fs/btrfs/transaction.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/fs.h>
   7 #include <linux/slab.h>
   8 #include <linux/sched.h>
   9 #include <linux/sched/mm.h>
  10 #include <linux/writeback.h>
  11 #include <linux/pagemap.h>
  12 #include <linux/blkdev.h>
  13 #include <linux/uuid.h>
  14 #include <linux/timekeeping.h>
  15 #include "misc.h"
  16 #include "ctree.h"
  17 #include "disk-io.h"
  18 #include "transaction.h"
  19 #include "locking.h"
  20 #include "tree-log.h"
  21 #include "volumes.h"
  22 #include "dev-replace.h"
  23 #include "qgroup.h"
  24 #include "block-group.h"
  25 #include "space-info.h"
  26 #include "zoned.h"
  27 #include "fs.h"
  28 #include "accessors.h"
  29 #include "extent-tree.h"
  30 #include "root-tree.h"
  31 #include "defrag.h"
  32 #include "dir-item.h"
  33 #include "uuid-tree.h"
  34 #include "ioctl.h"
  35 #include "relocation.h"
  36 #include "scrub.h"
  37
  38 static struct kmem_cache *btrfs_trans_handle_cachep;
  39
  40 #define BTRFS_ROOT_TRANS_TAG 0
  41
  42 /*
  43  * Transaction states and transitions
  44  *
  45  * No running transaction (fs tree blocks are not modified)
  46  * |
  47  * | To next stage:
  48  * |  Call start_transaction() variants. Except btrfs_join_transaction_nostart().
  49  * V
  50  * Transaction N [[TRANS_STATE_RUNNING]]
  51  * |
  52  * | New trans handles can be attached to transaction N by calling all
  53  * | start_transaction() variants.
  54  * |
  55  * | To next stage:
  56  * |  Call btrfs_commit_transaction() on any trans handle attached to
  57  * |  transaction N
  58  * V
  59  * Transaction N [[TRANS_STATE_COMMIT_START]]
  60  * |
  61  * | Will wait for previous running transaction to completely finish if there
  62  * | is one
  63  * |
  64  * | Then one of the following happes:
  65  * | - Wait for all other trans handle holders to release.
  66  * |   The btrfs_commit_transaction() caller will do the commit work.
  67  * | - Wait for current transaction to be committed by others.
  68  * |   Other btrfs_commit_transaction() caller will do the commit work.
  69  * |
  70  * | At this stage, only btrfs_join_transaction*() variants can attach
  71  * | to this running transaction.
  72  * | All other variants will wait for current one to finish and attach to
  73  * | transaction N+1.
  74  * |
  75  * | To next stage:
  76  * |  Caller is chosen to commit transaction N, and all other trans handle
  77  * |  haven been released.
  78  * V
  79  * Transaction N [[TRANS_STATE_COMMIT_DOING]]
  80  * |
  81  * | The heavy lifting transaction work is started.
  82  * | From running delayed refs (modifying extent tree) to creating pending
  83  * | snapshots, running qgroups.
  84  * | In short, modify supporting trees to reflect modifications of subvolume
  85  * | trees.
  86  * |
  87  * | At this stage, all start_transaction() calls will wait for this
  88  * | transaction to finish and attach to transaction N+1.
  89  * |
  90  * | To next stage:
  91  * |  Until all supporting trees are updated.
  92  * V
  93  * Transaction N [[TRANS_STATE_UNBLOCKED]]
  94  * |                                                Transaction N+1
  95  * | All needed trees are modified, thus we only    [[TRANS_STATE_RUNNING]]
  96  * | need to write them back to disk and update     |
  97  * | super blocks.                                  |
  98  * |                                                |
  99  * | At this stage, new transaction is allowed to   |
 100  * | start.                                         |
 101  * | All new start_transaction() calls will be      |
 102  * | attached to transid N+1.                       |
 103  * |                                                |
 104  * | To next stage:                                 |
 105  * |  Until all tree blocks are super blocks are    |
 106  * |  written to block devices                      |
 107  * V                                                |
 108  * Transaction N [[TRANS_STATE_COMPLETED]]          V
 109  *   All tree blocks and super blocks are written.  Transaction N+1
 110  *   This transaction is finished and all its       [[TRANS_STATE_COMMIT_START]]
 111  *   data structures will be cleaned up.            | Life goes on
 112  */
 113 static const unsigned int btrfs_blocked_trans_types[TRANS_STATE_MAX] = {
 114         [TRANS_STATE_RUNNING]           = 0U,
 115         [TRANS_STATE_COMMIT_START]      = (__TRANS_START | __TRANS_ATTACH),
 116         [TRANS_STATE_COMMIT_DOING]      = (__TRANS_START |
 117                                            __TRANS_ATTACH |
 118                                            __TRANS_JOIN |
 119                                            __TRANS_JOIN_NOSTART),
 120         [TRANS_STATE_UNBLOCKED]         = (__TRANS_START |
 121                                            __TRANS_ATTACH |
 122                                            __TRANS_JOIN |
 123                                            __TRANS_JOIN_NOLOCK |
 124                                            __TRANS_JOIN_NOSTART),
 125         [TRANS_STATE_SUPER_COMMITTED]   = (__TRANS_START |
 126                                            __TRANS_ATTACH |
 127                                            __TRANS_JOIN |
 128                                            __TRANS_JOIN_NOLOCK |
 129                                            __TRANS_JOIN_NOSTART),
 130         [TRANS_STATE_COMPLETED]         = (__TRANS_START |
 131                                            __TRANS_ATTACH |
 132                                            __TRANS_JOIN |
 133                                            __TRANS_JOIN_NOLOCK |
 134                                            __TRANS_JOIN_NOSTART),
 135 };
 136
 137 void btrfs_put_transaction(struct btrfs_transaction *transaction)
 138 {
 139         WARN_ON(refcount_read(&transaction->use_count) == 0);
 140         if (refcount_dec_and_test(&transaction->use_count)) {
 141                 BUG_ON(!list_empty(&transaction->list));
 142                 WARN_ON(!RB_EMPTY_ROOT(
 143                                 &transaction->delayed_refs.href_root.rb_root));
 144                 WARN_ON(!RB_EMPTY_ROOT(
 145                                 &transaction->delayed_refs.dirty_extent_root));
 146                 if (transaction->delayed_refs.pending_csums)
 147                         btrfs_err(transaction->fs_info,
 148                                   "pending csums is %llu",
 149                                   transaction->delayed_refs.pending_csums);
 150                 /*
 151                  * If any block groups are found in ->deleted_bgs then it's
 152                  * because the transaction was aborted and a commit did not
 153                  * happen (things failed before writing the new superblock
 154                  * and calling btrfs_finish_extent_commit()), so we can not
 155                  * discard the physical locations of the block groups.
 156                  */
 157                 while (!list_empty(&transaction->deleted_bgs)) {
 158                         struct btrfs_block_group *cache;
 159
 160                         cache = list_first_entry(&transaction->deleted_bgs,
 161                                                  struct btrfs_block_group,
 162                                                  bg_list);
 163                         list_del_init(&cache->bg_list);
 164                         btrfs_unfreeze_block_group(cache);
 165                         btrfs_put_block_group(cache);
 166                 }
 167                 WARN_ON(!list_empty(&transaction->dev_update_list));
 168                 kfree(transaction);
 169         }
 170 }
 171
 172 static noinline void switch_commit_roots(struct btrfs_trans_handle *trans)
 173 {
 174         struct btrfs_transaction *cur_trans = trans->transaction;
 175         struct btrfs_fs_info *fs_info = trans->fs_info;
 176         struct btrfs_root *root, *tmp;
 177
 178         /*
 179          * At this point no one can be using this transaction to modify any tree
 180          * and no one can start another transaction to modify any tree either.
 181          */
 182         ASSERT(cur_trans->state == TRANS_STATE_COMMIT_DOING);
 183
 184         down_write(&fs_info->commit_root_sem);
 185
 186         if (test_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags))
 187                 fs_info->last_reloc_trans = trans->transid;
 188
 189         list_for_each_entry_safe(root, tmp, &cur_trans->switch_commits,
 190                                  dirty_list) {
 191                 list_del_init(&root->dirty_list);
 192                 free_extent_buffer(root->commit_root);
 193                 root->commit_root = btrfs_root_node(root);
 194                 extent_io_tree_release(&root->dirty_log_pages);
 195                 btrfs_qgroup_clean_swapped_blocks(root);
 196         }
 197
 198         /* We can free old roots now. */
 199         spin_lock(&cur_trans->dropped_roots_lock);
 200         while (!list_empty(&cur_trans->dropped_roots)) {
 201                 root = list_first_entry(&cur_trans->dropped_roots,
 202                                         struct btrfs_root, root_list);
 203                 list_del_init(&root->root_list);
 204                 spin_unlock(&cur_trans->dropped_roots_lock);
 205                 btrfs_free_log(trans, root);
 206                 btrfs_drop_and_free_fs_root(fs_info, root);
 207                 spin_lock(&cur_trans->dropped_roots_lock);
 208         }
 209         spin_unlock(&cur_trans->dropped_roots_lock);
 210
 211         up_write(&fs_info->commit_root_sem);
 212 }
 213
 214 static inline void extwriter_counter_inc(struct btrfs_transaction *trans,
 215                                          unsigned int type)
 216 {
 217         if (type & TRANS_EXTWRITERS)
 218                 atomic_inc(&trans->num_extwriters);
 219 }
 220
 221 static inline void extwriter_counter_dec(struct btrfs_transaction *trans,
 222                                          unsigned int type)
 223 {
 224         if (type & TRANS_EXTWRITERS)
 225                 atomic_dec(&trans->num_extwriters);
 226 }
 227
 228 static inline void extwriter_counter_init(struct btrfs_transaction *trans,
 229                                           unsigned int type)
 230 {
 231         atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0));
 232 }
 233
 234 static inline int extwriter_counter_read(struct btrfs_transaction *trans)
 235 {
 236         return atomic_read(&trans->num_extwriters);
 237 }
 238
 239 /*
 240  * To be called after doing the chunk btree updates right after allocating a new
 241  * chunk (after btrfs_chunk_alloc_add_chunk_item() is called), when removing a
 242  * chunk after all chunk btree updates and after finishing the second phase of
 243  * chunk allocation (btrfs_create_pending_block_groups()) in case some block
 244  * group had its chunk item insertion delayed to the second phase.
 245  */
 246 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
 247 {
 248         struct btrfs_fs_info *fs_info = trans->fs_info;
 249
 250         if (!trans->chunk_bytes_reserved)
 251                 return;
 252
 253         btrfs_block_rsv_release(fs_info, &fs_info->chunk_block_rsv,
 254                                 trans->chunk_bytes_reserved, NULL);
 255         trans->chunk_bytes_reserved = 0;
 256 }
 257
 258 /*
 259  * either allocate a new transaction or hop into the existing one
 260  */
 261 static noinline int join_transaction(struct btrfs_fs_info *fs_info,
 262                                      unsigned int type)
 263 {
 264         struct btrfs_transaction *cur_trans;
 265
 266         spin_lock(&fs_info->trans_lock);
 267 loop:
 268         /* The file system has been taken offline. No new transactions. */
 269         if (BTRFS_FS_ERROR(fs_info)) {
 270                 spin_unlock(&fs_info->trans_lock);
 271                 return -EROFS;
 272         }
 273
 274         cur_trans = fs_info->running_transaction;
 275         if (cur_trans) {
 276                 if (TRANS_ABORTED(cur_trans)) {
 277                         spin_unlock(&fs_info->trans_lock);
 278                         return cur_trans->aborted;
 279                 }
 280                 if (btrfs_blocked_trans_types[cur_trans->state] & type) {
 281                         spin_unlock(&fs_info->trans_lock);
 282                         return -EBUSY;
 283                 }
 284                 refcount_inc(&cur_trans->use_count);
 285                 atomic_inc(&cur_trans->num_writers);
 286                 extwriter_counter_inc(cur_trans, type);
 287                 spin_unlock(&fs_info->trans_lock);
 288                 btrfs_lockdep_acquire(fs_info, btrfs_trans_num_writers);
 289                 btrfs_lockdep_acquire(fs_info, btrfs_trans_num_extwriters);
 290                 return 0;
 291         }
 292         spin_unlock(&fs_info->trans_lock);
 293
 294         /*
 295          * If we are ATTACH or TRANS_JOIN_NOSTART, we just want to catch the
 296          * current transaction, and commit it. If there is no transaction, just
 297          * return ENOENT.
