fs/btrfs/transaction.c

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