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
  35 #define BTRFS_ROOT_TRANS_TAG 0
  36
  37 void put_transaction(struct btrfs_transaction *transaction)
  38 {
  39         WARN_ON(atomic_read(&transaction->use_count) == 0);
  40         if (atomic_dec_and_test(&transaction->use_count)) {
  41                 BUG_ON(!list_empty(&transaction->list));
  42                 WARN_ON(transaction->delayed_refs.root.rb_node);
  43                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  44         }
  45 }
  46
  47 static noinline void switch_commit_root(struct btrfs_root *root)
  48 {
  49         free_extent_buffer(root->commit_root);
  50         root->commit_root = btrfs_root_node(root);
  51 }
  52
  53 static inline int can_join_transaction(struct btrfs_transaction *trans,
  54                                        int type)
  55 {
  56         return !(trans->in_commit &&
  57                  type != TRANS_JOIN &&
  58                  type != TRANS_JOIN_NOLOCK);
  59 }
  60
  61 /*
  62  * either allocate a new transaction or hop into the existing one
  63  */
  64 static noinline int join_transaction(struct btrfs_root *root, int type)
  65 {
  66         struct btrfs_transaction *cur_trans;
  67         struct btrfs_fs_info *fs_info = root->fs_info;
  68
  69         spin_lock(&fs_info->trans_lock);
  70 loop:
  71         /* The file system has been taken offline. No new transactions. */
  72         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
  73                 spin_unlock(&fs_info->trans_lock);
  74                 return -EROFS;
  75         }
  76
  77         if (fs_info->trans_no_join) {
  78                 /*
  79                  * If we are JOIN_NOLOCK we're already committing a current
  80                  * transaction, we just need a handle to deal with something
  81                  * when committing the transaction, such as inode cache and
  82                  * space cache. It is a special case.
  83                  */
  84                 if (type != TRANS_JOIN_NOLOCK) {
  85                         spin_unlock(&fs_info->trans_lock);
  86                         return -EBUSY;
  87                 }
  88         }
  89
  90         cur_trans = fs_info->running_transaction;
  91         if (cur_trans) {
  92                 if (cur_trans->aborted) {
  93                         spin_unlock(&fs_info->trans_lock);
  94                         return cur_trans->aborted;
  95                 }
  96                 if (!can_join_transaction(cur_trans, type)) {
  97                         spin_unlock(&fs_info->trans_lock);
  98                         return -EBUSY;
  99                 }
 100                 atomic_inc(&cur_trans->use_count);
 101                 atomic_inc(&cur_trans->num_writers);
 102                 cur_trans->num_joined++;
 103                 spin_unlock(&fs_info->trans_lock);
 104                 return 0;
 105         }
 106         spin_unlock(&fs_info->trans_lock);
 107
 108         /*
 109          * If we are ATTACH, we just want to catch the current transaction,
 110          * and commit it. If there is no transaction, just return ENOENT.
 111          */
 112         if (type == TRANS_ATTACH)
 113                 return -ENOENT;
 114
 115         cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
 116         if (!cur_trans)
 117                 return -ENOMEM;
 118
 119         spin_lock(&fs_info->trans_lock);
 120         if (fs_info->running_transaction) {
 121                 /*
 122                  * someone started a transaction after we unlocked.  Make sure
 123                  * to redo the trans_no_join checks above
 124                  */
 125                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 126                 goto loop;
 127         } else if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
 128                 spin_unlock(&fs_info->trans_lock);
 129                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 130                 return -EROFS;
 131         }
 132
 133         atomic_set(&cur_trans->num_writers, 1);
 134         cur_trans->num_joined = 0;
 135         init_waitqueue_head(&cur_trans->writer_wait);
 136         init_waitqueue_head(&cur_trans->commit_wait);
 137         cur_trans->in_commit = 0;
 138         cur_trans->blocked = 0;
 139         /*
 140          * One for this trans handle, one so it will live on until we
 141          * commit the transaction.
 142          */
 143         atomic_set(&cur_trans->use_count, 2);
 144         cur_trans->commit_done = 0;
 145         cur_trans->start_time = get_seconds();
 146
 147         cur_trans->delayed_refs.root = RB_ROOT;
 148         cur_trans->delayed_refs.num_entries = 0;
 149         cur_trans->delayed_refs.num_heads_ready = 0;
 150         cur_trans->delayed_refs.num_heads = 0;
 151         cur_trans->delayed_refs.flushing = 0;
 152         cur_trans->delayed_refs.run_delayed_start = 0;
 153
 154         /*
 155          * although the tree mod log is per file system and not per transaction,
 156          * the log must never go across transaction boundaries.
 157          */
 158         smp_mb();
 159         if (!list_empty(&fs_info->tree_mod_seq_list))
 160                 WARN(1, KERN_ERR "btrfs: tree_mod_seq_list not empty when "
 161                         "creating a fresh transaction\n");
 162         if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
 163                 WARN(1, KERN_ERR "btrfs: tree_mod_log rb tree not empty when "
 164                         "creating a fresh transaction\n");
 165         atomic_set(&fs_info->tree_mod_seq, 0);
 166
 167         spin_lock_init(&cur_trans->commit_lock);
 168         spin_lock_init(&cur_trans->delayed_refs.lock);
 169         atomic_set(&cur_trans->delayed_refs.procs_running_refs, 0);
 170         atomic_set(&cur_trans->delayed_refs.ref_seq, 0);
 171         init_waitqueue_head(&cur_trans->delayed_refs.wait);
 172
 173         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 174         INIT_LIST_HEAD(&cur_trans->ordered_operations);
 175         list_add_tail(&cur_trans->list, &fs_info->trans_list);
 176         extent_io_tree_init(&cur_trans->dirty_pages,
 177                              fs_info->btree_inode->i_mapping);
 178         fs_info->generation++;
 179         cur_trans->transid = fs_info->generation;
 180         fs_info->running_transaction = cur_trans;
 181         cur_trans->aborted = 0;
 182         spin_unlock(&fs_info->trans_lock);
 183
 184         return 0;
 185 }
 186
 187 /*
 188  * this does all the record keeping required to make sure that a reference
 189  * counted root is properly recorded in a given transaction.  This is required
 190  * to make sure the old root from before we joined the transaction is deleted
 191  * when the transaction commits
 192  */
 193 static int record_root_in_trans(struct btrfs_trans_handle *trans,
 194                                struct btrfs_root *root)
 195 {
 196         if (root->ref_cows && root->last_trans < trans->transid) {
 197                 WARN_ON(root == root->fs_info->extent_root);
 198                 WARN_ON(root->commit_root != root->node);
 199
 200                 /*
 201                  * see below for in_trans_setup usage rules
 202                  * we have the reloc mutex held now, so there
 203                  * is only one writer in this function
 204                  */
 205                 root->in_trans_setup = 1;
 206
 207                 /* make sure readers find in_trans_setup before
 208                  * they find our root->last_trans update
 209                  */
 210                 smp_wmb();
 211
 212                 spin_lock(&root->fs_info->fs_roots_radix_lock);
 213                 if (root->last_trans == trans->transid) {
 214                         spin_unlock(&root->fs_info->fs_roots_radix_lock);
 215                         return 0;
 216                 }
 217                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 218                            (unsigned long)root->root_key.objectid,
 219                            BTRFS_ROOT_TRANS_TAG);
 220                 spin_unlock(&root->fs_info->fs_roots_radix_lock);
 221                 root->last_trans = trans->transid;
 222
 223                 /* this is pretty tricky.  We don't want to
 224                  * take the relocation lock in btrfs_record_root_in_trans
 225                  * unless we're really doing the first setup for this root in
 226                  * this transaction.
 227                  *
 228                  * Normally we'd use root->last_trans as a flag to decide
 229                  * if we want to take the expensive mutex.
 230                  *
 231                  * But, we have to set root->last_trans before we
 232                  * init the relocation root, otherwise, we trip over warnings
 233                  * in ctree.c.  The solution used here is to flag ourselves
 234                  * with root->in_trans_setup.  When this is 1, we're still
 235                  * fixing up the reloc trees and everyone must wait.
