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 static unsigned int btrfs_blocked_trans_types[TRANS_STATE_MAX] = {
  38         [TRANS_STATE_RUNNING]           = 0U,
  39         [TRANS_STATE_BLOCKED]           = (__TRANS_USERSPACE |
  40                                            __TRANS_START),
  41         [TRANS_STATE_COMMIT_START]      = (__TRANS_USERSPACE |
  42                                            __TRANS_START |
  43                                            __TRANS_ATTACH),
  44         [TRANS_STATE_COMMIT_DOING]      = (__TRANS_USERSPACE |
  45                                            __TRANS_START |
  46                                            __TRANS_ATTACH |
  47                                            __TRANS_JOIN),
  48         [TRANS_STATE_UNBLOCKED]         = (__TRANS_USERSPACE |
  49                                            __TRANS_START |
  50                                            __TRANS_ATTACH |
  51                                            __TRANS_JOIN |
  52                                            __TRANS_JOIN_NOLOCK),
  53         [TRANS_STATE_COMPLETED]         = (__TRANS_USERSPACE |
  54                                            __TRANS_START |
  55                                            __TRANS_ATTACH |
  56                                            __TRANS_JOIN |
  57                                            __TRANS_JOIN_NOLOCK),
  58 };
  59
  60 static void put_transaction(struct btrfs_transaction *transaction)
  61 {
  62         WARN_ON(atomic_read(&transaction->use_count) == 0);
  63         if (atomic_dec_and_test(&transaction->use_count)) {
  64                 BUG_ON(!list_empty(&transaction->list));
  65                 WARN_ON(transaction->delayed_refs.root.rb_node);
  66                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  67         }
  68 }
  69
  70 static noinline void switch_commit_root(struct btrfs_root *root)
  71 {
  72         free_extent_buffer(root->commit_root);
  73         root->commit_root = btrfs_root_node(root);
  74 }
  75
  76 static inline void extwriter_counter_inc(struct btrfs_transaction *trans,
  77                                          unsigned int type)
  78 {
  79         if (type & TRANS_EXTWRITERS)
  80                 atomic_inc(&trans->num_extwriters);
  81 }
  82
  83 static inline void extwriter_counter_dec(struct btrfs_transaction *trans,
  84                                          unsigned int type)
  85 {
  86         if (type & TRANS_EXTWRITERS)
  87                 atomic_dec(&trans->num_extwriters);
  88 }
  89
  90 static inline void extwriter_counter_init(struct btrfs_transaction *trans,
  91                                           unsigned int type)
  92 {
  93         atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0));
  94 }
  95
  96 static inline int extwriter_counter_read(struct btrfs_transaction *trans)
  97 {
  98         return atomic_read(&trans->num_extwriters);
  99 }
 100
 101 /*
 102  * either allocate a new transaction or hop into the existing one
 103  */
 104 static noinline int join_transaction(struct btrfs_root *root, unsigned int type)
 105 {
 106         struct btrfs_transaction *cur_trans;
 107         struct btrfs_fs_info *fs_info = root->fs_info;
 108
 109         spin_lock(&fs_info->trans_lock);
 110 loop:
 111         /* The file system has been taken offline. No new transactions. */
 112         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
 113                 spin_unlock(&fs_info->trans_lock);
 114                 return -EROFS;
 115         }
 116
 117         cur_trans = fs_info->running_transaction;
 118         if (cur_trans) {
 119                 if (cur_trans->aborted) {
 120                         spin_unlock(&fs_info->trans_lock);
 121                         return cur_trans->aborted;
 122                 }
 123                 if (btrfs_blocked_trans_types[cur_trans->state] & type) {
 124                         spin_unlock(&fs_info->trans_lock);
 125                         return -EBUSY;
 126                 }
 127                 atomic_inc(&cur_trans->use_count);
 128                 atomic_inc(&cur_trans->num_writers);
 129                 extwriter_counter_inc(cur_trans, type);
 130                 spin_unlock(&fs_info->trans_lock);
 131                 return 0;
 132         }
 133         spin_unlock(&fs_info->trans_lock);
 134
 135         /*
 136          * If we are ATTACH, we just want to catch the current transaction,
 137          * and commit it. If there is no transaction, just return ENOENT.
 138          */
 139         if (type == TRANS_ATTACH)
 140                 return -ENOENT;
 141
 142         /*
 143          * JOIN_NOLOCK only happens during the transaction commit, so
 144          * it is impossible that ->running_transaction is NULL
 145          */
 146         BUG_ON(type == TRANS_JOIN_NOLOCK);
 147
 148         cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
 149         if (!cur_trans)
 150                 return -ENOMEM;
 151
 152         spin_lock(&fs_info->trans_lock);
 153         if (fs_info->running_transaction) {
 154                 /*
 155                  * someone started a transaction after we unlocked.  Make sure
 156                  * to redo the checks above
 157                  */
 158                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 159                 goto loop;
 160         } else if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
 161                 spin_unlock(&fs_info->trans_lock);
 162                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 163                 return -EROFS;
 164         }
 165
 166         atomic_set(&cur_trans->num_writers, 1);
 167         extwriter_counter_init(cur_trans, type);
 168         init_waitqueue_head(&cur_trans->writer_wait);
 169         init_waitqueue_head(&cur_trans->commit_wait);
 170         cur_trans->state = TRANS_STATE_RUNNING;
 171         /*
 172          * One for this trans handle, one so it will live on until we
 173          * commit the transaction.
 174          */
 175         atomic_set(&cur_trans->use_count, 2);
 176         cur_trans->start_time = get_seconds();
 177
 178         cur_trans->delayed_refs.root = RB_ROOT;
 179         cur_trans->delayed_refs.num_entries = 0;
 180         cur_trans->delayed_refs.num_heads_ready = 0;
 181         cur_trans->delayed_refs.num_heads = 0;
 182         cur_trans->delayed_refs.flushing = 0;
 183         cur_trans->delayed_refs.run_delayed_start = 0;
 184
 185         /*
 186          * although the tree mod log is per file system and not per transaction,
 187          * the log must never go across transaction boundaries.
 188          */
 189         smp_mb();
 190         if (!list_empty(&fs_info->tree_mod_seq_list))
 191                 WARN(1, KERN_ERR "btrfs: tree_mod_seq_list not empty when "
 192                         "creating a fresh transaction\n");
 193         if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
 194                 WARN(1, KERN_ERR "btrfs: tree_mod_log rb tree not empty when "
 195                         "creating a fresh transaction\n");
 196         atomic64_set(&fs_info->tree_mod_seq, 0);
 197
 198         spin_lock_init(&cur_trans->delayed_refs.lock);
 199         atomic_set(&cur_trans->delayed_refs.procs_running_refs, 0);
 200         atomic_set(&cur_trans->delayed_refs.ref_seq, 0);
 201         init_waitqueue_head(&cur_trans->delayed_refs.wait);
 202
 203         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 204         INIT_LIST_HEAD(&cur_trans->ordered_operations);
 205         list_add_tail(&cur_trans->list, &fs_info->trans_list);
 206         extent_io_tree_init(&cur_trans->dirty_pages,
 207                              fs_info->btree_inode->i_mapping);
 208         fs_info->generation++;
 209         cur_trans->transid = fs_info->generation;
 210         fs_info->running_transaction = cur_trans;
 211         cur_trans->aborted = 0;
 212         spin_unlock(&fs_info->trans_lock);
 213
 214         return 0;
 215 }
 216
 217 /*
 218  * this does all the record keeping required to make sure that a reference
 219  * counted root is properly recorded in a given transaction.  This is required
 220  * to make sure the old root from before we joined the transaction is deleted
 221  * when the transaction commits
 222  */
 223 static int record_root_in_trans(struct btrfs_trans_handle *trans,
 224                                struct btrfs_root *root)
 225 {
 226         if (root->ref_cows && root->last_trans < trans->transid) {
 227                 WARN_ON(root == root->fs_info->extent_root);
 228                 WARN_ON(root->commit_root != root->node);
 229
 230                 /*
 231                  * see below for in_trans_setup usage rules
 232                  * we have the reloc mutex held now, so there
 233                  * is only one writer in this function
 234                  */
 235                 root->in_trans_setup = 1;
 236
 237                 /* make sure readers find in_trans_setup before
 238                  * they find our root->last_trans update
 239                  */
 240                 smp_wmb();
 241
 242                 spin_lock(&root->fs_info->fs_roots_radix_lock);
 243                 if (root->last_trans == trans->transid) {
 244                         spin_unlock(&root->fs_info->fs_roots_radix_lock);
 245                         return 0;
 246                 }
 247                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 248                            (unsigned long)root->root_key.objectid,
 249                            BTRFS_ROOT_TRANS_TAG);
 250                 spin_unlock(&root->fs_info->fs_roots_radix_lock);
 251                 root->last_trans = trans->transid;
 252
 253                 /* this is pretty tricky.  We don't want to
 254                  * take the relocation lock in btrfs_record_root_in_trans
 255                  * unless we're really doing the first setup for this root in
 256                  * this transaction.
