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                 while (!list_empty(&transaction->pending_chunks)) {
  67                         struct extent_map *em;
  68
  69                         em = list_first_entry(&transaction->pending_chunks,
  70                                               struct extent_map, list);
  71                         list_del_init(&em->list);
  72                         free_extent_map(em);
  73                 }
  74                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  75         }
  76 }
  77
  78 static noinline void switch_commit_root(struct btrfs_root *root)
  79 {
  80         free_extent_buffer(root->commit_root);
  81         root->commit_root = btrfs_root_node(root);
  82 }
  83
  84 static inline void extwriter_counter_inc(struct btrfs_transaction *trans,
  85                                          unsigned int type)
  86 {
  87         if (type & TRANS_EXTWRITERS)
  88                 atomic_inc(&trans->num_extwriters);
  89 }
  90
  91 static inline void extwriter_counter_dec(struct btrfs_transaction *trans,
  92                                          unsigned int type)
  93 {
  94         if (type & TRANS_EXTWRITERS)
  95                 atomic_dec(&trans->num_extwriters);
  96 }
  97
  98 static inline void extwriter_counter_init(struct btrfs_transaction *trans,
  99                                           unsigned int type)
 100 {
 101         atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0));
 102 }
 103
 104 static inline int extwriter_counter_read(struct btrfs_transaction *trans)
 105 {
 106         return atomic_read(&trans->num_extwriters);
 107 }
 108
 109 /*
 110  * either allocate a new transaction or hop into the existing one
 111  */
 112 static noinline int join_transaction(struct btrfs_root *root, unsigned int type)
 113 {
 114         struct btrfs_transaction *cur_trans;
 115         struct btrfs_fs_info *fs_info = root->fs_info;
 116
 117         spin_lock(&fs_info->trans_lock);
 118 loop:
 119         /* The file system has been taken offline. No new transactions. */
 120         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
 121                 spin_unlock(&fs_info->trans_lock);
 122                 return -EROFS;
 123         }
 124
 125         cur_trans = fs_info->running_transaction;
 126         if (cur_trans) {
 127                 if (cur_trans->aborted) {
 128                         spin_unlock(&fs_info->trans_lock);
 129                         return cur_trans->aborted;
 130                 }
 131                 if (btrfs_blocked_trans_types[cur_trans->state] & type) {
 132                         spin_unlock(&fs_info->trans_lock);
 133                         return -EBUSY;
 134                 }
 135                 atomic_inc(&cur_trans->use_count);
 136                 atomic_inc(&cur_trans->num_writers);
 137                 extwriter_counter_inc(cur_trans, type);
 138                 spin_unlock(&fs_info->trans_lock);
 139                 return 0;
 140         }
 141         spin_unlock(&fs_info->trans_lock);
 142
 143         /*
 144          * If we are ATTACH, we just want to catch the current transaction,
 145          * and commit it. If there is no transaction, just return ENOENT.
 146          */
 147         if (type == TRANS_ATTACH)
 148                 return -ENOENT;
 149
 150         /*
 151          * JOIN_NOLOCK only happens during the transaction commit, so
 152          * it is impossible that ->running_transaction is NULL
 153          */
 154         BUG_ON(type == TRANS_JOIN_NOLOCK);
 155
 156         cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
 157         if (!cur_trans)
 158                 return -ENOMEM;
 159
 160         spin_lock(&fs_info->trans_lock);
 161         if (fs_info->running_transaction) {
 162                 /*
 163                  * someone started a transaction after we unlocked.  Make sure
 164                  * to redo the checks above
 165                  */
 166                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 167                 goto loop;
 168         } else if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
 169                 spin_unlock(&fs_info->trans_lock);
 170                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 171                 return -EROFS;
 172         }
 173
 174         atomic_set(&cur_trans->num_writers, 1);
 175         extwriter_counter_init(cur_trans, type);
 176         init_waitqueue_head(&cur_trans->writer_wait);
 177         init_waitqueue_head(&cur_trans->commit_wait);
 178         cur_trans->state = TRANS_STATE_RUNNING;
 179         /*
 180          * One for this trans handle, one so it will live on until we
 181          * commit the transaction.
 182          */
 183         atomic_set(&cur_trans->use_count, 2);
 184         cur_trans->start_time = get_seconds();
 185
 186         cur_trans->delayed_refs.root = RB_ROOT;
 187         cur_trans->delayed_refs.num_entries = 0;
 188         cur_trans->delayed_refs.num_heads_ready = 0;
 189         cur_trans->delayed_refs.num_heads = 0;
 190         cur_trans->delayed_refs.flushing = 0;
 191         cur_trans->delayed_refs.run_delayed_start = 0;
 192
 193         /*
 194          * although the tree mod log is per file system and not per transaction,
 195          * the log must never go across transaction boundaries.
 196          */
 197         smp_mb();
 198         if (!list_empty(&fs_info->tree_mod_seq_list))
 199                 WARN(1, KERN_ERR "btrfs: tree_mod_seq_list not empty when "
 200                         "creating a fresh transaction\n");
 201         if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
 202                 WARN(1, KERN_ERR "btrfs: tree_mod_log rb tree not empty when "
 203                         "creating a fresh transaction\n");
 204         atomic64_set(&fs_info->tree_mod_seq, 0);
 205
 206         spin_lock_init(&cur_trans->delayed_refs.lock);
 207         atomic_set(&cur_trans->delayed_refs.procs_running_refs, 0);
 208         atomic_set(&cur_trans->delayed_refs.ref_seq, 0);
 209         init_waitqueue_head(&cur_trans->delayed_refs.wait);
 210
 211         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 212         INIT_LIST_HEAD(&cur_trans->ordered_operations);
 213         INIT_LIST_HEAD(&cur_trans->pending_chunks);
 214         list_add_tail(&cur_trans->list, &fs_info->trans_list);
 215         extent_io_tree_init(&cur_trans->dirty_pages,
 216                              fs_info->btree_inode->i_mapping);
 217         fs_info->generation++;
 218         cur_trans->transid = fs_info->generation;
 219         fs_info->running_transaction = cur_trans;
 220         cur_trans->aborted = 0;
 221         spin_unlock(&fs_info->trans_lock);
 222
 223         return 0;
 224 }
 225
 226 /*
 227  * this does all the record keeping required to make sure that a reference
 228  * counted root is properly recorded in a given transaction.  This is required
 229  * to make sure the old root from before we joined the transaction is deleted
 230  * when the transaction commits
 231  */
 232 static int record_root_in_trans(struct btrfs_trans_handle *trans,
 233                                struct btrfs_root *root)
 234 {
 235         if (root->ref_cows && root->last_trans < trans->transid) {
 236                 WARN_ON(root == root->fs_info->extent_root);
 237                 WARN_ON(root->commit_root != root->node);
 238
 239                 /*
 240                  * see below for in_trans_setup usage rules
 241                  * we have the reloc mutex held now, so there
 242                  * is only one writer in this function
 243                  */
 244                 root->in_trans_setup = 1;
 245
 246                 /* make sure readers find in_trans_setup before
 247                  * they find our root->last_trans update
 248                  */
 249                 smp_wmb();
 250
 251                 spin_lock(&root->fs_info->fs_roots_radix_lock);
 252                 if (root->last_trans == trans->transid) {
 253                         spin_unlock(&root->fs_info->fs_roots_radix_lock);
 254                         return 0;
 255                 }
 256                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 257                            (unsigned long)root->root_key.objectid,
 258                            BTRFS_ROOT_TRANS_TAG);
 259                 spin_unlock(&root->fs_info->fs_roots_radix_lock);
 260                 root->last_trans = trans->transid;
 261
 262                 /* this is pretty tricky.  We don't want to
 263                  * take the relocation lock in btrfs_record_root_in_trans
 264                  * unless we're really doing the first setup for this root in
 265                  * this transaction.
 266                  *
 267                  * Normally we'd use root->last_trans as a flag to decide
 268                  * if we want to take the expensive mutex.
 269                  *
 270                  * But, we have to set root->last_trans before we
 271                  * init the relocation root, otherwise, we trip over warnings
 272                  * in ctree.c.  The solution used here is to flag ourselves
 273                  * with root->in_trans_setup.  When this is 1, we're still
 274                  * fixing up the reloc trees and everyone must wait.
