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                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  79         }
  80 }
  81
  82 static void clear_btree_io_tree(struct extent_io_tree *tree)
  83 {
  84         spin_lock(&tree->lock);
  85         while (!RB_EMPTY_ROOT(&tree->state)) {
  86                 struct rb_node *node;
  87                 struct extent_state *state;
  88
  89                 node = rb_first(&tree->state);
  90                 state = rb_entry(node, struct extent_state, rb_node);
  91                 rb_erase(&state->rb_node, &tree->state);
  92                 RB_CLEAR_NODE(&state->rb_node);
  93                 /*
  94                  * btree io trees aren't supposed to have tasks waiting for
  95                  * changes in the flags of extent states ever.
  96                  */
  97                 ASSERT(!waitqueue_active(&state->wq));
  98                 free_extent_state(state);
  99
 100                 cond_resched_lock(&tree->lock);
 101         }
 102         spin_unlock(&tree->lock);
 103 }
 104
 105 static noinline void switch_commit_roots(struct btrfs_transaction *trans,
 106                                          struct btrfs_fs_info *fs_info)
 107 {
 108         struct btrfs_root *root, *tmp;
 109
 110         down_write(&fs_info->commit_root_sem);
 111         list_for_each_entry_safe(root, tmp, &trans->switch_commits,
 112                                  dirty_list) {
 113                 list_del_init(&root->dirty_list);
 114                 free_extent_buffer(root->commit_root);
 115                 root->commit_root = btrfs_root_node(root);
 116                 if (is_fstree(root->objectid))
 117                         btrfs_unpin_free_ino(root);
 118                 clear_btree_io_tree(&root->dirty_log_pages);
 119         }
 120
 121         /* We can free old roots now. */
 122         spin_lock(&trans->dropped_roots_lock);
 123         while (!list_empty(&trans->dropped_roots)) {
 124                 root = list_first_entry(&trans->dropped_roots,
 125                                         struct btrfs_root, root_list);
 126                 list_del_init(&root->root_list);
 127                 spin_unlock(&trans->dropped_roots_lock);
 128                 btrfs_drop_and_free_fs_root(fs_info, root);
 129                 spin_lock(&trans->dropped_roots_lock);
 130         }
 131         spin_unlock(&trans->dropped_roots_lock);
 132         up_write(&fs_info->commit_root_sem);
 133 }
 134
 135 static inline void extwriter_counter_inc(struct btrfs_transaction *trans,
 136                                          unsigned int type)
 137 {
 138         if (type & TRANS_EXTWRITERS)
 139                 atomic_inc(&trans->num_extwriters);
 140 }
 141
 142 static inline void extwriter_counter_dec(struct btrfs_transaction *trans,
 143                                          unsigned int type)
 144 {
 145         if (type & TRANS_EXTWRITERS)
 146                 atomic_dec(&trans->num_extwriters);
 147 }
 148
 149 static inline void extwriter_counter_init(struct btrfs_transaction *trans,
 150                                           unsigned int type)
 151 {
 152         atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0));
 153 }
 154
 155 static inline int extwriter_counter_read(struct btrfs_transaction *trans)
 156 {
 157         return atomic_read(&trans->num_extwriters);
 158 }
 159
 160 /*
 161  * either allocate a new transaction or hop into the existing one
 162  */
 163 static noinline int join_transaction(struct btrfs_root *root, unsigned int type)
 164 {
 165         struct btrfs_transaction *cur_trans;
 166         struct btrfs_fs_info *fs_info = root->fs_info;
 167
 168         spin_lock(&fs_info->trans_lock);
 169 loop:
 170         /* The file system has been taken offline. No new transactions. */
 171         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
 172                 spin_unlock(&fs_info->trans_lock);
 173                 return -EROFS;
 174         }
 175
 176         cur_trans = fs_info->running_transaction;
 177         if (cur_trans) {
 178                 if (cur_trans->aborted) {
 179                         spin_unlock(&fs_info->trans_lock);
 180                         return cur_trans->aborted;
 181                 }
 182                 if (btrfs_blocked_trans_types[cur_trans->state] & type) {
 183                         spin_unlock(&fs_info->trans_lock);
 184                         return -EBUSY;
 185                 }
 186                 atomic_inc(&cur_trans->use_count);
 187                 atomic_inc(&cur_trans->num_writers);
 188                 extwriter_counter_inc(cur_trans, type);
 189                 spin_unlock(&fs_info->trans_lock);
 190                 return 0;
 191         }
 192         spin_unlock(&fs_info->trans_lock);
 193
 194         /*
 195          * If we are ATTACH, we just want to catch the current transaction,
 196          * and commit it. If there is no transaction, just return ENOENT.
 197          */
 198         if (type == TRANS_ATTACH)
 199                 return -ENOENT;
 200
 201         /*
 202          * JOIN_NOLOCK only happens during the transaction commit, so
 203          * it is impossible that ->running_transaction is NULL
 204          */
 205         BUG_ON(type == TRANS_JOIN_NOLOCK);
 206
 207         cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
 208         if (!cur_trans)
 209                 return -ENOMEM;
 210
 211         spin_lock(&fs_info->trans_lock);
 212         if (fs_info->running_transaction) {
 213                 /*
 214                  * someone started a transaction after we unlocked.  Make sure
 215                  * to redo the checks above
 216                  */
 217                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 218                 goto loop;
 219         } else if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
 220                 spin_unlock(&fs_info->trans_lock);
 221                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 222                 return -EROFS;
 223         }
 224
 225         atomic_set(&cur_trans->num_writers, 1);
 226         extwriter_counter_init(cur_trans, type);
 227         init_waitqueue_head(&cur_trans->writer_wait);
 228         init_waitqueue_head(&cur_trans->commit_wait);
 229         cur_trans->state = TRANS_STATE_RUNNING;
 230         /*
 231          * One for this trans handle, one so it will live on until we
 232          * commit the transaction.
 233          */
 234         atomic_set(&cur_trans->use_count, 2);
 235         cur_trans->have_free_bgs = 0;
 236         cur_trans->start_time = get_seconds();
 237         cur_trans->dirty_bg_run = 0;
 238
 239         cur_trans->delayed_refs.href_root = RB_ROOT;
 240         cur_trans->delayed_refs.dirty_extent_root = RB_ROOT;
 241         atomic_set(&cur_trans->delayed_refs.num_entries, 0);
 242         cur_trans->delayed_refs.num_heads_ready = 0;
 243         cur_trans->delayed_refs.pending_csums = 0;
 244         cur_trans->delayed_refs.num_heads = 0;
 245         cur_trans->delayed_refs.flushing = 0;
 246         cur_trans->delayed_refs.run_delayed_start = 0;
 247         cur_trans->delayed_refs.qgroup_to_skip = 0;
 248
 249         /*
 250          * although the tree mod log is per file system and not per transaction,
 251          * the log must never go across transaction boundaries.
 252          */
 253         smp_mb();
 254         if (!list_empty(&fs_info->tree_mod_seq_list))
 255                 WARN(1, KERN_ERR "BTRFS: tree_mod_seq_list not empty when "
 256                         "creating a fresh transaction\n");
 257         if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
 258                 WARN(1, KERN_ERR "BTRFS: tree_mod_log rb tree not empty when "
 259                         "creating a fresh transaction\n");
 260         atomic64_set(&fs_info->tree_mod_seq, 0);
 261
 262         spin_lock_init(&cur_trans->delayed_refs.lock);
 263
 264         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 265         INIT_LIST_HEAD(&cur_trans->pending_chunks);
 266         INIT_LIST_HEAD(&cur_trans->switch_commits);
 267         INIT_LIST_HEAD(&cur_trans->pending_ordered);
 268         INIT_LIST_HEAD(&cur_trans->dirty_bgs);
 269         INIT_LIST_HEAD(&cur_trans->io_bgs);
 270         INIT_LIST_HEAD(&cur_trans->dropped_roots);
 271         mutex_init(&cur_trans->cache_write_mutex);
 272         cur_trans->num_dirty_bgs = 0;
 273         spin_lock_init(&cur_trans->dirty_bgs_lock);
 274         INIT_LIST_HEAD(&cur_trans->deleted_bgs);
 275         spin_lock_init(&cur_trans->deleted_bgs_lock);
 276         spin_lock_init(&cur_trans->dropped_roots_lock);
 277         list_add_tail(&cur_trans->list, &fs_info->trans_list);
 278         extent_io_tree_init(&cur_trans->dirty_pages,
 279                              fs_info->btree_inode->i_mapping);
 280         fs_info->generation++;
 281         cur_trans->transid = fs_info->generation;
 282         fs_info->running_transaction = cur_trans;
 283         cur_trans->aborted = 0;
 284         spin_unlock(&fs_info->trans_lock);
 285
 286         return 0;
 287 }
 288
 289 /*
 290  * this does all the record keeping required to make sure that a reference
 291  * counted root is properly recorded in a given transaction.  This is required
 292  * to make sure the old root from before we joined the transaction is deleted
 293  * when the transaction commits
 294  */
 295 static int record_root_in_trans(struct btrfs_trans_handle *trans,
 296                                struct btrfs_root *root)
 297 {
 298         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
 299             root->last_trans < trans->transid) {
 300                 WARN_ON(root == root->fs_info->extent_root);
 301                 WARN_ON(root->commit_root != root->node);
 302
 303                 /*
 304                  * see below for IN_TRANS_SETUP usage rules
 305                  * we have the reloc mutex held now, so there
 306                  * is only one writer in this function
 307                  */
 308                 set_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
 309
 310                 /* make sure readers find IN_TRANS_SETUP before
 311                  * they find our root->last_trans update
 312                  */
 313                 smp_wmb();
 314
 315                 spin_lock(&root->fs_info->fs_roots_radix_lock);
 316                 if (root->last_trans == trans->transid) {
 317                         spin_unlock(&root->fs_info->fs_roots_radix_lock);
 318                         return 0;
 319                 }
 320                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 321                            (unsigned long)root->root_key.objectid,
 322                            BTRFS_ROOT_TRANS_TAG);
 323                 spin_unlock(&root->fs_info->fs_roots_radix_lock);
 324                 root->last_trans = trans->transid;
 325
 326                 /* this is pretty tricky.  We don't want to
 327                  * take the relocation lock in btrfs_record_root_in_trans
 328                  * unless we're really doing the first setup for this root in
 329                  * this transaction.