 298          */
 299         if (type == TRANS_ATTACH || type == TRANS_JOIN_NOSTART)
 300                 return -ENOENT;
 301
 302         /*
 303          * JOIN_NOLOCK only happens during the transaction commit, so
 304          * it is impossible that ->running_transaction is NULL
 305          */
 306         BUG_ON(type == TRANS_JOIN_NOLOCK);
 307
 308         cur_trans = kmalloc(sizeof(*cur_trans), GFP_NOFS);
 309         if (!cur_trans)
 310                 return -ENOMEM;
 311
 312         btrfs_lockdep_acquire(fs_info, btrfs_trans_num_writers);
 313         btrfs_lockdep_acquire(fs_info, btrfs_trans_num_extwriters);
 314
 315         spin_lock(&fs_info->trans_lock);
 316         if (fs_info->running_transaction) {
 317                 /*
 318                  * someone started a transaction after we unlocked.  Make sure
 319                  * to redo the checks above
 320                  */
 321                 btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters);
 322                 btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
 323                 kfree(cur_trans);
 324                 goto loop;
 325         } else if (BTRFS_FS_ERROR(fs_info)) {
 326                 spin_unlock(&fs_info->trans_lock);
 327                 btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters);
 328                 btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
 329                 kfree(cur_trans);
 330                 return -EROFS;
 331         }
 332
 333         cur_trans->fs_info = fs_info;
 334         atomic_set(&cur_trans->pending_ordered, 0);
 335         init_waitqueue_head(&cur_trans->pending_wait);
 336         atomic_set(&cur_trans->num_writers, 1);
 337         extwriter_counter_init(cur_trans, type);
 338         init_waitqueue_head(&cur_trans->writer_wait);
 339         init_waitqueue_head(&cur_trans->commit_wait);
 340         cur_trans->state = TRANS_STATE_RUNNING;
 341         /*
 342          * One for this trans handle, one so it will live on until we
 343          * commit the transaction.
 344          */
 345         refcount_set(&cur_trans->use_count, 2);
 346         cur_trans->flags = 0;
 347         cur_trans->start_time = ktime_get_seconds();
 348
 349         memset(&cur_trans->delayed_refs, 0, sizeof(cur_trans->delayed_refs));
 350
 351         cur_trans->delayed_refs.href_root = RB_ROOT_CACHED;
 352         cur_trans->delayed_refs.dirty_extent_root = RB_ROOT;
 353         atomic_set(&cur_trans->delayed_refs.num_entries, 0);
 354
 355         /*
 356          * although the tree mod log is per file system and not per transaction,
 357          * the log must never go across transaction boundaries.
 358          */
 359         smp_mb();
 360         if (!list_empty(&fs_info->tree_mod_seq_list))
 361                 WARN(1, KERN_ERR "BTRFS: tree_mod_seq_list not empty when creating a fresh transaction\n");
 362         if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
 363                 WARN(1, KERN_ERR "BTRFS: tree_mod_log rb tree not empty when creating a fresh transaction\n");
 364         atomic64_set(&fs_info->tree_mod_seq, 0);
 365
 366         spin_lock_init(&cur_trans->delayed_refs.lock);
 367
 368         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 369         INIT_LIST_HEAD(&cur_trans->dev_update_list);
 370         INIT_LIST_HEAD(&cur_trans->switch_commits);
 371         INIT_LIST_HEAD(&cur_trans->dirty_bgs);
 372         INIT_LIST_HEAD(&cur_trans->io_bgs);
 373         INIT_LIST_HEAD(&cur_trans->dropped_roots);
 374         mutex_init(&cur_trans->cache_write_mutex);
 375         spin_lock_init(&cur_trans->dirty_bgs_lock);
 376         INIT_LIST_HEAD(&cur_trans->deleted_bgs);
 377         spin_lock_init(&cur_trans->dropped_roots_lock);
 378         list_add_tail(&cur_trans->list, &fs_info->trans_list);
 379         extent_io_tree_init(fs_info, &cur_trans->dirty_pages,
 380                         IO_TREE_TRANS_DIRTY_PAGES);
 381         extent_io_tree_init(fs_info, &cur_trans->pinned_extents,
 382                         IO_TREE_FS_PINNED_EXTENTS);
 383         fs_info->generation++;
 384         cur_trans->transid = fs_info->generation;
 385         fs_info->running_transaction = cur_trans;
 386         cur_trans->aborted = 0;
 387         spin_unlock(&fs_info->trans_lock);
 388
 389         return 0;
 390 }
 391
 392 /*
 393  * This does all the record keeping required to make sure that a shareable root
 394  * is properly recorded in a given transaction.  This is required to make sure
 395  * the old root from before we joined the transaction is deleted when the
 396  * transaction commits.
 397  */
 398 static int record_root_in_trans(struct btrfs_trans_handle *trans,
 399                                struct btrfs_root *root,
 400                                int force)
 401 {
 402         struct btrfs_fs_info *fs_info = root->fs_info;
 403         int ret = 0;
 404
 405         if ((test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
 406             root->last_trans < trans->transid) || force) {
 407                 WARN_ON(!force && root->commit_root != root->node);
 408
 409                 /*
 410                  * see below for IN_TRANS_SETUP usage rules
 411                  * we have the reloc mutex held now, so there
 412                  * is only one writer in this function
 413                  */
 414                 set_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
 415
 416                 /* make sure readers find IN_TRANS_SETUP before
 417                  * they find our root->last_trans update
 418                  */
 419                 smp_wmb();
 420
 421                 spin_lock(&fs_info->fs_roots_radix_lock);
 422                 if (root->last_trans == trans->transid && !force) {
 423                         spin_unlock(&fs_info->fs_roots_radix_lock);
 424                         return 0;
 425                 }
 426                 radix_tree_tag_set(&fs_info->fs_roots_radix,
 427                                    (unsigned long)root->root_key.objectid,
 428                                    BTRFS_ROOT_TRANS_TAG);
 429                 spin_unlock(&fs_info->fs_roots_radix_lock);
 430                 root->last_trans = trans->transid;
 431
 432                 /* this is pretty tricky.  We don't want to
 433                  * take the relocation lock in btrfs_record_root_in_trans
 434                  * unless we're really doing the first setup for this root in
 435                  * this transaction.
 436                  *
 437                  * Normally we'd use root->last_trans as a flag to decide
 438                  * if we want to take the expensive mutex.
 439                  *
 440                  * But, we have to set root->last_trans before we
 441                  * init the relocation root, otherwise, we trip over warnings
 442                  * in ctree.c.  The solution used here is to flag ourselves
 443                  * with root IN_TRANS_SETUP.  When this is 1, we're still
 444                  * fixing up the reloc trees and everyone must wait.
 445                  *
 446                  * When this is zero, they can trust root->last_trans and fly
 447                  * through btrfs_record_root_in_trans without having to take the
 448                  * lock.  smp_wmb() makes sure that all the writes above are
 449                  * done before we pop in the zero below
 450                  */
 451                 ret = btrfs_init_reloc_root(trans, root);
 452                 smp_mb__before_atomic();
 453                 clear_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
 454         }
 455         return ret;
 456 }
 457
 458
 459 void btrfs_add_dropped_root(struct btrfs_trans_handle *trans,
 460                             struct btrfs_root *root)
 461 {
 462         struct btrfs_fs_info *fs_info = root->fs_info;
 463         struct btrfs_transaction *cur_trans = trans->transaction;
 464
 465         /* Add ourselves to the transaction dropped list */
 466         spin_lock(&cur_trans->dropped_roots_lock);
 467         list_add_tail(&root->root_list, &cur_trans->dropped_roots);
 468         spin_unlock(&cur_trans->dropped_roots_lock);
 469
 470         /* Make sure we don't try to update the root at commit time */
 471         spin_lock(&fs_info->fs_roots_radix_lock);
 472         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 473                              (unsigned long)root->root_key.objectid,
 474                              BTRFS_ROOT_TRANS_TAG);
 475         spin_unlock(&fs_info->fs_roots_radix_lock);
 476 }
 477
 478 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 479                                struct btrfs_root *root)
 480 {
 481         struct btrfs_fs_info *fs_info = root->fs_info;
 482         int ret;
 483
 484         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
 485                 return 0;
 486
 487         /*
 488          * see record_root_in_trans for comments about IN_TRANS_SETUP usage
 489          * and barriers
 490          */
 491         smp_rmb();
 492         if (root->last_trans == trans->transid &&
 493             !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state))
 494                 return 0;
 495
 496         mutex_lock(&fs_info->reloc_mutex);
 497         ret = record_root_in_trans(trans, root, 0);
 498         mutex_unlock(&fs_info->reloc_mutex);
 499
 500         return ret;
 501 }
 502
 503 static inline int is_transaction_blocked(struct btrfs_transaction *trans)
 504 {
 505         return (trans->state >= TRANS_STATE_COMMIT_START &&
 506                 trans->state < TRANS_STATE_UNBLOCKED &&
 507                 !TRANS_ABORTED(trans));
 508 }
 509
 510 /* wait for commit against the current transaction to become unblocked
 511  * when this is done, it is safe to start a new transaction, but the current
 512  * transaction might not be fully on disk.
 513  */
 514 static void wait_current_trans(struct btrfs_fs_info *fs_info)
 515 {
 516         struct btrfs_transaction *cur_trans;
 517
 518         spin_lock(&fs_info->trans_lock);
 519         cur_trans = fs_info->running_transaction;
 520         if (cur_trans && is_transaction_blocked(cur_trans)) {
 521                 refcount_inc(&cur_trans->use_count);
 522                 spin_unlock(&fs_info->trans_lock);
 523
 524                 btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
 525                 wait_event(fs_info->transaction_wait,
 526                            cur_trans->state >= TRANS_STATE_UNBLOCKED ||
 527                            TRANS_ABORTED(cur_trans));
 528                 btrfs_put_transaction(cur_trans);
 529         } else {
 530                 spin_unlock(&fs_info->trans_lock);
 531         }
 532 }
 533
 534 static int may_wait_transaction(struct btrfs_fs_info *fs_info, int type)
 535 {
 536         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
 537                 return 0;
 538
 539         if (type == TRANS_START)
 540                 return 1;
 541
 542         return 0;
 543 }
 544
 545 static inline bool need_reserve_reloc_root(struct btrfs_root *root)
 546 {
 547         struct btrfs_fs_info *fs_info = root->fs_info;
 548
 549         if (!fs_info->reloc_ctl ||
 550             !test_bit(BTRFS_ROOT_SHAREABLE, &root->state) ||
 551             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 552             root->reloc_root)
 553                 return false;
 554
 555         return true;
 556 }
 557
 558 static struct btrfs_trans_handle *
 559 start_transaction(struct btrfs_root *root, unsigned int num_items,
 560                   unsigned int type, enum btrfs_reserve_flush_enum flush,
 561                   bool enforce_qgroups)
 562 {
 563         struct btrfs_fs_info *fs_info = root->fs_info;
 564         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
 565         struct btrfs_trans_handle *h;
 566         struct btrfs_transaction *cur_trans;
 567         u64 num_bytes = 0;
 568         u64 qgroup_reserved = 0;
 569         bool reloc_reserved = false;
 570         bool do_chunk_alloc = false;
 571         int ret;
 572
 573         if (BTRFS_FS_ERROR(fs_info))
 574                 return ERR_PTR(-EROFS);
 575
 576         if (current->journal_info) {
 577                 WARN_ON(type & TRANS_EXTWRITERS);
 578                 h = current->journal_info;
 579                 refcount_inc(&h->use_count);
 580                 WARN_ON(refcount_read(&h->use_count) > 2);
 581                 h->orig_rsv = h->block_rsv;
 582                 h->block_rsv = NULL;
 583                 goto got_it;
 584         }
 585
 586         /*
 587          * Do the reservation before we join the transaction so we can do all
 588          * the appropriate flushing if need be.
 589          */
 590         if (num_items && root != fs_info->chunk_root) {
 591                 struct btrfs_block_rsv *rsv = &fs_info->trans_block_rsv;
 592                 u64 delayed_refs_bytes = 0;
 593
 594                 qgroup_reserved = num_items * fs_info->nodesize;
 595                 /*
 596                  * Use prealloc for now, as there might be a currently running
 597                  * transaction that could free this reserved space prematurely
 598                  * by committing.
 599                  */
 600                 ret = btrfs_qgroup_reserve_meta_prealloc(root, qgroup_reserved,
 601                                                          enforce_qgroups, false);
 602                 if (ret)
 603                         return ERR_PTR(ret);
 604
 605                 /*
 606                  * We want to reserve all the bytes we may need all at once, so
 607                  * we only do 1 enospc flushing cycle per transaction start.  We
 608                  * accomplish this by simply assuming we'll do num_items worth
 609                  * of delayed refs updates in this trans handle, and refill that
 610                  * amount for whatever is missing in the reserve.
