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