 330                  *
 331                  * Normally we'd use root->last_trans as a flag to decide
 332                  * if we want to take the expensive mutex.
 333                  *
 334                  * But, we have to set root->last_trans before we
 335                  * init the relocation root, otherwise, we trip over warnings
 336                  * in ctree.c.  The solution used here is to flag ourselves
 337                  * with root IN_TRANS_SETUP.  When this is 1, we're still
 338                  * fixing up the reloc trees and everyone must wait.
 339                  *
 340                  * When this is zero, they can trust root->last_trans and fly
 341                  * through btrfs_record_root_in_trans without having to take the
 342                  * lock.  smp_wmb() makes sure that all the writes above are
 343                  * done before we pop in the zero below
 344                  */
 345                 btrfs_init_reloc_root(trans, root);
 346                 smp_mb__before_atomic();
 347                 clear_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
 348         }
 349         return 0;
 350 }
 351
 352
 353 void btrfs_add_dropped_root(struct btrfs_trans_handle *trans,
 354                             struct btrfs_root *root)
 355 {
 356         struct btrfs_transaction *cur_trans = trans->transaction;
 357
 358         /* Add ourselves to the transaction dropped list */
 359         spin_lock(&cur_trans->dropped_roots_lock);
 360         list_add_tail(&root->root_list, &cur_trans->dropped_roots);
 361         spin_unlock(&cur_trans->dropped_roots_lock);
 362
 363         /* Make sure we don't try to update the root at commit time */
 364         spin_lock(&root->fs_info->fs_roots_radix_lock);
 365         radix_tree_tag_clear(&root->fs_info->fs_roots_radix,
 366                              (unsigned long)root->root_key.objectid,
 367                              BTRFS_ROOT_TRANS_TAG);
 368         spin_unlock(&root->fs_info->fs_roots_radix_lock);
 369 }
 370
 371 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 372                                struct btrfs_root *root)
 373 {
 374         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
 375                 return 0;
 376
 377         /*
 378          * see record_root_in_trans for comments about IN_TRANS_SETUP usage
 379          * and barriers
 380          */
 381         smp_rmb();
 382         if (root->last_trans == trans->transid &&
 383             !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state))
 384                 return 0;
 385
 386         mutex_lock(&root->fs_info->reloc_mutex);
 387         record_root_in_trans(trans, root);
 388         mutex_unlock(&root->fs_info->reloc_mutex);
 389
 390         return 0;
 391 }
 392
 393 static inline int is_transaction_blocked(struct btrfs_transaction *trans)
 394 {
 395         return (trans->state >= TRANS_STATE_BLOCKED &&
 396                 trans->state < TRANS_STATE_UNBLOCKED &&
 397                 !trans->aborted);
 398 }
 399
 400 /* wait for commit against the current transaction to become unblocked
 401  * when this is done, it is safe to start a new transaction, but the current
 402  * transaction might not be fully on disk.
 403  */
 404 static void wait_current_trans(struct btrfs_root *root)
 405 {
 406         struct btrfs_transaction *cur_trans;
 407
 408         spin_lock(&root->fs_info->trans_lock);
 409         cur_trans = root->fs_info->running_transaction;
 410         if (cur_trans && is_transaction_blocked(cur_trans)) {
 411                 atomic_inc(&cur_trans->use_count);
 412                 spin_unlock(&root->fs_info->trans_lock);
 413
 414                 wait_event(root->fs_info->transaction_wait,
 415                            cur_trans->state >= TRANS_STATE_UNBLOCKED ||
 416                            cur_trans->aborted);
 417                 btrfs_put_transaction(cur_trans);
 418         } else {
 419                 spin_unlock(&root->fs_info->trans_lock);
 420         }
 421 }
 422
 423 static int may_wait_transaction(struct btrfs_root *root, int type)
 424 {
 425         if (root->fs_info->log_root_recovering)
 426                 return 0;
 427
 428         if (type == TRANS_USERSPACE)
 429                 return 1;
 430
 431         if (type == TRANS_START &&
 432             !atomic_read(&root->fs_info->open_ioctl_trans))
 433                 return 1;
 434
 435         return 0;
 436 }
 437
 438 static inline bool need_reserve_reloc_root(struct btrfs_root *root)
 439 {
 440         if (!root->fs_info->reloc_ctl ||
 441             !test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
 442             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 443             root->reloc_root)
 444                 return false;
 445
 446         return true;
 447 }
 448
 449 static struct btrfs_trans_handle *
 450 start_transaction(struct btrfs_root *root, u64 num_items, unsigned int type,
 451                   enum btrfs_reserve_flush_enum flush)
 452 {
 453         struct btrfs_trans_handle *h;
 454         struct btrfs_transaction *cur_trans;
 455         u64 num_bytes = 0;
 456         u64 qgroup_reserved = 0;
 457         bool reloc_reserved = false;
 458         int ret;
 459
 460         /* Send isn't supposed to start transactions. */
 461         ASSERT(current->journal_info != BTRFS_SEND_TRANS_STUB);
 462
 463         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
 464                 return ERR_PTR(-EROFS);
 465
 466         if (current->journal_info) {
 467                 WARN_ON(type & TRANS_EXTWRITERS);
 468                 h = current->journal_info;
 469                 h->use_count++;
 470                 WARN_ON(h->use_count > 2);
 471                 h->orig_rsv = h->block_rsv;
 472                 h->block_rsv = NULL;
 473                 goto got_it;
 474         }
 475
 476         /*
 477          * Do the reservation before we join the transaction so we can do all
 478          * the appropriate flushing if need be.
 479          */
 480         if (num_items > 0 && root != root->fs_info->chunk_root) {
 481                 if (root->fs_info->quota_enabled &&
 482                     is_fstree(root->root_key.objectid)) {
 483                         qgroup_reserved = num_items * root->nodesize;
 484                         ret = btrfs_qgroup_reserve(root, qgroup_reserved);
 485                         if (ret)
 486                                 return ERR_PTR(ret);
 487                 }
 488
 489                 num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
 490                 /*
 491                  * Do the reservation for the relocation root creation
 492                  */
 493                 if (need_reserve_reloc_root(root)) {
 494                         num_bytes += root->nodesize;
 495                         reloc_reserved = true;
 496                 }
 497
 498                 ret = btrfs_block_rsv_add(root,
 499                                           &root->fs_info->trans_block_rsv,
 500                                           num_bytes, flush);
 501                 if (ret)
 502                         goto reserve_fail;
 503         }
 504 again:
 505         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 506         if (!h) {
 507                 ret = -ENOMEM;
 508                 goto alloc_fail;
 509         }
 510
 511         /*
 512          * If we are JOIN_NOLOCK we're already committing a transaction and
 513          * waiting on this guy, so we don't need to do the sb_start_intwrite
 514          * because we're already holding a ref.  We need this because we could
 515          * have raced in and did an fsync() on a file which can kick a commit
 516          * and then we deadlock with somebody doing a freeze.
 517          *
 518          * If we are ATTACH, it means we just want to catch the current
 519          * transaction and commit it, so we needn't do sb_start_intwrite().
 520          */
 521         if (type & __TRANS_FREEZABLE)
 522                 sb_start_intwrite(root->fs_info->sb);
 523
 524         if (may_wait_transaction(root, type))
 525                 wait_current_trans(root);
 526
 527         do {
 528                 ret = join_transaction(root, type);
 529                 if (ret == -EBUSY) {
 530                         wait_current_trans(root);
 531                         if (unlikely(type == TRANS_ATTACH))
 532                                 ret = -ENOENT;
 533                 }
 534         } while (ret == -EBUSY);
 535
 536         if (ret < 0) {
 537                 /* We must get the transaction if we are JOIN_NOLOCK. */
 538                 BUG_ON(type == TRANS_JOIN_NOLOCK);
 539                 goto join_fail;
 540         }
 541
 542         cur_trans = root->fs_info->running_transaction;
 543
 544         h->transid = cur_trans->transid;
 545         h->transaction = cur_trans;
 546         h->blocks_used = 0;
 547         h->bytes_reserved = 0;
 548         h->chunk_bytes_reserved = 0;
 549         h->root = root;
 550         h->delayed_ref_updates = 0;
 551         h->use_count = 1;
 552         h->adding_csums = 0;
 553         h->block_rsv = NULL;
 554         h->orig_rsv = NULL;
 555         h->aborted = 0;
 556         h->qgroup_reserved = 0;
 557         h->delayed_ref_elem.seq = 0;
 558         h->type = type;
 559         h->allocating_chunk = false;
 560         h->can_flush_pending_bgs = true;
 561         h->reloc_reserved = false;
 562         h->sync = false;
 563         INIT_LIST_HEAD(&h->qgroup_ref_list);
 564         INIT_LIST_HEAD(&h->new_bgs);
 565         INIT_LIST_HEAD(&h->ordered);
 566
 567         smp_mb();
 568         if (cur_trans->state >= TRANS_STATE_BLOCKED &&
 569             may_wait_transaction(root, type)) {
 570                 current->journal_info = h;
 571                 btrfs_commit_transaction(h, root);
 572                 goto again;
 573         }
 574
 575         if (num_bytes) {
 576                 trace_btrfs_space_reservation(root->fs_info, "transaction",
 577                                               h->transid, num_bytes, 1);
 578                 h->block_rsv = &root->fs_info->trans_block_rsv;
 579                 h->bytes_reserved = num_bytes;
 580                 h->reloc_reserved = reloc_reserved;
 581         }
 582         h->qgroup_reserved = qgroup_reserved;
 583
 584 got_it:
 585         btrfs_record_root_in_trans(h, root);
 586
 587         if (!current->journal_info && type != TRANS_USERSPACE)
 588                 current->journal_info = h;
 589         return h;
 590
 591 join_fail:
 592         if (type & __TRANS_FREEZABLE)
 593                 sb_end_intwrite(root->fs_info->sb);
 594         kmem_cache_free(btrfs_trans_handle_cachep, h);
 595 alloc_fail:
 596         if (num_bytes)
 597                 btrfs_block_rsv_release(root, &root->fs_info->trans_block_rsv,
 598                                         num_bytes);
 599 reserve_fail:
 600         if (qgroup_reserved)
 601                 btrfs_qgroup_free(root, qgroup_reserved);
 602         return ERR_PTR(ret);
 603 }
 604
 605 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 606                                                    int num_items)
 607 {
 608         return start_transaction(root, num_items, TRANS_START,
 609                                  BTRFS_RESERVE_FLUSH_ALL);
 610 }
 611
 612 struct btrfs_trans_handle *btrfs_start_transaction_lflush(
 613                                         struct btrfs_root *root, int num_items)
 614 {
 615         return start_transaction(root, num_items, TRANS_START,
 616                                  BTRFS_RESERVE_FLUSH_LIMIT);
 617 }
 618
 619 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 620 {
 621         return start_transaction(root, 0, TRANS_JOIN, 0);
 622 }
 623
 624 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 625 {
 626         return start_transaction(root, 0, TRANS_JOIN_NOLOCK, 0);
 627 }
 628
 629 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 630 {
 631         return start_transaction(root, 0, TRANS_USERSPACE, 0);
 632 }
 633
 634 /*
 635  * btrfs_attach_transaction() - catch the running transaction
 636  *
 637  * It is used when we want to commit the current the transaction, but
 638  * don't want to start a new one.
