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/sched/mm.h>
  10#include <linux/writeback.h>
  11#include <linux/pagemap.h>
  12#include <linux/blkdev.h>
  13#include <linux/uuid.h>
  14#include <linux/timekeeping.h>
  15#include "misc.h"
  16#include "ctree.h"
  17#include "disk-io.h"
  18#include "transaction.h"
  19#include "locking.h"
  20#include "tree-log.h"
  21#include "volumes.h"
  22#include "dev-replace.h"
  23#include "qgroup.h"
  24#include "block-group.h"
  25#include "space-info.h"
  26#include "fs.h"
  27#include "accessors.h"
  28#include "extent-tree.h"
  29#include "root-tree.h"
  30#include "dir-item.h"
  31#include "uuid-tree.h"
  32#include "ioctl.h"
  33#include "relocation.h"
  34#include "scrub.h"
  35#include "ordered-data.h"
  36#include "delayed-inode.h"
  37
  38static struct kmem_cache *btrfs_trans_handle_cachep;
  39
  40/*
  41 * Transaction states and transitions
  42 *
  43 * No running transaction (fs tree blocks are not modified)
  44 * |
  45 * | To next stage:
  46 * |  Call start_transaction() variants. Except btrfs_join_transaction_nostart().
  47 * V
  48 * Transaction N [[TRANS_STATE_RUNNING]]
  49 * |
  50 * | New trans handles can be attached to transaction N by calling all
  51 * | start_transaction() variants.
  52 * |
  53 * | To next stage:
  54 * |  Call btrfs_commit_transaction() on any trans handle attached to
  55 * |  transaction N
  56 * V
  57 * Transaction N [[TRANS_STATE_COMMIT_PREP]]
  58 * |
  59 * | If there are simultaneous calls to btrfs_commit_transaction() one will win
  60 * | the race and the rest will wait for the winner to commit the transaction.
  61 * |
  62 * | The winner will wait for previous running transaction to completely finish
  63 * | if there is one.
  64 * |
  65 * Transaction N [[TRANS_STATE_COMMIT_START]]
  66 * |
  67 * | Then one of the following happens:
  68 * | - Wait for all other trans handle holders to release.
  69 * |   The btrfs_commit_transaction() caller will do the commit work.
  70 * | - Wait for current transaction to be committed by others.
  71 * |   Other btrfs_commit_transaction() caller will do the commit work.
  72 * |
  73 * | At this stage, only btrfs_join_transaction*() variants can attach
  74 * | to this running transaction.
  75 * | All other variants will wait for current one to finish and attach to
  76 * | transaction N+1.
  77 * |
  78 * | To next stage:
  79 * |  Caller is chosen to commit transaction N, and all other trans handle
  80 * |  haven been released.
  81 * V
  82 * Transaction N [[TRANS_STATE_COMMIT_DOING]]
  83 * |
  84 * | The heavy lifting transaction work is started.
  85 * | From running delayed refs (modifying extent tree) to creating pending
  86 * | snapshots, running qgroups.
  87 * | In short, modify supporting trees to reflect modifications of subvolume
  88 * | trees.
  89 * |
  90 * | At this stage, all start_transaction() calls will wait for this
  91 * | transaction to finish and attach to transaction N+1.
  92 * |
  93 * | To next stage:
  94 * |  Until all supporting trees are updated.
  95 * V
  96 * Transaction N [[TRANS_STATE_UNBLOCKED]]
  97 * |						    Transaction N+1
  98 * | All needed trees are modified, thus we only    [[TRANS_STATE_RUNNING]]
  99 * | need to write them back to disk and update	    |
 100 * | super blocks.				    |
 101 * |						    |
 102 * | At this stage, new transaction is allowed to   |
 103 * | start.					    |
 104 * | All new start_transaction() calls will be	    |
 105 * | attached to transid N+1.			    |
 106 * |						    |
 107 * | To next stage:				    |
 108 * |  Until all tree blocks and super blocks are    |
 109 * |  written to block devices			    |
 110 * V						    |
 111 * Transaction N [[TRANS_STATE_COMPLETED]]	    V
 112 *   All tree blocks and super blocks are written.  Transaction N+1
 113 *   This transaction is finished and all its	    [[TRANS_STATE_COMMIT_START]]
 114 *   data structures will be cleaned up.	    | Life goes on
 115 */
 116static const unsigned int btrfs_blocked_trans_types[TRANS_STATE_MAX] = {
 117	[TRANS_STATE_RUNNING]		= 0U,
 118	[TRANS_STATE_COMMIT_PREP]	= 0U,
 119	[TRANS_STATE_COMMIT_START]	= (__TRANS_START | __TRANS_ATTACH),
 120	[TRANS_STATE_COMMIT_DOING]	= (__TRANS_START |
 121					   __TRANS_ATTACH |
 122					   __TRANS_JOIN |
 123					   __TRANS_JOIN_NOSTART),
 124	[TRANS_STATE_UNBLOCKED]		= (__TRANS_START |
 125					   __TRANS_ATTACH |
 126					   __TRANS_JOIN |
 127					   __TRANS_JOIN_NOLOCK |
 128					   __TRANS_JOIN_NOSTART),
 129	[TRANS_STATE_SUPER_COMMITTED]	= (__TRANS_START |
 130					   __TRANS_ATTACH |
 131					   __TRANS_JOIN |
 132					   __TRANS_JOIN_NOLOCK |
 133					   __TRANS_JOIN_NOSTART),
 134	[TRANS_STATE_COMPLETED]		= (__TRANS_START |
 135					   __TRANS_ATTACH |
 136					   __TRANS_JOIN |
 137					   __TRANS_JOIN_NOLOCK |
 138					   __TRANS_JOIN_NOSTART),
 139};
 140
 141void btrfs_put_transaction(struct btrfs_transaction *transaction)
 142{
 143	if (refcount_dec_and_test(&transaction->use_count)) {
 144		BUG_ON(!list_empty(&transaction->list));
 145		WARN_ON(!xa_empty(&transaction->delayed_refs.head_refs));
 146		WARN_ON(!xa_empty(&transaction->delayed_refs.dirty_extents));
 147		if (transaction->delayed_refs.pending_csums)
 148			btrfs_err(transaction->fs_info,
 149				  "pending csums is %llu",
 150				  transaction->delayed_refs.pending_csums);
 151		/*
 152		 * If any block groups are found in ->deleted_bgs then it's
 153		 * because the transaction was aborted and a commit did not
 154		 * happen (things failed before writing the new superblock
 155		 * and calling btrfs_finish_extent_commit()), so we can not
 156		 * discard the physical locations of the block groups.
 157		 */
 158		while (!list_empty(&transaction->deleted_bgs)) {
 159			struct btrfs_block_group *cache;
 160
 161			cache = list_first_entry(&transaction->deleted_bgs,
 162						 struct btrfs_block_group,
 163						 bg_list);
 164			/*
 165			 * Not strictly necessary to lock, as no other task will be using a
 166			 * block_group on the deleted_bgs list during a transaction abort.
 167			 */
 168			spin_lock(&transaction->fs_info->unused_bgs_lock);
 169			list_del_init(&cache->bg_list);
 170			spin_unlock(&transaction->fs_info->unused_bgs_lock);
 171			btrfs_unfreeze_block_group(cache);
 172			btrfs_put_block_group(cache);
 173		}
 174		WARN_ON(!list_empty(&transaction->dev_update_list));
 175		kfree(transaction);
 176	}
 177}
 178
 179static noinline void switch_commit_roots(struct btrfs_trans_handle *trans)
 180{
 181	struct btrfs_transaction *cur_trans = trans->transaction;
 182	struct btrfs_fs_info *fs_info = trans->fs_info;
 183	struct btrfs_root *root, *tmp;
 184
 185	/*
 186	 * At this point no one can be using this transaction to modify any tree
 187	 * and no one can start another transaction to modify any tree either.
 188	 */
 189	ASSERT(cur_trans->state == TRANS_STATE_COMMIT_DOING,
 190	       "cur_trans->state=%d", cur_trans->state);
 191
 192	down_write(&fs_info->commit_root_sem);
 193
 194	if (test_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags))
 195		fs_info->last_reloc_trans = trans->transid;
 196
 197	list_for_each_entry_safe(root, tmp, &cur_trans->switch_commits,
 198				 dirty_list) {
 199		list_del_init(&root->dirty_list);
 200		free_extent_buffer(root->commit_root);
 201		root->commit_root = btrfs_root_node(root);
 202		btrfs_extent_io_tree_release(&root->dirty_log_pages);
 203		btrfs_qgroup_clean_swapped_blocks(root);
 204	}
 205
 206	/* We can free old roots now. */
 207	spin_lock(&cur_trans->dropped_roots_lock);
 208	while (!list_empty(&cur_trans->dropped_roots)) {
 209		root = list_first_entry(&cur_trans->dropped_roots,
 210					struct btrfs_root, root_list);
 211		list_del_init(&root->root_list);
 212		spin_unlock(&cur_trans->dropped_roots_lock);
 213		btrfs_free_log(trans, root);
 214		btrfs_drop_and_free_fs_root(fs_info, root);
 215		spin_lock(&cur_trans->dropped_roots_lock);
 216	}
 217	spin_unlock(&cur_trans->dropped_roots_lock);
 218
 219	up_write(&fs_info->commit_root_sem);
 220}
 221
 222static inline void extwriter_counter_inc(struct btrfs_transaction *trans,
 223					 unsigned int type)
 224{
 225	if (type & TRANS_EXTWRITERS)
 226		atomic_inc(&trans->num_extwriters);
 227}
 228
 229static inline void extwriter_counter_dec(struct btrfs_transaction *trans,
 230					 unsigned int type)
 231{
 232	if (type & TRANS_EXTWRITERS)
 233		atomic_dec(&trans->num_extwriters);
 234}
 235
 236static inline void extwriter_counter_init(struct btrfs_transaction *trans,
 237					  unsigned int type)
 238{
 239	atomic_set(&trans->num_extwriters, ((type & TRANS_EXTWRITERS) ? 1 : 0));
 240}
 241
 242static inline int extwriter_counter_read(struct btrfs_transaction *trans)
 243{
 244	return atomic_read(&trans->num_extwriters);
 245}
 246
 247/*
 248 * To be called after doing the chunk btree updates right after allocating a new
 249 * chunk (after btrfs_chunk_alloc_add_chunk_item() is called), when removing a
 250 * chunk after all chunk btree updates and after finishing the second phase of
 251 * chunk allocation (btrfs_create_pending_block_groups()) in case some block
 252 * group had its chunk item insertion delayed to the second phase.