 236                  *
 237                  * When this is zero, they can trust root->last_trans and fly
 238                  * through btrfs_record_root_in_trans without having to take the
 239                  * lock.  smp_wmb() makes sure that all the writes above are
 240                  * done before we pop in the zero below
 241                  */
 242                 btrfs_init_reloc_root(trans, root);
 243                 smp_wmb();
 244                 root->in_trans_setup = 0;
 245         }
 246         return 0;
 247 }
 248
 249
 250 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 251                                struct btrfs_root *root)
 252 {
 253         if (!root->ref_cows)
 254                 return 0;
 255
 256         /*
 257          * see record_root_in_trans for comments about in_trans_setup usage
 258          * and barriers
 259          */
 260         smp_rmb();
 261         if (root->last_trans == trans->transid &&
 262             !root->in_trans_setup)
 263                 return 0;
 264
 265         mutex_lock(&root->fs_info->reloc_mutex);
 266         record_root_in_trans(trans, root);
 267         mutex_unlock(&root->fs_info->reloc_mutex);
 268
 269         return 0;
 270 }
 271
 272 /* wait for commit against the current transaction to become unblocked
 273  * when this is done, it is safe to start a new transaction, but the current
 274  * transaction might not be fully on disk.
 275  */
 276 static void wait_current_trans(struct btrfs_root *root)
 277 {
 278         struct btrfs_transaction *cur_trans;
 279
 280         spin_lock(&root->fs_info->trans_lock);
 281         cur_trans = root->fs_info->running_transaction;
 282         if (cur_trans && cur_trans->blocked) {
 283                 atomic_inc(&cur_trans->use_count);
 284                 spin_unlock(&root->fs_info->trans_lock);
 285
 286                 wait_event(root->fs_info->transaction_wait,
 287                            !cur_trans->blocked);
 288                 put_transaction(cur_trans);
 289         } else {
 290                 spin_unlock(&root->fs_info->trans_lock);
 291         }
 292 }
 293
 294 static int may_wait_transaction(struct btrfs_root *root, int type)
 295 {
 296         if (root->fs_info->log_root_recovering)
 297                 return 0;
 298
 299         if (type == TRANS_USERSPACE)
 300                 return 1;
 301
 302         if (type == TRANS_START &&
 303             !atomic_read(&root->fs_info->open_ioctl_trans))
 304                 return 1;
 305
 306         return 0;
 307 }
 308
 309 static struct btrfs_trans_handle *
 310 start_transaction(struct btrfs_root *root, u64 num_items, int type,
 311                   enum btrfs_reserve_flush_enum flush)
 312 {
 313         struct btrfs_trans_handle *h;
 314         struct btrfs_transaction *cur_trans;
 315         u64 num_bytes = 0;
 316         int ret;
 317         u64 qgroup_reserved = 0;
 318
 319         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
 320                 return ERR_PTR(-EROFS);
 321
 322         if (current->journal_info) {
 323                 WARN_ON(type != TRANS_JOIN && type != TRANS_JOIN_NOLOCK);
 324                 h = current->journal_info;
 325                 h->use_count++;
 326                 WARN_ON(h->use_count > 2);
 327                 h->orig_rsv = h->block_rsv;
 328                 h->block_rsv = NULL;
 329                 goto got_it;
 330         }
 331
 332         /*
 333          * Do the reservation before we join the transaction so we can do all
 334          * the appropriate flushing if need be.
 335          */
 336         if (num_items > 0 && root != root->fs_info->chunk_root) {
 337                 if (root->fs_info->quota_enabled &&
 338                     is_fstree(root->root_key.objectid)) {
 339                         qgroup_reserved = num_items * root->leafsize;
 340                         ret = btrfs_qgroup_reserve(root, qgroup_reserved);
 341                         if (ret)
 342                                 return ERR_PTR(ret);
 343                 }
 344
 345                 num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
 346                 ret = btrfs_block_rsv_add(root,
 347                                           &root->fs_info->trans_block_rsv,
 348                                           num_bytes, flush);
 349                 if (ret)
 350                         goto reserve_fail;
 351         }
 352 again:
 353         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 354         if (!h) {
 355                 ret = -ENOMEM;
 356                 goto alloc_fail;
 357         }
 358
 359         /*
 360          * If we are JOIN_NOLOCK we're already committing a transaction and
 361          * waiting on this guy, so we don't need to do the sb_start_intwrite
 362          * because we're already holding a ref.  We need this because we could
 363          * have raced in and did an fsync() on a file which can kick a commit
 364          * and then we deadlock with somebody doing a freeze.
 365          *
 366          * If we are ATTACH, it means we just want to catch the current
 367          * transaction and commit it, so we needn't do sb_start_intwrite().
 368          */
 369         if (type < TRANS_JOIN_NOLOCK)
 370                 sb_start_intwrite(root->fs_info->sb);
 371
 372         if (may_wait_transaction(root, type))
 373                 wait_current_trans(root);
 374
 375         do {
 376                 ret = join_transaction(root, type);
 377                 if (ret == -EBUSY) {
 378                         wait_current_trans(root);
 379                         if (unlikely(type == TRANS_ATTACH))
 380                                 ret = -ENOENT;
 381                 }
 382         } while (ret == -EBUSY);
 383
 384         if (ret < 0) {
 385                 /* We must get the transaction if we are JOIN_NOLOCK. */
 386                 BUG_ON(type == TRANS_JOIN_NOLOCK);
 387                 goto join_fail;
 388         }
 389
 390         cur_trans = root->fs_info->running_transaction;
 391
 392         h->transid = cur_trans->transid;
 393         h->transaction = cur_trans;
 394         h->blocks_used = 0;
 395         h->bytes_reserved = 0;
 396         h->root = root;
 397         h->delayed_ref_updates = 0;
 398         h->use_count = 1;
 399         h->adding_csums = 0;
 400         h->block_rsv = NULL;
 401         h->orig_rsv = NULL;
 402         h->aborted = 0;
 403         h->qgroup_reserved = 0;
 404         h->delayed_ref_elem.seq = 0;
 405         h->type = type;
 406         h->allocating_chunk = false;
 407         INIT_LIST_HEAD(&h->qgroup_ref_list);
 408         INIT_LIST_HEAD(&h->new_bgs);
 409
 410         smp_mb();
 411         if (cur_trans->blocked && may_wait_transaction(root, type)) {
 412                 btrfs_commit_transaction(h, root);
 413                 goto again;
 414         }
 415
 416         if (num_bytes) {
 417                 trace_btrfs_space_reservation(root->fs_info, "transaction",
 418                                               h->transid, num_bytes, 1);
 419                 h->block_rsv = &root->fs_info->trans_block_rsv;
 420                 h->bytes_reserved = num_bytes;
 421         }
 422         h->qgroup_reserved = qgroup_reserved;
 423
 424 got_it:
 425         btrfs_record_root_in_trans(h, root);
 426
 427         if (!current->journal_info && type != TRANS_USERSPACE)
 428                 current->journal_info = h;
 429         return h;
 430
 431 join_fail:
 432         if (type < TRANS_JOIN_NOLOCK)
 433                 sb_end_intwrite(root->fs_info->sb);
 434         kmem_cache_free(btrfs_trans_handle_cachep, h);
 435 alloc_fail:
 436         if (num_bytes)
 437                 btrfs_block_rsv_release(root, &root->fs_info->trans_block_rsv,
 438                                         num_bytes);
 439 reserve_fail:
 440         if (qgroup_reserved)
 441                 btrfs_qgroup_free(root, qgroup_reserved);
 442         return ERR_PTR(ret);
 443 }
 444
 445 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 446                                                    int num_items)
 447 {
 448         return start_transaction(root, num_items, TRANS_START,
 449                                  BTRFS_RESERVE_FLUSH_ALL);
 450 }
 451
 452 struct btrfs_trans_handle *btrfs_start_transaction_lflush(
 453                                         struct btrfs_root *root, int num_items)
 454 {
 455         return start_transaction(root, num_items, TRANS_START,
 456                                  BTRFS_RESERVE_FLUSH_LIMIT);
 457 }
 458
 459 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 460 {
 461         return start_transaction(root, 0, TRANS_JOIN, 0);
 462 }
 463
 464 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 465 {
 466         return start_transaction(root, 0, TRANS_JOIN_NOLOCK, 0);
 467 }
 468
 469 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 470 {
 471         return start_transaction(root, 0, TRANS_USERSPACE, 0);
 472 }
 473
 474 /*
 475  * btrfs_attach_transaction() - catch the running transaction
 476  *
 477  * It is used when we want to commit the current the transaction, but
 478  * don't want to start a new one.