 639  *
 640  * Note: If this function return -ENOENT, it just means there is no
 641  * running transaction. But it is possible that the inactive transaction
 642  * is still in the memory, not fully on disk. If you hope there is no
 643  * inactive transaction in the fs when -ENOENT is returned, you should
 644  * invoke
 645  *     btrfs_attach_transaction_barrier()
 646  */
 647 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
 648 {
 649         return start_transaction(root, 0, TRANS_ATTACH, 0);
 650 }
 651
 652 /*
 653  * btrfs_attach_transaction_barrier() - catch the running transaction
 654  *
 655  * It is similar to the above function, the differentia is this one
 656  * will wait for all the inactive transactions until they fully
 657  * complete.
 658  */
 659 struct btrfs_trans_handle *
 660 btrfs_attach_transaction_barrier(struct btrfs_root *root)
 661 {
 662         struct btrfs_trans_handle *trans;
 663
 664         trans = start_transaction(root, 0, TRANS_ATTACH, 0);
 665         if (IS_ERR(trans) && PTR_ERR(trans) == -ENOENT)
 666                 btrfs_wait_for_commit(root, 0);
 667
 668         return trans;
 669 }
 670
 671 /* wait for a transaction commit to be fully complete */
 672 static noinline void wait_for_commit(struct btrfs_root *root,
 673                                     struct btrfs_transaction *commit)
 674 {
 675         wait_event(commit->commit_wait, commit->state == TRANS_STATE_COMPLETED);
 676 }
 677
 678 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 679 {
 680         struct btrfs_transaction *cur_trans = NULL, *t;
 681         int ret = 0;
 682
 683         if (transid) {
 684                 if (transid <= root->fs_info->last_trans_committed)
 685                         goto out;
 686
 687                 /* find specified transaction */
 688                 spin_lock(&root->fs_info->trans_lock);
 689                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 690                         if (t->transid == transid) {
 691                                 cur_trans = t;
 692                                 atomic_inc(&cur_trans->use_count);
 693                                 ret = 0;
 694                                 break;
 695                         }
 696                         if (t->transid > transid) {
 697                                 ret = 0;
 698                                 break;
 699                         }
 700                 }
 701                 spin_unlock(&root->fs_info->trans_lock);
 702
 703                 /*
 704                  * The specified transaction doesn't exist, or we
 705                  * raced with btrfs_commit_transaction
 706                  */
 707                 if (!cur_trans) {
 708                         if (transid > root->fs_info->last_trans_committed)
 709                                 ret = -EINVAL;
 710                         goto out;
 711                 }
 712         } else {
 713                 /* find newest transaction that is committing | committed */
 714                 spin_lock(&root->fs_info->trans_lock);
 715                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 716                                             list) {
 717                         if (t->state >= TRANS_STATE_COMMIT_START) {
 718                                 if (t->state == TRANS_STATE_COMPLETED)
 719                                         break;
 720                                 cur_trans = t;
 721                                 atomic_inc(&cur_trans->use_count);
 722                                 break;
 723                         }
 724                 }
 725                 spin_unlock(&root->fs_info->trans_lock);
 726                 if (!cur_trans)
 727                         goto out;  /* nothing committing|committed */
 728         }
 729
 730         wait_for_commit(root, cur_trans);
 731         btrfs_put_transaction(cur_trans);
 732 out:
 733         return ret;
 734 }
 735
 736 void btrfs_throttle(struct btrfs_root *root)
 737 {
 738         if (!atomic_read(&root->fs_info->open_ioctl_trans))
 739                 wait_current_trans(root);
 740 }
 741
 742 static int should_end_transaction(struct btrfs_trans_handle *trans,
 743                                   struct btrfs_root *root)
 744 {
 745         if (root->fs_info->global_block_rsv.space_info->full &&
 746             btrfs_check_space_for_delayed_refs(trans, root))
 747                 return 1;
 748
 749         return !!btrfs_block_rsv_check(root, &root->fs_info->global_block_rsv, 5);
 750 }
 751
 752 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 753                                  struct btrfs_root *root)
 754 {
 755         struct btrfs_transaction *cur_trans = trans->transaction;
 756         int updates;
 757         int err;
 758
 759         smp_mb();
 760         if (cur_trans->state >= TRANS_STATE_BLOCKED ||
 761             cur_trans->delayed_refs.flushing)
 762                 return 1;
 763
 764         updates = trans->delayed_ref_updates;
 765         trans->delayed_ref_updates = 0;
 766         if (updates) {
 767                 err = btrfs_run_delayed_refs(trans, root, updates * 2);
 768                 if (err) /* Error code will also eval true */
 769                         return err;
 770         }
 771
 772         return should_end_transaction(trans, root);
 773 }
 774
 775 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 776                           struct btrfs_root *root, int throttle)
 777 {
 778         struct btrfs_transaction *cur_trans = trans->transaction;
 779         struct btrfs_fs_info *info = root->fs_info;
 780         unsigned long cur = trans->delayed_ref_updates;
 781         int lock = (trans->type != TRANS_JOIN_NOLOCK);
 782         int err = 0;
 783         int must_run_delayed_refs = 0;
 784
 785         if (trans->use_count > 1) {
 786                 trans->use_count--;
 787                 trans->block_rsv = trans->orig_rsv;
 788                 return 0;
 789         }
 790
 791         btrfs_trans_release_metadata(trans, root);
 792         trans->block_rsv = NULL;
 793
 794         if (!list_empty(&trans->new_bgs))
 795                 btrfs_create_pending_block_groups(trans, root);
 796
 797         if (!list_empty(&trans->ordered)) {
 798                 spin_lock(&info->trans_lock);
 799                 list_splice_init(&trans->ordered, &cur_trans->pending_ordered);
 800                 spin_unlock(&info->trans_lock);
 801         }
 802
 803         trans->delayed_ref_updates = 0;
 804         if (!trans->sync) {
 805                 must_run_delayed_refs =
 806                         btrfs_should_throttle_delayed_refs(trans, root);
 807                 cur = max_t(unsigned long, cur, 32);
 808
 809                 /*
 810                  * don't make the caller wait if they are from a NOLOCK
 811                  * or ATTACH transaction, it will deadlock with commit
 812                  */
 813                 if (must_run_delayed_refs == 1 &&
 814                     (trans->type & (__TRANS_JOIN_NOLOCK | __TRANS_ATTACH)))
 815                         must_run_delayed_refs = 2;
 816         }
 817
 818         if (trans->qgroup_reserved) {
 819                 /*
 820                  * the same root has to be passed here between start_transaction
 821                  * and end_transaction. Subvolume quota depends on this.
