fs/btrfs/transaction.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/fs.h>
   7 #include <linux/slab.h>
   8 #include <linux/sched.h>
   9 #include <linux/writeback.h>
  10 #include <linux/pagemap.h>
  11 #include <linux/blkdev.h>
  12 #include <linux/uuid.h>
  13 #include "misc.h"
  14 #include "ctree.h"
  15 #include "disk-io.h"
  16 #include "transaction.h"
  17 #include "locking.h"
  18 #include "tree-log.h"
  19 #include "inode-map.h"
  20 #include "volumes.h"
  21 #include "dev-replace.h"
  22 #include "qgroup.h"
  23 #include "block-group.h"
  24
  25 #define BTRFS_ROOT_TRANS_TAG 0
  26
  27 /*
  28  * Transaction states and transitions
  29  *
  30  * No running transaction (fs tree blocks are not modified)
  31  * |
  32  * | To next stage:
  33  * |  Call start_transaction() variants. Except btrfs_join_transaction_nostart().
  34  * V
  35  * Transaction N [[TRANS_STATE_RUNNING]]
  36  * |
  37  * | New trans handles can be attached to transaction N by calling all
  38  * | start_transaction() variants.
  39  * |
  40  * | To next stage:
  41  * |  Call btrfs_commit_transaction() on any trans handle attached to
  42  * |  transaction N
  43  * V
  44  * Transaction N [[TRANS_STATE_COMMIT_START]]
  45  * |
  46  * | Will wait for previous running transaction to completely finish if there
  47  * | is one
  48  * |
  49  * | Then one of the following happes:
  50  * | - Wait for all other trans handle holders to release.
  51  * |   The btrfs_commit_transaction() caller will do the commit work.
  52  * | - Wait for current transaction to be committed by others.
  53  * |   Other btrfs_commit_transaction() caller will do the commit work.
  54  * |
  55  * | At this stage, only btrfs_join_transaction*() variants can attach
  56  * | to this running transaction.
  57  * | All other variants will wait for current one to finish and attach to
  58  * | transaction N+1.
  59  * |
  60  * | To next stage:
  61  * |  Caller is chosen to commit transaction N, and all other trans handle
  62  * |  haven been released.
  63  * V
  64  * Transaction N [[TRANS_STATE_COMMIT_DOING]]
  65  * |
  66  * | The heavy lifting transaction work is started.
  67  * | From running delayed refs (modifying extent tree) to creating pending
  68  * | snapshots, running qgroups.
  69  * | In short, modify supporting trees to reflect modifications of subvolume
  70  * | trees.
  71  * |
  72  * | At this stage, all start_transaction() calls will wait for this
  73  * | transaction to finish and attach to transaction N+1.
  74  * |
  75  * | To next stage:
  76  * |  Until all supporting trees are updated.
  77  * V
  78  * Transaction N [[TRANS_STATE_UNBLOCKED]]
  79  * |                                                Transaction N+1
  80  * | All needed trees are modified, thus we only    [[TRANS_STATE_RUNNING]]
  81  * | need to write them back to disk and update     |
  82  * | super blocks.                                  |
  83  * |                                                |
  84  * | At this stage, new transaction is allowed to   |
  85  * | start.                                         |
  86  * | All new start_transaction() calls will be      |
  87  * | attached to transid N+1.                       |
  88  * |                                                |
  89  * | To next stage:                                 |
  90  * |  Until all tree blocks are super blocks are    |
  91  * |  written to block devices                      |
  92  * V                                                |
  93  * Transaction N [[TRANS_STATE_COMPLETED]]          V
  94  *   All tree blocks and super blocks are written.  Transaction N+1
  95  *   This transaction is finished and all its       [[TRANS_STATE_COMMIT_START]]
  96  *   data structures will be cleaned up.            | Life goes on
  97  */
  98 static const unsigned int btrfs_blocked_trans_types[TRANS_STATE_MAX] = {
  99         [TRANS_STATE_RUNNING]           = 0U,
 100         [TRANS_STATE_COMMIT_START]      = (__TRANS_START | __TRANS_ATTACH),
 101         [TRANS_STATE_COMMIT_DOING]      = (__TRANS_START |
 102                                            __TRANS_ATTACH |
 103                                            __TRANS_JOIN |
 104                                            __TRANS_JOIN_NOSTART),
 105         [TRANS_STATE_UNBLOCKED]         = (__TRANS_START |
 106                                            __TRANS_ATTACH |
 107                                            __TRANS_JOIN |
 108                                            __TRANS_JOIN_NOLOCK |
 109                                            __TRANS_JOIN_NOSTART),
 110         [TRANS_STATE_COMPLETED]         = (__TRANS_START |
 111                                            __TRANS_ATTACH |
 112                                            __TRANS_JOIN |
 113                                            __TRANS_JOIN_NOLOCK |
 114                                            __TRANS_JOIN_NOSTART),
 115 };
 116
 117 void btrfs_put_transaction(struct btrfs_transaction *transaction)
 118 {
 119         WARN_ON(refcount_read(&transaction->use_count) == 0);
 120         if (refcount_dec_and_test(&transaction->use_count)) {
 121                 BUG_ON(!list_empty(&transaction->list));
 122                 WARN_ON(!RB_EMPTY_ROOT(
 123                                 &transaction->delayed_refs.href_root.rb_root));
 124                 WARN_ON(!RB_EMPTY_ROOT(
 125                                 &transaction->delayed_refs.dirty_extent_root));
 126                 if (transaction->delayed_refs.pending_csums)
 127                         btrfs_err(transaction->fs_info,
 128                                   "pending csums is %llu",
 129                                   transaction->delayed_refs.pending_csums);
 130                 /*
 131                  * If any block groups are found in ->deleted_bgs then it's
 132                  * because the transaction was aborted and a commit did not
 133                  * happen (things failed before writing the new superblock
 134                  * and calling btrfs_finish_extent_commit()), so we can not
 135                  * discard the physical locations of the block groups.
 136                  */
 137                 while (!list_empty(&transaction->deleted_bgs)) {
 138                         struct btrfs_block_group *cache;
 139
 140                         cache = list_first_entry(&transaction->deleted_bgs,
 141                                                  struct btrfs_block_group,
 142                                                  bg_list);
 143                         list_del_init(&cache->bg_list);
 144                         btrfs_put_block_group_trimming(cache);
 145                         btrfs_put_block_group(cache);
 146                 }
 147                 WARN_ON(!list_empty(&transaction->dev_update_list));
 148                 kfree(transaction);
 149         }
 150 }
 151
 152 static noinline void switch_commit_roots(struct btrfs_trans_handle *trans)
 153 {
 154         struct btrfs_transaction *cur_trans = trans->transaction;
 155         struct btrfs_fs_info *fs_info = trans->fs_info;
 156         struct btrfs_root *root, *tmp;
 157
 158         down_write(&fs_info->commit_root_sem);
 159         list_for_each_entry_safe(root, tmp, &cur_trans->switch_commits,
 160                                  dirty_list) {
 161                 list_del_init(&root->dirty_list);
 162                 free_extent_buffer(root->commit_root);
 163                 root->commit_root = btrfs_root_node(root);
 164                 if (is_fstree(root->root_key.objectid))
 165                         btrfs_unpin_free_ino(root);
 166                 extent_io_tree_release(&root->dirty_log_pages);
 167                 btrfs_qgroup_clean_swapped_blocks(root);
 168         }
 169
 170         /* We can free old roots now. */
 171         spin_lock(&cur_trans->dropped_roots_lock);
 172         while (!list_empty(&cur_trans->dropped_roots)) {
 173                 root = list_first_entry(&cur_trans->dropped_roots,
 174                                         struct btrfs_root, root_list);
 175                 list_del_init(&root->root_list);
 176                 spin_unlock(&cur_trans->dropped_roots_lock);
 177                 btrfs_free_log(trans, root);
 178                 btrfs_drop_and_free_fs_root(fs_info, root);
 179                 spin_lock(&cur_trans->dropped_roots_lock);
 180         }
 181         spin_unlock(&cur_trans->dropped_roots_lock);
 182         up_write(&fs_info->commit_root_sem);
 183 }
 184
 185 static inline void extwriter_counter_inc(struct btrfs_transaction *trans,
 186                                          unsigned int type)
 187 {
 188         if (type & TRANS_EXTWRITERS)
 189                 atomic_inc(&trans->num_extwriters);
 190 }
 191
 192 static inline void extwriter_counter_dec(struct btrfs_transaction *trans,
 193                                          unsigned int type)
 194 {
 195         if (type & TRANS_EXTWRITERS)
 196                 atomic_dec(&trans->num_extwriters);
 197 }
 198
 199 static inline void extwriter_counter_init(struct btrfs_transaction *trans,
 200                                           unsigned int type)
 201 {
 202         atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0));
 203 }
 204
 205 static inline int extwriter_counter_read(struct btrfs_transaction *trans)
 206 {
 207         return atomic_read(&trans->num_extwriters);
 208 }
 209
 210 /*
 211  * To be called after all the new block groups attached to the transaction
 212  * handle have been created (btrfs_create_pending_block_groups()).
 213  */
 214 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
 215 {
 216         struct btrfs_fs_info *fs_info = trans->fs_info;
 217
 218         if (!trans->chunk_bytes_reserved)
 219                 return;
 220
 221         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
 222
 223         btrfs_block_rsv_release(fs_info, &fs_info->chunk_block_rsv,
 224                                 trans->chunk_bytes_reserved, NULL);
 225         trans->chunk_bytes_reserved = 0;
 226 }
 227
 228 /*
 229  * either allocate a new transaction or hop into the existing one
 230  */
 231 static noinline int join_transaction(struct btrfs_fs_info *fs_info,
 232                                      unsigned int type)
 233 {
 234         struct btrfs_transaction *cur_trans;
 235
 236         spin_lock(&fs_info->trans_lock);
 237 loop:
 238         /* The file system has been taken offline. No new transactions. */
 239         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
 240                 spin_unlock(&fs_info->trans_lock);
 241                 return -EROFS;
 242         }
 243
 244         cur_trans = fs_info->running_transaction;
 245         if (cur_trans) {
 246                 if (TRANS_ABORTED(cur_trans)) {
 247                         spin_unlock(&fs_info->trans_lock);
 248                         return cur_trans->aborted;
 249                 }
 250                 if (btrfs_blocked_trans_types[cur_trans->state] & type) {
 251                         spin_unlock(&fs_info->trans_lock);
 252                         return -EBUSY;
 253                 }
 254                 refcount_inc(&cur_trans->use_count);
 255                 atomic_inc(&cur_trans->num_writers);
 256                 extwriter_counter_inc(cur_trans, type);
 257                 spin_unlock(&fs_info->trans_lock);
 258                 return 0;
 259         }
 260         spin_unlock(&fs_info->trans_lock);
 261
 262         /*
 263          * If we are ATTACH, we just want to catch the current transaction,
 264          * and commit it. If there is no transaction, just return ENOENT.
