutex);
2169                 goto scrub_continue;
2170         }
2171
2172         ret = commit_cowonly_roots(trans, root);
2173         if (ret) {
2174                 mutex_unlock(&root->fs_info->tree_log_mutex);
2175                 mutex_unlock(&root->fs_info->reloc_mutex);
2176                 goto scrub_continue;
2177         }
2178
2179         /*
2180          * The tasks which save the space cache and inode cache may also
2181          * update ->aborted, check it.
2182          */
2183         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
2184                 ret = cur_trans->aborted;
2185                 mutex_unlock(&root->fs_info->tree_log_mutex);
2186                 mutex_unlock(&root->fs_info->reloc_mutex);
2187                 goto scrub_continue;
2188         }
2189
2190         btrfs_prepare_extent_commit(trans, root);
2191
2192         cur_trans = root->fs_info->running_transaction;
2193
2194         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
2195                             root->fs_info->tree_root->node);
2196         list_add_tail(&root->fs_info->tree_root->dirty_list,
2197                       &cur_trans->switch_commits);
2198
2199         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
2200                             root->fs_info->chunk_root->node);
2201         list_add_tail(&root->fs_info->chunk_root->dirty_list,
2202                       &cur_trans->switch_commits);
2203
2204         switch_commit_roots(cur_trans, root->fs_info);
2205
2206         assert_qgroups_uptodate(trans);
2207         ASSERT(list_empty(&cur_trans->dirty_bgs));
2208         ASSERT(list_empty(&cur_trans->io_bgs));
2209         update_super_roots(root);
2210
2211         btrfs_set_super_log_root(root->fs_info->super_copy, 0);
2212         btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
2213         memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
2214                sizeof(*root->fs_info->super_copy));
2215
2216         btrfs_update_commit_device_size(root->fs_info);
2217         btrfs_update_commit_device_bytes_used(root, cur_trans);
2218
2219         clear_bit(BTRFS_INODE_BTREE_LOG1_ERR, &btree_ino->runtime_flags);
2220         clear_bit(BTRFS_INODE_BTREE_LOG2_ERR, &btree_ino->runtime_flags);
2221
2222         btrfs_trans_release_chunk_metadata(trans);
2223
2224         spin_lock(&root->fs_info->trans_lock);
2225         cur_trans->state = TRANS_STATE_UNBLOCKED;
2226         root->fs_info->running_transaction = NULL;
2227         spin_unlock(&root->fs_info->trans_lock);
2228         mutex_unlock(&root->fs_info->reloc_mutex);
2229
2230         wake_up(&root->fs_info->transaction_wait);
2231
2232         ret = btrfs_write_and_wait_transaction(trans, root);
2233         if (ret) {
2234                 btrfs_handle_fs_error(root->fs_info, ret,
2235                             "Error while writing out transaction");
2236                 mutex_unlock(&root->fs_info->tree_log_mutex);
2237                 goto scrub_continue;
2238         }
2239
2240         ret = write_ctree_super(trans, root, 0);
2241         if (ret) {
2242                 mutex_unlock(&root->fs_info->tree_log_mutex);
2243                 goto scrub_continue;
2244         }
2245
2246         /*
2247          * the super is written, we can safely allow the tree-loggers
2248          * to go about their business
2249          */
2250         mutex_unlock(&root->fs_info->tree_log_mutex);
2251
2252         btrfs_finish_extent_commit(trans, root);
2253
2254         if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &cur_trans->flags))
2255                 btrfs_clear_space_info_full(root->fs_info);
2256
2257         root->fs_info->last_trans_committed = cur_trans->transid;
2258         /*
2259          * We needn't acquire the lock here because there is no other task
2260          * which can change it.
