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