 265          */
 266         if (type == TRANS_ATTACH)
 267                 return -ENOENT;
 268
 269         /*
 270          * JOIN_NOLOCK only happens during the transaction commit, so
 271          * it is impossible that ->running_transaction is NULL
 272          */
 273         BUG_ON(type == TRANS_JOIN_NOLOCK);
 274
 275         cur_trans = kmalloc(sizeof(*cur_trans), GFP_NOFS);
 276         if (!cur_trans)
 277                 return -ENOMEM;
 278
 279         spin_lock(&fs_info->trans_lock);
 280         if (fs_info->running_transaction) {
 281                 /*
 282                  * someone started a transaction after we unlocked.  Make sure
 283                  * to redo the checks above
 284                  */
 285                 kfree(cur_trans);
 286                 goto loop;
 287         } else if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
 288                 spin_unlock(&fs_info->trans_lock);
 289                 kfree(cur_trans);
 290                 return -EROFS;
 291         }
 292
 293         cur_trans->fs_info = fs_info;
 294         atomic_set(&cur_trans->num_writers, 1);
 295         extwriter_counter_init(cur_trans, type);
 296         init_waitqueue_head(&cur_trans->writer_wait);
 297         init_waitqueue_head(&cur_trans->commit_wait);
 298         cur_trans->state = TRANS_STATE_RUNNING;
 299         /*
 300          * One for this trans handle, one so it will live on until we
 301          * commit the transaction.
 302          */
 303         refcount_set(&cur_trans->use_count, 2);
 304         cur_trans->flags = 0;
 305         cur_trans->start_time = ktime_get_seconds();
 306
 307         memset(&cur_trans->delayed_refs, 0, sizeof(cur_trans->delayed_refs));
 308
 309         cur_trans->delayed_refs.href_root = RB_ROOT_CACHED;
 310         cur_trans->delayed_refs.dirty_extent_root = RB_ROOT;
 311         atomic_set(&cur_trans->delayed_refs.num_entries, 0);
 312
 313         /*
 314          * although the tree mod log is per file system and not per transaction,
 315          * the log must never go across transaction boundaries.
 316          */
 317         smp_mb();
 318         if (!list_empty(&fs_info->tree_mod_seq_list))
 319                 WARN(1, KERN_ERR "BTRFS: tree_mod_seq_list not empty when creating a fresh transaction\n");
 320         if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
 321                 WARN(1, KERN_ERR "BTRFS: tree_mod_log rb tree not empty when creating a fresh transaction\n");
 322         atomic64_set(&fs_info->tree_mod_seq, 0);
 323
 324         spin_lock_init(&cur_trans->delayed_refs.lock);
 325
 326         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 327         INIT_LIST_HEAD(&cur_trans->dev_update_list);
 328         INIT_LIST_HEAD(&cur_trans->switch_commits);
 329         INIT_LIST_HEAD(&cur_trans->dirty_bgs);
 330         INIT_LIST_HEAD(&cur_trans->io_bgs);
 331         INIT_LIST_HEAD(&cur_trans->dropped_roots);
 332         mutex_init(&cur_trans->cache_write_mutex);
 333         spin_lock_init(&cur_trans->dirty_bgs_lock);
 334         INIT_LIST_HEAD(&cur_trans->deleted_bgs);
 335         spin_lock_init(&cur_trans->dropped_roots_lock);
 336         list_add_tail(&cur_trans->list, &fs_info->trans_list);
 337         extent_io_tree_init(fs_info, &cur_trans->dirty_pages,
 338                         IO_TREE_TRANS_DIRTY_PAGES, fs_info->btree_inode);
 339         extent_io_tree_init(fs_info, &cur_trans->pinned_extents,
 340                         IO_TREE_FS_PINNED_EXTENTS, NULL);
 341         fs_info->generation++;
 342         cur_trans->transid = fs_info->generation;
 343         fs_info->running_transaction = cur_trans;
 344         cur_trans->aborted = 0;
 345         spin_unlock(&fs_info->trans_lock);
 346
 347         return 0;
 348 }
 349
 350 /*
 351  * this does all the record keeping required to make sure that a reference
 352  * counted root is properly recorded in a given transaction.  This is required
 353  * to make sure the old root from before we joined the transaction is deleted
 354  * when the transaction commits
 355  */
 356 static int record_root_in_trans(struct btrfs_trans_handle *trans,
 357                                struct btrfs_root *root,
 358                                int force)
 359 {
 360         struct btrfs_fs_info *fs_info = root->fs_info;
 361
 362         if ((test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
 363             root->last_trans < trans->transid) || force) {
 364                 WARN_ON(root == fs_info->extent_root);
 365                 WARN_ON(!force && root->commit_root != root->node);
 366
 367                 /*
 368                  * see below for IN_TRANS_SETUP usage rules
 369                  * we have the reloc mutex held now, so there
 370                  * is only one writer in this function
 371                  */
 372                 set_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
 373
 374                 /* make sure readers find IN_TRANS_SETUP before
 375                  * they find our root->last_trans update
 376                  */
 377                 smp_wmb();
 378
 379                 spin_lock(&fs_info->fs_roots_radix_lock);
 380                 if (root->last_trans == trans->transid && !force) {
 381                         spin_unlock(&fs_info->fs_roots_radix_lock);
 382                         return 0;
 383                 }
 384                 radix_tree_tag_set(&fs_info->fs_roots_radix,
 385                                    (unsigned long)root->root_key.objectid,
 386                                    BTRFS_ROOT_TRANS_TAG);
 387                 spin_unlock(&fs_info->fs_roots_radix_lock);
 388                 root->last_trans = trans->transid;
 389
 390                 /* this is pretty tricky.  We don't want to
 391                  * take the relocation lock in btrfs_record_root_in_trans
 392                  * unless we're really doing the first setup for this root in
 393                  * this transaction.
 394                  *
 395                  * Normally we'd use root->last_trans as a flag to decide
 396                  * if we want to take the expensive mutex.
 397                  *
 398                  * But, we have to set root->last_trans before we
 399                  * init the relocation root, otherwise, we trip over warnings
 400                  * in ctree.c.  The solution used here is to flag ourselves
 401                  * with root IN_TRANS_SETUP.  When this is 1, we're still
 402                  * fixing up the reloc trees and everyone must wait.
 403                  *
 404                  * When this is zero, they can trust root->last_trans and fly
 405                  * through btrfs_record_root_in_trans without having to take the
 406                  * lock.  smp_wmb() makes sure that all the writes above are
 407                  * done before we pop in the zero below
 408                  */
 409                 btrfs_init_reloc_root(trans, root);
 410                 smp_mb__before_atomic();
 411                 clear_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
 412         }
 413         return 0;
 414 }
 415
 416
 417 void btrfs_add_dropped_root(struct btrfs_trans_handle *trans,
 418                             struct btrfs_root *root)
 419 {
 420         struct btrfs_fs_info *fs_info = root->fs_info;
 421         struct btrfs_transaction *cur_trans = trans->transaction;
 422
 423         /* Add ourselves to the transaction dropped list */
 424         spin_lock(&cur_trans->dropped_roots_lock);
 425         list_add_tail(&root->root_list, &cur_trans->dropped_roots);
 426         spin_unlock(&cur_trans->dropped_roots_lock);
 427
 428         /* Make sure we don't try to update the root at commit time */
 429         spin_lock(&fs_info->fs_roots_radix_lock);
 430         radix_tree_tag_clear(&fs_info->fs_roots_radix,
 431                              (unsigned long)root->root_key.objectid,
 432                              BTRFS_ROOT_TRANS_TAG);
 433         spin_unlock(&fs_info->fs_roots_radix_lock);
 434 }
 435
 436 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 437                                struct btrfs_root *root)
 438 {
 439         struct btrfs_fs_info *fs_info = root->fs_info;
 440
 441         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
 442                 return 0;
 443
 444         /*
 445          * see record_root_in_trans for comments about IN_TRANS_SETUP usage
 446          * and barriers
 447          */
 448         smp_rmb();
 449         if (root->last_trans == trans->transid &&
 450             !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state))
 451                 return 0;
 452
 453         mutex_lock(&fs_info->reloc_mutex);
 454         record_root_in_trans(trans, root, 0);
 455         mutex_unlock(&fs_info->reloc_mutex);
 456
 457         return 0;
 458 }
 459
 460 static inline int is_transaction_blocked(struct btrfs_transaction *trans)
 461 {
 462         return (trans->state >= TRANS_STATE_COMMIT_START &&
 463                 trans->state < TRANS_STATE_UNBLOCKED &&
 464                 !TRANS_ABORTED(trans));
 465 }
 466
 467 /* wait for commit against the current transaction to become unblocked
 468  * when this is done, it is safe to start a new transaction, but the current
 469  * transaction might not be fully on disk.
 470  */
 471 static void wait_current_trans(struct btrfs_fs_info *fs_info)
 472 {
 473         struct btrfs_transaction *cur_trans;
 474
 475         spin_lock(&fs_info->trans_lock);
 476         cur_trans = fs_info->running_transaction;
 477         if (cur_trans && is_transaction_blocked(cur_trans)) {
 478                 refcount_inc(&cur_trans->use_count);
 479                 spin_unlock(&fs_info->trans_lock);
 480
 481                 wait_event(fs_info->transaction_wait,
 482                            cur_trans->state >= TRANS_STATE_UNBLOCKED ||
 483                            TRANS_ABORTED(cur_trans));
 484                 btrfs_put_transaction(cur_trans);
 485         } else {
 486                 spin_unlock(&fs_info->trans_lock);
 487         }
 488 }
 489
 490 static int may_wait_transaction(struct btrfs_fs_info *fs_info, int type)
 491 {
 492         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
 493                 return 0;
 494
 495         if (type == TRANS_START)
 496                 return 1;
 497
 498         return 0;
 499 }
 500
 501 static inline bool need_reserve_reloc_root(struct btrfs_root *root)
 502 {
 503         struct btrfs_fs_info *fs_info = root->fs_info;
 504
 505         if (!fs_info->reloc_ctl ||
 506             !test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
 507             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 508             root->reloc_root)
 509                 return false;
 510
 511         return true;
 512 }
 513
 514 static struct btrfs_trans_handle *
 515 start_transaction(struct btrfs_root *root, unsigned int num_items,
 516                   unsigned int type, enum btrfs_reserve_flush_enum flush,
 517                   bool enforce_qgroups)
 518 {
 519         struct btrfs_fs_info *fs_info = root->fs_info;
 520         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
 521         struct btrfs_trans_handle *h;
 522         struct btrfs_transaction *cur_trans;
 523         u64 num_bytes = 0;
 524         u64 qgroup_reserved = 0;
 525         bool reloc_reserved = false;
 526         int ret;
 527
 528         /* Send isn't supposed to start transactions. */
 529         ASSERT(current->journal_info != BTRFS_SEND_TRANS_STUB);
 530
 531         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
 532                 return ERR_PTR(-EROFS);
 533
 534         if (current->journal_info) {
 535                 WARN_ON(type & TRANS_EXTWRITERS);
 536                 h = current->journal_info;
 537                 refcount_inc(&h->use_count);
 538                 WARN_ON(refcount_read(&h->use_count) > 2);
 539                 h->orig_rsv = h->block_rsv;
 540                 h->block_rsv = NULL;
 541                 goto got_it;
 542         }
 543
 544         /*
 545          * Do the reservation before we join the transaction so we can do all
 546          * the appropriate flushing if need be.
 547          */
 548         if (num_items && root != fs_info->chunk_root) {
 549                 struct btrfs_block_rsv *rsv = &fs_info->trans_block_rsv;
 550                 u64 delayed_refs_bytes = 0;
 551
 552                 qgroup_reserved = num_items * fs_info->nodesize;
 553                 ret = btrfs_qgroup_reserve_meta_pertrans(root, qgroup_reserved,
 554                                 enforce_qgroups);
 555                 if (ret)
 556                         return ERR_PTR(ret);
 557
 558                 /*
 559                  * We want to reserve all the bytes we may need all at once, so
 560                  * we only do 1 enospc flushing cycle per transaction start.  We
 561                  * accomplish this by simply assuming we'll do 2 x num_items
 562                  * worth of delayed refs updates in this trans handle, and
 563                  * refill that amount for whatever is missing in the reserve.