 611                  */
 612                 num_bytes = btrfs_calc_insert_metadata_size(fs_info, num_items);
 613                 if (flush == BTRFS_RESERVE_FLUSH_ALL &&
 614                     !btrfs_block_rsv_full(delayed_refs_rsv)) {
 615                         delayed_refs_bytes = btrfs_calc_delayed_ref_bytes(fs_info,
 616                                                                           num_items);
 617                         num_bytes += delayed_refs_bytes;
 618                 }
 619
 620                 /*
 621                  * Do the reservation for the relocation root creation
 622                  */
 623                 if (need_reserve_reloc_root(root)) {
 624                         num_bytes += fs_info->nodesize;
 625                         reloc_reserved = true;
 626                 }
 627
 628                 ret = btrfs_block_rsv_add(fs_info, rsv, num_bytes, flush);
 629                 if (ret)
 630                         goto reserve_fail;
 631                 if (delayed_refs_bytes) {
 632                         btrfs_migrate_to_delayed_refs_rsv(fs_info, rsv,
 633                                                           delayed_refs_bytes);
 634                         num_bytes -= delayed_refs_bytes;
 635                 }
 636
 637                 if (rsv->space_info->force_alloc)
 638                         do_chunk_alloc = true;
 639         } else if (num_items == 0 && flush == BTRFS_RESERVE_FLUSH_ALL &&
 640                    !btrfs_block_rsv_full(delayed_refs_rsv)) {
 641                 /*
 642                  * Some people call with btrfs_start_transaction(root, 0)
 643                  * because they can be throttled, but have some other mechanism
 644                  * for reserving space.  We still want these guys to refill the
 645                  * delayed block_rsv so just add 1 items worth of reservation
 646                  * here.
 647                  */
 648                 ret = btrfs_delayed_refs_rsv_refill(fs_info, flush);
 649                 if (ret)
 650                         goto reserve_fail;
 651         }
 652 again:
 653         h = kmem_cache_zalloc(btrfs_trans_handle_cachep, GFP_NOFS);
 654         if (!h) {
 655                 ret = -ENOMEM;
 656                 goto alloc_fail;
 657         }
 658
 659         /*
 660          * If we are JOIN_NOLOCK we're already committing a transaction and
 661          * waiting on this guy, so we don't need to do the sb_start_intwrite
 662          * because we're already holding a ref.  We need this because we could
 663          * have raced in and did an fsync() on a file which can kick a commit
 664          * and then we deadlock with somebody doing a freeze.
 665          *
 666          * If we are ATTACH, it means we just want to catch the current
 667          * transaction and commit it, so we needn't do sb_start_intwrite().
 668          */
 669         if (type & __TRANS_FREEZABLE)
 670                 sb_start_intwrite(fs_info->sb);
 671
 672         if (may_wait_transaction(fs_info, type))
 673                 wait_current_trans(fs_info);
 674
 675         do {
 676                 ret = join_transaction(fs_info, type);
 677                 if (ret == -EBUSY) {
 678                         wait_current_trans(fs_info);
 679                         if (unlikely(type == TRANS_ATTACH ||
 680                                      type == TRANS_JOIN_NOSTART))
 681                                 ret = -ENOENT;
 682                 }
 683         } while (ret == -EBUSY);
 684
 685         if (ret < 0)
 686                 goto join_fail;
 687
 688         cur_trans = fs_info->running_transaction;
 689
 690         h->transid = cur_trans->transid;
 691         h->transaction = cur_trans;
 692         refcount_set(&h->use_count, 1);
 693         h->fs_info = root->fs_info;
 694
 695         h->type = type;
 696         INIT_LIST_HEAD(&h->new_bgs);
 697
 698         smp_mb();
 699         if (cur_trans->state >= TRANS_STATE_COMMIT_START &&
 700             may_wait_transaction(fs_info, type)) {
 701                 current->journal_info = h;
 702                 btrfs_commit_transaction(h);
 703                 goto again;
 704         }
 705
 706         if (num_bytes) {
 707                 trace_btrfs_space_reservation(fs_info, "transaction",
 708                                               h->transid, num_bytes, 1);
 709                 h->block_rsv = &fs_info->trans_block_rsv;
 710                 h->bytes_reserved = num_bytes;
 711                 h->reloc_reserved = reloc_reserved;
 712         }
 713
 714         /*
 715          * Now that we have found a transaction to be a part of, convert the
 716          * qgroup reservation from prealloc to pertrans. A different transaction
 717          * can't race in and free our pertrans out from under us.
 718          */
 719         if (qgroup_reserved)
 720                 btrfs_qgroup_convert_reserved_meta(root, qgroup_reserved);
 721
 722 got_it:
 723         if (!current->journal_info)
 724                 current->journal_info = h;
 725
 726         /*
 727          * If the space_info is marked ALLOC_FORCE then we'll get upgraded to
 728          * ALLOC_FORCE the first run through, and then we won't allocate for
 729          * anybody else who races in later.  We don't care about the return
 730          * value here.
 731          */
 732         if (do_chunk_alloc && num_bytes) {
 733                 u64 flags = h->block_rsv->space_info->flags;
 734
 735                 btrfs_chunk_alloc(h, btrfs_get_alloc_profile(fs_info, flags),
 736                                   CHUNK_ALLOC_NO_FORCE);
 737         }
 738
 739         /*
 740          * btrfs_record_root_in_trans() needs to alloc new extents, and may
 741          * call btrfs_join_transaction() while we're also starting a
 742          * transaction.
 743          *
 744          * Thus it need to be called after current->journal_info initialized,
 745          * or we can deadlock.
 746          */
 747         ret = btrfs_record_root_in_trans(h, root);
 748         if (ret) {
 749                 /*
 750                  * The transaction handle is fully initialized and linked with
 751                  * other structures so it needs to be ended in case of errors,
 752                  * not just freed.
 753                  */
 754                 btrfs_end_transaction(h);
 755                 return ERR_PTR(ret);
 756         }
 757
 758         return h;
 759
 760 join_fail:
 761         if (type & __TRANS_FREEZABLE)
 762                 sb_end_intwrite(fs_info->sb);
 763         kmem_cache_free(btrfs_trans_handle_cachep, h);
 764 alloc_fail:
 765         if (num_bytes)
 766                 btrfs_block_rsv_release(fs_info, &fs_info->trans_block_rsv,
 767                                         num_bytes, NULL);
 768 reserve_fail:
 769         btrfs_qgroup_free_meta_prealloc(root, qgroup_reserved);
 770         return ERR_PTR(ret);
 771 }
 772
 773 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 774                                                    unsigned int num_items)
 775 {
 776         return start_transaction(root, num_items, TRANS_START,
 777                                  BTRFS_RESERVE_FLUSH_ALL, true);
 778 }
 779
 780 struct btrfs_trans_handle *btrfs_start_transaction_fallback_global_rsv(
 781                                         struct btrfs_root *root,
 782                                         unsigned int num_items)
 783 {
 784         return start_transaction(root, num_items, TRANS_START,
 785                                  BTRFS_RESERVE_FLUSH_ALL_STEAL, false);
 786 }
 787
 788 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 789 {
 790         return start_transaction(root, 0, TRANS_JOIN, BTRFS_RESERVE_NO_FLUSH,
 791                                  true);
 792 }
 793
 794 struct btrfs_trans_handle *btrfs_join_transaction_spacecache(struct btrfs_root *root)
 795 {
 796         return start_transaction(root, 0, TRANS_JOIN_NOLOCK,
 797                                  BTRFS_RESERVE_NO_FLUSH, true);
 798 }
 799
 800 /*
 801  * Similar to regular join but it never starts a transaction when none is
 802  * running or when there's a running one at a state >= TRANS_STATE_UNBLOCKED.
 803  * This is similar to btrfs_attach_transaction() but it allows the join to
 804  * happen if the transaction commit already started but it's not yet in the
 805  * "doing" phase (the state is < TRANS_STATE_COMMIT_DOING).
 806  */
 807 struct btrfs_trans_handle *btrfs_join_transaction_nostart(struct btrfs_root *root)
 808 {
 809         return start_transaction(root, 0, TRANS_JOIN_NOSTART,
 810                                  BTRFS_RESERVE_NO_FLUSH, true);
 811 }
 812
 813 /*
 814  * btrfs_attach_transaction() - catch the running transaction
 815  *
 816  * It is used when we want to commit the current the transaction, but
 817  * don't want to start a new one.
 818  *
 819  * Note: If this function return -ENOENT, it just means there is no
 820  * running transaction. But it is possible that the inactive transaction
 821  * is still in the memory, not fully on disk. If you hope there is no
 822  * inactive transaction in the fs when -ENOENT is returned, you should
 823  * invoke
 824  *     btrfs_attach_transaction_barrier()
 825  */
 826 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
 827 {
 828         return start_transaction(root, 0, TRANS_ATTACH,
 829                                  BTRFS_RESERVE_NO_FLUSH, true);
 830 }
 831
 832 /*
 833  * btrfs_attach_transaction_barrier() - catch the running transaction
 834  *
 835  * It is similar to the above function, the difference is this one
 836  * will wait for all the inactive transactions until they fully
 837  * complete.
 838  */
 839 struct btrfs_trans_handle *
 840 btrfs_attach_transaction_barrier(struct btrfs_root *root)
 841 {
 842         struct btrfs_trans_handle *trans;
 843
 844         trans = start_transaction(root, 0, TRANS_ATTACH,
 845                                   BTRFS_RESERVE_NO_FLUSH, true);
 846         if (trans == ERR_PTR(-ENOENT)) {
 847                 int ret;
 848
 849                 ret = btrfs_wait_for_commit(root->fs_info, 0);
 850                 if (ret)
 851                         return ERR_PTR(ret);
 852         }
 853
 854         return trans;
 855 }
 856
 857 /* Wait for a transaction commit to reach at least the given state. */
 858 static noinline void wait_for_commit(struct btrfs_transaction *commit,
 859                                      const enum btrfs_trans_state min_state)
 860 {
 861         struct btrfs_fs_info *fs_info = commit->fs_info;
 862         u64 transid = commit->transid;
 863         bool put = false;
 864
 865         /*
 866          * At the moment this function is called with min_state either being
 867          * TRANS_STATE_COMPLETED or TRANS_STATE_SUPER_COMMITTED.
 868          */
 869         if (min_state == TRANS_STATE_COMPLETED)
 870                 btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED);
 871         else
 872                 btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED);
 873
 874         while (1) {
 875                 wait_event(commit->commit_wait, commit->state >= min_state);
 876                 if (put)
 877                         btrfs_put_transaction(commit);
 878
 879                 if (min_state < TRANS_STATE_COMPLETED)
 880                         break;
 881
 882                 /*
 883                  * A transaction isn't really completed until all of the
 884                  * previous transactions are completed, but with fsync we can
 885                  * end up with SUPER_COMMITTED transactions before a COMPLETED
 886                  * transaction. Wait for those.