 253 */
 254void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
 255{
 256	struct btrfs_fs_info *fs_info = trans->fs_info;
 257
 258	if (!trans->chunk_bytes_reserved)
 259		return;
 260
 261	btrfs_block_rsv_release(fs_info, &fs_info->chunk_block_rsv,
 262				trans->chunk_bytes_reserved, NULL);
 263	trans->chunk_bytes_reserved = 0;
 264}
 265
 266/*
 267 * either allocate a new transaction or hop into the existing one
 268 */
 269static noinline int join_transaction(struct btrfs_fs_info *fs_info,
 270				     unsigned int type)
 271{
 272	struct btrfs_transaction *cur_trans;
 273
 274	spin_lock(&fs_info->trans_lock);
 275loop:
 276	/* The file system has been taken offline. No new transactions. */
 277	if (BTRFS_FS_ERROR(fs_info)) {
 278		spin_unlock(&fs_info->trans_lock);
 279		return -EROFS;
 280	}
 281
 282	cur_trans = fs_info->running_transaction;
 283	if (cur_trans) {
 284		if (TRANS_ABORTED(cur_trans)) {
 285			const int abort_error = cur_trans->aborted;
 286
 287			spin_unlock(&fs_info->trans_lock);
 288			return abort_error;
 289		}
 290		if (btrfs_blocked_trans_types[cur_trans->state] & type) {
 291			spin_unlock(&fs_info->trans_lock);
 292			return -EBUSY;
 293		}
 294		refcount_inc(&cur_trans->use_count);
 295		atomic_inc(&cur_trans->num_writers);
 296		extwriter_counter_inc(cur_trans, type);
 297		spin_unlock(&fs_info->trans_lock);
 298		btrfs_lockdep_acquire(fs_info, btrfs_trans_num_writers);
 299		btrfs_lockdep_acquire(fs_info, btrfs_trans_num_extwriters);
 300		return 0;
 301	}
 302	spin_unlock(&fs_info->trans_lock);
 303
 304	/*
 305	 * If we are ATTACH or TRANS_JOIN_NOSTART, we just want to catch the
 306	 * current transaction, and commit it. If there is no transaction, just
 307	 * return ENOENT.
 308	 */
 309	if (type == TRANS_ATTACH || type == TRANS_JOIN_NOSTART)
 310		return -ENOENT;
 311
 312	/*
 313	 * JOIN_NOLOCK only happens during the transaction commit, so
 314	 * it is impossible that ->running_transaction is NULL
 315	 */
 316	BUG_ON(type == TRANS_JOIN_NOLOCK);
 317
 318	cur_trans = kmalloc(sizeof(*cur_trans), GFP_NOFS);
 319	if (!cur_trans)
 320		return -ENOMEM;
 321
 322	btrfs_lockdep_acquire(fs_info, btrfs_trans_num_writers);
 323	btrfs_lockdep_acquire(fs_info, btrfs_trans_num_extwriters);
 324
 325	spin_lock(&fs_info->trans_lock);
 326	if (fs_info->running_transaction) {
 327		/*
 328		 * someone started a transaction after we unlocked.  Make sure
 329		 * to redo the checks above
 330		 */
 331		btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters);
 332		btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
 333		kfree(cur_trans);
 334		goto loop;
 335	} else if (BTRFS_FS_ERROR(fs_info)) {
 336		spin_unlock(&fs_info->trans_lock);
 337		btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters);
 338		btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
 339		kfree(cur_trans);
 340		return -EROFS;
 341	}
 342
 343	cur_trans->fs_info = fs_info;
 344	atomic_set(&cur_trans->pending_ordered, 0);
 345	init_waitqueue_head(&cur_trans->pending_wait);
 346	atomic_set(&cur_trans->num_writers, 1);
 347	extwriter_counter_init(cur_trans, type);
 348	init_waitqueue_head(&cur_trans->writer_wait);
 349	init_waitqueue_head(&cur_trans->commit_wait);
 350	cur_trans->state = TRANS_STATE_RUNNING;
 351	/*
 352	 * One for this trans handle, one so it will live on until we
 353	 * commit the transaction.
 354	 */
 355	refcount_set(&cur_trans->use_count, 2);
 356	cur_trans->flags = 0;
 357	cur_trans->start_time = ktime_get_seconds();
 358
 359	memset(&cur_trans->delayed_refs, 0, sizeof(cur_trans->delayed_refs));
 360
 361	xa_init(&cur_trans->delayed_refs.head_refs);
 362	xa_init(&cur_trans->delayed_refs.dirty_extents);
 363
 364	/*
 365	 * although the tree mod log is per file system and not per transaction,
 366	 * the log must never go across transaction boundaries.
 367	 */
 368	smp_mb();
 369	if (!list_empty(&fs_info->tree_mod_seq_list))
 370		WARN(1, KERN_ERR "BTRFS: tree_mod_seq_list not empty when creating a fresh transaction\n");
 371	if (!RB_EMPTY_ROOT(&fs_info->tree_mod_log))
 372		WARN(1, KERN_ERR "BTRFS: tree_mod_log rb tree not empty when creating a fresh transaction\n");
 373	atomic64_set(&fs_info->tree_mod_seq, 0);
 374
 375	spin_lock_init(&cur_trans->delayed_refs.lock);
 376
 377	INIT_LIST_HEAD(&cur_trans->pending_snapshots);
 378	INIT_LIST_HEAD(&cur_trans->dev_update_list);
 379	INIT_LIST_HEAD(&cur_trans->switch_commits);
 380	INIT_LIST_HEAD(&cur_trans->dirty_bgs);
 381	INIT_LIST_HEAD(&cur_trans->io_bgs);
 382	INIT_LIST_HEAD(&cur_trans->dropped_roots);
 383	mutex_init(&cur_trans->cache_write_mutex);
 384	spin_lock_init(&cur_trans->dirty_bgs_lock);
 385	INIT_LIST_HEAD(&cur_trans->deleted_bgs);
 386	spin_lock_init(&cur_trans->dropped_roots_lock);
 387	list_add_tail(&cur_trans->list, &fs_info->trans_list);
 388	btrfs_extent_io_tree_init(fs_info, &cur_trans->dirty_pages,
 389				  IO_TREE_TRANS_DIRTY_PAGES);
 390	btrfs_extent_io_tree_init(fs_info, &cur_trans->pinned_extents,
 391				  IO_TREE_FS_PINNED_EXTENTS);
 392	btrfs_set_fs_generation(fs_info, fs_info->generation + 1);
 393	cur_trans->transid = fs_info->generation;
 394	fs_info->running_transaction = cur_trans;
 395	cur_trans->aborted = 0;
 396	spin_unlock(&fs_info->trans_lock);
 397
 398	return 0;
 399}
 400
 401/*
 402 * This does all the record keeping required to make sure that a shareable root
 403 * is properly recorded in a given transaction.  This is required to make sure
 404 * the old root from before we joined the transaction is deleted when the
 405 * transaction commits.
 406 */
 407static int record_root_in_trans(struct btrfs_trans_handle *trans,
 408			       struct btrfs_root *root,
 409			       bool force)
 410{
 411	struct btrfs_fs_info *fs_info = root->fs_info;
 412	int ret = 0;
 413
 414	if ((test_bit(BTRFS_ROOT_SHAREABLE, &root->state) &&
 415	    btrfs_get_root_last_trans(root) < trans->transid) || force) {
 416		WARN_ON(!force && root->commit_root != root->node);
 417
 418		/*
 419		 * see below for IN_TRANS_SETUP usage rules
 420		 * we have the reloc mutex held now, so there
 421		 * is only one writer in this function
 422		 */
 423		set_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
 424
 425		/* make sure readers find IN_TRANS_SETUP before
 426		 * they find our root->last_trans update
 427		 */
 428		smp_wmb();
 429
 430		spin_lock(&fs_info->fs_roots_radix_lock);
 431		if (btrfs_get_root_last_trans(root) == trans->transid && !force) {
 432			spin_unlock(&fs_info->fs_roots_radix_lock);
 433			return 0;
 434		}
 435		radix_tree_tag_set(&fs_info->fs_roots_radix,
 436				   (unsigned long)btrfs_root_id(root),
 437				   BTRFS_ROOT_TRANS_TAG);
 438		spin_unlock(&fs_info->fs_roots_radix_lock);
 439		btrfs_set_root_last_trans(root, trans->transid);
 440
 441		/* this is pretty tricky.  We don't want to
 442		 * take the relocation lock in btrfs_record_root_in_trans
 443		 * unless we're really doing the first setup for this root in
 444		 * this transaction.
 445		 *
 446		 * Normally we'd use root->last_trans as a flag to decide
 447		 * if we want to take the expensive mutex.
 448		 *
 449		 * But, we have to set root->last_trans before we
 450		 * init the relocation root, otherwise, we trip over warnings
 451		 * in ctree.c.  The solution used here is to flag ourselves
 452		 * with root IN_TRANS_SETUP.  When this is 1, we're still
 453		 * fixing up the reloc trees and everyone must wait.
 454		 *
 455		 * When this is zero, they can trust root->last_trans and fly
 456		 * through btrfs_record_root_in_trans without having to take the
 457		 * lock.  smp_wmb() makes sure that all the writes above are
 458		 * done before we pop in the zero below
 459		 */
 460		ret = btrfs_init_reloc_root(trans, root);
 461		smp_mb__before_atomic();
 462		clear_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state);
 463	}
 464	return ret;
 465}
 466
 467
 468void btrfs_add_dropped_root(struct btrfs_trans_handle *trans,
 469			    struct btrfs_root *root)
 470{
 471	struct btrfs_fs_info *fs_info = root->fs_info;
 472	struct btrfs_transaction *cur_trans = trans->transaction;
 473
 474	/* Add ourselves to the transaction dropped list */
 475	spin_lock(&cur_trans->dropped_roots_lock);
 476	list_add_tail(&root->root_list, &cur_trans->dropped_roots);
 477	spin_unlock(&cur_trans->dropped_roots_lock);
 478
 479	/* Make sure we don't try to update the root at commit time */
 480	spin_lock(&fs_info->fs_roots_radix_lock);
 481	radix_tree_tag_clear(&fs_info->fs_roots_radix,
 482			     (unsigned long)btrfs_root_id(root),
 483			     BTRFS_ROOT_TRANS_TAG);
 484	spin_unlock(&fs_info->fs_roots_radix_lock);
 485}
 486
 487int btrfs_record_root_in_trans(struct btrfs_trans_handle *trans,
 488			       struct btrfs_root *root)
 489{
 490	struct btrfs_fs_info *fs_info = root->fs_info;
 491	int ret;
 492
 493	if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
 494		return 0;
 495
 496	/*
 497	 * see record_root_in_trans for comments about IN_TRANS_SETUP usage
 498	 * and barriers
 499	 */
 500	smp_rmb();
 501	if (btrfs_get_root_last_trans(root) == trans->transid &&
 502	    !test_bit(BTRFS_ROOT_IN_TRANS_SETUP, &root->state))
 503		return 0;
 504
 505	mutex_lock(&fs_info->reloc_mutex);
 506	ret = record_root_in_trans(trans, root, 0);
 507	mutex_unlock(&fs_info->reloc_mutex);
 508
 509	return ret;
 510}
 511
 512static inline int is_transaction_blocked(struct btrfs_transaction *trans)
 513{
 514	return (trans->state >= TRANS_STATE_COMMIT_START &&
 515		trans->state < TRANS_STATE_UNBLOCKED &&
 516		!TRANS_ABORTED(trans));
 517}
 518
 519/* wait for commit against the current transaction to become unblocked
 520 * when this is done, it is safe to start a new transaction, but the current
 521 * transaction might not be fully on disk.