 822                  */
 823                 btrfs_qgroup_free(trans->root, trans->qgroup_reserved);
 824                 trans->qgroup_reserved = 0;
 825         }
 826
 827         btrfs_trans_release_metadata(trans, root);
 828         trans->block_rsv = NULL;
 829
 830         if (!list_empty(&trans->new_bgs))
 831                 btrfs_create_pending_block_groups(trans, root);
 832
 833         btrfs_trans_release_chunk_metadata(trans);
 834
 835         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
 836             should_end_transaction(trans, root) &&
 837             ACCESS_ONCE(cur_trans->state) == TRANS_STATE_RUNNING) {
 838                 spin_lock(&info->trans_lock);
 839                 if (cur_trans->state == TRANS_STATE_RUNNING)
 840                         cur_trans->state = TRANS_STATE_BLOCKED;
 841                 spin_unlock(&info->trans_lock);
 842         }
 843
 844         if (lock && ACCESS_ONCE(cur_trans->state) == TRANS_STATE_BLOCKED) {
 845                 if (throttle)
 846                         return btrfs_commit_transaction(trans, root);
 847                 else
 848                         wake_up_process(info->transaction_kthread);
 849         }
 850
 851         if (trans->type & __TRANS_FREEZABLE)
 852                 sb_end_intwrite(root->fs_info->sb);
 853
 854         WARN_ON(cur_trans != info->running_transaction);
 855         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 856         atomic_dec(&cur_trans->num_writers);
 857         extwriter_counter_dec(cur_trans, trans->type);
 858
 859         smp_mb();
 860         if (waitqueue_active(&cur_trans->writer_wait))
 861                 wake_up(&cur_trans->writer_wait);
 862         btrfs_put_transaction(cur_trans);
 863
 864         if (current->journal_info == trans)
 865                 current->journal_info = NULL;
 866
 867         if (throttle)
 868                 btrfs_run_delayed_iputs(root);
 869
 870         if (trans->aborted ||
 871             test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state)) {
 872                 wake_up_process(info->transaction_kthread);
 873                 err = -EIO;
 874         }
 875         assert_qgroups_uptodate(trans);
 876
 877         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 878         if (must_run_delayed_refs) {
 879                 btrfs_async_run_delayed_refs(root, cur,
 880                                              must_run_delayed_refs == 1);
 881         }
 882         return err;
 883 }
 884
 885 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 886                           struct btrfs_root *root)
 887 {
 888         return __btrfs_end_transaction(trans, root, 0);
 889 }
 890
 891 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 892                                    struct btrfs_root *root)
 893 {
 894         return __btrfs_end_transaction(trans, root, 1);
 895 }
 896
 897 /*
 898  * when btree blocks are allocated, they have some corresponding bits set for
 899  * them in one of two extent_io trees.  This is used to make sure all of
 900  * those extents are sent to disk but does not wait on them
 901  */
 902 int btrfs_write_marked_extents(struct btrfs_root *root,
 903                                struct extent_io_tree *dirty_pages, int mark)
 904 {
 905         int err = 0;
 906         int werr = 0;
 907         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 908         struct extent_state *cached_state = NULL;
 909         u64 start = 0;
 910         u64 end;
 911
 912         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 913                                       mark, &cached_state)) {
 914                 bool wait_writeback = false;
 915
 916                 err = convert_extent_bit(dirty_pages, start, end,
 917                                          EXTENT_NEED_WAIT,
 918                                          mark, &cached_state, GFP_NOFS);
 919                 /*
 920                  * convert_extent_bit can return -ENOMEM, which is most of the
 921                  * time a temporary error. So when it happens, ignore the error
 922                  * and wait for writeback of this range to finish - because we
 923                  * failed to set the bit EXTENT_NEED_WAIT for the range, a call
 924                  * to btrfs_wait_marked_extents() would not know that writeback
 925                  * for this range started and therefore wouldn't wait for it to
 926                  * finish - we don't want to commit a superblock that points to
 927                  * btree nodes/leafs for which writeback hasn't finished yet
 928                  * (and without errors).
 929                  * We cleanup any entries left in the io tree when committing
 930                  * the transaction (through clear_btree_io_tree()).
 931                  */
 932                 if (err == -ENOMEM) {
 933                         err = 0;
 934                         wait_writeback = true;
 935                 }
 936                 if (!err)
 937                         err = filemap_fdatawrite_range(mapping, start, end);
 938                 if (err)
 939                         werr = err;
 940                 else if (wait_writeback)
 941                         werr = filemap_fdatawait_range(mapping, start, end);
 942                 free_extent_state(cached_state);
 943                 cached_state = NULL;
 944                 cond_resched();
 945                 start = end + 1;
 946         }
 947         return werr;
 948 }
 949
 950 /*
 951  * when btree blocks are allocated, they have some corresponding bits set for
 952  * them in one of two extent_io trees.  This is used to make sure all of
 953  * those extents are on disk for transaction or log commit.  We wait
 954  * on all the pages and clear them from the dirty pages state tree
 955  */
 956 int btrfs_wait_marked_extents(struct btrfs_root *root,
 957                               struct extent_io_tree *dirty_pages, int mark)
 958 {
 959         int err = 0;
 960         int werr = 0;
 961         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 962         struct extent_state *cached_state = NULL;
 963         u64 start = 0;
 964         u64 end;
 965         struct btrfs_inode *btree_ino = BTRFS_I(root->fs_info->btree_inode);
 966         bool errors = false;
 967
 968         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 969                                       EXTENT_NEED_WAIT, &cached_state)) {
 970                 /*
 971                  * Ignore -ENOMEM errors returned by clear_extent_bit().
 972                  * When committing the transaction, we'll remove any entries
 973                  * left in the io tree. For a log commit, we don't remove them
 974                  * after committing the log because the tree can be accessed
 975                  * concurrently - we do it only at transaction commit time when
 976                  * it's safe to do it (through clear_btree_io_tree()).
 977                  */
 978                 err = clear_extent_bit(dirty_pages, start, end,
 979                                        EXTENT_NEED_WAIT,
 980                                        0, 0, &cached_state, GFP_NOFS);
 981                 if (err == -ENOMEM)
 982                         err = 0;
 983                 if (!err)
 984                         err = filemap_fdatawait_range(mapping, start, end);
 985                 if (err)
 986                         werr = err;
 987                 free_extent_state(cached_state);
 988                 cached_state = NULL;
 989                 cond_resched();
 990                 start = end + 1;
 991         }
 992         if (err)
 993                 werr = err;
 994
 995         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
 996                 if ((mark & EXTENT_DIRTY) &&
 997                     test_and_clear_bit(BTRFS_INODE_BTREE_LOG1_ERR,
 998                                        &btree_ino->runtime_flags))
 999                         errors = true;
1000
1001                 if ((mark & EXTENT_NEW) &&
1002                     test_and_clear_bit(BTRFS_INODE_BTREE_LOG2_ERR,
1003                                        &btree_ino->runtime_flags))
1004                         errors = true;
1005         } else {
1006                 if (test_and_clear_bit(BTRFS_INODE_BTREE_ERR,
1007                                        &btree_ino->runtime_flags))
1008                         errors = true;
1009         }
1010
1011         if (errors && !werr)
1012                 werr = -EIO;
1013
1014         return werr;
1015 }
1016
1017 /*
1018  * when btree blocks are allocated, they have some corresponding bits set for
1019  * them in one of two extent_io trees.  This is used to make sure all of
1020  * those extents are on disk for transaction or log commit
1021  */
1022 static int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
1023                                 struct extent_io_tree *dirty_pages, int mark)
1024 {
1025         int ret;
1026         int ret2;
1027         struct blk_plug plug;
1028
1029         blk_start_plug(&plug);
1030         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
1031         blk_finish_plug(&plug);
1032         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
1033
1034         if (ret)
1035                 return ret;
1036         if (ret2)
1037                 return ret2;
1038         return 0;
1039 }
1040
1041 static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
1042                                      struct btrfs_root *root)
1043 {
1044         int ret;
1045
1046         ret = btrfs_write_and_wait_marked_extents(root,
1047                                            &trans->transaction->dirty_pages,
1048                                            EXTENT_DIRTY);
1049         clear_btree_io_tree(&trans->transaction->dirty_pages);
1050
1051         return ret;
1052 }
1053
1054 /*
1055  * this is used to update the root pointer in the tree of tree roots.
1056  *
1057  * But, in the case of the extent allocation tree, updating the root
1058  * pointer may allocate blocks which may change the root of the extent
1059  * allocation tree.
1060  *
1061  * So, this loops and repeats and makes sure the cowonly root didn't
1062  * change while the root pointer was being updated in the metadata.
1063  */
1064 static int update_cowonly_root(struct btrfs_trans_handle *trans,
1065                                struct btrfs_root *root)
1066 {
1067         int ret;
1068         u64 old_root_bytenr;
1069         u64 old_root_used;
1070         struct btrfs_root *tree_root = root->fs_info->tree_root;
1071
1072         old_root_used = btrfs_root_used(&root->root_item);
1073
1074         while (1) {
1075                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1076                 if (old_root_bytenr == root->node->start &&
1077                     old_root_used == btrfs_root_used(&root->root_item))
1078                         break;
1079
1080                 btrfs_set_root_node(&root->root_item, root->node);
1081                 ret = btrfs_update_root(trans, tree_root,
1082                                         &root->root_key,
1083                                         &root->root_item);
1084                 if (ret)
1085                         return ret;
1086
1087                 old_root_used = btrfs_root_used(&root->root_item);
1088         }
1089
1090         return 0;
1091 }
1092
1093 /*
1094  * update all the cowonly tree roots on disk
1095  *
1096  * The error handling in this function may not be obvious. Any of the
1097  * failures will cause the file system to go offline. We still need
1098  * to clean up the delayed refs.