 479  *
 480  * Note: If this function return -ENOENT, it just means there is no
 481  * running transaction. But it is possible that the inactive transaction
 482  * is still in the memory, not fully on disk. If you hope there is no
 483  * inactive transaction in the fs when -ENOENT is returned, you should
 484  * invoke
 485  *     btrfs_attach_transaction_barrier()
 486  */
 487 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
 488 {
 489         return start_transaction(root, 0, TRANS_ATTACH, 0);
 490 }
 491
 492 /*
 493  * btrfs_attach_transaction() - catch the running transaction
 494  *
 495  * It is similar to the above function, the differentia is this one
 496  * will wait for all the inactive transactions until they fully
 497  * complete.
 498  */
 499 struct btrfs_trans_handle *
 500 btrfs_attach_transaction_barrier(struct btrfs_root *root)
 501 {
 502         struct btrfs_trans_handle *trans;
 503
 504         trans = start_transaction(root, 0, TRANS_ATTACH, 0);
 505         if (IS_ERR(trans) && PTR_ERR(trans) == -ENOENT)
 506                 btrfs_wait_for_commit(root, 0);
 507
 508         return trans;
 509 }
 510
 511 /* wait for a transaction commit to be fully complete */
 512 static noinline void wait_for_commit(struct btrfs_root *root,
 513                                     struct btrfs_transaction *commit)
 514 {
 515         wait_event(commit->commit_wait, commit->commit_done);
 516 }
 517
 518 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 519 {
 520         struct btrfs_transaction *cur_trans = NULL, *t;
 521         int ret = 0;
 522
 523         if (transid) {
 524                 if (transid <= root->fs_info->last_trans_committed)
 525                         goto out;
 526
 527                 ret = -EINVAL;
 528                 /* find specified transaction */
 529                 spin_lock(&root->fs_info->trans_lock);
 530                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 531                         if (t->transid == transid) {
 532                                 cur_trans = t;
 533                                 atomic_inc(&cur_trans->use_count);
 534                                 ret = 0;
 535                                 break;
 536                         }
 537                         if (t->transid > transid) {
 538                                 ret = 0;
 539                                 break;
 540                         }
 541                 }
 542                 spin_unlock(&root->fs_info->trans_lock);
 543                 /* The specified transaction doesn't exist */
 544                 if (!cur_trans)
 545                         goto out;
 546         } else {
 547                 /* find newest transaction that is committing | committed */
 548                 spin_lock(&root->fs_info->trans_lock);
 549                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 550                                             list) {
 551                         if (t->in_commit) {
 552                                 if (t->commit_done)
 553                                         break;
 554                                 cur_trans = t;
 555                                 atomic_inc(&cur_trans->use_count);
 556                                 break;
 557                         }
 558                 }
 559                 spin_unlock(&root->fs_info->trans_lock);
 560                 if (!cur_trans)
 561                         goto out;  /* nothing committing|committed */
 562         }
 563
 564         wait_for_commit(root, cur_trans);
 565         put_transaction(cur_trans);
 566 out:
 567         return ret;
 568 }
 569
 570 void btrfs_throttle(struct btrfs_root *root)
 571 {
 572         if (!atomic_read(&root->fs_info->open_ioctl_trans))
 573                 wait_current_trans(root);
 574 }
 575
 576 static int should_end_transaction(struct btrfs_trans_handle *trans,
 577                                   struct btrfs_root *root)
 578 {
 579         int ret;
 580
 581         ret = btrfs_block_rsv_check(root, &root->fs_info->global_block_rsv, 5);
 582         return ret ? 1 : 0;
 583 }
 584
 585 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 586                                  struct btrfs_root *root)
 587 {
 588         struct btrfs_transaction *cur_trans = trans->transaction;
 589         int updates;
 590         int err;
 591
 592         smp_mb();
 593         if (cur_trans->blocked || cur_trans->delayed_refs.flushing)
 594                 return 1;
 595
 596         updates = trans->delayed_ref_updates;
 597         trans->delayed_ref_updates = 0;
 598         if (updates) {
 599                 err = btrfs_run_delayed_refs(trans, root, updates);
 600                 if (err) /* Error code will also eval true */
 601                         return err;
 602         }
 603
 604         return should_end_transaction(trans, root);
 605 }
 606
 607 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 608                           struct btrfs_root *root, int throttle)
 609 {
 610         struct btrfs_transaction *cur_trans = trans->transaction;
 611         struct btrfs_fs_info *info = root->fs_info;
 612         int count = 0;
 613         int lock = (trans->type != TRANS_JOIN_NOLOCK);
 614         int err = 0;
 615
 616         if (--trans->use_count) {
 617                 trans->block_rsv = trans->orig_rsv;
 618                 return 0;
 619         }
 620
 621         /*
 622          * do the qgroup accounting as early as possible
 623          */
 624         err = btrfs_delayed_refs_qgroup_accounting(trans, info);
 625
 626         btrfs_trans_release_metadata(trans, root);
 627         trans->block_rsv = NULL;
 628         /*
 629          * the same root has to be passed to start_transaction and
 630          * end_transaction. Subvolume quota depends on this.
 631          */
 632         WARN_ON(trans->root != root);
 633
 634         if (trans->qgroup_reserved) {
 635                 btrfs_qgroup_free(root, trans->qgroup_reserved);
 636                 trans->qgroup_reserved = 0;
 637         }
 638
 639         if (!list_empty(&trans->new_bgs))
 640                 btrfs_create_pending_block_groups(trans, root);
 641
 642         while (count < 1) {
 643                 unsigned long cur = trans->delayed_ref_updates;
 644                 trans->delayed_ref_updates = 0;
 645                 if (cur &&
 646                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 647                         trans->delayed_ref_updates = 0;
 648                         btrfs_run_delayed_refs(trans, root, cur);
 649                 } else {
 650                         break;
 651                 }
 652                 count++;
 653         }
 654
 655         btrfs_trans_release_metadata(trans, root);
 656         trans->block_rsv = NULL;
 657
 658         if (!list_empty(&trans->new_bgs))
 659                 btrfs_create_pending_block_groups(trans, root);
 660
 661         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
 662             should_end_transaction(trans, root)) {
 663                 trans->transaction->blocked = 1;
 664                 smp_wmb();
 665         }
 666
 667         if (lock && cur_trans->blocked && !cur_trans->in_commit) {
 668                 if (throttle) {
 669                         /*
 670                          * We may race with somebody else here so end up having
 671                          * to call end_transaction on ourselves again, so inc
 672                          * our use_count.