 887                  */
 888
 889                 spin_lock(&fs_info->trans_lock);
 890                 commit = list_first_entry_or_null(&fs_info->trans_list,
 891                                                   struct btrfs_transaction,
 892                                                   list);
 893                 if (!commit || commit->transid > transid) {
 894                         spin_unlock(&fs_info->trans_lock);
 895                         break;
 896                 }
 897                 refcount_inc(&commit->use_count);
 898                 put = true;
 899                 spin_unlock(&fs_info->trans_lock);
 900         }
 901 }
 902
 903 int btrfs_wait_for_commit(struct btrfs_fs_info *fs_info, u64 transid)
 904 {
 905         struct btrfs_transaction *cur_trans = NULL, *t;
 906         int ret = 0;
 907
 908         if (transid) {
 909                 if (transid <= fs_info->last_trans_committed)
 910                         goto out;
 911
 912                 /* find specified transaction */
 913                 spin_lock(&fs_info->trans_lock);
 914                 list_for_each_entry(t, &fs_info->trans_list, list) {
 915                         if (t->transid == transid) {
 916                                 cur_trans = t;
 917                                 refcount_inc(&cur_trans->use_count);
 918                                 ret = 0;
 919                                 break;
 920                         }
 921                         if (t->transid > transid) {
 922                                 ret = 0;
 923                                 break;
 924                         }
 925                 }
 926                 spin_unlock(&fs_info->trans_lock);
 927
 928                 /*
 929                  * The specified transaction doesn't exist, or we
 930                  * raced with btrfs_commit_transaction
 931                  */
 932                 if (!cur_trans) {
 933                         if (transid > fs_info->last_trans_committed)
 934                                 ret = -EINVAL;
 935                         goto out;
 936                 }
 937         } else {
 938                 /* find newest transaction that is committing | committed */
 939                 spin_lock(&fs_info->trans_lock);
 940                 list_for_each_entry_reverse(t, &fs_info->trans_list,
 941                                             list) {
 942                         if (t->state >= TRANS_STATE_COMMIT_START) {
 943                                 if (t->state == TRANS_STATE_COMPLETED)
 944                                         break;
 945                                 cur_trans = t;
 946                                 refcount_inc(&cur_trans->use_count);
 947                                 break;
 948                         }
 949                 }
 950                 spin_unlock(&fs_info->trans_lock);
 951                 if (!cur_trans)
 952                         goto out;  /* nothing committing|committed */
 953         }
 954
 955         wait_for_commit(cur_trans, TRANS_STATE_COMPLETED);
 956         ret = cur_trans->aborted;
 957         btrfs_put_transaction(cur_trans);
 958 out:
 959         return ret;
 960 }
 961
 962 void btrfs_throttle(struct btrfs_fs_info *fs_info)
 963 {
 964         wait_current_trans(fs_info);
 965 }
 966
 967 bool btrfs_should_end_transaction(struct btrfs_trans_handle *trans)
 968 {
 969         struct btrfs_transaction *cur_trans = trans->transaction;
 970
 971         if (cur_trans->state >= TRANS_STATE_COMMIT_START ||
 972             test_bit(BTRFS_DELAYED_REFS_FLUSHING, &cur_trans->delayed_refs.flags))
 973                 return true;
 974
 975         if (btrfs_check_space_for_delayed_refs(trans->fs_info))
 976                 return true;
 977
 978         return !!btrfs_block_rsv_check(&trans->fs_info->global_block_rsv, 50);
 979 }
 980
 981 static void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans)
 982
 983 {
 984         struct btrfs_fs_info *fs_info = trans->fs_info;
 985
 986         if (!trans->block_rsv) {
 987                 ASSERT(!trans->bytes_reserved);
 988                 return;
 989         }
 990
 991         if (!trans->bytes_reserved)
 992                 return;
 993
 994         ASSERT(trans->block_rsv == &fs_info->trans_block_rsv);
 995         trace_btrfs_space_reservation(fs_info, "transaction",
 996                                       trans->transid, trans->bytes_reserved, 0);
 997         btrfs_block_rsv_release(fs_info, trans->block_rsv,
 998                                 trans->bytes_reserved, NULL);
 999         trans->bytes_reserved = 0;
1000 }
1001
1002 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
1003                                    int throttle)
1004 {
1005         struct btrfs_fs_info *info = trans->fs_info;
1006         struct btrfs_transaction *cur_trans = trans->transaction;
1007         int err = 0;
1008
1009         if (refcount_read(&trans->use_count) > 1) {
1010                 refcount_dec(&trans->use_count);
1011                 trans->block_rsv = trans->orig_rsv;
1012                 return 0;
1013         }
1014
1015         btrfs_trans_release_metadata(trans);
1016         trans->block_rsv = NULL;
1017
1018         btrfs_create_pending_block_groups(trans);
1019
1020         btrfs_trans_release_chunk_metadata(trans);
1021
1022         if (trans->type & __TRANS_FREEZABLE)
1023                 sb_end_intwrite(info->sb);
1024
1025         WARN_ON(cur_trans != info->running_transaction);
1026         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
1027         atomic_dec(&cur_trans->num_writers);
1028         extwriter_counter_dec(cur_trans, trans->type);
1029
1030         cond_wake_up(&cur_trans->writer_wait);
1031
1032         btrfs_lockdep_release(info, btrfs_trans_num_extwriters);
1033         btrfs_lockdep_release(info, btrfs_trans_num_writers);
1034
1035         btrfs_put_transaction(cur_trans);
1036
1037         if (current->journal_info == trans)
1038                 current->journal_info = NULL;
1039
1040         if (throttle)
1041                 btrfs_run_delayed_iputs(info);
1042
1043         if (TRANS_ABORTED(trans) || BTRFS_FS_ERROR(info)) {
1044                 wake_up_process(info->transaction_kthread);
1045                 if (TRANS_ABORTED(trans))
1046                         err = trans->aborted;
1047                 else
1048                         err = -EROFS;
1049         }
1050
1051         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1052         return err;
1053 }
1054
1055 int btrfs_end_transaction(struct btrfs_trans_handle *trans)
1056 {
1057         return __btrfs_end_transaction(trans, 0);
1058 }
1059
1060 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans)
1061 {
1062         return __btrfs_end_transaction(trans, 1);
1063 }
1064
1065 /*
1066  * when btree blocks are allocated, they have some corresponding bits set for
1067  * them in one of two extent_io trees.  This is used to make sure all of
1068  * those extents are sent to disk but does not wait on them
1069  */
1070 int btrfs_write_marked_extents(struct btrfs_fs_info *fs_info,
1071                                struct extent_io_tree *dirty_pages, int mark)
1072 {
1073         int err = 0;
1074         int werr = 0;
1075         struct address_space *mapping = fs_info->btree_inode->i_mapping;
1076         struct extent_state *cached_state = NULL;
1077         u64 start = 0;
1078         u64 end;
1079
1080         while (find_first_extent_bit(dirty_pages, start, &start, &end,
1081                                      mark, &cached_state)) {
1082                 bool wait_writeback = false;
1083
1084                 err = convert_extent_bit(dirty_pages, start, end,
1085                                          EXTENT_NEED_WAIT,
1086                                          mark, &cached_state);
1087                 /*
1088                  * convert_extent_bit can return -ENOMEM, which is most of the
1089                  * time a temporary error. So when it happens, ignore the error
1090                  * and wait for writeback of this range to finish - because we
1091                  * failed to set the bit EXTENT_NEED_WAIT for the range, a call
1092                  * to __btrfs_wait_marked_extents() would not know that
1093                  * writeback for this range started and therefore wouldn't
1094                  * wait for it to finish - we don't want to commit a
1095                  * superblock that points to btree nodes/leafs for which
1096                  * writeback hasn't finished yet (and without errors).
1097                  * We cleanup any entries left in the io tree when committing
1098                  * the transaction (through extent_io_tree_release()).
1099                  */
1100                 if (err == -ENOMEM) {
1101                         err = 0;
1102                         wait_writeback = true;
1103                 }
1104                 if (!err)
1105                         err = filemap_fdatawrite_range(mapping, start, end);
1106                 if (err)
1107                         werr = err;
1108                 else if (wait_writeback)
1109                         werr = filemap_fdatawait_range(mapping, start, end);
1110                 free_extent_state(cached_state);
1111                 cached_state = NULL;
1112                 cond_resched();
1113                 start = end + 1;
1114         }
1115         return werr;
1116 }
1117
1118 /*
1119  * when btree blocks are allocated, they have some corresponding bits set for
1120  * them in one of two extent_io trees.  This is used to make sure all of
1121  * those extents are on disk for transaction or log commit.  We wait
1122  * on all the pages and clear them from the dirty pages state tree
1123  */
1124 static int __btrfs_wait_marked_extents(struct btrfs_fs_info *fs_info,
1125                                        struct extent_io_tree *dirty_pages)
1126 {
1127         int err = 0;
1128         int werr = 0;
1129         struct address_space *mapping = fs_info->btree_inode->i_mapping;
1130         struct extent_state *cached_state = NULL;
1131         u64 start = 0;
1132         u64 end;
1133
1134         while (find_first_extent_bit(dirty_pages, start, &start, &end,
1135                                      EXTENT_NEED_WAIT, &cached_state)) {
1136                 /*
1137                  * Ignore -ENOMEM errors returned by clear_extent_bit().
1138                  * When committing the transaction, we'll remove any entries
1139                  * left in the io tree. For a log commit, we don't remove them
1140                  * after committing the log because the tree can be accessed
1141                  * concurrently - we do it only at transaction commit time when
1142                  * it's safe to do it (through extent_io_tree_release()).
1143                  */
1144                 err = clear_extent_bit(dirty_pages, start, end,
1145                                        EXTENT_NEED_WAIT, &cached_state);
1146                 if (err == -ENOMEM)
1147                         err = 0;
1148                 if (!err)
1149                         err = filemap_fdatawait_range(mapping, start, end);
1150                 if (err)
1151                         werr = err;
1152                 free_extent_state(cached_state);
1153                 cached_state = NULL;
1154                 cond_resched();
1155                 start = end + 1;
1156         }
1157         if (err)
1158                 werr = err;
1159         return werr;
1160 }
1161
1162 static int btrfs_wait_extents(struct btrfs_fs_info *fs_info,
1163                        struct extent_io_tree *dirty_pages)
1164 {
1165         bool errors = false;
1166         int err;
1167
1168         err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1169         if (test_and_clear_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags))
1170                 errors = true;
1171
1172         if (errors && !err)
1173                 err = -EIO;
1174         return err;
1175 }
1176
1177 int btrfs_wait_tree_log_extents(struct btrfs_root *log_root, int mark)
1178 {
1179         struct btrfs_fs_info *fs_info = log_root->fs_info;
1180         struct extent_io_tree *dirty_pages = &log_root->dirty_log_pages;
1181         bool errors = false;
1182         int err;
1183
1184         ASSERT(log_root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
1185
1186         err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1187         if ((mark & EXTENT_DIRTY) &&
1188             test_and_clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags))
1189                 errors = true;
1190
1191         if ((mark & EXTENT_NEW) &&
1192             test_and_clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags))
1193                 errors = true;
1194
1195         if (errors && !err)
1196                 err = -EIO;
1197         return err;
1198 }
1199
1200 /*
1201  * When btree blocks are allocated the corresponding extents are marked dirty.
1202  * This function ensures such extents are persisted on disk for transaction or
1203  * log commit.
1204  *
1205  * @trans: transaction whose dirty pages we'd like to write
1206  */
1207 static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans)
1208 {
1209         int ret;
1210         int ret2;
1211         struct extent_io_tree *dirty_pages = &trans->transaction->dirty_pages;
1212         struct btrfs_fs_info *fs_info = trans->fs_info;
1213         struct blk_plug plug;
1214
1215         blk_start_plug(&plug);
1216         ret = btrfs_write_marked_extents(fs_info, dirty_pages, EXTENT_DIRTY);
1217         blk_finish_plug(&plug);
1218         ret2 = btrfs_wait_extents(fs_info, dirty_pages);
1219
1220         extent_io_tree_release(&trans->transaction->dirty_pages);
1221
1222         if (ret)
1223                 return ret;
1224         else if (ret2)
1225                 return ret2;
1226         else
1227                 return 0;
1228 }
1229
1230 /*
1231  * this is used to update the root pointer in the tree of tree roots.
1232  *
1233  * But, in the case of the extent allocation tree, updating the root
1234  * pointer may allocate blocks which may change the root of the extent
1235  * allocation tree.
1236  *
1237  * So, this loops and repeats and makes sure the cowonly root didn't
1238  * change while the root pointer was being updated in the metadata.
1239  */
1240 static int update_cowonly_root(struct btrfs_trans_handle *trans,
1241                                struct btrfs_root *root)
1242 {
1243         int ret;
1244         u64 old_root_bytenr;
1245         u64 old_root_used;
1246         struct btrfs_fs_info *fs_info = root->fs_info;
1247         struct btrfs_root *tree_root = fs_info->tree_root;
1248
1249         old_root_used = btrfs_root_used(&root->root_item);
1250
1251         while (1) {
1252                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1253                 if (old_root_bytenr == root->node->start &&
1254                     old_root_used == btrfs_root_used(&root->root_item))
1255                         break;
1256
1257                 btrfs_set_root_node(&root->root_item, root->node);
1258                 ret = btrfs_update_root(trans, tree_root,
1259                                         &root->root_key,
1260                                         &root->root_item);
1261                 if (ret)
1262                         return ret;
1263
1264                 old_root_used = btrfs_root_used(&root->root_item);
1265         }
1266
1267         return 0;
1268 }
1269
1270 /*
1271  * update all the cowonly tree roots on disk
1272  *
1273  * The error handling in this function may not be obvious. Any of the
1274  * failures will cause the file system to go offline. We still need
1275  * to clean up the delayed refs.
1276  */
1277 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans)
1278 {
1279         struct btrfs_fs_info *fs_info = trans->fs_info;
1280         struct list_head *dirty_bgs = &trans->transaction->dirty_bgs;
1281         struct list_head *io_bgs = &trans->transaction->io_bgs;
1282         struct list_head *next;
1283         struct extent_buffer *eb;
1284         int ret;
1285
1286         /*
1287          * At this point no one can be using this transaction to modify any tree
1288          * and no one can start another transaction to modify any tree either.