1099  */
1100 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
1101                                          struct btrfs_root *root)
1102 {
1103         struct btrfs_fs_info *fs_info = root->fs_info;
1104         struct list_head *dirty_bgs = &trans->transaction->dirty_bgs;
1105         struct list_head *io_bgs = &trans->transaction->io_bgs;
1106         struct list_head *next;
1107         struct extent_buffer *eb;
1108         int ret;
1109
1110         eb = btrfs_lock_root_node(fs_info->tree_root);
1111         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
1112                               0, &eb);
1113         btrfs_tree_unlock(eb);
1114         free_extent_buffer(eb);
1115
1116         if (ret)
1117                 return ret;
1118
1119         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1120         if (ret)
1121                 return ret;
1122
1123         ret = btrfs_run_dev_stats(trans, root->fs_info);
1124         if (ret)
1125                 return ret;
1126         ret = btrfs_run_dev_replace(trans, root->fs_info);
1127         if (ret)
1128                 return ret;
1129         ret = btrfs_run_qgroups(trans, root->fs_info);
1130         if (ret)
1131                 return ret;
1132
1133         ret = btrfs_setup_space_cache(trans, root);
1134         if (ret)
1135                 return ret;
1136
1137         /* run_qgroups might have added some more refs */
1138         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1139         if (ret)
1140                 return ret;
1141 again:
1142         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1143                 next = fs_info->dirty_cowonly_roots.next;
1144                 list_del_init(next);
1145                 root = list_entry(next, struct btrfs_root, dirty_list);
1146                 clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1147
1148                 if (root != fs_info->extent_root)
1149                         list_add_tail(&root->dirty_list,
1150                                       &trans->transaction->switch_commits);
1151                 ret = update_cowonly_root(trans, root);
1152                 if (ret)
1153                         return ret;
1154                 ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1155                 if (ret)
1156                         return ret;
1157         }
1158
1159         while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1160                 ret = btrfs_write_dirty_block_groups(trans, root);
1161                 if (ret)
1162                         return ret;
1163                 ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1164                 if (ret)
1165                         return ret;
1166         }
1167
1168         if (!list_empty(&fs_info->dirty_cowonly_roots))
1169                 goto again;
1170
1171         list_add_tail(&fs_info->extent_root->dirty_list,
1172                       &trans->transaction->switch_commits);
1173         btrfs_after_dev_replace_commit(fs_info);
1174
1175         return 0;
1176 }
1177
1178 /*
1179  * dead roots are old snapshots that need to be deleted.  This allocates
1180  * a dirty root struct and adds it into the list of dead roots that need to
1181  * be deleted
1182  */
1183 void btrfs_add_dead_root(struct btrfs_root *root)
1184 {
1185         spin_lock(&root->fs_info->trans_lock);
1186         if (list_empty(&root->root_list))
1187                 list_add_tail(&root->root_list, &root->fs_info->dead_roots);
1188         spin_unlock(&root->fs_info->trans_lock);
1189 }
1190
1191 /*
1192  * update all the cowonly tree roots on disk
1193  */
1194 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
1195                                     struct btrfs_root *root)
1196 {
1197         struct btrfs_root *gang[8];
1198         struct btrfs_fs_info *fs_info = root->fs_info;
1199         int i;
1200         int ret;
1201         int err = 0;
1202
1203         spin_lock(&fs_info->fs_roots_radix_lock);
1204         while (1) {
1205                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1206                                                  (void **)gang, 0,
1207                                                  ARRAY_SIZE(gang),
1208                                                  BTRFS_ROOT_TRANS_TAG);
1209                 if (ret == 0)
1210                         break;
1211                 for (i = 0; i < ret; i++) {
1212                         root = gang[i];
1213                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
1214                                         (unsigned long)root->root_key.objectid,
1215                                         BTRFS_ROOT_TRANS_TAG);
1216                         spin_unlock(&fs_info->fs_roots_radix_lock);
1217
1218                         btrfs_free_log(trans, root);
1219                         btrfs_update_reloc_root(trans, root);
1220                         btrfs_orphan_commit_root(trans, root);
1221
1222                         btrfs_save_ino_cache(root, trans);
1223
1224                         /* see comments in should_cow_block() */
1225                         clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1226                         smp_mb__after_atomic();
1227
1228                         if (root->commit_root != root->node) {
1229                                 list_add_tail(&root->dirty_list,
1230                                         &trans->transaction->switch_commits);
1231                                 btrfs_set_root_node(&root->root_item,
1232                                                     root->node);
1233                         }
1234
1235                         err = btrfs_update_root(trans, fs_info->tree_root,
1236                                                 &root->root_key,
1237                                                 &root->root_item);
1238                         spin_lock(&fs_info->fs_roots_radix_lock);
1239                         if (err)
1240                                 break;
1241                 }
1242         }
1243         spin_unlock(&fs_info->fs_roots_radix_lock);
1244         return err;
1245 }
1246
1247 /*
1248  * defrag a given btree.
1249  * Every leaf in the btree is read and defragged.
1250  */
1251 int btrfs_defrag_root(struct btrfs_root *root)
1252 {
1253         struct btrfs_fs_info *info = root->fs_info;
1254         struct btrfs_trans_handle *trans;
1255         int ret;
1256
1257         if (test_and_set_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state))
1258                 return 0;
1259
1260         while (1) {
1261                 trans = btrfs_start_transaction(root, 0);
1262                 if (IS_ERR(trans))
1263                         return PTR_ERR(trans);
1264
1265                 ret = btrfs_defrag_leaves(trans, root);
1266
1267                 btrfs_end_transaction(trans, root);
1268                 btrfs_btree_balance_dirty(info->tree_root);
1269                 cond_resched();
1270
1271                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
1272                         break;
1273
1274                 if (btrfs_defrag_cancelled(root->fs_info)) {
1275                         pr_debug("BTRFS: defrag_root cancelled\n");
1276                         ret = -EAGAIN;
1277                         break;
1278                 }
1279         }
1280         clear_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state);
1281         return ret;
1282 }
1283
1284 /*
1285  * new snapshots need to be created at a very specific time in the
1286  * transaction commit.  This does the actual creation.
1287  *
1288  * Note:
1289  * If the error which may affect the commitment of the current transaction
1290  * happens, we should return the error number. If the error which just affect
1291  * the creation of the pending snapshots, just return 0.
1292  */
1293 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1294                                    struct btrfs_fs_info *fs_info,
1295                                    struct btrfs_pending_snapshot *pending)
1296 {
1297         struct btrfs_key key;
1298         struct btrfs_root_item *new_root_item;
1299         struct btrfs_root *tree_root = fs_info->tree_root;
1300         struct btrfs_root *root = pending->root;
1301         struct btrfs_root *parent_root;
1302         struct btrfs_block_rsv *rsv;
1303         struct inode *parent_inode;
1304         struct btrfs_path *path;
1305         struct btrfs_dir_item *dir_item;
1306         struct dentry *dentry;
1307         struct extent_buffer *tmp;
1308         struct extent_buffer *old;
1309         struct timespec cur_time = CURRENT_TIME;
1310         int ret = 0;
1311         u64 to_reserve = 0;
1312         u64 index = 0;
1313         u64 objectid;
1314         u64 root_flags;
1315         uuid_le new_uuid;
1316
1317         path = btrfs_alloc_path();
1318         if (!path) {
1319                 pending->error = -ENOMEM;
1320                 return 0;
1321         }
1322
1323         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
1324         if (!new_root_item) {
1325                 pending->error = -ENOMEM;
1326                 goto root_item_alloc_fail;
1327         }
1328
1329         pending->error = btrfs_find_free_objectid(tree_root, &objectid);
1330         if (pending->error)
1331                 goto no_free_objectid;
1332
1333         /*
1334          * Make qgroup to skip current new snapshot's qgroupid, as it is
1335          * accounted by later btrfs_qgroup_inherit().
1336          */
1337         btrfs_set_skip_qgroup(trans, objectid);
1338
1339         btrfs_reloc_pre_snapshot(pending, &to_reserve);
1340
1341         if (to_reserve > 0) {
1342                 pending->error = btrfs_block_rsv_add(root,
1343                                                      &pending->block_rsv,
1344                                                      to_reserve,
1345                                                      BTRFS_RESERVE_NO_FLUSH);
1346                 if (pending->error)
1347                         goto clear_skip_qgroup;
1348         }
1349
1350         key.objectid = objectid;
1351         key.offset = (u64)-1;
1352         key.type = BTRFS_ROOT_ITEM_KEY;
1353
1354         rsv = trans->block_rsv;
1355         trans->block_rsv = &pending->block_rsv;
1356         trans->bytes_reserved = trans->block_rsv->reserved;
1357
1358         dentry = pending->dentry;
1359         parent_inode = pending->dir;
1360         parent_root = BTRFS_I(parent_inode)->root;
1361         record_root_in_trans(trans, parent_root);
1362
1363         /*
1364          * insert the directory item
1365          */
1366         ret = btrfs_set_inode_index(parent_inode, &index);
1367         BUG_ON(ret); /* -ENOMEM */
1368
1369         /* check if there is a file/dir which has the same name. */
1370         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1371                                          btrfs_ino(parent_inode),
1372                                          dentry->d_name.name,
1373                                          dentry->d_name.len, 0);
1374         if (dir_item != NULL && !IS_ERR(dir_item)) {
1375                 pending->error = -EEXIST;
1376                 goto dir_item_existed;
1377         } else if (IS_ERR(dir_item)) {
1378                 ret = PTR_ERR(dir_item);
1379                 btrfs_abort_transaction(trans, root, ret);
1380                 goto fail;
1381         }
1382         btrfs_release_path(path);
1383
1384         /*
1385          * pull in the delayed directory update
1386          * and the delayed inode item
1387          * otherwise we corrupt the FS during
1388          * snapshot
1389          */
1390         ret = btrfs_run_delayed_items(trans, root);
1391         if (ret) {      /* Transaction aborted */
1392                 btrfs_abort_transaction(trans, root, ret);
1393                 goto fail;
1394         }
1395
1396         record_root_in_trans(trans, root);
1397         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1398         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1399         btrfs_check_and_init_root_item(new_root_item);
1400
1401         root_flags = btrfs_root_flags(new_root_item);
1402         if (pending->readonly)
1403                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1404         else
1405                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1406         btrfs_set_root_flags(new_root_item, root_flags);
1407
1408         btrfs_set_root_generation_v2(new_root_item,
1409                         trans->transid);
1410         uuid_le_gen(&new_uuid);
1411         memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1412         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1413                         BTRFS_UUID_SIZE);