 673                          */
 674                         trans->use_count++;
 675                         return btrfs_commit_transaction(trans, root);
 676                 } else {
 677                         wake_up_process(info->transaction_kthread);
 678                 }
 679         }
 680
 681         if (trans->type < TRANS_JOIN_NOLOCK)
 682                 sb_end_intwrite(root->fs_info->sb);
 683
 684         WARN_ON(cur_trans != info->running_transaction);
 685         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 686         atomic_dec(&cur_trans->num_writers);
 687
 688         smp_mb();
 689         if (waitqueue_active(&cur_trans->writer_wait))
 690                 wake_up(&cur_trans->writer_wait);
 691         put_transaction(cur_trans);
 692
 693         if (current->journal_info == trans)
 694                 current->journal_info = NULL;
 695
 696         if (throttle)
 697                 btrfs_run_delayed_iputs(root);
 698
 699         if (trans->aborted ||
 700             test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
 701                 err = -EIO;
 702         assert_qgroups_uptodate(trans);
 703
 704         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 705         return err;
 706 }
 707
 708 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 709                           struct btrfs_root *root)
 710 {
 711         int ret;
 712
 713         ret = __btrfs_end_transaction(trans, root, 0);
 714         if (ret)
 715                 return ret;
 716         return 0;
 717 }
 718
 719 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 720                                    struct btrfs_root *root)
 721 {
 722         int ret;
 723
 724         ret = __btrfs_end_transaction(trans, root, 1);
 725         if (ret)
 726                 return ret;
 727         return 0;
 728 }
 729
 730 int btrfs_end_transaction_dmeta(struct btrfs_trans_handle *trans,
 731                                 struct btrfs_root *root)
 732 {
 733         return __btrfs_end_transaction(trans, root, 1);
 734 }
 735
 736 /*
 737  * when btree blocks are allocated, they have some corresponding bits set for
 738  * them in one of two extent_io trees.  This is used to make sure all of
 739  * those extents are sent to disk but does not wait on them
 740  */
 741 int btrfs_write_marked_extents(struct btrfs_root *root,
 742                                struct extent_io_tree *dirty_pages, int mark)
 743 {
 744         int err = 0;
 745         int werr = 0;
 746         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 747         struct extent_state *cached_state = NULL;
 748         u64 start = 0;
 749         u64 end;
 750         struct blk_plug plug;
 751
 752         blk_start_plug(&plug);
 753         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 754                                       mark, &cached_state)) {
 755                 convert_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT,
 756                                    mark, &cached_state, GFP_NOFS);
 757                 cached_state = NULL;
 758                 err = filemap_fdatawrite_range(mapping, start, end);
 759                 if (err)
 760                         werr = err;
 761                 cond_resched();
 762                 start = end + 1;
 763         }
 764         if (err)
 765                 werr = err;
 766         blk_finish_plug(&plug);
 767         return werr;
 768 }
 769
 770 /*
 771  * when btree blocks are allocated, they have some corresponding bits set for
 772  * them in one of two extent_io trees.  This is used to make sure all of
 773  * those extents are on disk for transaction or log commit.  We wait
 774  * on all the pages and clear them from the dirty pages state tree
 775  */
 776 int btrfs_wait_marked_extents(struct btrfs_root *root,
 777                               struct extent_io_tree *dirty_pages, int mark)
 778 {
 779         int err = 0;
 780         int werr = 0;
 781         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 782         struct extent_state *cached_state = NULL;
 783         u64 start = 0;
 784         u64 end;
 785
 786         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 787                                       EXTENT_NEED_WAIT, &cached_state)) {
 788                 clear_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT,
 789                                  0, 0, &cached_state, GFP_NOFS);
 790                 err = filemap_fdatawait_range(mapping, start, end);
 791                 if (err)
 792                         werr = err;
 793                 cond_resched();
 794                 start = end + 1;
 795         }
 796         if (err)
 797                 werr = err;
 798         return werr;
 799 }
 800
 801 /*
 802  * when btree blocks are allocated, they have some corresponding bits set for
 803  * them in one of two extent_io trees.  This is used to make sure all of
 804  * those extents are on disk for transaction or log commit
 805  */
 806 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 807                                 struct extent_io_tree *dirty_pages, int mark)
 808 {
 809         int ret;
 810         int ret2;
 811
 812         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 813         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 814
 815         if (ret)
 816                 return ret;
 817         if (ret2)
 818                 return ret2;
 819         return 0;
 820 }
 821
 822 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 823                                      struct btrfs_root *root)
 824 {
 825         if (!trans || !trans->transaction) {
 826                 struct inode *btree_inode;
 827                 btree_inode = root->fs_info->btree_inode;
 828                 return filemap_write_and_wait(btree_inode->i_mapping);
 829         }
 830         return btrfs_write_and_wait_marked_extents(root,
 831                                            &trans->transaction->dirty_pages,
 832                                            EXTENT_DIRTY);
 833 }
 834
 835 /*
 836  * this is used to update the root pointer in the tree of tree roots.
 837  *
 838  * But, in the case of the extent allocation tree, updating the root
 839  * pointer may allocate blocks which may change the root of the extent
 840  * allocation tree.
 841  *
 842  * So, this loops and repeats and makes sure the cowonly root didn't
 843  * change while the root pointer was being updated in the metadata.
 844  */
 845 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 846                                struct btrfs_root *root)
 847 {
 848         int ret;
 849         u64 old_root_bytenr;
 850         u64 old_root_used;
 851         struct btrfs_root *tree_root = root->fs_info->tree_root;
 852
 853         old_root_used = btrfs_root_used(&root->root_item);
 854         btrfs_write_dirty_block_groups(trans, root);
 855
 856         while (1) {
 857                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 858                 if (old_root_bytenr == root->node->start &&
 859                     old_root_used == btrfs_root_used(&root->root_item))
 860                         break;
 861
 862                 btrfs_set_root_node(&root->root_item, root->node);
 863                 ret = btrfs_update_root(trans, tree_root,
 864                                         &root->root_key,
 865                                         &root->root_item);
 866                 if (ret)
 867                         return ret;
 868
 869                 old_root_used = btrfs_root_used(&root->root_item);
 870                 ret = btrfs_write_dirty_block_groups(trans, root);
 871                 if (ret)
 872                         return ret;
 873         }
 874
 875         if (root != root->fs_info->extent_root)
 876                 switch_commit_root(root);
 877
 878         return 0;
 879 }
 880
 881 /*
 882  * update all the cowonly tree roots on disk
 883  *
 884  * The error handling in this function may not be obvious. Any of the
 885  * failures will cause the file system to go offline. We still need
 886  * to clean up the delayed refs.
 887  */
 888 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 889                                          struct btrfs_root *root)
 890 {
 891         struct btrfs_fs_info *fs_info = root->fs_info;
 892         struct list_head *next;
 893         struct extent_buffer *eb;
 894         int ret;
 895
 896         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 897         if (ret)
 898                 return ret;
 899
 900         eb = btrfs_lock_root_node(fs_info->tree_root);
 901         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
 902                               0, &eb);
 903         btrfs_tree_unlock(eb);
 904         free_extent_buffer(eb);
 905
 906         if (ret)
 907                 return ret;
 908
 909         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 910         if (ret)
 911                 return ret;
 912
 913         ret = btrfs_run_dev_stats(trans, root->fs_info);
 914         WARN_ON(ret);
 915         ret = btrfs_run_dev_replace(trans, root->fs_info);
 916         WARN_ON(ret);
 917
 918         ret = btrfs_run_qgroups(trans, root->fs_info);
 919         BUG_ON(ret);
 920
 921         /* run_qgroups might have added some more refs */
 922         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 923         BUG_ON(ret);
 924
 925         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 926                 next = fs_info->dirty_cowonly_roots.next;
 927                 list_del_init(next);
 928                 root = list_entry(next, struct btrfs_root, dirty_list);
 929
 930                 ret = update_cowonly_root(trans, root);
 931                 if (ret)
 932                         return ret;
 933         }
 934
 935         down_write(&fs_info->extent_commit_sem);
 936         switch_commit_root(fs_info->extent_root);
 937         up_write(&fs_info->extent_commit_sem);
 938
 939         btrfs_after_dev_replace_commit(fs_info);
 940
 941         return 0;
 942 }
 943
 944 /*
 945  * dead roots are old snapshots that need to be deleted.  This allocates
 946  * a dirty root struct and adds it into the list of dead roots that need to
 947  * be deleted
 948  */
 949 int btrfs_add_dead_root(struct btrfs_root *root)
 950 {
 951         spin_lock(&root->fs_info->trans_lock);
 952         list_add(&root->root_list, &root->fs_info->dead_roots);
 953         spin_unlock(&root->fs_info->trans_lock);
 954         return 0;
 955 }
 956
 957 /*
 958  * update all the cowonly tree roots on disk
 959  */
 960 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 961                                     struct btrfs_root *root)
 962 {
 963         struct btrfs_root *gang[8];
 964         struct btrfs_fs_info *fs_info = root->fs_info;
 965         int i;
 966         int ret;
 967         int err = 0;
 968
 969         spin_lock(&fs_info->fs_roots_radix_lock);
 970         while (1) {
 971                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
 972                                                  (void **)gang, 0,
 973                                                  ARRAY_SIZE(gang),
 974                                                  BTRFS_ROOT_TRANS_TAG);
 975                 if (ret == 0)
 976                         break;
 977                 for (i = 0; i < ret; i++) {
 978                         root = gang[i];
 979                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 980                                         (unsigned long)root->root_key.objectid,
 981                                         BTRFS_ROOT_TRANS_TAG);
 982                         spin_unlock(&fs_info->fs_roots_radix_lock);
 983
 984                         btrfs_free_log(trans, root);
 985                         btrfs_update_reloc_root(trans, root);
 986                         btrfs_orphan_commit_root(trans, root);
 987
 988                         btrfs_save_ino_cache(root, trans);
 989
 990                         /* see comments in should_cow_block() */
 991                         root->force_cow = 0;
 992                         smp_wmb();
 993
 994                         if (root->commit_root != root->node) {
 995                                 mutex_lock(&root->fs_commit_mutex);
 996                                 switch_commit_root(root);
 997                                 btrfs_unpin_free_ino(root);
 998                                 mutex_unlock(&root->fs_commit_mutex);
 999
1000                                 btrfs_set_root_node(&root->root_item,
1001                                                     root->node);
1002                         }
1003
1004                         err = btrfs_update_root(trans, fs_info->tree_root,
1005                                                 &root->root_key,
1006                                                 &root->root_item);
1007                         spin_lock(&fs_info->fs_roots_radix_lock);
1008                         if (err)
1009                                 break;
1010                 }
1011         }
1012         spin_unlock(&fs_info->fs_roots_radix_lock);
1013         return err;
1014 }
1015
1016 /*
1017  * defrag a given btree.
