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 /*
  41  * Transaction states and transitions
  42  *
  43  * No running transaction (fs tree blocks are not modified)
  44  * |
  45  * | To next stage:
  46  * |  Call start_transaction() variants. Except btrfs_join_transaction_nostart().
  47  * V
  48  * Transaction N [[TRANS_STATE_RUNNING]]
  49  * |
  50  * | New trans handles can be attached to transaction N by calling all
  51  * | start_transaction() variants.
  52  * |
  53  * | To next stage:
  54  * |  Call btrfs_commit_transaction() on any trans handle attached to
  55  * |  transaction N
  56  * V
  57  * Transaction N [[TRANS_STATE_COMMIT_PREP]]
  58  * |
  59  * | If there are simultaneous calls to btrfs_commit_transaction() one will win
  60  * | the race and the rest will wait for the winner to commit the transaction.
  61  * |
  62  * | The winner will wait for previous running transaction to completely finish
  63  * | if there is one.
  64  * |
  65  * Transaction N [[TRANS_STATE_COMMIT_START]]
  66  * |
  67  * | Then one of the following happens:
  68  * | - Wait for all other trans handle holders to release.
  69  * |   The btrfs_commit_transaction() caller will do the commit work.
  70  * | - Wait for current transaction to be committed by others.
  71  * |   Other btrfs_commit_transaction() caller will do the commit work.
  72  * |
  73  * | At this stage, only btrfs_join_transaction*() variants can attach
  74  * | to this running transaction.
  75  * | All other variants will wait for current one to finish and attach to
  76  * | transaction N+1.
  77  * |
  78  * | To next stage:
  79  * |  Caller is chosen to commit transaction N, and all other trans handle
  80  * |  haven been released.
  81  * V
  82  * Transaction N [[TRANS_STATE_COMMIT_DOING]]
  83  * |
  84  * | The heavy lifting transaction work is started.
  85  * | From running delayed refs (modifying extent tree) to creating pending
  86  * | snapshots, running qgroups.
  87  * | In short, modify supporting trees to reflect modifications of subvolume
  88  * | trees.
  89  * |
  90  * | At this stage, all start_transaction() calls will wait for this
  91  * | transaction to finish and attach to transaction N+1.
  92  * |
  93  * | To next stage:
  94  * |  Until all supporting trees are updated.
  95  * V
  96  * Transaction N [[TRANS_STATE_UNBLOCKED]]
  97  * |                                                Transaction N+1
  98  * | All needed trees are modified, thus we only    [[TRANS_STATE_RUNNING]]
  99  * | need to write them back to disk and update     |
 100  * | super blocks.                                  |
 101  * |                                                |
 102  * | At this stage, new transaction is allowed to   |
 103  * | start.                                         |
 104  * | All new start_transaction() calls will be      |
 105  * | attached to transid N+1.                       |
 106  * |                                                |
 107  * | To next stage:                                 |
 108  * |  Until all tree blocks are super blocks are    |
 109  * |  written to block devices                      |
 110  * V                                                |
 111  * Transaction N [[TRANS_STATE_COMPLETED]]          V
 112  *   All tree blocks and super blocks are written.  Transaction N+1
 113  *   This transaction is finished and all its       [[TRANS_STATE_COMMIT_START]]
 114  *   data structures will be cleaned up.            | Life goes on
 115  */
 116 static const unsigned int btrfs_blocked_trans_types[TRANS_STATE_MAX] = {
 117         [TRANS_STATE_RUNNING]           = 0U,
 118         [TRANS_STATE_COMMIT_PREP]       = 0U,
 119         [TRANS_STATE_COMMIT_START]      = (__TRANS_START | __TRANS_ATTACH),
 120         [TRANS_STATE_COMMIT_DOING]      = (__TRANS_START |
 121                                            __TRANS_ATTACH |
 122                                            __TRANS_JOIN |
 123                                            __TRANS_JOIN_NOSTART),
 124         [TRANS_STATE_UNBLOCKED]         = (__TRANS_START |
 125                                            __TRANS_ATTACH |
 126                                            __TRANS_JOIN |
 127                                            __TRANS_JOIN_NOLOCK |
 128                                            __TRANS_JOIN_NOSTART),
 129         [TRANS_STATE_SUPER_COMMITTED]   = (__TRANS_START |
 130                                            __TRANS_ATTACH |
 131                                            __TRANS_JOIN |
 132                                            __TRANS_JOIN_NOLOCK |
 133                                            __TRANS_JOIN_NOSTART),
 134         [TRANS_STATE_COMPLETED]         = (__TRANS_START |
 135                                            __TRANS_ATTACH |
 136                                            __TRANS_JOIN |
 137                                            __TRANS_JOIN_NOLOCK |
 138                                            __TRANS_JOIN_NOSTART),
 139 };
 140
 141 void btrfs_put_transaction(struct btrfs_transaction *transaction)
 142 {
 143         WARN_ON(refcount_read(&transaction->use_count) == 0);
 144         if (refcount_dec_and_test(&transaction->use_count)) {
 145                 BUG_ON(!list_empty(&transaction->list));
 146                 WARN_ON(!RB_EMPTY_ROOT(
 147                                 &transaction->delayed_refs.href_root.rb_root));
 148                 WARN_ON(!RB_EMPTY_ROOT(
 149                                 &transaction->delayed_refs.dirty_extent_root));
 150                 if (transaction->delayed_refs.pending_csums)
 151                         btrfs_err(transaction->fs_info,
 152                                   "pending csums is %llu",
 153                                   transaction->delayed_refs.pending_csums);
 154                 /*
 155                  * If any block groups are found in ->deleted_bgs then it's
 156                  * because the transaction was aborted and a commit did not
 157                  * happen (things failed before writing the new superblock
 158                  * and calling btrfs_finish_extent_commit()), so we can not
 159                  * discard the physical locations of the block groups.
 160                  */
 161                 while (!list_empty(&transaction->deleted_bgs)) {
 162                         struct btrfs_block_group *cache;
 163
 164                         cache = list_first_entry(&transaction->deleted_bgs,
 165                                                  struct btrfs_block_group,
 166                                                  bg_list);
 167                         list_del_init(&cache->bg_list);
 168                         btrfs_unfreeze_block_group(cache);
 169                         btrfs_put_block_group(cache);
 170                 }
 171                 WARN_ON(!list_empty(&transaction->dev_update_list));
 172                 kfree(transaction);
 173         }
 174 }
 175
 176 static noinline void switch_commit_roots(struct btrfs_trans_handle *trans)
 177 {
 178         struct btrfs_transaction *cur_trans = trans->transaction;
 179         struct btrfs_fs_info *fs_info = trans->fs_info;
 180         struct btrfs_root *root, *tmp;
 181
 182         /*
 183          * At this point no one can be using this transaction to modify any tree
 184          * and no one can start another transaction to modify any tree either.
 185          */
 186         ASSERT(cur_trans->state == TRANS_STATE_COMMIT_DOING);
 187
 188         down_write(&fs_info->commit_root_sem);
 189
 190         if (test_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags))
 191                 fs_info->last_reloc_trans = trans->transid;
 192
 193         list_for_each_entry_safe(root, tmp, &cur_trans->switch_commits,
 194                                  dirty_list) {
 195                 list_del_init(&root->dirty_list);
 196                 free_extent_buffer(root->commit_root);
 197                 root->commit_root = btrfs_root_node(root);
 198                 extent_io_tree_release(&root->dirty_log_pages);
 199                 btrfs_qgroup_clean_swapped_blocks(root);
 200         }
 201
 202         /* We can free old roots now. */
 203         spin_lock(&cur_trans->dropped_roots_lock);
 204         while (!list_empty(&cur_trans->dropped_roots)) {
 205                 root = list_first_entry(&cur_trans->dropped_roots,
 206                                         struct btrfs_root, root_list);
 207                 list_del_init(&root->root_list);
 208                 spin_unlock(&cur_trans->dropped_roots_lock);
 209                 btrfs_free_log(trans, root);
 210                 btrfs_drop_and_free_fs_root(fs_info, root);
 211                 spin_lock(&cur_trans->dropped_roots_lock);
 212         }
 213         spin_unlock(&cur_trans->dropped_roots_lock);
 214
 215         up_write(&fs_info->commit_root_sem);
 216 }
 217
 218 static inline void extwriter_counter_inc(struct btrfs_transaction *trans,
 219                                          unsigned int type)
 220 {
 221         if (type & TRANS_EXTWRITERS)
 222                 atomic_inc(&trans->num_extwriters);
 223 }
 224
 225 static inline void extwriter_counter_dec(struct btrfs_transaction *trans,
 226                                          unsigned int type)
 227 {
 228         if (type & TRANS_EXTWRITERS)
 229                 atomic_dec(&trans->num_extwriters);
 230 }
 231
 232 static inline void extwriter_counter_init(struct btrfs_transaction *trans,
 233                                           unsigned int type)
 234 {
 235         atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0));
 236 }
 237
 238 static inline int extwriter_counter_read(struct btrfs_transaction *trans)
 239 {
 240         return atomic_read(&trans->num_extwriters);
 241 }
 242
 243 /*
 244  * To be called after doing the chunk btree updates right after allocating a new
 245  * chunk (after btrfs_chunk_alloc_add_chunk_item() is called), when removing a
 246  * chunk after all chunk btree updates and after finishing the second phase of
 247  * chunk allocation (btrfs_create_pending_block_groups()) in case some block
 248  * group had its chunk item insertion delayed to the second phase.
 249  */
 250 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
 251 {
 252         struct btrfs_fs_info *fs_info = trans->fs_info;
 253
 254         if (!trans->chunk_bytes_reserved)
 255                 return;
 256
 257         btrfs_block_rsv_release(fs_info, &fs_info->chunk_block_rsv,
 258                                 trans->chunk_bytes_reserved, NULL);
 259         trans->chunk_bytes_reserved = 0;
 260 }
 261
 262 /*
 263  * either allocate a new transaction or hop into the existing one
 264  */
 265 static noinline int join_transaction(struct btrfs_fs_info *fs_info,
 266                                      unsigned int type)
 267 {
 268         struct btrfs_transaction *cur_trans;
 269
 270         spin_lock(&fs_info->trans_lock);
 271 loop:
 272         /* The file system has been taken offline. No new transactions. */
 273         if (BTRFS_FS_ERROR(fs_info)) {
 274                 spin_unlock(&fs_info->trans_lock);
 275                 return -EROFS;
 276         }
 277
 278         cur_trans = fs_info->running_transaction;
 279         if (cur_trans) {
 280                 if (TRANS_ABORTED(cur_trans)) {
 281                         spin_unlock(&fs_info->trans_lock);
 282                         return cur_trans->aborted;
 283                 }
 284                 if (btrfs_blocked_trans_types[cur_trans->state] & type) {
 285                         spin_unlock(&fs_info->trans_lock);
 286                         return -EBUSY;
 287                 }
 288                 refcount_inc(&cur_trans->use_count);
 289                 atomic_inc(&cur_trans->num_writers);
 290                 extwriter_counter_inc(cur_trans, type);
 291                 spin_unlock(&fs_info->trans_lock);
 292                 btrfs_lockdep_acquire(fs_info, btrfs_trans_num_writers);
 293                 btrfs_lockdep_acquire(fs_info, btrfs_trans_num_extwriters);
 294                 return 0;
 295         }
 296         spin_unlock(&fs_info->trans_lock);
 297
 298         /*
 299          * If we are ATTACH or TRANS_JOIN_NOSTART, we just want to catch the
 300          * current transaction, and commit it. If there is no transaction, just
 301          * return ENOENT.
 302          */
 303         if (type == TRANS_ATTACH || type == TRANS_JOIN_NOSTART)
 304                 return -ENOENT;
 305
 306         /*
 307          * JOIN_NOLOCK only happens during the transaction commit, so
 308          * it is impossible that ->running_transaction is NULL
 309          */
 310         BUG_ON(type == TRANS_JOIN_NOLOCK);
 311
 312         cur_trans = kmalloc(sizeof(*cur_trans), GFP_NOFS);
 313         if (!cur_trans)
 314                 return -ENOMEM;
 315
 316         btrfs_lockdep_acquire(fs_info, btrfs_trans_num_writers);
 317         btrfs_lockdep_acquire(fs_info, btrfs_trans_num_extwriters);
 318
 319         spin_lock(&fs_info->trans_lock);
 320         if (fs_info->running_transaction) {
 321                 /*
 322                  * someone started a transaction after we unlocked.  Make sure
 323                  * to redo the checks above
 324                  */
 325                 btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters);
 326                 btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
 327                 kfree(cur_trans);
 328                 goto loop;
 329         } else if (BTRFS_FS_ERROR(fs_info)) {
 330                 spin_unlock(&fs_info->trans_lock);
 331                 btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters);
 332                 btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
 333                 kfree(cur_trans);
 334                 return -EROFS;
 335         }
 336
 337         cur_trans->fs_info = fs_info;
 338         atomic_set(&cur_trans->pending_ordered, 0);
 339         init_waitqueue_head(&cur_trans->pending_wait);
 340         atomic_set(&cur_trans->num_writers, 1);
 341         extwriter_counter_init(cur_trans, type);
 342         init_waitqueue_head(&cur_trans->writer_wait);
 343         init_waitqueue_head(&cur_trans->commit_wait);
 344         cur_trans->state = TRANS_STATE_RUNNING;
 345         /*
 346          * One for this trans handle, one so it will live on until we
 347          * commit the transaction.