1289          */
1290         ASSERT(trans->transaction->state == TRANS_STATE_COMMIT_DOING);
1291
1292         eb = btrfs_lock_root_node(fs_info->tree_root);
1293         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
1294                               0, &eb, BTRFS_NESTING_COW);
1295         btrfs_tree_unlock(eb);
1296         free_extent_buffer(eb);
1297
1298         if (ret)
1299                 return ret;
1300
1301         ret = btrfs_run_dev_stats(trans);
1302         if (ret)
1303                 return ret;
1304         ret = btrfs_run_dev_replace(trans);
1305         if (ret)
1306                 return ret;
1307         ret = btrfs_run_qgroups(trans);
1308         if (ret)
1309                 return ret;
1310
1311         ret = btrfs_setup_space_cache(trans);
1312         if (ret)
1313                 return ret;
1314
1315 again:
1316         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1317                 struct btrfs_root *root;
1318                 next = fs_info->dirty_cowonly_roots.next;
1319                 list_del_init(next);
1320                 root = list_entry(next, struct btrfs_root, dirty_list);
1321                 clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1322
1323                 list_add_tail(&root->dirty_list,
1324                               &trans->transaction->switch_commits);
1325                 ret = update_cowonly_root(trans, root);
1326                 if (ret)
1327                         return ret;
1328         }
1329
1330         /* Now flush any delayed refs generated by updating all of the roots */
1331         ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1332         if (ret)
1333                 return ret;
1334
1335         while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1336                 ret = btrfs_write_dirty_block_groups(trans);
1337                 if (ret)
1338                         return ret;
1339
1340                 /*
1341                  * We're writing the dirty block groups, which could generate
1342                  * delayed refs, which could generate more dirty block groups,
1343                  * so we want to keep this flushing in this loop to make sure
1344                  * everything gets run.
1345                  */
1346                 ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1347                 if (ret)
1348                         return ret;
1349         }
1350
1351         if (!list_empty(&fs_info->dirty_cowonly_roots))
1352                 goto again;
1353
1354         /* Update dev-replace pointer once everything is committed */
1355         fs_info->dev_replace.committed_cursor_left =
1356                 fs_info->dev_replace.cursor_left_last_write_of_item;
1357
1358         return 0;
1359 }
1360
1361 /*
1362  * If we had a pending drop we need to see if there are any others left in our
1363  * dead roots list, and if not clear our bit and wake any waiters.
1364  */
1365 void btrfs_maybe_wake_unfinished_drop(struct btrfs_fs_info *fs_info)
1366 {
1367         /*
1368          * We put the drop in progress roots at the front of the list, so if the
1369          * first entry doesn't have UNFINISHED_DROP set we can wake everybody
1370          * up.
1371          */
1372         spin_lock(&fs_info->trans_lock);
1373         if (!list_empty(&fs_info->dead_roots)) {
1374                 struct btrfs_root *root = list_first_entry(&fs_info->dead_roots,
1375                                                            struct btrfs_root,
1376                                                            root_list);
1377                 if (test_bit(BTRFS_ROOT_UNFINISHED_DROP, &root->state)) {
1378                         spin_unlock(&fs_info->trans_lock);
1379                         return;
1380                 }
1381         }
1382         spin_unlock(&fs_info->trans_lock);
1383
1384         btrfs_wake_unfinished_drop(fs_info);
1385 }
1386
1387 /*
1388  * dead roots are old snapshots that need to be deleted.  This allocates
1389  * a dirty root struct and adds it into the list of dead roots that need to
1390  * be deleted
1391  */
1392 void btrfs_add_dead_root(struct btrfs_root *root)
1393 {
1394         struct btrfs_fs_info *fs_info = root->fs_info;
1395
1396         spin_lock(&fs_info->trans_lock);
1397         if (list_empty(&root->root_list)) {
1398                 btrfs_grab_root(root);
1399
1400                 /* We want to process the partially complete drops first. */
1401                 if (test_bit(BTRFS_ROOT_UNFINISHED_DROP, &root->state))
1402                         list_add(&root->root_list, &fs_info->dead_roots);
1403                 else
1404                         list_add_tail(&root->root_list, &fs_info->dead_roots);
1405         }
1406         spin_unlock(&fs_info->trans_lock);
1407 }
1408
1409 /*
1410  * Update each subvolume root and its relocation root, if it exists, in the tree
1411  * of tree roots. Also free log roots if they exist.
1412  */
1413 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans)
1414 {
1415         struct btrfs_fs_info *fs_info = trans->fs_info;
1416         struct btrfs_root *gang[8];
1417         int i;
1418         int ret;
1419
1420         /*
1421          * At this point no one can be using this transaction to modify any tree
1422          * and no one can start another transaction to modify any tree either.
1423          */
1424         ASSERT(trans->transaction->state == TRANS_STATE_COMMIT_DOING);
1425
1426         spin_lock(&fs_info->fs_roots_radix_lock);
1427         while (1) {
1428                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1429                                                  (void **)gang, 0,
1430                                                  ARRAY_SIZE(gang),
1431                                                  BTRFS_ROOT_TRANS_TAG);
1432                 if (ret == 0)
1433                         break;
1434                 for (i = 0; i < ret; i++) {
1435                         struct btrfs_root *root = gang[i];
1436                         int ret2;
1437
1438                         /*
1439                          * At this point we can neither have tasks logging inodes
1440                          * from a root nor trying to commit a log tree.
1441                          */
1442                         ASSERT(atomic_read(&root->log_writers) == 0);
1443                         ASSERT(atomic_read(&root->log_commit[0]) == 0);
1444                         ASSERT(atomic_read(&root->log_commit[1]) == 0);
1445
1446                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
1447                                         (unsigned long)root->root_key.objectid,
1448                                         BTRFS_ROOT_TRANS_TAG);
1449                         spin_unlock(&fs_info->fs_roots_radix_lock);
1450
1451                         btrfs_free_log(trans, root);
1452                         ret2 = btrfs_update_reloc_root(trans, root);
1453                         if (ret2)
1454                                 return ret2;
1455
1456                         /* see comments in should_cow_block() */
1457                         clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1458                         smp_mb__after_atomic();
1459
1460                         if (root->commit_root != root->node) {
1461                                 list_add_tail(&root->dirty_list,
1462                                         &trans->transaction->switch_commits);
1463                                 btrfs_set_root_node(&root->root_item,
1464                                                     root->node);
1465                         }
1466
1467                         ret2 = btrfs_update_root(trans, fs_info->tree_root,
1468                                                 &root->root_key,
1469                                                 &root->root_item);
1470                         if (ret2)
1471                                 return ret2;
1472                         spin_lock(&fs_info->fs_roots_radix_lock);
1473                         btrfs_qgroup_free_meta_all_pertrans(root);
1474                 }
1475         }
1476         spin_unlock(&fs_info->fs_roots_radix_lock);
1477         return 0;
1478 }
1479
1480 /*
1481  * defrag a given btree.
1482  * Every leaf in the btree is read and defragged.
1483  */
1484 int btrfs_defrag_root(struct btrfs_root *root)
1485 {
1486         struct btrfs_fs_info *info = root->fs_info;
1487         struct btrfs_trans_handle *trans;
1488         int ret;
1489
1490         if (test_and_set_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state))
1491                 return 0;
1492
1493         while (1) {
1494                 trans = btrfs_start_transaction(root, 0);
1495                 if (IS_ERR(trans)) {
1496                         ret = PTR_ERR(trans);
1497                         break;
1498                 }
1499
1500                 ret = btrfs_defrag_leaves(trans, root);
1501
1502                 btrfs_end_transaction(trans);
1503                 btrfs_btree_balance_dirty(info);
1504                 cond_resched();
1505
1506                 if (btrfs_fs_closing(info) || ret != -EAGAIN)
1507                         break;
1508
1509                 if (btrfs_defrag_cancelled(info)) {
1510                         btrfs_debug(info, "defrag_root cancelled");
1511                         ret = -EAGAIN;
1512                         break;
1513                 }
1514         }
1515         clear_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state);
1516         return ret;
1517 }
1518
1519 /*
1520  * Do all special snapshot related qgroup dirty hack.
1521  *
1522  * Will do all needed qgroup inherit and dirty hack like switch commit
1523  * roots inside one transaction and write all btree into disk, to make
1524  * qgroup works.
1525  */
1526 static int qgroup_account_snapshot(struct btrfs_trans_handle *trans,
1527                                    struct btrfs_root *src,
1528                                    struct btrfs_root *parent,
1529                                    struct btrfs_qgroup_inherit *inherit,
1530                                    u64 dst_objectid)
1531 {
1532         struct btrfs_fs_info *fs_info = src->fs_info;
1533         int ret;
1534
1535         /*
1536          * Save some performance in the case that qgroups are not
1537          * enabled. If this check races with the ioctl, rescan will
1538          * kick in anyway.
1539          */
1540         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
1541                 return 0;
1542
1543         /*
1544          * Ensure dirty @src will be committed.  Or, after coming
1545          * commit_fs_roots() and switch_commit_roots(), any dirty but not
1546          * recorded root will never be updated again, causing an outdated root
1547          * item.
1548          */
1549         ret = record_root_in_trans(trans, src, 1);
1550         if (ret)
1551                 return ret;
1552
1553         /*
1554          * btrfs_qgroup_inherit relies on a consistent view of the usage for the
1555          * src root, so we must run the delayed refs here.
1556          *
1557          * However this isn't particularly fool proof, because there's no
1558          * synchronization keeping us from changing the tree after this point
1559          * before we do the qgroup_inherit, or even from making changes while
1560          * we're doing the qgroup_inherit.  But that's a problem for the future,
1561          * for now flush the delayed refs to narrow the race window where the
1562          * qgroup counters could end up wrong.
1563          */
1564         ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1565         if (ret) {
1566                 btrfs_abort_transaction(trans, ret);
1567                 return ret;
1568         }
1569
1570         ret = commit_fs_roots(trans);
1571         if (ret)
1572                 goto out;
1573         ret = btrfs_qgroup_account_extents(trans);
1574         if (ret < 0)
1575                 goto out;
1576
1577         /* Now qgroup are all updated, we can inherit it to new qgroups */
1578         ret = btrfs_qgroup_inherit(trans, src->root_key.objectid, dst_objectid,
1579                                    inherit);
1580         if (ret < 0)
1581                 goto out;
1582
1583         /*
1584          * Now we do a simplified commit transaction, which will:
1585          * 1) commit all subvolume and extent tree
1586          *    To ensure all subvolume and extent tree have a valid
1587          *    commit_root to accounting later insert_dir_item()
1588          * 2) write all btree blocks onto disk
1589          *    This is to make sure later btree modification will be cowed
1590          *    Or commit_root can be populated and cause wrong qgroup numbers
1591          * In this simplified commit, we don't really care about other trees
1592          * like chunk and root tree, as they won't affect qgroup.
1593          * And we don't write super to avoid half committed status.
1594          */
1595         ret = commit_cowonly_roots(trans);
1596         if (ret)
1597                 goto out;
1598         switch_commit_roots(trans);
1599         ret = btrfs_write_and_wait_transaction(trans);
1600         if (ret)
1601                 btrfs_handle_fs_error(fs_info, ret,
1602                         "Error while writing out transaction for qgroup");
1603
1604 out:
1605         /*
1606          * Force parent root to be updated, as we recorded it before so its
1607          * last_trans == cur_transid.
1608          * Or it won't be committed again onto disk after later
1609          * insert_dir_item()
1610          */
1611         if (!ret)
1612                 ret = record_root_in_trans(trans, parent, 1);
1613         return ret;
1614 }
1615
1616 /*
1617  * new snapshots need to be created at a very specific time in the
1618  * transaction commit.  This does the actual creation.
1619  *
1620  * Note:
1621  * If the error which may affect the commitment of the current transaction
1622  * happens, we should return the error number. If the error which just affect
1623  * the creation of the pending snapshots, just return 0.
1624  */
1625 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1626                                    struct btrfs_pending_snapshot *pending)
1627 {
1628
1629         struct btrfs_fs_info *fs_info = trans->fs_info;
1630         struct btrfs_key key;
1631         struct btrfs_root_item *new_root_item;
1632         struct btrfs_root *tree_root = fs_info->tree_root;
1633         struct btrfs_root *root = pending->root;
1634         struct btrfs_root *parent_root;
1635         struct btrfs_block_rsv *rsv;
1636         struct inode *parent_inode = pending->dir;
1637         struct btrfs_path *path;
1638         struct btrfs_dir_item *dir_item;
1639         struct extent_buffer *tmp;
1640         struct extent_buffer *old;
1641         struct timespec64 cur_time;
1642         int ret = 0;
1643         u64 to_reserve = 0;
1644         u64 index = 0;
1645         u64 objectid;
1646         u64 root_flags;
1647         unsigned int nofs_flags;
1648         struct fscrypt_name fname;
1649
1650         ASSERT(pending->path);
1651         path = pending->path;
1652
1653         ASSERT(pending->root_item);
1654         new_root_item = pending->root_item;
1655
1656         /*
1657          * We're inside a transaction and must make sure that any potential
1658          * allocations with GFP_KERNEL in fscrypt won't recurse back to
1659          * filesystem.