 522 */
 523static void wait_current_trans(struct btrfs_fs_info *fs_info, unsigned int type)
 524{
 525	struct btrfs_transaction *cur_trans;
 526
 527	spin_lock(&fs_info->trans_lock);
 528	cur_trans = fs_info->running_transaction;
 529	if (cur_trans && is_transaction_blocked(cur_trans) &&
 530	    (btrfs_blocked_trans_types[cur_trans->state] & type)) {
 531		refcount_inc(&cur_trans->use_count);
 532		spin_unlock(&fs_info->trans_lock);
 533
 534		btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
 535		wait_event(fs_info->transaction_wait,
 536			   cur_trans->state >= TRANS_STATE_UNBLOCKED ||
 537			   TRANS_ABORTED(cur_trans));
 538		btrfs_put_transaction(cur_trans);
 539	} else {
 540		spin_unlock(&fs_info->trans_lock);
 541	}
 542}
 543
 544static bool may_wait_transaction(struct btrfs_fs_info *fs_info, int type)
 545{
 546	if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
 547		return false;
 548
 549	if (type == TRANS_START)
 550		return true;
 551
 552	return false;
 553}
 554
 555static inline bool need_reserve_reloc_root(struct btrfs_root *root)
 556{
 557	struct btrfs_fs_info *fs_info = root->fs_info;
 558
 559	if (!fs_info->reloc_ctl ||
 560	    !test_bit(BTRFS_ROOT_SHAREABLE, &root->state) ||
 561	    btrfs_root_id(root) == BTRFS_TREE_RELOC_OBJECTID ||
 562	    root->reloc_root)
 563		return false;
 564
 565	return true;
 566}
 567
 568static int btrfs_reserve_trans_metadata(struct btrfs_fs_info *fs_info,
 569					enum btrfs_reserve_flush_enum flush,
 570					u64 num_bytes,
 571					u64 *delayed_refs_bytes)
 572{
 573	struct btrfs_space_info *si = fs_info->trans_block_rsv.space_info;
 574	u64 bytes = num_bytes + *delayed_refs_bytes;
 575	int ret;
 576
 577	/*
 578	 * We want to reserve all the bytes we may need all at once, so we only
 579	 * do 1 enospc flushing cycle per transaction start.
 580	 */
 581	ret = btrfs_reserve_metadata_bytes(si, bytes, flush);
 582
 583	/*
 584	 * If we are an emergency flush, which can steal from the global block
 585	 * reserve, then attempt to not reserve space for the delayed refs, as
 586	 * we will consume space for them from the global block reserve.
 587	 */
 588	if (ret && flush == BTRFS_RESERVE_FLUSH_ALL_STEAL) {
 589		bytes -= *delayed_refs_bytes;
 590		*delayed_refs_bytes = 0;
 591		ret = btrfs_reserve_metadata_bytes(si, bytes, flush);
 592	}
 593
 594	return ret;
 595}
 596
 597static struct btrfs_trans_handle *
 598start_transaction(struct btrfs_root *root, unsigned int num_items,
 599		  unsigned int type, enum btrfs_reserve_flush_enum flush,
 600		  bool enforce_qgroups)
 601{
 602	struct btrfs_fs_info *fs_info = root->fs_info;
 603	struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
 604	struct btrfs_block_rsv *trans_rsv = &fs_info->trans_block_rsv;
 605	struct btrfs_trans_handle *h;
 606	struct btrfs_transaction *cur_trans;
 607	u64 num_bytes = 0;
 608	u64 qgroup_reserved = 0;
 609	u64 delayed_refs_bytes = 0;
 610	bool reloc_reserved = false;
 611	bool do_chunk_alloc = false;
 612	int ret;
 613
 614	if (BTRFS_FS_ERROR(fs_info))
 615		return ERR_PTR(-EROFS);
 616
 617	if (current->journal_info) {
 618		WARN_ON(type & TRANS_EXTWRITERS);
 619		h = current->journal_info;
 620		refcount_inc(&h->use_count);
 621		WARN_ON(refcount_read(&h->use_count) > 2);
 622		h->orig_rsv = h->block_rsv;
 623		h->block_rsv = NULL;
 624		goto got_it;
 625	}
 626
 627	/*
 628	 * Do the reservation before we join the transaction so we can do all
 629	 * the appropriate flushing if need be.
 630	 */
 631	if (num_items && root != fs_info->chunk_root) {
 632		qgroup_reserved = num_items * fs_info->nodesize;
 633		/*
 634		 * Use prealloc for now, as there might be a currently running
 635		 * transaction that could free this reserved space prematurely
 636		 * by committing.
 637		 */
 638		ret = btrfs_qgroup_reserve_meta_prealloc(root, qgroup_reserved,
 639							 enforce_qgroups, false);
 640		if (ret)
 641			return ERR_PTR(ret);
 642
 643		num_bytes = btrfs_calc_insert_metadata_size(fs_info, num_items);
 644		/*
 645		 * If we plan to insert/update/delete "num_items" from a btree,
 646		 * we will also generate delayed refs for extent buffers in the
 647		 * respective btree paths, so reserve space for the delayed refs
 648		 * that will be generated by the caller as it modifies btrees.
 649		 * Try to reserve them to avoid excessive use of the global
 650		 * block reserve.
 651		 */
 652		delayed_refs_bytes = btrfs_calc_delayed_ref_bytes(fs_info, num_items);
 653
 654		/*
 655		 * Do the reservation for the relocation root creation
 656		 */
 657		if (need_reserve_reloc_root(root)) {
 658			num_bytes += fs_info->nodesize;
 659			reloc_reserved = true;
 660		}
 661
 662		ret = btrfs_reserve_trans_metadata(fs_info, flush, num_bytes,
 663						   &delayed_refs_bytes);
 664		if (ret)
 665			goto reserve_fail;
 666
 667		btrfs_block_rsv_add_bytes(trans_rsv, num_bytes, true);
 668
 669		if (trans_rsv->space_info->force_alloc)
 670			do_chunk_alloc = true;
 671	} else if (num_items == 0 && flush == BTRFS_RESERVE_FLUSH_ALL &&
 672		   !btrfs_block_rsv_full(delayed_refs_rsv)) {
 673		/*
 674		 * Some people call with btrfs_start_transaction(root, 0)
 675		 * because they can be throttled, but have some other mechanism
 676		 * for reserving space.  We still want these guys to refill the
 677		 * delayed block_rsv so just add 1 items worth of reservation
 678		 * here.
 679		 */
 680		ret = btrfs_delayed_refs_rsv_refill(fs_info, flush);
 681		if (ret)
 682			goto reserve_fail;
 683	}
 684again:
 685	h = kmem_cache_zalloc(btrfs_trans_handle_cachep, GFP_NOFS);
 686	if (!h) {
 687		ret = -ENOMEM;
 688		goto alloc_fail;
 689	}
 690
 691	/*
 692	 * If we are JOIN_NOLOCK we're already committing a transaction and
 693	 * waiting on this guy, so we don't need to do the sb_start_intwrite
 694	 * because we're already holding a ref.  We need this because we could
 695	 * have raced in and did an fsync() on a file which can kick a commit
 696	 * and then we deadlock with somebody doing a freeze.
 697	 *
 698	 * If we are ATTACH, it means we just want to catch the current
 699	 * transaction and commit it, so we needn't do sb_start_intwrite(). 
 700	 */
 701	if (type & __TRANS_FREEZABLE)
 702		sb_start_intwrite(fs_info->sb);
 703
 704	if (may_wait_transaction(fs_info, type))
 705		wait_current_trans(fs_info, type);
 706
 707	do {
 708		ret = join_transaction(fs_info, type);
 709		if (ret == -EBUSY) {
 710			wait_current_trans(fs_info, type);
 711			if (unlikely(type == TRANS_ATTACH ||
 712				     type == TRANS_JOIN_NOSTART))
 713				ret = -ENOENT;
 714		}
 715	} while (ret == -EBUSY);
 716
 717	if (ret < 0)
 718		goto join_fail;
 719
 720	cur_trans = fs_info->running_transaction;
 721
 722	h->transid = cur_trans->transid;
 723	h->transaction = cur_trans;
 724	refcount_set(&h->use_count, 1);
 725	h->fs_info = root->fs_info;
 726
 727	h->type = type;
 728	INIT_LIST_HEAD(&h->new_bgs);
 729	btrfs_init_metadata_block_rsv(fs_info, &h->delayed_rsv, BTRFS_BLOCK_RSV_DELREFS);
 730
 731	smp_mb();
 732	if (cur_trans->state >= TRANS_STATE_COMMIT_START &&
 733	    may_wait_transaction(fs_info, type)) {
 734		current->journal_info = h;
 735		btrfs_commit_transaction(h);
 736		goto again;
 737	}
 738
 739	if (num_bytes) {
 740		trace_btrfs_space_reservation(fs_info, "transaction",
 741					      h->transid, num_bytes, 1);
 742		h->block_rsv = trans_rsv;
 743		h->bytes_reserved = num_bytes;
 744		if (delayed_refs_bytes > 0) {
 745			trace_btrfs_space_reservation(fs_info,
 746						      "local_delayed_refs_rsv",
 747						      h->transid,
 748						      delayed_refs_bytes, 1);
 749			h->delayed_refs_bytes_reserved = delayed_refs_bytes;
 750			btrfs_block_rsv_add_bytes(&h->delayed_rsv, delayed_refs_bytes, true);
 751			delayed_refs_bytes = 0;
 752		}
 753		h->reloc_reserved = reloc_reserved;
 754	}
 755
 756got_it:
 757	if (!current->journal_info)
 758		current->journal_info = h;
 759
 760	/*
 761	 * If the space_info is marked ALLOC_FORCE then we'll get upgraded to
 762	 * ALLOC_FORCE the first run through, and then we won't allocate for
 763	 * anybody else who races in later.  We don't care about the return
 764	 * value here.
 765	 */
 766	if (do_chunk_alloc && num_bytes) {
 767		struct btrfs_space_info *space_info = h->block_rsv->space_info;
 768		u64 flags = space_info->flags;
 769
 770		btrfs_chunk_alloc(h, space_info, btrfs_get_alloc_profile(fs_info, flags),
 771				  CHUNK_ALLOC_NO_FORCE);
 772	}
 773
 774	/*
 775	 * btrfs_record_root_in_trans() needs to alloc new extents, and may
 776	 * call btrfs_join_transaction() while we're also starting a
 777	 * transaction.
 778	 *
 779	 * Thus it need to be called after current->journal_info initialized,
 780	 * or we can deadlock.
 781	 */
 782	ret = btrfs_record_root_in_trans(h, root);
 783	if (ret) {
 784		/*
 785		 * The transaction handle is fully initialized and linked with
 786		 * other structures so it needs to be ended in case of errors,
 787		 * not just freed.