 275                  *
 276                  * When this is zero, they can trust root->last_trans and fly
 277                  * through btrfs_record_root_in_trans without having to take the
 278                  * lock.  smp_wmb() makes sure that all the writes above are
 279                  * done before we pop in the zero below
 280                  */
 281                 btrfs_init_reloc_root(trans, root);
 282                 smp_wmb();
 283                 root->in_trans_setup = 0;
 284         }
 285         return 0;
 286 }
 287
 288
 289 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 290                                struct btrfs_root *root)
 291 {
 292         if (!root->ref_cows)
 293                 return 0;
 294
 295         /*
 296          * see record_root_in_trans for comments about in_trans_setup usage
 297          * and barriers
 298          */
 299         smp_rmb();
 300         if (root->last_trans == trans->transid &&
 301             !root->in_trans_setup)
 302                 return 0;
 303
 304         mutex_lock(&root->fs_info->reloc_mutex);
 305         record_root_in_trans(trans, root);
 306         mutex_unlock(&root->fs_info->reloc_mutex);
 307
 308         return 0;
 309 }
 310
 311 static inline int is_transaction_blocked(struct btrfs_transaction *trans)
 312 {
 313         return (trans->state >= TRANS_STATE_BLOCKED &&
 314                 trans->state < TRANS_STATE_UNBLOCKED &&
 315                 !trans->aborted);
 316 }
 317
 318 /* wait for commit against the current transaction to become unblocked
 319  * when this is done, it is safe to start a new transaction, but the current
 320  * transaction might not be fully on disk.
 321  */
 322 static void wait_current_trans(struct btrfs_root *root)
 323 {
 324         struct btrfs_transaction *cur_trans;
 325
 326         spin_lock(&root->fs_info->trans_lock);
 327         cur_trans = root->fs_info->running_transaction;
 328         if (cur_trans && is_transaction_blocked(cur_trans)) {
 329                 atomic_inc(&cur_trans->use_count);
 330                 spin_unlock(&root->fs_info->trans_lock);
 331
 332                 wait_event(root->fs_info->transaction_wait,
 333                            cur_trans->state >= TRANS_STATE_UNBLOCKED ||
 334                            cur_trans->aborted);
 335                 put_transaction(cur_trans);
 336         } else {
 337                 spin_unlock(&root->fs_info->trans_lock);
 338         }
 339 }
 340
 341 static int may_wait_transaction(struct btrfs_root *root, int type)
 342 {
 343         if (root->fs_info->log_root_recovering)
 344                 return 0;
 345
 346         if (type == TRANS_USERSPACE)
 347                 return 1;
 348
 349         if (type == TRANS_START &&
 350             !atomic_read(&root->fs_info->open_ioctl_trans))
 351                 return 1;
 352
 353         return 0;
 354 }
 355
 356 static struct btrfs_trans_handle *
 357 start_transaction(struct btrfs_root *root, u64 num_items, unsigned int type,
 358                   enum btrfs_reserve_flush_enum flush)
 359 {
 360         struct btrfs_trans_handle *h;
 361         struct btrfs_transaction *cur_trans;
 362         u64 num_bytes = 0;
 363         int ret;
 364         u64 qgroup_reserved = 0;
 365
 366         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
 367                 return ERR_PTR(-EROFS);
 368
 369         if (current->journal_info) {
 370                 WARN_ON(type & TRANS_EXTWRITERS);
 371                 h = current->journal_info;
 372                 h->use_count++;
 373                 WARN_ON(h->use_count > 2);
 374                 h->orig_rsv = h->block_rsv;
 375                 h->block_rsv = NULL;
 376                 goto got_it;
 377         }
 378
 379         /*
 380          * Do the reservation before we join the transaction so we can do all
 381          * the appropriate flushing if need be.
 382          */
 383         if (num_items > 0 && root != root->fs_info->chunk_root) {
 384                 if (root->fs_info->quota_enabled &&
 385                     is_fstree(root->root_key.objectid)) {
 386                         qgroup_reserved = num_items * root->leafsize;
 387                         ret = btrfs_qgroup_reserve(root, qgroup_reserved);
 388                         if (ret)
 389                                 return ERR_PTR(ret);
 390                 }
 391
 392                 num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
 393                 ret = btrfs_block_rsv_add(root,
 394                                           &root->fs_info->trans_block_rsv,
 395                                           num_bytes, flush);
 396                 if (ret)
 397                         goto reserve_fail;
 398         }
 399 again:
 400         h = kmem_cache_alloc(btrfs_trans_handle_cachep, GFP_NOFS);
 401         if (!h) {
 402                 ret = -ENOMEM;
 403                 goto alloc_fail;
 404         }
 405
 406         /*
 407          * If we are JOIN_NOLOCK we're already committing a transaction and
 408          * waiting on this guy, so we don't need to do the sb_start_intwrite
 409          * because we're already holding a ref.  We need this because we could
 410          * have raced in and did an fsync() on a file which can kick a commit
 411          * and then we deadlock with somebody doing a freeze.
 412          *
 413          * If we are ATTACH, it means we just want to catch the current
 414          * transaction and commit it, so we needn't do sb_start_intwrite().
 415          */
 416         if (type & __TRANS_FREEZABLE)
 417                 sb_start_intwrite(root->fs_info->sb);
 418
 419         if (may_wait_transaction(root, type))
 420                 wait_current_trans(root);
 421
 422         do {
 423                 ret = join_transaction(root, type);
 424                 if (ret == -EBUSY) {
 425                         wait_current_trans(root);
 426                         if (unlikely(type == TRANS_ATTACH))
 427                                 ret = -ENOENT;
 428                 }
 429         } while (ret == -EBUSY);
 430
 431         if (ret < 0) {
 432                 /* We must get the transaction if we are JOIN_NOLOCK. */
 433                 BUG_ON(type == TRANS_JOIN_NOLOCK);
 434                 goto join_fail;
 435         }
 436
 437         cur_trans = root->fs_info->running_transaction;
 438
 439         h->transid = cur_trans->transid;
 440         h->transaction = cur_trans;
 441         h->blocks_used = 0;
 442         h->bytes_reserved = 0;
 443         h->root = root;
 444         h->delayed_ref_updates = 0;
 445         h->use_count = 1;
 446         h->adding_csums = 0;
 447         h->block_rsv = NULL;
 448         h->orig_rsv = NULL;
 449         h->aborted = 0;
 450         h->qgroup_reserved = 0;
 451         h->delayed_ref_elem.seq = 0;
 452         h->type = type;
 453         h->allocating_chunk = false;
 454         INIT_LIST_HEAD(&h->qgroup_ref_list);
 455         INIT_LIST_HEAD(&h->new_bgs);
 456
 457         smp_mb();
 458         if (cur_trans->state >= TRANS_STATE_BLOCKED &&
 459             may_wait_transaction(root, type)) {
 460                 btrfs_commit_transaction(h, root);
 461                 goto again;
 462         }
 463
 464         if (num_bytes) {
 465                 trace_btrfs_space_reservation(root->fs_info, "transaction",
 466                                               h->transid, num_bytes, 1);
 467                 h->block_rsv = &root->fs_info->trans_block_rsv;
 468                 h->bytes_reserved = num_bytes;
 469         }
 470         h->qgroup_reserved = qgroup_reserved;
 471
 472 got_it:
 473         btrfs_record_root_in_trans(h, root);
 474
 475         if (!current->journal_info && type != TRANS_USERSPACE)
 476                 current->journal_info = h;
 477         return h;
 478
 479 join_fail:
 480         if (type & __TRANS_FREEZABLE)
 481                 sb_end_intwrite(root->fs_info->sb);
 482         kmem_cache_free(btrfs_trans_handle_cachep, h);
 483 alloc_fail:
 484         if (num_bytes)
 485                 btrfs_block_rsv_release(root, &root->fs_info->trans_block_rsv,
 486                                         num_bytes);
 487 reserve_fail:
 488         if (qgroup_reserved)
 489                 btrfs_qgroup_free(root, qgroup_reserved);
 490         return ERR_PTR(ret);
 491 }
 492
 493 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 494                                                    int num_items)
 495 {
 496         return start_transaction(root, num_items, TRANS_START,
 497                                  BTRFS_RESERVE_FLUSH_ALL);
 498 }
 499
 500 struct btrfs_trans_handle *btrfs_start_transaction_lflush(
 501                                         struct btrfs_root *root, int num_items)
 502 {
 503         return start_transaction(root, num_items, TRANS_START,
 504                                  BTRFS_RESERVE_FLUSH_LIMIT);
 505 }
 506
 507 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 508 {
 509         return start_transaction(root, 0, TRANS_JOIN, 0);
 510 }
 511
 512 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 513 {
 514         return start_transaction(root, 0, TRANS_JOIN_NOLOCK, 0);
 515 }
 516
 517 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 518 {
 519         return start_transaction(root, 0, TRANS_USERSPACE, 0);
 520 }
 521
 522 /*
 523  * btrfs_attach_transaction() - catch the running transaction
 524  *
 525  * It is used when we want to commit the current the transaction, but
 526  * don't want to start a new one.