1660          */
1661         nofs_flags = memalloc_nofs_save();
1662         pending->error = fscrypt_setup_filename(parent_inode,
1663                                                 &pending->dentry->d_name, 0,
1664                                                 &fname);
1665         memalloc_nofs_restore(nofs_flags);
1666         if (pending->error)
1667                 goto free_pending;
1668
1669         pending->error = btrfs_get_free_objectid(tree_root, &objectid);
1670         if (pending->error)
1671                 goto free_fname;
1672
1673         /*
1674          * Make qgroup to skip current new snapshot's qgroupid, as it is
1675          * accounted by later btrfs_qgroup_inherit().
1676          */
1677         btrfs_set_skip_qgroup(trans, objectid);
1678
1679         btrfs_reloc_pre_snapshot(pending, &to_reserve);
1680
1681         if (to_reserve > 0) {
1682                 pending->error = btrfs_block_rsv_add(fs_info,
1683                                                      &pending->block_rsv,
1684                                                      to_reserve,
1685                                                      BTRFS_RESERVE_NO_FLUSH);
1686                 if (pending->error)
1687                         goto clear_skip_qgroup;
1688         }
1689
1690         key.objectid = objectid;
1691         key.offset = (u64)-1;
1692         key.type = BTRFS_ROOT_ITEM_KEY;
1693
1694         rsv = trans->block_rsv;
1695         trans->block_rsv = &pending->block_rsv;
1696         trans->bytes_reserved = trans->block_rsv->reserved;
1697         trace_btrfs_space_reservation(fs_info, "transaction",
1698                                       trans->transid,
1699                                       trans->bytes_reserved, 1);
1700         parent_root = BTRFS_I(parent_inode)->root;
1701         ret = record_root_in_trans(trans, parent_root, 0);
1702         if (ret)
1703                 goto fail;
1704         cur_time = current_time(parent_inode);
1705
1706         /*
1707          * insert the directory item
1708          */
1709         ret = btrfs_set_inode_index(BTRFS_I(parent_inode), &index);
1710         if (ret) {
1711                 btrfs_abort_transaction(trans, ret);
1712                 goto fail;
1713         }
1714
1715         /* check if there is a file/dir which has the same name. */
1716         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1717                                          btrfs_ino(BTRFS_I(parent_inode)),
1718                                          &fname.disk_name, 0);
1719         if (dir_item != NULL && !IS_ERR(dir_item)) {
1720                 pending->error = -EEXIST;
1721                 goto dir_item_existed;
1722         } else if (IS_ERR(dir_item)) {
1723                 ret = PTR_ERR(dir_item);
1724                 btrfs_abort_transaction(trans, ret);
1725                 goto fail;
1726         }
1727         btrfs_release_path(path);
1728
1729         /*
1730          * pull in the delayed directory update
1731          * and the delayed inode item
1732          * otherwise we corrupt the FS during
1733          * snapshot
1734          */
1735         ret = btrfs_run_delayed_items(trans);
1736         if (ret) {      /* Transaction aborted */
1737                 btrfs_abort_transaction(trans, ret);
1738                 goto fail;
1739         }
1740
1741         ret = record_root_in_trans(trans, root, 0);
1742         if (ret) {
1743                 btrfs_abort_transaction(trans, ret);
1744                 goto fail;
1745         }
1746         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1747         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1748         btrfs_check_and_init_root_item(new_root_item);
1749
1750         root_flags = btrfs_root_flags(new_root_item);
1751         if (pending->readonly)
1752                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1753         else
1754                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1755         btrfs_set_root_flags(new_root_item, root_flags);
1756
1757         btrfs_set_root_generation_v2(new_root_item,
1758                         trans->transid);
1759         generate_random_guid(new_root_item->uuid);
1760         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1761                         BTRFS_UUID_SIZE);
1762         if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1763                 memset(new_root_item->received_uuid, 0,
1764                        sizeof(new_root_item->received_uuid));
1765                 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1766                 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1767                 btrfs_set_root_stransid(new_root_item, 0);
1768                 btrfs_set_root_rtransid(new_root_item, 0);
1769         }
1770         btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1771         btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1772         btrfs_set_root_otransid(new_root_item, trans->transid);
1773
1774         old = btrfs_lock_root_node(root);
1775         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old,
1776                               BTRFS_NESTING_COW);
1777         if (ret) {
1778                 btrfs_tree_unlock(old);
1779                 free_extent_buffer(old);
1780                 btrfs_abort_transaction(trans, ret);
1781                 goto fail;
1782         }
1783
1784         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1785         /* clean up in any case */
1786         btrfs_tree_unlock(old);
1787         free_extent_buffer(old);
1788         if (ret) {
1789                 btrfs_abort_transaction(trans, ret);
1790                 goto fail;
1791         }
1792         /* see comments in should_cow_block() */
1793         set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1794         smp_wmb();
1795
1796         btrfs_set_root_node(new_root_item, tmp);
1797         /* record when the snapshot was created in key.offset */
1798         key.offset = trans->transid;
1799         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1800         btrfs_tree_unlock(tmp);
1801         free_extent_buffer(tmp);
1802         if (ret) {
1803                 btrfs_abort_transaction(trans, ret);
1804                 goto fail;
1805         }
1806
1807         /*
1808          * insert root back/forward references
1809          */
1810         ret = btrfs_add_root_ref(trans, objectid,
1811                                  parent_root->root_key.objectid,
1812                                  btrfs_ino(BTRFS_I(parent_inode)), index,
1813                                  &fname.disk_name);
1814         if (ret) {
1815                 btrfs_abort_transaction(trans, ret);
1816                 goto fail;
1817         }
1818
1819         key.offset = (u64)-1;
1820         pending->snap = btrfs_get_new_fs_root(fs_info, objectid, pending->anon_dev);
1821         if (IS_ERR(pending->snap)) {
1822                 ret = PTR_ERR(pending->snap);
1823                 pending->snap = NULL;
1824                 btrfs_abort_transaction(trans, ret);
1825                 goto fail;
1826         }
1827
1828         ret = btrfs_reloc_post_snapshot(trans, pending);
1829         if (ret) {
1830                 btrfs_abort_transaction(trans, ret);
1831                 goto fail;
1832         }
1833
1834         /*
1835          * Do special qgroup accounting for snapshot, as we do some qgroup
1836          * snapshot hack to do fast snapshot.
1837          * To co-operate with that hack, we do hack again.
1838          * Or snapshot will be greatly slowed down by a subtree qgroup rescan
1839          */
1840         ret = qgroup_account_snapshot(trans, root, parent_root,
1841                                       pending->inherit, objectid);
1842         if (ret < 0)
1843                 goto fail;
1844
1845         ret = btrfs_insert_dir_item(trans, &fname.disk_name,
1846                                     BTRFS_I(parent_inode), &key, BTRFS_FT_DIR,
1847                                     index);
1848         /* We have check then name at the beginning, so it is impossible. */
1849         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1850         if (ret) {
1851                 btrfs_abort_transaction(trans, ret);
1852                 goto fail;
1853         }
1854
1855         btrfs_i_size_write(BTRFS_I(parent_inode), parent_inode->i_size +
1856                                                   fname.disk_name.len * 2);
1857         parent_inode->i_mtime = inode_set_ctime_current(parent_inode);
1858         ret = btrfs_update_inode_fallback(trans, parent_root, BTRFS_I(parent_inode));
1859         if (ret) {
1860                 btrfs_abort_transaction(trans, ret);
1861                 goto fail;
1862         }
1863         ret = btrfs_uuid_tree_add(trans, new_root_item->uuid,
1864                                   BTRFS_UUID_KEY_SUBVOL,
1865                                   objectid);
1866         if (ret) {
1867                 btrfs_abort_transaction(trans, ret);
1868                 goto fail;
1869         }
1870         if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1871                 ret = btrfs_uuid_tree_add(trans, new_root_item->received_uuid,
1872                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1873                                           objectid);
1874                 if (ret && ret != -EEXIST) {
1875                         btrfs_abort_transaction(trans, ret);
1876                         goto fail;
1877                 }
1878         }
1879
1880 fail:
1881         pending->error = ret;
1882 dir_item_existed:
1883         trans->block_rsv = rsv;
1884         trans->bytes_reserved = 0;
1885 clear_skip_qgroup:
1886         btrfs_clear_skip_qgroup(trans);
1887 free_fname:
1888         fscrypt_free_filename(&fname);
1889 free_pending:
1890         kfree(new_root_item);
1891         pending->root_item = NULL;
1892         btrfs_free_path(path);
1893         pending->path = NULL;
1894
1895         return ret;
1896 }
1897
1898 /*
1899  * create all the snapshots we've scheduled for creation
1900  */
1901 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans)
1902 {
1903         struct btrfs_pending_snapshot *pending, *next;
1904         struct list_head *head = &trans->transaction->pending_snapshots;
1905         int ret = 0;
1906
1907         list_for_each_entry_safe(pending, next, head, list) {
1908                 list_del(&pending->list);
1909                 ret = create_pending_snapshot(trans, pending);
1910                 if (ret)
1911                         break;
1912         }
1913         return ret;
1914 }
1915
1916 static void update_super_roots(struct btrfs_fs_info *fs_info)
1917 {
1918         struct btrfs_root_item *root_item;
1919         struct btrfs_super_block *super;
1920
1921         super = fs_info->super_copy;
1922
1923         root_item = &fs_info->chunk_root->root_item;
1924         super->chunk_root = root_item->bytenr;
1925         super->chunk_root_generation = root_item->generation;
1926         super->chunk_root_level = root_item->level;
1927
1928         root_item = &fs_info->tree_root->root_item;
1929         super->root = root_item->bytenr;
1930         super->generation = root_item->generation;
1931         super->root_level = root_item->level;
1932         if (btrfs_test_opt(fs_info, SPACE_CACHE))
1933                 super->cache_generation = root_item->generation;
1934         else if (test_bit(BTRFS_FS_CLEANUP_SPACE_CACHE_V1, &fs_info->flags))
1935                 super->cache_generation = 0;
1936         if (test_bit(BTRFS_FS_UPDATE_UUID_TREE_GEN, &fs_info->flags))
1937                 super->uuid_tree_generation = root_item->generation;
1938 }
1939
1940 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1941 {
1942         struct btrfs_transaction *trans;
1943         int ret = 0;
1944
1945         spin_lock(&info->trans_lock);
1946         trans = info->running_transaction;
1947         if (trans)
1948                 ret = (trans->state >= TRANS_STATE_COMMIT_START);
1949         spin_unlock(&info->trans_lock);
1950         return ret;
1951 }
1952
1953 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1954 {
1955         struct btrfs_transaction *trans;
1956         int ret = 0;
1957
1958         spin_lock(&info->trans_lock);
1959         trans = info->running_transaction;
1960         if (trans)
1961                 ret = is_transaction_blocked(trans);
1962         spin_unlock(&info->trans_lock);
1963         return ret;
1964 }
1965
1966 void btrfs_commit_transaction_async(struct btrfs_trans_handle *trans)
1967 {
1968         struct btrfs_fs_info *fs_info = trans->fs_info;
1969         struct btrfs_transaction *cur_trans;
1970
1971         /* Kick the transaction kthread. */
1972         set_bit(BTRFS_FS_COMMIT_TRANS, &fs_info->flags);
1973         wake_up_process(fs_info->transaction_kthread);
1974
1975         /* take transaction reference */
1976         cur_trans = trans->transaction;
1977         refcount_inc(&cur_trans->use_count);
1978
1979         btrfs_end_transaction(trans);
1980
1981         /*
1982          * Wait for the current transaction commit to start and block
1983          * subsequent transaction joins
1984          */
1985         btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_START);
1986         wait_event(fs_info->transaction_blocked_wait,
1987                    cur_trans->state >= TRANS_STATE_COMMIT_START ||
1988                    TRANS_ABORTED(cur_trans));
1989         btrfs_put_transaction(cur_trans);
1990 }
1991
1992 static void cleanup_transaction(struct btrfs_trans_handle *trans, int err)
1993 {
1994         struct btrfs_fs_info *fs_info = trans->fs_info;
1995         struct btrfs_transaction *cur_trans = trans->transaction;
1996
1997         WARN_ON(refcount_read(&trans->use_count) > 1);
1998
1999         btrfs_abort_transaction(trans, err);
2000
2001         spin_lock(&fs_info->trans_lock);
2002
2003         /*
2004          * If the transaction is removed from the list, it means this
2005          * transaction has been committed successfully, so it is impossible
2006          * to call the cleanup function.
2007          */
2008         BUG_ON(list_empty(&cur_trans->list));
2009
2010         if (cur_trans == fs_info->running_transaction) {
2011                 cur_trans->state = TRANS_STATE_COMMIT_DOING;
2012                 spin_unlock(&fs_info->trans_lock);
2013
2014                 /*
2015                  * The thread has already released the lockdep map as reader
2016                  * already in btrfs_commit_transaction().