 788		 */
 789		btrfs_end_transaction(h);
 790		goto reserve_fail;
 791	}
 792	/*
 793	 * Now that we have found a transaction to be a part of, convert the
 794	 * qgroup reservation from prealloc to pertrans. A different transaction
 795	 * can't race in and free our pertrans out from under us.
 796	 */
 797	if (qgroup_reserved)
 798		btrfs_qgroup_convert_reserved_meta(root, qgroup_reserved);
 799
 800	return h;
 801
 802join_fail:
 803	if (type & __TRANS_FREEZABLE)
 804		sb_end_intwrite(fs_info->sb);
 805	kmem_cache_free(btrfs_trans_handle_cachep, h);
 806alloc_fail:
 807	if (num_bytes)
 808		btrfs_block_rsv_release(fs_info, trans_rsv, num_bytes, NULL);
 809	if (delayed_refs_bytes)
 810		btrfs_space_info_free_bytes_may_use(trans_rsv->space_info, delayed_refs_bytes);
 811reserve_fail:
 812	btrfs_qgroup_free_meta_prealloc(root, qgroup_reserved);
 813	return ERR_PTR(ret);
 814}
 815
 816struct btrfs_trans_handle *btrfs_start_transaction(struct btrfs_root *root,
 817						   unsigned int num_items)
 818{
 819	return start_transaction(root, num_items, TRANS_START,
 820				 BTRFS_RESERVE_FLUSH_ALL, true);
 821}
 822
 823struct btrfs_trans_handle *btrfs_start_transaction_fallback_global_rsv(
 824					struct btrfs_root *root,
 825					unsigned int num_items)
 826{
 827	return start_transaction(root, num_items, TRANS_START,
 828				 BTRFS_RESERVE_FLUSH_ALL_STEAL, false);
 829}
 830
 831struct btrfs_trans_handle *btrfs_join_transaction(struct btrfs_root *root)
 832{
 833	return start_transaction(root, 0, TRANS_JOIN, BTRFS_RESERVE_NO_FLUSH,
 834				 true);
 835}
 836
 837struct btrfs_trans_handle *btrfs_join_transaction_spacecache(struct btrfs_root *root)
 838{
 839	return start_transaction(root, 0, TRANS_JOIN_NOLOCK,
 840				 BTRFS_RESERVE_NO_FLUSH, true);
 841}
 842
 843/*
 844 * Similar to regular join but it never starts a transaction when none is
 845 * running or when there's a running one at a state >= TRANS_STATE_UNBLOCKED.
 846 * This is similar to btrfs_attach_transaction() but it allows the join to
 847 * happen if the transaction commit already started but it's not yet in the
 848 * "doing" phase (the state is < TRANS_STATE_COMMIT_DOING).
 849 */
 850struct btrfs_trans_handle *btrfs_join_transaction_nostart(struct btrfs_root *root)
 851{
 852	return start_transaction(root, 0, TRANS_JOIN_NOSTART,
 853				 BTRFS_RESERVE_NO_FLUSH, true);
 854}
 855
 856/*
 857 * Catch the running transaction.
 858 *
 859 * It is used when we want to commit the current the transaction, but
 860 * don't want to start a new one.
 861 *
 862 * Note: If this function return -ENOENT, it just means there is no
 863 * running transaction. But it is possible that the inactive transaction
 864 * is still in the memory, not fully on disk. If you hope there is no
 865 * inactive transaction in the fs when -ENOENT is returned, you should
 866 * invoke
 867 *     btrfs_attach_transaction_barrier()
 868 */
 869struct btrfs_trans_handle *btrfs_attach_transaction(struct btrfs_root *root)
 870{
 871	return start_transaction(root, 0, TRANS_ATTACH,
 872				 BTRFS_RESERVE_NO_FLUSH, true);
 873}
 874
 875/*
 876 * Catch the running transaction.
 877 *
 878 * It is similar to the above function, the difference is this one
 879 * will wait for all the inactive transactions until they fully
 880 * complete.
 881 */
 882struct btrfs_trans_handle *
 883btrfs_attach_transaction_barrier(struct btrfs_root *root)
 884{
 885	struct btrfs_trans_handle *trans;
 886
 887	trans = start_transaction(root, 0, TRANS_ATTACH,
 888				  BTRFS_RESERVE_NO_FLUSH, true);
 889	if (trans == ERR_PTR(-ENOENT)) {
 890		int ret;
 891
 892		ret = btrfs_wait_for_commit(root->fs_info, 0);
 893		if (ret)
 894			return ERR_PTR(ret);
 895	}
 896
 897	return trans;
 898}
 899
 900/* Wait for a transaction commit to reach at least the given state. */
 901static noinline void wait_for_commit(struct btrfs_transaction *commit,
 902				     const enum btrfs_trans_state min_state)
 903{
 904	struct btrfs_fs_info *fs_info = commit->fs_info;
 905	u64 transid = commit->transid;
 906	bool put = false;
 907
 908	/*
 909	 * At the moment this function is called with min_state either being
 910	 * TRANS_STATE_COMPLETED or TRANS_STATE_SUPER_COMMITTED.
 911	 */
 912	if (min_state == TRANS_STATE_COMPLETED)
 913		btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED);
 914	else
 915		btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED);
 916
 917	while (1) {
 918		wait_event(commit->commit_wait, commit->state >= min_state);
 919		if (put)
 920			btrfs_put_transaction(commit);
 921
 922		if (min_state < TRANS_STATE_COMPLETED)
 923			break;
 924
 925		/*
 926		 * A transaction isn't really completed until all of the
 927		 * previous transactions are completed, but with fsync we can
 928		 * end up with SUPER_COMMITTED transactions before a COMPLETED
 929		 * transaction. Wait for those.
 930		 */
 931
 932		spin_lock(&fs_info->trans_lock);
 933		commit = list_first_entry_or_null(&fs_info->trans_list,
 934						  struct btrfs_transaction,
 935						  list);
 936		if (!commit || commit->transid > transid) {
 937			spin_unlock(&fs_info->trans_lock);
 938			break;
 939		}
 940		refcount_inc(&commit->use_count);
 941		put = true;
 942		spin_unlock(&fs_info->trans_lock);
 943	}
 944}
 945
 946int btrfs_wait_for_commit(struct btrfs_fs_info *fs_info, u64 transid)
 947{
 948	struct btrfs_transaction *cur_trans = NULL, *t;
 949	int ret = 0;
 950
 951	if (transid) {
 952		if (transid <= btrfs_get_last_trans_committed(fs_info))
 953			goto out;
 954
 955		/* find specified transaction */
 956		spin_lock(&fs_info->trans_lock);
 957		list_for_each_entry(t, &fs_info->trans_list, list) {
 958			if (t->transid == transid) {
 959				cur_trans = t;
 960				refcount_inc(&cur_trans->use_count);
 961				ret = 0;
 962				break;
 963			}
 964			if (t->transid > transid) {
 965				ret = 0;
 966				break;
 967			}
 968		}
 969		spin_unlock(&fs_info->trans_lock);
 970
 971		/*
 972		 * The specified transaction doesn't exist, or we
 973		 * raced with btrfs_commit_transaction
 974		 */
 975		if (!cur_trans) {
 976			if (transid > btrfs_get_last_trans_committed(fs_info))
 977				ret = -EINVAL;
 978			goto out;
 979		}
 980	} else {
 981		/* find newest transaction that is committing | committed */
 982		spin_lock(&fs_info->trans_lock);
 983		list_for_each_entry_reverse(t, &fs_info->trans_list,
 984					    list) {
 985			if (t->state >= TRANS_STATE_COMMIT_START) {
 986				if (t->state == TRANS_STATE_COMPLETED)
 987					break;
 988				cur_trans = t;
 989				refcount_inc(&cur_trans->use_count);
 990				break;
 991			}
 992		}
 993		spin_unlock(&fs_info->trans_lock);
 994		if (!cur_trans)
 995			goto out;  /* nothing committing|committed */
 996	}
 997
 998	wait_for_commit(cur_trans, TRANS_STATE_COMPLETED);
 999	ret = cur_trans->aborted;
1000	btrfs_put_transaction(cur_trans);
1001out:
1002	return ret;
1003}
1004
1005void btrfs_throttle(struct btrfs_fs_info *fs_info)
1006{
1007	wait_current_trans(fs_info, TRANS_START);
1008}
1009
1010bool btrfs_should_end_transaction(struct btrfs_trans_handle *trans)
1011{
1012	struct btrfs_transaction *cur_trans = trans->transaction;
1013
1014	if (cur_trans->state >= TRANS_STATE_COMMIT_START ||
1015	    test_bit(BTRFS_DELAYED_REFS_FLUSHING, &cur_trans->delayed_refs.flags))
1016		return true;
1017
1018	if (btrfs_check_space_for_delayed_refs(trans->fs_info))
1019		return true;
1020
1021	return !!btrfs_block_rsv_check(&trans->fs_info->global_block_rsv, 50);
1022}
1023
1024static void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans)
1025
1026{
1027	struct btrfs_fs_info *fs_info = trans->fs_info;
1028
1029	if (!trans->block_rsv) {
1030		ASSERT(trans->bytes_reserved == 0,
1031		       "trans->bytes_reserved=%llu", trans->bytes_reserved);
1032		ASSERT(trans->delayed_refs_bytes_reserved == 0,
1033		       "trans->delayed_refs_bytes_reserved=%llu",
1034		       trans->delayed_refs_bytes_reserved);
1035		return;
1036	}
1037
1038	if (!trans->bytes_reserved) {
1039		ASSERT(trans->delayed_refs_bytes_reserved == 0,
1040		       "trans->delayed_refs_bytes_reserved=%llu",
1041		       trans->delayed_refs_bytes_reserved);
1042		return;
1043	}
1044
1045	ASSERT(trans->block_rsv == &fs_info->trans_block_rsv);
1046	trace_btrfs_space_reservation(fs_info, "transaction",
1047				      trans->transid, trans->bytes_reserved, 0);
1048	btrfs_block_rsv_release(fs_info, trans->block_rsv,
1049				trans->bytes_reserved, NULL);
1050	trans->bytes_reserved = 0;
1051
1052	if (!trans->delayed_refs_bytes_reserved)
1053		return;
1054
1055	trace_btrfs_space_reservation(fs_info, "local_delayed_refs_rsv",
1056				      trans->transid,
1057				      trans->delayed_refs_bytes_reserved, 0);
1058	btrfs_block_rsv_release(fs_info, &trans->delayed_rsv,
1059				trans->delayed_refs_bytes_reserved, NULL);
1060	trans->delayed_refs_bytes_reserved = 0;
1061}
1062
1063static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
1064				   int throttle)
1065{
1066	struct btrfs_fs_info *info = trans->fs_info;
1067	struct btrfs_transaction *cur_trans = trans->transaction;
1068	int ret = 0;
1069