 257                  *
 258                  * Normally we'd use root->last_trans as a flag to decide
 259                  * if we want to take the expensive mutex.
 260                  *
 261                  * But, we have to set root->last_trans before we
 262                  * init the relocation root, otherwise, we trip over warnings
 263                  * in ctree.c.  The solution used here is to flag ourselves
 264                  * with root->in_trans_setup.  When this is 1, we're still
 265                  * fixing up the reloc trees and everyone must wait.
 266                  *
 267                  * When this is zero, they can trust root->last_trans and fly
 268                  * through btrfs_record_root_in_trans without having to take the
 269                  * lock.  smp_wmb() makes sure that all the writes above are
 270                  * done before we pop in the zero below
 271                  */
 272                 btrfs_init_reloc_root(trans, root);
 273                 smp_wmb();
 274                 root->in_trans_setup = 0;
 275         }
 276         return 0;
 277 }
 278
 279
 280 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 281                                struct btrfs_root *root)
 282 {
 283         if (!root->ref_cows)
 284                 return 0;
 285
 286         /*
 287          * see record_root_in_trans for comments about in_trans_setup usage
 288          * and barriers
 289          */
 290         smp_rmb();
 291         if (root->last_trans == trans->transid &&
 292             !root->in_trans_setup)
 293                 return 0;
 294
 295         mutex_lock(&root->fs_info->reloc_mutex);
 296         record_root_in_trans(trans, root);
 297         mutex_unlock(&root->fs_info->reloc_mutex);
 298
 299         return 0;
 300 }
 301
 302 static inline int is_transaction_blocked(struct btrfs_transaction *trans)
 303 {
 304         return (trans->state >= TRANS_STATE_BLOCKED &&
 305                 trans->state < TRANS_STATE_UNBLOCKED);
 306 }
 307
 308 /* wait for commit against the current transaction to become unblocked
 309  * when this is done, it is safe to start a new transaction, but the current
 310  * transaction might not be fully on disk.
 311  */
 312 static void wait_current_trans(struct btrfs_root *root)
 313 {
 314         struct btrfs_transaction *cur_trans;
 315
 316         spin_lock(&root->fs_info->trans_lock);
 317         cur_trans = root->fs_info->running_transaction;
 318         if (cur_trans && is_transaction_blocked(cur_trans)) {
 319                 atomic_inc(&cur_trans->use_count);
 320                 spin_unlock(&root->fs_info->trans_lock);
 321
 322                 wait_event(root->fs_info->transaction_wait,
 323                            cur_trans->state >= TRANS_STATE_UNBLOCKED);
 324                 put_transaction(cur_trans);
 325         } else {
 326                 spin_unlock(&root->fs_info->trans_lock);
 327         }
 328 }
 329
 330 static int may_wait_transaction(struct btrfs_root *root, int type)
 331 {
 332         if (root->fs_info->log_root_recovering)
 333                 return 0;
 334
 335         if (type == TRANS_USERSPACE)
 336                 return 1;
 337
 338         if (type == TRANS_START &&
 339             !atomic_read(&root->fs_info->open_ioctl_trans))
 340                 return 1;
 341
 342         return 0;
 343 }
 344
 345 static struct btrfs_trans_handle *
 346 start_transaction(struct btrfs_root *root, u64 num_items, unsigned int type,
 347                   enum btrfs_reserve_flush_enum flush)
 348 {
 349         struct btrfs_trans_handle *h;
 350         struct btrfs_transaction *cur_trans;
 351         u64 num_bytes = 0;
 352         int ret;
 353         u64 qgroup_reserved = 0;
 354
 355         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
 356                 return ERR_PTR(-EROFS);
 357
 358         if (current->journal_info) {
 359                 WARN_ON(type & TRANS_EXTWRITERS);
 360                 h = current->journal_info;
 361                 h->use_count++;
 362                 WARN_ON(h->use_count > 2);
 363                 h->orig_rsv = h->block_rsv;
 364                 h->block_rsv = NULL;
 365                 goto got_it;
 366         }
 367
 368         /*
 369          * Do the reservation before we join the transaction so we can do all
 370          * the appropriate flushing if need be.
 371          */
 372         if (num_items > 0 && root != root->fs_info->chunk_root) {
 373                 if (root->fs_info->quota_enabled &&
 374                     is_fstree(root->root_key.objectid)) {
 375                         qgroup_reserved = num_items * root->leafsize;
 376                         ret = btrfs_qgroup_reserve(root, qgroup_reserved);
 377                         if (ret)
 378                                 return ERR_PTR(ret);
 379                 }
 380
 381                 num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
 382                 ret = btrfs_block_rsv_add(root,
 383                                           &root->fs_info->trans_block_rsv,
 384                                           num_bytes, flush);
 385                 if (ret)
 386                         goto reserve_fail;
 387         }
 388 again:
 389         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 390         if (!h) {
 391                 ret = -ENOMEM;
 392                 goto alloc_fail;
 393         }
 394
 395         /*
 396          * If we are JOIN_NOLOCK we're already committing a transaction and
 397          * waiting on this guy, so we don't need to do the sb_start_intwrite
 398          * because we're already holding a ref.  We need this because we could
 399          * have raced in and did an fsync() on a file which can kick a commit
 400          * and then we deadlock with somebody doing a freeze.
 401          *
 402          * If we are ATTACH, it means we just want to catch the current
 403          * transaction and commit it, so we needn't do sb_start_intwrite().
 404          */
 405         if (type & __TRANS_FREEZABLE)
 406                 sb_start_intwrite(root->fs_info->sb);
 407
 408         if (may_wait_transaction(root, type))
 409                 wait_current_trans(root);
 410
 411         do {
 412                 ret = join_transaction(root, type);
 413                 if (ret == -EBUSY) {
 414                         wait_current_trans(root);
 415                         if (unlikely(type == TRANS_ATTACH))
 416                                 ret = -ENOENT;
 417                 }
 418         } while (ret == -EBUSY);
 419
 420         if (ret < 0) {
 421                 /* We must get the transaction if we are JOIN_NOLOCK. */
 422                 BUG_ON(type == TRANS_JOIN_NOLOCK);
 423                 goto join_fail;
 424         }
 425
 426         cur_trans = root->fs_info->running_transaction;
 427
 428         h->transid = cur_trans->transid;
 429         h->transaction = cur_trans;
 430         h->blocks_used = 0;
 431         h->bytes_reserved = 0;
 432         h->root = root;
 433         h->delayed_ref_updates = 0;
 434         h->use_count = 1;
 435         h->adding_csums = 0;
 436         h->block_rsv = NULL;
 437         h->orig_rsv = NULL;
 438         h->aborted = 0;
 439         h->qgroup_reserved = 0;
 440         h->delayed_ref_elem.seq = 0;
 441         h->type = type;
 442         h->allocating_chunk = false;
 443         INIT_LIST_HEAD(&h->qgroup_ref_list);
 444         INIT_LIST_HEAD(&h->new_bgs);
 445
 446         smp_mb();
 447         if (cur_trans->state >= TRANS_STATE_BLOCKED &&
 448             may_wait_transaction(root, type)) {
 449                 btrfs_commit_transaction(h, root);
 450                 goto again;
 451         }
 452
 453         if (num_bytes) {
 454                 trace_btrfs_space_reservation(root->fs_info, "transaction",
 455                                               h->transid, num_bytes, 1);
 456                 h->block_rsv = &root->fs_info->trans_block_rsv;
 457                 h->bytes_reserved = num_bytes;
 458         }
 459         h->qgroup_reserved = qgroup_reserved;
 460
 461 got_it:
 462         btrfs_record_root_in_trans(h, root);
 463
 464         if (!current->journal_info && type != TRANS_USERSPACE)
 465                 current->journal_info = h;
 466         return h;
 467
 468 join_fail:
 469         if (type & __TRANS_FREEZABLE)
 470                 sb_end_intwrite(root->fs_info->sb);
 471         kmem_cache_free(btrfs_trans_handle_cachep, h);
 472 alloc_fail:
 473         if (num_bytes)
 474                 btrfs_block_rsv_release(root, &root->fs_info->trans_block_rsv,
 475                                         num_bytes);
 476 reserve_fail:
 477         if (qgroup_reserved)
 478                 btrfs_qgroup_free(root, qgroup_reserved);
 479         return ERR_PTR(ret);
 480 }
 481
 482 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 483                                                    int num_items)
 484 {
 485         return start_transaction(root, num_items, TRANS_START,
 486                                  BTRFS_RESERVE_FLUSH_ALL);
 487 }
 488
 489 struct btrfs_trans_handle *btrfs_start_transaction_lflush(
 490                                         struct btrfs_root *root, int num_items)
 491 {
 492         return start_transaction(root, num_items, TRANS_START,
 493                                  BTRFS_RESERVE_FLUSH_LIMIT);
 494 }
 495
 496 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 497 {
 498         return start_transaction(root, 0, TRANS_JOIN, 0);
 499 }
 500
 501 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 502 {
 503         return start_transaction(root, 0, TRANS_JOIN_NOLOCK, 0);
 504 }
 505
 506 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 507 {
 508         return start_transaction(root, 0, TRANS_USERSPACE, 0);
 509 }
 510
 511 /*
 512  * btrfs_attach_transaction() - catch the running transaction
 513  *
 514  * It is used when we want to commit the current the transaction, but
 515  * don't want to start a new one.