 527  *
 528  * Note: If this function return -ENOENT, it just means there is no
 529  * running transaction. But it is possible that the inactive transaction
 530  * is still in the memory, not fully on disk. If you hope there is no
 531  * inactive transaction in the fs when -ENOENT is returned, you should
 532  * invoke
 533  *     btrfs_attach_transaction_barrier()
 534  */
 535 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
 536 {
 537         return start_transaction(root, 0, TRANS_ATTACH, 0);
 538 }
 539
 540 /*
 541  * btrfs_attach_transaction_barrier() - catch the running transaction
 542  *
 543  * It is similar to the above function, the differentia is this one
 544  * will wait for all the inactive transactions until they fully
 545  * complete.
 546  */
 547 struct btrfs_trans_handle *
 548 btrfs_attach_transaction_barrier(struct btrfs_root *root)
 549 {
 550         struct btrfs_trans_handle *trans;
 551
 552         trans = start_transaction(root, 0, TRANS_ATTACH, 0);
 553         if (IS_ERR(trans) && PTR_ERR(trans) == -ENOENT)
 554                 btrfs_wait_for_commit(root, 0);
 555
 556         return trans;
 557 }
 558
 559 /* wait for a transaction commit to be fully complete */
 560 static noinline void wait_for_commit(struct btrfs_root *root,
 561                                     struct btrfs_transaction *commit)
 562 {
 563         wait_event(commit->commit_wait, commit->state == TRANS_STATE_COMPLETED);
 564 }
 565
 566 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 567 {
 568         struct btrfs_transaction *cur_trans = NULL, *t;
 569         int ret = 0;
 570
 571         if (transid) {
 572                 if (transid <= root->fs_info->last_trans_committed)
 573                         goto out;
 574
 575                 ret = -EINVAL;
 576                 /* find specified transaction */
 577                 spin_lock(&root->fs_info->trans_lock);
 578                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 579                         if (t->transid == transid) {
 580                                 cur_trans = t;
 581                                 atomic_inc(&cur_trans->use_count);
 582                                 ret = 0;
 583                                 break;
 584                         }
 585                         if (t->transid > transid) {
 586                                 ret = 0;
 587                                 break;
 588                         }
 589                 }
 590                 spin_unlock(&root->fs_info->trans_lock);
 591                 /* The specified transaction doesn't exist */
 592                 if (!cur_trans)
 593                         goto out;
 594         } else {
 595                 /* find newest transaction that is committing | committed */
 596                 spin_lock(&root->fs_info->trans_lock);
 597                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 598                                             list) {
 599                         if (t->state >= TRANS_STATE_COMMIT_START) {
 600                                 if (t->state == TRANS_STATE_COMPLETED)
 601                                         break;
 602                                 cur_trans = t;
 603                                 atomic_inc(&cur_trans->use_count);
 604                                 break;
 605                         }
 606                 }
 607                 spin_unlock(&root->fs_info->trans_lock);
 608                 if (!cur_trans)
 609                         goto out;  /* nothing committing|committed */
 610         }
 611
 612         wait_for_commit(root, cur_trans);
 613         put_transaction(cur_trans);
 614 out:
 615         return ret;
 616 }
 617
 618 void btrfs_throttle(struct btrfs_root *root)
 619 {
 620         if (!atomic_read(&root->fs_info->open_ioctl_trans))
 621                 wait_current_trans(root);
 622 }
 623
 624 static int should_end_transaction(struct btrfs_trans_handle *trans,
 625                                   struct btrfs_root *root)
 626 {
 627         if (root->fs_info->global_block_rsv.space_info->full &&
 628             btrfs_should_throttle_delayed_refs(trans, root))
 629                 return 1;
 630
 631         return !!btrfs_block_rsv_check(root, &root->fs_info->global_block_rsv, 5);
 632 }
 633
 634 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 635                                  struct btrfs_root *root)
 636 {
 637         struct btrfs_transaction *cur_trans = trans->transaction;
 638         int updates;
 639         int err;
 640
 641         smp_mb();
 642         if (cur_trans->state >= TRANS_STATE_BLOCKED ||
 643             cur_trans->delayed_refs.flushing)
 644                 return 1;
 645
 646         updates = trans->delayed_ref_updates;
 647         trans->delayed_ref_updates = 0;
 648         if (updates) {
 649                 err = btrfs_run_delayed_refs(trans, root, updates);
 650                 if (err) /* Error code will also eval true */
 651                         return err;
 652         }
 653
 654         return should_end_transaction(trans, root);
 655 }
 656
 657 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 658                           struct btrfs_root *root, int throttle)
 659 {
 660         struct btrfs_transaction *cur_trans = trans->transaction;
 661         struct btrfs_fs_info *info = root->fs_info;
 662         unsigned long cur = trans->delayed_ref_updates;
 663         int lock = (trans->type != TRANS_JOIN_NOLOCK);
 664         int err = 0;
 665
 666         if (--trans->use_count) {
 667                 trans->block_rsv = trans->orig_rsv;
 668                 return 0;
 669         }
 670
 671         /*
 672          * do the qgroup accounting as early as possible
 673          */
 674         err = btrfs_delayed_refs_qgroup_accounting(trans, info);
 675
 676         btrfs_trans_release_metadata(trans, root);
 677         trans->block_rsv = NULL;
 678
 679         if (trans->qgroup_reserved) {
 680                 /*
 681                  * the same root has to be passed here between start_transaction
 682                  * and end_transaction. Subvolume quota depends on this.
 683                  */
 684                 btrfs_qgroup_free(trans->root, trans->qgroup_reserved);
 685                 trans->qgroup_reserved = 0;
 686         }
 687
 688         if (!list_empty(&trans->new_bgs))
 689                 btrfs_create_pending_block_groups(trans, root);
 690
 691         trans->delayed_ref_updates = 0;
 692         if (btrfs_should_throttle_delayed_refs(trans, root)) {
 693                 cur = max_t(unsigned long, cur, 1);
 694                 trans->delayed_ref_updates = 0;
 695                 btrfs_run_delayed_refs(trans, root, cur);
 696         }
 697
 698         btrfs_trans_release_metadata(trans, root);
 699         trans->block_rsv = NULL;
 700
 701         if (!list_empty(&trans->new_bgs))
 702                 btrfs_create_pending_block_groups(trans, root);
 703
 704         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
 705             should_end_transaction(trans, root) &&
 706             ACCESS_ONCE(cur_trans->state) == TRANS_STATE_RUNNING) {
 707                 spin_lock(&info->trans_lock);
 708                 if (cur_trans->state == TRANS_STATE_RUNNING)
 709                         cur_trans->state = TRANS_STATE_BLOCKED;
 710                 spin_unlock(&info->trans_lock);
 711         }
 712
 713         if (lock && ACCESS_ONCE(cur_trans->state) == TRANS_STATE_BLOCKED) {
 714                 if (throttle) {
 715                         /*
 716                          * We may race with somebody else here so end up having
 717                          * to call end_transaction on ourselves again, so inc
 718                          * our use_count.