1070	if (refcount_read(&trans->use_count) > 1) {
1071		refcount_dec(&trans->use_count);
1072		trans->block_rsv = trans->orig_rsv;
1073		return 0;
1074	}
1075
1076	btrfs_trans_release_metadata(trans);
1077	trans->block_rsv = NULL;
1078
1079	btrfs_create_pending_block_groups(trans);
1080
1081	btrfs_trans_release_chunk_metadata(trans);
1082
1083	if (trans->type & __TRANS_FREEZABLE)
1084		sb_end_intwrite(info->sb);
1085
1086	WARN_ON(cur_trans != info->running_transaction);
1087	WARN_ON(atomic_read(&cur_trans->num_writers) < 1);
1088	atomic_dec(&cur_trans->num_writers);
1089	extwriter_counter_dec(cur_trans, trans->type);
1090
1091	cond_wake_up(&cur_trans->writer_wait);
1092
1093	btrfs_lockdep_release(info, btrfs_trans_num_extwriters);
1094	btrfs_lockdep_release(info, btrfs_trans_num_writers);
1095
1096	btrfs_put_transaction(cur_trans);
1097
1098	if (current->journal_info == trans)
1099		current->journal_info = NULL;
1100
1101	if (throttle)
1102		btrfs_run_delayed_iputs(info);
1103
1104	if (TRANS_ABORTED(trans) || BTRFS_FS_ERROR(info)) {
1105		wake_up_process(info->transaction_kthread);
1106		if (TRANS_ABORTED(trans))
1107			ret = trans->aborted;
1108		else
1109			ret = -EROFS;
1110	}
1111
1112	kmem_cache_free(btrfs_trans_handle_cachep, trans);
1113	return ret;
1114}
1115
1116int btrfs_end_transaction(struct btrfs_trans_handle *trans)
1117{
1118	return __btrfs_end_transaction(trans, 0);
1119}
1120
1121int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans)
1122{
1123	return __btrfs_end_transaction(trans, 1);
1124}
1125
1126/*
1127 * when btree blocks are allocated, they have some corresponding bits set for
1128 * them in one of two extent_io trees.  This is used to make sure all of
1129 * those extents are sent to disk but does not wait on them
1130 */
1131int btrfs_write_marked_extents(struct btrfs_fs_info *fs_info,
1132			       struct extent_io_tree *dirty_pages, int mark)
1133{
1134	int ret = 0;
1135	struct address_space *mapping = fs_info->btree_inode->i_mapping;
1136	struct extent_state *cached_state = NULL;
1137	u64 start = 0;
1138	u64 end;
1139
1140	while (btrfs_find_first_extent_bit(dirty_pages, start, &start, &end,
1141					   mark, &cached_state)) {
1142		bool wait_writeback = false;
1143
1144		ret = btrfs_convert_extent_bit(dirty_pages, start, end,
1145					       EXTENT_NEED_WAIT,
1146					       mark, &cached_state);
1147		/*
1148		 * convert_extent_bit can return -ENOMEM, which is most of the
1149		 * time a temporary error. So when it happens, ignore the error
1150		 * and wait for writeback of this range to finish - because we
1151		 * failed to set the bit EXTENT_NEED_WAIT for the range, a call
1152		 * to __btrfs_wait_marked_extents() would not know that
1153		 * writeback for this range started and therefore wouldn't
1154		 * wait for it to finish - we don't want to commit a
1155		 * superblock that points to btree nodes/leafs for which
1156		 * writeback hasn't finished yet (and without errors).
1157		 * We cleanup any entries left in the io tree when committing
1158		 * the transaction (through extent_io_tree_release()).
1159		 */
1160		if (ret == -ENOMEM) {
1161			ret = 0;
1162			wait_writeback = true;
1163		}
1164		if (!ret)
1165			ret = filemap_fdatawrite_range(mapping, start, end);
1166		if (!ret && wait_writeback)
1167			btrfs_btree_wait_writeback_range(fs_info, start, end);
1168		btrfs_free_extent_state(cached_state);
1169		if (ret)
1170			break;
1171		cached_state = NULL;
1172		cond_resched();
1173		start = end + 1;
1174	}
1175	return ret;
1176}
1177
1178/*
1179 * when btree blocks are allocated, they have some corresponding bits set for
1180 * them in one of two extent_io trees.  This is used to make sure all of
1181 * those extents are on disk for transaction or log commit.  We wait
1182 * on all the pages and clear them from the dirty pages state tree
1183 */
1184static int __btrfs_wait_marked_extents(struct btrfs_fs_info *fs_info,
1185				       struct extent_io_tree *dirty_pages)
1186{
1187	struct extent_state *cached_state = NULL;
1188	u64 start = 0;
1189	u64 end;
1190	int ret = 0;
1191
1192	while (btrfs_find_first_extent_bit(dirty_pages, start, &start, &end,
1193					   EXTENT_NEED_WAIT, &cached_state)) {
1194		/*
1195		 * Ignore -ENOMEM errors returned by clear_extent_bit().
1196		 * When committing the transaction, we'll remove any entries
1197		 * left in the io tree. For a log commit, we don't remove them
1198		 * after committing the log because the tree can be accessed
1199		 * concurrently - we do it only at transaction commit time when
1200		 * it's safe to do it (through extent_io_tree_release()).
1201		 */
1202		ret = btrfs_clear_extent_bit(dirty_pages, start, end,
1203					     EXTENT_NEED_WAIT, &cached_state);
1204		if (ret == -ENOMEM)
1205			ret = 0;
1206		if (!ret)
1207			btrfs_btree_wait_writeback_range(fs_info, start, end);
1208		btrfs_free_extent_state(cached_state);
1209		if (ret)
1210			break;
1211		cached_state = NULL;
1212		cond_resched();
1213		start = end + 1;
1214	}
1215	return ret;
1216}
1217
1218static int btrfs_wait_extents(struct btrfs_fs_info *fs_info,
1219		       struct extent_io_tree *dirty_pages)
1220{
1221	bool errors = false;
1222	int ret;
1223
1224	ret = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1225	if (test_and_clear_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags))
1226		errors = true;
1227
1228	if (errors && !ret)
1229		ret = -EIO;
1230	return ret;
1231}
1232
1233int btrfs_wait_tree_log_extents(struct btrfs_root *log_root, int mark)
1234{
1235	struct btrfs_fs_info *fs_info = log_root->fs_info;
1236	struct extent_io_tree *dirty_pages = &log_root->dirty_log_pages;
1237	bool errors = false;
1238	int ret;
1239
1240	ASSERT(btrfs_root_id(log_root) == BTRFS_TREE_LOG_OBJECTID,
1241	       "root_id(log_root)=%llu", btrfs_root_id(log_root));
1242
1243	ret = __btrfs_wait_marked_extents(fs_info, dirty_pages);
1244	if ((mark & EXTENT_DIRTY_LOG1) &&
1245	    test_and_clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags))
1246		errors = true;
1247
1248	if ((mark & EXTENT_DIRTY_LOG2) &&
1249	    test_and_clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags))
1250		errors = true;
1251
1252	if (errors && !ret)
1253		ret = -EIO;
1254	return ret;
1255}
1256
1257/*
1258 * When btree blocks are allocated the corresponding extents are marked dirty.
1259 * This function ensures such extents are persisted on disk for transaction or
1260 * log commit.
1261 *
1262 * @trans: transaction whose dirty pages we'd like to write
1263 */
1264static int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans)
1265{
1266	int ret;
1267	int ret2;
1268	struct extent_io_tree *dirty_pages = &trans->transaction->dirty_pages;
1269	struct btrfs_fs_info *fs_info = trans->fs_info;
1270	struct blk_plug plug;
1271
1272	blk_start_plug(&plug);
1273	ret = btrfs_write_marked_extents(fs_info, dirty_pages, EXTENT_DIRTY);
1274	blk_finish_plug(&plug);
1275	ret2 = btrfs_wait_extents(fs_info, dirty_pages);
1276
1277	btrfs_extent_io_tree_release(&trans->transaction->dirty_pages);
1278
1279	if (ret)
1280		return ret;
1281	else if (ret2)
1282		return ret2;
1283	else
1284		return 0;
1285}
1286
1287/*
1288 * this is used to update the root pointer in the tree of tree roots.
1289 *
1290 * But, in the case of the extent allocation tree, updating the root
1291 * pointer may allocate blocks which may change the root of the extent
1292 * allocation tree.
1293 *
1294 * So, this loops and repeats and makes sure the cowonly root didn't
1295 * change while the root pointer was being updated in the metadata.
1296 */
1297static int update_cowonly_root(struct btrfs_trans_handle *trans,
1298			       struct btrfs_root *root)
1299{
1300	int ret;
1301	u64 old_root_bytenr;
1302	u64 old_root_used;
1303	struct btrfs_fs_info *fs_info = root->fs_info;
1304	struct btrfs_root *tree_root = fs_info->tree_root;
1305
1306	old_root_used = btrfs_root_used(&root->root_item);
1307
1308	while (1) {
1309		old_root_bytenr = btrfs_root_bytenr(&root->root_item);
1310		if (old_root_bytenr == root->node->start &&
1311		    old_root_used == btrfs_root_used(&root->root_item))
1312			break;
1313
1314		btrfs_set_root_node(&root->root_item, root->node);
1315		ret = btrfs_update_root(trans, tree_root,
1316					&root->root_key,
1317					&root->root_item);
1318		if (ret)
1319			return ret;
1320
1321		old_root_used = btrfs_root_used(&root->root_item);
1322	}
1323
1324	return 0;
1325}
1326
1327/*
1328 * update all the cowonly tree roots on disk
1329 *
1330 * The error handling in this function may not be obvious. Any of the
1331 * failures will cause the file system to go offline. We still need
1332 * to clean up the delayed refs.
1333 */
1334static noinline int commit_cowonly_roots(struct btrfs_trans_handle *trans)
1335{
1336	struct btrfs_fs_info *fs_info = trans->fs_info;
1337	struct list_head *dirty_bgs = &trans->transaction->dirty_bgs;
1338	struct list_head *io_bgs = &trans->transaction->io_bgs;
1339	struct extent_buffer *eb;
1340	int ret;
1341
1342	/*
1343	 * At this point no one can be using this transaction to modify any tree
1344	 * and no one can start another transaction to modify any tree either.
