fs/btrfs/transaction.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/fs.h>
  20 #include <linux/slab.h>
  21 #include <linux/sched.h>
  22 #include <linux/writeback.h>
  23 #include <linux/pagemap.h>
  24 #include <linux/blkdev.h>
  25 #include <linux/uuid.h>
  26 #include "ctree.h"
  27 #include "disk-io.h"
  28 #include "transaction.h"
  29 #include "locking.h"
  30 #include "tree-log.h"
  31 #include "inode-map.h"
  32 #include "volumes.h"
  33 #include "dev-replace.h"
  34 #include "qgroup.h"
  35
  36 #define BTRFS_ROOT_TRANS_TAG 0
  37
  38 static const unsigned int btrfs_blocked_trans_types[TRANS_STATE_MAX] = {
  39         [TRANS_STATE_RUNNING]           = 0U,
  40         [TRANS_STATE_BLOCKED]           = (__TRANS_USERSPACE |
  41                                            __TRANS_START),
  42         [TRANS_STATE_COMMIT_START]      = (__TRANS_USERSPACE |
  43                                            __TRANS_START |
  44                                            __TRANS_ATTACH),
  45         [TRANS_STATE_COMMIT_DOING]      = (__TRANS_USERSPACE |
  46                                            __TRANS_START |
  47                                            __TRANS_ATTACH |
  48                                            __TRANS_JOIN),
  49         [TRANS_STATE_UNBLOCKED]         = (__TRANS_USERSPACE |
  50                                            __TRANS_START |
  51                                            __TRANS_ATTACH |
  52                                            __TRANS_JOIN |
  53                                            __TRANS_JOIN_NOLOCK),
  54         [TRANS_STATE_COMPLETED]         = (__TRANS_USERSPACE |
  55                                            __TRANS_START |
  56                                            __TRANS_ATTACH |
  57                                            __TRANS_JOIN |
  58                                            __TRANS_JOIN_NOLOCK),
  59 };
  60
  61 void btrfs_put_transaction(struct btrfs_transaction *transaction)
  62 {
  63         WARN_ON(atomic_read(&transaction->use_count) == 0);
  64         if (atomic_dec_and_test(&transaction->use_count)) {
  65                 BUG_ON(!list_empty(&transaction->list));
  66                 WARN_ON(!RB_EMPTY_ROOT(&transaction->delayed_refs.href_root));
  67                 if (transaction->delayed_refs.pending_csums)
  68                         printk(KERN_ERR "pending csums is %llu\n",
  69                                transaction->delayed_refs.pending_csums);
  70                 while (!list_empty(&transaction->pending_chunks)) {
  71                         struct extent_map *em;
  72
  73                         em = list_first_entry(&transaction->pending_chunks,
  74                                               struct extent_map, list);
  75                         list_del_init(&em->list);
  76                         free_extent_map(em);
  77                 }
  78                 /*
  79                  * If any block groups are found in ->deleted_bgs then it's
  80                  * because the transaction was aborted and a commit did not
  81                  * happen (things failed before writing the new superblock
  82                  * and calling btrfs_finish_extent_commit()), so we can not
  83                  * discard the physical locations of the block groups.
  84                  */
  85                 while (!list_empty(&transaction->deleted_bgs)) {
  86                         struct btrfs_block_group_cache *cache;
  87
  88                         cache = list_first_entry(&transaction->deleted_bgs,
  89                                                  struct btrfs_block_group_cache,
  90                                                  bg_list);
  91                         list_del_init(&cache->bg_list);
  92                         btrfs_put_block_group_trimming(cache);
  93                         btrfs_put_block_group(cache);
  94                 }
  95                 kmem_cache_free(btrfs_transaction_cachep, transaction);
  96         }
  97 }
  98
  99 static void clear_btree_io_tree(struct extent_io_tree *tree)
 100 {
 101         spin_lock(&tree->lock);
 102         /*
 103          * Do a single barrier for the waitqueue_active check here, the state
 104          * of the waitqueue should not change once clear_btree_io_tree is
 105          * called.
 106          */
 107         smp_mb();
 108         while (!RB_EMPTY_ROOT(&tree->state)) {
 109                 struct rb_node *node;
 110                 struct extent_state *state;
 111
 112                 node = rb_first(&tree->state);
 113                 state = rb_entry(node, struct extent_state, rb_node);
 114                 rb_erase(&state->rb_node, &tree->state);
 115                 RB_CLEAR_NODE(&state->rb_node);
 116                 /*
 117                  * btree io trees aren't supposed to have tasks waiting for
 118                  * changes in the flags of extent states ever.
 119                  */
 120                 ASSERT(!waitqueue_active(&state->wq));
 121                 free_extent_state(state);
 122
 123                 cond_resched_lock(&tree->lock);
 124         }
 125         spin_unlock(&tree->lock);
 126 }
 127
 128 static noinline void switch_commit_roots(struct btrfs_transaction *trans,
 129                                          struct btrfs_fs_info *fs_info)
 130 {
 131         struct btrfs_root *root, *tmp;
 132
 133         down_write(&fs_info->commit_root_sem);
 134         list_for_each_entry_safe(root, tmp, &trans->switch_commits,
 135                                  dirty_list) {
 136                 list_del_init(&root->dirty_list);
 137                 free_extent_buffer(root->commit_root);
 138                 root->commit_root = btrfs_root_node(root);
 139                 if (is_fstree(root->objectid))
 140                         btrfs_unpin_free_ino(root);
 141                 clear_btree_io_tree(&root->dirty_log_pages);
 142         }
 143
 144         /* We can free old roots now. */
 145         spin_lock(&trans->dropped_roots_lock);
 146         while (!list_empty(&trans->dropped_roots)) {
 147                 root = list_first_entry(&trans->dropped_roots,
 148                                         struct btrfs_root, root_list);
 149                 list_del_init(&root->root_list);
 150                 spin_unlock(&trans->dropped_roots_lock);
 151                 btrfs_drop_and_free_fs_root(fs_info, root);
 152                 spin_lock(&trans->dropped_roots_lock);
 153         }
 154         spin_unlock(&trans->dropped_roots_lock);
 155         up_write(&fs_info->commit_root_sem);
 156 }
 157
 158 static inline void extwriter_counter_inc(struct btrfs_transaction *trans,
 159                                          unsigned int type)
 160 {
 161         if (type & TRANS_EXTWRITERS)
 162                 atomic_inc(&trans->num_extwriters);
 163 }
 164
 165 static inline void extwriter_counter_dec(struct btrfs_transaction *trans,
 166                                          unsigned int type)
 167 {
 168         if (type & TRANS_EXTWRITERS)
 169                 atomic_dec(&trans->num_extwriters);
 170 }
 171
 172 static inline void extwriter_counter_init(struct btrfs_transaction *trans,
 173                                           unsigned int type)
 174 {
 175         atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0));
 176 }
 177
 178 static inline int extwriter_counter_read(struct btrfs_transaction *trans)
 179 {
 180         return atomic_read(&trans->num_extwriters);
 181 }
 182
 183 /*
 184  * either allocate a new transaction or hop into the existing one
 185  */
 186 static noinline int join_transaction(struct btrfs_root *root, unsigned int type)
 187 {
 188         struct btrfs_transaction *cur_trans;
 189         struct btrfs_fs_info *fs_info = root->fs_info;
 190
 191         spin_lock(&fs_info->trans_lock);
 192 loop:
 193         /* The file system has been taken offline. No new transactions. */
 194         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
 195                 spin_unlock(&fs_info->trans_lock);
 196                 return -EROFS;
 197         }
 198
 199         cur_trans = fs_info->running_transaction;
 200         if (cur_trans) {
 201                 if (cur_trans->aborted) {
 202                         spin_unlock(&fs_info->trans_lock);
 203                         return cur_trans->aborted;
 204                 }
 205                 if (btrfs_blocked_trans_types[cur_trans->state] & type) {
 206                         spin_unlock(&fs_info->trans_lock);
 207                         return -EBUSY;
 208                 }
 209                 atomic_inc(&cur_trans->use_count);
 210                 atomic_inc(&cur_trans->num_writers);
 211                 extwriter_counter_inc(cur_trans, type);
 212                 spin_unlock(&fs_info->trans_lock);
 213                 return 0;
 214         }
 215         spin_unlock(&fs_info->trans_lock);
 216
 217         /*
 218          * If we are ATTACH, we just want to catch the current transaction,
 219          * and commit it. If there is no transaction, just return ENOENT.
 220          */
 221         if (type == TRANS_ATTACH)
 222                 return -ENOENT;
 223
 224         /*
 225          * JOIN_NOLOCK only happens during the transaction commit, so
 226          * it is impossible that ->running_transaction is NULL
 227          */
 228         BUG_ON(type == TRANS_JOIN_NOLOCK);
 229
 230         cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
 231         if (!cur_trans)
 232                 return -ENOMEM;
 233
 234         spin_lock(&fs_info->trans_lock);
 235         if (fs_info->running_transaction) {
 236                 /*
 237                  * someone started a transaction after we unlocked.  Make sure
 238                  * to redo the checks above
 239                  */
 240                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 241                 goto loop;
 242         } else if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
 243                 spin_unlock(&fs_info->trans_lock);
 244                 kmem_cache_free(btrfs_transaction_cachep, cur_trans);
 245                 return -EROFS;
 246         }
 247
 248         atomic_set(&cur_trans->num_writers, 1);
 249         extwriter_counter_init(cur_trans, type);
 250         init_waitqueue_head(&cur_trans->writer_wait);
 251         init_waitqueue_head(&cur_trans->commit_wait);
 252         init_waitqueue_head(&cur_trans->pending_wait);
 253         cur_trans->state = TRANS_STATE_RUNNING;
 254         /*
 255          * One for this trans handle, one so it will live on until we
 256          * commit the transaction.
 257          */
 258         atomic_set(&cur_trans->use_count, 2);
 259         atomic_set(&cur_trans->pending_ordered, 0);
 260         cur_trans->flags = 0;
 261         cur_trans->start_time = get_seconds();
 262
 263         memset(&cur_trans->delayed_refs, 0, sizeof(cur_trans->delayed_refs));
 264
 265         cur_trans->delayed_refs.href_root = RB_ROOT;
 266         cur_trans->delayed_refs.dirty_extent_root = RB_ROOT;
 267         atomic_set(&cur_trans->delayed_refs.num_entries, 0);
 268
 269         /*
 270          * although the tree mod log is per file system and not per transaction,
 271          * the log must never go across transaction boundaries.