 564                  */
 565                 num_bytes = btrfs_calc_insert_metadata_size(fs_info, num_items);
 566                 if (delayed_refs_rsv->full == 0) {
 567                         delayed_refs_bytes = num_bytes;
 568                         num_bytes <<= 1;
 569                 }
 570
 571                 /*
 572                  * Do the reservation for the relocation root creation
 573                  */
 574                 if (need_reserve_reloc_root(root)) {
 575                         num_bytes += fs_info->nodesize;
 576                         reloc_reserved = true;
 577                 }
 578
 579                 ret = btrfs_block_rsv_add(root, rsv, num_bytes, flush);
 580                 if (ret)
 581                         goto reserve_fail;
 582                 if (delayed_refs_bytes) {
 583                         btrfs_migrate_to_delayed_refs_rsv(fs_info, rsv,
 584                                                           delayed_refs_bytes);
 585                         num_bytes -= delayed_refs_bytes;
 586                 }
 587         } else if (num_items == 0 && flush == BTRFS_RESERVE_FLUSH_ALL &&
 588                    !delayed_refs_rsv->full) {
 589                 /*
 590                  * Some people call with btrfs_start_transaction(root, 0)
 591                  * because they can be throttled, but have some other mechanism
 592                  * for reserving space.  We still want these guys to refill the
 593                  * delayed block_rsv so just add 1 items worth of reservation
 594                  * here.
 595                  */
 596                 ret = btrfs_delayed_refs_rsv_refill(fs_info, flush);
 597                 if (ret)
 598                         goto reserve_fail;
 599         }
 600 again:
 601         h = kmem_cache_zalloc(btrfs_trans_handle_cachep, GFP_NOFS);
 602         if (!h) {
 603                 ret = -ENOMEM;
 604                 goto alloc_fail;
 605         }
 606
 607         /*
 608          * If we are JOIN_NOLOCK we're already committing a transaction and
 609          * waiting on this guy, so we don't need to do the sb_start_intwrite
 610          * because we're already holding a ref.  We need this because we could
 611          * have raced in and did an fsync() on a file which can kick a commit
 612          * and then we deadlock with somebody doing a freeze.
 613          *
 614          * If we are ATTACH, it means we just want to catch the current
 615          * transaction and commit it, so we needn't do sb_start_intwrite().
 616          */
 617         if (type & __TRANS_FREEZABLE)
 618                 sb_start_intwrite(fs_info->sb);
 619
 620         if (may_wait_transaction(fs_info, type))
 621                 wait_current_trans(fs_info);
 622
 623         do {
 624                 ret = join_transaction(fs_info, type);
 625                 if (ret == -EBUSY) {
 626                         wait_current_trans(fs_info);
 627                         if (unlikely(type == TRANS_ATTACH ||
 628                                      type == TRANS_JOIN_NOSTART))
 629                                 ret = -ENOENT;
 630                 }
 631         } while (ret == -EBUSY);
 632
 633         if (ret < 0)
 634                 goto join_fail;
 635
 636         cur_trans = fs_info->running_transaction;
 637
 638         h->transid = cur_trans->transid;
 639         h->transaction = cur_trans;
 640         h->root = root;
 641         refcount_set(&h->use_count, 1);
 642         h->fs_info = root->fs_info;
 643
 644         h->type = type;
 645         h->can_flush_pending_bgs = true;
 646         INIT_LIST_HEAD(&h->new_bgs);
 647
 648         smp_mb();
 649         if (cur_trans->state >= TRANS_STATE_COMMIT_START &&
 650             may_wait_transaction(fs_info, type)) {
 651                 current->journal_info = h;
 652                 btrfs_commit_transaction(h);
 653                 goto again;
 654         }
 655
 656         if (num_bytes) {
 657                 trace_btrfs_space_reservation(fs_info, "transaction",
 658                                               h->transid, num_bytes, 1);
 659                 h->block_rsv = &fs_info->trans_block_rsv;
 660                 h->bytes_reserved = num_bytes;
 661                 h->reloc_reserved = reloc_reserved;
 662         }
 663
 664 got_it:
 665         if (!current->journal_info)
 666                 current->journal_info = h;
 667
 668         /*
 669          * btrfs_record_root_in_trans() needs to alloc new extents, and may
 670          * call btrfs_join_transaction() while we're also starting a
 671          * transaction.
 672          *
 673          * Thus it need to be called after current->journal_info initialized,
 674          * or we can deadlock.
 675          */
 676         btrfs_record_root_in_trans(h, root);
 677
 678         return h;
 679
 680 join_fail:
 681         if (type & __TRANS_FREEZABLE)
 682                 sb_end_intwrite(fs_info->sb);
 683         kmem_cache_free(btrfs_trans_handle_cachep, h);
 684 alloc_fail:
 685         if (num_bytes)
 686                 btrfs_block_rsv_release(fs_info, &fs_info->trans_block_rsv,
 687                                         num_bytes, NULL);
 688 reserve_fail:
 689         btrfs_qgroup_free_meta_pertrans(root, qgroup_reserved);
 690         return ERR_PTR(ret);
 691 }
 692
 693 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 694                                                    unsigned int num_items)
 695 {
 696         return start_transaction(root, num_items, TRANS_START,
 697                                  BTRFS_RESERVE_FLUSH_ALL, true);
 698 }
 699
 700 struct btrfs_trans_handle *btrfs_start_transaction_fallback_global_rsv(
 701                                         struct btrfs_root *root,
 702                                         unsigned int num_items,
 703                                         int min_factor)
 704 {
 705         struct btrfs_fs_info *fs_info = root->fs_info;
 706         struct btrfs_trans_handle *trans;
 707         u64 num_bytes;
 708         int ret;
 709
 710         /*
 711          * We have two callers: unlink and block group removal.  The
 712          * former should succeed even if we will temporarily exceed
 713          * quota and the latter operates on the extent root so
 714          * qgroup enforcement is ignored anyway.
 715          */
 716         trans = start_transaction(root, num_items, TRANS_START,
 717                                   BTRFS_RESERVE_FLUSH_ALL, false);
 718         if (!IS_ERR(trans) || PTR_ERR(trans) != -ENOSPC)
 719                 return trans;
 720
 721         trans = btrfs_start_transaction(root, 0);
 722         if (IS_ERR(trans))
 723                 return trans;
 724
 725         num_bytes = btrfs_calc_insert_metadata_size(fs_info, num_items);
 726         ret = btrfs_cond_migrate_bytes(fs_info, &fs_info->trans_block_rsv,
 727                                        num_bytes, min_factor);
 728         if (ret) {
 729                 btrfs_end_transaction(trans);
 730                 return ERR_PTR(ret);
 731         }
 732
 733         trans->block_rsv = &fs_info->trans_block_rsv;
 734         trans->bytes_reserved = num_bytes;
 735         trace_btrfs_space_reservation(fs_info, "transaction",
 736                                       trans->transid, num_bytes, 1);
 737
 738         return trans;
 739 }
 740
 741 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 742 {
 743         return start_transaction(root, 0, TRANS_JOIN, BTRFS_RESERVE_NO_FLUSH,
 744                                  true);
 745 }
 746
 747 struct btrfs_trans_handle *btrfs_join_transaction_spacecache(struct btrfs_root *root)
 748 {
 749         return start_transaction(root, 0, TRANS_JOIN_NOLOCK,
 750                                  BTRFS_RESERVE_NO_FLUSH, true);
 751 }
 752
 753 /*
 754  * Similar to regular join but it never starts a transaction when none is
 755  * running or after waiting for the current one to finish.
 756  */
 757 struct btrfs_trans_handle *btrfs_join_transaction_nostart(struct btrfs_root *root)
 758 {
 759         return start_transaction(root, 0, TRANS_JOIN_NOSTART,
 760                                  BTRFS_RESERVE_NO_FLUSH, true);
 761 }
 762
 763 /*
 764  * btrfs_attach_transaction() - catch the running transaction
 765  *
 766  * It is used when we want to commit the current the transaction, but
 767  * don't want to start a new one.
 768  *
 769  * Note: If this function return -ENOENT, it just means there is no
 770  * running transaction. But it is possible that the inactive transaction
 771  * is still in the memory, not fully on disk. If you hope there is no
 772  * inactive transaction in the fs when -ENOENT is returned, you should
 773  * invoke
 774  *     btrfs_attach_transaction_barrier()
 775  */
 776 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
 777 {
 778         return start_transaction(root, 0, TRANS_ATTACH,
 779                                  BTRFS_RESERVE_NO_FLUSH, true);
 780 }
 781
 782 /*
 783  * btrfs_attach_transaction_barrier() - catch the running transaction
 784  *
 785  * It is similar to the above function, the difference is this one
 786  * will wait for all the inactive transactions until they fully
 787  * complete.