1018  * Every leaf in the btree is read and defragged.
1019  */
1020 int btrfs_defrag_root(struct btrfs_root *root)
1021 {
1022         struct btrfs_fs_info *info = root->fs_info;
1023         struct btrfs_trans_handle *trans;
1024         int ret;
1025
1026         if (xchg(&root->defrag_running, 1))
1027                 return 0;
1028
1029         while (1) {
1030                 trans = btrfs_start_transaction(root, 0);
1031                 if (IS_ERR(trans))
1032                         return PTR_ERR(trans);
1033
1034                 ret = btrfs_defrag_leaves(trans, root);
1035
1036                 btrfs_end_transaction(trans, root);
1037                 btrfs_btree_balance_dirty(info->tree_root);
1038                 cond_resched();
1039
1040                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
1041                         break;
1042
1043                 if (btrfs_defrag_cancelled(root->fs_info)) {
1044                         printk(KERN_DEBUG "btrfs: defrag_root cancelled\n");
1045                         ret = -EAGAIN;
1046                         break;
1047                 }
1048         }
1049         root->defrag_running = 0;
1050         return ret;
1051 }
1052
1053 /*
1054  * new snapshots need to be created at a very specific time in the
1055  * transaction commit.  This does the actual creation.
1056  *
1057  * Note:
1058  * If the error which may affect the commitment of the current transaction
1059  * happens, we should return the error number. If the error which just affect
1060  * the creation of the pending snapshots, just return 0.
1061  */
1062 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1063                                    struct btrfs_fs_info *fs_info,
1064                                    struct btrfs_pending_snapshot *pending)
1065 {
1066         struct btrfs_key key;
1067         struct btrfs_root_item *new_root_item;
1068         struct btrfs_root *tree_root = fs_info->tree_root;
1069         struct btrfs_root *root = pending->root;
1070         struct btrfs_root *parent_root;
1071         struct btrfs_block_rsv *rsv;
1072         struct inode *parent_inode;
1073         struct btrfs_path *path;
1074         struct btrfs_dir_item *dir_item;
1075         struct dentry *dentry;
1076         struct extent_buffer *tmp;
1077         struct extent_buffer *old;
1078         struct timespec cur_time = CURRENT_TIME;
1079         int ret = 0;
1080         u64 to_reserve = 0;
1081         u64 index = 0;
1082         u64 objectid;
1083         u64 root_flags;
1084         uuid_le new_uuid;
1085
1086         path = btrfs_alloc_path();
1087         if (!path) {
1088                 pending->error = -ENOMEM;
1089                 return 0;
1090         }
1091
1092         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
1093         if (!new_root_item) {
1094                 pending->error = -ENOMEM;
1095                 goto root_item_alloc_fail;
1096         }
1097
1098         pending->error = btrfs_find_free_objectid(tree_root, &objectid);
1099         if (pending->error)
1100                 goto no_free_objectid;
1101
1102         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
1103
1104         if (to_reserve > 0) {
1105                 pending->error = btrfs_block_rsv_add(root,
1106                                                      &pending->block_rsv,
1107                                                      to_reserve,
1108                                                      BTRFS_RESERVE_NO_FLUSH);
1109                 if (pending->error)
1110                         goto no_free_objectid;
1111         }
1112
1113         pending->error = btrfs_qgroup_inherit(trans, fs_info,
1114                                               root->root_key.objectid,
1115                                               objectid, pending->inherit);
1116         if (pending->error)
1117                 goto no_free_objectid;
1118
1119         key.objectid = objectid;
1120         key.offset = (u64)-1;
1121         key.type = BTRFS_ROOT_ITEM_KEY;
1122
1123         rsv = trans->block_rsv;
1124         trans->block_rsv = &pending->block_rsv;
1125         trans->bytes_reserved = trans->block_rsv->reserved;
1126
1127         dentry = pending->dentry;
1128         parent_inode = pending->dir;
1129         parent_root = BTRFS_I(parent_inode)->root;
1130         record_root_in_trans(trans, parent_root);
1131
1132         /*
1133          * insert the directory item
1134          */
1135         ret = btrfs_set_inode_index(parent_inode, &index);
1136         BUG_ON(ret); /* -ENOMEM */
1137
1138         /* check if there is a file/dir which has the same name. */
1139         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1140                                          btrfs_ino(parent_inode),
1141                                          dentry->d_name.name,
1142                                          dentry->d_name.len, 0);
1143         if (dir_item != NULL && !IS_ERR(dir_item)) {
1144                 pending->error = -EEXIST;
1145                 goto dir_item_existed;
1146         } else if (IS_ERR(dir_item)) {
1147                 ret = PTR_ERR(dir_item);
1148                 btrfs_abort_transaction(trans, root, ret);
1149                 goto fail;
1150         }
1151         btrfs_release_path(path);
1152
1153         /*
1154          * pull in the delayed directory update
1155          * and the delayed inode item
1156          * otherwise we corrupt the FS during
1157          * snapshot
1158          */
1159         ret = btrfs_run_delayed_items(trans, root);
1160         if (ret) {      /* Transaction aborted */
1161                 btrfs_abort_transaction(trans, root, ret);
1162                 goto fail;
1163         }
1164
1165         record_root_in_trans(trans, root);
1166         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1167         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1168         btrfs_check_and_init_root_item(new_root_item);
1169
1170         root_flags = btrfs_root_flags(new_root_item);
1171         if (pending->readonly)
1172                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1173         else
1174                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1175         btrfs_set_root_flags(new_root_item, root_flags);
1176
1177         btrfs_set_root_generation_v2(new_root_item,
1178                         trans->transid);
1179         uuid_le_gen(&new_uuid);
1180         memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1181         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1182                         BTRFS_UUID_SIZE);
1183         new_root_item->otime.sec = cpu_to_le64(cur_time.tv_sec);
1184         new_root_item->otime.nsec = cpu_to_le32(cur_time.tv_nsec);
1185         btrfs_set_root_otransid(new_root_item, trans->transid);
1186         memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1187         memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1188         btrfs_set_root_stransid(new_root_item, 0);
1189         btrfs_set_root_rtransid(new_root_item, 0);
1190
1191         old = btrfs_lock_root_node(root);
1192         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1193         if (ret) {
1194                 btrfs_tree_unlock(old);
1195                 free_extent_buffer(old);
1196                 btrfs_abort_transaction(trans, root, ret);
1197                 goto fail;
1198         }
1199
1200         btrfs_set_lock_blocking(old);
1201
1202         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1203         /* clean up in any case */
1204         btrfs_tree_unlock(old);
1205         free_extent_buffer(old);
1206         if (ret) {
1207                 btrfs_abort_transaction(trans, root, ret);
1208                 goto fail;
1209         }
1210
1211         /* see comments in should_cow_block() */
1212         root->force_cow = 1;
1213         smp_wmb();
1214