 272          */
 273         smp_mb();
 274         if (!list_empty(&fs_info->tree_mod_seq_list))
 275                 WARN(1, KERN_ERR "BTRFS: tree_mod_seq_list not empty when "
 276                         "creating a fresh transaction\n");
 277         if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
 278                 WARN(1, KERN_ERR "BTRFS: tree_mod_log rb tree not empty when "
 279                         "creating a fresh transaction\n");
 280         atomic64_set(&fs_info->tree_mod_seq, 0);
 281
 282         spin_lock_init(&cur_trans->delayed_refs.lock);
 283
 284         INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 285         INIT_LIST_HEAD(&cur_trans->pending_chunks);
 286         INIT_LIST_HEAD(&cur_trans->switch_commits);
 287         INIT_LIST_HEAD(&cur_trans->dirty_bgs);
 288         INIT_LIST_HEAD(&cur_trans->io_bgs);
 289         INIT_LIST_HEAD(&cur_trans->dropped_roots);
 290         mutex_init(&cur_trans->cache_write_mutex);
 291         cur_trans->num_dirty_bgs = 0;
 292         spin_lock_init(&cur_trans->dirty_bgs_lock);
 293         INIT_LIST_HEAD(&cur_trans->deleted_bgs);
 294         spin_lock_init(&cur_trans->dropped_roots_lock);
 295         list_add_tail(&cur_trans->list, &fs_info->trans_list);
 296         extent_io_tree_init(&cur_trans->dirty_pages,
 297                              fs_info->btree_inode->i_mapping);
 298         fs_info->generation++;
 299         cur_trans->transid = fs_info->generation;
 300         fs_info->running_transaction = cur_trans;
 301         cur_trans->aborted = 0;
 302         spin_unlock(&fs_info->trans_lock);
 303
 304         return 0;
 305 }
 306
 307 /*
 308  * this does all the record keeping required to make sure that a reference
 309  * counted root is properly recorded in a given transaction.  This is required
 310  * to make sure the old root from before we joined the transaction is deleted
 311  * when the transaction commits
 312  */
 313 static int record_root_in_trans(struct btrfs_trans_handle *trans,
 314                                struct btrfs_root *root,
 315                                int force)
 316 {
 317         if ((test_bit(BTRFS_ROOT_REF_COWS, &root->state) &&
 318             root->last_trans < trans->transid) || force) {
 319                 WARN_ON(root == root->fs_info->extent_root);
 320                 WARN_ON(root->commit_root != root->node);
 321
 322                 /*
 323                  * see below for IN_TRANS_SETUP usage rules
 324                  * we have the reloc mutex held now, so there
 325                  * is only one writer in this function
 326                  */
 327                 set_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
 328
 329                 /* make sure readers find IN_TRANS_SETUP before
 330                  * they find our root->last_trans update
 331                  */
 332                 smp_wmb();
 333
 334                 spin_lock(&root->fs_info->fs_roots_radix_lock);
 335                 if (root->last_trans == trans->transid && !force) {
 336                         spin_unlock(&root->fs_info->fs_roots_radix_lock);
 337                         return 0;
 338                 }
 339                 radix_tree_tag_set(&root->fs_info->fs_roots_radix,
 340                            (unsigned long)root->root_key.objectid,
 341                            BTRFS_ROOT_TRANS_TAG);
 342                 spin_unlock(&root->fs_info->fs_roots_radix_lock);
 343                 root->last_trans = trans->transid;
 344
 345                 /* this is pretty tricky.  We don't want to
 346                  * take the relocation lock in btrfs_record_root_in_trans
 347                  * unless we're really doing the first setup for this root in
 348                  * this transaction.
 349                  *
 350                  * Normally we'd use root->last_trans as a flag to decide
 351                  * if we want to take the expensive mutex.
 352                  *
 353                  * But, we have to set root->last_trans before we
 354                  * init the relocation root, otherwise, we trip over warnings
 355                  * in ctree.c.  The solution used here is to flag ourselves
 356                  * with root IN_TRANS_SETUP.  When this is 1, we're still
 357                  * fixing up the reloc trees and everyone must wait.
 358                  *
 359                  * When this is zero, they can trust root->last_trans and fly
 360                  * through btrfs_record_root_in_trans without having to take the
 361                  * lock.  smp_wmb() makes sure that all the writes above are
 362                  * done before we pop in the zero below
 363                  */
 364                 btrfs_init_reloc_root(trans, root);
 365                 smp_mb__before_atomic();
 366                 clear_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
 367         }
 368         return 0;
 369 }
 370
 371
 372 void btrfs_add_dropped_root(struct btrfs_trans_handle *trans,
 373                             struct btrfs_root *root)
 374 {
 375         struct btrfs_transaction *cur_trans = trans->transaction;
 376
 377         /* Add ourselves to the transaction dropped list */
 378         spin_lock(&cur_trans->dropped_roots_lock);
 379         list_add_tail(&root->root_list, &cur_trans->dropped_roots);
 380         spin_unlock(&cur_trans->dropped_roots_lock);
 381
 382         /* Make sure we don't try to update the root at commit time */
 383         spin_lock(&root->fs_info->fs_roots_radix_lock);
 384         radix_tree_tag_clear(&root->fs_info->fs_roots_radix,
 385                              (unsigned long)root->root_key.objectid,
 386                              BTRFS_ROOT_TRANS_TAG);
 387         spin_unlock(&root->fs_info->fs_roots_radix_lock);
 388 }
 389
 390 int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 391                                struct btrfs_root *root)
 392 {
 393         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state))
 394                 return 0;
 395
 396         /*
 397          * see record_root_in_trans for comments about IN_TRANS_SETUP usage
 398          * and barriers
 399          */
 400         smp_rmb();
 401         if (root->last_trans == trans->transid &&
 402             !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state))
 403                 return 0;
 404
 405         mutex_lock(&root->fs_info->reloc_mutex);
 406         record_root_in_trans(trans, root, 0);
 407         mutex_unlock(&root->fs_info->reloc_mutex);
 408
 409         return 0;
 410 }
 411
 412 static inline int is_transaction_blocked(struct btrfs_transaction *trans)
 413 {
 414         return (trans->state >= TRANS_STATE_BLOCKED &&
 415                 trans->state < TRANS_STATE_UNBLOCKED &&
 416                 !trans->aborted);
 417 }
 418
 419 /* wait for commit against the current transaction to become unblocked
 420  * when this is done, it is safe to start a new transaction, but the current
 421  * transaction might not be fully on disk.
 422  */
 423 static void wait_current_trans(struct btrfs_root *root)
 424 {
 425         struct btrfs_transaction *cur_trans;
 426
 427         spin_lock(&root->fs_info->trans_lock);
 428         cur_trans = root->fs_info->running_transaction;
 429         if (cur_trans && is_transaction_blocked(cur_trans)) {
 430                 atomic_inc(&cur_trans->use_count);
 431                 spin_unlock(&root->fs_info->trans_lock);
 432
 433                 wait_event(root->fs_info->transaction_wait,
 434                            cur_trans->state >= TRANS_STATE_UNBLOCKED ||
 435                            cur_trans->aborted);
 436                 btrfs_put_transaction(cur_trans);
 437         } else {
 438                 spin_unlock(&root->fs_info->trans_lock);
 439         }
 440 }
 441
 442 static int may_wait_transaction(struct btrfs_root *root, int type)
 443 {
 444         if (root->fs_info->log_root_recovering)
 445                 return 0;
 446
 447         if (type == TRANS_USERSPACE)
 448                 return 1;
 449
 450         if (type == TRANS_START &&
 451             !atomic_read(&root->fs_info->open_ioctl_trans))
 452                 return 1;
 453
 454         return 0;
 455 }
 456
 457 static inline bool need_reserve_reloc_root(struct btrfs_root *root)
 458 {
 459         if (!root->fs_info->reloc_ctl ||
 460             !test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
 461             root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID ||
 462             root->reloc_root)
 463                 return false;
 464
 465         return true;
 466 }
 467
 468 static struct btrfs_trans_handle *
 469 start_transaction(struct btrfs_root *root, unsigned int num_items,
 470                   unsigned int type, enum btrfs_reserve_flush_enum flush)
 471 {
 472         struct btrfs_trans_handle *h;
 473         struct btrfs_transaction *cur_trans;
 474         u64 num_bytes = 0;
 475         u64 qgroup_reserved = 0;
 476         bool reloc_reserved = false;
 477         int ret;
 478
 479         /* Send isn't supposed to start transactions. */
 480         ASSERT(current->journal_info != BTRFS_SEND_TRANS_STUB);
 481
 482         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
 483                 return ERR_PTR(-EROFS);
 484
 485         if (current->journal_info) {
 486                 WARN_ON(type & TRANS_EXTWRITERS);
 487                 h = current->journal_info;
 488                 h->use_count++;
 489                 WARN_ON(h->use_count > 2);
 490                 h->orig_rsv = h->block_rsv;
 491                 h->block_rsv = NULL;
 492                 goto got_it;
 493         }
 494
 495         /*
 496          * Do the reservation before we join the transaction so we can do all
 497          * the appropriate flushing if need be.
 498          */
 499         if (num_items > 0 && root != root->fs_info->chunk_root) {
 500                 qgroup_reserved = num_items * root->nodesize;
 501                 ret = btrfs_qgroup_reserve_meta(root, qgroup_reserved);
 502                 if (ret)
 503                         return ERR_PTR(ret);
 504
 505                 num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
 506                 /*
 507                  * Do the reservation for the relocation root creation
 508                  */
 509                 if (need_reserve_reloc_root(root)) {
 510                         num_bytes += root->nodesize;
 511                         reloc_reserved = true;
 512                 }
 513
 514                 ret = btrfs_block_rsv_add(root,
 515                                           &root->fs_info->trans_block_rsv,
 516                                           num_bytes, flush);
 517                 if (ret)
 518                         goto reserve_fail;
 519         }
 520 again:
 521         h = kmem_cache_zalloc(btrfs_trans_handle_cachep, GFP_NOFS);
 522         if (!h) {
 523                 ret = -ENOMEM;
 524                 goto alloc_fail;
 525         }
 526
 527         /*
 528          * If we are JOIN_NOLOCK we're already committing a transaction and
 529          * waiting on this guy, so we don't need to do the sb_start_intwrite
 530          * because we're already holding a ref.  We need this because we could
 531          * have raced in and did an fsync() on a file which can kick a commit
 532          * and then we deadlock with somebody doing a freeze.
 533          *
 534          * If we are ATTACH, it means we just want to catch the current
 535          * transaction and commit it, so we needn't do sb_start_intwrite().