2261          */
2262         cur_trans->state = TRANS_STATE_COMPLETED;
2263         wake_up(&cur_trans->commit_wait);
2264
2265         spin_lock(&root->fs_info->trans_lock);
2266         list_del_init(&cur_trans->list);
2267         spin_unlock(&root->fs_info->trans_lock);
2268
2269         btrfs_put_transaction(cur_trans);
2270         btrfs_put_transaction(cur_trans);
2271
2272         if (trans->type & __TRANS_FREEZABLE)
2273                 sb_end_intwrite(root->fs_info->sb);
2274
2275         trace_btrfs_transaction_commit(root);
2276
2277         btrfs_scrub_continue(root);
2278
2279         if (current->journal_info == trans)
2280                 current->journal_info = NULL;
2281
2282         kmem_cache_free(btrfs_trans_handle_cachep, trans);
2283
2284         if (current != root->fs_info->transaction_kthread &&
2285             current != root->fs_info->cleaner_kthread)
2286                 btrfs_run_delayed_iputs(root);
2287
2288         return ret;
2289
2290 scrub_continue:
2291         btrfs_scrub_continue(root);
2292 cleanup_transaction:
2293         btrfs_trans_release_metadata(trans, root);
2294         btrfs_trans_release_chunk_metadata(trans);
2295         trans->block_rsv = NULL;
2296         btrfs_warn(root->fs_info, "Skipping commit of aborted transaction.");
2297         if (current->journal_info == trans)
2298                 current->journal_info = NULL;
2299         cleanup_transaction(trans, root, ret);
2300
2301         return ret;
2302 }
2303
2304 /*
2305  * return < 0 if error
2306  * 0 if there are no more dead_roots at the time of call
2307  * 1 there are more to be processed, call me again
2308  *
2309  * The return value indicates there are certainly more snapshots to delete, but
2310  * if there comes a new one during processing, it may return 0. We don't mind,
2311  * because btrfs_commit_super will poke cleaner thread and it will process it a
2312  * few seconds later.
2313  */
2314 int btrfs_clean_one_deleted_snapshot(struct btrfs_root *root)
2315 {
2316         int ret;
2317         struct btrfs_fs_info *fs_info = root->fs_info;
2318
2319         spin_lock(&fs_info->trans_lock);
2320         if (list_empty(&fs_info->dead_roots)) {
2321                 spin_unlock(&fs_info->trans_lock);
2322                 return 0;
2323         }
2324         root = list_first_entry(&fs_info->dead_roots,
2325                         struct btrfs_root, root_list);
2326         list_del_init(&root->root_list);
2327         spin_unlock(&fs_info->trans_lock);
2328
2329         pr_debug("BTRFS: cleaner removing %llu\n", root->objectid);
2330
2331         btrfs_kill_all_delayed_nodes(root);
2332
2333         if (btrfs_header_backref_rev(root->node) <
2334                         BTRFS_MIXED_BACKREF_REV)
2335                 ret = btrfs_drop_snapshot(root, NULL, 0, 0);
2336         else
2337                 ret = btrfs_drop_snapshot(root, NULL, 1, 0);
2338
2339         return (ret < 0) ? 0 : 1;
2340 }
2341
2342 void btrfs_apply_pending_changes(struct btrfs_fs_info *fs_info)
2343 {
2344         unsigned long prev;
2345         unsigned long bit;
2346
2347         prev = xchg(&fs_info->pending_changes, 0);
2348         if (!prev)
2349                 return;
2350
2351         bit = 1 << BTRFS_PENDING_SET_INODE_MAP_CACHE;
2352         if (prev & bit)
2353                 btrfs_set_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2354         prev &= ~bit;
2355
2356         bit = 1 << BTRFS_PENDING_CLEAR_INODE_MAP_CACHE;
2357         if (prev & bit)
2358                 btrfs_clear_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2359         prev &= ~bit;
2360
2361         bit = 1 << BTRFS_PENDING_COMMIT;
2362         if (prev & bit)
2363                 btrfs_debug(fs_info, "pending commit done");
2364         prev &= ~bit;
2365
2366         if (prev)
2367                 btrfs_warn(fs_info,
2368                         "unknown pending changes left 0x%lx, ignoring", prev);
2369 }