 516  *
 517  * Note: If this function return -ENOENT, it just means there is no
 518  * running transaction. But it is possible that the inactive transaction
 519  * is still in the memory, not fully on disk. If you hope there is no
 520  * inactive transaction in the fs when -ENOENT is returned, you should
 521  * invoke
 522  *     btrfs_attach_transaction_barrier()
 523  */
 524 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
 525 {
 526         return start_transaction(root, 0, TRANS_ATTACH, 0);
 527 }
 528
 529 /*
 530  * btrfs_attach_transaction() - catch the running transaction
 531  *
 532  * It is similar to the above function, the differentia is this one
 533  * will wait for all the inactive transactions until they fully
 534  * complete.
 535  */
 536 struct btrfs_trans_handle *
 537 btrfs_attach_transaction_barrier(struct btrfs_root *root)
 538 {
 539         struct btrfs_trans_handle *trans;
 540
 541         trans = start_transaction(root, 0, TRANS_ATTACH, 0);
 542         if (IS_ERR(trans) && PTR_ERR(trans) == -ENOENT)
 543                 btrfs_wait_for_commit(root, 0);
 544
 545         return trans;
 546 }
 547
 548 /* wait for a transaction commit to be fully complete */
 549 static noinline void wait_for_commit(struct btrfs_root *root,
 550                                     struct btrfs_transaction *commit)
 551 {
 552         wait_event(commit->commit_wait, commit->state == TRANS_STATE_COMPLETED);
 553 }
 554
 555 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 556 {
 557         struct btrfs_transaction *cur_trans = NULL, *t;
 558         int ret = 0;
 559
 560         if (transid) {
 561                 if (transid <= root->fs_info->last_trans_committed)
 562                         goto out;
 563
 564                 ret = -EINVAL;
 565                 /* find specified transaction */
 566                 spin_lock(&root->fs_info->trans_lock);
 567                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 568                         if (t->transid == transid) {
 569                                 cur_trans = t;
 570                                 atomic_inc(&cur_trans->use_count);
 571                                 ret = 0;
 572                                 break;
 573                         }
 574                         if (t->transid > transid) {
 575                                 ret = 0;
 576                                 break;
 577                         }
 578                 }
 579                 spin_unlock(&root->fs_info->trans_lock);
 580                 /* The specified transaction doesn't exist */
 581                 if (!cur_trans)
 582                         goto out;
 583         } else {
 584                 /* find newest transaction that is committing | committed */
 585                 spin_lock(&root->fs_info->trans_lock);
 586                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 587                                             list) {
 588                         if (t->state >= TRANS_STATE_COMMIT_START) {
 589                                 if (t->state == TRANS_STATE_COMPLETED)
 590                                         break;
 591                                 cur_trans = t;
 592                                 atomic_inc(&cur_trans->use_count);
 593                                 break;
 594                         }
 595                 }
 596                 spin_unlock(&root->fs_info->trans_lock);
 597                 if (!cur_trans)
 598                         goto out;  /* nothing committing|committed */
 599         }
 600
 601         wait_for_commit(root, cur_trans);
 602         put_transaction(cur_trans);
 603 out:
 604         return ret;
 605 }
 606
 607 void btrfs_throttle(struct btrfs_root *root)
 608 {
 609         if (!atomic_read(&root->fs_info->open_ioctl_trans))
 610                 wait_current_trans(root);
 611 }
 612
 613 static int should_end_transaction(struct btrfs_trans_handle *trans,
 614                                   struct btrfs_root *root)
 615 {
 616         int ret;
 617
 618         ret = btrfs_block_rsv_check(root, &root->fs_info->global_block_rsv, 5);
 619         return ret ? 1 : 0;
 620 }
 621
 622 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 623                                  struct btrfs_root *root)
 624 {
 625         struct btrfs_transaction *cur_trans = trans->transaction;
 626         int updates;
 627         int err;
 628
 629         smp_mb();
 630         if (cur_trans->state >= TRANS_STATE_BLOCKED ||
 631             cur_trans->delayed_refs.flushing)
 632                 return 1;
 633
 634         updates = trans->delayed_ref_updates;
 635         trans->delayed_ref_updates = 0;
 636         if (updates) {
 637                 err = btrfs_run_delayed_refs(trans, root, updates);
 638                 if (err) /* Error code will also eval true */
 639                         return err;
 640         }
 641
 642         return should_end_transaction(trans, root);
 643 }
 644
 645 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 646                           struct btrfs_root *root, int throttle)
 647 {
 648         struct btrfs_transaction *cur_trans = trans->transaction;
 649         struct btrfs_fs_info *info = root->fs_info;
 650         int count = 0;
 651         int lock = (trans->type != TRANS_JOIN_NOLOCK);
 652         int err = 0;
 653
 654         if (--trans->use_count) {
 655                 trans->block_rsv = trans->orig_rsv;
 656                 return 0;
 657         }
 658
 659         /*
 660          * do the qgroup accounting as early as possible
 661          */
 662         err = btrfs_delayed_refs_qgroup_accounting(trans, info);
 663
 664         btrfs_trans_release_metadata(trans, root);
 665         trans->block_rsv = NULL;
 666
 667         if (trans->qgroup_reserved) {
 668                 /*
 669                  * the same root has to be passed here between start_transaction
 670                  * and end_transaction. Subvolume quota depends on this.
 671                  */
 672                 btrfs_qgroup_free(trans->root, trans->qgroup_reserved);
 673                 trans->qgroup_reserved = 0;
 674         }
 675
 676         if (!list_empty(&trans->new_bgs))
 677                 btrfs_create_pending_block_groups(trans, root);
 678
 679         while (count < 1) {
 680                 unsigned long cur = trans->delayed_ref_updates;
 681                 trans->delayed_ref_updates = 0;
 682                 if (cur &&
 683                     trans->transaction->delayed_refs.num_heads_ready > 64) {
 684                         trans->delayed_ref_updates = 0;
 685                         btrfs_run_delayed_refs(trans, root, cur);
 686                 } else {
 687                         break;
 688                 }
 689                 count++;
 690         }
 691
 692         btrfs_trans_release_metadata(trans, root);
 693         trans->block_rsv = NULL;
 694
 695         if (!list_empty(&trans->new_bgs))
 696                 btrfs_create_pending_block_groups(trans, root);
 697
 698         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
 699             should_end_transaction(trans, root) &&
 700             ACCESS_ONCE(cur_trans->state) == TRANS_STATE_RUNNING) {
 701                 spin_lock(&info->trans_lock);
 702                 if (cur_trans->state == TRANS_STATE_RUNNING)
 703                         cur_trans->state = TRANS_STATE_BLOCKED;
 704                 spin_unlock(&info->trans_lock);
 705         }
 706
 707         if (lock && ACCESS_ONCE(cur_trans->state) == TRANS_STATE_BLOCKED) {
 708                 if (throttle) {
 709                         /*
 710                          * We may race with somebody else here so end up having
 711                          * to call end_transaction on ourselves again, so inc
 712                          * our use_count.