 348          */
 349         refcount_set(&cur_trans->use_count, 2);
 350         cur_trans->flags = 0;
 351         cur_trans->start_time = ktime_get_seconds();
 352
 353         memset(&cur_trans->delayed_refs, 0, sizeof(cur_trans->delayed_refs));
 354
 355         cur_trans->delayed_refs.href_root = RB_ROOT_CACHED;
 356         cur_trans->delayed_refs.dirty_extent_root = RB_ROOT;
 357         atomic_set(&cur_trans->delayed_refs.num_entries, 0);
 358
 359         /*
 360          * although the tree mod log is per file system and not per transaction,
 361          * the log must never go across transaction boundaries.
 362          */
 363         smp_mb();
 364         if (!list_empty(&fs_info->tree_mod_seq_list))
 365                 WARN(1, KERN_ERR "BTRFS: tree_mod_seq_list not empty when creating a fresh transaction\n");
 366         if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
 367                 WARN(1, KERN_ERR "BTRFS: tree_mod_log rb tree not empty when creating a fresh transaction\n");
 368         atomic64_set(&fs_info->tree_mod_seq, 0);
 369
 370         spin_lock_init(&cur_trans->delayed_refs.lock);
 371
 372         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 373         INIT_LIST_HEAD(&cur_trans->dev_update_list);
 374         INIT_LIST_HEAD(&cur_trans->switch_commits);
 375         INIT_LIST_HEAD(&cur_trans->dirty_bgs);
 376         INIT_LIST_HEAD(&cur_trans->io_bgs);
 377         INIT_LIST_HEAD(&cur_trans->dropped_roots);
 378         mutex_init(&cur_trans->cache_write_mutex);
 379         spin_lock_init(&cur_trans->dirty_bgs_lock);
 380         INIT_LIST_HEAD(&cur_trans->deleted_bgs);
 381         spin_lock_init(&cur_trans->dropped_roots_lock);
 382         list_add_tail(&cur_trans->list, &fs_info->trans_list);
 383         extent_io_tree_init(fs_info, &cur_trans->dirty_pages,
 384                         IO_TREE_TRANS_DIRTY_PAGES);
 385         extent_io_tree_init(fs_info, &cur_trans->pinned_extents,
 386                         IO_TREE_FS_PINNED_EXTENTS);
 387         btrfs_set_fs_generation(fs_info, fs_info->generation + 1);
 388         cur_trans->transid = fs_info->generation;
 389         fs_info->running_transaction = cur_trans;
 390         cur_trans->aborted = 0;
 391         spin_unlock(&fs_info->trans_lock);
 392
 393         return 0;
 394 }
 395
 396 /*
 397  * This does all the record keeping required to make sure that a shareable root
 398  * is properly recorded in a given transaction.  This is required to make sure
 399  * the old root from before we joined the transaction is deleted when the
 400  * transaction commits.
 401  */
 402 static int record_root_in_trans(struct btrfs_trans_handle *trans,
 403                                struct btrfs_root *root,
 404                                int force)
 405 {
 406         struct btrfs_fs_info *fs_info = root->fs_info;
 407         int ret = 0;
 408
 409         if ((test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
 410             root->last_trans < trans->transid) || force) {
 411                 WARN_ON(!force && root->commit_root != root->node);
 412
 413                 /*
 414                  * see below for IN_TRANS_SETUP usage rules
 415                  * we have the reloc mutex held now, so there
 416                  * is only one writer in this function
 417                  */
 418                 set_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
 419
 420                 /* make sure readers find IN_TRANS_SETUP before
 421                  * they find our root->last_trans update
 422                  */
 423                 smp_wmb();
 424
 425                 spin_lock(&fs_info->fs_roots_radix_lock);
 426                 if (root->last_trans == trans->transid && !force) {
 427                         spin_unlock(&fs_info->fs_roots_radix_lock);
 428                         return 0;
 429                 }
 430                 radix_tree_tag_set(&fs_info->fs_roots_radix,
 431                                    (unsigned long)root->root_key.objectid,
 432                                    BTRFS_ROOT_TRANS_TAG);
 433                 spin_unlock(&fs_info->fs_roots_radix_lock);
 434                 root->last_trans = trans->transid;
 435
 436                 /* this is pretty tricky.  We don't want to
 437                  * take the relocation lock in btrfs_record_root_in_trans
 438                  * unless we're really doing the first setup for this root in
 439                  * this transaction.
 440                  *
 441                  * Normally we'd use root->last_trans as a flag to decide
 442                  * if we want to take the expensive mutex.
 443                  *
 444                  * But, we have to set root->last_trans before we
 445                  * init the relocation root, otherwise, we trip over warnings
 446                  * in ctree.c.  The solution used here is to flag ourselves
 447                  * with root IN_TRANS_SETUP.  When this is 1, we're still
 448                  * fixing up the reloc trees and everyone must wait.
 449                  *
 450                  * When this is zero, they can trust root->last_trans and fly
 451                  * through btrfs_record_root_in_trans without having to take the
 452                  * lock.  smp_wmb() makes sure that all the writes above are
 453                  * done before we pop in the zero below
 454                  */
 455                 ret = btrfs_init_reloc_root(trans, root);
 456                 smp_mb__before_atomic();
 457                 clear_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
 458         }
 459         return ret;
 460 }
 461
 462
 463 void btrfs_add_dropped_root(struct btrfs_trans_handle *trans,
 464                             struct btrfs_root *root)
 465 {
 466         struct btrfs_fs_info *fs_info = root->fs_info;
 467         struct btrfs_transaction *cur_trans = trans->transaction;
 468
 469         /* Add ourselves to the transaction dropped list */
 470         spin_lock(&cur_trans->dropped_roots_lock);
 471         list_add_tail(&root->root_list, &cur_trans->dropped_roots);
 472         spin_unlock(&cur_trans->dropped_roots_lock);
 473
 474         /* Make sure we don't try to update the root at commit time */
 475         spin_lock(&fs_info->fs_roots_radix_lock);
 476         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 477                              (unsigned long)root->root_key.objectid,
 478                              BTRFS_ROOT_TRANS_TAG);
 479         spin_unlock(&fs_info->fs_roots_radix_lock);
 480 }
 481
 482 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 483                                struct btrfs_root *root)
 484 {
 485         struct btrfs_fs_info *fs_info = root->fs_info;
 486         int ret;
 487
 488         if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
 489                 return 0;
 490
 491         /*
 492          * see record_root_in_trans for comments about IN_TRANS_SETUP usage
 493          * and barriers
 494          */
 495         smp_rmb();
 496         if (root->last_trans == trans->transid &&
 497             !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state))
 498                 return 0;
 499
 500         mutex_lock(&fs_info->reloc_mutex);
 501         ret = record_root_in_trans(trans, root, 0);
 502         mutex_unlock(&fs_info->reloc_mutex);
 503
 504         return ret;
 505 }
 506
 507 static inline int is_transaction_blocked(struct btrfs_transaction *trans)
 508 {
 509         return (trans->state >= TRANS_STATE_COMMIT_START &&
 510                 trans->state < TRANS_STATE_UNBLOCKED &&
 511                 !TRANS_ABORTED(trans));
 512 }
 513
 514 /* wait for commit against the current transaction to become unblocked
 515  * when this is done, it is safe to start a new transaction, but the current
 516  * transaction might not be fully on disk.
 517  */
 518 static void wait_current_trans(struct btrfs_fs_info *fs_info)
 519 {
 520         struct btrfs_transaction *cur_trans;
 521
 522         spin_lock(&fs_info->trans_lock);
 523         cur_trans = fs_info->running_transaction;
 524         if (cur_trans && is_transaction_blocked(cur_trans)) {
 525                 refcount_inc(&cur_trans->use_count);
 526                 spin_unlock(&fs_info->trans_lock);
 527
 528                 btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
 529                 wait_event(fs_info->transaction_wait,
 530                            cur_trans->state >= TRANS_STATE_UNBLOCKED ||
 531                            TRANS_ABORTED(cur_trans));
 532                 btrfs_put_transaction(cur_trans);
 533         } else {
 534                 spin_unlock(&fs_info->trans_lock);
 535         }
 536 }
 537
 538 static int may_wait_transaction(struct btrfs_fs_info *fs_info, int type)
 539 {
 540         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
 541                 return 0;
 542
 543         if (type == TRANS_START)
 544                 return 1;
 545
 546         return 0;
 547 }
 548
 549 static inline bool need_reserve_reloc_root(struct btrfs_root *root)
 550 {
 551         struct btrfs_fs_info *fs_info = root->fs_info;
 552
 553         if (!fs_info->reloc_ctl ||
 554             !test_bit(BTRFS_ROOT_SHAREABLE, &root->state) ||
 555             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 556             root->reloc_root)
 557                 return false;
 558
 559         return true;
 560 }
 561
 562 static int btrfs_reserve_trans_metadata(struct btrfs_fs_info *fs_info,
 563                                         enum btrfs_reserve_flush_enum flush,
 564                                         u64 num_bytes,
 565                                         u64 *delayed_refs_bytes)
 566 {
 567         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
 568         struct btrfs_space_info *si = fs_info->trans_block_rsv.space_info;
 569         u64 extra_delayed_refs_bytes = 0;
 570         u64 bytes;
 571         int ret;
 572
 573         /*
 574          * If there's a gap between the size of the delayed refs reserve and
 575          * its reserved space, than some tasks have added delayed refs or bumped
 576          * its size otherwise (due to block group creation or removal, or block
 577          * group item update). Also try to allocate that gap in order to prevent
 578          * using (and possibly abusing) the global reserve when committing the
 579          * transaction.
 580          */
 581         if (flush == BTRFS_RESERVE_FLUSH_ALL &&
 582             !btrfs_block_rsv_full(delayed_refs_rsv)) {
 583                 spin_lock(&delayed_refs_rsv->lock);
 584                 if (delayed_refs_rsv->size > delayed_refs_rsv->reserved)
 585                         extra_delayed_refs_bytes = delayed_refs_rsv->size -
 586                                 delayed_refs_rsv->reserved;
 587                 spin_unlock(&delayed_refs_rsv->lock);
 588         }
 589
 590         bytes = num_bytes + *delayed_refs_bytes + extra_delayed_refs_bytes;
 591
 592         /*
 593          * We want to reserve all the bytes we may need all at once, so we only
 594          * do 1 enospc flushing cycle per transaction start.
 595          */
 596         ret = btrfs_reserve_metadata_bytes(fs_info, si, bytes, flush);
 597         if (ret == 0) {
 598                 if (extra_delayed_refs_bytes > 0)
 599                         btrfs_migrate_to_delayed_refs_rsv(fs_info,
 600                                                           extra_delayed_refs_bytes);
 601                 return 0;
 602         }
 603
 604         if (extra_delayed_refs_bytes > 0) {
 605                 bytes -= extra_delayed_refs_bytes;
 606                 ret = btrfs_reserve_metadata_bytes(fs_info, si, bytes, flush);
 607                 if (ret == 0)
 608                         return 0;
 609         }
 610
 611         /*
 612          * If we are an emergency flush, which can steal from the global block
 613          * reserve, then attempt to not reserve space for the delayed refs, as
 614          * we will consume space for them from the global block reserve.
 615          */
 616         if (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL) {
 617                 bytes -= *delayed_refs_bytes;
 618                 *delayed_refs_bytes = 0;
 619                 ret = btrfs_reserve_metadata_bytes(fs_info, si, bytes, flush);
 620         }
 621
 622         return ret;
 623 }
 624
 625 static struct btrfs_trans_handle *
 626 start_transaction(struct btrfs_root *root, unsigned int num_items,
 627                   unsigned int type, enum btrfs_reserve_flush_enum flush,
 628                   bool enforce_qgroups)
 629 {
 630         struct btrfs_fs_info *fs_info = root->fs_info;
 631         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
 632         struct btrfs_block_rsv *trans_rsv = &fs_info->trans_block_rsv;
 633         struct btrfs_trans_handle *h;
 634         struct btrfs_transaction *cur_trans;
 635         u64 num_bytes = 0;
 636         u64 qgroup_reserved = 0;
 637         u64 delayed_refs_bytes = 0;
 638         bool reloc_reserved = false;
 639         bool do_chunk_alloc = false;
 640         int ret;
 641
 642         if (BTRFS_FS_ERROR(fs_info))
 643                 return ERR_PTR(-EROFS);
 644
 645         if (current->journal_info) {
 646                 WARN_ON(type & TRANS_EXTWRITERS);
 647                 h = current->journal_info;
 648                 refcount_inc(&h->use_count);
 649                 WARN_ON(refcount_read(&h->use_count) > 2);
 650                 h->orig_rsv = h->block_rsv;
 651                 h->block_rsv = NULL;
 652                 goto got_it;
 653         }
 654
 655         /*
 656          * Do the reservation before we join the transaction so we can do all
 657          * the appropriate flushing if need be.
