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