 788  */
 789 struct btrfs_trans_handle *
 790 btrfs_attach_transaction_barrier(struct btrfs_root *root)
 791 {
 792         struct btrfs_trans_handle *trans;
 793
 794         trans = start_transaction(root, 0, TRANS_ATTACH,
 795                                   BTRFS_RESERVE_NO_FLUSH, true);
 796         if (trans == ERR_PTR(-ENOENT))
 797                 btrfs_wait_for_commit(root->fs_info, 0);
 798
 799         return trans;
 800 }
 801
 802 /* wait for a transaction commit to be fully complete */
 803 static noinline void wait_for_commit(struct btrfs_transaction *commit)
 804 {
 805         wait_event(commit->commit_wait, commit->state == TRANS_STATE_COMPLETED);
 806 }
 807
 808 int btrfs_wait_for_commit(struct btrfs_fs_info *fs_info, u64 transid)
 809 {
 810         struct btrfs_transaction *cur_trans = NULL, *t;
 811         int ret = 0;
 812
 813         if (transid) {
 814                 if (transid <= fs_info->last_trans_committed)
 815                         goto out;
 816
 817                 /* find specified transaction */
 818                 spin_lock(&fs_info->trans_lock);
 819                 list_for_each_entry(t, &fs_info->trans_list, list) {
 820                         if (t->transid == transid) {
 821                                 cur_trans = t;
 822                                 refcount_inc(&cur_trans->use_count);
 823                                 ret = 0;
 824                                 break;
 825                         }
 826                         if (t->transid > transid) {
 827                                 ret = 0;
 828                                 break;
 829                         }
 830                 }
 831                 spin_unlock(&fs_info->trans_lock);
 832
 833                 /*
 834                  * The specified transaction doesn't exist, or we
 835                  * raced with btrfs_commit_transaction
 836                  */
 837                 if (!cur_trans) {
 838                         if (transid > fs_info->last_trans_committed)
 839                                 ret = -EINVAL;
 840                         goto out;
 841                 }
 842         } else {
 843                 /* find newest transaction that is committing | committed */
 844                 spin_lock(&fs_info->trans_lock);
 845                 list_for_each_entry_reverse(t, &fs_info->trans_list,
 846                                             list) {
 847                         if (t->state >= TRANS_STATE_COMMIT_START) {
 848                                 if (t->state == TRANS_STATE_COMPLETED)
 849                                         break;
 850                                 cur_trans = t;
 851                                 refcount_inc(&cur_trans->use_count);
 852                                 break;
 853                         }
 854                 }
 855                 spin_unlock(&fs_info->trans_lock);
 856                 if (!cur_trans)
 857                         goto out;  /* nothing committing|committed */
 858         }
 859
 860         wait_for_commit(cur_trans);
 861         btrfs_put_transaction(cur_trans);
 862 out:
 863         return ret;
 864 }
 865
 866 void btrfs_throttle(struct btrfs_fs_info *fs_info)
 867 {
 868         wait_current_trans(fs_info);
 869 }
 870
 871 static int should_end_transaction(struct btrfs_trans_handle *trans)
 872 {
 873         struct btrfs_fs_info *fs_info = trans->fs_info;
 874
 875         if (btrfs_check_space_for_delayed_refs(fs_info))
 876                 return 1;
 877
 878         return !!btrfs_block_rsv_check(&fs_info->global_block_rsv, 5);
 879 }
 880
 881 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans)
 882 {
 883         struct btrfs_transaction *cur_trans = trans->transaction;
 884
 885         smp_mb();
 886         if (cur_trans->state >= TRANS_STATE_COMMIT_START ||
 887             cur_trans->delayed_refs.flushing)
 888                 return 1;
 889
 890         return should_end_transaction(trans);
 891 }
 892
 893 static void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans)
 894
 895 {
 896         struct btrfs_fs_info *fs_info = trans->fs_info;
 897
 898         if (!trans->block_rsv) {
 899                 ASSERT(!trans->bytes_reserved);
 900                 return;
 901         }
 902
 903         if (!trans->bytes_reserved)
 904                 return;
 905
 906         ASSERT(trans->block_rsv == &fs_info->trans_block_rsv);
 907         trace_btrfs_space_reservation(fs_info, "transaction",
 908                                       trans->transid, trans->bytes_reserved, 0);
 909         btrfs_block_rsv_release(fs_info, trans->block_rsv,
 910                                 trans->bytes_reserved, NULL);
 911         trans->bytes_reserved = 0;
 912 }
 913
 914 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 915                                    int throttle)
 916 {
 917         struct btrfs_fs_info *info = trans->fs_info;
 918         struct btrfs_transaction *cur_trans = trans->transaction;
 919         int err = 0;
 920
 921         if (refcount_read(&trans->use_count) > 1) {
 922                 refcount_dec(&trans->use_count);
 923                 trans->block_rsv = trans->orig_rsv;
 924                 return 0;
 925         }
 926
 927         btrfs_trans_release_metadata(trans);
 928         trans->block_rsv = NULL;
 929
 930         btrfs_create_pending_block_groups(trans);
 931
 932         btrfs_trans_release_chunk_metadata(trans);
 933
 934         if (trans->type & __TRANS_FREEZABLE)
 935                 sb_end_intwrite(info->sb);
 936
 937         WARN_ON(cur_trans != info->running_transaction);
 938         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 939         atomic_dec(&cur_trans->num_writers);
 940         extwriter_counter_dec(cur_trans, trans->type);
 941
 942         cond_wake_up(&cur_trans->writer_wait);
 943         btrfs_put_transaction(cur_trans);
 944
 945         if (current->journal_info == trans)
 946                 current->journal_info = NULL;
 947
 948         if (throttle)
 949                 btrfs_run_delayed_iputs(info);
 950
 951         if (TRANS_ABORTED(trans) ||
 952             test_bit(BTRFS_FS_STATE_ERROR, &info->fs_state)) {
 953                 wake_up_process(info->transaction_kthread);
 954                 err = -EIO;
 955         }
 956
 957         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 958         return err;
 959 }
 960
 961 int btrfs_end_transaction(struct btrfs_trans_handle *trans)
 962 {
 963         return __btrfs_end_transaction(trans, 0);
 964 }
 965
 966 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans)
 967 {
 968         return __btrfs_end_transaction(trans, 1);
 969 }
 970
 971 /*
 972  * when btree blocks are allocated, they have some corresponding bits set for
 973  * them in one of two extent_io trees.  This is used to make sure all of
 974  * those extents are sent to disk but does not wait on them
 975  */
 976 int btrfs_write_marked_extents(struct btrfs_fs_info *fs_info,
 977                                struct extent_io_tree *dirty_pages, int mark)
 978 {
 979         int err = 0;
 980         int werr = 0;
 981         struct address_space *mapping = fs_info->btree_inode->i_mapping;
 982         struct extent_state *cached_state = NULL;
 983         u64 start = 0;
 984         u64 end;
 985
 986         atomic_inc(&BTRFS_I(fs_info->btree_inode)->sync_writers);
 987         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 988                                       mark, &cached_state)) {
 989                 bool wait_writeback = false;
 990
 991                 err = convert_extent_bit(dirty_pages, start, end,
 992                                          EXTENT_NEED_WAIT,
 993                                          mark, &cached_state);
 994                 /*
 995                  * convert_extent_bit can return -ENOMEM, which is most of the
 996                  * time a temporary error. So when it happens, ignore the error
 997                  * and wait for writeback of this range to finish - because we
 998                  * failed to set the bit EXTENT_NEED_WAIT for the range, a call
 999                  * to __btrfs_wait_marked_extents() would not know that
1000                  * writeback for this range started and therefore wouldn't
1001                  * wait for it to finish - we don't want to commit a
1002                  * superblock that points to btree nodes/leafs for which
1003                  * writeback hasn't finished yet (and without errors).
1004                  * We cleanup any entries left in the io tree when committing
1005                  * the transaction (through extent_io_tree_release()).
1006                  */
1007                 if (err == -ENOMEM) {
1008                         err = 0;
1009                         wait_writeback = true;
1010                 }
1011                 if (!err)
1012                         err = filemap_fdatawrite_range(mapping, start, end);
1013                 if (err)
1014                         werr = err;
1015                 else if (wait_writeback)
1016                         werr = filemap_fdatawait_range(mapping, start, end);
1017                 free_extent_state(cached_state);
1018                 cached_state = NULL;
1019                 cond_resched();
1020                 start = end + 1;
1021         }
1022         atomic_dec(&BTRFS_I(fs_info->btree_inode)->sync_writers);
1023         return werr;
1024 }
1025
1026 /*
1027  * when btree blocks are allocated, they have some corresponding bits set for
1028  * them in one of two extent_io trees.  This is used to make sure all of
1029  * those extents are on disk for transaction or log commit.  We wait
1030  * on all the pages and clear them from the dirty pages state tree
1031  */
1032 static int __btrfs_wait_marked_extents(struct btrfs_fs_info *fs_info,
1033                                        struct extent_io_tree *dirty_pages)
1034 {
1035         int err = 0;
1036         int werr = 0;
1037         struct address_space *mapping = fs_info->btree_inode->i_mapping;
1038         struct extent_state *cached_state = NULL;
1039         u64 start = 0;
1040         u64 end;
1041
1042         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
1043                                       EXTENT_NEED_WAIT, &cached_state)) {
1044                 /*
1045                  * Ignore -ENOMEM errors returned by clear_extent_bit().
1046                  * When committing the transaction, we'll remove any entries
1047                  * left in the io tree. For a log commit, we don't remove them
1048                  * after committing the log because the tree can be accessed
1049                  * concurrently - we do it only at transaction commit time when
1050                  * it's safe to do it (through extent_io_tree_release()).
1051                  */
1052                 err = clear_extent_bit(dirty_pages, start, end,
1053                                        EXTENT_NEED_WAIT, 0, 0, &cached_state);
1054                 if (err == -ENOMEM)
1055                         err = 0;
1056                 if (!err)
1057                         err = filemap_fdatawait_range(mapping, start, end);
1058                 if (err)
1059                         werr = err;
1060                 free_extent_state(cached_state);
1061                 cached_state = NULL;
1062                 cond_resched();
1063                 start = end + 1;
1064         }
1065         if (err)
1066                 werr = err;
1067         return werr;
1068 }
1069
1070 static int btrfs_wait_extents(struct btrfs_fs_info *fs_info,
1071                        struct extent_io_tree *dirty_pages)
1072 {
1073         bool errors = false;
1074         int err;
1075
1076         err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1077         if (test_and_clear_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags))
1078                 errors = true;
1079
1080         if (errors && !err)
1081                 err = -EIO;
1082         return err;
1083 }
1084
1085 int btrfs_wait_tree_log_extents(struct btrfs_root *log_root, int mark)
1086 {
1087         struct btrfs_fs_info *fs_info = log_root->fs_info;
1088         struct extent_io_tree *dirty_pages = &log_root->dirty_log_pages;
1089         bool errors = false;
1090         int err;
1091
1092         ASSERT(log_root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID);
1093
1094         err = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1095         if ((mark & EXTENT_DIRTY) &&
1096             test_and_clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags))
1097                 errors = true;
1098
1099         if ((mark & EXTENT_NEW) &&
1100             test_and_clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags))
1101                 errors = true;
1102
1103         if (errors && !err)
1104                 err = -EIO;
1105         return err;
1106 }
1107
1108 /*
1109  * When btree blocks are allocated the corresponding extents are marked dirty.
1110  * This function ensures such extents are persisted on disk for transaction or
1111  * log commit.
1112  *
1113  * @trans: transaction whose dirty pages we'd like to write
1114  */
1115 static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans)
1116 {
1117         int ret;
1118         int ret2;
1119         struct extent_io_tree *dirty_pages = &trans->transaction->dirty_pages;
1120         struct btrfs_fs_info *fs_info = trans->fs_info;
1121         struct blk_plug plug;
1122
1123         blk_start_plug(&plug);
1124         ret = btrfs_write_marked_extents(fs_info, dirty_pages, EXTENT_DIRTY);
1125         blk_finish_plug(&plug);
1126         ret2 = btrfs_wait_extents(fs_info, dirty_pages);
1127
1128         extent_io_tree_release(&trans->transaction->dirty_pages);
1129
1130         if (ret)
1131                 return ret;
1132         else if (ret2)
1133                 return ret2;
1134         else
1135                 return 0;
1136 }
1137
1138 /*
1139  * this is used to update the root pointer in the tree of tree roots.
1140  *
1141  * But, in the case of the extent allocation tree, updating the root
1142  * pointer may allocate blocks which may change the root of the extent
1143  * allocation tree.
1144  *
1145  * So, this loops and repeats and makes sure the cowonly root didn't
1146  * change while the root pointer was being updated in the metadata.
1147  */
1148 static int update_cowonly_root(struct btrfs_trans_handle *trans,
1149                                struct btrfs_root *root)
1150 {
1151         int ret;
1152         u64 old_root_bytenr;
1153         u64 old_root_used;
1154         struct btrfs_fs_info *fs_info = root->fs_info;
1155         struct btrfs_root *tree_root = fs_info->tree_root;
1156
1157         old_root_used = btrfs_root_used(&root->root_item);
1158
1159         while (1) {
1160                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1161                 if (old_root_bytenr == root->node->start &&
1162                     old_root_used == btrfs_root_used(&root->root_item))
1163                         break;
1164
1165                 btrfs_set_root_node(&root->root_item, root->node);
1166                 ret = btrfs_update_root(trans, tree_root,
1167                                         &root->root_key,
1168                                         &root->root_item);
1169                 if (ret)
1170                         return ret;
1171
1172                 old_root_used = btrfs_root_used(&root->root_item);
1173         }
1174
1175         return 0;
1176 }
1177
1178 /*
1179  * update all the cowonly tree roots on disk
1180  *
1181  * The error handling in this function may not be obvious. Any of the
1182  * failures will cause the file system to go offline. We still need
1183  * to clean up the delayed refs.