 658          */
 659         if (num_items && root != fs_info->chunk_root) {
 660                 qgroup_reserved = num_items * fs_info->nodesize;
 661                 /*
 662                  * Use prealloc for now, as there might be a currently running
 663                  * transaction that could free this reserved space prematurely
 664                  * by committing.
 665                  */
 666                 ret = btrfs_qgroup_reserve_meta_prealloc(root, qgroup_reserved,
 667                                                          enforce_qgroups, false);
 668                 if (ret)
 669                         return ERR_PTR(ret);
 670
 671                 num_bytes = btrfs_calc_insert_metadata_size(fs_info, num_items);
 672                 /*
 673                  * If we plan to insert/update/delete "num_items" from a btree,
 674                  * we will also generate delayed refs for extent buffers in the
 675                  * respective btree paths, so reserve space for the delayed refs
 676                  * that will be generated by the caller as it modifies btrees.
 677                  * Try to reserve them to avoid excessive use of the global
 678                  * block reserve.
 679                  */
 680                 delayed_refs_bytes = btrfs_calc_delayed_ref_bytes(fs_info, num_items);
 681
 682                 /*
 683                  * Do the reservation for the relocation root creation
 684                  */
 685                 if (need_reserve_reloc_root(root)) {
 686                         num_bytes += fs_info->nodesize;
 687                         reloc_reserved = true;
 688                 }
 689
 690                 ret = btrfs_reserve_trans_metadata(fs_info, flush, num_bytes,
 691                                                    &delayed_refs_bytes);
 692                 if (ret)
 693                         goto reserve_fail;
 694
 695                 btrfs_block_rsv_add_bytes(trans_rsv, num_bytes, true);
 696
 697                 if (trans_rsv->space_info->force_alloc)
 698                         do_chunk_alloc = true;
 699         } else if (num_items == 0 && flush == BTRFS_RESERVE_FLUSH_ALL &&
 700                    !btrfs_block_rsv_full(delayed_refs_rsv)) {
 701                 /*
 702                  * Some people call with btrfs_start_transaction(root, 0)
 703                  * because they can be throttled, but have some other mechanism
 704                  * for reserving space.  We still want these guys to refill the
 705                  * delayed block_rsv so just add 1 items worth of reservation
 706                  * here.
 707                  */
 708                 ret = btrfs_delayed_refs_rsv_refill(fs_info, flush);
 709                 if (ret)
 710                         goto reserve_fail;
 711         }
 712 again:
 713         h = kmem_cache_zalloc(btrfs_trans_handle_cachep, GFP_NOFS);
 714         if (!h) {
 715                 ret = -ENOMEM;
 716                 goto alloc_fail;
 717         }
 718
 719         /*
 720          * If we are JOIN_NOLOCK we're already committing a transaction and
 721          * waiting on this guy, so we don't need to do the sb_start_intwrite
 722          * because we're already holding a ref.  We need this because we could
 723          * have raced in and did an fsync() on a file which can kick a commit
 724          * and then we deadlock with somebody doing a freeze.
 725          *
 726          * If we are ATTACH, it means we just want to catch the current
 727          * transaction and commit it, so we needn't do sb_start_intwrite().
 728          */
 729         if (type & __TRANS_FREEZABLE)
 730                 sb_start_intwrite(fs_info->sb);
 731
 732         if (may_wait_transaction(fs_info, type))
 733                 wait_current_trans(fs_info);
 734
 735         do {
 736                 ret = join_transaction(fs_info, type);
 737                 if (ret == -EBUSY) {
 738                         wait_current_trans(fs_info);
 739                         if (unlikely(type == TRANS_ATTACH ||
 740                                      type == TRANS_JOIN_NOSTART))
 741                                 ret = -ENOENT;
 742                 }
 743         } while (ret == -EBUSY);
 744
 745         if (ret < 0)
 746                 goto join_fail;
 747
 748         cur_trans = fs_info->running_transaction;
 749
 750         h->transid = cur_trans->transid;
 751         h->transaction = cur_trans;
 752         refcount_set(&h->use_count, 1);
 753         h->fs_info = root->fs_info;
 754
 755         h->type = type;
 756         INIT_LIST_HEAD(&h->new_bgs);
 757         btrfs_init_metadata_block_rsv(fs_info, &h->delayed_rsv, BTRFS_BLOCK_RSV_DELOPS);
 758
 759         smp_mb();
 760         if (cur_trans->state >= TRANS_STATE_COMMIT_START &&
 761             may_wait_transaction(fs_info, type)) {
 762                 current->journal_info = h;
 763                 btrfs_commit_transaction(h);
 764                 goto again;
 765         }
 766
 767         if (num_bytes) {
 768                 trace_btrfs_space_reservation(fs_info, "transaction",
 769                                               h->transid, num_bytes, 1);
 770                 h->block_rsv = trans_rsv;
 771                 h->bytes_reserved = num_bytes;
 772                 if (delayed_refs_bytes > 0) {
 773                         trace_btrfs_space_reservation(fs_info,
 774                                                       "local_delayed_refs_rsv",
 775                                                       h->transid,
 776                                                       delayed_refs_bytes, 1);
 777                         h->delayed_refs_bytes_reserved = delayed_refs_bytes;
 778                         btrfs_block_rsv_add_bytes(&h->delayed_rsv, delayed_refs_bytes, true);
 779                         delayed_refs_bytes = 0;
 780                 }
 781                 h->reloc_reserved = reloc_reserved;
 782         }
 783
 784         /*
 785          * Now that we have found a transaction to be a part of, convert the
 786          * qgroup reservation from prealloc to pertrans. A different transaction
 787          * can't race in and free our pertrans out from under us.
 788          */
 789         if (qgroup_reserved)
 790                 btrfs_qgroup_convert_reserved_meta(root, qgroup_reserved);
 791
 792 got_it:
 793         if (!current->journal_info)
 794                 current->journal_info = h;
 795
 796         /*
 797          * If the space_info is marked ALLOC_FORCE then we'll get upgraded to
 798          * ALLOC_FORCE the first run through, and then we won't allocate for
 799          * anybody else who races in later.  We don't care about the return
 800          * value here.
 801          */
 802         if (do_chunk_alloc && num_bytes) {
 803                 u64 flags = h->block_rsv->space_info->flags;
 804
 805                 btrfs_chunk_alloc(h, btrfs_get_alloc_profile(fs_info, flags),
 806                                   CHUNK_ALLOC_NO_FORCE);
 807         }
 808
 809         /*
 810          * btrfs_record_root_in_trans() needs to alloc new extents, and may
 811          * call btrfs_join_transaction() while we're also starting a
 812          * transaction.
 813          *
 814          * Thus it need to be called after current->journal_info initialized,
 815          * or we can deadlock.
 816          */
 817         ret = btrfs_record_root_in_trans(h, root);
 818         if (ret) {
 819                 /*
 820                  * The transaction handle is fully initialized and linked with
 821                  * other structures so it needs to be ended in case of errors,
 822                  * not just freed.
 823                  */
 824                 btrfs_end_transaction(h);
 825                 return ERR_PTR(ret);
 826         }
 827
 828         return h;
 829
 830 join_fail:
 831         if (type & __TRANS_FREEZABLE)
 832                 sb_end_intwrite(fs_info->sb);
 833         kmem_cache_free(btrfs_trans_handle_cachep, h);
 834 alloc_fail:
 835         if (num_bytes)
 836                 btrfs_block_rsv_release(fs_info, trans_rsv, num_bytes, NULL);
 837         if (delayed_refs_bytes)
 838                 btrfs_space_info_free_bytes_may_use(fs_info, trans_rsv->space_info,
 839                                                     delayed_refs_bytes);
 840 reserve_fail:
 841         btrfs_qgroup_free_meta_prealloc(root, qgroup_reserved);
 842         return ERR_PTR(ret);
 843 }
 844
 845 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 846                                                    unsigned int num_items)
 847 {
 848         return start_transaction(root, num_items, TRANS_START,
 849                                  BTRFS_RESERVE_FLUSH_ALL, true);
 850 }
 851
 852 struct btrfs_trans_handle *btrfs_start_transaction_fallback_global_rsv(
 853                                         struct btrfs_root *root,
 854                                         unsigned int num_items)
 855 {
 856         return start_transaction(root, num_items, TRANS_START,
 857                                  BTRFS_RESERVE_FLUSH_ALL_STEAL, false);
 858 }
 859
 860 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 861 {
 862         return start_transaction(root, 0, TRANS_JOIN, BTRFS_RESERVE_NO_FLUSH,
 863                                  true);
 864 }
 865
 866 struct btrfs_trans_handle *btrfs_join_transaction_spacecache(struct btrfs_root *root)
 867 {
 868         return start_transaction(root, 0, TRANS_JOIN_NOLOCK,
 869                                  BTRFS_RESERVE_NO_FLUSH, true);
 870 }
 871
 872 /*
 873  * Similar to regular join but it never starts a transaction when none is
 874  * running or when there's a running one at a state >= TRANS_STATE_UNBLOCKED.
 875  * This is similar to btrfs_attach_transaction() but it allows the join to
 876  * happen if the transaction commit already started but it's not yet in the
 877  * "doing" phase (the state is < TRANS_STATE_COMMIT_DOING).
 878  */
 879 struct btrfs_trans_handle *btrfs_join_transaction_nostart(struct btrfs_root *root)
 880 {
 881         return start_transaction(root, 0, TRANS_JOIN_NOSTART,
 882                                  BTRFS_RESERVE_NO_FLUSH, true);
 883 }
 884
 885 /*
 886  * Catch the running transaction.
 887  *
 888  * It is used when we want to commit the current the transaction, but
 889  * don't want to start a new one.
 890  *
 891  * Note: If this function return -ENOENT, it just means there is no
 892  * running transaction. But it is possible that the inactive transaction
 893  * is still in the memory, not fully on disk. If you hope there is no
 894  * inactive transaction in the fs when -ENOENT is returned, you should
 895  * invoke
 896  *     btrfs_attach_transaction_barrier()
 897  */
 898 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
 899 {
 900         return start_transaction(root, 0, TRANS_ATTACH,
 901                                  BTRFS_RESERVE_NO_FLUSH, true);
 902 }
 903
 904 /*
 905  * Catch the running transaction.
 906  *
 907  * It is similar to the above function, the difference is this one
 908  * will wait for all the inactive transactions until they fully
 909  * complete.
 910  */
 911 struct btrfs_trans_handle *
 912 btrfs_attach_transaction_barrier(struct btrfs_root *root)
 913 {
 914         struct btrfs_trans_handle *trans;
 915
 916         trans = start_transaction(root, 0, TRANS_ATTACH,
 917                                   BTRFS_RESERVE_NO_FLUSH, true);
 918         if (trans == ERR_PTR(-ENOENT)) {
 919                 int ret;
 920
 921                 ret = btrfs_wait_for_commit(root->fs_info, 0);
 922                 if (ret)
 923                         return ERR_PTR(ret);
 924         }
 925
 926         return trans;
 927 }
 928
 929 /* Wait for a transaction commit to reach at least the given state. */
 930 static noinline void wait_for_commit(struct btrfs_transaction *commit,
 931                                      const enum btrfs_trans_state min_state)
 932 {
 933         struct btrfs_fs_info *fs_info = commit->fs_info;
 934         u64 transid = commit->transid;
 935         bool put = false;
 936
 937         /*
 938          * At the moment this function is called with min_state either being
 939          * TRANS_STATE_COMPLETED or TRANS_STATE_SUPER_COMMITTED.
 940          */
 941         if (min_state == TRANS_STATE_COMPLETED)
 942                 btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED);
 943         else
 944                 btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED);
 945
 946         while (1) {
 947                 wait_event(commit->commit_wait, commit->state >= min_state);
 948                 if (put)
 949                         btrfs_put_transaction(commit);
 950
 951                 if (min_state < TRANS_STATE_COMPLETED)
 952                         break;
 953
 954                 /*
 955                  * A transaction isn't really completed until all of the
 956                  * previous transactions are completed, but with fsync we can
 957                  * end up with SUPER_COMMITTED transactions before a COMPLETED
 958                  * transaction. Wait for those.