 536          */
 537         if (type & __TRANS_FREEZABLE)
 538                 sb_start_intwrite(root->fs_info->sb);
 539
 540         if (may_wait_transaction(root, type))
 541                 wait_current_trans(root);
 542
 543         do {
 544                 ret = join_transaction(root, type);
 545                 if (ret == -EBUSY) {
 546                         wait_current_trans(root);
 547                         if (unlikely(type == TRANS_ATTACH))
 548                                 ret = -ENOENT;
 549                 }
 550         } while (ret == -EBUSY);
 551
 552         if (ret < 0) {
 553                 /* We must get the transaction if we are JOIN_NOLOCK. */
 554                 BUG_ON(type == TRANS_JOIN_NOLOCK);
 555                 goto join_fail;
 556         }
 557
 558         cur_trans = root->fs_info->running_transaction;
 559
 560         h->transid = cur_trans->transid;
 561         h->transaction = cur_trans;
 562         h->root = root;
 563         h->use_count = 1;
 564         h->fs_info = root->fs_info;
 565
 566         h->type = type;
 567         h->can_flush_pending_bgs = true;
 568         INIT_LIST_HEAD(&h->qgroup_ref_list);
 569         INIT_LIST_HEAD(&h->new_bgs);
 570
 571         smp_mb();
 572         if (cur_trans->state >= TRANS_STATE_BLOCKED &&
 573             may_wait_transaction(root, type)) {
 574                 current->journal_info = h;
 575                 btrfs_commit_transaction(h, root);
 576                 goto again;
 577         }
 578
 579         if (num_bytes) {
 580                 trace_btrfs_space_reservation(root->fs_info, "transaction",
 581                                               h->transid, num_bytes, 1);
 582                 h->block_rsv = &root->fs_info->trans_block_rsv;
 583                 h->bytes_reserved = num_bytes;
 584                 h->reloc_reserved = reloc_reserved;
 585         }
 586
 587 got_it:
 588         btrfs_record_root_in_trans(h, root);
 589
 590         if (!current->journal_info && type != TRANS_USERSPACE)
 591                 current->journal_info = h;
 592         return h;
 593
 594 join_fail:
 595         if (type & __TRANS_FREEZABLE)
 596                 sb_end_intwrite(root->fs_info->sb);
 597         kmem_cache_free(btrfs_trans_handle_cachep, h);
 598 alloc_fail:
 599         if (num_bytes)
 600                 btrfs_block_rsv_release(root, &root->fs_info->trans_block_rsv,
 601                                         num_bytes);
 602 reserve_fail:
 603         btrfs_qgroup_free_meta(root, qgroup_reserved);
 604         return ERR_PTR(ret);
 605 }
 606
 607 struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 608                                                    unsigned int num_items)
 609 {
 610         return start_transaction(root, num_items, TRANS_START,
 611                                  BTRFS_RESERVE_FLUSH_ALL);
 612 }
 613 struct btrfs_trans_handle *btrfs_start_transaction_fallback_global_rsv(
 614                                         struct btrfs_root *root,
 615                                         unsigned int num_items,
 616                                         int min_factor)
 617 {
 618         struct btrfs_trans_handle *trans;
 619         u64 num_bytes;
 620         int ret;
 621
 622         trans = btrfs_start_transaction(root, num_items);
 623         if (!IS_ERR(trans) || PTR_ERR(trans) != -ENOSPC)
 624                 return trans;
 625
 626         trans = btrfs_start_transaction(root, 0);
 627         if (IS_ERR(trans))
 628                 return trans;
 629
 630         num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
 631         ret = btrfs_cond_migrate_bytes(root->fs_info,
 632                                        &root->fs_info->trans_block_rsv,
 633                                        num_bytes,
 634                                        min_factor);
 635         if (ret) {
 636                 btrfs_end_transaction(trans, root);
 637                 return ERR_PTR(ret);
 638         }
 639
 640         trans->block_rsv = &root->fs_info->trans_block_rsv;
 641         trans->bytes_reserved = num_bytes;
 642         trace_btrfs_space_reservation(root->fs_info, "transaction",
 643                                       trans->transid, num_bytes, 1);
 644
 645         return trans;
 646 }
 647
 648 struct btrfs_trans_handle *btrfs_start_transaction_lflush(
 649                                         struct btrfs_root *root,
 650                                         unsigned int num_items)
 651 {
 652         return start_transaction(root, num_items, TRANS_START,
 653                                  BTRFS_RESERVE_FLUSH_LIMIT);
 654 }
 655
 656 struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 657 {
 658         return start_transaction(root, 0, TRANS_JOIN,
 659                                  BTRFS_RESERVE_NO_FLUSH);
 660 }
 661
 662 struct btrfs_trans_handle *btrfs_join_transaction_nolock(struct btrfs_root *root)
 663 {
 664         return start_transaction(root, 0, TRANS_JOIN_NOLOCK,
 665                                  BTRFS_RESERVE_NO_FLUSH);
 666 }
 667
 668 struct btrfs_trans_handle *btrfs_start_ioctl_transaction(struct btrfs_root *root)
 669 {
 670         return start_transaction(root, 0, TRANS_USERSPACE,
 671                                  BTRFS_RESERVE_NO_FLUSH);
 672 }
 673
 674 /*
 675  * btrfs_attach_transaction() - catch the running transaction
 676  *
 677  * It is used when we want to commit the current the transaction, but
 678  * don't want to start a new one.
 679  *
 680  * Note: If this function return -ENOENT, it just means there is no
 681  * running transaction. But it is possible that the inactive transaction
 682  * is still in the memory, not fully on disk. If you hope there is no
 683  * inactive transaction in the fs when -ENOENT is returned, you should
 684  * invoke
 685  *     btrfs_attach_transaction_barrier()
 686  */
 687 struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
 688 {
 689         return start_transaction(root, 0, TRANS_ATTACH,
 690                                  BTRFS_RESERVE_NO_FLUSH);
 691 }
 692
 693 /*
 694  * btrfs_attach_transaction_barrier() - catch the running transaction
 695  *
 696  * It is similar to the above function, the differentia is this one
 697  * will wait for all the inactive transactions until they fully
 698  * complete.
 699  */
 700 struct btrfs_trans_handle *
 701 btrfs_attach_transaction_barrier(struct btrfs_root *root)
 702 {
 703         struct btrfs_trans_handle *trans;
 704
 705         trans = start_transaction(root, 0, TRANS_ATTACH,
 706                                   BTRFS_RESERVE_NO_FLUSH);
 707         if (IS_ERR(trans) && PTR_ERR(trans) == -ENOENT)
 708                 btrfs_wait_for_commit(root, 0);
 709
 710         return trans;
 711 }
 712
 713 /* wait for a transaction commit to be fully complete */
 714 static noinline void wait_for_commit(struct btrfs_root *root,
 715                                     struct btrfs_transaction *commit)
 716 {
 717         wait_event(commit->commit_wait, commit->state == TRANS_STATE_COMPLETED);
 718 }
 719
 720 int btrfs_wait_for_commit(struct btrfs_root *root, u64 transid)
 721 {
 722         struct btrfs_transaction *cur_trans = NULL, *t;
 723         int ret = 0;
 724
 725         if (transid) {
 726                 if (transid <= root->fs_info->last_trans_committed)
 727                         goto out;
 728
 729                 /* find specified transaction */
 730                 spin_lock(&root->fs_info->trans_lock);
 731                 list_for_each_entry(t, &root->fs_info->trans_list, list) {
 732                         if (t->transid == transid) {
 733                                 cur_trans = t;
 734                                 atomic_inc(&cur_trans->use_count);
 735                                 ret = 0;
 736                                 break;
 737                         }
 738                         if (t->transid > transid) {
 739                                 ret = 0;
 740                                 break;
 741                         }
 742                 }
 743                 spin_unlock(&root->fs_info->trans_lock);
 744
 745                 /*
 746                  * The specified transaction doesn't exist, or we
 747                  * raced with btrfs_commit_transaction
 748                  */
 749                 if (!cur_trans) {
 750                         if (transid > root->fs_info->last_trans_committed)
 751                                 ret = -EINVAL;
 752                         goto out;
 753                 }
 754         } else {
 755                 /* find newest transaction that is committing | committed */
 756                 spin_lock(&root->fs_info->trans_lock);
 757                 list_for_each_entry_reverse(t, &root->fs_info->trans_list,
 758                                             list) {
 759                         if (t->state >= TRANS_STATE_COMMIT_START) {
 760                                 if (t->state == TRANS_STATE_COMPLETED)
 761                                         break;
 762                                 cur_trans = t;
 763                                 atomic_inc(&cur_trans->use_count);
 764                                 break;
 765                         }
 766                 }
 767                 spin_unlock(&root->fs_info->trans_lock);
 768                 if (!cur_trans)
 769                         goto out;  /* nothing committing|committed */
 770         }
 771
 772         wait_for_commit(root, cur_trans);
 773         btrfs_put_transaction(cur_trans);
 774 out:
 775         return ret;
 776 }
 777
 778 void btrfs_throttle(struct btrfs_root *root)
 779 {
 780         if (!atomic_read(&root->fs_info->open_ioctl_trans))
 781                 wait_current_trans(root);
 782 }
 783
 784 static int should_end_transaction(struct btrfs_trans_handle *trans,
 785                                   struct btrfs_root *root)
 786 {
 787         if (root->fs_info->global_block_rsv.space_info->full &&
 788             btrfs_check_space_for_delayed_refs(trans, root))
 789                 return 1;
 790
 791         return !!btrfs_block_rsv_check(root, &root->fs_info->global_block_rsv, 5);
 792 }
 793
 794 int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 795                                  struct btrfs_root *root)
 796 {
 797         struct btrfs_transaction *cur_trans = trans->transaction;
 798         int updates;
 799         int err;
 800
 801         smp_mb();
 802         if (cur_trans->state >= TRANS_STATE_BLOCKED ||
 803             cur_trans->delayed_refs.flushing)
 804                 return 1;
 805
 806         updates = trans->delayed_ref_updates;
 807         trans->delayed_ref_updates = 0;
 808         if (updates) {
 809                 err = btrfs_run_delayed_refs(trans, root, updates * 2);
 810                 if (err) /* Error code will also eval true */
 811                         return err;
 812         }
 813
 814         return should_end_transaction(trans, root);
 815 }
 816
 817 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 818                           struct btrfs_root *root, int throttle)
 819 {
 820         struct btrfs_transaction *cur_trans = trans->transaction;
 821         struct btrfs_fs_info *info = root->fs_info;
 822         u64 transid = trans->transid;
 823         unsigned long cur = trans->delayed_ref_updates;
 824         int lock = (trans->type != TRANS_JOIN_NOLOCK);
 825         int err = 0;
 826         int must_run_delayed_refs = 0;
 827
 828         if (trans->use_count > 1) {
 829                 trans->use_count--;
 830                 trans->block_rsv = trans->orig_rsv;
 831                 return 0;
 832         }
 833
 834         btrfs_trans_release_metadata(trans, root);
 835         trans->block_rsv = NULL;
 836
 837         if (!list_empty(&trans->new_bgs))
 838                 btrfs_create_pending_block_groups(trans, root);
 839
 840         trans->delayed_ref_updates = 0;
 841         if (!trans->sync) {
 842                 must_run_delayed_refs =
 843                         btrfs_should_throttle_delayed_refs(trans, root);
 844                 cur = max_t(unsigned long, cur, 32);
 845
 846                 /*
 847                  * don't make the caller wait if they are from a NOLOCK
 848                  * or ATTACH transaction, it will deadlock with commit
 849                  */
 850                 if (must_run_delayed_refs == 1 &&
 851                     (trans->type & (__TRANS_JOIN_NOLOCK | __TRANS_ATTACH)))
 852                         must_run_delayed_refs = 2;
 853         }
 854
 855         btrfs_trans_release_metadata(trans, root);
 856         trans->block_rsv = NULL;
 857
 858         if (!list_empty(&trans->new_bgs))
 859                 btrfs_create_pending_block_groups(trans, root);
 860
 861         btrfs_trans_release_chunk_metadata(trans);
 862
 863         if (lock && !atomic_read(&root->fs_info->open_ioctl_trans) &&
 864             should_end_transaction(trans, root) &&
 865             ACCESS_ONCE(cur_trans->state) == TRANS_STATE_RUNNING) {
 866                 spin_lock(&info->trans_lock);
 867                 if (cur_trans->state == TRANS_STATE_RUNNING)
 868                         cur_trans->state = TRANS_STATE_BLOCKED;
 869                 spin_unlock(&info->trans_lock);
 870         }
 871
 872         if (lock && ACCESS_ONCE(cur_trans->state) == TRANS_STATE_BLOCKED) {
 873                 if (throttle)
 874                         return btrfs_commit_transaction(trans, root);
 875                 else
 876                         wake_up_process(info->transaction_kthread);
 877         }
 878
 879         if (trans->type & __TRANS_FREEZABLE)
 880                 sb_end_intwrite(root->fs_info->sb);
 881
 882         WARN_ON(cur_trans != info->running_transaction);
 883         WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
 884         atomic_dec(&cur_trans->num_writers);
 885         extwriter_counter_dec(cur_trans, trans->type);
 886
 887         /*
 888          * Make sure counter is updated before we wake up waiters.