 713                          */
 714                         trans->use_count++;
 715                         return btrfs_commit_transaction(trans, root);
 716                 } else {
 717                         wake_up_process(info->transaction_kthread);
 718                 }
 719         }
 720
 721         if (trans->type & __TRANS_FREEZABLE)
 722                 sb_end_intwrite(root->fs_info->sb);
 723
 724         WARN_ON(cur_trans != info->running_transaction);
 725         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 726         atomic_dec(&cur_trans->num_writers);
 727         extwriter_counter_dec(cur_trans, trans->type);
 728
 729         smp_mb();
 730         if (waitqueue_active(&cur_trans->writer_wait))
 731                 wake_up(&cur_trans->writer_wait);
 732         put_transaction(cur_trans);
 733
 734         if (current->journal_info == trans)
 735                 current->journal_info = NULL;
 736
 737         if (throttle)
 738                 btrfs_run_delayed_iputs(root);
 739
 740         if (trans->aborted ||
 741             test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
 742                 err = -EIO;
 743         assert_qgroups_uptodate(trans);
 744
 745         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 746         return err;
 747 }
 748
 749 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 750                           struct btrfs_root *root)
 751 {
 752         return __btrfs_end_transaction(trans, root, 0);
 753 }
 754
 755 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 756                                    struct btrfs_root *root)
 757 {
 758         return __btrfs_end_transaction(trans, root, 1);
 759 }
 760
 761 int btrfs_end_transaction_dmeta(struct btrfs_trans_handle *trans,
 762                                 struct btrfs_root *root)
 763 {
 764         return __btrfs_end_transaction(trans, root, 1);
 765 }
 766
 767 /*
 768  * when btree blocks are allocated, they have some corresponding bits set for
 769  * them in one of two extent_io trees.  This is used to make sure all of
 770  * those extents are sent to disk but does not wait on them
 771  */
 772 int btrfs_write_marked_extents(struct btrfs_root *root,
 773                                struct extent_io_tree *dirty_pages, int mark)
 774 {
 775         int err = 0;
 776         int werr = 0;
 777         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 778         struct extent_state *cached_state = NULL;
 779         u64 start = 0;
 780         u64 end;
 781         struct blk_plug plug;
 782
 783         blk_start_plug(&plug);
 784         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 785                                       mark, &cached_state)) {
 786                 convert_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT,
 787                                    mark, &cached_state, GFP_NOFS);
 788                 cached_state = NULL;
 789                 err = filemap_fdatawrite_range(mapping, start, end);
 790                 if (err)
 791                         werr = err;
 792                 cond_resched();
 793                 start = end + 1;
 794         }
 795         if (err)
 796                 werr = err;
 797         blk_finish_plug(&plug);
 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.  We wait
 805  * on all the pages and clear them from the dirty pages state tree
 806  */
 807 int btrfs_wait_marked_extents(struct btrfs_root *root,
 808                               struct extent_io_tree *dirty_pages, int mark)
 809 {
 810         int err = 0;
 811         int werr = 0;
 812         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 813         struct extent_state *cached_state = NULL;
 814         u64 start = 0;
 815         u64 end;
 816
 817         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 818                                       EXTENT_NEED_WAIT, &cached_state)) {
 819                 clear_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT,
 820                                  0, 0, &cached_state, GFP_NOFS);
 821                 err = filemap_fdatawait_range(mapping, start, end);
 822                 if (err)
 823                         werr = err;
 824                 cond_resched();
 825                 start = end + 1;
 826         }
 827         if (err)
 828                 werr = err;
 829         return werr;
 830 }
 831
 832 /*
 833  * when btree blocks are allocated, they have some corresponding bits set for
 834  * them in one of two extent_io trees.  This is used to make sure all of
 835  * those extents are on disk for transaction or log commit
 836  */
 837 int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 838                                 struct extent_io_tree *dirty_pages, int mark)
 839 {
 840         int ret;
 841         int ret2;
 842
 843         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 844         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 845
 846         if (ret)
 847                 return ret;
 848         if (ret2)
 849                 return ret2;
 850         return 0;
 851 }
 852
 853 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 854                                      struct btrfs_root *root)
 855 {
 856         if (!trans || !trans->transaction) {
 857                 struct inode *btree_inode;
 858                 btree_inode = root->fs_info->btree_inode;
 859                 return filemap_write_and_wait(btree_inode->i_mapping);
 860         }
 861         return btrfs_write_and_wait_marked_extents(root,
 862                                            &trans->transaction->dirty_pages,
 863                                            EXTENT_DIRTY);
 864 }
 865
 866 /*
 867  * this is used to update the root pointer in the tree of tree roots.
 868  *
 869  * But, in the case of the extent allocation tree, updating the root
 870  * pointer may allocate blocks which may change the root of the extent
 871  * allocation tree.
 872  *
 873  * So, this loops and repeats and makes sure the cowonly root didn't
 874  * change while the root pointer was being updated in the metadata.
 875  */
 876 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 877                                struct btrfs_root *root)
 878 {
 879         int ret;
 880         u64 old_root_bytenr;
 881         u64 old_root_used;
 882         struct btrfs_root *tree_root = root->fs_info->tree_root;
 883
 884         old_root_used = btrfs_root_used(&root->root_item);
 885         btrfs_write_dirty_block_groups(trans, root);
 886
 887         while (1) {
 888                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 889                 if (old_root_bytenr == root->node->start &&
 890                     old_root_used == btrfs_root_used(&root->root_item))
 891                         break;
 892
 893                 btrfs_set_root_node(&root->root_item, root->node);
 894                 ret = btrfs_update_root(trans, tree_root,
 895                                         &root->root_key,
 896                                         &root->root_item);
 897                 if (ret)
 898                         return ret;
 899
 900                 old_root_used = btrfs_root_used(&root->root_item);
 901                 ret = btrfs_write_dirty_block_groups(trans, root);
 902                 if (ret)
 903                         return ret;
 904         }
 905
 906         if (root != root->fs_info->extent_root)
 907                 switch_commit_root(root);
 908
 909         return 0;
 910 }
 911
 912 /*
 913  * update all the cowonly tree roots on disk
 914  *
 915  * The error handling in this function may not be obvious. Any of the
 916  * failures will cause the file system to go offline. We still need
 917  * to clean up the delayed refs.