1215         btrfs_set_root_node(new_root_item, tmp);
1216         /* record when the snapshot was created in key.offset */
1217         key.offset = trans->transid;
1218         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1219         btrfs_tree_unlock(tmp);
1220         free_extent_buffer(tmp);
1221         if (ret) {
1222                 btrfs_abort_transaction(trans, root, ret);
1223                 goto fail;
1224         }
1225
1226         /*
1227          * insert root back/forward references
1228          */
1229         ret = btrfs_add_root_ref(trans, tree_root, objectid,
1230                                  parent_root->root_key.objectid,
1231                                  btrfs_ino(parent_inode), index,
1232                                  dentry->d_name.name, dentry->d_name.len);
1233         if (ret) {
1234                 btrfs_abort_transaction(trans, root, ret);
1235                 goto fail;
1236         }
1237
1238         key.offset = (u64)-1;
1239         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1240         if (IS_ERR(pending->snap)) {
1241                 ret = PTR_ERR(pending->snap);
1242                 btrfs_abort_transaction(trans, root, ret);
1243                 goto fail;
1244         }
1245
1246         ret = btrfs_reloc_post_snapshot(trans, pending);
1247         if (ret) {
1248                 btrfs_abort_transaction(trans, root, ret);
1249                 goto fail;
1250         }
1251
1252         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1253         if (ret) {
1254                 btrfs_abort_transaction(trans, root, ret);
1255                 goto fail;
1256         }
1257
1258         ret = btrfs_insert_dir_item(trans, parent_root,
1259                                     dentry->d_name.name, dentry->d_name.len,
1260                                     parent_inode, &key,
1261                                     BTRFS_FT_DIR, index);
1262         /* We have check then name at the beginning, so it is impossible. */
1263         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1264         if (ret) {
1265                 btrfs_abort_transaction(trans, root, ret);
1266                 goto fail;
1267         }
1268
1269         btrfs_i_size_write(parent_inode, parent_inode->i_size +
1270                                          dentry->d_name.len * 2);
1271         parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
1272         ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1273         if (ret)
1274                 btrfs_abort_transaction(trans, root, ret);
1275 fail:
1276         pending->error = ret;
1277 dir_item_existed:
1278         trans->block_rsv = rsv;
1279         trans->bytes_reserved = 0;
1280 no_free_objectid:
1281         kfree(new_root_item);
1282 root_item_alloc_fail:
1283         btrfs_free_path(path);
1284         return ret;
1285 }
1286
1287 /*
1288  * create all the snapshots we've scheduled for creation
1289  */
1290 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1291                                              struct btrfs_fs_info *fs_info)
1292 {
1293         struct btrfs_pending_snapshot *pending, *next;
1294         struct list_head *head = &trans->transaction->pending_snapshots;
1295         int ret = 0;
1296
1297         list_for_each_entry_safe(pending, next, head, list) {
1298                 list_del(&pending->list);
1299                 ret = create_pending_snapshot(trans, fs_info, pending);
1300                 if (ret)
1301                         break;
1302         }
1303         return ret;
1304 }
1305
1306 static void update_super_roots(struct btrfs_root *root)
1307 {
1308         struct btrfs_root_item *root_item;
1309         struct btrfs_super_block *super;
1310
1311         super = root->fs_info->super_copy;
1312
1313         root_item = &root->fs_info->chunk_root->root_item;
1314         super->chunk_root = root_item->bytenr;
1315         super->chunk_root_generation = root_item->generation;
1316         super->chunk_root_level = root_item->level;
1317
1318         root_item = &root->fs_info->tree_root->root_item;
1319         super->root = root_item->bytenr;
1320         super->generation = root_item->generation;
1321         super->root_level = root_item->level;
1322         if (btrfs_test_opt(root, SPACE_CACHE))
1323                 super->cache_generation = root_item->generation;
1324 }
1325
1326 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1327 {
1328         int ret = 0;
1329         spin_lock(&info->trans_lock);
1330         if (info->running_transaction)
1331                 ret = info->running_transaction->in_commit;
1332         spin_unlock(&info->trans_lock);
1333         return ret;
1334 }
1335
1336 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1337 {
1338         int ret = 0;
1339         spin_lock(&info->trans_lock);
1340         if (info->running_transaction)
1341                 ret = info->running_transaction->blocked;
1342         spin_unlock(&info->trans_lock);
1343         return ret;
1344 }
1345
1346 /*
1347  * wait for the current transaction commit to start and block subsequent
1348  * transaction joins
1349  */
1350 static void wait_current_trans_commit_start(struct btrfs_root *root,
1351                                             struct btrfs_transaction *trans)
1352 {
1353         wait_event(root->fs_info->transaction_blocked_wait, trans->in_commit);
1354 }
1355
1356 /*
1357  * wait for the current transaction to start and then become unblocked.
1358  * caller holds ref.
1359  */
1360 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1361                                          struct btrfs_transaction *trans)
1362 {
1363         wait_event(root->fs_info->transaction_wait,
1364                    trans->commit_done || (trans->in_commit && !trans->blocked));
1365 }
1366
1367 /*
1368  * commit transactions asynchronously. once btrfs_commit_transaction_async
1369  * returns, any subsequent transaction will not be allowed to join.
1370  */
1371 struct btrfs_async_commit {
1372         struct btrfs_trans_handle *newtrans;
1373         struct btrfs_root *root;
1374         struct work_struct work;
1375 };
1376
1377 static void do_async_commit(struct work_struct *work)
1378 {
1379         struct btrfs_async_commit *ac =
1380                 container_of(work, struct btrfs_async_commit, work);
1381
1382         /*
1383          * We've got freeze protection passed with the transaction.
1384          * Tell lockdep about it.
1385          */
1386         if (ac->newtrans->type < TRANS_JOIN_NOLOCK)
1387                 rwsem_acquire_read(
1388                      &ac->root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
1389                      0, 1, _THIS_IP_);
1390
1391         current->journal_info = ac->newtrans;
1392
1393         btrfs_commit_transaction(ac->newtrans, ac->root);
1394         kfree(ac);
1395 }
1396
1397 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1398                                    struct btrfs_root *root,
1399                                    int wait_for_unblock)
1400 {
1401         struct btrfs_async_commit *ac;
1402         struct btrfs_transaction *cur_trans;
1403
1404         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1405         if (!ac)
1406                 return -ENOMEM;
1407
1408         INIT_WORK(&ac->work, do_async_commit);
1409         ac->root = root;
1410         ac->newtrans = btrfs_join_transaction(root);
1411         if (IS_ERR(ac->newtrans)) {
1412                 int err = PTR_ERR(ac->newtrans);
1413                 kfree(ac);
1414                 return err;
1415         }
1416
1417         /* take transaction reference */
1418         cur_trans = trans->transaction;
1419         atomic_inc(&cur_trans->use_count);
1420
1421         btrfs_end_transaction(trans, root);
1422
1423         /*
1424          * Tell lockdep we've released the freeze rwsem, since the
1425          * async commit thread will be the one to unlock it.