 889          */
 890         smp_mb();
 891         if (waitqueue_active(&cur_trans->writer_wait))
 892                 wake_up(&cur_trans->writer_wait);
 893         btrfs_put_transaction(cur_trans);
 894
 895         if (current->journal_info == trans)
 896                 current->journal_info = NULL;
 897
 898         if (throttle)
 899                 btrfs_run_delayed_iputs(root);
 900
 901         if (trans->aborted ||
 902             test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state)) {
 903                 wake_up_process(info->transaction_kthread);
 904                 err = -EIO;
 905         }
 906         assert_qgroups_uptodate(trans);
 907
 908         kmem_cache_free(btrfs_trans_handle_cachep, trans);
 909         if (must_run_delayed_refs) {
 910                 btrfs_async_run_delayed_refs(root, cur, transid,
 911                                              must_run_delayed_refs == 1);
 912         }
 913         return err;
 914 }
 915
 916 int btrfs_end_transaction(struct btrfs_trans_handle *trans,
 917                           struct btrfs_root *root)
 918 {
 919         return __btrfs_end_transaction(trans, root, 0);
 920 }
 921
 922 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 923                                    struct btrfs_root *root)
 924 {
 925         return __btrfs_end_transaction(trans, root, 1);
 926 }
 927
 928 /*
 929  * when btree blocks are allocated, they have some corresponding bits set for
 930  * them in one of two extent_io trees.  This is used to make sure all of
 931  * those extents are sent to disk but does not wait on them
 932  */
 933 int btrfs_write_marked_extents(struct btrfs_root *root,
 934                                struct extent_io_tree *dirty_pages, int mark)
 935 {
 936         int err = 0;
 937         int werr = 0;
 938         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 939         struct extent_state *cached_state = NULL;
 940         u64 start = 0;
 941         u64 end;
 942
 943         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
 944                                       mark, &cached_state)) {
 945                 bool wait_writeback = false;
 946
 947                 err = convert_extent_bit(dirty_pages, start, end,
 948                                          EXTENT_NEED_WAIT,
 949                                          mark, &cached_state);
 950                 /*
 951                  * convert_extent_bit can return -ENOMEM, which is most of the
 952                  * time a temporary error. So when it happens, ignore the error
 953                  * and wait for writeback of this range to finish - because we
 954                  * failed to set the bit EXTENT_NEED_WAIT for the range, a call
 955                  * to btrfs_wait_marked_extents() would not know that writeback
 956                  * for this range started and therefore wouldn't wait for it to
 957                  * finish - we don't want to commit a superblock that points to
 958                  * btree nodes/leafs for which writeback hasn't finished yet
 959                  * (and without errors).
 960                  * We cleanup any entries left in the io tree when committing
 961                  * the transaction (through clear_btree_io_tree()).
 962                  */
 963                 if (err == -ENOMEM) {
 964                         err = 0;
 965                         wait_writeback = true;
 966                 }
 967                 if (!err)
 968                         err = filemap_fdatawrite_range(mapping, start, end);
 969                 if (err)
 970                         werr = err;
 971                 else if (wait_writeback)
 972                         werr = filemap_fdatawait_range(mapping, start, end);
 973                 free_extent_state(cached_state);
 974                 cached_state = NULL;
 975                 cond_resched();
 976                 start = end + 1;
 977         }
 978         return werr;
 979 }
 980
 981 /*
 982  * when btree blocks are allocated, they have some corresponding bits set for
 983  * them in one of two extent_io trees.  This is used to make sure all of
 984  * those extents are on disk for transaction or log commit.  We wait
 985  * on all the pages and clear them from the dirty pages state tree
 986  */
 987 int btrfs_wait_marked_extents(struct btrfs_root *root,
 988                               struct extent_io_tree *dirty_pages, int mark)
 989 {
 990         int err = 0;
 991         int werr = 0;
 992         struct address_space *mapping = root->fs_info->btree_inode->i_mapping;
 993         struct extent_state *cached_state = NULL;
 994         u64 start = 0;
 995         u64 end;
 996         struct btrfs_inode *btree_ino = BTRFS_I(root->fs_info->btree_inode);
 997         bool errors = false;
 998
 999         while (!find_first_extent_bit(dirty_pages, start, &start, &end,
1000                                       EXTENT_NEED_WAIT, &cached_state)) {
1001                 /*
1002                  * Ignore -ENOMEM errors returned by clear_extent_bit().
1003                  * When committing the transaction, we'll remove any entries
1004                  * left in the io tree. For a log commit, we don't remove them
1005                  * after committing the log because the tree can be accessed
1006                  * concurrently - we do it only at transaction commit time when
1007                  * it's safe to do it (through clear_btree_io_tree()).
1008                  */
1009                 err = clear_extent_bit(dirty_pages, start, end,
1010                                        EXTENT_NEED_WAIT,
1011                                        0, 0, &cached_state, GFP_NOFS);
1012                 if (err == -ENOMEM)
1013                         err = 0;
1014                 if (!err)
1015                         err = filemap_fdatawait_range(mapping, start, end);
1016                 if (err)
1017                         werr = err;
1018                 free_extent_state(cached_state);
1019                 cached_state = NULL;
1020                 cond_resched();
1021                 start = end + 1;
1022         }
1023         if (err)
1024                 werr = err;
1025
1026         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
1027                 if ((mark & EXTENT_DIRTY) &&
1028                     test_and_clear_bit(BTRFS_INODE_BTREE_LOG1_ERR,
1029                                        &btree_ino->runtime_flags))
1030                         errors = true;
1031
1032                 if ((mark & EXTENT_NEW) &&
1033                     test_and_clear_bit(BTRFS_INODE_BTREE_LOG2_ERR,
1034                                        &btree_ino->runtime_flags))
1035                         errors = true;
1036         } else {
1037                 if (test_and_clear_bit(BTRFS_INODE_BTREE_ERR,
1038                                        &btree_ino->runtime_flags))
1039                         errors = true;
1040         }
1041
1042         if (errors && !werr)
1043                 werr = -EIO;
1044
1045         return werr;
1046 }
1047
1048 /*
1049  * when btree blocks are allocated, they have some corresponding bits set for
1050  * them in one of two extent_io trees.  This is used to make sure all of
1051  * those extents are on disk for transaction or log commit
1052  */
1053 static int btrfs_write_and_wait_marked_extents(struct btrfs_root *root,
1054                                 struct extent_io_tree *dirty_pages, int mark)
1055 {
1056         int ret;
1057         int ret2;
1058         struct blk_plug plug;
1059
1060         blk_start_plug(&plug);
1061         ret = btrfs_write_marked_extents(root, dirty_pages, mark);
1062         blk_finish_plug(&plug);
1063         ret2 = btrfs_wait_marked_extents(root, dirty_pages, mark);
1064
1065         if (ret)
1066                 return ret;
1067         if (ret2)
1068                 return ret2;
1069         return 0;
1070 }
1071
1072 static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
1073                                      struct btrfs_root *root)
1074 {
1075         int ret;
1076
1077         ret = btrfs_write_and_wait_marked_extents(root,
1078                                            &trans->transaction->dirty_pages,
1079                                            EXTENT_DIRTY);
1080         clear_btree_io_tree(&trans->transaction->dirty_pages);
1081
1082         return ret;
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_root *tree_root = root->fs_info->tree_root;
1102
1103         old_root_used = btrfs_root_used(&root->root_item);
1104
1105         while (1) {
1106                 old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1107                 if (old_root_bytenr == root->node->start &&
1108                     old_root_used == btrfs_root_used(&root->root_item))
1109                         break;
1110
1111                 btrfs_set_root_node(&root->root_item, root->node);
1112                 ret = btrfs_update_root(trans, tree_root,
1113                                         &root->root_key,
1114                                         &root->root_item);
1115                 if (ret)
1116                         return ret;
1117
1118                 old_root_used = btrfs_root_used(&root->root_item);
1119         }
1120
1121         return 0;
1122 }
1123
1124 /*
1125  * update all the cowonly tree roots on disk
1126  *
1127  * The error handling in this function may not be obvious. Any of the
1128  * failures will cause the file system to go offline. We still need
1129  * to clean up the delayed refs.