1414         if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1415                 memset(new_root_item->received_uuid, 0,
1416                        sizeof(new_root_item->received_uuid));
1417                 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1418                 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1419                 btrfs_set_root_stransid(new_root_item, 0);
1420                 btrfs_set_root_rtransid(new_root_item, 0);
1421         }
1422         btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1423         btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1424         btrfs_set_root_otransid(new_root_item, trans->transid);
1425
1426         old = btrfs_lock_root_node(root);
1427         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1428         if (ret) {
1429                 btrfs_tree_unlock(old);
1430                 free_extent_buffer(old);
1431                 btrfs_abort_transaction(trans, root, ret);
1432                 goto fail;
1433         }
1434
1435         btrfs_set_lock_blocking(old);
1436
1437         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1438         /* clean up in any case */
1439         btrfs_tree_unlock(old);
1440         free_extent_buffer(old);
1441         if (ret) {
1442                 btrfs_abort_transaction(trans, root, ret);
1443                 goto fail;
1444         }
1445         /* see comments in should_cow_block() */
1446         set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1447         smp_wmb();
1448
1449         btrfs_set_root_node(new_root_item, tmp);
1450         /* record when the snapshot was created in key.offset */
1451         key.offset = trans->transid;
1452         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1453         btrfs_tree_unlock(tmp);
1454         free_extent_buffer(tmp);
1455         if (ret) {
1456                 btrfs_abort_transaction(trans, root, ret);
1457                 goto fail;
1458         }
1459
1460         /*
1461          * insert root back/forward references
1462          */
1463         ret = btrfs_add_root_ref(trans, tree_root, objectid,
1464                                  parent_root->root_key.objectid,
1465                                  btrfs_ino(parent_inode), index,
1466                                  dentry->d_name.name, dentry->d_name.len);
1467         if (ret) {
1468                 btrfs_abort_transaction(trans, root, ret);
1469                 goto fail;
1470         }
1471
1472         key.offset = (u64)-1;
1473         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1474         if (IS_ERR(pending->snap)) {
1475                 ret = PTR_ERR(pending->snap);
1476                 btrfs_abort_transaction(trans, root, ret);
1477                 goto fail;
1478         }
1479
1480         ret = btrfs_reloc_post_snapshot(trans, pending);
1481         if (ret) {
1482                 btrfs_abort_transaction(trans, root, ret);
1483                 goto fail;
1484         }
1485
1486         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1487         if (ret) {
1488                 btrfs_abort_transaction(trans, root, ret);
1489                 goto fail;
1490         }
1491
1492         ret = btrfs_insert_dir_item(trans, parent_root,
1493                                     dentry->d_name.name, dentry->d_name.len,
1494                                     parent_inode, &key,
1495                                     BTRFS_FT_DIR, index);
1496         /* We have check then name at the beginning, so it is impossible. */
1497         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1498         if (ret) {
1499                 btrfs_abort_transaction(trans, root, ret);
1500                 goto fail;
1501         }
1502
1503         btrfs_i_size_write(parent_inode, parent_inode->i_size +
1504                                          dentry->d_name.len * 2);
1505         parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
1506         ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1507         if (ret) {
1508                 btrfs_abort_transaction(trans, root, ret);
1509                 goto fail;
1510         }
1511         ret = btrfs_uuid_tree_add(trans, fs_info->uuid_root, new_uuid.b,
1512                                   BTRFS_UUID_KEY_SUBVOL, objectid);
1513         if (ret) {
1514                 btrfs_abort_transaction(trans, root, ret);
1515                 goto fail;
1516         }
1517         if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1518                 ret = btrfs_uuid_tree_add(trans, fs_info->uuid_root,
1519                                           new_root_item->received_uuid,
1520                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1521                                           objectid);
1522                 if (ret && ret != -EEXIST) {
1523                         btrfs_abort_transaction(trans, root, ret);
1524                         goto fail;
1525                 }
1526         }
1527
1528         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1529         if (ret) {
1530                 btrfs_abort_transaction(trans, root, ret);
1531                 goto fail;
1532         }
1533
1534         /*
1535          * account qgroup counters before qgroup_inherit()
1536          */
1537         ret = btrfs_qgroup_prepare_account_extents(trans, fs_info);
1538         if (ret)
1539                 goto fail;
1540         ret = btrfs_qgroup_account_extents(trans, fs_info);
1541         if (ret)
1542                 goto fail;
1543         ret = btrfs_qgroup_inherit(trans, fs_info,
1544                                    root->root_key.objectid,
1545                                    objectid, pending->inherit);
1546         if (ret) {
1547                 btrfs_abort_transaction(trans, root, ret);
1548                 goto fail;
1549         }
1550
1551 fail:
1552         pending->error = ret;
1553 dir_item_existed:
1554         trans->block_rsv = rsv;
1555         trans->bytes_reserved = 0;
1556 clear_skip_qgroup:
1557         btrfs_clear_skip_qgroup(trans);
1558 no_free_objectid:
1559         kfree(new_root_item);
1560 root_item_alloc_fail:
1561         btrfs_free_path(path);
1562         return ret;
1563 }
1564
1565 /*
1566  * create all the snapshots we've scheduled for creation
1567  */
1568 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1569                                              struct btrfs_fs_info *fs_info)
1570 {
1571         struct btrfs_pending_snapshot *pending, *next;
1572         struct list_head *head = &trans->transaction->pending_snapshots;
1573         int ret = 0;
1574
1575         list_for_each_entry_safe(pending, next, head, list) {
1576                 list_del(&pending->list);
1577                 ret = create_pending_snapshot(trans, fs_info, pending);
1578                 if (ret)
1579                         break;
1580         }
1581         return ret;
1582 }
1583
1584 static void update_super_roots(struct btrfs_root *root)
1585 {
1586         struct btrfs_root_item *root_item;
1587         struct btrfs_super_block *super;
1588
1589         super = root->fs_info->super_copy;
1590
1591         root_item = &root->fs_info->chunk_root->root_item;
1592         super->chunk_root = root_item->bytenr;
1593         super->chunk_root_generation = root_item->generation;
1594         super->chunk_root_level = root_item->level;
1595
1596         root_item = &root->fs_info->tree_root->root_item;
1597         super->root = root_item->bytenr;
1598         super->generation = root_item->generation;
1599         super->root_level = root_item->level;
1600         if (btrfs_test_opt(root, SPACE_CACHE))
1601                 super->cache_generation = root_item->generation;
1602         if (root->fs_info->update_uuid_tree_gen)
1603                 super->uuid_tree_generation = root_item->generation;
1604 }
1605
1606 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1607 {
1608         struct btrfs_transaction *trans;
1609         int ret = 0;
1610
1611         spin_lock(&info->trans_lock);
1612         trans = info->running_transaction;
1613         if (trans)
1614                 ret = (trans->state >= TRANS_STATE_COMMIT_START);
1615         spin_unlock(&info->trans_lock);
1616         return ret;
1617 }
1618
1619 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1620 {
1621         struct btrfs_transaction *trans;
1622         int ret = 0;
1623
1624         spin_lock(&info->trans_lock);
1625         trans = info->running_transaction;
1626         if (trans)
1627                 ret = is_transaction_blocked(trans);
1628         spin_unlock(&info->trans_lock);
1629         return ret;
1630 }
1631
1632 /*
1633  * wait for the current transaction commit to start and block subsequent
1634  * transaction joins
1635  */
1636 static void wait_current_trans_commit_start(struct btrfs_root *root,
1637                                             struct btrfs_transaction *trans)
1638 {
1639         wait_event(root->fs_info->transaction_blocked_wait,
1640                    trans->state >= TRANS_STATE_COMMIT_START ||
1641                    trans->aborted);
1642 }
1643
1644 /*
1645  * wait for the current transaction to start and then become unblocked.
1646  * caller holds ref.
1647  */
1648 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1649                                          struct btrfs_transaction *trans)
1650 {
1651         wait_event(root->fs_info->transaction_wait,
1652                    trans->state >= TRANS_STATE_UNBLOCKED ||
1653                    trans->aborted);
1654 }
1655
1656 /*
1657  * commit transactions asynchronously. once btrfs_commit_transaction_async
1658  * returns, any subsequent transaction will not be allowed to join.
1659  */
1660 struct btrfs_async_commit {
1661         struct btrfs_trans_handle *newtrans;
1662         struct btrfs_root *root;
1663         struct work_struct work;
1664 };
1665
1666 static void do_async_commit(struct work_struct *work)
1667 {
1668         struct btrfs_async_commit *ac =
1669                 container_of(work, struct btrfs_async_commit, work);
1670
1671         /*
1672          * We've got freeze protection passed with the transaction.
1673          * Tell lockdep about it.
1674          */
1675         if (ac->newtrans->type & __TRANS_FREEZABLE)
1676                 __sb_writers_acquired(ac->root->fs_info->sb, SB_FREEZE_FS);
1677
1678         current->journal_info = ac->newtrans;
1679
1680         btrfs_commit_transaction(ac->newtrans, ac->root);
1681         kfree(ac);
1682 }
1683
1684 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1685                                    struct btrfs_root *root,
1686                                    int wait_for_unblock)
1687 {
1688         struct btrfs_async_commit *ac;
1689         struct btrfs_transaction *cur_trans;
1690
1691         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1692         if (!ac)
1693                 return -ENOMEM;
1694
1695         INIT_WORK(&ac->work, do_async_commit);
1696         ac->root = root;
1697         ac->newtrans = btrfs_join_transaction(root);
1698         if (IS_ERR(ac->newtrans)) {
1699                 int err = PTR_ERR(ac->newtrans);
1700                 kfree(ac);
1701                 return err;
1702         }
1703
1704         /* take transaction reference */
1705         cur_trans = trans->transaction;
1706         atomic_inc(&cur_trans->use_count);
1707
1708         btrfs_end_transaction(trans, root);
1709
1710         /*
1711          * Tell lockdep we've released the freeze rwsem, since the
1712          * async commit thread will be the one to unlock it.
1713          */
1714         if (ac->newtrans->type & __TRANS_FREEZABLE)
1715                 __sb_writers_release(root->fs_info->sb, SB_FREEZE_FS);
1716
1717         schedule_work(&ac->work);
1718
1719         /* wait for transaction to start and unblock */
1720         if (wait_for_unblock)
1721                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1722         else
1723                 wait_current_trans_commit_start(root, cur_trans);
1724
1725         if (current->journal_info == trans)
1726                 current->journal_info = NULL;
1727
1728         btrfs_put_transaction(cur_trans);
1729         return 0;
1730 }
1731
1732
1733 static void cleanup_transaction(struct btrfs_trans_handle *trans,
1734                                 struct btrfs_root *root, int err)
1735 {
1736         struct btrfs_transaction *cur_trans = trans->transaction;
1737         DEFINE_WAIT(wait);
1738
1739         WARN_ON(trans->use_count > 1);
1740
1741         btrfs_abort_transaction(trans, root, err);
1742
1743         spin_lock(&root->fs_info->trans_lock);
1744
1745         /*
1746          * If the transaction is removed from the list, it means this
1747          * transaction has been committed successfully, so it is impossible
1748          * to call the cleanup function.
