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