 959                  */
 960
 961                 spin_lock(&fs_info->trans_lock);
 962                 commit = list_first_entry_or_null(&fs_info->trans_list,
 963                                                   struct btrfs_transaction,
 964                                                   list);
 965                 if (!commit || commit->transid > transid) {
 966                         spin_unlock(&fs_info->trans_lock);
 967                         break;
 968                 }
 969                 refcount_inc(&commit->use_count);
 970                 put = true;
 971                 spin_unlock(&fs_info->trans_lock);
 972         }
 973 }
 974
 975 int btrfs_wait_for_commit(struct btrfs_fs_info *fs_info, u64 transid)
 976 {
 977         struct btrfs_transaction *cur_trans = NULL, *t;
 978         int ret = 0;
 979
 980         if (transid) {
 981                 if (transid <= btrfs_get_last_trans_committed(fs_info))
 982                         goto out;
 983
 984                 /* find specified transaction */
 985                 spin_lock(&fs_info->trans_lock);
 986                 list_for_each_entry(t, &fs_info->trans_list, list) {
 987                         if (t->transid == transid) {
 988                                 cur_trans = t;
 989                                 refcount_inc(&cur_trans->use_count);
 990                                 ret = 0;
 991                                 break;
 992                         }
 993                         if (t->transid > transid) {
 994                                 ret = 0;
 995                                 break;
 996                         }
 997                 }
 998                 spin_unlock(&fs_info->trans_lock);
 999
1000                 /*
1001                  * The specified transaction doesn't exist, or we
1002                  * raced with btrfs_commit_transaction
1003                  */
1004                 if (!cur_trans) {
1005                         if (transid > btrfs_get_last_trans_committed(fs_info))
1006                                 ret = -EINVAL;
1007                         goto out;
1008                 }
1009         } else {
1010                 /* find newest transaction that is committing | committed */
1011                 spin_lock(&fs_info->trans_lock);
1012                 list_for_each_entry_reverse(t, &fs_info->trans_list,
1013                                             list) {
1014                         if (t->state >= TRANS_STATE_COMMIT_START) {
1015                                 if (t->state == TRANS_STATE_COMPLETED)
1016                                         break;
1017                                 cur_trans = t;
1018                                 refcount_inc(&cur_trans->use_count);
1019                                 break;
1020                         }
1021                 }
1022                 spin_unlock(&fs_info->trans_lock);
1023                 if (!cur_trans)
1024                         goto out;  /* nothing committing|committed */
1025         }
1026
1027         wait_for_commit(cur_trans, TRANS_STATE_COMPLETED);
1028         ret = cur_trans->aborted;
1029         btrfs_put_transaction(cur_trans);
1030 out:
1031         return ret;
1032 }
1033
1034 void btrfs_throttle(struct btrfs_fs_info *fs_info)
1035 {
1036         wait_current_trans(fs_info);
1037 }
1038
1039 bool btrfs_should_end_transaction(struct btrfs_trans_handle *trans)
1040 {
1041         struct btrfs_transaction *cur_trans = trans->transaction;
1042
1043         if (cur_trans->state >= TRANS_STATE_COMMIT_START ||
1044             test_bit(BTRFS_DELAYED_REFS_FLUSHING, &cur_trans->delayed_refs.flags))
1045                 return true;
1046
1047         if (btrfs_check_space_for_delayed_refs(trans->fs_info))
1048                 return true;
1049
1050         return !!btrfs_block_rsv_check(&trans->fs_info->global_block_rsv, 50);
1051 }
1052
1053 static void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans)
1054
1055 {
1056         struct btrfs_fs_info *fs_info = trans->fs_info;
1057
1058         if (!trans->block_rsv) {
1059                 ASSERT(!trans->bytes_reserved);
1060                 ASSERT(!trans->delayed_refs_bytes_reserved);
1061                 return;
1062         }
1063
1064         if (!trans->bytes_reserved) {
1065                 ASSERT(!trans->delayed_refs_bytes_reserved);
1066                 return;
1067         }
1068
1069         ASSERT(trans->block_rsv == &fs_info->trans_block_rsv);
1070         trace_btrfs_space_reservation(fs_info, "transaction",
1071                                       trans->transid, trans->bytes_reserved, 0);
1072         btrfs_block_rsv_release(fs_info, trans->block_rsv,
1073                                 trans->bytes_reserved, NULL);
1074         trans->bytes_reserved = 0;
1075
1076         if (!trans->delayed_refs_bytes_reserved)
1077                 return;
1078
1079         trace_btrfs_space_reservation(fs_info, "local_delayed_refs_rsv",
1080                                       trans->transid,
1081                                       trans->delayed_refs_bytes_reserved, 0);
1082         btrfs_block_rsv_release(fs_info, &trans->delayed_rsv,
1083                                 trans->delayed_refs_bytes_reserved, NULL);
1084         trans->delayed_refs_bytes_reserved = 0;
1085 }
1086
1087 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
1088                                    int throttle)
1089 {
1090         struct btrfs_fs_info *info = trans->fs_info;
1091         struct btrfs_transaction *cur_trans = trans->transaction;
1092         int err = 0;
1093
1094         if (refcount_read(&trans->use_count) > 1) {
1095                 refcount_dec(&trans->use_count);
1096                 trans->block_rsv = trans->orig_rsv;
1097                 return 0;
1098         }
1099
1100         btrfs_trans_release_metadata(trans);
1101         trans->block_rsv = NULL;
1102
1103         btrfs_create_pending_block_groups(trans);
1104
1105         btrfs_trans_release_chunk_metadata(trans);
1106
1107         if (trans->type & __TRANS_FREEZABLE)
1108                 sb_end_intwrite(info->sb);
1109
1110         WARN_ON(cur_trans != info->running_transaction);
1111         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
1112         atomic_dec(&cur_trans->num_writers);
1113         extwriter_counter_dec(cur_trans, trans->type);
1114
1115         cond_wake_up(&cur_trans->writer_wait);
1116
1117         btrfs_lockdep_release(info, btrfs_trans_num_extwriters);
1118         btrfs_lockdep_release(info, btrfs_trans_num_writers);
1119
1120         btrfs_put_transaction(cur_trans);
1121
1122         if (current->journal_info == trans)
1123                 current->journal_info = NULL;
1124
1125         if (throttle)
1126                 btrfs_run_delayed_iputs(info);
1127
1128         if (TRANS_ABORTED(trans) || BTRFS_FS_ERROR(info)) {
1129                 wake_up_process(info->transaction_kthread);
1130                 if (TRANS_ABORTED(trans))
1131                         err = trans->aborted;
1132                 else
1133                         err = -EROFS;
1134         }
1135
1136         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1137         return err;
1138 }
1139
1140 int btrfs_end_transaction(struct btrfs_trans_handle *trans)
1141 {
1142         return __btrfs_end_transaction(trans, 0);
1143 }
1144
1145 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans)
1146 {
1147         return __btrfs_end_transaction(trans, 1);
1148 }
1149
1150 /*
1151  * when btree blocks are allocated, they have some corresponding bits set for
1152  * them in one of two extent_io trees.  This is used to make sure all of
1153  * those extents are sent to disk but does not wait on them
1154  */
1155 int btrfs_write_marked_extents(struct btrfs_fs_info *fs_info,
1156                                struct extent_io_tree *dirty_pages, int mark)
1157 {
1158         int err = 0;
1159         int werr = 0;
1160         struct address_space *mapping = fs_info->btree_inode->i_mapping;
1161         struct extent_state *cached_state = NULL;
1162         u64 start = 0;
1163         u64 end;
1164
1165         while (find_first_extent_bit(dirty_pages, start, &start, &end,
1166                                      mark, &cached_state)) {
1167                 bool wait_writeback = false;
1168
1169                 err = convert_extent_bit(dirty_pages, start, end,
1170                                          EXTENT_NEED_WAIT,
1171                                          mark, &cached_state);
1172                 /*
1173                  * convert_extent_bit can return -ENOMEM, which is most of the
1174                  * time a temporary error. So when it happens, ignore the error
1175                  * and wait for writeback of this range to finish - because we
1176                  * failed to set the bit EXTENT_NEED_WAIT for the range, a call
1177                  * to __btrfs_wait_marked_extents() would not know that
1178                  * writeback for this range started and therefore wouldn't
1179                  * wait for it to finish - we don't want to commit a
1180                  * superblock that points to btree nodes/leafs for which
1181                  * writeback hasn't finished yet (and without errors).
1182                  * We cleanup any entries left in the io tree when committing
1183                  * the transaction (through extent_io_tree_release()).
1184                  */
1185                 if (err == -ENOMEM) {
1186                         err = 0;
1187                         wait_writeback = true;
1188                 }
1189                 if (!err)
1190                         err = filemap_fdatawrite_range(mapping, start, end);
1191                 if (err)
1192                         werr = err;
1193                 else if (wait_writeback)
1194                         werr = filemap_fdatawait_range(mapping, start, end);
1195                 free_extent_state(cached_state);
1196                 cached_state = NULL;
1197                 cond_resched();
1198                 start = end + 1;
1199         }
1200         return werr;
1201 }
1202
1203 /*
1204  * when btree blocks are allocated, they have some corresponding bits set for
1205  * them in one of two extent_io trees.  This is used to make sure all of
1206  * those extents are on disk for transaction or log commit.  We wait
1207  * on all the pages and clear them from the dirty pages state tree
1208  */
1209 static int __btrfs_wait_marked_extents(struct btrfs_fs_info *fs_info,
1210                                        struct extent_io_tree *dirty_pages)
1211 {
1212         int err = 0;
1213         int werr = 0;
1214         struct address_space *mapping = fs_info->btree_inode->i_mapping;
1215         struct extent_state *cached_state = NULL;
1216         u64 start = 0;
1217         u64 end;
1218
1219         while (find_first_extent_bit(dirty_pages, start, &start, &end,
1220                                      EXTENT_NEED_WAIT, &cached_state)) {
1221                 /*
1222                  * Ignore -ENOMEM errors returned by clear_extent_bit().
1223                  * When committing the transaction, we'll remove any entries
1224                  * left in the io tree. For a log commit, we don't remove them
1225                  * after committing the log because the tree can be accessed
1226                  * concurrently - we do it only at transaction commit time when
1227                  * it's safe to do it (through extent_io_tree_release()).
1228                  */
1229                 err = clear_extent_bit(dirty_pages, start, end,
1230                                        EXTENT_NEED_WAIT, &cached_state);
1231                 if (err == -ENOMEM)
1232                         err = 0;
1233                 if (!err)
1234                         err = filemap_fdatawait_range(mapping, start, end);
1235                 if (err)
1236                         werr = err;
1237                 free_extent_state(cached_state);
1238                 cached_state = NULL;
1239                 cond_resched();
1240                 start = end + 1;
1241         }
1242         if (err)
1243                 werr = err;
1244         return werr;
1245 }
1246
1247 static int btrfs_wait_extents(struct btrfs_fs_info *fs_info,
1248                        struct extent_io_tree *dirty_pages)
1249 {
1250         bool errors = false;
1251         int err;
1252
1253         err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1254         if (test_and_clear_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags))
1255                 errors = true;
1256
1257         if (errors && !err)
1258                 err = -EIO;
1259         return err;
1260 }
1261
1262 int btrfs_wait_tree_log_extents(struct btrfs_root *log_root, int mark)
1263 {
1264         struct btrfs_fs_info *fs_info = log_root->fs_info;
1265         struct extent_io_tree *dirty_pages = &log_root->dirty_log_pages;
1266         bool errors = false;
1267         int err;
1268
1269         ASSERT(log_root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
1270
1271         err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1272         if ((mark & EXTENT_DIRTY) &&
1273             test_and_clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags))
1274                 errors = true;
1275
1276         if ((mark & EXTENT_NEW) &&
1277             test_and_clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags))
1278                 errors = true;
1279
1280         if (errors && !err)
1281                 err = -EIO;
1282         return err;
1283 }
1284
1285 /*
1286  * When btree blocks are allocated the corresponding extents are marked dirty.
1287  * This function ensures such extents are persisted on disk for transaction or
1288  * log commit.
1289  *
1290  * @trans: transaction whose dirty pages we'd like to write
1291  */
1292 static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans)
1293 {
1294         int ret;
1295         int ret2;
1296         struct extent_io_tree *dirty_pages = &trans->transaction->dirty_pages;
1297         struct btrfs_fs_info *fs_info = trans->fs_info;
1298         struct blk_plug plug;
1299
1300         blk_start_plug(&plug);
1301         ret = btrfs_write_marked_extents(fs_info, dirty_pages, EXTENT_DIRTY);
1302         blk_finish_plug(&plug);
1303         ret2 = btrfs_wait_extents(fs_info, dirty_pages);
1304
1305         extent_io_tree_release(&trans->transaction->dirty_pages);
1306
1307         if (ret)
1308                 return ret;
1309         else if (ret2)
1310                 return ret2;
1311         else
1312                 return 0;
1313 }
1314
1315 /*
1316  * this is used to update the root pointer in the tree of tree roots.
1317  *
1318  * But, in the case of the extent allocation tree, updating the root
1319  * pointer may allocate blocks which may change the root of the extent
1320  * allocation tree.
1321  *
1322  * So, this loops and repeats and makes sure the cowonly root didn't
1323  * change while the root pointer was being updated in the metadata.
1324  */
1325 static int update_cowonly_root(struct btrfs_trans_handle *trans,
1326                                struct btrfs_root *root)
1327 {
1328         int ret;
1329         u64 old_root_bytenr;
1330         u64 old_root_used;
1331         struct btrfs_fs_info *fs_info = root->fs_info;
1332         struct btrfs_root *tree_root = fs_info->tree_root;
1333
1334         old_root_used = btrfs_root_used(&root->root_item);
1335
1336         while (1) {
1337                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1338                 if (old_root_bytenr == root->node->start &&
1339                     old_root_used == btrfs_root_used(&root->root_item))
1340                         break;
1341
1342                 btrfs_set_root_node(&root->root_item, root->node);
1343                 ret = btrfs_update_root(trans, tree_root,
1344                                         &root->root_key,
1345                                         &root->root_item);
1346                 if (ret)
1347                         return ret;
1348
1349                 old_root_used = btrfs_root_used(&root->root_item);
1350         }
1351
1352         return 0;
1353 }
1354
1355 /*
1356  * update all the cowonly tree roots on disk
1357  *
1358  * The error handling in this function may not be obvious. Any of the
1359  * failures will cause the file system to go offline. We still need
1360  * to clean up the delayed refs.