1184  */
1185 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans)
1186 {
1187         struct btrfs_fs_info *fs_info = trans->fs_info;
1188         struct list_head *dirty_bgs = &trans->transaction->dirty_bgs;
1189         struct list_head *io_bgs = &trans->transaction->io_bgs;
1190         struct list_head *next;
1191         struct extent_buffer *eb;
1192         int ret;
1193
1194         eb = btrfs_lock_root_node(fs_info->tree_root);
1195         ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
1196                               0, &eb);
1197         btrfs_tree_unlock(eb);
1198         free_extent_buffer(eb);
1199
1200         if (ret)
1201                 return ret;
1202
1203         ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1204         if (ret)
1205                 return ret;
1206
1207         ret = btrfs_run_dev_stats(trans);
1208         if (ret)
1209                 return ret;
1210         ret = btrfs_run_dev_replace(trans);
1211         if (ret)
1212                 return ret;
1213         ret = btrfs_run_qgroups(trans);
1214         if (ret)
1215                 return ret;
1216
1217         ret = btrfs_setup_space_cache(trans);
1218         if (ret)
1219                 return ret;
1220
1221         /* run_qgroups might have added some more refs */
1222         ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1223         if (ret)
1224                 return ret;
1225 again:
1226         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1227                 struct btrfs_root *root;
1228                 next = fs_info->dirty_cowonly_roots.next;
1229                 list_del_init(next);
1230                 root = list_entry(next, struct btrfs_root, dirty_list);
1231                 clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1232
1233                 if (root != fs_info->extent_root)
1234                         list_add_tail(&root->dirty_list,
1235                                       &trans->transaction->switch_commits);
1236                 ret = update_cowonly_root(trans, root);
1237                 if (ret)
1238                         return ret;
1239                 ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1240                 if (ret)
1241                         return ret;
1242         }
1243
1244         while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1245                 ret = btrfs_write_dirty_block_groups(trans);
1246                 if (ret)
1247                         return ret;
1248                 ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1249                 if (ret)
1250                         return ret;
1251         }
1252
1253         if (!list_empty(&fs_info->dirty_cowonly_roots))
1254                 goto again;
1255
1256         list_add_tail(&fs_info->extent_root->dirty_list,
1257                       &trans->transaction->switch_commits);
1258
1259         /* Update dev-replace pointer once everything is committed */
1260         fs_info->dev_replace.committed_cursor_left =
1261                 fs_info->dev_replace.cursor_left_last_write_of_item;
1262
1263         return 0;
1264 }
1265
1266 /*
1267  * dead roots are old snapshots that need to be deleted.  This allocates
1268  * a dirty root struct and adds it into the list of dead roots that need to
1269  * be deleted
1270  */
1271 void btrfs_add_dead_root(struct btrfs_root *root)
1272 {
1273         struct btrfs_fs_info *fs_info = root->fs_info;
1274
1275         spin_lock(&fs_info->trans_lock);
1276         if (list_empty(&root->root_list)) {
1277                 btrfs_grab_root(root);
1278                 list_add_tail(&root->root_list, &fs_info->dead_roots);
1279         }
1280         spin_unlock(&fs_info->trans_lock);
1281 }
1282
1283 /*
1284  * update all the cowonly tree roots on disk
1285  */
1286 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans)
1287 {
1288         struct btrfs_fs_info *fs_info = trans->fs_info;
1289         struct btrfs_root *gang[8];
1290         int i;
1291         int ret;
1292         int err = 0;
1293
1294         spin_lock(&fs_info->fs_roots_radix_lock);
1295         while (1) {
1296                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1297                                                  (void **)gang, 0,
1298                                                  ARRAY_SIZE(gang),
1299                                                  BTRFS_ROOT_TRANS_TAG);
1300                 if (ret == 0)
1301                         break;
1302                 for (i = 0; i < ret; i++) {
1303                         struct btrfs_root *root = gang[i];
1304                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
1305                                         (unsigned long)root->root_key.objectid,
1306                                         BTRFS_ROOT_TRANS_TAG);
1307                         spin_unlock(&fs_info->fs_roots_radix_lock);
1308
1309                         btrfs_free_log(trans, root);
1310                         btrfs_update_reloc_root(trans, root);
1311
1312                         btrfs_save_ino_cache(root, trans);
1313
1314                         /* see comments in should_cow_block() */
1315                         clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1316                         smp_mb__after_atomic();
1317
1318                         if (root->commit_root != root->node) {
1319                                 list_add_tail(&root->dirty_list,
1320                                         &trans->transaction->switch_commits);
1321                                 btrfs_set_root_node(&root->root_item,
1322                                                     root->node);
1323                         }
1324
1325                         err = btrfs_update_root(trans, fs_info->tree_root,
1326                                                 &root->root_key,
1327                                                 &root->root_item);
1328                         spin_lock(&fs_info->fs_roots_radix_lock);
1329                         if (err)
1330                                 break;
1331                         btrfs_qgroup_free_meta_all_pertrans(root);
1332                 }
1333         }
1334         spin_unlock(&fs_info->fs_roots_radix_lock);
1335         return err;
1336 }
1337
1338 /*
1339  * defrag a given btree.
1340  * Every leaf in the btree is read and defragged.
1341  */
1342 int btrfs_defrag_root(struct btrfs_root *root)
1343 {
1344         struct btrfs_fs_info *info = root->fs_info;
1345         struct btrfs_trans_handle *trans;
1346         int ret;
1347
1348         if (test_and_set_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state))
1349                 return 0;
1350
1351         while (1) {
1352                 trans = btrfs_start_transaction(root, 0);
1353                 if (IS_ERR(trans))
1354                         return PTR_ERR(trans);
1355
1356                 ret = btrfs_defrag_leaves(trans, root);
1357
1358                 btrfs_end_transaction(trans);
1359                 btrfs_btree_balance_dirty(info);
1360                 cond_resched();
1361
1362                 if (btrfs_fs_closing(info) || ret != -EAGAIN)
1363                         break;
1364
1365                 if (btrfs_defrag_cancelled(info)) {
1366                         btrfs_debug(info, "defrag_root cancelled");
1367                         ret = -EAGAIN;
1368                         break;
1369                 }
1370         }
1371         clear_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state);
1372         return ret;
1373 }
1374
1375 /*
1376  * Do all special snapshot related qgroup dirty hack.
1377  *
1378  * Will do all needed qgroup inherit and dirty hack like switch commit
1379  * roots inside one transaction and write all btree into disk, to make
1380  * qgroup works.
1381  */
1382 static int qgroup_account_snapshot(struct btrfs_trans_handle *trans,
1383                                    struct btrfs_root *src,
1384                                    struct btrfs_root *parent,
1385                                    struct btrfs_qgroup_inherit *inherit,
1386                                    u64 dst_objectid)
1387 {
1388         struct btrfs_fs_info *fs_info = src->fs_info;
1389         int ret;
1390
1391         /*
1392          * Save some performance in the case that qgroups are not
1393          * enabled. If this check races with the ioctl, rescan will
1394          * kick in anyway.
1395          */
1396         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
1397                 return 0;
1398
1399         /*
1400          * Ensure dirty @src will be committed.  Or, after coming
1401          * commit_fs_roots() and switch_commit_roots(), any dirty but not
1402          * recorded root will never be updated again, causing an outdated root
1403          * item.
1404          */
1405         record_root_in_trans(trans, src, 1);
1406
1407         /*
1408          * We are going to commit transaction, see btrfs_commit_transaction()
1409          * comment for reason locking tree_log_mutex
1410          */
1411         mutex_lock(&fs_info->tree_log_mutex);
1412
1413         ret = commit_fs_roots(trans);
1414         if (ret)
1415                 goto out;
1416         ret = btrfs_qgroup_account_extents(trans);
1417         if (ret < 0)
1418                 goto out;
1419
1420         /* Now qgroup are all updated, we can inherit it to new qgroups */
1421         ret = btrfs_qgroup_inherit(trans, src->root_key.objectid, dst_objectid,
1422                                    inherit);
1423         if (ret < 0)
1424                 goto out;
1425
1426         /*
1427          * Now we do a simplified commit transaction, which will:
1428          * 1) commit all subvolume and extent tree
1429          *    To ensure all subvolume and extent tree have a valid
1430          *    commit_root to accounting later insert_dir_item()
1431          * 2) write all btree blocks onto disk
1432          *    This is to make sure later btree modification will be cowed
1433          *    Or commit_root can be populated and cause wrong qgroup numbers
1434          * In this simplified commit, we don't really care about other trees
1435          * like chunk and root tree, as they won't affect qgroup.
1436          * And we don't write super to avoid half committed status.
1437          */
1438         ret = commit_cowonly_roots(trans);
1439         if (ret)
1440                 goto out;
1441         switch_commit_roots(trans);
1442         ret = btrfs_write_and_wait_transaction(trans);
1443         if (ret)
1444                 btrfs_handle_fs_error(fs_info, ret,
1445                         "Error while writing out transaction for qgroup");
1446
1447 out:
1448         mutex_unlock(&fs_info->tree_log_mutex);
1449
1450         /*
1451          * Force parent root to be updated, as we recorded it before so its
1452          * last_trans == cur_transid.
1453          * Or it won't be committed again onto disk after later
1454          * insert_dir_item()
1455          */
1456         if (!ret)
1457                 record_root_in_trans(trans, parent, 1);
1458         return ret;
1459 }
1460
1461 /*
1462  * new snapshots need to be created at a very specific time in the
1463  * transaction commit.  This does the actual creation.
1464  *
1465  * Note:
1466  * If the error which may affect the commitment of the current transaction
1467  * happens, we should return the error number. If the error which just affect
1468  * the creation of the pending snapshots, just return 0.
1469  */
1470 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1471                                    struct btrfs_pending_snapshot *pending)
1472 {
1473
1474         struct btrfs_fs_info *fs_info = trans->fs_info;
1475         struct btrfs_key key;
1476         struct btrfs_root_item *new_root_item;
1477         struct btrfs_root *tree_root = fs_info->tree_root;
1478         struct btrfs_root *root = pending->root;
1479         struct btrfs_root *parent_root;
1480         struct btrfs_block_rsv *rsv;
1481         struct inode *parent_inode;
1482         struct btrfs_path *path;
1483         struct btrfs_dir_item *dir_item;
1484         struct dentry *dentry;
1485         struct extent_buffer *tmp;
1486         struct extent_buffer *old;
1487         struct timespec64 cur_time;
1488         int ret = 0;
1489         u64 to_reserve = 0;
1490         u64 index = 0;
1491         u64 objectid;
1492         u64 root_flags;
1493
1494         ASSERT(pending->path);
1495         path = pending->path;
1496
1497         ASSERT(pending->root_item);
1498         new_root_item = pending->root_item;
1499
1500         pending->error = btrfs_find_free_objectid(tree_root, &objectid);
1501         if (pending->error)
1502                 goto no_free_objectid;
1503
1504         /*
1505          * Make qgroup to skip current new snapshot's qgroupid, as it is
1506          * accounted by later btrfs_qgroup_inherit().
