fs/btrfs/transaction.c

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