1361  */
1362 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans)
1363 {
1364         struct btrfs_fs_info *fs_info = trans->fs_info;
1365         struct list_head *dirty_bgs = &trans->transaction->dirty_bgs;
1366         struct list_head *io_bgs = &trans->transaction->io_bgs;
1367         struct list_head *next;
1368         struct extent_buffer *eb;
1369         int ret;
1370
1371         /*
1372          * At this point no one can be using this transaction to modify any tree
1373          * and no one can start another transaction to modify any tree either.
1374          */
1375         ASSERT(trans->transaction->state == TRANS_STATE_COMMIT_DOING);
1376
1377         eb = btrfs_lock_root_node(fs_info->tree_root);
1378         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
1379                               0, &eb, BTRFS_NESTING_COW);
1380         btrfs_tree_unlock(eb);
1381         free_extent_buffer(eb);
1382
1383         if (ret)
1384                 return ret;
1385
1386         ret = btrfs_run_dev_stats(trans);
1387         if (ret)
1388                 return ret;
1389         ret = btrfs_run_dev_replace(trans);
1390         if (ret)
1391                 return ret;
1392         ret = btrfs_run_qgroups(trans);
1393         if (ret)
1394                 return ret;
1395
1396         ret = btrfs_setup_space_cache(trans);
1397         if (ret)
1398                 return ret;
1399
1400 again:
1401         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1402                 struct btrfs_root *root;
1403                 next = fs_info->dirty_cowonly_roots.next;
1404                 list_del_init(next);
1405                 root = list_entry(next, struct btrfs_root, dirty_list);
1406                 clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1407
1408                 list_add_tail(&root->dirty_list,
1409                               &trans->transaction->switch_commits);
1410                 ret = update_cowonly_root(trans, root);
1411                 if (ret)
1412                         return ret;
1413         }
1414
1415         /* Now flush any delayed refs generated by updating all of the roots */
1416         ret = btrfs_run_delayed_refs(trans, U64_MAX);
1417         if (ret)
1418                 return ret;
1419
1420         while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1421                 ret = btrfs_write_dirty_block_groups(trans);
1422                 if (ret)
1423                         return ret;
1424
1425                 /*
1426                  * We're writing the dirty block groups, which could generate
1427                  * delayed refs, which could generate more dirty block groups,
1428                  * so we want to keep this flushing in this loop to make sure
1429                  * everything gets run.
1430                  */
1431                 ret = btrfs_run_delayed_refs(trans, U64_MAX);
1432                 if (ret)
1433                         return ret;
1434         }
1435
1436         if (!list_empty(&fs_info->dirty_cowonly_roots))
1437                 goto again;
1438
1439         /* Update dev-replace pointer once everything is committed */
1440         fs_info->dev_replace.committed_cursor_left =
1441                 fs_info->dev_replace.cursor_left_last_write_of_item;
1442
1443         return 0;
1444 }
1445
1446 /*
1447  * If we had a pending drop we need to see if there are any others left in our
1448  * dead roots list, and if not clear our bit and wake any waiters.
1449  */
1450 void btrfs_maybe_wake_unfinished_drop(struct btrfs_fs_info *fs_info)
1451 {
1452         /*
1453          * We put the drop in progress roots at the front of the list, so if the
1454          * first entry doesn't have UNFINISHED_DROP set we can wake everybody
1455          * up.
1456          */
1457         spin_lock(&fs_info->trans_lock);
1458         if (!list_empty(&fs_info->dead_roots)) {
1459                 struct btrfs_root *root = list_first_entry(&fs_info->dead_roots,
1460                                                            struct btrfs_root,
1461                                                            root_list);
1462                 if (test_bit(BTRFS_ROOT_UNFINISHED_DROP, &root->state)) {
1463                         spin_unlock(&fs_info->trans_lock);
1464                         return;
1465                 }
1466         }
1467         spin_unlock(&fs_info->trans_lock);
1468
1469         btrfs_wake_unfinished_drop(fs_info);
1470 }
1471
1472 /*
1473  * dead roots are old snapshots that need to be deleted.  This allocates
1474  * a dirty root struct and adds it into the list of dead roots that need to
1475  * be deleted
1476  */
1477 void btrfs_add_dead_root(struct btrfs_root *root)
1478 {
1479         struct btrfs_fs_info *fs_info = root->fs_info;
1480
1481         spin_lock(&fs_info->trans_lock);
1482         if (list_empty(&root->root_list)) {
1483                 btrfs_grab_root(root);
1484
1485                 /* We want to process the partially complete drops first. */
1486                 if (test_bit(BTRFS_ROOT_UNFINISHED_DROP, &root->state))
1487                         list_add(&root->root_list, &fs_info->dead_roots);
1488                 else
1489                         list_add_tail(&root->root_list, &fs_info->dead_roots);
1490         }
1491         spin_unlock(&fs_info->trans_lock);
1492 }
1493
1494 /*
1495  * Update each subvolume root and its relocation root, if it exists, in the tree
1496  * of tree roots. Also free log roots if they exist.
1497  */
1498 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans)
1499 {
1500         struct btrfs_fs_info *fs_info = trans->fs_info;
1501         struct btrfs_root *gang[8];
1502         int i;
1503         int ret;
1504
1505         /*
1506          * At this point no one can be using this transaction to modify any tree
1507          * and no one can start another transaction to modify any tree either.
1508          */
1509         ASSERT(trans->transaction->state == TRANS_STATE_COMMIT_DOING);
1510
1511         spin_lock(&fs_info->fs_roots_radix_lock);
1512         while (1) {
1513                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1514                                                  (void **)gang, 0,
1515                                                  ARRAY_SIZE(gang),
1516                                                  BTRFS_ROOT_TRANS_TAG);
1517                 if (ret == 0)
1518                         break;
1519                 for (i = 0; i < ret; i++) {
1520                         struct btrfs_root *root = gang[i];
1521                         int ret2;
1522
1523                         /*
1524                          * At this point we can neither have tasks logging inodes
1525                          * from a root nor trying to commit a log tree.
1526                          */
1527                         ASSERT(atomic_read(&root->log_writers) == 0);
1528                         ASSERT(atomic_read(&root->log_commit[0]) == 0);
1529                         ASSERT(atomic_read(&root->log_commit[1]) == 0);
1530
1531                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
1532                                         (unsigned long)root->root_key.objectid,
1533                                         BTRFS_ROOT_TRANS_TAG);
1534                         spin_unlock(&fs_info->fs_roots_radix_lock);
1535
1536                         btrfs_free_log(trans, root);
1537                         ret2 = btrfs_update_reloc_root(trans, root);
1538                         if (ret2)
1539                                 return ret2;
1540
1541                         /* see comments in should_cow_block() */
1542                         clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1543                         smp_mb__after_atomic();
1544
1545                         if (root->commit_root != root->node) {
1546                                 list_add_tail(&root->dirty_list,
1547                                         &trans->transaction->switch_commits);
1548                                 btrfs_set_root_node(&root->root_item,
1549                                                     root->node);
1550                         }
1551
1552                         ret2 = btrfs_update_root(trans, fs_info->tree_root,
1553                                                 &root->root_key,
1554                                                 &root->root_item);
1555                         if (ret2)
1556                                 return ret2;
1557                         spin_lock(&fs_info->fs_roots_radix_lock);
1558                         btrfs_qgroup_free_meta_all_pertrans(root);
1559                 }
1560         }
1561         spin_unlock(&fs_info->fs_roots_radix_lock);
1562         return 0;
1563 }
1564
1565 /*
1566  * Do all special snapshot related qgroup dirty hack.
1567  *
1568  * Will do all needed qgroup inherit and dirty hack like switch commit
1569  * roots inside one transaction and write all btree into disk, to make
1570  * qgroup works.
1571  */
1572 static int qgroup_account_snapshot(struct btrfs_trans_handle *trans,
1573                                    struct btrfs_root *src,
1574                                    struct btrfs_root *parent,
1575                                    struct btrfs_qgroup_inherit *inherit,
1576                                    u64 dst_objectid)
1577 {
1578         struct btrfs_fs_info *fs_info = src->fs_info;
1579         int ret;
1580
1581         /*
1582          * Save some performance in the case that qgroups are not enabled. If
1583          * this check races with the ioctl, rescan will kick in anyway.
1584          */
1585         if (!btrfs_qgroup_full_accounting(fs_info))
1586                 return 0;
1587
1588         /*
1589          * Ensure dirty @src will be committed.  Or, after coming
1590          * commit_fs_roots() and switch_commit_roots(), any dirty but not
1591          * recorded root will never be updated again, causing an outdated root
1592          * item.
1593          */
1594         ret = record_root_in_trans(trans, src, 1);
1595         if (ret)
1596                 return ret;
1597
1598         /*
1599          * btrfs_qgroup_inherit relies on a consistent view of the usage for the
1600          * src root, so we must run the delayed refs here.
1601          *
1602          * However this isn't particularly fool proof, because there's no
1603          * synchronization keeping us from changing the tree after this point
1604          * before we do the qgroup_inherit, or even from making changes while
1605          * we're doing the qgroup_inherit.  But that's a problem for the future,
1606          * for now flush the delayed refs to narrow the race window where the
1607          * qgroup counters could end up wrong.
1608          */
1609         ret = btrfs_run_delayed_refs(trans, U64_MAX);
1610         if (ret) {
1611                 btrfs_abort_transaction(trans, ret);
1612                 return ret;
1613         }
1614
1615         ret = commit_fs_roots(trans);
1616         if (ret)
1617                 goto out;
1618         ret = btrfs_qgroup_account_extents(trans);
1619         if (ret < 0)
1620                 goto out;
1621
1622         /* Now qgroup are all updated, we can inherit it to new qgroups */
1623         ret = btrfs_qgroup_inherit(trans, src->root_key.objectid, dst_objectid,
1624                                    parent->root_key.objectid, inherit);
1625         if (ret < 0)
1626                 goto out;
1627
1628         /*
1629          * Now we do a simplified commit transaction, which will:
1630          * 1) commit all subvolume and extent tree
1631          *    To ensure all subvolume and extent tree have a valid
1632          *    commit_root to accounting later insert_dir_item()
1633          * 2) write all btree blocks onto disk
1634          *    This is to make sure later btree modification will be cowed
1635          *    Or commit_root can be populated and cause wrong qgroup numbers
1636          * In this simplified commit, we don't really care about other trees
1637          * like chunk and root tree, as they won't affect qgroup.
1638          * And we don't write super to avoid half committed status.
1639          */
1640         ret = commit_cowonly_roots(trans);
1641         if (ret)
1642                 goto out;
1643         switch_commit_roots(trans);
1644         ret = btrfs_write_and_wait_transaction(trans);
1645         if (ret)
1646                 btrfs_handle_fs_error(fs_info, ret,
1647                         "Error while writing out transaction for qgroup");
1648
1649 out:
1650         /*
1651          * Force parent root to be updated, as we recorded it before so its
1652          * last_trans == cur_transid.
1653          * Or it won't be committed again onto disk after later
1654          * insert_dir_item()
1655          */
1656         if (!ret)
1657                 ret = record_root_in_trans(trans, parent, 1);
1658         return ret;
1659 }
1660
1661 /*
1662  * new snapshots need to be created at a very specific time in the
1663  * transaction commit.  This does the actual creation.
1664  *
1665  * Note:
1666  * If the error which may affect the commitment of the current transaction
1667  * happens, we should return the error number. If the error which just affect
1668  * the creation of the pending snapshots, just return 0.
1669  */
1670 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1671                                    struct btrfs_pending_snapshot *pending)
1672 {
1673
1674         struct btrfs_fs_info *fs_info = trans->fs_info;
1675         struct btrfs_key key;
1676         struct btrfs_root_item *new_root_item;
1677         struct btrfs_root *tree_root = fs_info->tree_root;
1678         struct btrfs_root *root = pending->root;
1679         struct btrfs_root *parent_root;
1680         struct btrfs_block_rsv *rsv;
1681         struct inode *parent_inode = pending->dir;
1682         struct btrfs_path *path;
1683         struct btrfs_dir_item *dir_item;
1684         struct extent_buffer *tmp;
1685         struct extent_buffer *old;
1686         struct timespec64 cur_time;
1687         int ret = 0;
1688         u64 to_reserve = 0;
1689         u64 index = 0;
1690         u64 objectid;
1691         u64 root_flags;
1692         unsigned int nofs_flags;
1693         struct fscrypt_name fname;
1694
1695         ASSERT(pending->path);
1696         path = pending->path;
1697
1698         ASSERT(pending->root_item);
1699         new_root_item = pending->root_item;
1700
1701         /*
1702          * We're inside a transaction and must make sure that any potential
1703          * allocations with GFP_KERNEL in fscrypt won't recurse back to
1704          * filesystem.