1130  */
1131 static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans,
1132                                          struct btrfs_root *root)
1133 {
1134         struct btrfs_fs_info *fs_info = root->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, root, (unsigned long)-1);
1151         if (ret)
1152                 return ret;
1153
1154         ret = btrfs_run_dev_stats(trans, root->fs_info);
1155         if (ret)
1156                 return ret;
1157         ret = btrfs_run_dev_replace(trans, root->fs_info);
1158         if (ret)
1159                 return ret;
1160         ret = btrfs_run_qgroups(trans, root->fs_info);
1161         if (ret)
1162                 return ret;
1163
1164         ret = btrfs_setup_space_cache(trans, root);
1165         if (ret)
1166                 return ret;
1167
1168         /* run_qgroups might have added some more refs */
1169         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1170         if (ret)
1171                 return ret;
1172 again:
1173         while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1174                 next = fs_info->dirty_cowonly_roots.next;
1175                 list_del_init(next);
1176                 root = list_entry(next, struct btrfs_root, dirty_list);
1177                 clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1178
1179                 if (root != fs_info->extent_root)
1180                         list_add_tail(&root->dirty_list,
1181                                       &trans->transaction->switch_commits);
1182                 ret = update_cowonly_root(trans, root);
1183                 if (ret)
1184                         return ret;
1185                 ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1186                 if (ret)
1187                         return ret;
1188         }
1189
1190         while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1191                 ret = btrfs_write_dirty_block_groups(trans, root);
1192                 if (ret)
1193                         return ret;
1194                 ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1195                 if (ret)
1196                         return ret;
1197         }
1198
1199         if (!list_empty(&fs_info->dirty_cowonly_roots))
1200                 goto again;
1201
1202         list_add_tail(&fs_info->extent_root->dirty_list,
1203                       &trans->transaction->switch_commits);
1204         btrfs_after_dev_replace_commit(fs_info);
1205
1206         return 0;
1207 }
1208
1209 /*
1210  * dead roots are old snapshots that need to be deleted.  This allocates
1211  * a dirty root struct and adds it into the list of dead roots that need to
1212  * be deleted
1213  */
1214 void btrfs_add_dead_root(struct btrfs_root *root)
1215 {
1216         spin_lock(&root->fs_info->trans_lock);
1217         if (list_empty(&root->root_list))
1218                 list_add_tail(&root->root_list, &root->fs_info->dead_roots);
1219         spin_unlock(&root->fs_info->trans_lock);
1220 }
1221
1222 /*
1223  * update all the cowonly tree roots on disk
1224  */
1225 static noinline int commit_fs_roots(struct btrfs_trans_handle *trans,
1226                                     struct btrfs_root *root)
1227 {
1228         struct btrfs_root *gang[8];
1229         struct btrfs_fs_info *fs_info = root->fs_info;
1230         int i;
1231         int ret;
1232         int err = 0;
1233
1234         spin_lock(&fs_info->fs_roots_radix_lock);
1235         while (1) {
1236                 ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1237                                                  (void **)gang, 0,
1238                                                  ARRAY_SIZE(gang),
1239                                                  BTRFS_ROOT_TRANS_TAG);
1240                 if (ret == 0)
1241                         break;
1242                 for (i = 0; i < ret; i++) {
1243                         root = gang[i];
1244                         radix_tree_tag_clear(&fs_info->fs_roots_radix,
1245                                         (unsigned long)root->root_key.objectid,
1246                                         BTRFS_ROOT_TRANS_TAG);
1247                         spin_unlock(&fs_info->fs_roots_radix_lock);
1248
1249                         btrfs_free_log(trans, root);
1250                         btrfs_update_reloc_root(trans, root);
1251                         btrfs_orphan_commit_root(trans, root);
1252
1253                         btrfs_save_ino_cache(root, trans);
1254
1255                         /* see comments in should_cow_block() */
1256                         clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1257                         smp_mb__after_atomic();
1258
1259                         if (root->commit_root != root->node) {
1260                                 list_add_tail(&root->dirty_list,
1261                                         &trans->transaction->switch_commits);
1262                                 btrfs_set_root_node(&root->root_item,
1263                                                     root->node);
1264                         }
1265
1266                         err = btrfs_update_root(trans, fs_info->tree_root,
1267                                                 &root->root_key,
1268                                                 &root->root_item);
1269                         spin_lock(&fs_info->fs_roots_radix_lock);
1270                         if (err)
1271                                 break;
1272                         btrfs_qgroup_free_meta_all(root);
1273                 }
1274         }
1275         spin_unlock(&fs_info->fs_roots_radix_lock);
1276         return err;
1277 }
1278
1279 /*
1280  * defrag a given btree.
1281  * Every leaf in the btree is read and defragged.
1282  */
1283 int btrfs_defrag_root(struct btrfs_root *root)
1284 {
1285         struct btrfs_fs_info *info = root->fs_info;
1286         struct btrfs_trans_handle *trans;
1287         int ret;
1288
1289         if (test_and_set_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state))
1290                 return 0;
1291
1292         while (1) {
1293                 trans = btrfs_start_transaction(root, 0);
1294                 if (IS_ERR(trans))
1295                         return PTR_ERR(trans);
1296
1297                 ret = btrfs_defrag_leaves(trans, root);
1298
1299                 btrfs_end_transaction(trans, root);
1300                 btrfs_btree_balance_dirty(info->tree_root);
1301                 cond_resched();
1302
1303                 if (btrfs_fs_closing(root->fs_info) || ret != -EAGAIN)
1304                         break;
1305
1306                 if (btrfs_defrag_cancelled(root->fs_info)) {
1307                         pr_debug("BTRFS: defrag_root cancelled\n");
1308                         ret = -EAGAIN;
1309                         break;
1310                 }
1311         }
1312         clear_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state);
1313         return ret;
1314 }
1315
1316 /*
1317  * Do all special snapshot related qgroup dirty hack.
1318  *
1319  * Will do all needed qgroup inherit and dirty hack like switch commit
1320  * roots inside one transaction and write all btree into disk, to make
1321  * qgroup works.
1322  */
1323 static int qgroup_account_snapshot(struct btrfs_trans_handle *trans,
1324                                    struct btrfs_root *src,
1325                                    struct btrfs_root *parent,
1326                                    struct btrfs_qgroup_inherit *inherit,
1327                                    u64 dst_objectid)
1328 {
1329         struct btrfs_fs_info *fs_info = src->fs_info;
1330         int ret;
1331
1332         /*
1333          * Save some performance in the case that qgroups are not
1334          * enabled. If this check races with the ioctl, rescan will
1335          * kick in anyway.
1336          */
1337         mutex_lock(&fs_info->qgroup_ioctl_lock);
1338         if (!fs_info->quota_enabled) {
1339                 mutex_unlock(&fs_info->qgroup_ioctl_lock);
1340                 return 0;
1341         }
1342         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1343
1344         /*
1345          * We are going to commit transaction, see btrfs_commit_transaction()
1346          * comment for reason locking tree_log_mutex
1347          */
1348         mutex_lock(&fs_info->tree_log_mutex);
1349
1350         ret = commit_fs_roots(trans, src);
1351         if (ret)
1352                 goto out;
1353         ret = btrfs_qgroup_prepare_account_extents(trans, fs_info);
1354         if (ret < 0)
1355                 goto out;
1356         ret = btrfs_qgroup_account_extents(trans, fs_info);
1357         if (ret < 0)
1358                 goto out;
1359
1360         /* Now qgroup are all updated, we can inherit it to new qgroups */
1361         ret = btrfs_qgroup_inherit(trans, fs_info,
1362                                    src->root_key.objectid, dst_objectid,
1363                                    inherit);
1364         if (ret < 0)
1365                 goto out;
1366
1367         /*
1368          * Now we do a simplified commit transaction, which will:
1369          * 1) commit all subvolume and extent tree
1370          *    To ensure all subvolume and extent tree have a valid
1371          *    commit_root to accounting later insert_dir_item()
1372          * 2) write all btree blocks onto disk
1373          *    This is to make sure later btree modification will be cowed
1374          *    Or commit_root can be populated and cause wrong qgroup numbers
1375          * In this simplified commit, we don't really care about other trees
1376          * like chunk and root tree, as they won't affect qgroup.
1377          * And we don't write super to avoid half committed status.
1378          */
1379         ret = commit_cowonly_roots(trans, src);
1380         if (ret)
1381                 goto out;
1382         switch_commit_roots(trans->transaction, fs_info);
1383         ret = btrfs_write_and_wait_transaction(trans, src);
1384         if (ret)
1385                 btrfs_handle_fs_error(fs_info, ret,
1386                         "Error while writing out transaction for qgroup");
1387
1388 out:
1389         mutex_unlock(&fs_info->tree_log_mutex);
1390
1391         /*
1392          * Force parent root to be updated, as we recorded it before so its
1393          * last_trans == cur_transid.
1394          * Or it won't be committed again onto disk after later
1395          * insert_dir_item()
1396          */
1397         if (!ret)
1398                 record_root_in_trans(trans, parent, 1);
1399         return ret;
1400 }
1401
1402 /*
1403  * new snapshots need to be created at a very specific time in the
1404  * transaction commit.  This does the actual creation.
1405  *
1406  * Note:
1407  * If the error which may affect the commitment of the current transaction
1408  * happens, we should return the error number. If the error which just affect
1409  * the creation of the pending snapshots, just return 0.
1410  */
1411 static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1412                                    struct btrfs_fs_info *fs_info,
1413                                    struct btrfs_pending_snapshot *pending)
1414 {
1415         struct btrfs_key key;
1416         struct btrfs_root_item *new_root_item;
1417         struct btrfs_root *tree_root = fs_info->tree_root;
1418         struct btrfs_root *root = pending->root;
1419         struct btrfs_root *parent_root;
1420         struct btrfs_block_rsv *rsv;
1421         struct inode *parent_inode;
1422         struct btrfs_path *path;
1423         struct btrfs_dir_item *dir_item;
1424         struct dentry *dentry;
1425         struct extent_buffer *tmp;
1426         struct extent_buffer *old;
1427         struct timespec cur_time;
1428         int ret = 0;
1429         u64 to_reserve = 0;
1430         u64 index = 0;
1431         u64 objectid;
1432         u64 root_flags;
1433         uuid_le new_uuid;
1434
1435         ASSERT(pending->path);
1436         path = pending->path;
1437
1438         ASSERT(pending->root_item);
1439         new_root_item = pending->root_item;
1440
1441         pending->error = btrfs_find_free_objectid(tree_root, &objectid);
1442         if (pending->error)
1443                 goto no_free_objectid;
1444
1445         /*
1446          * Make qgroup to skip current new snapshot's qgroupid, as it is
1447          * accounted by later btrfs_qgroup_inherit().