 719                          */
 720                         trans->use_count++;
 721                         return btrfs_commit_transaction(trans, root);
 722                 } else {
 723                         wake_up_process(info->transaction_kthread);
 724                 }
 725         }
 726
 727         if (trans->type & __TRANS_FREEZABLE)
 728                 sb_end_intwrite(root->fs_info->sb);
 729
 730         WARN_ON(cur_trans != info->running_transaction);
 731         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 732         atomic_dec(&cur_trans->num_writers);
 733         extwriter_counter_dec(cur_trans, trans->type);
 734
 735         smp_mb();
 736         if (waitqueue_active(&cur_trans->writer_wait))
 737                 wake_up(&cur_trans->writer_wait);
 738         put_transaction(cur_trans);
 739
 740         if (current->journal_info == trans)
 741                 current->journal_info = NULL;
 742
 743         if (throttle)
 744                 btrfs_run_delayed_iputs(root);
 745
 746         if (trans->aborted ||
 747             test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
 748                 err = -EIO;
 749         assert_qgroups_uptodate(trans);
 750
 751         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 752         return err;
 753 }
 754
 755 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 756                           struct btrfs_root *root)
 757 {
 758         return __btrfs_end_transaction(trans, root, 0);
 759 }
 760
 761 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 762                                    struct btrfs_root *root)
 763 {
 764         return __btrfs_end_transaction(trans, root, 1);
 765 }
 766
 767 int btrfs_end_transaction_dmeta(struct btrfs_trans_handle *trans,
 768                                 struct btrfs_root *root)
 769 {
 770         return __btrfs_end_transaction(trans, root, 1);
 771 }
 772
 773 /*
 774  * when btree blocks are allocated, they have some corresponding bits set for
 775  * them in one of two extent_io trees.  This is used to make sure all of
 776  * those extents are sent to disk but does not wait on them
 777  */
 778 int btrfs_write_marked_extents(struct btrfs_root *root,
 779                                struct extent_io_tree *dirty_pages, int mark)
 780 {
 781         int err = 0;
 782         int werr = 0;
 783         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 784         struct extent_state *cached_state = NULL;
 785         u64 start = 0;
 786         u64 end;
 787
 788         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 789                                       mark, &cached_state)) {
 790                 convert_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT,
 791                                    mark, &cached_state, GFP_NOFS);
 792                 cached_state = NULL;
 793                 err = filemap_fdatawrite_range(mapping, start, end);
 794                 if (err)
 795                         werr = err;
 796                 cond_resched();
 797                 start = end + 1;
 798         }
 799         if (err)
 800                 werr = err;
 801         return werr;
 802 }
 803
 804 /*
 805  * when btree blocks are allocated, they have some corresponding bits set for
 806  * them in one of two extent_io trees.  This is used to make sure all of
 807  * those extents are on disk for transaction or log commit.  We wait
 808  * on all the pages and clear them from the dirty pages state tree
 809  */
 810 int btrfs_wait_marked_extents(struct btrfs_root *root,
 811                               struct extent_io_tree *dirty_pages, int mark)
 812 {
 813         int err = 0;
 814         int werr = 0;
 815         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 816         struct extent_state *cached_state = NULL;
 817         u64 start = 0;
 818         u64 end;
 819
 820         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 821                                       EXTENT_NEED_WAIT, &cached_state)) {
 822                 clear_extent_bit(dirty_pages, start, end, EXTENT_NEED_WAIT,
 823                                  0, 0, &cached_state, GFP_NOFS);
 824                 err = filemap_fdatawait_range(mapping, start, end);
 825                 if (err)
 826                         werr = err;
 827                 cond_resched();
 828                 start = end + 1;
 829         }
 830         if (err)
 831                 werr = err;
 832         return werr;
 833 }
 834
 835 /*
 836  * when btree blocks are allocated, they have some corresponding bits set for
 837  * them in one of two extent_io trees.  This is used to make sure all of
 838  * those extents are on disk for transaction or log commit
 839  */
 840 static int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
 841                                 struct extent_io_tree *dirty_pages, int mark)
 842 {
 843         int ret;
 844         int ret2;
 845         struct blk_plug plug;
 846
 847         blk_start_plug(&plug);
 848         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
 849         blk_finish_plug(&plug);
 850         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
 851
 852         if (ret)
 853                 return ret;
 854         if (ret2)
 855                 return ret2;
 856         return 0;
 857 }
 858
 859 int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 860                                      struct btrfs_root *root)
 861 {
 862         if (!trans || !trans->transaction) {
 863                 struct inode *btree_inode;
 864                 btree_inode = root->fs_info->btree_inode;
 865                 return filemap_write_and_wait(btree_inode->i_mapping);
 866         }
 867         return btrfs_write_and_wait_marked_extents(root,
 868                                            &trans->transaction->dirty_pages,
 869                                            EXTENT_DIRTY);
 870 }
 871
 872 /*
 873  * this is used to update the root pointer in the tree of tree roots.
 874  *
 875  * But, in the case of the extent allocation tree, updating the root
 876  * pointer may allocate blocks which may change the root of the extent
 877  * allocation tree.
 878  *
 879  * So, this loops and repeats and makes sure the cowonly root didn't
 880  * change while the root pointer was being updated in the metadata.
 881  */
 882 static int update_cowonly_root(struct btrfs_trans_handle *trans,
 883                                struct btrfs_root *root)
 884 {
 885         int ret;
 886         u64 old_root_bytenr;
 887         u64 old_root_used;
 888         struct btrfs_root *tree_root = root->fs_info->tree_root;
 889
 890         old_root_used = btrfs_root_used(&root->root_item);
 891         btrfs_write_dirty_block_groups(trans, root);
 892
 893         while (1) {
 894                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
 895                 if (old_root_bytenr == root->node->start &&
 896                     old_root_used == btrfs_root_used(&root->root_item))
 897                         break;
 898
 899                 btrfs_set_root_node(&root->root_item, root->node);
 900                 ret = btrfs_update_root(trans, tree_root,
 901                                         &root->root_key,
 902                                         &root->root_item);
 903                 if (ret)
 904                         return ret;
 905
 906                 old_root_used = btrfs_root_used(&root->root_item);
 907                 ret = btrfs_write_dirty_block_groups(trans, root);
 908                 if (ret)
 909                         return ret;
 910         }
 911
 912         if (root != root->fs_info->extent_root)
 913                 switch_commit_root(root);
 914
 915         return 0;
 916 }
 917
 918 /*
 919  * update all the cowonly tree roots on disk
 920  *
 921  * The error handling in this function may not be obvious. Any of the
 922  * failures will cause the file system to go offline. We still need
 923  * to clean up the delayed refs.
 924  */
 925 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
 926                                          struct btrfs_root *root)
 927 {
 928         struct btrfs_fs_info *fs_info = root->fs_info;
 929         struct list_head *next;
 930         struct extent_buffer *eb;
 931         int ret;
 932
 933         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 934         if (ret)
 935                 return ret;
 936
 937         eb = btrfs_lock_root_node(fs_info->tree_root);
 938         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
 939                               0, &eb);
 940         btrfs_tree_unlock(eb);
 941         free_extent_buffer(eb);
 942
 943         if (ret)
 944                 return ret;
 945
 946         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 947         if (ret)
 948                 return ret;
 949
 950         ret = btrfs_run_dev_stats(trans, root->fs_info);
 951         WARN_ON(ret);
 952         ret = btrfs_run_dev_replace(trans, root->fs_info);
 953         WARN_ON(ret);
 954
 955         ret = btrfs_run_qgroups(trans, root->fs_info);
 956         BUG_ON(ret);
 957
 958         /* run_qgroups might have added some more refs */
 959         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
 960         BUG_ON(ret);
 961
 962         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
 963                 next = fs_info->dirty_cowonly_roots.next;
 964                 list_del_init(next);
 965                 root = list_entry(next, struct btrfs_root, dirty_list);
 966
 967                 ret = update_cowonly_root(trans, root);
 968                 if (ret)
 969                         return ret;
 970         }
 971
 972         down_write(&fs_info->extent_commit_sem);
 973         switch_commit_root(fs_info->extent_root);
 974         up_write(&fs_info->extent_commit_sem);
 975
 976         btrfs_after_dev_replace_commit(fs_info);
 977
 978         return 0;
 979 }
 980
 981 /*
 982  * dead roots are old snapshots that need to be deleted.  This allocates
 983  * a dirty root struct and adds it into the list of dead roots that need to
 984  * be deleted
 985  */
 986 void btrfs_add_dead_root(struct btrfs_root *root)
 987 {
 988         spin_lock(&root->fs_info->trans_lock);
 989         if (list_empty(&root->root_list))
 990                 list_add_tail(&root->root_list, &root->fs_info->dead_roots);
 991         spin_unlock(&root->fs_info->trans_lock);
 992 }
 993
 994 /*
 995  * update all the cowonly tree roots on disk
 996  */
 997 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
 998                                     struct btrfs_root *root)
 999 {
1000         struct btrfs_root *gang[8];
1001         struct btrfs_fs_info *fs_info = root->fs_info;
1002         int i;
1003         int ret;
1004         int err = 0;
1005
1006         spin_lock(&fs_info->fs_roots_radix_lock);
1007         while (1) {
1008                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1009                                                  (void **)gang, 0,
1010                                                  ARRAY_SIZE(gang),
1011                                                  BTRFS_ROOT_TRANS_TAG);
1012                 if (ret == 0)
1013                         break;
1014                 for (i = 0; i < ret; i++) {
1015                         root = gang[i];
1016                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
1017                                         (unsigned long)root->root_key.objectid,
1018                                         BTRFS_ROOT_TRANS_TAG);
1019                         spin_unlock(&fs_info->fs_roots_radix_lock);
1020
1021                         btrfs_free_log(trans, root);
1022                         btrfs_update_reloc_root(trans, root);
1023                         btrfs_orphan_commit_root(trans, root);
1024
1025                         btrfs_save_ino_cache(root, trans);
1026
1027                         /* see comments in should_cow_block() */
1028                         root->force_cow = 0;
1029                         smp_wmb();
1030
1031                         if (root->commit_root != root->node) {
1032                                 mutex_lock(&root->fs_commit_mutex);
1033                                 switch_commit_root(root);
1034                                 btrfs_unpin_free_ino(root);
1035                                 mutex_unlock(&root->fs_commit_mutex);
1036
1037                                 btrfs_set_root_node(&root->root_item,
1038                                                     root->node);
1039                         }
1040
1041                         err = btrfs_update_root(trans, fs_info->tree_root,
1042                                                 &root->root_key,
1043                                                 &root->root_item);
1044                         spin_lock(&fs_info->fs_roots_radix_lock);
1045                         if (err)
1046                                 break;
1047                 }
1048         }
1049         spin_unlock(&fs_info->fs_roots_radix_lock);
1050         return err;
1051 }
1052
1053 /*
1054  * defrag a given btree.