2017                  */
2018                 btrfs_might_wait_for_event(fs_info, btrfs_trans_num_writers);
2019                 wait_event(cur_trans->writer_wait,
2020                            atomic_read(&cur_trans->num_writers) == 1);
2021
2022                 spin_lock(&fs_info->trans_lock);
2023         }
2024
2025         /*
2026          * Now that we know no one else is still using the transaction we can
2027          * remove the transaction from the list of transactions. This avoids
2028          * the transaction kthread from cleaning up the transaction while some
2029          * other task is still using it, which could result in a use-after-free
2030          * on things like log trees, as it forces the transaction kthread to
2031          * wait for this transaction to be cleaned up by us.
2032          */
2033         list_del_init(&cur_trans->list);
2034
2035         spin_unlock(&fs_info->trans_lock);
2036
2037         btrfs_cleanup_one_transaction(trans->transaction, fs_info);
2038
2039         spin_lock(&fs_info->trans_lock);
2040         if (cur_trans == fs_info->running_transaction)
2041                 fs_info->running_transaction = NULL;
2042         spin_unlock(&fs_info->trans_lock);
2043
2044         if (trans->type & __TRANS_FREEZABLE)
2045                 sb_end_intwrite(fs_info->sb);
2046         btrfs_put_transaction(cur_trans);
2047         btrfs_put_transaction(cur_trans);
2048
2049         trace_btrfs_transaction_commit(fs_info);
2050
2051         if (current->journal_info == trans)
2052                 current->journal_info = NULL;
2053
2054         /*
2055          * If relocation is running, we can't cancel scrub because that will
2056          * result in a deadlock. Before relocating a block group, relocation
2057          * pauses scrub, then starts and commits a transaction before unpausing
2058          * scrub. If the transaction commit is being done by the relocation
2059          * task or triggered by another task and the relocation task is waiting
2060          * for the commit, and we end up here due to an error in the commit
2061          * path, then calling btrfs_scrub_cancel() will deadlock, as we are
2062          * asking for scrub to stop while having it asked to be paused higher
2063          * above in relocation code.
2064          */
2065         if (!test_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags))
2066                 btrfs_scrub_cancel(fs_info);
2067
2068         kmem_cache_free(btrfs_trans_handle_cachep, trans);
2069 }
2070
2071 /*
2072  * Release reserved delayed ref space of all pending block groups of the
2073  * transaction and remove them from the list
2074  */
2075 static void btrfs_cleanup_pending_block_groups(struct btrfs_trans_handle *trans)
2076 {
2077        struct btrfs_fs_info *fs_info = trans->fs_info;
2078        struct btrfs_block_group *block_group, *tmp;
2079
2080        list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
2081                btrfs_delayed_refs_rsv_release(fs_info, 1);
2082                list_del_init(&block_group->bg_list);
2083        }
2084 }
2085
2086 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
2087 {
2088         /*
2089          * We use try_to_writeback_inodes_sb() here because if we used
2090          * btrfs_start_delalloc_roots we would deadlock with fs freeze.
2091          * Currently are holding the fs freeze lock, if we do an async flush
2092          * we'll do btrfs_join_transaction() and deadlock because we need to
2093          * wait for the fs freeze lock.  Using the direct flushing we benefit
2094          * from already being in a transaction and our join_transaction doesn't
2095          * have to re-take the fs freeze lock.
2096          *
2097          * Note that try_to_writeback_inodes_sb() will only trigger writeback
2098          * if it can read lock sb->s_umount. It will always be able to lock it,
2099          * except when the filesystem is being unmounted or being frozen, but in
2100          * those cases sync_filesystem() is called, which results in calling
2101          * writeback_inodes_sb() while holding a write lock on sb->s_umount.
2102          * Note that we don't call writeback_inodes_sb() directly, because it
2103          * will emit a warning if sb->s_umount is not locked.
2104          */
2105         if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
2106                 try_to_writeback_inodes_sb(fs_info->sb, WB_REASON_SYNC);
2107         return 0;
2108 }
2109
2110 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
2111 {
2112         if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
2113                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0, (u64)-1);
2114 }
2115
2116 /*
2117  * Add a pending snapshot associated with the given transaction handle to the
2118  * respective handle. This must be called after the transaction commit started
2119  * and while holding fs_info->trans_lock.
2120  * This serves to guarantee a caller of btrfs_commit_transaction() that it can
2121  * safely free the pending snapshot pointer in case btrfs_commit_transaction()
2122  * returns an error.
2123  */
2124 static void add_pending_snapshot(struct btrfs_trans_handle *trans)
2125 {
2126         struct btrfs_transaction *cur_trans = trans->transaction;
2127
2128         if (!trans->pending_snapshot)
2129                 return;
2130
2131         lockdep_assert_held(&trans->fs_info->trans_lock);
2132         ASSERT(cur_trans->state >= TRANS_STATE_COMMIT_START);
2133
2134         list_add(&trans->pending_snapshot->list, &cur_trans->pending_snapshots);
2135 }
2136
2137 static void update_commit_stats(struct btrfs_fs_info *fs_info, ktime_t interval)
2138 {
2139         fs_info->commit_stats.commit_count++;
2140         fs_info->commit_stats.last_commit_dur = interval;
2141         fs_info->commit_stats.max_commit_dur =
2142                         max_t(u64, fs_info->commit_stats.max_commit_dur, interval);
2143         fs_info->commit_stats.total_commit_dur += interval;
2144 }
2145
2146 int btrfs_commit_transaction(struct btrfs_trans_handle *trans)
2147 {
2148         struct btrfs_fs_info *fs_info = trans->fs_info;
2149         struct btrfs_transaction *cur_trans = trans->transaction;
2150         struct btrfs_transaction *prev_trans = NULL;
2151         int ret;
2152         ktime_t start_time;
2153         ktime_t interval;
2154
2155         ASSERT(refcount_read(&trans->use_count) == 1);
2156         btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_START);
2157
2158         clear_bit(BTRFS_FS_NEED_TRANS_COMMIT, &fs_info->flags);
2159
2160         /* Stop the commit early if ->aborted is set */
2161         if (TRANS_ABORTED(cur_trans)) {
2162                 ret = cur_trans->aborted;
2163                 goto lockdep_trans_commit_start_release;
2164         }
2165
2166         btrfs_trans_release_metadata(trans);
2167         trans->block_rsv = NULL;
2168
2169         /*
2170          * We only want one transaction commit doing the flushing so we do not
2171          * waste a bunch of time on lock contention on the extent root node.
2172          */
2173         if (!test_and_set_bit(BTRFS_DELAYED_REFS_FLUSHING,
2174                               &cur_trans->delayed_refs.flags)) {
2175                 /*
2176                  * Make a pass through all the delayed refs we have so far.
2177                  * Any running threads may add more while we are here.
2178                  */
2179                 ret = btrfs_run_delayed_refs(trans, 0);
2180                 if (ret)
2181                         goto lockdep_trans_commit_start_release;
2182         }
2183
2184         btrfs_create_pending_block_groups(trans);
2185
2186         if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) {
2187                 int run_it = 0;
2188
2189                 /* this mutex is also taken before trying to set
2190                  * block groups readonly.  We need to make sure
2191                  * that nobody has set a block group readonly
2192                  * after a extents from that block group have been
2193                  * allocated for cache files.  btrfs_set_block_group_ro
2194                  * will wait for the transaction to commit if it
2195                  * finds BTRFS_TRANS_DIRTY_BG_RUN set.
2196                  *
2197                  * The BTRFS_TRANS_DIRTY_BG_RUN flag is also used to make sure
2198                  * only one process starts all the block group IO.  It wouldn't
2199                  * hurt to have more than one go through, but there's no
2200                  * real advantage to it either.
2201                  */
2202                 mutex_lock(&fs_info->ro_block_group_mutex);
2203                 if (!test_and_set_bit(BTRFS_TRANS_DIRTY_BG_RUN,
2204                                       &cur_trans->flags))
2205                         run_it = 1;
2206                 mutex_unlock(&fs_info->ro_block_group_mutex);
2207
2208                 if (run_it) {
2209                         ret = btrfs_start_dirty_block_groups(trans);
2210                         if (ret)
2211                                 goto lockdep_trans_commit_start_release;
2212                 }
2213         }
2214
2215         spin_lock(&fs_info->trans_lock);
2216         if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
2217                 enum btrfs_trans_state want_state = TRANS_STATE_COMPLETED;
2218
2219                 add_pending_snapshot(trans);
2220
2221                 spin_unlock(&fs_info->trans_lock);
2222                 refcount_inc(&cur_trans->use_count);
2223
2224                 if (trans->in_fsync)
2225                         want_state = TRANS_STATE_SUPER_COMMITTED;
2226
2227                 btrfs_trans_state_lockdep_release(fs_info,
2228                                                   BTRFS_LOCKDEP_TRANS_COMMIT_START);
2229                 ret = btrfs_end_transaction(trans);
2230                 wait_for_commit(cur_trans, want_state);
2231
2232                 if (TRANS_ABORTED(cur_trans))
2233                         ret = cur_trans->aborted;
2234
2235                 btrfs_put_transaction(cur_trans);
2236
2237                 return ret;
2238         }
2239
2240         cur_trans->state = TRANS_STATE_COMMIT_START;
2241         wake_up(&fs_info->transaction_blocked_wait);
2242         btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_START);
2243
2244         if (cur_trans->list.prev != &fs_info->trans_list) {
2245                 enum btrfs_trans_state want_state = TRANS_STATE_COMPLETED;
2246
2247                 if (trans->in_fsync)
2248                         want_state = TRANS_STATE_SUPER_COMMITTED;
2249
2250                 prev_trans = list_entry(cur_trans->list.prev,
2251                                         struct btrfs_transaction, list);
2252                 if (prev_trans->state < want_state) {
2253                         refcount_inc(&prev_trans->use_count);
2254                         spin_unlock(&fs_info->trans_lock);
2255
2256                         wait_for_commit(prev_trans, want_state);
2257
2258                         ret = READ_ONCE(prev_trans->aborted);
2259
2260                         btrfs_put_transaction(prev_trans);
2261                         if (ret)
2262                                 goto lockdep_release;
2263                 } else {
2264                         spin_unlock(&fs_info->trans_lock);
2265                 }
2266         } else {
2267                 spin_unlock(&fs_info->trans_lock);
2268                 /*
2269                  * The previous transaction was aborted and was already removed
2270                  * from the list of transactions at fs_info->trans_list. So we
2271                  * abort to prevent writing a new superblock that reflects a
2272                  * corrupt state (pointing to trees with unwritten nodes/leafs).
2273                  */
2274                 if (BTRFS_FS_ERROR(fs_info)) {
2275                         ret = -EROFS;
2276                         goto lockdep_release;
2277                 }
2278         }
2279
2280         /*
2281          * Get the time spent on the work done by the commit thread and not
2282          * the time spent waiting on a previous commit
2283          */
2284         start_time = ktime_get_ns();
2285
2286         extwriter_counter_dec(cur_trans, trans->type);
2287
2288         ret = btrfs_start_delalloc_flush(fs_info);
2289         if (ret)
2290                 goto lockdep_release;
2291
2292         ret = btrfs_run_delayed_items(trans);
2293         if (ret)
2294                 goto lockdep_release;
2295
2296         /*
2297          * The thread has started/joined the transaction thus it holds the
2298          * lockdep map as a reader. It has to release it before acquiring the
2299          * lockdep map as a writer.
2300          */
2301         btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters);
2302         btrfs_might_wait_for_event(fs_info, btrfs_trans_num_extwriters);
2303         wait_event(cur_trans->writer_wait,
2304                    extwriter_counter_read(cur_trans) == 0);
2305
2306         /* some pending stuffs might be added after the previous flush. */
2307         ret = btrfs_run_delayed_items(trans);
2308         if (ret) {
2309                 btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
2310                 goto cleanup_transaction;
2311         }
2312
2313         btrfs_wait_delalloc_flush(fs_info);
2314
2315         /*
2316          * Wait for all ordered extents started by a fast fsync that joined this
2317          * transaction. Otherwise if this transaction commits before the ordered
2318          * extents complete we lose logged data after a power failure.
2319          */
2320         btrfs_might_wait_for_event(fs_info, btrfs_trans_pending_ordered);
2321         wait_event(cur_trans->pending_wait,
2322                    atomic_read(&cur_trans->pending_ordered) == 0);
2323
2324         btrfs_scrub_pause(fs_info);
2325         /*
2326          * Ok now we need to make sure to block out any other joins while we
2327          * commit the transaction.  We could have started a join before setting
2328          * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
2329          */
2330         spin_lock(&fs_info->trans_lock);
2331         add_pending_snapshot(trans);
2332         cur_trans->state = TRANS_STATE_COMMIT_DOING;
2333         spin_unlock(&fs_info->trans_lock);
2334
2335         /*
2336          * The thread has started/joined the transaction thus it holds the
2337          * lockdep map as a reader. It has to release it before acquiring the
2338          * lockdep map as a writer.