1448          */
1449         btrfs_set_skip_qgroup(trans, objectid);
1450
1451         btrfs_reloc_pre_snapshot(pending, &to_reserve);
1452
1453         if (to_reserve > 0) {
1454                 pending->error = btrfs_block_rsv_add(root,
1455                                                      &pending->block_rsv,
1456                                                      to_reserve,
1457                                                      BTRFS_RESERVE_NO_FLUSH);
1458                 if (pending->error)
1459                         goto clear_skip_qgroup;
1460         }
1461
1462         key.objectid = objectid;
1463         key.offset = (u64)-1;
1464         key.type = BTRFS_ROOT_ITEM_KEY;
1465
1466         rsv = trans->block_rsv;
1467         trans->block_rsv = &pending->block_rsv;
1468         trans->bytes_reserved = trans->block_rsv->reserved;
1469         trace_btrfs_space_reservation(root->fs_info, "transaction",
1470                                       trans->transid,
1471                                       trans->bytes_reserved, 1);
1472         dentry = pending->dentry;
1473         parent_inode = pending->dir;
1474         parent_root = BTRFS_I(parent_inode)->root;
1475         record_root_in_trans(trans, parent_root, 0);
1476
1477         cur_time = current_time(parent_inode);
1478
1479         /*
1480          * insert the directory item
1481          */
1482         ret = btrfs_set_inode_index(parent_inode, &index);
1483         BUG_ON(ret); /* -ENOMEM */
1484
1485         /* check if there is a file/dir which has the same name. */
1486         dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1487                                          btrfs_ino(parent_inode),
1488                                          dentry->d_name.name,
1489                                          dentry->d_name.len, 0);
1490         if (dir_item != NULL && !IS_ERR(dir_item)) {
1491                 pending->error = -EEXIST;
1492                 goto dir_item_existed;
1493         } else if (IS_ERR(dir_item)) {
1494                 ret = PTR_ERR(dir_item);
1495                 btrfs_abort_transaction(trans, ret);
1496                 goto fail;
1497         }
1498         btrfs_release_path(path);
1499
1500         /*
1501          * pull in the delayed directory update
1502          * and the delayed inode item
1503          * otherwise we corrupt the FS during
1504          * snapshot
1505          */
1506         ret = btrfs_run_delayed_items(trans, root);
1507         if (ret) {      /* Transaction aborted */
1508                 btrfs_abort_transaction(trans, ret);
1509                 goto fail;
1510         }
1511
1512         record_root_in_trans(trans, root, 0);
1513         btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1514         memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1515         btrfs_check_and_init_root_item(new_root_item);
1516
1517         root_flags = btrfs_root_flags(new_root_item);
1518         if (pending->readonly)
1519                 root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1520         else
1521                 root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1522         btrfs_set_root_flags(new_root_item, root_flags);
1523
1524         btrfs_set_root_generation_v2(new_root_item,
1525                         trans->transid);
1526         uuid_le_gen(&new_uuid);
1527         memcpy(new_root_item->uuid, new_uuid.b, BTRFS_UUID_SIZE);
1528         memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1529                         BTRFS_UUID_SIZE);
1530         if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1531                 memset(new_root_item->received_uuid, 0,
1532                        sizeof(new_root_item->received_uuid));
1533                 memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1534                 memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1535                 btrfs_set_root_stransid(new_root_item, 0);
1536                 btrfs_set_root_rtransid(new_root_item, 0);
1537         }
1538         btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1539         btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1540         btrfs_set_root_otransid(new_root_item, trans->transid);
1541
1542         old = btrfs_lock_root_node(root);
1543         ret = btrfs_cow_block(trans, root, old, NULL, 0, &old);
1544         if (ret) {
1545                 btrfs_tree_unlock(old);
1546                 free_extent_buffer(old);
1547                 btrfs_abort_transaction(trans, ret);
1548                 goto fail;
1549         }
1550
1551         btrfs_set_lock_blocking(old);
1552
1553         ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1554         /* clean up in any case */
1555         btrfs_tree_unlock(old);
1556         free_extent_buffer(old);
1557         if (ret) {
1558                 btrfs_abort_transaction(trans, ret);
1559                 goto fail;
1560         }
1561         /* see comments in should_cow_block() */
1562         set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1563         smp_wmb();
1564
1565         btrfs_set_root_node(new_root_item, tmp);
1566         /* record when the snapshot was created in key.offset */
1567         key.offset = trans->transid;
1568         ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1569         btrfs_tree_unlock(tmp);
1570         free_extent_buffer(tmp);
1571         if (ret) {
1572                 btrfs_abort_transaction(trans, ret);
1573                 goto fail;
1574         }
1575
1576         /*
1577          * insert root back/forward references
1578          */
1579         ret = btrfs_add_root_ref(trans, tree_root, objectid,
1580                                  parent_root->root_key.objectid,
1581                                  btrfs_ino(parent_inode), index,
1582                                  dentry->d_name.name, dentry->d_name.len);
1583         if (ret) {
1584                 btrfs_abort_transaction(trans, ret);
1585                 goto fail;
1586         }
1587
1588         key.offset = (u64)-1;
1589         pending->snap = btrfs_read_fs_root_no_name(root->fs_info, &key);
1590         if (IS_ERR(pending->snap)) {
1591                 ret = PTR_ERR(pending->snap);
1592                 btrfs_abort_transaction(trans, ret);
1593                 goto fail;
1594         }
1595
1596         ret = btrfs_reloc_post_snapshot(trans, pending);
1597         if (ret) {
1598                 btrfs_abort_transaction(trans, ret);
1599                 goto fail;
1600         }
1601
1602         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1603         if (ret) {
1604                 btrfs_abort_transaction(trans, ret);
1605                 goto fail;
1606         }
1607
1608         /*
1609          * Do special qgroup accounting for snapshot, as we do some qgroup
1610          * snapshot hack to do fast snapshot.
1611          * To co-operate with that hack, we do hack again.
1612          * Or snapshot will be greatly slowed down by a subtree qgroup rescan
1613          */
1614         ret = qgroup_account_snapshot(trans, root, parent_root,
1615                                       pending->inherit, objectid);
1616         if (ret < 0)
1617                 goto fail;
1618
1619         ret = btrfs_insert_dir_item(trans, parent_root,
1620                                     dentry->d_name.name, dentry->d_name.len,
1621                                     parent_inode, &key,
1622                                     BTRFS_FT_DIR, index);
1623         /* We have check then name at the beginning, so it is impossible. */
1624         BUG_ON(ret == -EEXIST || ret == -EOVERFLOW);
1625         if (ret) {
1626                 btrfs_abort_transaction(trans, ret);
1627                 goto fail;
1628         }
1629
1630         btrfs_i_size_write(parent_inode, parent_inode->i_size +
1631                                          dentry->d_name.len * 2);
1632         parent_inode->i_mtime = parent_inode->i_ctime =
1633                 current_time(parent_inode);
1634         ret = btrfs_update_inode_fallback(trans, parent_root, parent_inode);
1635         if (ret) {
1636                 btrfs_abort_transaction(trans, ret);
1637                 goto fail;
1638         }
1639         ret = btrfs_uuid_tree_add(trans, fs_info->uuid_root, new_uuid.b,
1640                                   BTRFS_UUID_KEY_SUBVOL, objectid);
1641         if (ret) {
1642                 btrfs_abort_transaction(trans, ret);
1643                 goto fail;
1644         }
1645         if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1646                 ret = btrfs_uuid_tree_add(trans, fs_info->uuid_root,
1647                                           new_root_item->received_uuid,
1648                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1649                                           objectid);
1650                 if (ret && ret != -EEXIST) {
1651                         btrfs_abort_transaction(trans, ret);
1652                         goto fail;
1653                 }
1654         }
1655
1656         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
1657         if (ret) {
1658                 btrfs_abort_transaction(trans, ret);
1659                 goto fail;
1660         }
1661
1662 fail:
1663         pending->error = ret;
1664 dir_item_existed:
1665         trans->block_rsv = rsv;
1666         trans->bytes_reserved = 0;
1667 clear_skip_qgroup:
1668         btrfs_clear_skip_qgroup(trans);
1669 no_free_objectid:
1670         kfree(new_root_item);
1671         pending->root_item = NULL;
1672         btrfs_free_path(path);
1673         pending->path = NULL;
1674
1675         return ret;
1676 }
1677
1678 /*
1679  * create all the snapshots we've scheduled for creation
1680  */
1681 static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans,
1682                                              struct btrfs_fs_info *fs_info)
1683 {
1684         struct btrfs_pending_snapshot *pending, *next;
1685         struct list_head *head = &trans->transaction->pending_snapshots;
1686         int ret = 0;
1687
1688         list_for_each_entry_safe(pending, next, head, list) {
1689                 list_del(&pending->list);
1690                 ret = create_pending_snapshot(trans, fs_info, pending);
1691                 if (ret)
1692                         break;
1693         }
1694         return ret;
1695 }
1696
1697 static void update_super_roots(struct btrfs_root *root)
1698 {
1699         struct btrfs_root_item *root_item;
1700         struct btrfs_super_block *super;
1701
1702         super = root->fs_info->super_copy;
1703
1704         root_item = &root->fs_info->chunk_root->root_item;
1705         super->chunk_root = root_item->bytenr;
1706         super->chunk_root_generation = root_item->generation;
1707         super->chunk_root_level = root_item->level;
1708
1709         root_item = &root->fs_info->tree_root->root_item;
1710         super->root = root_item->bytenr;
1711         super->generation = root_item->generation;
1712         super->root_level = root_item->level;
1713         if (btrfs_test_opt(root->fs_info, SPACE_CACHE))
1714                 super->cache_generation = root_item->generation;
1715         if (root->fs_info->update_uuid_tree_gen)
1716                 super->uuid_tree_generation = root_item->generation;
1717 }
1718
1719 int btrfs_transaction_in_commit(struct btrfs_fs_info *info)
1720 {
1721         struct btrfs_transaction *trans;
1722         int ret = 0;
1723
1724         spin_lock(&info->trans_lock);
1725         trans = info->running_transaction;
1726         if (trans)
1727                 ret = (trans->state >= TRANS_STATE_COMMIT_START);
1728         spin_unlock(&info->trans_lock);
1729         return ret;
1730 }
1731
1732 int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1733 {
1734         struct btrfs_transaction *trans;
1735         int ret = 0;
1736
1737         spin_lock(&info->trans_lock);
1738         trans = info->running_transaction;
1739         if (trans)
1740                 ret = is_transaction_blocked(trans);
1741         spin_unlock(&info->trans_lock);
1742         return ret;
1743 }
1744
1745 /*
1746  * wait for the current transaction commit to start and block subsequent
1747  * transaction joins
1748  */
1749 static void wait_current_trans_commit_start(struct btrfs_root *root,
1750                                             struct btrfs_transaction *trans)
1751 {
1752         wait_event(root->fs_info->transaction_blocked_wait,
1753                    trans->state >= TRANS_STATE_COMMIT_START ||
1754                    trans->aborted);
1755 }
1756
1757 /*
1758  * wait for the current transaction to start and then become unblocked.
1759  * caller holds ref.
1760  */
1761 static void wait_current_trans_commit_start_and_unblock(struct btrfs_root *root,
1762                                          struct btrfs_transaction *trans)
1763 {
1764         wait_event(root->fs_info->transaction_wait,
1765                    trans->state >= TRANS_STATE_UNBLOCKED ||
1766                    trans->aborted);
1767 }
1768
1769 /*
1770  * commit transactions asynchronously. once btrfs_commit_transaction_async
1771  * returns, any subsequent transaction will not be allowed to join.