1426          */
1427         if (trans->type < TRANS_JOIN_NOLOCK)
1428                 rwsem_release(
1429                         &root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
1430                         1, _THIS_IP_);
1431
1432         schedule_work(&ac->work);
1433
1434         /* wait for transaction to start and unblock */
1435         if (wait_for_unblock)
1436                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1437         else
1438                 wait_current_trans_commit_start(root, cur_trans);
1439
1440         if (current->journal_info == trans)
1441                 current->journal_info = NULL;
1442
1443         put_transaction(cur_trans);
1444         return 0;
1445 }
1446
1447
1448 static void cleanup_transaction(struct btrfs_trans_handle *trans,
1449                                 struct btrfs_root *root, int err)
1450 {
1451         struct btrfs_transaction *cur_trans = trans->transaction;
1452         DEFINE_WAIT(wait);
1453
1454         WARN_ON(trans->use_count > 1);
1455
1456         btrfs_abort_transaction(trans, root, err);
1457
1458         spin_lock(&root->fs_info->trans_lock);
1459
1460         if (list_empty(&cur_trans->list)) {
1461                 spin_unlock(&root->fs_info->trans_lock);
1462                 btrfs_end_transaction(trans, root);
1463                 return;
1464         }
1465
1466         list_del_init(&cur_trans->list);
1467         if (cur_trans == root->fs_info->running_transaction) {
1468                 root->fs_info->trans_no_join = 1;
1469                 spin_unlock(&root->fs_info->trans_lock);
1470                 wait_event(cur_trans->writer_wait,
1471                            atomic_read(&cur_trans->num_writers) == 1);
1472
1473                 spin_lock(&root->fs_info->trans_lock);
1474                 root->fs_info->running_transaction = NULL;
1475         }
1476         spin_unlock(&root->fs_info->trans_lock);
1477
1478         btrfs_cleanup_one_transaction(trans->transaction, root);
1479
1480         put_transaction(cur_trans);
1481         put_transaction(cur_trans);
1482
1483         trace_btrfs_transaction_commit(root);
1484
1485         btrfs_scrub_continue(root);
1486
1487         if (current->journal_info == trans)
1488                 current->journal_info = NULL;
1489
1490         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1491 }
1492
1493 static int btrfs_flush_all_pending_stuffs(struct btrfs_trans_handle *trans,
1494                                           struct btrfs_root *root)
1495 {
1496         int flush_on_commit = btrfs_test_opt(root, FLUSHONCOMMIT);
1497         int snap_pending = 0;
1498         int ret;
1499
1500         if (!flush_on_commit) {
1501                 spin_lock(&root->fs_info->trans_lock);
1502                 if (!list_empty(&trans->transaction->pending_snapshots))
1503                         snap_pending = 1;
1504                 spin_unlock(&root->fs_info->trans_lock);
1505         }
1506
1507         if (flush_on_commit || snap_pending) {
1508                 ret = btrfs_start_delalloc_inodes(root, 1);
1509                 if (ret)
1510                         return ret;
1511                 btrfs_wait_ordered_extents(root, 1);
1512         }
1513
1514         ret = btrfs_run_delayed_items(trans, root);
1515         if (ret)
1516                 return ret;
1517
1518         /*
1519          * running the delayed items may have added new refs. account
1520          * them now so that they hinder processing of more delayed refs
1521          * as little as possible.
1522          */
1523         btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
1524
1525         /*
1526          * rename don't use btrfs_join_transaction, so, once we
1527          * set the transaction to blocked above, we aren't going
1528          * to get any new ordered operations.  We can safely run
1529          * it here and no for sure that nothing new will be added
1530          * to the list
1531          */
1532         ret = btrfs_run_ordered_operations(trans, root, 1);
1533
1534         return ret;
1535 }
1536
1537 /*
1538  * btrfs_transaction state sequence:
1539  *    in_commit = 0, blocked = 0  (initial)
1540  *    in_commit = 1, blocked = 1
1541  *    blocked = 0
1542  *    commit_done = 1
1543  */
1544 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1545                              struct btrfs_root *root)
1546 {
1547         unsigned long joined = 0;
1548         struct btrfs_transaction *cur_trans = trans->transaction;
1549         struct btrfs_transaction *prev_trans = NULL;
1550         DEFINE_WAIT(wait);
1551         int ret;
1552         int should_grow = 0;
1553         unsigned long now = get_seconds();
1554
1555         ret = btrfs_run_ordered_operations(trans, root, 0);
1556         if (ret) {
1557                 btrfs_abort_transaction(trans, root, ret);
1558                 btrfs_end_transaction(trans, root);
1559                 return ret;
1560         }
1561
1562         /* Stop the commit early if ->aborted is set */
1563         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1564                 ret = cur_trans->aborted;
1565                 btrfs_end_transaction(trans, root);
1566                 return ret;
1567         }
1568
1569         /* make a pass through all the delayed refs we have so far
1570          * any runnings procs may add more while we are here
1571          */
1572         ret = btrfs_run_delayed_refs(trans, root, 0);
1573         if (ret) {
1574                 btrfs_end_transaction(trans, root);
1575                 return ret;
1576         }
1577
1578         btrfs_trans_release_metadata(trans, root);
1579         trans->block_rsv = NULL;
1580         if (trans->qgroup_reserved) {
1581                 btrfs_qgroup_free(root, trans->qgroup_reserved);
1582                 trans->qgroup_reserved = 0;
1583         }
1584
1585         cur_trans = trans->transaction;
1586
1587         /*
1588          * set the flushing flag so procs in this transaction have to
1589          * start sending their work down.
1590          */
1591         cur_trans->delayed_refs.flushing = 1;
1592
1593         if (!list_empty(&trans->new_bgs))
1594                 btrfs_create_pending_block_groups(trans, root);
1595
1596         ret = btrfs_run_delayed_refs(trans, root, 0);
1597         if (ret) {
1598                 btrfs_end_transaction(trans, root);
1599                 return ret;
1600         }
1601
1602         spin_lock(&cur_trans->commit_lock);
1603         if (cur_trans->in_commit) {
1604                 spin_unlock(&cur_trans->commit_lock);
1605                 atomic_inc(&cur_trans->use_count);
1606                 ret = btrfs_end_transaction(trans, root);
1607
1608                 wait_for_commit(root, cur_trans);
1609
1610                 put_transaction(cur_trans);
1611
1612                 return ret;
1613         }
1614
1615         trans->transaction->in_commit = 1;
1616         trans->transaction->blocked = 1;
1617         spin_unlock(&cur_trans->commit_lock);
1618         wake_up(&root->fs_info->transaction_blocked_wait);
1619
1620         spin_lock(&root->fs_info->trans_lock);
1621         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1622                 prev_trans = list_entry(cur_trans->list.prev,
1623                                         struct btrfs_transaction, list);
1624                 if (!prev_trans->commit_done) {
1625                         atomic_inc(&prev_trans->use_count);
1626                         spin_unlock(&root->fs_info->trans_lock);
1627
1628                         wait_for_commit(root, prev_trans);
1629
1630                         put_transaction(prev_trans);
1631                 } else {
1632                         spin_unlock(&root->fs_info->trans_lock);
1633                 }
1634         } else {
1635                 spin_unlock(&root->fs_info->trans_lock);
1636         }
1637
1638         if (!btrfs_test_opt(root, SSD) &&
1639             (now < cur_trans->start_time || now - cur_trans->start_time < 1))
1640                 should_grow = 1;
1641
1642         do {
1643                 joined = cur_trans->num_joined;
1644
1645                 WARN_ON(cur_trans != trans->transaction);
1646
1647                 ret = btrfs_flush_all_pending_stuffs(trans, root);
1648                 if (ret)
1649                         goto cleanup_transaction;
1650
1651                 prepare_to_wait(&cur_trans->writer_wait, &wait,
1652                                 TASK_UNINTERRUPTIBLE);
1653
1654                 if (atomic_read(&cur_trans->num_writers) > 1)
1655                         schedule_timeout(MAX_SCHEDULE_TIMEOUT);
1656                 else if (should_grow)
1657                         schedule_timeout(1);
1658
1659                 finish_wait(&cur_trans->writer_wait, &wait);
1660         } while (atomic_read(&cur_trans->num_writers) > 1 ||
1661                  (should_grow && cur_trans->num_joined != joined));
1662
1663         ret = btrfs_flush_all_pending_stuffs(trans, root);
1664         if (ret)
1665                 goto cleanup_transaction;
1666
1667         /*
1668          * Ok now we need to make sure to block out any other joins while we
1669          * commit the transaction.  We could have started a join before setting
1670          * no_join so make sure to wait for num_writers to == 1 again.
1671          */
1672         spin_lock(&root->fs_info->trans_lock);
1673         root->fs_info->trans_no_join = 1;
1674         spin_unlock(&root->fs_info->trans_lock);
1675         wait_event(cur_trans->writer_wait,
1676                    atomic_read(&cur_trans->num_writers) == 1);
1677
1678         /* ->aborted might be set after the previous check, so check it */
1679         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1680                 ret = cur_trans->aborted;
1681                 goto cleanup_transaction;
1682         }
1683         /*
1684          * the reloc mutex makes sure that we stop
1685          * the balancing code from coming in and moving
1686          * extents around in the middle of the commit
1687          */
1688         mutex_lock(&root->fs_info->reloc_mutex);
1689
1690         /*
1691          * We needn't worry about the delayed items because we will
1692          * deal with them in create_pending_snapshot(), which is the
1693          * core function of the snapshot creation.