1507          */
1508         btrfs_set_skip_qgroup(trans, objectid);
1509
1510         btrfs_reloc_pre_snapshot(pending, &to_reserve);
1511
1512         if (to_reserve > 0) {
1513                 pending->error = btrfs_block_rsv_add(root,
1514                                                      &pending->block_rsv,
1515                                                      to_reserve,
1516                                                      BTRFS_RESERVE_NO_FLUSH);
1517                 if (pending->error)
1518                         goto clear_skip_qgroup;
1519         }
1520
1521         key.objectid = objectid;
1522         key.offset = (u64)-1;
1523         key.type = BTRFS_ROOT_ITEM_KEY;
1524
1525         rsv = trans->block_rsv;
1526         trans->block_rsv = &pending->block_rsv;
1527         trans->bytes_reserved = trans->block_rsv->reserved;
1528         trace_btrfs_space_reservation(fs_info, "transaction",
1529                                       trans->transid,
1530                                       trans->bytes_reserved, 1);
1531         dentry = pending->dentry;
1532         parent_inode = pending->dir;
1533         parent_root = BTRFS_I(parent_inode)->root;
1534         record_root_in_trans(trans, parent_root, 0);
1535
1536         cur_time = current_time(parent_inode);
1537
1538         /*
1539          * insert the directory item
1540          */
1541         ret = btrfs_set_inode_index(BTRFS_I(parent_inode), &index);
1542         BUG_ON(ret); /* -ENOMEM */
1543
1544         /* check if there is a file/dir which has the same name. */
1545         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1546                                          btrfs_ino(BTRFS_I(parent_inode)),
1547                                          dentry->d_name.name,
1548                                          dentry->d_name.len, 0);
1549         if (dir_item != NULL && !IS_ERR(dir_item)) {
1550                 pending->error = -EEXIST;
1551                 goto dir_item_existed;
1552         } else if (IS_ERR(dir_item)) {
1553                 ret = PTR_ERR(dir_item);
1554                 btrfs_abort_transaction(trans, ret);
1555                 goto fail;
1556         }
1557         btrfs_release_path(path);
1558
1559         /*
1560          * pull in the delayed directory update
1561          * and the delayed inode item
1562          * otherwise we corrupt the FS during
1563          * snapshot
1564          */
1565         ret = btrfs_run_delayed_items(trans);
1566         if (ret) {      /* Transaction aborted */
1567                 btrfs_abort_transaction(trans, ret);
1568                 goto fail;
1569         }
1570
1571         record_root_in_trans(trans, root, 0);
1572         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1573         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1574         btrfs_check_and_init_root_item(new_root_item);
1575
1576         root_flags = btrfs_root_flags(new_root_item);
1577         if (pending->readonly)
1578                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1579         else
1580                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1581         btrfs_set_root_flags(new_root_item, root_flags);
1582
1583         btrfs_set_root_generation_v2(new_root_item,
1584                         trans->transid);
1585         generate_random_guid(new_root_item->uuid);
1586         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1587                         BTRFS_UUID_SIZE);
1588         if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1589                 memset(new_root_item->received_uuid, 0,
1590                        sizeof(new_root_item->received_uuid));
1591                 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1592                 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1593                 btrfs_set_root_stransid(new_root_item, 0);
1594                 btrfs_set_root_rtransid(new_root_item, 0);
1595         }
1596         btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1597         btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1598         btrfs_set_root_otransid(new_root_item, trans->transid);
1599
1600         old = btrfs_lock_root_node(root);
1601         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1602         if (ret) {
1603                 btrfs_tree_unlock(old);
1604                 free_extent_buffer(old);
1605                 btrfs_abort_transaction(trans, ret);
1606                 goto fail;
1607         }
1608
1609         btrfs_set_lock_blocking_write(old);
1610
1611         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1612         /* clean up in any case */
1613         btrfs_tree_unlock(old);
1614         free_extent_buffer(old);
1615         if (ret) {
1616                 btrfs_abort_transaction(trans, ret);
1617                 goto fail;
1618         }
1619         /* see comments in should_cow_block() */
1620         set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1621         smp_wmb();
1622
1623         btrfs_set_root_node(new_root_item, tmp);
1624         /* record when the snapshot was created in key.offset */
1625         key.offset = trans->transid;
1626         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1627         btrfs_tree_unlock(tmp);
1628         free_extent_buffer(tmp);
1629         if (ret) {
1630                 btrfs_abort_transaction(trans, ret);
1631                 goto fail;
1632         }
1633
1634         /*
1635          * insert root back/forward references
1636          */
1637         ret = btrfs_add_root_ref(trans, objectid,
1638                                  parent_root->root_key.objectid,
1639                                  btrfs_ino(BTRFS_I(parent_inode)), index,
1640                                  dentry->d_name.name, dentry->d_name.len);
1641         if (ret) {
1642                 btrfs_abort_transaction(trans, ret);
1643                 goto fail;
1644         }
1645
1646         key.offset = (u64)-1;
1647         pending->snap = btrfs_get_fs_root(fs_info, &key, true);
1648         if (IS_ERR(pending->snap)) {
1649                 ret = PTR_ERR(pending->snap);
1650                 btrfs_abort_transaction(trans, ret);
1651                 goto fail;
1652         }
1653
1654         ret = btrfs_reloc_post_snapshot(trans, pending);
1655         if (ret) {
1656                 btrfs_abort_transaction(trans, ret);
1657                 goto fail;
1658         }
1659
1660         ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1661         if (ret) {
1662                 btrfs_abort_transaction(trans, ret);
1663                 goto fail;
1664         }
1665
1666         /*
1667          * Do special qgroup accounting for snapshot, as we do some qgroup
1668          * snapshot hack to do fast snapshot.
1669          * To co-operate with that hack, we do hack again.
1670          * Or snapshot will be greatly slowed down by a subtree qgroup rescan
1671          */
1672         ret = qgroup_account_snapshot(trans, root, parent_root,
1673                                       pending->inherit, objectid);
1674         if (ret < 0)
1675                 goto fail;
1676
1677         ret = btrfs_insert_dir_item(trans, dentry->d_name.name,
1678                                     dentry->d_name.len, BTRFS_I(parent_inode),
1679                                     &key, BTRFS_FT_DIR, index);
1680         /* We have check then name at the beginning, so it is impossible. */
1681         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1682         if (ret) {
1683                 btrfs_abort_transaction(trans, ret);
1684                 goto fail;
1685         }
1686
1687         btrfs_i_size_write(BTRFS_I(parent_inode), parent_inode->i_size +
1688                                          dentry->d_name.len * 2);
1689         parent_inode->i_mtime = parent_inode->i_ctime =
1690                 current_time(parent_inode);
1691         ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1692         if (ret) {
1693                 btrfs_abort_transaction(trans, ret);
1694                 goto fail;
1695         }
1696         ret = btrfs_uuid_tree_add(trans, new_root_item->uuid,
1697                                   BTRFS_UUID_KEY_SUBVOL,
1698                                   objectid);
1699         if (ret) {
1700                 btrfs_abort_transaction(trans, ret);
1701                 goto fail;
1702         }
1703         if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1704                 ret = btrfs_uuid_tree_add(trans, new_root_item->received_uuid,
1705                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1706                                           objectid);
1707                 if (ret && ret != -EEXIST) {
1708                         btrfs_abort_transaction(trans, ret);
1709                         goto fail;
1710                 }
1711         }
1712
1713         ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
1714         if (ret) {
1715                 btrfs_abort_transaction(trans, ret);
1716                 goto fail;
1717         }
1718
1719 fail:
1720         pending->error = ret;
1721 dir_item_existed:
1722         trans->block_rsv = rsv;
1723         trans->bytes_reserved = 0;
1724 clear_skip_qgroup:
1725         btrfs_clear_skip_qgroup(trans);
1726 no_free_objectid:
1727         kfree(new_root_item);
1728         pending->root_item = NULL;
1729         btrfs_free_path(path);
1730         pending->path = NULL;
1731
1732         return ret;
1733 }
1734
1735 /*
1736  * create all the snapshots we've scheduled for creation
1737  */
1738 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans)
1739 {
1740         struct btrfs_pending_snapshot *pending, *next;
1741         struct list_head *head = &trans->transaction->pending_snapshots;
1742         int ret = 0;
1743
1744         list_for_each_entry_safe(pending, next, head, list) {
1745                 list_del(&pending->list);
1746                 ret = create_pending_snapshot(trans, pending);
1747                 if (ret)
1748                         break;
1749         }
1750         return ret;
1751 }
1752
1753 static void update_super_roots(struct btrfs_fs_info *fs_info)
1754 {
1755         struct btrfs_root_item *root_item;
1756         struct btrfs_super_block *super;
1757
1758         super = fs_info->super_copy;
1759
1760         root_item = &fs_info->chunk_root->root_item;
1761         super->chunk_root = root_item->bytenr;
1762         super->chunk_root_generation = root_item->generation;
1763         super->chunk_root_level = root_item->level;
1764
1765         root_item = &fs_info->tree_root->root_item;
1766         super->root = root_item->bytenr;
1767         super->generation = root_item->generation;
1768         super->root_level = root_item->level;
1769         if (btrfs_test_opt(fs_info, SPACE_CACHE))
1770                 super->cache_generation = root_item->generation;
1771         if (test_bit(BTRFS_FS_UPDATE_UUID_TREE_GEN, &fs_info->flags))
1772                 super->uuid_tree_generation = root_item->generation;
1773 }
1774
1775 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1776 {
1777         struct btrfs_transaction *trans;
1778         int ret = 0;
1779
1780         spin_lock(&info->trans_lock);
1781         trans = info->running_transaction;
1782         if (trans)
1783                 ret = (trans->state >= TRANS_STATE_COMMIT_START);
1784         spin_unlock(&info->trans_lock);
1785         return ret;
1786 }
1787
1788 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1789 {
1790         struct btrfs_transaction *trans;
1791         int ret = 0;
1792
1793         spin_lock(&info->trans_lock);
1794         trans = info->running_transaction;
1795         if (trans)
1796                 ret = is_transaction_blocked(trans);
1797         spin_unlock(&info->trans_lock);
1798         return ret;
1799 }
1800
1801 /*
1802  * wait for the current transaction commit to start and block subsequent
1803  * transaction joins
1804  */
1805 static void wait_current_trans_commit_start(struct btrfs_fs_info *fs_info,
1806                                             struct btrfs_transaction *trans)
1807 {
1808         wait_event(fs_info->transaction_blocked_wait,
1809                    trans->state >= TRANS_STATE_COMMIT_START ||
1810                    TRANS_ABORTED(trans));
1811 }
1812
1813 /*
1814  * wait for the current transaction to start and then become unblocked.
1815  * caller holds ref.
1816  */
1817 static void wait_current_trans_commit_start_and_unblock(
1818                                         struct btrfs_fs_info *fs_info,
1819                                         struct btrfs_transaction *trans)
1820 {
1821         wait_event(fs_info->transaction_wait,
1822                    trans->state >= TRANS_STATE_UNBLOCKED ||
1823                    TRANS_ABORTED(trans));
1824 }
1825
1826 /*
1827  * commit transactions asynchronously. once btrfs_commit_transaction_async
1828  * returns, any subsequent transaction will not be allowed to join.
1829  */
1830 struct btrfs_async_commit {
1831         struct btrfs_trans_handle *newtrans;
1832         struct work_struct work;
1833 };
1834
1835 static void do_async_commit(struct work_struct *work)
1836 {
1837         struct btrfs_async_commit *ac =
1838                 container_of(work, struct btrfs_async_commit, work);
1839
1840         /*
1841          * We've got freeze protection passed with the transaction.
1842          * Tell lockdep about it.
1843          */
1844         if (ac->newtrans->type & __TRANS_FREEZABLE)
1845                 __sb_writers_acquired(ac->newtrans->fs_info->sb, SB_FREEZE_FS);
1846
1847         current->journal_info = ac->newtrans;
1848
1849         btrfs_commit_transaction(ac->newtrans);
1850         kfree(ac);
1851 }
1852
1853 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1854                                    int wait_for_unblock)
1855 {
1856         struct btrfs_fs_info *fs_info = trans->fs_info;
1857         struct btrfs_async_commit *ac;
1858         struct btrfs_transaction *cur_trans;
1859
1860         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1861         if (!ac)
1862                 return -ENOMEM;
1863
1864         INIT_WORK(&ac->work, do_async_commit);
1865         ac->newtrans = btrfs_join_transaction(trans->root);
1866         if (IS_ERR(ac->newtrans)) {
1867                 int err = PTR_ERR(ac->newtrans);
1868                 kfree(ac);
1869                 return err;
1870         }
1871
1872         /* take transaction reference */
1873         cur_trans = trans->transaction;
1874         refcount_inc(&cur_trans->use_count);
1875
1876         btrfs_end_transaction(trans);
1877
1878         /*
1879          * Tell lockdep we've released the freeze rwsem, since the
1880          * async commit thread will be the one to unlock it.