1772  */
1773 struct btrfs_async_commit {
1774         struct btrfs_trans_handle *newtrans;
1775         struct btrfs_root *root;
1776         struct work_struct work;
1777 };
1778
1779 static void do_async_commit(struct work_struct *work)
1780 {
1781         struct btrfs_async_commit *ac =
1782                 container_of(work, struct btrfs_async_commit, work);
1783
1784         /*
1785          * We've got freeze protection passed with the transaction.
1786          * Tell lockdep about it.
1787          */
1788         if (ac->newtrans->type & __TRANS_FREEZABLE)
1789                 __sb_writers_acquired(ac->root->fs_info->sb, SB_FREEZE_FS);
1790
1791         current->journal_info = ac->newtrans;
1792
1793         btrfs_commit_transaction(ac->newtrans, ac->root);
1794         kfree(ac);
1795 }
1796
1797 int btrfs_commit_transaction_async(struct btrfs_trans_handle *trans,
1798                                    struct btrfs_root *root,
1799                                    int wait_for_unblock)
1800 {
1801         struct btrfs_async_commit *ac;
1802         struct btrfs_transaction *cur_trans;
1803
1804         ac = kmalloc(sizeof(*ac), GFP_NOFS);
1805         if (!ac)
1806                 return -ENOMEM;
1807
1808         INIT_WORK(&ac->work, do_async_commit);
1809         ac->root = root;
1810         ac->newtrans = btrfs_join_transaction(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         atomic_inc(&cur_trans->use_count);
1820
1821         btrfs_end_transaction(trans, root);
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(root->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(root, cur_trans);
1835         else
1836                 wait_current_trans_commit_start(root, 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,
1847                                 struct btrfs_root *root, int err)
1848 {
1849         struct btrfs_transaction *cur_trans = trans->transaction;
1850         DEFINE_WAIT(wait);
1851
1852         WARN_ON(trans->use_count > 1);
1853
1854         btrfs_abort_transaction(trans, err);
1855
1856         spin_lock(&root->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 == root->fs_info->running_transaction) {
1867                 cur_trans->state = TRANS_STATE_COMMIT_DOING;
1868                 spin_unlock(&root->fs_info->trans_lock);
1869                 wait_event(cur_trans->writer_wait,
1870                            atomic_read(&cur_trans->num_writers) == 1);
1871
1872                 spin_lock(&root->fs_info->trans_lock);
1873         }
1874         spin_unlock(&root->fs_info->trans_lock);
1875
1876         btrfs_cleanup_one_transaction(trans->transaction, root);
1877
1878         spin_lock(&root->fs_info->trans_lock);
1879         if (cur_trans == root->fs_info->running_transaction)
1880                 root->fs_info->running_transaction = NULL;
1881         spin_unlock(&root->fs_info->trans_lock);
1882
1883         if (trans->type & __TRANS_FREEZABLE)
1884                 sb_end_intwrite(root->fs_info->sb);
1885         btrfs_put_transaction(cur_trans);
1886         btrfs_put_transaction(cur_trans);
1887
1888         trace_btrfs_transaction_commit(root);
1889
1890         if (current->journal_info == trans)
1891                 current->journal_info = NULL;
1892         btrfs_scrub_cancel(root->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         if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
1900                 return btrfs_start_delalloc_roots(fs_info, 1, -1);
1901         return 0;
1902 }
1903
1904 static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
1905 {
1906         if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
1907                 btrfs_wait_ordered_roots(fs_info, -1, 0, (u64)-1);
1908 }
1909
1910 static inline void
1911 btrfs_wait_pending_ordered(struct btrfs_transaction *cur_trans)
1912 {
1913         wait_event(cur_trans->pending_wait,
1914                    atomic_read(&cur_trans->pending_ordered) == 0);
1915 }
1916
1917 int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
1918                              struct btrfs_root *root)
1919 {
1920         struct btrfs_transaction *cur_trans = trans->transaction;
1921         struct btrfs_transaction *prev_trans = NULL;
1922         struct btrfs_inode *btree_ino = BTRFS_I(root->fs_info->btree_inode);
1923         int ret;
1924
1925         /* Stop the commit early if ->aborted is set */
1926         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
1927                 ret = cur_trans->aborted;
1928                 btrfs_end_transaction(trans, root);
1929                 return ret;
1930         }
1931
1932         /* make a pass through all the delayed refs we have so far
1933          * any runnings procs may add more while we are here
1934          */
1935         ret = btrfs_run_delayed_refs(trans, root, 0);
1936         if (ret) {
1937                 btrfs_end_transaction(trans, root);
1938                 return ret;
1939         }
1940
1941         btrfs_trans_release_metadata(trans, root);
1942         trans->block_rsv = NULL;
1943
1944         cur_trans = trans->transaction;
1945
1946         /*
1947          * set the flushing flag so procs in this transaction have to
1948          * start sending their work down.
1949          */
1950         cur_trans->delayed_refs.flushing = 1;
1951         smp_wmb();
1952
1953         if (!list_empty(&trans->new_bgs))
1954                 btrfs_create_pending_block_groups(trans, root);
1955
1956         ret = btrfs_run_delayed_refs(trans, root, 0);
1957         if (ret) {
1958                 btrfs_end_transaction(trans, root);
1959                 return ret;
1960         }
1961
1962         if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) {
1963                 int run_it = 0;
1964
1965                 /* this mutex is also taken before trying to set
1966                  * block groups readonly.  We need to make sure
1967                  * that nobody has set a block group readonly
1968                  * after a extents from that block group have been
1969                  * allocated for cache files.  btrfs_set_block_group_ro
1970                  * will wait for the transaction to commit if it
1971                  * finds BTRFS_TRANS_DIRTY_BG_RUN set.
1972                  *
1973                  * The BTRFS_TRANS_DIRTY_BG_RUN flag is also used to make sure
1974                  * only one process starts all the block group IO.  It wouldn't
1975                  * hurt to have more than one go through, but there's no
1976                  * real advantage to it either.
1977                  */
1978                 mutex_lock(&root->fs_info->ro_block_group_mutex);
1979                 if (!test_and_set_bit(BTRFS_TRANS_DIRTY_BG_RUN,
1980                                       &cur_trans->flags))
1981                         run_it = 1;
1982                 mutex_unlock(&root->fs_info->ro_block_group_mutex);
1983
1984                 if (run_it)
1985                         ret = btrfs_start_dirty_block_groups(trans, root);
1986         }
1987         if (ret) {
1988                 btrfs_end_transaction(trans, root);
1989                 return ret;
1990         }
1991
1992         spin_lock(&root->fs_info->trans_lock);
1993         if (cur_trans->state >= TRANS_STATE_COMMIT_START) {
1994                 spin_unlock(&root->fs_info->trans_lock);
1995                 atomic_inc(&cur_trans->use_count);
1996                 ret = btrfs_end_transaction(trans, root);
1997
1998                 wait_for_commit(root, cur_trans);
1999
2000                 if (unlikely(cur_trans->aborted))
2001                         ret = cur_trans->aborted;
2002
2003                 btrfs_put_transaction(cur_trans);
2004
2005                 return ret;
2006         }
2007
2008         cur_trans->state = TRANS_STATE_COMMIT_START;
2009         wake_up(&root->fs_info->transaction_blocked_wait);
2010
2011         if (cur_trans->list.prev != &root->fs_info->trans_list) {
2012                 prev_trans = list_entry(cur_trans->list.prev,
2013                                         struct btrfs_transaction, list);
2014                 if (prev_trans->state != TRANS_STATE_COMPLETED) {
2015                         atomic_inc(&prev_trans->use_count);
2016                         spin_unlock(&root->fs_info->trans_lock);
2017
2018                         wait_for_commit(root, prev_trans);
2019                         ret = prev_trans->aborted;
2020
2021                         btrfs_put_transaction(prev_trans);
2022                         if (ret)
2023                                 goto cleanup_transaction;
2024                 } else {
2025                         spin_unlock(&root->fs_info->trans_lock);
2026                 }
2027         } else {
2028                 spin_unlock(&root->fs_info->trans_lock);
2029         }
2030
2031         extwriter_counter_dec(cur_trans, trans->type);
2032
2033         ret = btrfs_start_delalloc_flush(root->fs_info);
2034         if (ret)
2035                 goto cleanup_transaction;
2036
2037         ret = btrfs_run_delayed_items(trans, root);
2038         if (ret)
2039                 goto cleanup_transaction;
2040
2041         wait_event(cur_trans->writer_wait,
2042                    extwriter_counter_read(cur_trans) == 0);
2043
2044         /* some pending stuffs might be added after the previous flush. */
2045         ret = btrfs_run_delayed_items(trans, root);
2046         if (ret)
2047                 goto cleanup_transaction;
2048
2049         btrfs_wait_delalloc_flush(root->fs_info);
2050
2051         btrfs_wait_pending_ordered(cur_trans);
2052
2053         btrfs_scrub_pause(root);
2054         /*
2055          * Ok now we need to make sure to block out any other joins while we
2056          * commit the transaction.  We could have started a join before setting
2057          * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
2058          */
2059         spin_lock(&root->fs_info->trans_lock);
2060         cur_trans->state = TRANS_STATE_COMMIT_DOING;
2061         spin_unlock(&root->fs_info->trans_lock);
2062         wait_event(cur_trans->writer_wait,
2063                    atomic_read(&cur_trans->num_writers) == 1);
2064
2065         /* ->aborted might be set after the previous check, so check it */
2066         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
2067                 ret = cur_trans->aborted;
2068                 goto scrub_continue;
2069         }
2070         /*
2071          * the reloc mutex makes sure that we stop
2072          * the balancing code from coming in and moving
2073          * extents around in the middle of the commit
2074          */
2075         mutex_lock(&root->fs_info->reloc_mutex);
2076
2077         /*
2078          * We needn't worry about the delayed items because we will
2079          * deal with them in create_pending_snapshot(), which is the
2080          * core function of the snapshot creation.
2081          */
2082         ret = create_pending_snapshots(trans, root->fs_info);
2083         if (ret) {
2084                 mutex_unlock(&root->fs_info->reloc_mutex);
2085                 goto scrub_continue;
2086         }
2087
2088         /*
2089          * We insert the dir indexes of the snapshots and update the inode
2090          * of the snapshots' parents after the snapshot creation, so there
2091          * are some delayed items which are not dealt with. Now deal with
2092          * them.
2093          *
2094          * We needn't worry that this operation will corrupt the snapshots,
2095          * because all the tree which are snapshoted will be forced to COW
2096          * the nodes and leaves.