 918  */
 919 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 920                                          struct btrfs_root *root)
 921 {
 922         struct btrfs_fs_info *fs_info = root->fs_info;
 923         struct list_head *next;
 924         struct extent_buffer *eb;
 925         int ret;
 926
 927         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 928         if (ret)
 929                 return ret;
 930
 931         eb = btrfs_lock_root_node(fs_info->tree_root);
 932         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
 933                               0, &eb);
 934         btrfs_tree_unlock(eb);
 935         free_extent_buffer(eb);
 936
 937         if (ret)
 938                 return ret;
 939
 940         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 941         if (ret)
 942                 return ret;
 943
 944         ret = btrfs_run_dev_stats(trans, root->fs_info);
 945         WARN_ON(ret);
 946         ret = btrfs_run_dev_replace(trans, root->fs_info);
 947         WARN_ON(ret);
 948
 949         ret = btrfs_run_qgroups(trans, root->fs_info);
 950         BUG_ON(ret);
 951
 952         /* run_qgroups might have added some more refs */
 953         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 954         BUG_ON(ret);
 955
 956         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 957                 next = fs_info->dirty_cowonly_roots.next;
 958                 list_del_init(next);
 959                 root = list_entry(next, struct btrfs_root, dirty_list);
 960
 961                 ret = update_cowonly_root(trans, root);
 962                 if (ret)
 963                         return ret;
 964         }
 965
 966         down_write(&fs_info->extent_commit_sem);
 967         switch_commit_root(fs_info->extent_root);
 968         up_write(&fs_info->extent_commit_sem);
 969
 970         btrfs_after_dev_replace_commit(fs_info);
 971
 972         return 0;
 973 }
 974
 975 /*
 976  * dead roots are old snapshots that need to be deleted.  This allocates
 977  * a dirty root struct and adds it into the list of dead roots that need to
 978  * be deleted
 979  */
 980 int btrfs_add_dead_root(struct btrfs_root *root)
 981 {
 982         spin_lock(&root->fs_info->trans_lock);
 983         list_add_tail(&root->root_list, &root->fs_info->dead_roots);
 984         spin_unlock(&root->fs_info->trans_lock);
 985         return 0;
 986 }
 987
 988 /*
 989  * update all the cowonly tree roots on disk
 990  */
 991 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 992                                     struct btrfs_root *root)
 993 {
 994         struct btrfs_root *gang[8];
 995         struct btrfs_fs_info *fs_info = root->fs_info;
 996         int i;
 997         int ret;
 998         int err = 0;
 999
1000         spin_lock(&fs_info->fs_roots_radix_lock);
1001         while (1) {
1002                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1003                                                  (void **)gang, 0,
1004                                                  ARRAY_SIZE(gang),
1005                                                  BTRFS_ROOT_TRANS_TAG);
1006                 if (ret == 0)
1007                         break;
1008                 for (i = 0; i < ret; i++) {
1009                         root = gang[i];
1010                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
1011                                         (unsigned long)root->root_key.objectid,
1012                                         BTRFS_ROOT_TRANS_TAG);
1013                         spin_unlock(&fs_info->fs_roots_radix_lock);
1014
1015                         btrfs_free_log(trans, root);
1016                         btrfs_update_reloc_root(trans, root);
1017                         btrfs_orphan_commit_root(trans, root);
1018
1019                         btrfs_save_ino_cache(root, trans);
1020
1021                         /* see comments in should_cow_block() */
1022                         root->force_cow = 0;
1023                         smp_wmb();
1024
1025                         if (root->commit_root != root->node) {
1026                                 mutex_lock(&root->fs_commit_mutex);
1027                                 switch_commit_root(root);
1028                                 btrfs_unpin_free_ino(root);
1029                                 mutex_unlock(&root->fs_commit_mutex);
1030
1031                                 btrfs_set_root_node(&root->root_item,
1032                                                     root->node);
1033                         }
1034
1035                         err = btrfs_update_root(trans, fs_info->tree_root,
1036                                                 &root->root_key,
1037                                                 &root->root_item);
1038                         spin_lock(&fs_info->fs_roots_radix_lock);
1039                         if (err)
1040                                 break;
1041                 }
1042         }
1043         spin_unlock(&fs_info->fs_roots_radix_lock);
1044         return err;
1045 }
1046
1047 /*
1048  * defrag a given btree.
1049  * Every leaf in the btree is read and defragged.
1050  */
1051 int btrfs_defrag_root(struct btrfs_root *root)
1052 {
1053         struct btrfs_fs_info *info = root->fs_info;
1054         struct btrfs_trans_handle *trans;
1055         int ret;
1056
1057         if (xchg(&root->defrag_running, 1))
1058                 return 0;
1059
1060         while (1) {
1061                 trans = btrfs_start_transaction(root, 0);
1062                 if (IS_ERR(trans))
1063                         return PTR_ERR(trans);
1064
1065                 ret = btrfs_defrag_leaves(trans, root);
1066
1067                 btrfs_end_transaction(trans, root);
1068                 btrfs_btree_balance_dirty(info->tree_root);
1069                 cond_resched();
1070
1071                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
1072                         break;
1073
1074                 if (btrfs_defrag_cancelled(root->fs_info)) {
1075                         printk(KERN_DEBUG "btrfs: defrag_root cancelled\n");
1076                         ret = -EAGAIN;
1077                         break;
1078                 }
1079         }
1080         root->defrag_running = 0;
1081         return ret;
1082 }
1083
1084 /*
1085  * new snapshots need to be created at a very specific time in the
1086  * transaction commit.  This does the actual creation.
1087  *
1088  * Note:
1089  * If the error which may affect the commitment of the current transaction
1090  * happens, we should return the error number. If the error which just affect
1091  * the creation of the pending snapshots, just return 0.
1092  */
1093 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1094                                    struct btrfs_fs_info *fs_info,
1095                                    struct btrfs_pending_snapshot *pending)
1096 {
1097         struct btrfs_key key;
1098         struct btrfs_root_item *new_root_item;
1099         struct btrfs_root *tree_root = fs_info->tree_root;
1100         struct btrfs_root *root = pending->root;
1101         struct btrfs_root *parent_root;
1102         struct btrfs_block_rsv *rsv;
1103         struct inode *parent_inode;
1104         struct btrfs_path *path;
1105         struct btrfs_dir_item *dir_item;
1106         struct dentry *dentry;
1107         struct extent_buffer *tmp;
1108         struct extent_buffer *old;
1109         struct timespec cur_time = CURRENT_TIME;
1110         int ret = 0;
1111         u64 to_reserve = 0;
1112         u64 index = 0;
1113         u64 objectid;
1114         u64 root_flags;
1115         uuid_le new_uuid;
1116
1117         path = btrfs_alloc_path();
1118         if (!path) {
1119                 pending->error = -ENOMEM;
1120                 return 0;
1121         }
1122
1123         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
1124         if (!new_root_item) {
1125                 pending->error = -ENOMEM;
1126                 goto root_item_alloc_fail;
1127         }
1128
1129         pending->error = btrfs_find_free_objectid(tree_root, &objectid);
1130         if (pending->error)
1131                 goto no_free_objectid;
1132
1133         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
1134
1135         if (to_reserve > 0) {
1136                 pending->error = btrfs_block_rsv_add(root,
1137                                                      &pending->block_rsv,
1138                                                      to_reserve,
1139                                                      BTRFS_RESERVE_NO_FLUSH);
1140                 if (pending->error)
1141                         goto no_free_objectid;
1142         }
1143
1144         pending->error = btrfs_qgroup_inherit(trans, fs_info,
1145                                               root->root_key.objectid,
1146                                               objectid, pending->inherit);
1147         if (pending->error)
1148                 goto no_free_objectid;
1149
1150         key.objectid = objectid;
1151         key.offset = (u64)-1;
1152         key.type = BTRFS_ROOT_ITEM_KEY;
1153
1154         rsv = trans->block_rsv;
1155         trans->block_rsv = &pending->block_rsv;
1156         trans->bytes_reserved = trans->block_rsv->reserved;
1157
1158         dentry = pending->dentry;
1159         parent_inode = pending->dir;
1160         parent_root = BTRFS_I(parent_inode)->root;
1161         record_root_in_trans(trans, parent_root);
1162
1163         /*
1164          * insert the directory item
1165          */
1166         ret = btrfs_set_inode_index(parent_inode, &index);
1167         BUG_ON(ret); /* -ENOMEM */
1168
1169         /* check if there is a file/dir which has the same name. */
1170         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1171                                          btrfs_ino(parent_inode),
1172                                          dentry->d_name.name,
1173                                          dentry->d_name.len, 0);
1174         if (dir_item != NULL && !IS_ERR(dir_item)) {
1175                 pending->error = -EEXIST;
1176                 goto dir_item_existed;
1177         } else if (IS_ERR(dir_item)) {
1178                 ret = PTR_ERR(dir_item);
1179                 btrfs_abort_transaction(trans, root, ret);
1180                 goto fail;
1181         }
1182         btrfs_release_path(path);
1183
1184         /*
1185          * pull in the delayed directory update
1186          * and the delayed inode item
1187          * otherwise we corrupt the FS during
1188          * snapshot
1189          */
1190         ret = btrfs_run_delayed_items(trans, root);