1055  * Every leaf in the btree is read and defragged.
1056  */
1057 int btrfs_defrag_root(struct btrfs_root *root)
1058 {
1059         struct btrfs_fs_info *info = root->fs_info;
1060         struct btrfs_trans_handle *trans;
1061         int ret;
1062
1063         if (xchg(&root->defrag_running, 1))
1064                 return 0;
1065
1066         while (1) {
1067                 trans = btrfs_start_transaction(root, 0);
1068                 if (IS_ERR(trans))
1069                         return PTR_ERR(trans);
1070
1071                 ret = btrfs_defrag_leaves(trans, root);
1072
1073                 btrfs_end_transaction(trans, root);
1074                 btrfs_btree_balance_dirty(info->tree_root);
1075                 cond_resched();
1076
1077                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
1078                         break;
1079
1080                 if (btrfs_defrag_cancelled(root->fs_info)) {
1081                         printk(KERN_DEBUG "btrfs: defrag_root cancelled\n");
1082                         ret = -EAGAIN;
1083                         break;
1084                 }
1085         }
1086         root->defrag_running = 0;
1087         return ret;
1088 }
1089
1090 /*
1091  * new snapshots need to be created at a very specific time in the
1092  * transaction commit.  This does the actual creation.
1093  *
1094  * Note:
1095  * If the error which may affect the commitment of the current transaction
1096  * happens, we should return the error number. If the error which just affect
1097  * the creation of the pending snapshots, just return 0.
1098  */
1099 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1100                                    struct btrfs_fs_info *fs_info,
1101                                    struct btrfs_pending_snapshot *pending)
1102 {
1103         struct btrfs_key key;
1104         struct btrfs_root_item *new_root_item;
1105         struct btrfs_root *tree_root = fs_info->tree_root;
1106         struct btrfs_root *root = pending->root;
1107         struct btrfs_root *parent_root;
1108         struct btrfs_block_rsv *rsv;
1109         struct inode *parent_inode;
1110         struct btrfs_path *path;
1111         struct btrfs_dir_item *dir_item;
1112         struct dentry *dentry;
1113         struct extent_buffer *tmp;
1114         struct extent_buffer *old;
1115         struct timespec cur_time = CURRENT_TIME;
1116         int ret = 0;
1117         u64 to_reserve = 0;
1118         u64 index = 0;
1119         u64 objectid;
1120         u64 root_flags;
1121         uuid_le new_uuid;
1122
1123         path = btrfs_alloc_path();
1124         if (!path) {
1125                 pending->error = -ENOMEM;
1126                 return 0;
1127         }
1128
1129         new_root_item = kmalloc(sizeof(*new_root_item), GFP_NOFS);
1130         if (!new_root_item) {
1131                 pending->error = -ENOMEM;
1132                 goto root_item_alloc_fail;
1133         }
1134
1135         pending->error = btrfs_find_free_objectid(tree_root, &objectid);
1136         if (pending->error)
1137                 goto no_free_objectid;
1138
1139         btrfs_reloc_pre_snapshot(trans, pending, &to_reserve);
1140
1141         if (to_reserve > 0) {
1142                 pending->error = btrfs_block_rsv_add(root,
1143                                                      &pending->block_rsv,
1144                                                      to_reserve,
1145                                                      BTRFS_RESERVE_NO_FLUSH);
1146                 if (pending->error)
1147                         goto no_free_objectid;
1148         }
1149
1150         pending->error = btrfs_qgroup_inherit(trans, fs_info,
1151                                               root->root_key.objectid,
1152                                               objectid, pending->inherit);
1153         if (pending->error)
1154                 goto no_free_objectid;
1155
1156         key.objectid = objectid;
1157         key.offset = (u64)-1;
1158         key.type = BTRFS_ROOT_ITEM_KEY;
1159
1160         rsv = trans->block_rsv;
1161         trans->block_rsv = &pending->block_rsv;
1162         trans->bytes_reserved = trans->block_rsv->reserved;
1163
1164         dentry = pending->dentry;
1165         parent_inode = pending->dir;
1166         parent_root = BTRFS_I(parent_inode)->root;
1167         record_root_in_trans(trans, parent_root);
1168
1169         /*
1170          * insert the directory item
1171          */
1172         ret = btrfs_set_inode_index(parent_inode, &index);
1173         BUG_ON(ret); /* -ENOMEM */
1174
1175         /* check if there is a file/dir which has the same name. */
1176         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1177                                          btrfs_ino(parent_inode),
1178                                          dentry->d_name.name,
1179                                          dentry->d_name.len, 0);
1180         if (dir_item != NULL && !IS_ERR(dir_item)) {
1181                 pending->error = -EEXIST;
1182                 goto dir_item_existed;
1183         } else if (IS_ERR(dir_item)) {
1184                 ret = PTR_ERR(dir_item);
1185                 btrfs_abort_transaction(trans, root, ret);
1186                 goto fail;
1187         }
1188         btrfs_release_path(path);
1189
1190         /*
1191          * pull in the delayed directory update
1192          * and the delayed inode item
1193          * otherwise we corrupt the FS during
1194          * snapshot
1195          */
1196         ret = btrfs_run_delayed_items(trans, root);
1197         if (ret) {      /* Transaction aborted */
1198                 btrfs_abort_transaction(trans, root, ret);
1199                 goto fail;
1200         }
1201
1202         record_root_in_trans(trans, root);
1203         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1204         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1205         btrfs_check_and_init_root_item(new_root_item);
1206
1207         root_flags = btrfs_root_flags(new_root_item);
1208         if (pending->readonly)
1209                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1210         else
1211                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1212         btrfs_set_root_flags(new_root_item, root_flags);
1213
1214         btrfs_set_root_generation_v2(new_root_item,
1215                         trans->transid);
1216         uuid_le_gen(&new_uuid);
1217         memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1218         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1219                         BTRFS_UUID_SIZE);
1220         if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1221                 memset(new_root_item->received_uuid, 0,
1222                        sizeof(new_root_item->received_uuid));
1223                 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1224                 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1225                 btrfs_set_root_stransid(new_root_item, 0);
1226                 btrfs_set_root_rtransid(new_root_item, 0);
1227         }
1228         btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1229         btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1230         btrfs_set_root_otransid(new_root_item, trans->transid);
1231
1232         old = btrfs_lock_root_node(root);
1233         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1234         if (ret) {
1235                 btrfs_tree_unlock(old);
1236                 free_extent_buffer(old);
1237                 btrfs_abort_transaction(trans, root, ret);
1238                 goto fail;
1239         }
1240
1241         btrfs_set_lock_blocking(old);
1242
1243         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1244         /* clean up in any case */
1245         btrfs_tree_unlock(old);
1246         free_extent_buffer(old);
1247         if (ret) {
1248                 btrfs_abort_transaction(trans, root, ret);
1249                 goto fail;
1250         }
1251
1252         /* see comments in should_cow_block() */
1253         root->force_cow = 1;
1254         smp_wmb();
1255
1256         btrfs_set_root_node(new_root_item, tmp);
1257         /* record when the snapshot was created in key.offset */
1258         key.offset = trans->transid;
1259         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1260         btrfs_tree_unlock(tmp);
1261         free_extent_buffer(tmp);
1262         if (ret) {
1263                 btrfs_abort_transaction(trans, root, ret);
1264                 goto fail;
1265         }
1266
1267         /*
1268          * insert root back/forward references
1269          */