1345	 */
1346	ASSERT(trans->transaction->state == TRANS_STATE_COMMIT_DOING,
1347	       "trans->transaction->state=%d", trans->transaction->state);
1348
1349	eb = btrfs_lock_root_node(fs_info->tree_root);
1350	ret = btrfs_cow_block(trans, fs_info->tree_root, eb, NULL,
1351			      0, &eb, BTRFS_NESTING_COW);
1352	btrfs_tree_unlock(eb);
1353	free_extent_buffer(eb);
1354
1355	if (ret)
1356		return ret;
1357
1358	ret = btrfs_run_dev_stats(trans);
1359	if (ret)
1360		return ret;
1361	ret = btrfs_run_dev_replace(trans);
1362	if (ret)
1363		return ret;
1364	ret = btrfs_run_qgroups(trans);
1365	if (ret)
1366		return ret;
1367
1368	ret = btrfs_setup_space_cache(trans);
1369	if (ret)
1370		return ret;
1371
1372again:
1373	while (!list_empty(&fs_info->dirty_cowonly_roots)) {
1374		struct btrfs_root *root;
1375
1376		root = list_first_entry(&fs_info->dirty_cowonly_roots,
1377					struct btrfs_root, dirty_list);
1378		clear_bit(BTRFS_ROOT_DIRTY, &root->state);
1379		list_move_tail(&root->dirty_list,
1380			       &trans->transaction->switch_commits);
1381
1382		ret = update_cowonly_root(trans, root);
1383		if (ret)
1384			return ret;
1385	}
1386
1387	/* Now flush any delayed refs generated by updating all of the roots */
1388	ret = btrfs_run_delayed_refs(trans, U64_MAX);
1389	if (ret)
1390		return ret;
1391
1392	while (!list_empty(dirty_bgs) || !list_empty(io_bgs)) {
1393		ret = btrfs_write_dirty_block_groups(trans);
1394		if (ret)
1395			return ret;
1396
1397		/*
1398		 * We're writing the dirty block groups, which could generate
1399		 * delayed refs, which could generate more dirty block groups,
1400		 * so we want to keep this flushing in this loop to make sure
1401		 * everything gets run.
1402		 */
1403		ret = btrfs_run_delayed_refs(trans, U64_MAX);
1404		if (ret)
1405			return ret;
1406	}
1407
1408	if (!list_empty(&fs_info->dirty_cowonly_roots))
1409		goto again;
1410
1411	/* Update dev-replace pointer once everything is committed */
1412	fs_info->dev_replace.committed_cursor_left =
1413		fs_info->dev_replace.cursor_left_last_write_of_item;
1414
1415	return 0;
1416}
1417
1418/*
1419 * If we had a pending drop we need to see if there are any others left in our
1420 * dead roots list, and if not clear our bit and wake any waiters.
1421 */
1422void btrfs_maybe_wake_unfinished_drop(struct btrfs_fs_info *fs_info)
1423{
1424	/*
1425	 * We put the drop in progress roots at the front of the list, so if the
1426	 * first entry doesn't have UNFINISHED_DROP set we can wake everybody
1427	 * up.
1428	 */
1429	spin_lock(&fs_info->trans_lock);
1430	if (!list_empty(&fs_info->dead_roots)) {
1431		struct btrfs_root *root = list_first_entry(&fs_info->dead_roots,
1432							   struct btrfs_root,
1433							   root_list);
1434		if (test_bit(BTRFS_ROOT_UNFINISHED_DROP, &root->state)) {
1435			spin_unlock(&fs_info->trans_lock);
1436			return;
1437		}
1438	}
1439	spin_unlock(&fs_info->trans_lock);
1440
1441	btrfs_wake_unfinished_drop(fs_info);
1442}
1443
1444/*
1445 * dead roots are old snapshots that need to be deleted.  This allocates
1446 * a dirty root struct and adds it into the list of dead roots that need to
1447 * be deleted
1448 */
1449void btrfs_add_dead_root(struct btrfs_root *root)
1450{
1451	struct btrfs_fs_info *fs_info = root->fs_info;
1452
1453	spin_lock(&fs_info->trans_lock);
1454	if (list_empty(&root->root_list)) {
1455		btrfs_grab_root(root);
1456
1457		/* We want to process the partially complete drops first. */
1458		if (test_bit(BTRFS_ROOT_UNFINISHED_DROP, &root->state))
1459			list_add(&root->root_list, &fs_info->dead_roots);
1460		else
1461			list_add_tail(&root->root_list, &fs_info->dead_roots);
1462	}
1463	spin_unlock(&fs_info->trans_lock);
1464}
1465
1466/*
1467 * Update each subvolume root and its relocation root, if it exists, in the tree
1468 * of tree roots. Also free log roots if they exist.
1469 */
1470static noinline int commit_fs_roots(struct btrfs_trans_handle *trans)
1471{
1472	struct btrfs_fs_info *fs_info = trans->fs_info;
1473	struct btrfs_root *gang[8];
1474	int i;
1475	int ret;
1476
1477	/*
1478	 * At this point no one can be using this transaction to modify any tree
1479	 * and no one can start another transaction to modify any tree either.
1480	 */
1481	ASSERT(trans->transaction->state == TRANS_STATE_COMMIT_DOING,
1482	       "trans->transaction->state=%d", trans->transaction->state);
1483
1484	spin_lock(&fs_info->fs_roots_radix_lock);
1485	while (1) {
1486		ret = radix_tree_gang_lookup_tag(&fs_info->fs_roots_radix,
1487						 (void **)gang, 0,
1488						 ARRAY_SIZE(gang),
1489						 BTRFS_ROOT_TRANS_TAG);
1490		if (ret == 0)
1491			break;
1492		for (i = 0; i < ret; i++) {
1493			struct btrfs_root *root = gang[i];
1494			int ret2;
1495
1496			/*
1497			 * At this point we can neither have tasks logging inodes
1498			 * from a root nor trying to commit a log tree.
1499			 */
1500			ASSERT(atomic_read(&root->log_writers) == 0,
1501			       "atomic_read(&root->log_writers)=%d",
1502			       atomic_read(&root->log_writers));
1503			ASSERT(atomic_read(&root->log_commit[0]) == 0,
1504			       "atomic_read(&root->log_commit[0])=%d",
1505			       atomic_read(&root->log_commit[0]));
1506			ASSERT(atomic_read(&root->log_commit[1]) == 0,
1507			       "atomic_read(&root->log_commit[1])=%d",
1508			       atomic_read(&root->log_commit[1]));
1509
1510			radix_tree_tag_clear(&fs_info->fs_roots_radix,
1511					(unsigned long)btrfs_root_id(root),
1512					BTRFS_ROOT_TRANS_TAG);
1513			btrfs_qgroup_free_meta_all_pertrans(root);
1514			spin_unlock(&fs_info->fs_roots_radix_lock);
1515
1516			btrfs_free_log(trans, root);
1517			ret2 = btrfs_update_reloc_root(trans, root);
1518			if (ret2)
1519				return ret2;
1520
1521			/* see comments in should_cow_block() */
1522			clear_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1523			smp_mb__after_atomic();
1524
1525			if (root->commit_root != root->node) {
1526				list_add_tail(&root->dirty_list,
1527					&trans->transaction->switch_commits);
1528				btrfs_set_root_node(&root->root_item,
1529						    root->node);
1530			}
1531
1532			ret2 = btrfs_update_root(trans, fs_info->tree_root,
1533						&root->root_key,
1534						&root->root_item);
1535			if (ret2)
1536				return ret2;
1537			spin_lock(&fs_info->fs_roots_radix_lock);
1538		}
1539	}
1540	spin_unlock(&fs_info->fs_roots_radix_lock);
1541	return 0;
1542}
1543
1544/*
1545 * Do all special snapshot related qgroup dirty hack.
1546 *
1547 * Will do all needed qgroup inherit and dirty hack like switch commit
1548 * roots inside one transaction and write all btree into disk, to make
1549 * qgroup works.
1550 */
1551static int qgroup_account_snapshot(struct btrfs_trans_handle *trans,
1552				   struct btrfs_root *src,
1553				   struct btrfs_root *parent,
1554				   struct btrfs_qgroup_inherit *inherit,
1555				   u64 dst_objectid)
1556{
1557	struct btrfs_fs_info *fs_info = src->fs_info;
1558	int ret;
1559
1560	/*
1561	 * Save some performance in the case that qgroups are not enabled. If
1562	 * this check races with the ioctl, rescan will kick in anyway.
1563	 */
1564	if (!btrfs_qgroup_full_accounting(fs_info))
1565		return 0;
1566
1567	/*
1568	 * Ensure dirty @src will be committed.  Or, after coming
1569	 * commit_fs_roots() and switch_commit_roots(), any dirty but not
1570	 * recorded root will never be updated again, causing an outdated root
1571	 * item.
1572	 */
1573	ret = record_root_in_trans(trans, src, 1);
1574	if (ret)
1575		return ret;
1576
1577	/*
1578	 * btrfs_qgroup_inherit relies on a consistent view of the usage for the
1579	 * src root, so we must run the delayed refs here.
1580	 *
1581	 * However this isn't particularly fool proof, because there's no
1582	 * synchronization keeping us from changing the tree after this point
1583	 * before we do the qgroup_inherit, or even from making changes while
1584	 * we're doing the qgroup_inherit.  But that's a problem for the future,
1585	 * for now flush the delayed refs to narrow the race window where the
1586	 * qgroup counters could end up wrong.
1587	 */
1588	ret = btrfs_run_delayed_refs(trans, U64_MAX);
1589	if (unlikely(ret)) {
1590		btrfs_abort_transaction(trans, ret);
1591		return ret;
1592	}
1593
1594	ret = commit_fs_roots(trans);
1595	if (ret)
1596		goto out;
1597	ret = btrfs_qgroup_account_extents(trans);
1598	if (ret < 0)
1599		goto out;
1600
1601	/* Now qgroup are all updated, we can inherit it to new qgroups */
1602	ret = btrfs_qgroup_inherit(trans, btrfs_root_id(src), dst_objectid,
1603				   btrfs_root_id(parent), inherit);
1604	if (ret < 0)
1605		goto out;
1606
1607	/*
1608	 * Now we do a simplified commit transaction, which will:
1609	 * 1) commit all subvolume and extent tree
1610	 *    To ensure all subvolume and extent tree have a valid
1611	 *    commit_root to accounting later insert_dir_item()
1612	 * 2) write all btree blocks onto disk
1613	 *    This is to make sure later btree modification will be cowed
1614	 *    Or commit_root can be populated and cause wrong qgroup numbers
1615	 * In this simplified commit, we don't really care about other trees
1616	 * like chunk and root tree, as they won't affect qgroup.
1617	 * And we don't write super to avoid half committed status.
1618	 */
1619	ret = commit_cowonly_roots(trans);
1620	if (ret)
1621		goto out;
1622	switch_commit_roots(trans);
1623	ret = btrfs_write_and_wait_transaction(trans);
1624	if (ret)
1625		btrfs_handle_fs_error(fs_info, ret,
1626			"Error while writing out transaction for qgroup");
1627
1628out:
1629	/*
1630	 * Force parent root to be updated, as we recorded it before so its
1631	 * last_trans == cur_transid.
1632	 * Or it won't be committed again onto disk after later
1633	 * insert_dir_item()
1634	 */
1635	if (!ret)
1636		ret = record_root_in_trans(trans, parent, 1);
1637	return ret;
1638}
1639
1640/*
1641 * new snapshots need to be created at a very specific time in the
1642 * transaction commit.  This does the actual creation.
1643 *
1644 * Note:
1645 * If the error which may affect the commitment of the current transaction
1646 * happens, we should return the error number. If the error which just affect
1647 * the creation of the pending snapshots, just return 0.