1191         if (ret) {      /* Transaction aborted */
1192                 btrfs_abort_transaction(trans, root, ret);
1193                 goto fail;
1194         }
1195
1196         record_root_in_trans(trans, root);
1197         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1198         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1199         btrfs_check_and_init_root_item(new_root_item);
1200
1201         root_flags = btrfs_root_flags(new_root_item);
1202         if (pending->readonly)
1203                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1204         else
1205                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1206         btrfs_set_root_flags(new_root_item, root_flags);
1207
1208         btrfs_set_root_generation_v2(new_root_item,
1209                         trans->transid);
1210         uuid_le_gen(&new_uuid);
1211         memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1212         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1213                         BTRFS_UUID_SIZE);
1214         if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1215                 memset(new_root_item->received_uuid, 0,
1216                        sizeof(new_root_item->received_uuid));
1217                 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1218                 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1219                 btrfs_set_root_stransid(new_root_item, 0);
1220                 btrfs_set_root_rtransid(new_root_item, 0);
1221         }
1222         new_root_item->otime.sec = cpu_to_le64(cur_time.tv_sec);
1223         new_root_item->otime.nsec = cpu_to_le32(cur_time.tv_nsec);
1224         btrfs_set_root_otransid(new_root_item, trans->transid);
1225
1226         old = btrfs_lock_root_node(root);
1227         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1228         if (ret) {
1229                 btrfs_tree_unlock(old);
1230                 free_extent_buffer(old);
1231                 btrfs_abort_transaction(trans, root, ret);
1232                 goto fail;
1233         }
1234
1235         btrfs_set_lock_blocking(old);
1236
1237         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1238         /* clean up in any case */
1239         btrfs_tree_unlock(old);
1240         free_extent_buffer(old);
1241         if (ret) {
1242                 btrfs_abort_transaction(trans, root, ret);
1243                 goto fail;
1244         }
1245
1246         /* see comments in should_cow_block() */
1247         root->force_cow = 1;
1248         smp_wmb();
1249
1250         btrfs_set_root_node(new_root_item, tmp);
1251         /* record when the snapshot was created in key.offset */
1252         key.offset = trans->transid;
1253         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1254         btrfs_tree_unlock(tmp);
1255         free_extent_buffer(tmp);
1256         if (ret) {
1257                 btrfs_abort_transaction(trans, root, ret);
1258                 goto fail;
1259         }
1260
1261         /*
1262          * insert root back/forward references
1263          */
1264         ret = btrfs_add_root_ref(trans, tree_root, objectid,
1265                                  parent_root->root_key.objectid,
1266                                  btrfs_ino(parent_inode), index,
1267                                  dentry->d_name.name, dentry->d_name.len);
1268         if (ret) {
1269                 btrfs_abort_transaction(trans, root, ret);
1270                 goto fail;
1271         }
1272
1273         key.offset = (u64)-1;
1274         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1275         if (IS_ERR(pending->snap)) {
1276                 ret = PTR_ERR(pending->snap);
1277                 btrfs_abort_transaction(trans, root, ret);
1278                 goto fail;
1279         }
1280
1281         ret = btrfs_reloc_post_snapshot(trans, pending);
1282         if (ret) {
1283                 btrfs_abort_transaction(trans, root, ret);
1284                 goto fail;
1285         }
1286
1287         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1288         if (ret) {
1289                 btrfs_abort_transaction(trans, root, ret);
1290                 goto fail;
1291         }
1292
1293         ret = btrfs_insert_dir_item(trans, parent_root,
1294                                     dentry->d_name.name, dentry->d_name.len,
1295                                     parent_inode, &key,
1296                                     BTRFS_FT_DIR, index);
1297         /* We have check then name at the beginning, so it is impossible. */
1298         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1299         if (ret) {
1300                 btrfs_abort_transaction(trans, root, ret);
1301                 goto fail;
1302         }
1303
1304         btrfs_i_size_write(parent_inode, parent_inode->i_size +
1305                                          dentry->d_name.len * 2);
1306         parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
1307         ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1308         if (ret)
1309                 btrfs_abort_transaction(trans, root, ret);
1310 fail:
1311         pending->error = ret;
1312 dir_item_existed:
1313         trans->block_rsv = rsv;
1314         trans->bytes_reserved = 0;
1315 no_free_objectid:
1316         kfree(new_root_item);
1317 root_item_alloc_fail:
1318         btrfs_free_path(path);
1319         return ret;
1320 }
1321
1322 /*
1323  * create all the snapshots we've scheduled for creation
1324  */
1325 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1326                                              struct btrfs_fs_info *fs_info)
1327 {
1328         struct btrfs_pending_snapshot *pending, *next;
1329         struct list_head *head = &trans->transaction->pending_snapshots;
1330         int ret = 0;
1331
1332         list_for_each_entry_safe(pending, next, head, list) {
1333                 list_del(&pending->list);
1334                 ret = create_pending_snapshot(trans, fs_info, pending);
1335                 if (ret)
1336                         break;
1337         }
1338         return ret;
1339 }
1340
1341 static void update_super_roots(struct btrfs_root *root)
1342 {
1343         struct btrfs_root_item *root_item;
1344         struct btrfs_super_block *super;
1345
1346         super = root->fs_info->super_copy;
1347
1348         root_item = &root->fs_info->chunk_root->root_item;
1349         super->chunk_root = root_item->bytenr;
1350         super->chunk_root_generation = root_item->generation;
1351         super->chunk_root_level = root_item->level;
1352
1353         root_item = &root->fs_info->tree_root->root_item;
1354         super->root = root_item->bytenr;
1355         super->generation = root_item->generation;
1356         super->root_level = root_item->level;
1357         if (btrfs_test_opt(root, SPACE_CACHE))
1358                 super->cache_generation = root_item->generation;
1359 }
1360
1361 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1362 {
1363         struct btrfs_transaction *trans;
1364         int ret = 0;
1365
1366         spin_lock(&info->trans_lock);
1367         trans = info->running_transaction;
1368         if (trans)
1369                 ret = (trans->state >= TRANS_STATE_COMMIT_START);
1370         spin_unlock(&info->trans_lock);
1371         return ret;
1372 }
1373
1374 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1375 {
1376         struct btrfs_transaction *trans;
1377         int ret = 0;
1378
1379         spin_lock(&info->trans_lock);
1380         trans = info->running_transaction;
1381         if (trans)
1382                 ret = is_transaction_blocked(trans);
1383         spin_unlock(&info->trans_lock);
1384         return ret;
1385 }
1386
1387 /*
1388  * wait for the current transaction commit to start and block subsequent
1389  * transaction joins
1390  */
1391 static void wait_current_trans_commit_start(struct btrfs_root *root,
1392                                             struct btrfs_transaction *trans)
1393 {
1394         wait_event(root->fs_info->transaction_blocked_wait,
1395                    trans->state >= TRANS_STATE_COMMIT_START);
1396 }
1397
1398 /*
1399  * wait for the current transaction to start and then become unblocked.
1400  * caller holds ref.
1401  */
1402 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1403                                          struct btrfs_transaction *trans)
1404 {
1405         wait_event(root->fs_info->transaction_wait,
1406                    trans->state >= TRANS_STATE_UNBLOCKED);
1407 }
1408
1409 /*
1410  * commit transactions asynchronously. once btrfs_commit_transaction_async
1411  * returns, any subsequent transaction will not be allowed to join.
1412  */
1413 struct btrfs_async_commit {
1414         struct btrfs_trans_handle *newtrans;
1415         struct btrfs_root *root;
1416         struct work_struct work;
1417 };
1418
1419 static void do_async_commit(struct work_struct *work)
1420 {
1421         struct btrfs_async_commit *ac =
1422                 container_of(work, struct btrfs_async_commit, work);
1423
1424         /*
1425          * We've got freeze protection passed with the transaction.
1426          * Tell lockdep about it.
1427          */
1428         if (ac->newtrans->type < TRANS_JOIN_NOLOCK)
1429                 rwsem_acquire_read(
1430                      &ac->root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
1431                      0, 1, _THIS_IP_);
1432
1433         current->journal_info = ac->newtrans;
1434
1435         btrfs_commit_transaction(ac->newtrans, ac->root);
1436         kfree(ac);
1437 }
1438
1439 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1440                                    struct btrfs_root *root,
1441                                    int wait_for_unblock)
1442 {
1443         struct btrfs_async_commit *ac;
1444         struct btrfs_transaction *cur_trans;
1445
1446         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1447         if (!ac)
1448                 return -ENOMEM;
1449
1450         INIT_WORK(&ac->work, do_async_commit);
1451         ac->root = root;
1452         ac->newtrans = btrfs_join_transaction(root);
1453         if (IS_ERR(ac->newtrans)) {
1454                 int err = PTR_ERR(ac->newtrans);
1455                 kfree(ac);
1456                 return err;
1457         }
1458
1459         /* take transaction reference */
1460         cur_trans = trans->transaction;
1461         atomic_inc(&cur_trans->use_count);
1462
1463         btrfs_end_transaction(trans, root);
1464
1465         /*
1466          * Tell lockdep we've released the freeze rwsem, since the
1467          * async commit thread will be the one to unlock it.