1270         ret = btrfs_add_root_ref(trans, tree_root, objectid,
1271                                  parent_root->root_key.objectid,
1272                                  btrfs_ino(parent_inode), index,
1273                                  dentry->d_name.name, dentry->d_name.len);
1274         if (ret) {
1275                 btrfs_abort_transaction(trans, root, ret);
1276                 goto fail;
1277         }
1278
1279         key.offset = (u64)-1;
1280         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1281         if (IS_ERR(pending->snap)) {
1282                 ret = PTR_ERR(pending->snap);
1283                 btrfs_abort_transaction(trans, root, ret);
1284                 goto fail;
1285         }
1286
1287         ret = btrfs_reloc_post_snapshot(trans, pending);
1288         if (ret) {
1289                 btrfs_abort_transaction(trans, root, ret);
1290                 goto fail;
1291         }
1292
1293         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1294         if (ret) {
1295                 btrfs_abort_transaction(trans, root, ret);
1296                 goto fail;
1297         }
1298
1299         ret = btrfs_insert_dir_item(trans, parent_root,
1300                                     dentry->d_name.name, dentry->d_name.len,
1301                                     parent_inode, &key,
1302                                     BTRFS_FT_DIR, index);
1303         /* We have check then name at the beginning, so it is impossible. */
1304         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1305         if (ret) {
1306                 btrfs_abort_transaction(trans, root, ret);
1307                 goto fail;
1308         }
1309
1310         btrfs_i_size_write(parent_inode, parent_inode->i_size +
1311                                          dentry->d_name.len * 2);
1312         parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
1313         ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1314         if (ret) {
1315                 btrfs_abort_transaction(trans, root, ret);
1316                 goto fail;
1317         }
1318         ret = btrfs_uuid_tree_add(trans, fs_info->uuid_root, new_uuid.b,
1319                                   BTRFS_UUID_KEY_SUBVOL, objectid);
1320         if (ret) {
1321                 btrfs_abort_transaction(trans, root, ret);
1322                 goto fail;
1323         }
1324         if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1325                 ret = btrfs_uuid_tree_add(trans, fs_info->uuid_root,
1326                                           new_root_item->received_uuid,
1327                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1328                                           objectid);
1329                 if (ret && ret != -EEXIST) {
1330                         btrfs_abort_transaction(trans, root, ret);
1331                         goto fail;
1332                 }
1333         }
1334 fail:
1335         pending->error = ret;
1336 dir_item_existed:
1337         trans->block_rsv = rsv;
1338         trans->bytes_reserved = 0;
1339 no_free_objectid:
1340         kfree(new_root_item);
1341 root_item_alloc_fail:
1342         btrfs_free_path(path);
1343         return ret;
1344 }
1345
1346 /*
1347  * create all the snapshots we've scheduled for creation
1348  */
1349 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1350                                              struct btrfs_fs_info *fs_info)
1351 {
1352         struct btrfs_pending_snapshot *pending, *next;
1353         struct list_head *head = &trans->transaction->pending_snapshots;
1354         int ret = 0;
1355
1356         list_for_each_entry_safe(pending, next, head, list) {
1357                 list_del(&pending->list);
1358                 ret = create_pending_snapshot(trans, fs_info, pending);
1359                 if (ret)
1360                         break;
1361         }
1362         return ret;
1363 }
1364
1365 static void update_super_roots(struct btrfs_root *root)
1366 {
1367         struct btrfs_root_item *root_item;
1368         struct btrfs_super_block *super;
1369
1370         super = root->fs_info->super_copy;
1371
1372         root_item = &root->fs_info->chunk_root->root_item;
1373         super->chunk_root = root_item->bytenr;
1374         super->chunk_root_generation = root_item->generation;
1375         super->chunk_root_level = root_item->level;
1376
1377         root_item = &root->fs_info->tree_root->root_item;
1378         super->root = root_item->bytenr;
1379         super->generation = root_item->generation;
1380         super->root_level = root_item->level;
1381         if (btrfs_test_opt(root, SPACE_CACHE))
1382                 super->cache_generation = root_item->generation;
1383         if (root->fs_info->update_uuid_tree_gen)
1384                 super->uuid_tree_generation = root_item->generation;
1385 }
1386
1387 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1388 {
1389         struct btrfs_transaction *trans;
1390         int ret = 0;
1391
1392         spin_lock(&info->trans_lock);
1393         trans = info->running_transaction;
1394         if (trans)
1395                 ret = (trans->state >= TRANS_STATE_COMMIT_START);
1396         spin_unlock(&info->trans_lock);
1397         return ret;
1398 }
1399
1400 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1401 {
1402         struct btrfs_transaction *trans;
1403         int ret = 0;
1404
1405         spin_lock(&info->trans_lock);
1406         trans = info->running_transaction;
1407         if (trans)
1408                 ret = is_transaction_blocked(trans);
1409         spin_unlock(&info->trans_lock);
1410         return ret;
1411 }
1412
1413 /*
1414  * wait for the current transaction commit to start and block subsequent
1415  * transaction joins
1416  */
1417 static void wait_current_trans_commit_start(struct btrfs_root *root,
1418                                             struct btrfs_transaction *trans)
1419 {
1420         wait_event(root->fs_info->transaction_blocked_wait,
1421                    trans->state >= TRANS_STATE_COMMIT_START ||
1422                    trans->aborted);
1423 }
1424
1425 /*
1426  * wait for the current transaction to start and then become unblocked.
1427  * caller holds ref.
1428  */
1429 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1430                                          struct btrfs_transaction *trans)
1431 {
1432         wait_event(root->fs_info->transaction_wait,
1433                    trans->state >= TRANS_STATE_UNBLOCKED ||
1434                    trans->aborted);
1435 }
1436
1437 /*
1438  * commit transactions asynchronously. once btrfs_commit_transaction_async
1439  * returns, any subsequent transaction will not be allowed to join.
1440  */
1441 struct btrfs_async_commit {
1442         struct btrfs_trans_handle *newtrans;
1443         struct btrfs_root *root;
1444         struct work_struct work;
1445 };
1446
1447 static void do_async_commit(struct work_struct *work)
1448 {
1449         struct btrfs_async_commit *ac =
1450                 container_of(work, struct btrfs_async_commit, work);
1451
1452         /*
1453          * We've got freeze protection passed with the transaction.
1454          * Tell lockdep about it.
1455          */
1456         if (ac->newtrans->type < TRANS_JOIN_NOLOCK)
1457                 rwsem_acquire_read(
1458                      &ac->root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
1459                      0, 1, _THIS_IP_);
1460
1461         current->journal_info = ac->newtrans;
1462
1463         btrfs_commit_transaction(ac->newtrans, ac->root);
1464         kfree(ac);
1465 }
1466
1467 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1468                                    struct btrfs_root *root,
1469                                    int wait_for_unblock)
1470 {
1471         struct btrfs_async_commit *ac;
1472         struct btrfs_transaction *cur_trans;
1473
1474         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1475         if (!ac)
1476                 return -ENOMEM;
1477
1478         INIT_WORK(&ac->work, do_async_commit);
1479         ac->root = root;
1480         ac->newtrans = btrfs_join_transaction(root);
1481         if (IS_ERR(ac->newtrans)) {
1482                 int err = PTR_ERR(ac->newtrans);
1483                 kfree(ac);
1484                 return err;
1485         }
1486
1487         /* take transaction reference */
1488         cur_trans = trans->transaction;
1489         atomic_inc(&cur_trans->use_count);
1490
1491         btrfs_end_transaction(trans, root);
1492
1493         /*
1494          * Tell lockdep we've released the freeze rwsem, since the
1495          * async commit thread will be the one to unlock it.