1881          */
1882         if (ac->newtrans->type & __TRANS_FREEZABLE)
1883                 __sb_writers_release(fs_info->sb, SB_FREEZE_FS);
1884
1885         schedule_work(&ac->work);
1886
1887         /* wait for transaction to start and unblock */
1888         if (wait_for_unblock)
1889                 wait_current_trans_commit_start_and_unblock(fs_info, cur_trans);
1890         else
1891                 wait_current_trans_commit_start(fs_info, cur_trans);
1892
1893         if (current->journal_info == trans)
1894                 current->journal_info = NULL;
1895
1896         btrfs_put_transaction(cur_trans);
1897         return 0;
1898 }
1899
1900
1901 static void cleanup_transaction(struct btrfs_trans_handle *trans, int err)
1902 {
1903         struct btrfs_fs_info *fs_info = trans->fs_info;
1904         struct btrfs_transaction *cur_trans = trans->transaction;
1905
1906         WARN_ON(refcount_read(&trans->use_count) > 1);
1907
1908         btrfs_abort_transaction(trans, err);
1909
1910         spin_lock(&fs_info->trans_lock);
1911
1912         /*
1913          * If the transaction is removed from the list, it means this
1914          * transaction has been committed successfully, so it is impossible
1915          * to call the cleanup function.
1916          */
1917         BUG_ON(list_empty(&cur_trans->list));
1918
1919         list_del_init(&cur_trans->list);
1920         if (cur_trans == fs_info->running_transaction) {
1921                 cur_trans->state = TRANS_STATE_COMMIT_DOING;
1922                 spin_unlock(&fs_info->trans_lock);
1923                 wait_event(cur_trans->writer_wait,
1924                            atomic_read(&cur_trans->num_writers) == 1);
1925
1926                 spin_lock(&fs_info->trans_lock);
1927         }
1928         spin_unlock(&fs_info->trans_lock);
1929
1930         btrfs_cleanup_one_transaction(trans->transaction, fs_info);
1931
1932         spin_lock(&fs_info->trans_lock);
1933         if (cur_trans == fs_info->running_transaction)
1934                 fs_info->running_transaction = NULL;
1935         spin_unlock(&fs_info->trans_lock);
1936
1937         if (trans->type & __TRANS_FREEZABLE)
1938                 sb_end_intwrite(fs_info->sb);
1939         btrfs_put_transaction(cur_trans);
1940         btrfs_put_transaction(cur_trans);
1941
1942         trace_btrfs_transaction_commit(trans->root);
1943
1944         if (current->journal_info == trans)
1945                 current->journal_info = NULL;
1946         btrfs_scrub_cancel(fs_info);
1947
1948         kmem_cache_free(btrfs_trans_handle_cachep, trans);
1949 }
1950
1951 /*
1952  * Release reserved delayed ref space of all pending block groups of the
1953  * transaction and remove them from the list
1954  */
1955 static void btrfs_cleanup_pending_block_groups(struct btrfs_trans_handle *trans)
1956 {
1957        struct btrfs_fs_info *fs_info = trans->fs_info;
1958        struct btrfs_block_group *block_group, *tmp;
1959
1960        list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
1961                btrfs_delayed_refs_rsv_release(fs_info, 1);
1962                list_del_init(&block_group->bg_list);
1963        }
1964 }
1965
1966 static inline int btrfs_start_delalloc_flush(struct btrfs_trans_handle *trans)
1967 {
1968         struct btrfs_fs_info *fs_info = trans->fs_info;
1969
1970         /*
1971          * We use writeback_inodes_sb here because if we used
1972          * btrfs_start_delalloc_roots we would deadlock with fs freeze.
1973          * Currently are holding the fs freeze lock, if we do an async flush
1974          * we'll do btrfs_join_transaction() and deadlock because we need to
1975          * wait for the fs freeze lock.  Using the direct flushing we benefit
1976          * from already being in a transaction and our join_transaction doesn't
1977          * have to re-take the fs freeze lock.
1978          */
1979         if (btrfs_test_opt(fs_info, FLUSHONCOMMIT)) {
1980                 writeback_inodes_sb(fs_info->sb, WB_REASON_SYNC);
1981         } else {
1982                 struct btrfs_pending_snapshot *pending;
1983                 struct list_head *head = &trans->transaction->pending_snapshots;
1984
1985                 /*
1986                  * Flush dellaloc for any root that is going to be snapshotted.
1987                  * This is done to avoid a corrupted version of files, in the
1988                  * snapshots, that had both buffered and direct IO writes (even
1989                  * if they were done sequentially) due to an unordered update of
1990                  * the inode's size on disk.
1991                  */
1992                 list_for_each_entry(pending, head, list) {
1993                         int ret;
1994
1995                         ret = btrfs_start_delalloc_snapshot(pending->root);
1996                         if (ret)
1997                                 return ret;
1998                 }
1999         }
2000         return 0;
2001 }
2002
2003 static inline void btrfs_wait_delalloc_flush(struct btrfs_trans_handle *trans)
2004 {
2005         struct btrfs_fs_info *fs_info = trans->fs_info;
2006
2007         if (btrfs_test_opt(fs_info, FLUSHONCOMMIT)) {
2008                 btrfs_wait_ordered_roots(fs_info, U64_MAX, 0, (u64)-1);
2009         } else {
2010                 struct btrfs_pending_snapshot *pending;
2011                 struct list_head *head = &trans->transaction->pending_snapshots;
2012
2013                 /*
2014                  * Wait for any dellaloc that we started previously for the roots
2015                  * that are going to be snapshotted. This is to avoid a corrupted
2016                  * version of files in the snapshots that had both buffered and
2017                  * direct IO writes (even if they were done sequentially).
2018                  */
2019                 list_for_each_entry(pending, head, list)
2020                         btrfs_wait_ordered_extents(pending->root,
2021                                                    U64_MAX, 0, U64_MAX);
2022         }
2023 }
2024
2025 int btrfs_commit_transaction(struct btrfs_trans_handle *trans)
2026 {
2027         struct btrfs_fs_info *fs_info = trans->fs_info;
2028         struct btrfs_transaction *cur_trans = trans->transaction;
2029         struct btrfs_transaction *prev_trans = NULL;
2030         int ret;
2031
2032         ASSERT(refcount_read(&trans->use_count) == 1);
2033
2034         /*
2035          * Some places just start a transaction to commit it.  We need to make
2036          * sure that if this commit fails that the abort code actually marks the
2037          * transaction as failed, so set trans->dirty to make the abort code do
2038          * the right thing.
2039          */
2040         trans->dirty = true;
2041
2042         /* Stop the commit early if ->aborted is set */
2043         if (TRANS_ABORTED(cur_trans)) {
2044                 ret = cur_trans->aborted;
2045                 btrfs_end_transaction(trans);
2046                 return ret;
2047         }
2048
2049         btrfs_trans_release_metadata(trans);
2050         trans->block_rsv = NULL;
2051
2052         /* make a pass through all the delayed refs we have so far
2053          * any runnings procs may add more while we are here
2054          */
2055         ret = btrfs_run_delayed_refs(trans, 0);
2056         if (ret) {
2057                 btrfs_end_transaction(trans);
2058                 return ret;
2059         }
2060
2061         cur_trans = trans->transaction;
2062
2063         /*
2064          * set the flushing flag so procs in this transaction have to
2065          * start sending their work down.
2066          */
2067         cur_trans->delayed_refs.flushing = 1;
2068         smp_wmb();
2069
2070         btrfs_create_pending_block_groups(trans);
2071
2072         ret = btrfs_run_delayed_refs(trans, 0);
2073         if (ret) {
2074                 btrfs_end_transaction(trans);
2075                 return ret;
2076         }
2077
2078         if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) {
2079                 int run_it = 0;
2080
2081                 /* this mutex is also taken before trying to set
2082                  * block groups readonly.  We need to make sure
2083                  * that nobody has set a block group readonly
2084                  * after a extents from that block group have been
2085                  * allocated for cache files.  btrfs_set_block_group_ro
2086                  * will wait for the transaction to commit if it
2087                  * finds BTRFS_TRANS_DIRTY_BG_RUN set.
2088                  *
2089                  * The BTRFS_TRANS_DIRTY_BG_RUN flag is also used to make sure
2090                  * only one process starts all the block group IO.  It wouldn't
2091                  * hurt to have more than one go through, but there's no
2092                  * real advantage to it either.
2093                  */
2094                 mutex_lock(&fs_info->ro_block_group_mutex);
2095                 if (!test_and_set_bit(BTRFS_TRANS_DIRTY_BG_RUN,
2096                                       &cur_trans->flags))
2097                         run_it = 1;
2098                 mutex_unlock(&fs_info->ro_block_group_mutex);
2099
2100                 if (run_it) {
2101                         ret = btrfs_start_dirty_block_groups(trans);
2102                         if (ret) {
2103                                 btrfs_end_transaction(trans);
2104                                 return ret;
2105                         }
2106                 }
2107         }
2108
2109         spin_lock(&fs_info->trans_lock);
2110         if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
2111                 spin_unlock(&fs_info->trans_lock);
2112                 refcount_inc(&cur_trans->use_count);
2113                 ret = btrfs_end_transaction(trans);
2114
2115                 wait_for_commit(cur_trans);
2116
2117                 if (TRANS_ABORTED(cur_trans))
2118                         ret = cur_trans->aborted;
2119
2120                 btrfs_put_transaction(cur_trans);
2121
2122                 return ret;
2123         }
2124
2125         cur_trans->state = TRANS_STATE_COMMIT_START;
2126         wake_up(&fs_info->transaction_blocked_wait);
2127
2128         if (cur_trans->list.prev != &fs_info->trans_list) {
2129                 prev_trans = list_entry(cur_trans->list.prev,
2130                                         struct btrfs_transaction, list);
2131                 if (prev_trans->state != TRANS_STATE_COMPLETED) {
2132                         refcount_inc(&prev_trans->use_count);
2133                         spin_unlock(&fs_info->trans_lock);
2134
2135                         wait_for_commit(prev_trans);
2136                         ret = READ_ONCE(prev_trans->aborted);
2137
2138                         btrfs_put_transaction(prev_trans);
2139                         if (ret)
2140                                 goto cleanup_transaction;
2141                 } else {
2142                         spin_unlock(&fs_info->trans_lock);
2143                 }
2144         } else {
2145                 spin_unlock(&fs_info->trans_lock);
2146                 /*
2147                  * The previous transaction was aborted and was already removed
2148                  * from the list of transactions at fs_info->trans_list. So we
2149                  * abort to prevent writing a new superblock that reflects a
2150                  * corrupt state (pointing to trees with unwritten nodes/leafs).