2097          */
2098         ret = btrfs_run_delayed_items(trans, root);
2099         if (ret) {
2100                 mutex_unlock(&root->fs_info->reloc_mutex);
2101                 goto scrub_continue;
2102         }
2103
2104         ret = btrfs_run_delayed_refs(trans, root, (unsigned long)-1);
2105         if (ret) {
2106                 mutex_unlock(&root->fs_info->reloc_mutex);
2107                 goto scrub_continue;
2108         }
2109
2110         /* Reocrd old roots for later qgroup accounting */
2111         ret = btrfs_qgroup_prepare_account_extents(trans, root->fs_info);
2112         if (ret) {
2113                 mutex_unlock(&root->fs_info->reloc_mutex);
2114                 goto scrub_continue;
2115         }
2116
2117         /*
2118          * make sure none of the code above managed to slip in a
2119          * delayed item
2120          */
2121         btrfs_assert_delayed_root_empty(root);
2122
2123         WARN_ON(cur_trans != trans->transaction);
2124
2125         /* btrfs_commit_tree_roots is responsible for getting the
2126          * various roots consistent with each other.  Every pointer
2127          * in the tree of tree roots has to point to the most up to date
2128          * root for every subvolume and other tree.  So, we have to keep
2129          * the tree logging code from jumping in and changing any
2130          * of the trees.
2131          *
2132          * At this point in the commit, there can't be any tree-log
2133          * writers, but a little lower down we drop the trans mutex
2134          * and let new people in.  By holding the tree_log_mutex
2135          * from now until after the super is written, we avoid races
2136          * with the tree-log code.
2137          */
2138         mutex_lock(&root->fs_info->tree_log_mutex);
2139
2140         ret = commit_fs_roots(trans, root);
2141         if (ret) {
2142                 mutex_unlock(&root->fs_info->tree_log_mutex);
2143                 mutex_unlock(&root->fs_info->reloc_mutex);
2144                 goto scrub_continue;
2145         }
2146
2147         /*
2148          * Since the transaction is done, we can apply the pending changes
2149          * before the next transaction.
2150          */
2151         btrfs_apply_pending_changes(root->fs_info);
2152
2153         /* commit_fs_roots gets rid of all the tree log roots, it is now
2154          * safe to free the root of tree log roots
2155          */
2156         btrfs_free_log_root_tree(trans, root->fs_info);
2157
2158         /*
2159          * Since fs roots are all committed, we can get a quite accurate
2160          * new_roots. So let's do quota accounting.
2161          */
2162         ret = btrfs_qgroup_account_extents(trans, root->fs_info);
2163         if (ret < 0) {
2164                 mutex_unlock(&root->fs_info->tree_log_mutex);
2165                 mutex_unlock(&root->fs_info->reloc_mutex);
2166                 goto scrub_continue;
2167         }
2168
2169         ret = commit_cowonly_roots(trans, root);
2170         if (ret) {
2171                 mutex_unlock(&root->fs_info->tree_log_mutex);
2172                 mutex_unlock(&root->fs_info->reloc_mutex);
2173                 goto scrub_continue;
2174         }
2175
2176         /*
2177          * The tasks which save the space cache and inode cache may also
2178          * update ->aborted, check it.
2179          */
2180         if (unlikely(ACCESS_ONCE(cur_trans->aborted))) {
2181                 ret = cur_trans->aborted;
2182                 mutex_unlock(&root->fs_info->tree_log_mutex);
2183                 mutex_unlock(&root->fs_info->reloc_mutex);
2184                 goto scrub_continue;
2185         }
2186
2187         btrfs_prepare_extent_commit(trans, root);
2188
2189         cur_trans = root->fs_info->running_transaction;
2190
2191         btrfs_set_root_node(&root->fs_info->tree_root->root_item,
2192                             root->fs_info->tree_root->node);
2193         list_add_tail(&root->fs_info->tree_root->dirty_list,
2194                       &cur_trans->switch_commits);
2195
2196         btrfs_set_root_node(&root->fs_info->chunk_root->root_item,
2197                             root->fs_info->chunk_root->node);
2198         list_add_tail(&root->fs_info->chunk_root->dirty_list,
2199                       &cur_trans->switch_commits);
2200
2201         switch_commit_roots(cur_trans, root->fs_info);
2202
2203         assert_qgroups_uptodate(trans);
2204         ASSERT(list_empty(&cur_trans->dirty_bgs));
2205         ASSERT(list_empty(&cur_trans->io_bgs));
2206         update_super_roots(root);
2207
2208         btrfs_set_super_log_root(root->fs_info->super_copy, 0);
2209         btrfs_set_super_log_root_level(root->fs_info->super_copy, 0);
2210         memcpy(root->fs_info->super_for_commit, root->fs_info->super_copy,
2211                sizeof(*root->fs_info->super_copy));
2212
2213         btrfs_update_commit_device_size(root->fs_info);
2214         btrfs_update_commit_device_bytes_used(root, cur_trans);
2215
2216         clear_bit(BTRFS_INODE_BTREE_LOG1_ERR, &btree_ino->runtime_flags);
2217         clear_bit(BTRFS_INODE_BTREE_LOG2_ERR, &btree_ino->runtime_flags);
2218
2219         btrfs_trans_release_chunk_metadata(trans);
2220
2221         spin_lock(&root->fs_info->trans_lock);
2222         cur_trans->state = TRANS_STATE_UNBLOCKED;
2223         root->fs_info->running_transaction = NULL;
2224         spin_unlock(&root->fs_info->trans_lock);
2225         mutex_unlock(&root->fs_info->reloc_mutex);
2226
2227         wake_up(&root->fs_info->transaction_wait);
2228
2229         ret = btrfs_write_and_wait_transaction(trans, root);
2230         if (ret) {
2231                 btrfs_handle_fs_error(root->fs_info, ret,
2232                             "Error while writing out transaction");
2233                 mutex_unlock(&root->fs_info->tree_log_mutex);
2234                 goto scrub_continue;
2235         }
2236
2237         ret = write_ctree_super(trans, root, 0);
2238         if (ret) {
2239                 mutex_unlock(&root->fs_info->tree_log_mutex);
2240                 goto scrub_continue;
2241         }
2242
2243         /*
2244          * the super is written, we can safely allow the tree-loggers
2245          * to go about their business
2246          */
2247         mutex_unlock(&root->fs_info->tree_log_mutex);
2248
2249         btrfs_finish_extent_commit(trans, root);
2250
2251         if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &cur_trans->flags))
2252                 btrfs_clear_space_info_full(root->fs_info);
2253
2254         root->fs_info->last_trans_committed = cur_trans->transid;
2255         /*
2256          * We needn't acquire the lock here because there is no other task
2257          * which can change it.
2258          */
2259         cur_trans->state = TRANS_STATE_COMPLETED;
2260         wake_up(&cur_trans->commit_wait);
2261
2262         spin_lock(&root->fs_info->trans_lock);
2263         list_del_init(&cur_trans->list);
2264         spin_unlock(&root->fs_info->trans_lock);
2265
2266         btrfs_put_transaction(cur_trans);
2267         btrfs_put_transaction(cur_trans);
2268
2269         if (trans->type & __TRANS_FREEZABLE)
2270                 sb_end_intwrite(root->fs_info->sb);
2271
2272         trace_btrfs_transaction_commit(root);
2273
2274         btrfs_scrub_continue(root);
2275
2276         if (current->journal_info == trans)
2277                 current->journal_info = NULL;
2278
2279         kmem_cache_free(btrfs_trans_handle_cachep, trans);
2280
2281         /*
2282          * If fs has been frozen, we can not handle delayed iputs, otherwise
2283          * it'll result in deadlock about SB_FREEZE_FS.
2284          */
2285         if (current != root->fs_info->transaction_kthread &&
2286             current != root->fs_info->cleaner_kthread &&
2287             !root->fs_info->fs_frozen)
2288                 btrfs_run_delayed_iputs(root);
2289
2290         return ret;
2291
2292 scrub_continue:
2293         btrfs_scrub_continue(root);
2294 cleanup_transaction:
2295         btrfs_trans_release_metadata(trans, root);
2296         btrfs_trans_release_chunk_metadata(trans);
2297         trans->block_rsv = NULL;
2298         btrfs_warn(root->fs_info, "Skipping commit of aborted transaction.");
2299         if (current->journal_info == trans)
2300                 current->journal_info = NULL;
2301         cleanup_transaction(trans, root, ret);
2302
2303         return ret;
2304 }
2305
2306 /*
2307  * return < 0 if error
2308  * 0 if there are no more dead_roots at the time of call
2309  * 1 there are more to be processed, call me again
2310  *
2311  * The return value indicates there are certainly more snapshots to delete, but
2312  * if there comes a new one during processing, it may return 0. We don't mind,
2313  * because btrfs_commit_super will poke cleaner thread and it will process it a
2314  * few seconds later.
2315  */
2316 int btrfs_clean_one_deleted_snapshot(struct btrfs_root *root)
2317 {
2318         int ret;
2319         struct btrfs_fs_info *fs_info = root->fs_info;
2320
2321         spin_lock(&fs_info->trans_lock);
2322         if (list_empty(&fs_info->dead_roots)) {
2323                 spin_unlock(&fs_info->trans_lock);
2324                 return 0;
2325         }
2326         root = list_first_entry(&fs_info->dead_roots,
2327                         struct btrfs_root, root_list);
2328         list_del_init(&root->root_list);
2329         spin_unlock(&fs_info->trans_lock);
2330
2331         pr_debug("BTRFS: cleaner removing %llu\n", root->objectid);
2332
2333         btrfs_kill_all_delayed_nodes(root);
2334
2335         if (btrfs_header_backref_rev(root->node) <
2336                         BTRFS_MIXED_BACKREF_REV)
2337                 ret = btrfs_drop_snapshot(root, NULL, 0, 0);
2338         else
2339                 ret = btrfs_drop_snapshot(root, NULL, 1, 0);
2340
2341         return (ret < 0) ? 0 : 1;
2342 }
2343
2344 void btrfs_apply_pending_changes(struct btrfs_fs_info *fs_info)
2345 {
2346         unsigned long prev;
2347         unsigned long bit;
2348
2349         prev = xchg(&fs_info->pending_changes, 0);
2350         if (!prev)
2351                 return;
2352
2353         bit = 1 << BTRFS_PENDING_SET_INODE_MAP_CACHE;
2354         if (prev & bit)
2355                 btrfs_set_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2356         prev &= ~bit;
2357
2358         bit = 1 << BTRFS_PENDING_CLEAR_INODE_MAP_CACHE;
2359         if (prev & bit)
2360                 btrfs_clear_opt(fs_info->mount_opt, INODE_MAP_CACHE);
2361         prev &= ~bit;
2362
2363         bit = 1 << BTRFS_PENDING_COMMIT;
2364         if (prev & bit)
2365                 btrfs_debug(fs_info, "pending commit done");
2366         prev &= ~bit;
2367
2368         if (prev)
2369                 btrfs_warn(fs_info,
2370                         "unknown pending changes left 0x%lx, ignoring", prev);
2371 }