1496          */
1497         if (trans->type < TRANS_JOIN_NOLOCK)
1498                 rwsem_release(
1499                         &root->fs_info->sb->s_writers.lock_map[SB_FREEZE_FS-1],
1500                         1, _THIS_IP_);
1501
1502         schedule_work(&ac->work);
1503
1504         /* wait for transaction to start and unblock */
1505         if (wait_for_unblock)
1506                 wait_current_trans_commit_start_and_unblock(root, cur_trans);
1507         else
1508                 wait_current_trans_commit_start(root, cur_trans);
1509
1510         if (current->journal_info == trans)
1511                 current->journal_info = NULL;
1512
1513         put_transaction(cur_trans);
1514         return 0;
1515 }
1516
1517
1518 static void cleanup_transaction(struct btrfs_trans_handle *trans,
1519                                 struct btrfs_root *root, int err)
1520 {
1521         struct btrfs_transaction *cur_trans = trans->transaction;
1522         DEFINE_WAIT(wait);
1523
1524         WARN_ON(trans->use_count > 1);
1525
1526         btrfs_abort_transaction(trans, root, err);
1527
1528         spin_lock(&root->fs_info->trans_lock);
1529
1530         /*
1531          * If the transaction is removed from the list, it means this
1532          * transaction has been committed successfully, so it is impossible
1533          * to call the cleanup function.
1534          */
1535         BUG_ON(list_empty(&cur_trans->list));
1536
1537         list_del_init(&cur_trans->list);
1538         if (cur_trans == root->fs_info->running_transaction) {
1539                 cur_trans->state = TRANS_STATE_COMMIT_DOING;
1540                 spin_unlock(&root->fs_info->trans_lock);
1541                 wait_event(cur_trans->writer_wait,
1542                            atomic_read(&cur_trans->num_writers) == 1);
1543
1544                 spin_lock(&root->fs_info->trans_lock);
1545         }
1546         spin_unlock(&root->fs_info->trans_lock);
1547
1548         btrfs_cleanup_one_transaction(trans->transaction, root);
1549
1550         spin_lock(&root->fs_info->trans_lock);
1551         if (cur_trans == root->fs_info->running_transaction)
1552                 root->fs_info->running_transaction = NULL;
1553         spin_unlock(&root->fs_info->trans_lock);
1554
1555         put_transaction(cur_trans);
1556         put_transaction(cur_trans);
1557
1558         trace_btrfs_transaction_commit(root);
1559
1560         btrfs_scrub_continue(root);
1561
1562         if (current->journal_info == trans)
1563                 current->journal_info = NULL;
1564
1565         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1566 }
1567
1568 static int btrfs_flush_all_pending_stuffs(struct btrfs_trans_handle *trans,
1569                                           struct btrfs_root *root)
1570 {
1571         int ret;
1572
1573         ret = btrfs_run_delayed_items(trans, root);
1574         if (ret)
1575                 return ret;
1576
1577         /*
1578          * running the delayed items may have added new refs. account
1579          * them now so that they hinder processing of more delayed refs
1580          * as little as possible.
1581          */
1582         btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
1583
1584         /*
1585          * rename don't use btrfs_join_transaction, so, once we
1586          * set the transaction to blocked above, we aren't going
1587          * to get any new ordered operations.  We can safely run
1588          * it here and no for sure that nothing new will be added
1589          * to the list
1590          */
1591         ret = btrfs_run_ordered_operations(trans, root, 1);
1592
1593         return ret;
1594 }
1595
1596 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
1597 {
1598         if (btrfs_test_opt(fs_info->tree_root, FLUSHONCOMMIT))
1599                 return btrfs_start_all_delalloc_inodes(fs_info, 1);
1600         return 0;
1601 }
1602
1603 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
1604 {
1605         if (btrfs_test_opt(fs_info->tree_root, FLUSHONCOMMIT))
1606                 btrfs_wait_all_ordered_extents(fs_info);
1607 }
1608
1609 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1610                              struct btrfs_root *root)
1611 {
1612         struct btrfs_transaction *cur_trans = trans->transaction;
1613         struct btrfs_transaction *prev_trans = NULL;
1614         int ret;
1615
1616         ret = btrfs_run_ordered_operations(trans, root, 0);
1617         if (ret) {
1618                 btrfs_abort_transaction(trans, root, ret);
1619                 btrfs_end_transaction(trans, root);
1620                 return ret;
1621         }
1622
1623         /* Stop the commit early if ->aborted is set */
1624         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1625                 ret = cur_trans->aborted;
1626                 btrfs_end_transaction(trans, root);
1627                 return ret;
1628         }
1629
1630         /* make a pass through all the delayed refs we have so far
1631          * any runnings procs may add more while we are here
1632          */
1633         ret = btrfs_run_delayed_refs(trans, root, 0);
1634         if (ret) {
1635                 btrfs_end_transaction(trans, root);
1636                 return ret;
1637         }
1638
1639         btrfs_trans_release_metadata(trans, root);
1640         trans->block_rsv = NULL;
1641         if (trans->qgroup_reserved) {
1642                 btrfs_qgroup_free(root, trans->qgroup_reserved);
1643                 trans->qgroup_reserved = 0;
1644         }
1645
1646         cur_trans = trans->transaction;
1647
1648         /*
1649          * set the flushing flag so procs in this transaction have to
1650          * start sending their work down.
1651          */
1652         cur_trans->delayed_refs.flushing = 1;
1653         smp_wmb();
1654
1655         if (!list_empty(&trans->new_bgs))
1656                 btrfs_create_pending_block_groups(trans, root);
1657
1658         ret = btrfs_run_delayed_refs(trans, root, 0);
1659         if (ret) {
1660                 btrfs_end_transaction(trans, root);
1661                 return ret;
1662         }
1663
1664         spin_lock(&root->fs_info->trans_lock);
1665         if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
1666                 spin_unlock(&root->fs_info->trans_lock);
1667                 atomic_inc(&cur_trans->use_count);
1668                 ret = btrfs_end_transaction(trans, root);
1669
1670                 wait_for_commit(root, cur_trans);
1671
1672                 put_transaction(cur_trans);
1673
1674                 return ret;
1675         }
1676
1677         cur_trans->state = TRANS_STATE_COMMIT_START;
1678         wake_up(&root->fs_info->transaction_blocked_wait);
1679
1680         if (cur_trans->list.prev != &root->fs_info->trans_list) {
1681                 prev_trans = list_entry(cur_trans->list.prev,
1682                                         struct btrfs_transaction, list);
1683                 if (prev_trans->state != TRANS_STATE_COMPLETED) {
1684                         atomic_inc(&prev_trans->use_count);
1685                         spin_unlock(&root->fs_info->trans_lock);
1686
1687                         wait_for_commit(root, prev_trans);
1688
1689                         put_transaction(prev_trans);
1690                 } else {
1691                         spin_unlock(&root->fs_info->trans_lock);
1692                 }
1693         } else {
1694                 spin_unlock(&root->fs_info->trans_lock);
1695         }
1696
1697         extwriter_counter_dec(cur_trans, trans->type);
1698
1699         ret = btrfs_start_delalloc_flush(root->fs_info);
1700         if (ret)
1701                 goto cleanup_transaction;
1702
1703         ret = btrfs_flush_all_pending_stuffs(trans, root);
1704         if (ret)
1705                 goto cleanup_transaction;
1706
1707         wait_event(cur_trans->writer_wait,
1708                    extwriter_counter_read(cur_trans) == 0);
1709
1710         /* some pending stuffs might be added after the previous flush. */
1711         ret = btrfs_flush_all_pending_stuffs(trans, root);
1712         if (ret)
1713                 goto cleanup_transaction;
1714
1715         btrfs_wait_delalloc_flush(root->fs_info);
1716         /*
1717          * Ok now we need to make sure to block out any other joins while we
1718          * commit the transaction.  We could have started a join before setting
1719          * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
1720          */
1721         spin_lock(&root->fs_info->trans_lock);
1722         cur_trans->state = TRANS_STATE_COMMIT_DOING;
1723         spin_unlock(&root->fs_info->trans_lock);
1724         wait_event(cur_trans->writer_wait,
1725                    atomic_read(&cur_trans->num_writers) == 1);
1726
1727         /* ->aborted might be set after the previous check, so check it */
1728         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1729                 ret = cur_trans->aborted;
1730                 goto cleanup_transaction;
1731         }
1732         /*
1733          * the reloc mutex makes sure that we stop
1734          * the balancing code from coming in and moving
1735          * extents around in the middle of the commit
1736          */
1737         mutex_lock(&root->fs_info->reloc_mutex);
1738
1739         /*
1740          * We needn't worry about the delayed items because we will
1741          * deal with them in create_pending_snapshot(), which is the
1742          * core function of the snapshot creation.