1705          */
1706         nofs_flags = memalloc_nofs_save();
1707         pending->error = fscrypt_setup_filename(parent_inode,
1708                                                 &pending->dentry->d_name, 0,
1709                                                 &fname);
1710         memalloc_nofs_restore(nofs_flags);
1711         if (pending->error)
1712                 goto free_pending;
1713
1714         pending->error = btrfs_get_free_objectid(tree_root, &objectid);
1715         if (pending->error)
1716                 goto free_fname;
1717
1718         /*
1719          * Make qgroup to skip current new snapshot's qgroupid, as it is
1720          * accounted by later btrfs_qgroup_inherit().
1721          */
1722         btrfs_set_skip_qgroup(trans, objectid);
1723
1724         btrfs_reloc_pre_snapshot(pending, &to_reserve);
1725
1726         if (to_reserve > 0) {
1727                 pending->error = btrfs_block_rsv_add(fs_info,
1728                                                      &pending->block_rsv,
1729                                                      to_reserve,
1730                                                      BTRFS_RESERVE_NO_FLUSH);
1731                 if (pending->error)
1732                         goto clear_skip_qgroup;
1733         }
1734
1735         key.objectid = objectid;
1736         key.offset = (u64)-1;
1737         key.type = BTRFS_ROOT_ITEM_KEY;
1738
1739         rsv = trans->block_rsv;
1740         trans->block_rsv = &pending->block_rsv;
1741         trans->bytes_reserved = trans->block_rsv->reserved;
1742         trace_btrfs_space_reservation(fs_info, "transaction",
1743                                       trans->transid,
1744                                       trans->bytes_reserved, 1);
1745         parent_root = BTRFS_I(parent_inode)->root;
1746         ret = record_root_in_trans(trans, parent_root, 0);
1747         if (ret)
1748                 goto fail;
1749         cur_time = current_time(parent_inode);
1750
1751         /*
1752          * insert the directory item
1753          */
1754         ret = btrfs_set_inode_index(BTRFS_I(parent_inode), &index);
1755         if (ret) {
1756                 btrfs_abort_transaction(trans, ret);
1757                 goto fail;
1758         }
1759
1760         /* check if there is a file/dir which has the same name. */
1761         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1762                                          btrfs_ino(BTRFS_I(parent_inode)),
1763                                          &fname.disk_name, 0);
1764         if (dir_item != NULL && !IS_ERR(dir_item)) {
1765                 pending->error = -EEXIST;
1766                 goto dir_item_existed;
1767         } else if (IS_ERR(dir_item)) {
1768                 ret = PTR_ERR(dir_item);
1769                 btrfs_abort_transaction(trans, ret);
1770                 goto fail;
1771         }
1772         btrfs_release_path(path);
1773
1774         ret = btrfs_create_qgroup(trans, objectid);
1775         if (ret && ret != -EEXIST) {
1776                 btrfs_abort_transaction(trans, ret);
1777                 goto fail;
1778         }
1779
1780         /*
1781          * pull in the delayed directory update
1782          * and the delayed inode item
1783          * otherwise we corrupt the FS during
1784          * snapshot
1785          */
1786         ret = btrfs_run_delayed_items(trans);
1787         if (ret) {      /* Transaction aborted */
1788                 btrfs_abort_transaction(trans, ret);
1789                 goto fail;
1790         }
1791
1792         ret = record_root_in_trans(trans, root, 0);
1793         if (ret) {
1794                 btrfs_abort_transaction(trans, ret);
1795                 goto fail;
1796         }
1797         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1798         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1799         btrfs_check_and_init_root_item(new_root_item);
1800
1801         root_flags = btrfs_root_flags(new_root_item);
1802         if (pending->readonly)
1803                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1804         else
1805                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1806         btrfs_set_root_flags(new_root_item, root_flags);
1807
1808         btrfs_set_root_generation_v2(new_root_item,
1809                         trans->transid);
1810         generate_random_guid(new_root_item->uuid);
1811         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1812                         BTRFS_UUID_SIZE);
1813         if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1814                 memset(new_root_item->received_uuid, 0,
1815                        sizeof(new_root_item->received_uuid));
1816                 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1817                 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1818                 btrfs_set_root_stransid(new_root_item, 0);
1819                 btrfs_set_root_rtransid(new_root_item, 0);
1820         }
1821         btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1822         btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1823         btrfs_set_root_otransid(new_root_item, trans->transid);
1824
1825         old = btrfs_lock_root_node(root);
1826         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old,
1827                               BTRFS_NESTING_COW);
1828         if (ret) {
1829                 btrfs_tree_unlock(old);
1830                 free_extent_buffer(old);
1831                 btrfs_abort_transaction(trans, ret);
1832                 goto fail;
1833         }
1834
1835         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1836         /* clean up in any case */
1837         btrfs_tree_unlock(old);
1838         free_extent_buffer(old);
1839         if (ret) {
1840                 btrfs_abort_transaction(trans, ret);
1841                 goto fail;
1842         }
1843         /* see comments in should_cow_block() */
1844         set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1845         smp_wmb();
1846
1847         btrfs_set_root_node(new_root_item, tmp);
1848         /* record when the snapshot was created in key.offset */
1849         key.offset = trans->transid;
1850         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1851         btrfs_tree_unlock(tmp);
1852         free_extent_buffer(tmp);
1853         if (ret) {
1854                 btrfs_abort_transaction(trans, ret);
1855                 goto fail;
1856         }
1857
1858         /*
1859          * insert root back/forward references
1860          */
1861         ret = btrfs_add_root_ref(trans, objectid,
1862                                  parent_root->root_key.objectid,
1863                                  btrfs_ino(BTRFS_I(parent_inode)), index,
1864                                  &fname.disk_name);
1865         if (ret) {
1866                 btrfs_abort_transaction(trans, ret);
1867                 goto fail;
1868         }
1869
1870         key.offset = (u64)-1;
1871         pending->snap = btrfs_get_new_fs_root(fs_info, objectid, pending->anon_dev);
1872         if (IS_ERR(pending->snap)) {
1873                 ret = PTR_ERR(pending->snap);
1874                 pending->snap = NULL;
1875                 btrfs_abort_transaction(trans, ret);
1876                 goto fail;
1877         }
1878
1879         ret = btrfs_reloc_post_snapshot(trans, pending);
1880         if (ret) {
1881                 btrfs_abort_transaction(trans, ret);
1882                 goto fail;
1883         }
1884
1885         /*
1886          * Do special qgroup accounting for snapshot, as we do some qgroup
1887          * snapshot hack to do fast snapshot.
1888          * To co-operate with that hack, we do hack again.
1889          * Or snapshot will be greatly slowed down by a subtree qgroup rescan
1890          */
1891         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_FULL)
1892                 ret = qgroup_account_snapshot(trans, root, parent_root,
1893                                               pending->inherit, objectid);
1894         else if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE)
1895                 ret = btrfs_qgroup_inherit(trans, root->root_key.objectid, objectid,
1896                                            parent_root->root_key.objectid, pending->inherit);
1897         if (ret < 0)
1898                 goto fail;
1899
1900         ret = btrfs_insert_dir_item(trans, &fname.disk_name,
1901                                     BTRFS_I(parent_inode), &key, BTRFS_FT_DIR,
1902                                     index);
1903         /* We have check then name at the beginning, so it is impossible. */
1904         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1905         if (ret) {
1906                 btrfs_abort_transaction(trans, ret);
1907                 goto fail;
1908         }
1909
1910         btrfs_i_size_write(BTRFS_I(parent_inode), parent_inode->i_size +
1911                                                   fname.disk_name.len * 2);
1912         inode_set_mtime_to_ts(parent_inode,
1913                               inode_set_ctime_current(parent_inode));
1914         ret = btrfs_update_inode_fallback(trans, BTRFS_I(parent_inode));
1915         if (ret) {
1916                 btrfs_abort_transaction(trans, ret);
1917                 goto fail;
1918         }
1919         ret = btrfs_uuid_tree_add(trans, new_root_item->uuid,
1920                                   BTRFS_UUID_KEY_SUBVOL,
1921                                   objectid);
1922         if (ret) {
1923                 btrfs_abort_transaction(trans, ret);
1924                 goto fail;
1925         }
1926         if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1927                 ret = btrfs_uuid_tree_add(trans, new_root_item->received_uuid,
1928                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1929                                           objectid);
1930                 if (ret && ret != -EEXIST) {
1931                         btrfs_abort_transaction(trans, ret);
1932                         goto fail;
1933                 }
1934         }
1935
1936 fail:
1937         pending->error = ret;
1938 dir_item_existed:
1939         trans->block_rsv = rsv;
1940         trans->bytes_reserved = 0;
1941 clear_skip_qgroup:
1942         btrfs_clear_skip_qgroup(trans);
1943 free_fname:
1944         fscrypt_free_filename(&fname);
1945 free_pending:
1946         kfree(new_root_item);
1947         pending->root_item = NULL;
1948         btrfs_free_path(path);
1949         pending->path = NULL;
1950
1951         return ret;
1952 }
1953
1954 /*
1955  * create all the snapshots we've scheduled for creation
1956  */
1957 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans)
1958 {
1959         struct btrfs_pending_snapshot *pending, *next;
1960         struct list_head *head = &trans->transaction->pending_snapshots;
1961         int ret = 0;
1962
1963         list_for_each_entry_safe(pending, next, head, list) {
1964                 list_del(&pending->list);
1965                 ret = create_pending_snapshot(trans, pending);
1966                 if (ret)
1967                         break;
1968         }
1969         return ret;
1970 }
1971
1972 static void update_super_roots(struct btrfs_fs_info *fs_info)
1973 {
1974         struct btrfs_root_item *root_item;
1975         struct btrfs_super_block *super;
1976
1977         super = fs_info->super_copy;
1978
1979         root_item = &fs_info->chunk_root->root_item;
1980         super->chunk_root = root_item->bytenr;
1981         super->chunk_root_generation = root_item->generation;
1982         super->chunk_root_level = root_item->level;
1983
1984         root_item = &fs_info->tree_root->root_item;
1985         super->root = root_item->bytenr;
1986         super->generation = root_item->generation;
1987         super->root_level = root_item->level;
1988         if (btrfs_test_opt(fs_info, SPACE_CACHE))
1989                 super->cache_generation = root_item->generation;
1990         else if (test_bit(BTRFS_FS_CLEANUP_SPACE_CACHE_V1, &fs_info->flags))
1991                 super->cache_generation = 0;
1992         if (test_bit(BTRFS_FS_UPDATE_UUID_TREE_GEN, &fs_info->flags))
1993                 super->uuid_tree_generation = root_item->generation;
1994 }
1995
1996 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1997 {
1998         struct btrfs_transaction *trans;
1999         int ret = 0;
2000
2001         spin_lock(&info->trans_lock);
2002         trans = info->running_transaction;
2003         if (trans)
2004                 ret = (trans->state >= TRANS_STATE_COMMIT_START);
2005         spin_unlock(&info->trans_lock);
2006         return ret;
2007 }
2008
2009 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
2010 {
2011         struct btrfs_transaction *trans;
2012         int ret = 0;
2013
2014         spin_lock(&info->trans_lock);
2015         trans = info->running_transaction;
2016         if (trans)
2017                 ret = is_transaction_blocked(trans);
2018         spin_unlock(&info->trans_lock);
2019         return ret;
2020 }
2021
2022 void btrfs_commit_transaction_async(struct btrfs_trans_handle *trans)
2023 {
2024         struct btrfs_fs_info *fs_info = trans->fs_info;
2025         struct btrfs_transaction *cur_trans;
2026
2027         /* Kick the transaction kthread. */
2028         set_bit(BTRFS_FS_COMMIT_TRANS, &fs_info->flags);
2029         wake_up_process(fs_info->transaction_kthread);
2030
2031         /* take transaction reference */
2032         cur_trans = trans->transaction;
2033         refcount_inc(&cur_trans->use_count);
2034
2035         btrfs_end_transaction(trans);
2036
2037         /*
2038          * Wait for the current transaction commit to start and block
2039          * subsequent transaction joins
2040          */
2041         btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_PREP);
2042         wait_event(fs_info->transaction_blocked_wait,
2043                    cur_trans->state >= TRANS_STATE_COMMIT_START ||
2044                    TRANS_ABORTED(cur_trans));
2045         btrfs_put_transaction(cur_trans);
2046 }
2047
2048 static void cleanup_transaction(struct btrfs_trans_handle *trans, int err)
2049 {
2050         struct btrfs_fs_info *fs_info = trans->fs_info;
2051         struct btrfs_transaction *cur_trans = trans->transaction;
2052
2053         WARN_ON(refcount_read(&trans->use_count) > 1);
2054
2055         btrfs_abort_transaction(trans, err);
2056
2057         spin_lock(&fs_info->trans_lock);
2058
2059         /*
2060          * If the transaction is removed from the list, it means this
2061          * transaction has been committed successfully, so it is impossible
2062          * to call the cleanup function.