1468          */
1469         if (trans->type < TRANS_JOIN_NOLOCK)
1470                 rwsem_release(
1471                         &root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
1472                         1, _THIS_IP_);
1473
1474         schedule_work(&ac->work);
1475
1476         /* wait for transaction to start and unblock */
1477         if (wait_for_unblock)
1478                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1479         else
1480                 wait_current_trans_commit_start(root, cur_trans);
1481
1482         if (current->journal_info == trans)
1483                 current->journal_info = NULL;
1484
1485         put_transaction(cur_trans);
1486         return 0;
1487 }
1488
1489
1490 static void cleanup_transaction(struct btrfs_trans_handle *trans,
1491                                 struct btrfs_root *root, int err)
1492 {
1493         struct btrfs_transaction *cur_trans = trans->transaction;
1494         DEFINE_WAIT(wait);
1495
1496         WARN_ON(trans->use_count > 1);
1497
1498         btrfs_abort_transaction(trans, root, err);
1499
1500         spin_lock(&root->fs_info->trans_lock);
1501
1502         /*
1503          * If the transaction is removed from the list, it means this
1504          * transaction has been committed successfully, so it is impossible
1505          * to call the cleanup function.
1506          */
1507         BUG_ON(list_empty(&cur_trans->list));
1508
1509         list_del_init(&cur_trans->list);
1510         if (cur_trans == root->fs_info->running_transaction) {
1511                 cur_trans->state = TRANS_STATE_COMMIT_DOING;
1512                 spin_unlock(&root->fs_info->trans_lock);
1513                 wait_event(cur_trans->writer_wait,
1514                            atomic_read(&cur_trans->num_writers) == 1);
1515
1516                 spin_lock(&root->fs_info->trans_lock);
1517         }
1518         spin_unlock(&root->fs_info->trans_lock);
1519
1520         btrfs_cleanup_one_transaction(trans->transaction, root);
1521
1522         spin_lock(&root->fs_info->trans_lock);
1523         if (cur_trans == root->fs_info->running_transaction)
1524                 root->fs_info->running_transaction = NULL;
1525         spin_unlock(&root->fs_info->trans_lock);
1526
1527         put_transaction(cur_trans);
1528         put_transaction(cur_trans);
1529
1530         trace_btrfs_transaction_commit(root);
1531
1532         btrfs_scrub_continue(root);
1533
1534         if (current->journal_info == trans)
1535                 current->journal_info = NULL;
1536
1537         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1538 }
1539
1540 static int btrfs_flush_all_pending_stuffs(struct btrfs_trans_handle *trans,
1541                                           struct btrfs_root *root)
1542 {
1543         int ret;
1544
1545         ret = btrfs_run_delayed_items(trans, root);
1546         if (ret)
1547                 return ret;
1548
1549         /*
1550          * running the delayed items may have added new refs. account
1551          * them now so that they hinder processing of more delayed refs
1552          * as little as possible.
1553          */
1554         btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
1555
1556         /*
1557          * rename don't use btrfs_join_transaction, so, once we
1558          * set the transaction to blocked above, we aren't going
1559          * to get any new ordered operations.  We can safely run
1560          * it here and no for sure that nothing new will be added
1561          * to the list
1562          */
1563         ret = btrfs_run_ordered_operations(trans, root, 1);
1564
1565         return ret;
1566 }
1567
1568 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
1569 {
1570         if (btrfs_test_opt(fs_info->tree_root, FLUSHONCOMMIT))
1571                 return btrfs_start_all_delalloc_inodes(fs_info, 1);
1572         return 0;
1573 }
1574
1575 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
1576 {
1577         if (btrfs_test_opt(fs_info->tree_root, FLUSHONCOMMIT))
1578                 btrfs_wait_all_ordered_extents(fs_info, 1);
1579 }
1580
1581 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1582                              struct btrfs_root *root)
1583 {
1584         struct btrfs_transaction *cur_trans = trans->transaction;
1585         struct btrfs_transaction *prev_trans = NULL;
1586         int ret;
1587
1588         ret = btrfs_run_ordered_operations(trans, root, 0);
1589         if (ret) {
1590                 btrfs_abort_transaction(trans, root, ret);
1591                 btrfs_end_transaction(trans, root);
1592                 return ret;
1593         }
1594
1595         /* Stop the commit early if ->aborted is set */
1596         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1597                 ret = cur_trans->aborted;
1598                 btrfs_end_transaction(trans, root);
1599                 return ret;
1600         }
1601
1602         /* make a pass through all the delayed refs we have so far
1603          * any runnings procs may add more while we are here
1604          */
1605         ret = btrfs_run_delayed_refs(trans, root, 0);
1606         if (ret) {
1607                 btrfs_end_transaction(trans, root);
1608                 return ret;
1609         }
1610
1611         btrfs_trans_release_metadata(trans, root);
1612         trans->block_rsv = NULL;
1613         if (trans->qgroup_reserved) {
1614                 btrfs_qgroup_free(root, trans->qgroup_reserved);
1615                 trans->qgroup_reserved = 0;
1616         }
1617
1618         cur_trans = trans->transaction;
1619
1620         /*
1621          * set the flushing flag so procs in this transaction have to
1622          * start sending their work down.
1623          */
1624         cur_trans->delayed_refs.flushing = 1;
1625
1626         if (!list_empty(&trans->new_bgs))
1627                 btrfs_create_pending_block_groups(trans, root);
1628
1629         ret = btrfs_run_delayed_refs(trans, root, 0);
1630         if (ret) {
1631                 btrfs_end_transaction(trans, root);
1632                 return ret;
1633         }
1634
1635         spin_lock(&root->fs_info->trans_lock);
1636         if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
1637                 spin_unlock(&root->fs_info->trans_lock);
1638                 atomic_inc(&cur_trans->use_count);
1639                 ret = btrfs_end_transaction(trans, root);
1640
1641                 wait_for_commit(root, cur_trans);
1642
1643                 put_transaction(cur_trans);
1644
1645                 return ret;
1646         }
1647
1648         cur_trans->state = TRANS_STATE_COMMIT_START;
1649         wake_up(&root->fs_info->transaction_blocked_wait);
1650
1651         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1652                 prev_trans = list_entry(cur_trans->list.prev,
1653                                         struct btrfs_transaction, list);
1654                 if (prev_trans->state != TRANS_STATE_COMPLETED) {
1655                         atomic_inc(&prev_trans->use_count);
1656                         spin_unlock(&root->fs_info->trans_lock);
1657
1658                         wait_for_commit(root, prev_trans);
1659
1660                         put_transaction(prev_trans);
1661                 } else {
1662                         spin_unlock(&root->fs_info->trans_lock);
1663                 }
1664         } else {
1665                 spin_unlock(&root->fs_info->trans_lock);
1666         }
1667
1668         extwriter_counter_dec(cur_trans, trans->type);
1669
1670         ret = btrfs_start_delalloc_flush(root->fs_info);
1671         if (ret)
1672                 goto cleanup_transaction;
1673
1674         ret = btrfs_flush_all_pending_stuffs(trans, root);
1675         if (ret)
1676                 goto cleanup_transaction;
1677
1678         wait_event(cur_trans->writer_wait,
1679                    extwriter_counter_read(cur_trans) == 0);
1680
1681         /* some pending stuffs might be added after the previous flush. */
1682         ret = btrfs_flush_all_pending_stuffs(trans, root);
1683         if (ret)
1684                 goto cleanup_transaction;
1685
1686         btrfs_wait_delalloc_flush(root->fs_info);
1687         /*
1688          * Ok now we need to make sure to block out any other joins while we
1689          * commit the transaction.  We could have started a join before setting
1690          * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
1691          */
1692         spin_lock(&root->fs_info->trans_lock);
1693         cur_trans->state = TRANS_STATE_COMMIT_DOING;
1694         spin_unlock(&root->fs_info->trans_lock);
1695         wait_event(cur_trans->writer_wait,
1696                    atomic_read(&cur_trans->num_writers) == 1);
1697
1698         /* ->aborted might be set after the previous check, so check it */
1699         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1700                 ret = cur_trans->aborted;
1701                 goto cleanup_transaction;
1702         }
1703         /*
1704          * the reloc mutex makes sure that we stop
1705          * the balancing code from coming in and moving
1706          * extents around in the middle of the commit
1707          */
1708         mutex_lock(&root->fs_info->reloc_mutex);
1709
1710         /*
1711          * We needn't worry about the delayed items because we will
1712          * deal with them in create_pending_snapshot(), which is the
1713          * core function of the snapshot creation.