1694          */
1695         ret = create_pending_snapshots(trans, root->fs_info);
1696         if (ret) {
1697                 mutex_unlock(&root->fs_info->reloc_mutex);
1698                 goto cleanup_transaction;
1699         }
1700
1701         /*
1702          * We insert the dir indexes of the snapshots and update the inode
1703          * of the snapshots' parents after the snapshot creation, so there
1704          * are some delayed items which are not dealt with. Now deal with
1705          * them.
1706          *
1707          * We needn't worry that this operation will corrupt the snapshots,
1708          * because all the tree which are snapshoted will be forced to COW
1709          * the nodes and leaves.
1710          */
1711         ret = btrfs_run_delayed_items(trans, root);
1712         if (ret) {
1713                 mutex_unlock(&root->fs_info->reloc_mutex);
1714                 goto cleanup_transaction;
1715         }
1716
1717         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1718         if (ret) {
1719                 mutex_unlock(&root->fs_info->reloc_mutex);
1720                 goto cleanup_transaction;
1721         }
1722
1723         /*
1724          * make sure none of the code above managed to slip in a
1725          * delayed item
1726          */
1727         btrfs_assert_delayed_root_empty(root);
1728
1729         WARN_ON(cur_trans != trans->transaction);
1730
1731         btrfs_scrub_pause(root);
1732         /* btrfs_commit_tree_roots is responsible for getting the
1733          * various roots consistent with each other.  Every pointer
1734          * in the tree of tree roots has to point to the most up to date
1735          * root for every subvolume and other tree.  So, we have to keep
1736          * the tree logging code from jumping in and changing any
1737          * of the trees.
1738          *
1739          * At this point in the commit, there can't be any tree-log
1740          * writers, but a little lower down we drop the trans mutex
1741          * and let new people in.  By holding the tree_log_mutex
1742          * from now until after the super is written, we avoid races
1743          * with the tree-log code.
1744          */
1745         mutex_lock(&root->fs_info->tree_log_mutex);
1746
1747         ret = commit_fs_roots(trans, root);
1748         if (ret) {
1749                 mutex_unlock(&root->fs_info->tree_log_mutex);
1750                 mutex_unlock(&root->fs_info->reloc_mutex);
1751                 goto cleanup_transaction;
1752         }
1753
1754         /* commit_fs_roots gets rid of all the tree log roots, it is now
1755          * safe to free the root of tree log roots
1756          */
1757         btrfs_free_log_root_tree(trans, root->fs_info);
1758
1759         ret = commit_cowonly_roots(trans, root);
1760         if (ret) {
1761                 mutex_unlock(&root->fs_info->tree_log_mutex);
1762                 mutex_unlock(&root->fs_info->reloc_mutex);
1763                 goto cleanup_transaction;
1764         }
1765
1766         /*
1767          * The tasks which save the space cache and inode cache may also
1768          * update ->aborted, check it.
1769          */
1770         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1771                 ret = cur_trans->aborted;
1772                 mutex_unlock(&root->fs_info->tree_log_mutex);
1773                 mutex_unlock(&root->fs_info->reloc_mutex);
1774                 goto cleanup_transaction;
1775         }
1776
1777         btrfs_prepare_extent_commit(trans, root);
1778
1779         cur_trans = root->fs_info->running_transaction;
1780
1781         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1782                             root->fs_info->tree_root->node);
1783         switch_commit_root(root->fs_info->tree_root);
1784
1785         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1786                             root->fs_info->chunk_root->node);
1787         switch_commit_root(root->fs_info->chunk_root);
1788
1789         assert_qgroups_uptodate(trans);
1790         update_super_roots(root);
1791
1792         if (!root->fs_info->log_root_recovering) {
1793                 btrfs_set_super_log_root(root->fs_info->super_copy, 0);
1794                 btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
1795         }
1796
1797         memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
1798                sizeof(*root->fs_info->super_copy));
1799
1800         trans->transaction->blocked = 0;
1801         spin_lock(&root->fs_info->trans_lock);
1802         root->fs_info->running_transaction = NULL;
1803         root->fs_info->trans_no_join = 0;
1804         spin_unlock(&root->fs_info->trans_lock);
1805         mutex_unlock(&root->fs_info->reloc_mutex);
1806
1807         wake_up(&root->fs_info->transaction_wait);
1808
1809         ret = btrfs_write_and_wait_transaction(trans, root);
1810         if (ret) {
1811                 btrfs_error(root->fs_info, ret,
1812                             "Error while writing out transaction.");
1813                 mutex_unlock(&root->fs_info->tree_log_mutex);
1814                 goto cleanup_transaction;
1815         }
1816
1817         ret = write_ctree_super(trans, root, 0);
1818         if (ret) {
1819                 mutex_unlock(&root->fs_info->tree_log_mutex);
1820                 goto cleanup_transaction;
1821         }
1822
1823         /*
1824          * the super is written, we can safely allow the tree-loggers
1825          * to go about their business
1826          */
1827         mutex_unlock(&root->fs_info->tree_log_mutex);
1828
1829         btrfs_finish_extent_commit(trans, root);
1830
1831         cur_trans->commit_done = 1;
1832
1833         root->fs_info->last_trans_committed = cur_trans->transid;
1834
1835         wake_up(&cur_trans->commit_wait);
1836
1837         spin_lock(&root->fs_info->trans_lock);
1838         list_del_init(&cur_trans->list);
1839         spin_unlock(&root->fs_info->trans_lock);
1840
1841         put_transaction(cur_trans);
1842         put_transaction(cur_trans);
1843
1844         if (trans->type < TRANS_JOIN_NOLOCK)
1845                 sb_end_intwrite(root->fs_info->sb);
1846
1847         trace_btrfs_transaction_commit(root);
1848
1849         btrfs_scrub_continue(root);
1850
1851         if (current->journal_info == trans)
1852                 current->journal_info = NULL;
1853
1854         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1855
1856         if (current != root->fs_info->transaction_kthread)
1857                 btrfs_run_delayed_iputs(root);
1858
1859         return ret;
1860
1861 cleanup_transaction:
1862         btrfs_trans_release_metadata(trans, root);
1863         trans->block_rsv = NULL;
1864         if (trans->qgroup_reserved) {
1865                 btrfs_qgroup_free(root, trans->qgroup_reserved);
1866                 trans->qgroup_reserved = 0;
1867         }
1868         btrfs_printk(root->fs_info, "Skipping commit of aborted transaction.\n");
1869 //      WARN_ON(1);
1870         if (current->journal_info == trans)
1871                 current->journal_info = NULL;
1872         cleanup_transaction(trans, root, ret);
1873
1874         return ret;
1875 }
1876
1877 /*
1878  * interface function to delete all the snapshots we have scheduled for deletion
1879  */
1880 int btrfs_clean_old_snapshots(struct btrfs_root *root)
1881 {
1882         LIST_HEAD(list);
1883         struct btrfs_fs_info *fs_info = root->fs_info;
1884
1885         spin_lock(&fs_info->trans_lock);
1886         list_splice_init(&fs_info->dead_roots, &list);
1887         spin_unlock(&fs_info->trans_lock);
1888
1889         while (!list_empty(&list)) {
1890                 int ret;
1891
1892                 root = list_entry(list.next, struct btrfs_root, root_list);
1893                 list_del(&root->root_list);
1894
1895                 btrfs_kill_all_delayed_nodes(root);
1896
1897                 if (btrfs_header_backref_rev(root->node) <
1898                     BTRFS_MIXED_BACKREF_REV)
1899                         ret = btrfs_drop_snapshot(root, NULL, 0, 0);
1900                 else
1901                         ret =btrfs_drop_snapshot(root, NULL, 1, 0);
1902                 BUG_ON(ret < 0);
1903         }
1904         return 0;
1905 }