1648 */
1649static noinline int create_pending_snapshot(struct btrfs_trans_handle *trans,
1650				   struct btrfs_pending_snapshot *pending)
1651{
1652
1653	struct btrfs_fs_info *fs_info = trans->fs_info;
1654	struct btrfs_key key;
1655	struct btrfs_root_item *new_root_item;
1656	struct btrfs_root *tree_root = fs_info->tree_root;
1657	struct btrfs_root *root = pending->root;
1658	struct btrfs_root *parent_root;
1659	struct btrfs_block_rsv *rsv;
1660	struct btrfs_inode *parent_inode = pending->dir;
1661	BTRFS_PATH_AUTO_FREE(path);
1662	struct btrfs_dir_item *dir_item;
1663	struct extent_buffer *tmp;
1664	struct extent_buffer *old;
1665	struct timespec64 cur_time;
1666	int ret = 0;
1667	u64 to_reserve = 0;
1668	u64 index = 0;
1669	u64 objectid;
1670	u64 root_flags;
1671	unsigned int nofs_flags;
1672	struct fscrypt_name fname;
1673
1674	ASSERT(pending->path);
1675	path = pending->path;
1676
1677	ASSERT(pending->root_item);
1678	new_root_item = pending->root_item;
1679
1680	/*
1681	 * We're inside a transaction and must make sure that any potential
1682	 * allocations with GFP_KERNEL in fscrypt won't recurse back to
1683	 * filesystem.
1684	 */
1685	nofs_flags = memalloc_nofs_save();
1686	pending->error = fscrypt_setup_filename(&parent_inode->vfs_inode,
1687						&pending->dentry->d_name, 0,
1688						&fname);
1689	memalloc_nofs_restore(nofs_flags);
1690	if (pending->error)
1691		goto free_pending;
1692
1693	pending->error = btrfs_get_free_objectid(tree_root, &objectid);
1694	if (pending->error)
1695		goto free_fname;
1696
1697	/*
1698	 * Make qgroup to skip current new snapshot's qgroupid, as it is
1699	 * accounted by later btrfs_qgroup_inherit().
1700	 */
1701	btrfs_set_skip_qgroup(trans, objectid);
1702
1703	btrfs_reloc_pre_snapshot(pending, &to_reserve);
1704
1705	if (to_reserve > 0) {
1706		pending->error = btrfs_block_rsv_add(fs_info,
1707						     &pending->block_rsv,
1708						     to_reserve,
1709						     BTRFS_RESERVE_NO_FLUSH);
1710		if (pending->error)
1711			goto clear_skip_qgroup;
1712	}
1713
1714	rsv = trans->block_rsv;
1715	trans->block_rsv = &pending->block_rsv;
1716	trans->bytes_reserved = trans->block_rsv->reserved;
1717	trace_btrfs_space_reservation(fs_info, "transaction",
1718				      trans->transid,
1719				      trans->bytes_reserved, 1);
1720	parent_root = parent_inode->root;
1721	ret = record_root_in_trans(trans, parent_root, 0);
1722	if (ret)
1723		goto fail;
1724	cur_time = current_time(&parent_inode->vfs_inode);
1725
1726	/*
1727	 * insert the directory item
1728	 */
1729	ret = btrfs_set_inode_index(parent_inode, &index);
1730	if (unlikely(ret)) {
1731		btrfs_abort_transaction(trans, ret);
1732		goto fail;
1733	}
1734
1735	/* check if there is a file/dir which has the same name. */
1736	dir_item = btrfs_lookup_dir_item(NULL, parent_root, path,
1737					 btrfs_ino(parent_inode),
1738					 &fname.disk_name, 0);
1739	if (dir_item != NULL && !IS_ERR(dir_item)) {
1740		pending->error = -EEXIST;
1741		goto dir_item_existed;
1742	} else if (IS_ERR(dir_item)) {
1743		ret = PTR_ERR(dir_item);
1744		btrfs_abort_transaction(trans, ret);
1745		goto fail;
1746	}
1747	btrfs_release_path(path);
1748
1749	ret = btrfs_create_qgroup(trans, objectid);
1750	if (ret && ret != -EEXIST) {
1751		if (unlikely(ret != -ENOTCONN || btrfs_qgroup_enabled(fs_info))) {
1752			btrfs_abort_transaction(trans, ret);
1753			goto fail;
1754		}
1755	}
1756
1757	/*
1758	 * pull in the delayed directory update
1759	 * and the delayed inode item
1760	 * otherwise we corrupt the FS during
1761	 * snapshot
1762	 */
1763	ret = btrfs_run_delayed_items(trans);
1764	if (unlikely(ret)) {
1765		btrfs_abort_transaction(trans, ret);
1766		goto fail;
1767	}
1768
1769	ret = record_root_in_trans(trans, root, 0);
1770	if (unlikely(ret)) {
1771		btrfs_abort_transaction(trans, ret);
1772		goto fail;
1773	}
1774	btrfs_set_root_last_snapshot(&root->root_item, trans->transid);
1775	memcpy(new_root_item, &root->root_item, sizeof(*new_root_item));
1776	btrfs_check_and_init_root_item(new_root_item);
1777
1778	root_flags = btrfs_root_flags(new_root_item);
1779	if (pending->readonly)
1780		root_flags |= BTRFS_ROOT_SUBVOL_RDONLY;
1781	else
1782		root_flags &= ~BTRFS_ROOT_SUBVOL_RDONLY;
1783	btrfs_set_root_flags(new_root_item, root_flags);
1784
1785	btrfs_set_root_generation_v2(new_root_item,
1786			trans->transid);
1787	generate_random_guid(new_root_item->uuid);
1788	memcpy(new_root_item->parent_uuid, root->root_item.uuid,
1789			BTRFS_UUID_SIZE);
1790	if (!(root_flags & BTRFS_ROOT_SUBVOL_RDONLY)) {
1791		memset(new_root_item->received_uuid, 0,
1792		       sizeof(new_root_item->received_uuid));
1793		memset(&new_root_item->stime, 0, sizeof(new_root_item->stime));
1794		memset(&new_root_item->rtime, 0, sizeof(new_root_item->rtime));
1795		btrfs_set_root_stransid(new_root_item, 0);
1796		btrfs_set_root_rtransid(new_root_item, 0);
1797	}
1798	btrfs_set_stack_timespec_sec(&new_root_item->otime, cur_time.tv_sec);
1799	btrfs_set_stack_timespec_nsec(&new_root_item->otime, cur_time.tv_nsec);
1800	btrfs_set_root_otransid(new_root_item, trans->transid);
1801
1802	old = btrfs_lock_root_node(root);
1803	ret = btrfs_cow_block(trans, root, old, NULL, 0, &old,
1804			      BTRFS_NESTING_COW);
1805	if (unlikely(ret)) {
1806		btrfs_tree_unlock(old);
1807		free_extent_buffer(old);
1808		btrfs_abort_transaction(trans, ret);
1809		goto fail;
1810	}
1811
1812	ret = btrfs_copy_root(trans, root, old, &tmp, objectid);
1813	/* clean up in any case */
1814	btrfs_tree_unlock(old);
1815	free_extent_buffer(old);
1816	if (unlikely(ret)) {
1817		btrfs_abort_transaction(trans, ret);
1818		goto fail;
1819	}
1820	/* see comments in should_cow_block() */
1821	set_bit(BTRFS_ROOT_FORCE_COW, &root->state);
1822	smp_mb__after_atomic();
1823
1824	btrfs_set_root_node(new_root_item, tmp);
1825	/* record when the snapshot was created in key.offset */
1826	key.objectid = objectid;
1827	key.type = BTRFS_ROOT_ITEM_KEY;
1828	key.offset = trans->transid;
1829	ret = btrfs_insert_root(trans, tree_root, &key, new_root_item);
1830	btrfs_tree_unlock(tmp);
1831	free_extent_buffer(tmp);
1832	if (unlikely(ret)) {
1833		btrfs_abort_transaction(trans, ret);
1834		goto fail;
1835	}
1836
1837	/*
1838	 * insert root back/forward references
1839	 */
1840	ret = btrfs_add_root_ref(trans, objectid,
1841				 btrfs_root_id(parent_root),
1842				 btrfs_ino(parent_inode), index,
1843				 &fname.disk_name);
1844	if (unlikely(ret)) {
1845		btrfs_abort_transaction(trans, ret);
1846		goto fail;
1847	}
1848
1849	key.offset = (u64)-1;
1850	pending->snap = btrfs_get_new_fs_root(fs_info, objectid, &pending->anon_dev);
1851	if (IS_ERR(pending->snap)) {
1852		ret = PTR_ERR(pending->snap);
1853		pending->snap = NULL;
1854		btrfs_abort_transaction(trans, ret);
1855		goto fail;
1856	}
1857
1858	ret = btrfs_reloc_post_snapshot(trans, pending);
1859	if (unlikely(ret)) {
1860		btrfs_abort_transaction(trans, ret);
1861		goto fail;
1862	}
1863
1864	/*
1865	 * Do special qgroup accounting for snapshot, as we do some qgroup
1866	 * snapshot hack to do fast snapshot.
1867	 * To co-operate with that hack, we do hack again.
1868	 * Or snapshot will be greatly slowed down by a subtree qgroup rescan
1869	 */
1870	if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_FULL)
1871		ret = qgroup_account_snapshot(trans, root, parent_root,
1872					      pending->inherit, objectid);
1873	else if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE)
1874		ret = btrfs_qgroup_inherit(trans, btrfs_root_id(root), objectid,
1875					   btrfs_root_id(parent_root), pending->inherit);
1876	if (ret < 0)
1877		goto fail;
1878
1879	ret = btrfs_insert_dir_item(trans, &fname.disk_name,
1880				    parent_inode, &key, BTRFS_FT_DIR,
1881				    index);
1882	if (unlikely(ret)) {
1883		btrfs_abort_transaction(trans, ret);
1884		goto fail;
1885	}
1886
1887	btrfs_i_size_write(parent_inode, parent_inode->vfs_inode.i_size +
1888						  fname.disk_name.len * 2);
1889	inode_set_mtime_to_ts(&parent_inode->vfs_inode,
1890			      inode_set_ctime_current(&parent_inode->vfs_inode));
1891	ret = btrfs_update_inode_fallback(trans, parent_inode);
1892	if (unlikely(ret)) {
1893		btrfs_abort_transaction(trans, ret);
1894		goto fail;
1895	}
1896	ret = btrfs_uuid_tree_add(trans, new_root_item->uuid,
1897				  BTRFS_UUID_KEY_SUBVOL,
1898				  objectid);
1899	if (unlikely(ret)) {
1900		btrfs_abort_transaction(trans, ret);
1901		goto fail;
1902	}
1903	if (!btrfs_is_empty_uuid(new_root_item->received_uuid)) {
1904		ret = btrfs_uuid_tree_add(trans, new_root_item->received_uuid,
1905					  BTRFS_UUID_KEY_RECEIVED_SUBVOL,
1906					  objectid);
1907		/*
1908		 * We are creating of lot of snapshots of the same root that was
1909		 * received (has a received UUID) and reached a leaf's limit for
1910		 * an item. We can safely ignore this and avoid a transaction
1911		 * abort. A deletion of this snapshot will still work since we
1912		 * ignore if an item with a BTRFS_UUID_KEY_RECEIVED_SUBVOL key
1913		 * is missing (see btrfs_delete_subvolume()). Send/receive will
1914		 * work too since it peeks the first root id from the existing
1915		 * item (it could peek any), and in case it's missing it
1916		 * falls back to search by BTRFS_UUID_KEY_SUBVOL keys.
