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