2063          */
2064         BUG_ON(list_empty(&cur_trans->list));
2065
2066         if (cur_trans == fs_info->running_transaction) {
2067                 cur_trans->state = TRANS_STATE_COMMIT_DOING;
2068                 spin_unlock(&fs_info->trans_lock);
2069
2070                 /*
2071                  * The thread has already released the lockdep map as reader
2072                  * already in btrfs_commit_transaction().
2073                  */
2074                 btrfs_might_wait_for_event(fs_info, btrfs_trans_num_writers);
2075                 wait_event(cur_trans->writer_wait,
2076                            atomic_read(&cur_trans->num_writers) == 1);
2077
2078                 spin_lock(&fs_info->trans_lock);
2079         }
2080
2081         /*
2082          * Now that we know no one else is still using the transaction we can
2083          * remove the transaction from the list of transactions. This avoids
2084          * the transaction kthread from cleaning up the transaction while some
2085          * other task is still using it, which could result in a use-after-free
2086          * on things like log trees, as it forces the transaction kthread to
2087          * wait for this transaction to be cleaned up by us.
2088          */
2089         list_del_init(&cur_trans->list);
2090
2091         spin_unlock(&fs_info->trans_lock);
2092
2093         btrfs_cleanup_one_transaction(trans->transaction, fs_info);
2094
2095         spin_lock(&fs_info->trans_lock);
2096         if (cur_trans == fs_info->running_transaction)
2097                 fs_info->running_transaction = NULL;
2098         spin_unlock(&fs_info->trans_lock);
2099
2100         if (trans->type & __TRANS_FREEZABLE)
2101                 sb_end_intwrite(fs_info->sb);
2102         btrfs_put_transaction(cur_trans);
2103         btrfs_put_transaction(cur_trans);
2104
2105         trace_btrfs_transaction_commit(fs_info);
2106
2107         if (current->journal_info == trans)
2108                 current->journal_info = NULL;
2109
2110         /*
2111          * If relocation is running, we can't cancel scrub because that will
2112          * result in a deadlock. Before relocating a block group, relocation
2113          * pauses scrub, then starts and commits a transaction before unpausing
2114          * scrub. If the transaction commit is being done by the relocation
2115          * task or triggered by another task and the relocation task is waiting
2116          * for the commit, and we end up here due to an error in the commit
2117          * path, then calling btrfs_scrub_cancel() will deadlock, as we are
2118          * asking for scrub to stop while having it asked to be paused higher
2119          * above in relocation code.
2120          */
2121         if (!test_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags))
2122                 btrfs_scrub_cancel(fs_info);
2123
2124         kmem_cache_free(btrfs_trans_handle_cachep, trans);
2125 }
2126
2127 /*
2128  * Release reserved delayed ref space of all pending block groups of the
2129  * transaction and remove them from the list
2130  */
2131 static void btrfs_cleanup_pending_block_groups(struct btrfs_trans_handle *trans)
2132 {
2133        struct btrfs_fs_info *fs_info = trans->fs_info;
2134        struct btrfs_block_group *block_group, *tmp;
2135
2136        list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
2137                btrfs_dec_delayed_refs_rsv_bg_inserts(fs_info);
2138                list_del_init(&block_group->bg_list);
2139        }
2140 }
2141
2142 static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
2143 {
2144         /*
2145          * We use try_to_writeback_inodes_sb() here because if we used
2146          * btrfs_start_delalloc_roots we would deadlock with fs freeze.
2147          * Currently are holding the fs freeze lock, if we do an async flush
2148          * we'll do btrfs_join_transaction() and deadlock because we need to
2149          * wait for the fs freeze lock.  Using the direct flushing we benefit
2150          * from already being in a transaction and our join_transaction doesn't
2151          * have to re-take the fs freeze lock.
2152          *
2153          * Note that try_to_writeback_inodes_sb() will only trigger writeback
2154          * if it can read lock sb->s_umount. It will always be able to lock it,
2155          * except when the filesystem is being unmounted or being frozen, but in
2156          * those cases sync_filesystem() is called, which results in calling
2157          * writeback_inodes_sb() while holding a write lock on sb->s_umount.
2158          * Note that we don't call writeback_inodes_sb() directly, because it
2159          * will emit a warning if sb->s_umount is not locked.
2160          */
2161         if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
2162                 try_to_writeback_inodes_sb(fs_info->sb, WB_REASON_SYNC);
2163         return 0;
2164 }
2165
2166 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
2167 {
2168         if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
2169                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0, (u64)-1);
2170 }
2171
2172 /*
2173  * Add a pending snapshot associated with the given transaction handle to the
2174  * respective handle. This must be called after the transaction commit started
2175  * and while holding fs_info->trans_lock.
2176  * This serves to guarantee a caller of btrfs_commit_transaction() that it can
2177  * safely free the pending snapshot pointer in case btrfs_commit_transaction()
2178  * returns an error.
2179  */
2180 static void add_pending_snapshot(struct btrfs_trans_handle *trans)
2181 {
2182         struct btrfs_transaction *cur_trans = trans->transaction;
2183
2184         if (!trans->pending_snapshot)
2185                 return;
2186
2187         lockdep_assert_held(&trans->fs_info->trans_lock);
2188         ASSERT(cur_trans->state >= TRANS_STATE_COMMIT_PREP);
2189
2190         list_add(&trans->pending_snapshot->list, &cur_trans->pending_snapshots);
2191 }
2192
2193 static void update_commit_stats(struct btrfs_fs_info *fs_info, ktime_t interval)
2194 {
2195         fs_info->commit_stats.commit_count++;
2196         fs_info->commit_stats.last_commit_dur = interval;
2197         fs_info->commit_stats.max_commit_dur =
2198                         max_t(u64, fs_info->commit_stats.max_commit_dur, interval);
2199         fs_info->commit_stats.total_commit_dur += interval;
2200 }
2201
2202 int btrfs_commit_transaction(struct btrfs_trans_handle *trans)
2203 {
2204         struct btrfs_fs_info *fs_info = trans->fs_info;
2205         struct btrfs_transaction *cur_trans = trans->transaction;
2206         struct btrfs_transaction *prev_trans = NULL;
2207         int ret;
2208         ktime_t start_time;
2209         ktime_t interval;
2210
2211         ASSERT(refcount_read(&trans->use_count) == 1);
2212         btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_PREP);
2213
2214         clear_bit(BTRFS_FS_NEED_TRANS_COMMIT, &fs_info->flags);
2215
2216         /* Stop the commit early if ->aborted is set */
2217         if (TRANS_ABORTED(cur_trans)) {
2218                 ret = cur_trans->aborted;
2219                 goto lockdep_trans_commit_start_release;
2220         }
2221
2222         btrfs_trans_release_metadata(trans);
2223         trans->block_rsv = NULL;
2224
2225         /*
2226          * We only want one transaction commit doing the flushing so we do not
2227          * waste a bunch of time on lock contention on the extent root node.
2228          */
2229         if (!test_and_set_bit(BTRFS_DELAYED_REFS_FLUSHING,
2230                               &cur_trans->delayed_refs.flags)) {
2231                 /*
2232                  * Make a pass through all the delayed refs we have so far.
2233                  * Any running threads may add more while we are here.
2234                  */
2235                 ret = btrfs_run_delayed_refs(trans, 0);
2236                 if (ret)
2237                         goto lockdep_trans_commit_start_release;
2238         }
2239
2240         btrfs_create_pending_block_groups(trans);
2241
2242         if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) {
2243                 int run_it = 0;
2244
2245                 /* this mutex is also taken before trying to set
2246                  * block groups readonly.  We need to make sure
2247                  * that nobody has set a block group readonly
2248                  * after a extents from that block group have been
2249                  * allocated for cache files.  btrfs_set_block_group_ro
2250                  * will wait for the transaction to commit if it
2251                  * finds BTRFS_TRANS_DIRTY_BG_RUN set.
2252                  *
2253                  * The BTRFS_TRANS_DIRTY_BG_RUN flag is also used to make sure
2254                  * only one process starts all the block group IO.  It wouldn't
2255                  * hurt to have more than one go through, but there's no
2256                  * real advantage to it either.
2257                  */
2258                 mutex_lock(&fs_info->ro_block_group_mutex);
2259                 if (!test_and_set_bit(BTRFS_TRANS_DIRTY_BG_RUN,
2260                                       &cur_trans->flags))
2261                         run_it = 1;
2262                 mutex_unlock(&fs_info->ro_block_group_mutex);
2263
2264                 if (run_it) {
2265                         ret = btrfs_start_dirty_block_groups(trans);
2266                         if (ret)
2267                                 goto lockdep_trans_commit_start_release;
2268                 }
2269         }
2270
2271         spin_lock(&fs_info->trans_lock);
2272         if (cur_trans->state >= TRANS_STATE_COMMIT_PREP) {
2273                 enum btrfs_trans_state want_state = TRANS_STATE_COMPLETED;
2274
2275                 add_pending_snapshot(trans);
2276
2277                 spin_unlock(&fs_info->trans_lock);
2278                 refcount_inc(&cur_trans->use_count);
2279
2280                 if (trans->in_fsync)
2281                         want_state = TRANS_STATE_SUPER_COMMITTED;
2282
2283                 btrfs_trans_state_lockdep_release(fs_info,
2284                                                   BTRFS_LOCKDEP_TRANS_COMMIT_PREP);
2285                 ret = btrfs_end_transaction(trans);
2286                 wait_for_commit(cur_trans, want_state);
2287
2288                 if (TRANS_ABORTED(cur_trans))
2289                         ret = cur_trans->aborted;
2290
2291                 btrfs_put_transaction(cur_trans);
2292
2293                 return ret;
2294         }
2295
2296         cur_trans->state = TRANS_STATE_COMMIT_PREP;
2297         wake_up(&fs_info->transaction_blocked_wait);
2298         btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_PREP);
2299
2300         if (cur_trans->list.prev != &fs_info->trans_list) {
2301                 enum btrfs_trans_state want_state = TRANS_STATE_COMPLETED;
2302
2303                 if (trans->in_fsync)
2304                         want_state = TRANS_STATE_SUPER_COMMITTED;
2305
2306                 prev_trans = list_entry(cur_trans->list.prev,
2307                                         struct btrfs_transaction, list);
2308                 if (prev_trans->state < want_state) {
2309                         refcount_inc(&prev_trans->use_count);
2310                         spin_unlock(&fs_info->trans_lock);
2311
2312                         wait_for_commit(prev_trans, want_state);
2313
2314                         ret = READ_ONCE(prev_trans->aborted);
2315
2316                         btrfs_put_transaction(prev_trans);
2317                         if (ret)
2318                                 goto lockdep_release;
2319                         spin_lock(&fs_info->trans_lock);
2320                 }
2321         } else {
2322                 /*
2323                  * The previous transaction was aborted and was already removed
2324                  * from the list of transactions at fs_info->trans_list. So we
2325                  * abort to prevent writing a new superblock that reflects a
2326                  * corrupt state (pointing to trees with unwritten nodes/leafs).
2327                  */
2328                 if (BTRFS_FS_ERROR(fs_info)) {
2329                         spin_unlock(&fs_info->trans_lock);
2330                         ret = -EROFS;
2331                         goto lockdep_release;
2332                 }
2333         }
2334
2335         cur_trans->state = TRANS_STATE_COMMIT_START;
2336         wake_up(&fs_info->transaction_blocked_wait);
2337         spin_unlock(&fs_info->trans_lock);
2338
2339         /*
2340          * Get the time spent on the work done by the commit thread and not
2341          * the time spent waiting on a previous commit
2342          */
2343         start_time = ktime_get_ns();
2344
2345         extwriter_counter_dec(cur_trans, trans->type);
2346
2347         ret = btrfs_start_delalloc_flush(fs_info);
2348         if (ret)
2349                 goto lockdep_release;
2350
2351         ret = btrfs_run_delayed_items(trans);
2352         if (ret)
2353                 goto lockdep_release;
2354
2355         /*
2356          * The thread has started/joined the transaction thus it holds the
2357          * lockdep map as a reader. It has to release it before acquiring the
2358          * lockdep map as a writer.
2359          */
2360         btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters);
2361         btrfs_might_wait_for_event(fs_info, btrfs_trans_num_extwriters);
2362         wait_event(cur_trans->writer_wait,
2363                    extwriter_counter_read(cur_trans) == 0);
2364
2365         /* some pending stuffs might be added after the previous flush. */
2366         ret = btrfs_run_delayed_items(trans);
2367         if (ret) {
2368                 btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
2369                 goto cleanup_transaction;
2370         }
2371
2372         btrfs_wait_delalloc_flush(fs_info);
2373
2374         /*
2375          * Wait for all ordered extents started by a fast fsync that joined this
2376          * transaction. Otherwise if this transaction commits before the ordered
2377          * extents complete we lose logged data after a power failure.
