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