1749          */
1750         BUG_ON(list_empty(&cur_trans->list));
1751
1752         list_del_init(&cur_trans->list);
1753         if (cur_trans == root->fs_info->running_transaction) {
1754                 cur_trans->state = TRANS_STATE_COMMIT_DOING;
1755                 spin_unlock(&root->fs_info->trans_lock);
1756                 wait_event(cur_trans->writer_wait,
1757                            atomic_read(&cur_trans->num_writers) == 1);
1758
1759                 spin_lock(&root->fs_info->trans_lock);
1760         }
1761         spin_unlock(&root->fs_info->trans_lock);
1762
1763         btrfs_cleanup_one_transaction(trans->transaction, root);
1764
1765         spin_lock(&root->fs_info->trans_lock);
1766         if (cur_trans == root->fs_info->running_transaction)
1767                 root->fs_info->running_transaction = NULL;
1768         spin_unlock(&root->fs_info->trans_lock);
1769
1770         if (trans->type & __TRANS_FREEZABLE)
1771                 sb_end_intwrite(root->fs_info->sb);
1772         btrfs_put_transaction(cur_trans);
1773         btrfs_put_transaction(cur_trans);
1774
1775         trace_btrfs_transaction_commit(root);
1776
1777         if (current->journal_info == trans)
1778                 current->journal_info = NULL;
1779         btrfs_scrub_cancel(root->fs_info);
1780
1781         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1782 }
1783
1784 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
1785 {
1786         if (btrfs_test_opt(fs_info->tree_root, FLUSHONCOMMIT))
1787                 return btrfs_start_delalloc_roots(fs_info, 1, -1);
1788         return 0;
1789 }
1790
1791 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
1792 {
1793         if (btrfs_test_opt(fs_info->tree_root, FLUSHONCOMMIT))
1794                 btrfs_wait_ordered_roots(fs_info, -1);
1795 }
1796
1797 static inline void
1798 btrfs_wait_pending_ordered(struct btrfs_transaction *cur_trans,
1799                            struct btrfs_fs_info *fs_info)
1800 {
1801         struct btrfs_ordered_extent *ordered;
1802
1803         spin_lock(&fs_info->trans_lock);
1804         while (!list_empty(&cur_trans->pending_ordered)) {
1805                 ordered = list_first_entry(&cur_trans->pending_ordered,
1806                                            struct btrfs_ordered_extent,
1807                                            trans_list);
1808                 list_del_init(&ordered->trans_list);
1809                 spin_unlock(&fs_info->trans_lock);
1810
1811                 wait_event(ordered->wait, test_bit(BTRFS_ORDERED_COMPLETE,
1812                                                    &ordered->flags));
1813                 btrfs_put_ordered_extent(ordered);
1814                 spin_lock(&fs_info->trans_lock);
1815         }
1816         spin_unlock(&fs_info->trans_lock);
1817 }
1818
1819 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1820                              struct btrfs_root *root)
1821 {
1822         struct btrfs_transaction *cur_trans = trans->transaction;
1823         struct btrfs_transaction *prev_trans = NULL;
1824         struct btrfs_inode *btree_ino = BTRFS_I(root->fs_info->btree_inode);
1825         int ret;
1826
1827         /* Stop the commit early if ->aborted is set */
1828         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1829                 ret = cur_trans->aborted;
1830                 btrfs_end_transaction(trans, root);
1831                 return ret;
1832         }
1833
1834         /* make a pass through all the delayed refs we have so far
1835          * any runnings procs may add more while we are here
1836          */
1837         ret = btrfs_run_delayed_refs(trans, root, 0);
1838         if (ret) {
1839                 btrfs_end_transaction(trans, root);
1840                 return ret;
1841         }
1842
1843         btrfs_trans_release_metadata(trans, root);
1844         trans->block_rsv = NULL;
1845         if (trans->qgroup_reserved) {
1846                 btrfs_qgroup_free(root, trans->qgroup_reserved);
1847                 trans->qgroup_reserved = 0;
1848         }
1849
1850         cur_trans = trans->transaction;
1851
1852         /*
1853          * set the flushing flag so procs in this transaction have to
1854          * start sending their work down.
1855          */
1856         cur_trans->delayed_refs.flushing = 1;
1857         smp_wmb();
1858
1859         if (!list_empty(&trans->new_bgs))
1860                 btrfs_create_pending_block_groups(trans, root);
1861
1862         ret = btrfs_run_delayed_refs(trans, root, 0);
1863         if (ret) {
1864                 btrfs_end_transaction(trans, root);
1865                 return ret;
1866         }
1867
1868         if (!cur_trans->dirty_bg_run) {
1869                 int run_it = 0;
1870
1871                 /* this mutex is also taken before trying to set
1872                  * block groups readonly.  We need to make sure
1873                  * that nobody has set a block group readonly
1874                  * after a extents from that block group have been
1875                  * allocated for cache files.  btrfs_set_block_group_ro
1876                  * will wait for the transaction to commit if it
1877                  * finds dirty_bg_run = 1
1878                  *
1879                  * The dirty_bg_run flag is also used to make sure only
1880                  * one process starts all the block group IO.  It wouldn't
1881                  * hurt to have more than one go through, but there's no
1882                  * real advantage to it either.
1883                  */
1884                 mutex_lock(&root->fs_info->ro_block_group_mutex);
1885                 if (!cur_trans->dirty_bg_run) {
1886                         run_it = 1;
1887                         cur_trans->dirty_bg_run = 1;
1888                 }
1889                 mutex_unlock(&root->fs_info->ro_block_group_mutex);
1890
1891                 if (run_it)
1892                         ret = btrfs_start_dirty_block_groups(trans, root);
1893         }
1894         if (ret) {
1895                 btrfs_end_transaction(trans, root);
1896                 return ret;
1897         }
1898
1899         spin_lock(&root->fs_info->trans_lock);
1900         list_splice_init(&trans->ordered, &cur_trans->pending_ordered);
1901         if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
1902                 spin_unlock(&root->fs_info->trans_lock);
1903                 atomic_inc(&cur_trans->use_count);
1904                 ret = btrfs_end_transaction(trans, root);
1905
1906                 wait_for_commit(root, cur_trans);
1907
1908                 if (unlikely(cur_trans->aborted))
1909                         ret = cur_trans->aborted;
1910
1911                 btrfs_put_transaction(cur_trans);
1912
1913                 return ret;
1914         }
1915
1916         cur_trans->state = TRANS_STATE_COMMIT_START;
1917         wake_up(&root->fs_info->transaction_blocked_wait);
1918
1919         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1920                 prev_trans = list_entry(cur_trans->list.prev,
1921                                         struct btrfs_transaction, list);
1922                 if (prev_trans->state != TRANS_STATE_COMPLETED) {
1923                         atomic_inc(&prev_trans->use_count);
1924                         spin_unlock(&root->fs_info->trans_lock);
1925
1926                         wait_for_commit(root, prev_trans);
1927                         ret = prev_trans->aborted;
1928
1929                         btrfs_put_transaction(prev_trans);
1930                         if (ret)
1931                                 goto cleanup_transaction;
1932                 } else {
1933                         spin_unlock(&root->fs_info->trans_lock);
1934                 }
1935         } else {
1936                 spin_unlock(&root->fs_info->trans_lock);
1937         }
1938
1939         extwriter_counter_dec(cur_trans, trans->type);
1940
1941         ret = btrfs_start_delalloc_flush(root->fs_info);
1942         if (ret)
1943                 goto cleanup_transaction;
1944
1945         ret = btrfs_run_delayed_items(trans, root);
1946         if (ret)
1947                 goto cleanup_transaction;
1948
1949         wait_event(cur_trans->writer_wait,
1950                    extwriter_counter_read(cur_trans) == 0);
1951
1952         /* some pending stuffs might be added after the previous flush. */
1953         ret = btrfs_run_delayed_items(trans, root);
1954         if (ret)
1955                 goto cleanup_transaction;
1956
1957         btrfs_wait_delalloc_flush(root->fs_info);
1958
1959         btrfs_wait_pending_ordered(cur_trans, root->fs_info);
1960
1961         btrfs_scrub_pause(root);
1962         /*
1963          * Ok now we need to make sure to block out any other joins while we
1964          * commit the transaction.  We could have started a join before setting
1965          * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
1966          */
1967         spin_lock(&root->fs_info->trans_lock);
1968         cur_trans->state = TRANS_STATE_COMMIT_DOING;
1969         spin_unlock(&root->fs_info->trans_lock);
1970         wait_event(cur_trans->writer_wait,
1971                    atomic_read(&cur_trans->num_writers) == 1);
1972
1973         /* ->aborted might be set after the previous check, so check it */
1974         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1975                 ret = cur_trans->aborted;
1976                 goto scrub_continue;
1977         }
1978         /*
1979          * the reloc mutex makes sure that we stop
1980          * the balancing code from coming in and moving
1981          * extents around in the middle of the commit
1982          */
1983         mutex_lock(&root->fs_info->reloc_mutex);
1984
1985         /*
1986          * We needn't worry about the delayed items because we will
1987          * deal with them in create_pending_snapshot(), which is the
1988          * core function of the snapshot creation.
1989          */
1990         ret = create_pending_snapshots(trans, root->fs_info);
1991         if (ret) {
1992                 mutex_unlock(&root->fs_info->reloc_mutex);
1993                 goto scrub_continue;
1994         }
1995
1996         /*
1997          * We insert the dir indexes of the snapshots and update the inode
1998          * of the snapshots' parents after the snapshot creation, so there
1999          * are some delayed items which are not dealt with. Now deal with
2000          * them.
2001          *
2002          * We needn't worry that this operation will corrupt the snapshots,
2003          * because all the tree which are snapshoted will be forced to COW
2004          * the nodes and leaves.