2339          */
2340         btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
2341         btrfs_might_wait_for_event(fs_info, btrfs_trans_num_writers);
2342         wait_event(cur_trans->writer_wait,
2343                    atomic_read(&cur_trans->num_writers) == 1);
2344
2345         /*
2346          * Make lockdep happy by acquiring the state locks after
2347          * btrfs_trans_num_writers is released. If we acquired the state locks
2348          * before releasing the btrfs_trans_num_writers lock then lockdep would
2349          * complain because we did not follow the reverse order unlocking rule.
2350          */
2351         btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED);
2352         btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED);
2353         btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
2354
2355         /*
2356          * We've started the commit, clear the flag in case we were triggered to
2357          * do an async commit but somebody else started before the transaction
2358          * kthread could do the work.
2359          */
2360         clear_bit(BTRFS_FS_COMMIT_TRANS, &fs_info->flags);
2361
2362         if (TRANS_ABORTED(cur_trans)) {
2363                 ret = cur_trans->aborted;
2364                 btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
2365                 goto scrub_continue;
2366         }
2367         /*
2368          * the reloc mutex makes sure that we stop
2369          * the balancing code from coming in and moving
2370          * extents around in the middle of the commit
2371          */
2372         mutex_lock(&fs_info->reloc_mutex);
2373
2374         /*
2375          * We needn't worry about the delayed items because we will
2376          * deal with them in create_pending_snapshot(), which is the
2377          * core function of the snapshot creation.
2378          */
2379         ret = create_pending_snapshots(trans);
2380         if (ret)
2381                 goto unlock_reloc;
2382
2383         /*
2384          * We insert the dir indexes of the snapshots and update the inode
2385          * of the snapshots' parents after the snapshot creation, so there
2386          * are some delayed items which are not dealt with. Now deal with
2387          * them.
2388          *
2389          * We needn't worry that this operation will corrupt the snapshots,
2390          * because all the tree which are snapshoted will be forced to COW
2391          * the nodes and leaves.
2392          */
2393         ret = btrfs_run_delayed_items(trans);
2394         if (ret)
2395                 goto unlock_reloc;
2396
2397         ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
2398         if (ret)
2399                 goto unlock_reloc;
2400
2401         /*
2402          * make sure none of the code above managed to slip in a
2403          * delayed item
2404          */
2405         btrfs_assert_delayed_root_empty(fs_info);
2406
2407         WARN_ON(cur_trans != trans->transaction);
2408
2409         ret = commit_fs_roots(trans);
2410         if (ret)
2411                 goto unlock_reloc;
2412
2413         /* commit_fs_roots gets rid of all the tree log roots, it is now
2414          * safe to free the root of tree log roots
2415          */
2416         btrfs_free_log_root_tree(trans, fs_info);
2417
2418         /*
2419          * Since fs roots are all committed, we can get a quite accurate
2420          * new_roots. So let's do quota accounting.
2421          */
2422         ret = btrfs_qgroup_account_extents(trans);
2423         if (ret < 0)
2424                 goto unlock_reloc;
2425
2426         ret = commit_cowonly_roots(trans);
2427         if (ret)
2428                 goto unlock_reloc;
2429
2430         /*
2431          * The tasks which save the space cache and inode cache may also
2432          * update ->aborted, check it.
2433          */
2434         if (TRANS_ABORTED(cur_trans)) {
2435                 ret = cur_trans->aborted;
2436                 goto unlock_reloc;
2437         }
2438
2439         cur_trans = fs_info->running_transaction;
2440
2441         btrfs_set_root_node(&fs_info->tree_root->root_item,
2442                             fs_info->tree_root->node);
2443         list_add_tail(&fs_info->tree_root->dirty_list,
2444                       &cur_trans->switch_commits);
2445
2446         btrfs_set_root_node(&fs_info->chunk_root->root_item,
2447                             fs_info->chunk_root->node);
2448         list_add_tail(&fs_info->chunk_root->dirty_list,
2449                       &cur_trans->switch_commits);
2450
2451         if (btrfs_fs_incompat(fs_info, EXTENT_TREE_V2)) {
2452                 btrfs_set_root_node(&fs_info->block_group_root->root_item,
2453                                     fs_info->block_group_root->node);
2454                 list_add_tail(&fs_info->block_group_root->dirty_list,
2455                               &cur_trans->switch_commits);
2456         }
2457
2458         switch_commit_roots(trans);
2459
2460         ASSERT(list_empty(&cur_trans->dirty_bgs));
2461         ASSERT(list_empty(&cur_trans->io_bgs));
2462         update_super_roots(fs_info);
2463
2464         btrfs_set_super_log_root(fs_info->super_copy, 0);
2465         btrfs_set_super_log_root_level(fs_info->super_copy, 0);
2466         memcpy(fs_info->super_for_commit, fs_info->super_copy,
2467                sizeof(*fs_info->super_copy));
2468
2469         btrfs_commit_device_sizes(cur_trans);
2470
2471         clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
2472         clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
2473
2474         btrfs_trans_release_chunk_metadata(trans);
2475
2476         /*
2477          * Before changing the transaction state to TRANS_STATE_UNBLOCKED and
2478          * setting fs_info->running_transaction to NULL, lock tree_log_mutex to
2479          * make sure that before we commit our superblock, no other task can
2480          * start a new transaction and commit a log tree before we commit our
2481          * superblock. Anyone trying to commit a log tree locks this mutex before
2482          * writing its superblock.
2483          */
2484         mutex_lock(&fs_info->tree_log_mutex);
2485
2486         spin_lock(&fs_info->trans_lock);
2487         cur_trans->state = TRANS_STATE_UNBLOCKED;
2488         fs_info->running_transaction = NULL;
2489         spin_unlock(&fs_info->trans_lock);
2490         mutex_unlock(&fs_info->reloc_mutex);
2491
2492         wake_up(&fs_info->transaction_wait);
2493         btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
2494
2495         /* If we have features changed, wake up the cleaner to update sysfs. */
2496         if (test_bit(BTRFS_FS_FEATURE_CHANGED, &fs_info->flags) &&
2497             fs_info->cleaner_kthread)
2498                 wake_up_process(fs_info->cleaner_kthread);
2499
2500         ret = btrfs_write_and_wait_transaction(trans);
2501         if (ret) {
2502                 btrfs_handle_fs_error(fs_info, ret,
2503                                       "Error while writing out transaction");
2504                 mutex_unlock(&fs_info->tree_log_mutex);
2505                 goto scrub_continue;
2506         }
2507
2508         ret = write_all_supers(fs_info, 0);
2509         /*
2510          * the super is written, we can safely allow the tree-loggers
2511          * to go about their business
2512          */
2513         mutex_unlock(&fs_info->tree_log_mutex);
2514         if (ret)
2515                 goto scrub_continue;
2516
2517         /*
2518          * We needn't acquire the lock here because there is no other task
2519          * which can change it.
2520          */
2521         cur_trans->state = TRANS_STATE_SUPER_COMMITTED;
2522         wake_up(&cur_trans->commit_wait);
2523         btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED);
2524
2525         btrfs_finish_extent_commit(trans);
2526
2527         if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &cur_trans->flags))
2528                 btrfs_clear_space_info_full(fs_info);
2529
2530         fs_info->last_trans_committed = cur_trans->transid;
2531         /*
2532          * We needn't acquire the lock here because there is no other task
2533          * which can change it.
2534          */
2535         cur_trans->state = TRANS_STATE_COMPLETED;
2536         wake_up(&cur_trans->commit_wait);
2537         btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED);
2538
2539         spin_lock(&fs_info->trans_lock);
2540         list_del_init(&cur_trans->list);
2541         spin_unlock(&fs_info->trans_lock);
2542
2543         btrfs_put_transaction(cur_trans);
2544         btrfs_put_transaction(cur_trans);
2545
2546         if (trans->type & __TRANS_FREEZABLE)
2547                 sb_end_intwrite(fs_info->sb);
2548
2549         trace_btrfs_transaction_commit(fs_info);
2550
2551         interval = ktime_get_ns() - start_time;
2552
2553         btrfs_scrub_continue(fs_info);
2554
2555         if (current->journal_info == trans)
2556                 current->journal_info = NULL;
2557
2558         kmem_cache_free(btrfs_trans_handle_cachep, trans);
2559
2560         update_commit_stats(fs_info, interval);
2561
2562         return ret;
2563
2564 unlock_reloc:
2565         mutex_unlock(&fs_info->reloc_mutex);
2566         btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
2567 scrub_continue:
2568         btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED);
2569         btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED);
2570         btrfs_scrub_continue(fs_info);
2571 cleanup_transaction:
2572         btrfs_trans_release_metadata(trans);
2573         btrfs_cleanup_pending_block_groups(trans);
2574         btrfs_trans_release_chunk_metadata(trans);
2575         trans->block_rsv = NULL;
2576         btrfs_warn(fs_info, "Skipping commit of aborted transaction.");
2577         if (current->journal_info == trans)
2578                 current->journal_info = NULL;
2579         cleanup_transaction(trans, ret);
2580
2581         return ret;
2582
2583 lockdep_release:
2584         btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters);
2585         btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
2586         goto cleanup_transaction;
2587
2588 lockdep_trans_commit_start_release:
2589         btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_START);
2590         btrfs_end_transaction(trans);
2591         return ret;
2592 }
2593
2594 /*
2595  * return < 0 if error
2596  * 0 if there are no more dead_roots at the time of call
2597  * 1 there are more to be processed, call me again
2598  *
2599  * The return value indicates there are certainly more snapshots to delete, but
2600  * if there comes a new one during processing, it may return 0. We don't mind,
2601  * because btrfs_commit_super will poke cleaner thread and it will process it a
2602  * few seconds later.
2603  */
2604 int btrfs_clean_one_deleted_snapshot(struct btrfs_fs_info *fs_info)
2605 {
2606         struct btrfs_root *root;
2607         int ret;
2608
2609         spin_lock(&fs_info->trans_lock);
2610         if (list_empty(&fs_info->dead_roots)) {
2611                 spin_unlock(&fs_info->trans_lock);
2612                 return 0;
2613         }
2614         root = list_first_entry(&fs_info->dead_roots,
2615                         struct btrfs_root, root_list);
2616         list_del_init(&root->root_list);
2617         spin_unlock(&fs_info->trans_lock);
2618
2619         btrfs_debug(fs_info, "cleaner removing %llu", root->root_key.objectid);
2620
2621         btrfs_kill_all_delayed_nodes(root);
2622
2623         if (btrfs_header_backref_rev(root->node) <
2624                         BTRFS_MIXED_BACKREF_REV)
2625                 ret = btrfs_drop_snapshot(root, 0, 0);
2626         else
2627                 ret = btrfs_drop_snapshot(root, 1, 0);
2628
2629         btrfs_put_root(root);
2630         return (ret < 0) ? 0 : 1;
2631 }
2632
2633 /*
2634  * We only mark the transaction aborted and then set the file system read-only.
2635  * This will prevent new transactions from starting or trying to join this
2636  * one.
2637  *
2638  * This means that error recovery at the call site is limited to freeing
2639  * any local memory allocations and passing the error code up without
2640  * further cleanup. The transaction should complete as it normally would
2641  * in the call path but will return -EIO.
2642  *
2643  * We'll complete the cleanup in btrfs_end_transaction and
2644  * btrfs_commit_transaction.
2645  */
2646 void __cold __btrfs_abort_transaction(struct btrfs_trans_handle *trans,
2647                                       const char *function,
2648                                       unsigned int line, int errno, bool first_hit)
2649 {
2650         struct btrfs_fs_info *fs_info = trans->fs_info;
2651
2652         WRITE_ONCE(trans->aborted, errno);
2653         WRITE_ONCE(trans->transaction->aborted, errno);
2654         if (first_hit && errno == -ENOSPC)
2655                 btrfs_dump_space_info_for_trans_abort(fs_info);
2656         /* Wake up anybody who may be waiting on this transaction */
2657         wake_up(&fs_info->transaction_wait);
2658         wake_up(&fs_info->transaction_blocked_wait);
2659         __btrfs_handle_fs_error(fs_info, function, line, errno, NULL);
2660 }
2661
2662 int __init btrfs_transaction_init(void)
2663 {
2664         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
2665                         sizeof(struct btrfs_trans_handle), 0,
2666                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
2667         if (!btrfs_trans_handle_cachep)
2668                 return -ENOMEM;
2669         return 0;
2670 }
2671
2672 void __cold btrfs_transaction_exit(void)
2673 {
2674         kmem_cache_destroy(btrfs_trans_handle_cachep);
2675 }