1743          */
1744         ret = create_pending_snapshots(trans, root->fs_info);
1745         if (ret) {
1746                 mutex_unlock(&root->fs_info->reloc_mutex);
1747                 goto cleanup_transaction;
1748         }
1749
1750         /*
1751          * We insert the dir indexes of the snapshots and update the inode
1752          * of the snapshots' parents after the snapshot creation, so there
1753          * are some delayed items which are not dealt with. Now deal with
1754          * them.
1755          *
1756          * We needn't worry that this operation will corrupt the snapshots,
1757          * because all the tree which are snapshoted will be forced to COW
1758          * the nodes and leaves.
1759          */
1760         ret = btrfs_run_delayed_items(trans, root);
1761         if (ret) {
1762                 mutex_unlock(&root->fs_info->reloc_mutex);
1763                 goto cleanup_transaction;
1764         }
1765
1766         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1767         if (ret) {
1768                 mutex_unlock(&root->fs_info->reloc_mutex);
1769                 goto cleanup_transaction;
1770         }
1771
1772         /*
1773          * make sure none of the code above managed to slip in a
1774          * delayed item
1775          */
1776         btrfs_assert_delayed_root_empty(root);
1777
1778         WARN_ON(cur_trans != trans->transaction);
1779
1780         btrfs_scrub_pause(root);
1781         /* btrfs_commit_tree_roots is responsible for getting the
1782          * various roots consistent with each other.  Every pointer
1783          * in the tree of tree roots has to point to the most up to date
1784          * root for every subvolume and other tree.  So, we have to keep
1785          * the tree logging code from jumping in and changing any
1786          * of the trees.
1787          *
1788          * At this point in the commit, there can't be any tree-log
1789          * writers, but a little lower down we drop the trans mutex
1790          * and let new people in.  By holding the tree_log_mutex
1791          * from now until after the super is written, we avoid races
1792          * with the tree-log code.
1793          */
1794         mutex_lock(&root->fs_info->tree_log_mutex);
1795
1796         ret = commit_fs_roots(trans, root);
1797         if (ret) {
1798                 mutex_unlock(&root->fs_info->tree_log_mutex);
1799                 mutex_unlock(&root->fs_info->reloc_mutex);
1800                 goto cleanup_transaction;
1801         }
1802
1803         /* commit_fs_roots gets rid of all the tree log roots, it is now
1804          * safe to free the root of tree log roots
1805          */
1806         btrfs_free_log_root_tree(trans, root->fs_info);
1807
1808         ret = commit_cowonly_roots(trans, root);
1809         if (ret) {
1810                 mutex_unlock(&root->fs_info->tree_log_mutex);
1811                 mutex_unlock(&root->fs_info->reloc_mutex);
1812                 goto cleanup_transaction;
1813         }
1814
1815         /*
1816          * The tasks which save the space cache and inode cache may also
1817          * update ->aborted, check it.
1818          */
1819         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1820                 ret = cur_trans->aborted;
1821                 mutex_unlock(&root->fs_info->tree_log_mutex);
1822                 mutex_unlock(&root->fs_info->reloc_mutex);
1823                 goto cleanup_transaction;
1824         }
1825
1826         btrfs_prepare_extent_commit(trans, root);
1827
1828         cur_trans = root->fs_info->running_transaction;
1829
1830         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
1831                             root->fs_info->tree_root->node);
1832         switch_commit_root(root->fs_info->tree_root);
1833
1834         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
1835                             root->fs_info->chunk_root->node);
1836         switch_commit_root(root->fs_info->chunk_root);
1837
1838         assert_qgroups_uptodate(trans);
1839         update_super_roots(root);
1840
1841         btrfs_set_super_log_root(root->fs_info->super_copy, 0);
1842         btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
1843         memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
1844                sizeof(*root->fs_info->super_copy));
1845
1846         spin_lock(&root->fs_info->trans_lock);
1847         cur_trans->state = TRANS_STATE_UNBLOCKED;
1848         root->fs_info->running_transaction = NULL;
1849         spin_unlock(&root->fs_info->trans_lock);
1850         mutex_unlock(&root->fs_info->reloc_mutex);
1851
1852         wake_up(&root->fs_info->transaction_wait);
1853
1854         ret = btrfs_write_and_wait_transaction(trans, root);
1855         if (ret) {
1856                 btrfs_error(root->fs_info, ret,
1857                             "Error while writing out transaction");
1858                 mutex_unlock(&root->fs_info->tree_log_mutex);
1859                 goto cleanup_transaction;
1860         }
1861
1862         ret = write_ctree_super(trans, root, 0);
1863         if (ret) {
1864                 mutex_unlock(&root->fs_info->tree_log_mutex);
1865                 goto cleanup_transaction;
1866         }
1867
1868         /*
1869          * the super is written, we can safely allow the tree-loggers
1870          * to go about their business
1871          */
1872         mutex_unlock(&root->fs_info->tree_log_mutex);
1873
1874         btrfs_finish_extent_commit(trans, root);
1875
1876         root->fs_info->last_trans_committed = cur_trans->transid;
1877         /*
1878          * We needn't acquire the lock here because there is no other task
1879          * which can change it.
1880          */
1881         cur_trans->state = TRANS_STATE_COMPLETED;
1882         wake_up(&cur_trans->commit_wait);
1883
1884         spin_lock(&root->fs_info->trans_lock);
1885         list_del_init(&cur_trans->list);
1886         spin_unlock(&root->fs_info->trans_lock);
1887
1888         put_transaction(cur_trans);
1889         put_transaction(cur_trans);
1890
1891         if (trans->type & __TRANS_FREEZABLE)
1892                 sb_end_intwrite(root->fs_info->sb);
1893
1894         trace_btrfs_transaction_commit(root);
1895
1896         btrfs_scrub_continue(root);
1897
1898         if (current->journal_info == trans)
1899                 current->journal_info = NULL;
1900
1901         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1902
1903         if (current != root->fs_info->transaction_kthread)
1904                 btrfs_run_delayed_iputs(root);
1905
1906         return ret;
1907
1908 cleanup_transaction:
1909         btrfs_trans_release_metadata(trans, root);
1910         trans->block_rsv = NULL;
1911         if (trans->qgroup_reserved) {
1912                 btrfs_qgroup_free(root, trans->qgroup_reserved);
1913                 trans->qgroup_reserved = 0;
1914         }
1915         btrfs_warn(root->fs_info, "Skipping commit of aborted transaction.");
1916         if (current->journal_info == trans)
1917                 current->journal_info = NULL;
1918         cleanup_transaction(trans, root, ret);
1919
1920         return ret;
1921 }
1922
1923 /*
1924  * return < 0 if error
1925  * 0 if there are no more dead_roots at the time of call
1926  * 1 there are more to be processed, call me again
1927  *
1928  * The return value indicates there are certainly more snapshots to delete, but
1929  * if there comes a new one during processing, it may return 0. We don't mind,
1930  * because btrfs_commit_super will poke cleaner thread and it will process it a
1931  * few seconds later.
1932  */
1933 int btrfs_clean_one_deleted_snapshot(struct btrfs_root *root)
1934 {
1935         int ret;
1936         struct btrfs_fs_info *fs_info = root->fs_info;
1937
1938         spin_lock(&fs_info->trans_lock);
1939         if (list_empty(&fs_info->dead_roots)) {
1940                 spin_unlock(&fs_info->trans_lock);
1941                 return 0;
1942         }
1943         root = list_first_entry(&fs_info->dead_roots,
1944                         struct btrfs_root, root_list);
1945         list_del_init(&root->root_list);
1946         spin_unlock(&fs_info->trans_lock);
1947
1948         pr_debug("btrfs: cleaner removing %llu\n", root->objectid);
1949
1950         btrfs_kill_all_delayed_nodes(root);
1951
1952         if (btrfs_header_backref_rev(root->node) <
1953                         BTRFS_MIXED_BACKREF_REV)
1954                 ret = btrfs_drop_snapshot(root, NULL, 0, 0);
1955         else
1956                 ret = btrfs_drop_snapshot(root, NULL, 1, 0);
1957         /*
1958          * If we encounter a transaction abort during snapshot cleaning, we
1959          * don't want to crash here
1960          */
1961         return (ret < 0) ? 0 : 1;
1962 }