1714          */
1715         ret = create_pending_snapshots(trans, root->fs_info);
1716         if (ret) {
1717                 mutex_unlock(&root->fs_info->reloc_mutex);
1718                 goto cleanup_transaction;
1719         }
1720
1721         /*
1722          * We insert the dir indexes of the snapshots and update the inode
1723          * of the snapshots' parents after the snapshot creation, so there
1724          * are some delayed items which are not dealt with. Now deal with
1725          * them.
1726          *
1727          * We needn't worry that this operation will corrupt the snapshots,
1728          * because all the tree which are snapshoted will be forced to COW
1729          * the nodes and leaves.
1730          */
1731         ret = btrfs_run_delayed_items(trans, root);
1732         if (ret) {
1733                 mutex_unlock(&root->fs_info->reloc_mutex);
1734                 goto cleanup_transaction;
1735         }
1736
1737         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1738         if (ret) {
1739                 mutex_unlock(&root->fs_info->reloc_mutex);
1740                 goto cleanup_transaction;
1741         }
1742
1743         /*
1744          * make sure none of the code above managed to slip in a
1745          * delayed item
1746          */
1747         btrfs_assert_delayed_root_empty(root);
1748
1749         WARN_ON(cur_trans != trans->transaction);
1750
1751         btrfs_scrub_pause(root);
1752         /* btrfs_commit_tree_roots is responsible for getting the
1753          * various roots consistent with each other.  Every pointer
1754          * in the tree of tree roots has to point to the most up to date
1755          * root for every subvolume and other tree.  So, we have to keep
1756          * the tree logging code from jumping in and changing any
1757          * of the trees.
1758          *
1759          * At this point in the commit, there can't be any tree-log
1760          * writers, but a little lower down we drop the trans mutex
1761          * and let new people in.  By holding the tree_log_mutex
1762          * from now until after the super is written, we avoid races
1763          * with the tree-log code.
1764          */
1765         mutex_lock(&root->fs_info->tree_log_mutex);
1766
1767         ret = commit_fs_roots(trans, root);
1768         if (ret) {
1769                 mutex_unlock(&root->fs_info->tree_log_mutex);
1770                 mutex_unlock(&root->fs_info->reloc_mutex);
1771                 goto cleanup_transaction;
1772         }
1773
1774         /* commit_fs_roots gets rid of all the tree log roots, it is now
1775          * safe to free the root of tree log roots
1776          */
1777         btrfs_free_log_root_tree(trans, root->fs_info);
1778
1779         ret = commit_cowonly_roots(trans, root);
1780         if (ret) {
1781                 mutex_unlock(&root->fs_info->tree_log_mutex);
1782                 mutex_unlock(&root->fs_info->reloc_mutex);
1783                 goto cleanup_transaction;
1784         }
1785
1786         /*
1787          * The tasks which save the space cache and inode cache may also
1788          * update ->aborted, check it.
1789          */
1790         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1791                 ret = cur_trans->aborted;
1792                 mutex_unlock(&root->fs_info->tree_log_mutex);
1793                 mutex_unlock(&root->fs_info->reloc_mutex);
1794                 goto cleanup_transaction;
1795         }
1796
1797         btrfs_prepare_extent_commit(trans, root);
1798
1799         cur_trans = root->fs_info->running_transaction;
1800
1801         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1802                             root->fs_info->tree_root->node);
1803         switch_commit_root(root->fs_info->tree_root);
1804
1805         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1806                             root->fs_info->chunk_root->node);
1807         switch_commit_root(root->fs_info->chunk_root);
1808
1809         assert_qgroups_uptodate(trans);
1810         update_super_roots(root);
1811
1812         if (!root->fs_info->log_root_recovering) {
1813                 btrfs_set_super_log_root(root->fs_info->super_copy, 0);
1814                 btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
1815         }
1816
1817         memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
1818                sizeof(*root->fs_info->super_copy));
1819
1820         spin_lock(&root->fs_info->trans_lock);
1821         cur_trans->state = TRANS_STATE_UNBLOCKED;
1822         root->fs_info->running_transaction = NULL;
1823         spin_unlock(&root->fs_info->trans_lock);
1824         mutex_unlock(&root->fs_info->reloc_mutex);
1825
1826         wake_up(&root->fs_info->transaction_wait);
1827
1828         ret = btrfs_write_and_wait_transaction(trans, root);
1829         if (ret) {
1830                 btrfs_error(root->fs_info, ret,
1831                             "Error while writing out transaction");
1832                 mutex_unlock(&root->fs_info->tree_log_mutex);
1833                 goto cleanup_transaction;
1834         }
1835
1836         ret = write_ctree_super(trans, root, 0);
1837         if (ret) {
1838                 mutex_unlock(&root->fs_info->tree_log_mutex);
1839                 goto cleanup_transaction;
1840         }
1841
1842         /*
1843          * the super is written, we can safely allow the tree-loggers
1844          * to go about their business
1845          */
1846         mutex_unlock(&root->fs_info->tree_log_mutex);
1847
1848         btrfs_finish_extent_commit(trans, root);
1849
1850         root->fs_info->last_trans_committed = cur_trans->transid;
1851         /*
1852          * We needn't acquire the lock here because there is no other task
1853          * which can change it.
1854          */
1855         cur_trans->state = TRANS_STATE_COMPLETED;
1856         wake_up(&cur_trans->commit_wait);
1857
1858         spin_lock(&root->fs_info->trans_lock);
1859         list_del_init(&cur_trans->list);
1860         spin_unlock(&root->fs_info->trans_lock);
1861
1862         put_transaction(cur_trans);
1863         put_transaction(cur_trans);
1864
1865         if (trans->type & __TRANS_FREEZABLE)
1866                 sb_end_intwrite(root->fs_info->sb);
1867
1868         trace_btrfs_transaction_commit(root);
1869
1870         btrfs_scrub_continue(root);
1871
1872         if (current->journal_info == trans)
1873                 current->journal_info = NULL;
1874
1875         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1876
1877         if (current != root->fs_info->transaction_kthread)
1878                 btrfs_run_delayed_iputs(root);
1879
1880         return ret;
1881
1882 cleanup_transaction:
1883         btrfs_trans_release_metadata(trans, root);
1884         trans->block_rsv = NULL;
1885         if (trans->qgroup_reserved) {
1886                 btrfs_qgroup_free(root, trans->qgroup_reserved);
1887                 trans->qgroup_reserved = 0;
1888         }
1889         btrfs_warn(root->fs_info, "Skipping commit of aborted transaction.");
1890         if (current->journal_info == trans)
1891                 current->journal_info = NULL;
1892         cleanup_transaction(trans, root, ret);
1893
1894         return ret;
1895 }
1896
1897 /*
1898  * return < 0 if error
1899  * 0 if there are no more dead_roots at the time of call
1900  * 1 there are more to be processed, call me again
1901  *
1902  * The return value indicates there are certainly more snapshots to delete, but
1903  * if there comes a new one during processing, it may return 0. We don't mind,
1904  * because btrfs_commit_super will poke cleaner thread and it will process it a
1905  * few seconds later.
1906  */
1907 int btrfs_clean_one_deleted_snapshot(struct btrfs_root *root)
1908 {
1909         int ret;
1910         struct btrfs_fs_info *fs_info = root->fs_info;
1911
1912         spin_lock(&fs_info->trans_lock);
1913         if (list_empty(&fs_info->dead_roots)) {
1914                 spin_unlock(&fs_info->trans_lock);
1915                 return 0;
1916         }
1917         root = list_first_entry(&fs_info->dead_roots,
1918                         struct btrfs_root, root_list);
1919         list_del(&root->root_list);
1920         spin_unlock(&fs_info->trans_lock);
1921
1922         pr_debug("btrfs: cleaner removing %llu\n",
1923                         (unsigned long long)root->objectid);
1924
1925         btrfs_kill_all_delayed_nodes(root);
1926
1927         if (btrfs_header_backref_rev(root->node) <
1928                         BTRFS_MIXED_BACKREF_REV)
1929                 ret = btrfs_drop_snapshot(root, NULL, 0, 0);
1930         else
1931                 ret = btrfs_drop_snapshot(root, NULL, 1, 0);
1932         /*
1933          * If we encounter a transaction abort during snapshot cleaning, we
1934          * don't want to crash here
1935          */
1936         BUG_ON(ret < 0 && ret != -EAGAIN && ret != -EROFS);
1937         return 1;
1938 }