fs/btrfs/transaction.c

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