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