2151                  */
2152                 if (test_bit(BTRFS_FS_STATE_TRANS_ABORTED, &fs_info->fs_state)) {
2153                         ret = -EROFS;
2154                         goto cleanup_transaction;
2155                 }
2156         }
2157
2158         extwriter_counter_dec(cur_trans, trans->type);
2159
2160         ret = btrfs_start_delalloc_flush(trans);
2161         if (ret)
2162                 goto cleanup_transaction;
2163
2164         ret = btrfs_run_delayed_items(trans);
2165         if (ret)
2166                 goto cleanup_transaction;
2167
2168         wait_event(cur_trans->writer_wait,
2169                    extwriter_counter_read(cur_trans) == 0);
2170
2171         /* some pending stuffs might be added after the previous flush. */
2172         ret = btrfs_run_delayed_items(trans);
2173         if (ret)
2174                 goto cleanup_transaction;
2175
2176         btrfs_wait_delalloc_flush(trans);
2177
2178         btrfs_scrub_pause(fs_info);
2179         /*
2180          * Ok now we need to make sure to block out any other joins while we
2181          * commit the transaction.  We could have started a join before setting
2182          * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
2183          */
2184         spin_lock(&fs_info->trans_lock);
2185         cur_trans->state = TRANS_STATE_COMMIT_DOING;
2186         spin_unlock(&fs_info->trans_lock);
2187         wait_event(cur_trans->writer_wait,
2188                    atomic_read(&cur_trans->num_writers) == 1);
2189
2190         if (TRANS_ABORTED(cur_trans)) {
2191                 ret = cur_trans->aborted;
2192                 goto scrub_continue;
2193         }
2194         /*
2195          * the reloc mutex makes sure that we stop
2196          * the balancing code from coming in and moving
2197          * extents around in the middle of the commit
2198          */
2199         mutex_lock(&fs_info->reloc_mutex);
2200
2201         /*
2202          * We needn't worry about the delayed items because we will
2203          * deal with them in create_pending_snapshot(), which is the
2204          * core function of the snapshot creation.
2205          */
2206         ret = create_pending_snapshots(trans);
2207         if (ret)
2208                 goto unlock_reloc;
2209
2210         /*
2211          * We insert the dir indexes of the snapshots and update the inode
2212          * of the snapshots' parents after the snapshot creation, so there
2213          * are some delayed items which are not dealt with. Now deal with
2214          * them.
2215          *
2216          * We needn't worry that this operation will corrupt the snapshots,
2217          * because all the tree which are snapshoted will be forced to COW
2218          * the nodes and leaves.
2219          */
2220         ret = btrfs_run_delayed_items(trans);
2221         if (ret)
2222                 goto unlock_reloc;
2223
2224         ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
2225         if (ret)
2226                 goto unlock_reloc;
2227
2228         /*
2229          * make sure none of the code above managed to slip in a
2230          * delayed item
2231          */
2232         btrfs_assert_delayed_root_empty(fs_info);
2233
2234         WARN_ON(cur_trans != trans->transaction);
2235
2236         /* btrfs_commit_tree_roots is responsible for getting the
2237          * various roots consistent with each other.  Every pointer
2238          * in the tree of tree roots has to point to the most up to date
2239          * root for every subvolume and other tree.  So, we have to keep
2240          * the tree logging code from jumping in and changing any
2241          * of the trees.
2242          *
2243          * At this point in the commit, there can't be any tree-log
2244          * writers, but a little lower down we drop the trans mutex
2245          * and let new people in.  By holding the tree_log_mutex
2246          * from now until after the super is written, we avoid races
2247          * with the tree-log code.
2248          */
2249         mutex_lock(&fs_info->tree_log_mutex);
2250
2251         ret = commit_fs_roots(trans);
2252         if (ret)
2253                 goto unlock_tree_log;
2254
2255         /*
2256          * Since the transaction is done, we can apply the pending changes
2257          * before the next transaction.
2258          */
2259         btrfs_apply_pending_changes(fs_info);
2260
2261         /* commit_fs_roots gets rid of all the tree log roots, it is now
2262          * safe to free the root of tree log roots
2263          */
2264         btrfs_free_log_root_tree(trans, fs_info);
2265
2266         /*
2267          * commit_fs_roots() can call btrfs_save_ino_cache(), which generates
2268          * new delayed refs. Must handle them or qgroup can be wrong.
2269          */
2270         ret = btrfs_run_delayed_refs(trans, (unsigned long)-1);
2271         if (ret)
2272                 goto unlock_tree_log;
2273
2274         /*
2275          * Since fs roots are all committed, we can get a quite accurate
2276          * new_roots. So let's do quota accounting.
2277          */
2278         ret = btrfs_qgroup_account_extents(trans);
2279         if (ret < 0)
2280                 goto unlock_tree_log;
2281
2282         ret = commit_cowonly_roots(trans);
2283         if (ret)
2284                 goto unlock_tree_log;
2285
2286         /*
2287          * The tasks which save the space cache and inode cache may also
2288          * update ->aborted, check it.
2289          */
2290         if (TRANS_ABORTED(cur_trans)) {
2291                 ret = cur_trans->aborted;
2292                 goto unlock_tree_log;
2293         }
2294
2295         btrfs_prepare_extent_commit(fs_info);
2296
2297         cur_trans = fs_info->running_transaction;
2298
2299         btrfs_set_root_node(&fs_info->tree_root->root_item,
2300                             fs_info->tree_root->node);
2301         list_add_tail(&fs_info->tree_root->dirty_list,
2302                       &cur_trans->switch_commits);
2303
2304         btrfs_set_root_node(&fs_info->chunk_root->root_item,
2305                             fs_info->chunk_root->node);
2306         list_add_tail(&fs_info->chunk_root->dirty_list,
2307                       &cur_trans->switch_commits);
2308
2309         switch_commit_roots(trans);
2310
2311         ASSERT(list_empty(&cur_trans->dirty_bgs));
2312         ASSERT(list_empty(&cur_trans->io_bgs));
2313         update_super_roots(fs_info);
2314
2315         btrfs_set_super_log_root(fs_info->super_copy, 0);
2316         btrfs_set_super_log_root_level(fs_info->super_copy, 0);
2317         memcpy(fs_info->super_for_commit, fs_info->super_copy,
2318                sizeof(*fs_info->super_copy));
2319
2320         btrfs_commit_device_sizes(cur_trans);
2321
2322         clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
2323         clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
2324
2325         btrfs_trans_release_chunk_metadata(trans);
2326
2327         spin_lock(&fs_info->trans_lock);
2328         cur_trans->state = TRANS_STATE_UNBLOCKED;
2329         fs_info->running_transaction = NULL;
2330         spin_unlock(&fs_info->trans_lock);
2331         mutex_unlock(&fs_info->reloc_mutex);
2332
2333         wake_up(&fs_info->transaction_wait);
2334
2335         ret = btrfs_write_and_wait_transaction(trans);
2336         if (ret) {
2337                 btrfs_handle_fs_error(fs_info, ret,
2338                                       "Error while writing out transaction");
2339                 /*
2340                  * reloc_mutex has been unlocked, tree_log_mutex is still held
2341                  * but we can't jump to unlock_tree_log causing double unlock
2342                  */
2343                 mutex_unlock(&fs_info->tree_log_mutex);
2344                 goto scrub_continue;
2345         }
2346
2347         ret = write_all_supers(fs_info, 0);
2348         /*
2349          * the super is written, we can safely allow the tree-loggers
2350          * to go about their business
2351          */
2352         mutex_unlock(&fs_info->tree_log_mutex);
2353         if (ret)
2354                 goto scrub_continue;
2355
2356         btrfs_finish_extent_commit(trans);
2357
2358         if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &cur_trans->flags))
2359                 btrfs_clear_space_info_full(fs_info);
2360
2361         fs_info->last_trans_committed = cur_trans->transid;
2362         /*
2363          * We needn't acquire the lock here because there is no other task
2364          * which can change it.
2365          */
2366         cur_trans->state = TRANS_STATE_COMPLETED;
2367         wake_up(&cur_trans->commit_wait);
2368         clear_bit(BTRFS_FS_NEED_ASYNC_COMMIT, &fs_info->flags);
2369
2370         spin_lock(&fs_info->trans_lock);
2371         list_del_init(&cur_trans->list);
2372         spin_unlock(&fs_info->trans_lock);
2373
2374         btrfs_put_transaction(cur_trans);
2375         btrfs_put_transaction(cur_trans);
2376
2377         if (trans->type & __TRANS_FREEZABLE)
2378                 sb_end_intwrite(fs_info->sb);
2379
2380         trace_btrfs_transaction_commit(trans->root);
2381
2382         btrfs_scrub_continue(fs_info);
2383
2384         if (current->journal_info == trans)
2385                 current->journal_info = NULL;
2386
2387         kmem_cache_free(btrfs_trans_handle_cachep, trans);
2388
2389         return ret;
2390
2391 unlock_tree_log:
2392         mutex_unlock(&fs_info->tree_log_mutex);
2393 unlock_reloc:
2394         mutex_unlock(&fs_info->reloc_mutex);
2395 scrub_continue:
2396         btrfs_scrub_continue(fs_info);
2397 cleanup_transaction:
2398         btrfs_trans_release_metadata(trans);
2399         btrfs_cleanup_pending_block_groups(trans);
2400         btrfs_trans_release_chunk_metadata(trans);
2401         trans->block_rsv = NULL;
2402         btrfs_warn(fs_info, "Skipping commit of aborted transaction.");
2403         if (current->journal_info == trans)
2404                 current->journal_info = NULL;
2405         cleanup_transaction(trans, ret);
2406
2407         return ret;
2408 }
2409
2410 /*
2411  * return < 0 if error
2412  * 0 if there are no more dead_roots at the time of call
2413  * 1 there are more to be processed, call me again
2414  *
2415  * The return value indicates there are certainly more snapshots to delete, but
2416  * if there comes a new one during processing, it may return 0. We don't mind,
2417  * because btrfs_commit_super will poke cleaner thread and it will process it a
2418  * few seconds later.
2419  */
2420 int btrfs_clean_one_deleted_snapshot(struct btrfs_root *root)
2421 {
2422         int ret;
2423         struct btrfs_fs_info *fs_info = root->fs_info;
2424
2425         spin_lock(&fs_info->trans_lock);
2426         if (list_empty(&fs_info->dead_roots)) {
2427                 spin_unlock(&fs_info->trans_lock);
2428                 return 0;
2429         }
2430         root = list_first_entry(&fs_info->dead_roots,
2431                         struct btrfs_root, root_list);
2432         list_del_init(&root->root_list);
2433         spin_unlock(&fs_info->trans_lock);
2434
2435         btrfs_debug(fs_info, "cleaner removing %llu", root->root_key.objectid);
2436
2437         btrfs_kill_all_delayed_nodes(root);
2438         if (root->ino_cache_inode) {
2439                 iput(root->ino_cache_inode);
2440                 root->ino_cache_inode = NULL;
2441         }
2442
2443         if (btrfs_header_backref_rev(root->node) <
2444                         BTRFS_MIXED_BACKREF_REV)
2445                 ret = btrfs_drop_snapshot(root, 0, 0);
2446         else
2447                 ret = btrfs_drop_snapshot(root, 1, 0);
2448
2449         btrfs_put_root(root);
2450         return (ret < 0) ? 0 : 1;
2451 }
2452
2453 void btrfs_apply_pending_changes(struct btrfs_fs_info *fs_info)
2454 {
2455         unsigned long prev;
2456         unsigned long bit;
2457
2458         prev = xchg(&fs_info->pending_changes, 0);
2459         if (!prev)
2460                 return;
2461
2462         bit = 1 << BTRFS_PENDING_SET_INODE_MAP_CACHE;
2463         if (prev & bit)
2464                 btrfs_set_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2465         prev &= ~bit;
2466
2467         bit = 1 << BTRFS_PENDING_CLEAR_INODE_MAP_CACHE;
2468         if (prev & bit)
2469                 btrfs_clear_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2470         prev &= ~bit;
2471
2472         bit = 1 << BTRFS_PENDING_COMMIT;
2473         if (prev & bit)
2474                 btrfs_debug(fs_info, "pending commit done");
2475         prev &= ~bit;
2476
2477         if (prev)
2478                 btrfs_warn(fs_info,
2479                         "unknown pending changes left 0x%lx, ignoring", prev);
2480 }