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