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