2378          */
2379         btrfs_might_wait_for_event(fs_info, btrfs_trans_pending_ordered);
2380         wait_event(cur_trans->pending_wait,
2381                    atomic_read(&cur_trans->pending_ordered) == 0);
2382
2383         btrfs_scrub_pause(fs_info);
2384         /*
2385          * Ok now we need to make sure to block out any other joins while we
2386          * commit the transaction.  We could have started a join before setting
2387          * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
2388          */
2389         spin_lock(&fs_info->trans_lock);
2390         add_pending_snapshot(trans);
2391         cur_trans->state = TRANS_STATE_COMMIT_DOING;
2392         spin_unlock(&fs_info->trans_lock);
2393
2394         /*
2395          * The thread has started/joined the transaction thus it holds the
2396          * lockdep map as a reader. It has to release it before acquiring the
2397          * lockdep map as a writer.
2398          */
2399         btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
2400         btrfs_might_wait_for_event(fs_info, btrfs_trans_num_writers);
2401         wait_event(cur_trans->writer_wait,
2402                    atomic_read(&cur_trans->num_writers) == 1);
2403
2404         /*
2405          * Make lockdep happy by acquiring the state locks after
2406          * btrfs_trans_num_writers is released. If we acquired the state locks
2407          * before releasing the btrfs_trans_num_writers lock then lockdep would
2408          * complain because we did not follow the reverse order unlocking rule.
2409          */
2410         btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED);
2411         btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED);
2412         btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
2413
2414         /*
2415          * We've started the commit, clear the flag in case we were triggered to
2416          * do an async commit but somebody else started before the transaction
2417          * kthread could do the work.
2418          */
2419         clear_bit(BTRFS_FS_COMMIT_TRANS, &fs_info->flags);
2420
2421         if (TRANS_ABORTED(cur_trans)) {
2422                 ret = cur_trans->aborted;
2423                 btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
2424                 goto scrub_continue;
2425         }
2426         /*
2427          * the reloc mutex makes sure that we stop
2428          * the balancing code from coming in and moving
2429          * extents around in the middle of the commit
2430          */
2431         mutex_lock(&fs_info->reloc_mutex);
2432
2433         /*
2434          * We needn't worry about the delayed items because we will
2435          * deal with them in create_pending_snapshot(), which is the
2436          * core function of the snapshot creation.
2437          */
2438         ret = create_pending_snapshots(trans);
2439         if (ret)
2440                 goto unlock_reloc;
2441
2442         /*
2443          * We insert the dir indexes of the snapshots and update the inode
2444          * of the snapshots' parents after the snapshot creation, so there
2445          * are some delayed items which are not dealt with. Now deal with
2446          * them.
2447          *
2448          * We needn't worry that this operation will corrupt the snapshots,
2449          * because all the tree which are snapshoted will be forced to COW
2450          * the nodes and leaves.
2451          */
2452         ret = btrfs_run_delayed_items(trans);
2453         if (ret)
2454                 goto unlock_reloc;
2455
2456         ret = btrfs_run_delayed_refs(trans, U64_MAX);
2457         if (ret)
2458                 goto unlock_reloc;
2459
2460         /*
2461          * make sure none of the code above managed to slip in a
2462          * delayed item
2463          */
2464         btrfs_assert_delayed_root_empty(fs_info);
2465
2466         WARN_ON(cur_trans != trans->transaction);
2467
2468         ret = commit_fs_roots(trans);
2469         if (ret)
2470                 goto unlock_reloc;
2471
2472         /* commit_fs_roots gets rid of all the tree log roots, it is now
2473          * safe to free the root of tree log roots
2474          */
2475         btrfs_free_log_root_tree(trans, fs_info);
2476
2477         /*
2478          * Since fs roots are all committed, we can get a quite accurate
2479          * new_roots. So let's do quota accounting.
2480          */
2481         ret = btrfs_qgroup_account_extents(trans);
2482         if (ret < 0)
2483                 goto unlock_reloc;
2484
2485         ret = commit_cowonly_roots(trans);
2486         if (ret)
2487                 goto unlock_reloc;
2488
2489         /*
2490          * The tasks which save the space cache and inode cache may also
2491          * update ->aborted, check it.
2492          */
2493         if (TRANS_ABORTED(cur_trans)) {
2494                 ret = cur_trans->aborted;
2495                 goto unlock_reloc;
2496         }
2497
2498         cur_trans = fs_info->running_transaction;
2499
2500         btrfs_set_root_node(&fs_info->tree_root->root_item,
2501                             fs_info->tree_root->node);
2502         list_add_tail(&fs_info->tree_root->dirty_list,
2503                       &cur_trans->switch_commits);
2504
2505         btrfs_set_root_node(&fs_info->chunk_root->root_item,
2506                             fs_info->chunk_root->node);
2507         list_add_tail(&fs_info->chunk_root->dirty_list,
2508                       &cur_trans->switch_commits);
2509
2510         if (btrfs_fs_incompat(fs_info, EXTENT_TREE_V2)) {
2511                 btrfs_set_root_node(&fs_info->block_group_root->root_item,
2512                                     fs_info->block_group_root->node);
2513                 list_add_tail(&fs_info->block_group_root->dirty_list,
2514                               &cur_trans->switch_commits);
2515         }
2516
2517         switch_commit_roots(trans);
2518
2519         ASSERT(list_empty(&cur_trans->dirty_bgs));
2520         ASSERT(list_empty(&cur_trans->io_bgs));
2521         update_super_roots(fs_info);
2522
2523         btrfs_set_super_log_root(fs_info->super_copy, 0);
2524         btrfs_set_super_log_root_level(fs_info->super_copy, 0);
2525         memcpy(fs_info->super_for_commit, fs_info->super_copy,
2526                sizeof(*fs_info->super_copy));
2527
2528         btrfs_commit_device_sizes(cur_trans);
2529
2530         clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
2531         clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
2532
2533         btrfs_trans_release_chunk_metadata(trans);
2534
2535         /*
2536          * Before changing the transaction state to TRANS_STATE_UNBLOCKED and
2537          * setting fs_info->running_transaction to NULL, lock tree_log_mutex to
2538          * make sure that before we commit our superblock, no other task can
2539          * start a new transaction and commit a log tree before we commit our
2540          * superblock. Anyone trying to commit a log tree locks this mutex before
2541          * writing its superblock.
2542          */
2543         mutex_lock(&fs_info->tree_log_mutex);
2544
2545         spin_lock(&fs_info->trans_lock);
2546         cur_trans->state = TRANS_STATE_UNBLOCKED;
2547         fs_info->running_transaction = NULL;
2548         spin_unlock(&fs_info->trans_lock);
2549         mutex_unlock(&fs_info->reloc_mutex);
2550
2551         wake_up(&fs_info->transaction_wait);
2552         btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
2553
2554         /* If we have features changed, wake up the cleaner to update sysfs. */
2555         if (test_bit(BTRFS_FS_FEATURE_CHANGED, &fs_info->flags) &&
2556             fs_info->cleaner_kthread)
2557                 wake_up_process(fs_info->cleaner_kthread);
2558
2559         ret = btrfs_write_and_wait_transaction(trans);
2560         if (ret) {
2561                 btrfs_handle_fs_error(fs_info, ret,
2562                                       "Error while writing out transaction");
2563                 mutex_unlock(&fs_info->tree_log_mutex);
2564                 goto scrub_continue;
2565         }
2566
2567         ret = write_all_supers(fs_info, 0);
2568         /*
2569          * the super is written, we can safely allow the tree-loggers
2570          * to go about their business
2571          */
2572         mutex_unlock(&fs_info->tree_log_mutex);
2573         if (ret)
2574                 goto scrub_continue;
2575
2576         /*
2577          * We needn't acquire the lock here because there is no other task
2578          * which can change it.
2579          */
2580         cur_trans->state = TRANS_STATE_SUPER_COMMITTED;
2581         wake_up(&cur_trans->commit_wait);
2582         btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED);
2583
2584         btrfs_finish_extent_commit(trans);
2585
2586         if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &cur_trans->flags))
2587                 btrfs_clear_space_info_full(fs_info);
2588
2589         btrfs_set_last_trans_committed(fs_info, cur_trans->transid);
2590         /*
2591          * We needn't acquire the lock here because there is no other task
2592          * which can change it.
2593          */
2594         cur_trans->state = TRANS_STATE_COMPLETED;
2595         wake_up(&cur_trans->commit_wait);
2596         btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED);
2597
2598         spin_lock(&fs_info->trans_lock);
2599         list_del_init(&cur_trans->list);
2600         spin_unlock(&fs_info->trans_lock);
2601
2602         btrfs_put_transaction(cur_trans);
2603         btrfs_put_transaction(cur_trans);
2604
2605         if (trans->type & __TRANS_FREEZABLE)
2606                 sb_end_intwrite(fs_info->sb);
2607
2608         trace_btrfs_transaction_commit(fs_info);
2609
2610         interval = ktime_get_ns() - start_time;
2611
2612         btrfs_scrub_continue(fs_info);
2613
2614         if (current->journal_info == trans)
2615                 current->journal_info = NULL;
2616
2617         kmem_cache_free(btrfs_trans_handle_cachep, trans);
2618
2619         update_commit_stats(fs_info, interval);
2620
2621         return ret;
2622
2623 unlock_reloc:
2624         mutex_unlock(&fs_info->reloc_mutex);
2625         btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
2626 scrub_continue:
2627         btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED);
2628         btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED);
2629         btrfs_scrub_continue(fs_info);
2630 cleanup_transaction:
2631         btrfs_trans_release_metadata(trans);
2632         btrfs_cleanup_pending_block_groups(trans);
2633         btrfs_trans_release_chunk_metadata(trans);
2634         trans->block_rsv = NULL;
2635         btrfs_warn(fs_info, "Skipping commit of aborted transaction.");
2636         if (current->journal_info == trans)
2637                 current->journal_info = NULL;
2638         cleanup_transaction(trans, ret);
2639
2640         return ret;
2641
2642 lockdep_release:
2643         btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters);
2644         btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
2645         goto cleanup_transaction;
2646
2647 lockdep_trans_commit_start_release:
2648         btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_PREP);
2649         btrfs_end_transaction(trans);
2650         return ret;
2651 }
2652
2653 /*
2654  * return < 0 if error
2655  * 0 if there are no more dead_roots at the time of call
2656  * 1 there are more to be processed, call me again
2657  *
2658  * The return value indicates there are certainly more snapshots to delete, but
2659  * if there comes a new one during processing, it may return 0. We don't mind,
2660  * because btrfs_commit_super will poke cleaner thread and it will process it a
2661  * few seconds later.
2662  */
2663 int btrfs_clean_one_deleted_snapshot(struct btrfs_fs_info *fs_info)
2664 {
2665         struct btrfs_root *root;
2666         int ret;
2667
2668         spin_lock(&fs_info->trans_lock);
2669         if (list_empty(&fs_info->dead_roots)) {
2670                 spin_unlock(&fs_info->trans_lock);
2671                 return 0;
2672         }
2673         root = list_first_entry(&fs_info->dead_roots,
2674                         struct btrfs_root, root_list);
2675         list_del_init(&root->root_list);
2676         spin_unlock(&fs_info->trans_lock);
2677
2678         btrfs_debug(fs_info, "cleaner removing %llu", root->root_key.objectid);
2679
2680         btrfs_kill_all_delayed_nodes(root);
2681
2682         if (btrfs_header_backref_rev(root->node) <
2683                         BTRFS_MIXED_BACKREF_REV)
2684                 ret = btrfs_drop_snapshot(root, 0, 0);
2685         else
2686                 ret = btrfs_drop_snapshot(root, 1, 0);
2687
2688         btrfs_put_root(root);
2689         return (ret < 0) ? 0 : 1;
2690 }
2691
2692 /*
2693  * We only mark the transaction aborted and then set the file system read-only.
2694  * This will prevent new transactions from starting or trying to join this
2695  * one.
2696  *
2697  * This means that error recovery at the call site is limited to freeing
2698  * any local memory allocations and passing the error code up without
2699  * further cleanup. The transaction should complete as it normally would
2700  * in the call path but will return -EIO.
2701  *
2702  * We'll complete the cleanup in btrfs_end_transaction and
2703  * btrfs_commit_transaction.
2704  */
2705 void __cold __btrfs_abort_transaction(struct btrfs_trans_handle *trans,
2706                                       const char *function,
2707                                       unsigned int line, int error, bool first_hit)
2708 {
2709         struct btrfs_fs_info *fs_info = trans->fs_info;
2710
2711         WRITE_ONCE(trans->aborted, error);
2712         WRITE_ONCE(trans->transaction->aborted, error);
2713         if (first_hit && error == -ENOSPC)
2714                 btrfs_dump_space_info_for_trans_abort(fs_info);
2715         /* Wake up anybody who may be waiting on this transaction */
2716         wake_up(&fs_info->transaction_wait);
2717         wake_up(&fs_info->transaction_blocked_wait);
2718         __btrfs_handle_fs_error(fs_info, function, line, error, NULL);
2719 }
2720
2721 int __init btrfs_transaction_init(void)
2722 {
2723         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
2724                         sizeof(struct btrfs_trans_handle), 0,
2725                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
2726         if (!btrfs_trans_handle_cachep)
2727                 return -ENOMEM;
2728         return 0;
2729 }
2730
2731 void __cold btrfs_transaction_exit(void)
2732 {
2733         kmem_cache_destroy(btrfs_trans_handle_cachep);
2734 }