2005          */
2006         ret = btrfs_run_delayed_items(trans, root);
2007         if (ret) {
2008                 mutex_unlock(&root->fs_info->reloc_mutex);
2009                 goto scrub_continue;
2010         }
2011
2012         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
2013         if (ret) {
2014                 mutex_unlock(&root->fs_info->reloc_mutex);
2015                 goto scrub_continue;
2016         }
2017
2018         /* Reocrd old roots for later qgroup accounting */
2019         ret = btrfs_qgroup_prepare_account_extents(trans, root->fs_info);
2020         if (ret) {
2021                 mutex_unlock(&root->fs_info->reloc_mutex);
2022                 goto scrub_continue;
2023         }
2024
2025         /*
2026          * make sure none of the code above managed to slip in a
2027          * delayed item
2028          */
2029         btrfs_assert_delayed_root_empty(root);
2030
2031         WARN_ON(cur_trans != trans->transaction);
2032
2033         /* btrfs_commit_tree_roots is responsible for getting the
2034          * various roots consistent with each other.  Every pointer
2035          * in the tree of tree roots has to point to the most up to date
2036          * root for every subvolume and other tree.  So, we have to keep
2037          * the tree logging code from jumping in and changing any
2038          * of the trees.
2039          *
2040          * At this point in the commit, there can't be any tree-log
2041          * writers, but a little lower down we drop the trans mutex
2042          * and let new people in.  By holding the tree_log_mutex
2043          * from now until after the super is written, we avoid races
2044          * with the tree-log code.
2045          */
2046         mutex_lock(&root->fs_info->tree_log_mutex);
2047
2048         ret = commit_fs_roots(trans, root);
2049         if (ret) {
2050                 mutex_unlock(&root->fs_info->tree_log_mutex);
2051                 mutex_unlock(&root->fs_info->reloc_mutex);
2052                 goto scrub_continue;
2053         }
2054
2055         /*
2056          * Since the transaction is done, we can apply the pending changes
2057          * before the next transaction.
2058          */
2059         btrfs_apply_pending_changes(root->fs_info);
2060
2061         /* commit_fs_roots gets rid of all the tree log roots, it is now
2062          * safe to free the root of tree log roots
2063          */
2064         btrfs_free_log_root_tree(trans, root->fs_info);
2065
2066         /*
2067          * Since fs roots are all committed, we can get a quite accurate
2068          * new_roots. So let's do quota accounting.
2069          */
2070         ret = btrfs_qgroup_account_extents(trans, root->fs_info);
2071         if (ret < 0) {
2072                 mutex_unlock(&root->fs_info->tree_log_mutex);
2073                 mutex_unlock(&root->fs_info->reloc_mutex);
2074                 goto scrub_continue;
2075         }
2076
2077         ret = commit_cowonly_roots(trans, root);
2078         if (ret) {
2079                 mutex_unlock(&root->fs_info->tree_log_mutex);
2080                 mutex_unlock(&root->fs_info->reloc_mutex);
2081                 goto scrub_continue;
2082         }
2083
2084         /*
2085          * The tasks which save the space cache and inode cache may also
2086          * update ->aborted, check it.
2087          */
2088         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
2089                 ret = cur_trans->aborted;
2090                 mutex_unlock(&root->fs_info->tree_log_mutex);
2091                 mutex_unlock(&root->fs_info->reloc_mutex);
2092                 goto scrub_continue;
2093         }
2094
2095         btrfs_prepare_extent_commit(trans, root);
2096
2097         cur_trans = root->fs_info->running_transaction;
2098
2099         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
2100                             root->fs_info->tree_root->node);
2101         list_add_tail(&root->fs_info->tree_root->dirty_list,
2102                       &cur_trans->switch_commits);
2103
2104         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
2105                             root->fs_info->chunk_root->node);
2106         list_add_tail(&root->fs_info->chunk_root->dirty_list,
2107                       &cur_trans->switch_commits);
2108
2109         switch_commit_roots(cur_trans, root->fs_info);
2110
2111         assert_qgroups_uptodate(trans);
2112         ASSERT(list_empty(&cur_trans->dirty_bgs));
2113         ASSERT(list_empty(&cur_trans->io_bgs));
2114         update_super_roots(root);
2115
2116         btrfs_set_super_log_root(root->fs_info->super_copy, 0);
2117         btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
2118         memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
2119                sizeof(*root->fs_info->super_copy));
2120
2121         btrfs_update_commit_device_size(root->fs_info);
2122         btrfs_update_commit_device_bytes_used(root, cur_trans);
2123
2124         clear_bit(BTRFS_INODE_BTREE_LOG1_ERR, &btree_ino->runtime_flags);
2125         clear_bit(BTRFS_INODE_BTREE_LOG2_ERR, &btree_ino->runtime_flags);
2126
2127         btrfs_trans_release_chunk_metadata(trans);
2128
2129         spin_lock(&root->fs_info->trans_lock);
2130         cur_trans->state = TRANS_STATE_UNBLOCKED;
2131         root->fs_info->running_transaction = NULL;
2132         spin_unlock(&root->fs_info->trans_lock);
2133         mutex_unlock(&root->fs_info->reloc_mutex);
2134
2135         wake_up(&root->fs_info->transaction_wait);
2136
2137         ret = btrfs_write_and_wait_transaction(trans, root);
2138         if (ret) {
2139                 btrfs_error(root->fs_info, ret,
2140                             "Error while writing out transaction");
2141                 mutex_unlock(&root->fs_info->tree_log_mutex);
2142                 goto scrub_continue;
2143         }
2144
2145         ret = write_ctree_super(trans, root, 0);
2146         if (ret) {
2147                 mutex_unlock(&root->fs_info->tree_log_mutex);
2148                 goto scrub_continue;
2149         }
2150
2151         /*
2152          * the super is written, we can safely allow the tree-loggers
2153          * to go about their business
2154          */
2155         mutex_unlock(&root->fs_info->tree_log_mutex);
2156
2157         btrfs_finish_extent_commit(trans, root);
2158
2159         if (cur_trans->have_free_bgs)
2160                 btrfs_clear_space_info_full(root->fs_info);
2161
2162         root->fs_info->last_trans_committed = cur_trans->transid;
2163         /*
2164          * We needn't acquire the lock here because there is no other task
2165          * which can change it.
2166          */
2167         cur_trans->state = TRANS_STATE_COMPLETED;
2168         wake_up(&cur_trans->commit_wait);
2169
2170         spin_lock(&root->fs_info->trans_lock);
2171         list_del_init(&cur_trans->list);
2172         spin_unlock(&root->fs_info->trans_lock);
2173
2174         btrfs_put_transaction(cur_trans);
2175         btrfs_put_transaction(cur_trans);
2176
2177         if (trans->type & __TRANS_FREEZABLE)
2178                 sb_end_intwrite(root->fs_info->sb);
2179
2180         trace_btrfs_transaction_commit(root);
2181
2182         btrfs_scrub_continue(root);
2183
2184         if (current->journal_info == trans)
2185                 current->journal_info = NULL;
2186
2187         kmem_cache_free(btrfs_trans_handle_cachep, trans);
2188
2189         if (current != root->fs_info->transaction_kthread &&
2190             current != root->fs_info->cleaner_kthread)
2191                 btrfs_run_delayed_iputs(root);
2192
2193         return ret;
2194
2195 scrub_continue:
2196         btrfs_scrub_continue(root);
2197 cleanup_transaction:
2198         btrfs_trans_release_metadata(trans, root);
2199         btrfs_trans_release_chunk_metadata(trans);
2200         trans->block_rsv = NULL;
2201         if (trans->qgroup_reserved) {
2202                 btrfs_qgroup_free(root, trans->qgroup_reserved);
2203                 trans->qgroup_reserved = 0;
2204         }
2205         btrfs_warn(root->fs_info, "Skipping commit of aborted transaction.");
2206         if (current->journal_info == trans)
2207                 current->journal_info = NULL;
2208         cleanup_transaction(trans, root, ret);
2209
2210         return ret;
2211 }
2212
2213 /*
2214  * return < 0 if error
2215  * 0 if there are no more dead_roots at the time of call
2216  * 1 there are more to be processed, call me again
2217  *
2218  * The return value indicates there are certainly more snapshots to delete, but
2219  * if there comes a new one during processing, it may return 0. We don't mind,
2220  * because btrfs_commit_super will poke cleaner thread and it will process it a
2221  * few seconds later.
2222  */
2223 int btrfs_clean_one_deleted_snapshot(struct btrfs_root *root)
2224 {
2225         int ret;
2226         struct btrfs_fs_info *fs_info = root->fs_info;
2227
2228         spin_lock(&fs_info->trans_lock);
2229         if (list_empty(&fs_info->dead_roots)) {
2230                 spin_unlock(&fs_info->trans_lock);
2231                 return 0;
2232         }
2233         root = list_first_entry(&fs_info->dead_roots,
2234                         struct btrfs_root, root_list);
2235         list_del_init(&root->root_list);
2236         spin_unlock(&fs_info->trans_lock);
2237
2238         pr_debug("BTRFS: cleaner removing %llu\n", root->objectid);
2239
2240         btrfs_kill_all_delayed_nodes(root);
2241
2242         if (btrfs_header_backref_rev(root->node) <
2243                         BTRFS_MIXED_BACKREF_REV)
2244                 ret = btrfs_drop_snapshot(root, NULL, 0, 0);
2245         else
2246                 ret = btrfs_drop_snapshot(root, NULL, 1, 0);
2247
2248         return (ret < 0) ? 0 : 1;
2249 }
2250
2251 void btrfs_apply_pending_changes(struct btrfs_fs_info *fs_info)
2252 {
2253         unsigned long prev;
2254         unsigned long bit;
2255
2256         prev = xchg(&fs_info->pending_changes, 0);
2257         if (!prev)
2258                 return;
2259
2260         bit = 1 << BTRFS_PENDING_SET_INODE_MAP_CACHE;
2261         if (prev & bit)
2262                 btrfs_set_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2263         prev &= ~bit;
2264
2265         bit = 1 << BTRFS_PENDING_CLEAR_INODE_MAP_CACHE;
2266         if (prev & bit)
2267                 btrfs_clear_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2268         prev &= ~bit;
2269
2270         bit = 1 << BTRFS_PENDING_COMMIT;
2271         if (prev & bit)
2272                 btrfs_debug(fs_info, "pending commit done");
2273         prev &= ~bit;
2274
2275         if (prev)
2276                 btrfs_warn(fs_info,
2277                         "unknown pending changes left 0x%lx, ignoring", prev);
2278 }