1917		 * Creation of a snapshot does not require CAP_SYS_ADMIN, so
1918		 * we don't want users triggering transaction aborts, either
1919		 * intentionally or not.
1920		 */
1921		if (ret == -EOVERFLOW)
1922			ret = 0;
1923		if (unlikely(ret && ret != -EEXIST)) {
1924			btrfs_abort_transaction(trans, ret);
1925			goto fail;
1926		}
1927	}
1928
1929fail:
1930	pending->error = ret;
1931dir_item_existed:
1932	trans->block_rsv = rsv;
1933	trans->bytes_reserved = 0;
1934clear_skip_qgroup:
1935	btrfs_clear_skip_qgroup(trans);
1936free_fname:
1937	fscrypt_free_filename(&fname);
1938free_pending:
1939	kfree(new_root_item);
1940	pending->root_item = NULL;
1941	pending->path = NULL;
1942
1943	return ret;
1944}
1945
1946/*
1947 * create all the snapshots we've scheduled for creation
1948 */
1949static noinline int create_pending_snapshots(struct btrfs_trans_handle *trans)
1950{
1951	struct btrfs_pending_snapshot *pending, *next;
1952	struct list_head *head = &trans->transaction->pending_snapshots;
1953	int ret = 0;
1954
1955	list_for_each_entry_safe(pending, next, head, list) {
1956		list_del(&pending->list);
1957		ret = create_pending_snapshot(trans, pending);
1958		if (ret)
1959			break;
1960	}
1961	return ret;
1962}
1963
1964static void update_super_roots(struct btrfs_fs_info *fs_info)
1965{
1966	struct btrfs_root_item *root_item;
1967	struct btrfs_super_block *super;
1968
1969	super = fs_info->super_copy;
1970
1971	root_item = &fs_info->chunk_root->root_item;
1972	super->chunk_root = root_item->bytenr;
1973	super->chunk_root_generation = root_item->generation;
1974	super->chunk_root_level = root_item->level;
1975
1976	root_item = &fs_info->tree_root->root_item;
1977	super->root = root_item->bytenr;
1978	super->generation = root_item->generation;
1979	super->root_level = root_item->level;
1980	if (btrfs_test_opt(fs_info, SPACE_CACHE))
1981		super->cache_generation = root_item->generation;
1982	else if (test_bit(BTRFS_FS_CLEANUP_SPACE_CACHE_V1, &fs_info->flags))
1983		super->cache_generation = 0;
1984	if (test_bit(BTRFS_FS_UPDATE_UUID_TREE_GEN, &fs_info->flags))
1985		super->uuid_tree_generation = root_item->generation;
1986}
1987
1988int btrfs_transaction_blocked(struct btrfs_fs_info *info)
1989{
1990	struct btrfs_transaction *trans;
1991	int ret = 0;
1992
1993	spin_lock(&info->trans_lock);
1994	trans = info->running_transaction;
1995	if (trans)
1996		ret = is_transaction_blocked(trans);
1997	spin_unlock(&info->trans_lock);
1998	return ret;
1999}
2000
2001void btrfs_commit_transaction_async(struct btrfs_trans_handle *trans)
2002{
2003	struct btrfs_fs_info *fs_info = trans->fs_info;
2004	struct btrfs_transaction *cur_trans;
2005
2006	/* Kick the transaction kthread. */
2007	set_bit(BTRFS_FS_COMMIT_TRANS, &fs_info->flags);
2008	wake_up_process(fs_info->transaction_kthread);
2009
2010	/* take transaction reference */
2011	cur_trans = trans->transaction;
2012	refcount_inc(&cur_trans->use_count);
2013
2014	btrfs_end_transaction(trans);
2015
2016	/*
2017	 * Wait for the current transaction commit to start and block
2018	 * subsequent transaction joins
2019	 */
2020	btrfs_might_wait_for_state(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_PREP);
2021	wait_event(fs_info->transaction_blocked_wait,
2022		   cur_trans->state >= TRANS_STATE_COMMIT_START ||
2023		   TRANS_ABORTED(cur_trans));
2024	btrfs_put_transaction(cur_trans);
2025}
2026
2027/*
2028 * If there is a running transaction commit it or if it's already committing,
2029 * wait for its commit to complete. Does not start and commit a new transaction
2030 * if there isn't any running.
2031 */
2032int btrfs_commit_current_transaction(struct btrfs_root *root)
2033{
2034	struct btrfs_trans_handle *trans;
2035
2036	trans = btrfs_attach_transaction_barrier(root);
2037	if (IS_ERR(trans)) {
2038		int ret = PTR_ERR(trans);
2039
2040		return (ret == -ENOENT) ? 0 : ret;
2041	}
2042
2043	return btrfs_commit_transaction(trans);
2044}
2045
2046static void cleanup_transaction(struct btrfs_trans_handle *trans, int err)
2047{
2048	struct btrfs_fs_info *fs_info = trans->fs_info;
2049	struct btrfs_transaction *cur_trans = trans->transaction;
2050
2051	WARN_ON(refcount_read(&trans->use_count) > 1);
2052
2053	btrfs_abort_transaction(trans, err);
2054
2055	spin_lock(&fs_info->trans_lock);
2056
2057	/*
2058	 * If the transaction is removed from the list, it means this
2059	 * transaction has been committed successfully, so it is impossible
2060	 * to call the cleanup function.
2061	 */
2062	BUG_ON(list_empty(&cur_trans->list));
2063
2064	if (cur_trans == fs_info->running_transaction) {
2065		cur_trans->state = TRANS_STATE_COMMIT_DOING;
2066		spin_unlock(&fs_info->trans_lock);
2067
2068		/*
2069		 * The thread has already released the lockdep map as reader
2070		 * already in btrfs_commit_transaction().
2071		 */
2072		btrfs_might_wait_for_event(fs_info, btrfs_trans_num_writers);
2073		wait_event(cur_trans->writer_wait,
2074			   atomic_read(&cur_trans->num_writers) == 1);
2075
2076		spin_lock(&fs_info->trans_lock);
2077	}
2078
2079	/*
2080	 * Now that we know no one else is still using the transaction we can
2081	 * remove the transaction from the list of transactions. This avoids
2082	 * the transaction kthread from cleaning up the transaction while some
2083	 * other task is still using it, which could result in a use-after-free
2084	 * on things like log trees, as it forces the transaction kthread to
2085	 * wait for this transaction to be cleaned up by us.
2086	 */
2087	list_del_init(&cur_trans->list);
2088
2089	spin_unlock(&fs_info->trans_lock);
2090
2091	btrfs_cleanup_one_transaction(trans->transaction);
2092
2093	spin_lock(&fs_info->trans_lock);
2094	if (cur_trans == fs_info->running_transaction)
2095		fs_info->running_transaction = NULL;
2096	spin_unlock(&fs_info->trans_lock);
2097
2098	if (trans->type & __TRANS_FREEZABLE)
2099		sb_end_intwrite(fs_info->sb);
2100	btrfs_put_transaction(cur_trans);
2101	btrfs_put_transaction(cur_trans);
2102
2103	trace_btrfs_transaction_commit(fs_info);
2104
2105	if (current->journal_info == trans)
2106		current->journal_info = NULL;
2107
2108	/*
2109	 * If relocation is running, we can't cancel scrub because that will
2110	 * result in a deadlock. Before relocating a block group, relocation
2111	 * pauses scrub, then starts and commits a transaction before unpausing
2112	 * scrub. If the transaction commit is being done by the relocation
2113	 * task or triggered by another task and the relocation task is waiting
2114	 * for the commit, and we end up here due to an error in the commit
2115	 * path, then calling btrfs_scrub_cancel() will deadlock, as we are
2116	 * asking for scrub to stop while having it asked to be paused higher
2117	 * above in relocation code.
2118	 */
2119	if (!test_bit(BTRFS_FS_RELOC_RUNNING, &fs_info->flags))
2120		btrfs_scrub_cancel(fs_info);
2121
2122	kmem_cache_free(btrfs_trans_handle_cachep, trans);
2123}
2124
2125/*
2126 * Release reserved delayed ref space of all pending block groups of the
2127 * transaction and remove them from the list
2128 */
2129static void btrfs_cleanup_pending_block_groups(struct btrfs_trans_handle *trans)
2130{
2131       struct btrfs_fs_info *fs_info = trans->fs_info;
2132       struct btrfs_block_group *block_group, *tmp;
2133
2134       list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
2135               btrfs_dec_delayed_refs_rsv_bg_inserts(fs_info);
2136		/*
2137		* Not strictly necessary to lock, as no other task will be using a
2138		* block_group on the new_bgs list during a transaction abort.
2139		*/
2140	       spin_lock(&fs_info->unused_bgs_lock);
2141               list_del_init(&block_group->bg_list);
2142	       btrfs_put_block_group(block_group);
2143	       spin_unlock(&fs_info->unused_bgs_lock);
2144       }
2145}
2146
2147static inline int btrfs_start_delalloc_flush(struct btrfs_fs_info *fs_info)
2148{
2149	/*
2150	 * We use try_to_writeback_inodes_sb() here because if we used
2151	 * btrfs_start_delalloc_roots we would deadlock with fs freeze.
2152	 * Currently are holding the fs freeze lock, if we do an async flush
2153	 * we'll do btrfs_join_transaction() and deadlock because we need to
2154	 * wait for the fs freeze lock.  Using the direct flushing we benefit
2155	 * from already being in a transaction and our join_transaction doesn't
2156	 * have to re-take the fs freeze lock.
2157	 *
2158	 * Note that try_to_writeback_inodes_sb() will only trigger writeback
2159	 * if it can read lock sb->s_umount. It will always be able to lock it,
2160	 * except when the filesystem is being unmounted or being frozen, but in
2161	 * those cases sync_filesystem() is called, which results in calling
2162	 * writeback_inodes_sb() while holding a write lock on sb->s_umount.
2163	 * Note that we don't call writeback_inodes_sb() directly, because it
2164	 * will emit a warning if sb->s_umount is not locked.
2165	 */
2166	if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
2167		try_to_writeback_inodes_sb(fs_info->sb, WB_REASON_SYNC);
2168	return 0;
2169}
2170
2171static inline void btrfs_wait_delalloc_flush(struct btrfs_fs_info *fs_info)
2172{
2173	if (btrfs_test_opt(fs_info, FLUSHONCOMMIT))
2174		btrfs_wait_ordered_roots(fs_info, U64_MAX, NULL);
2175}
2176
2177/*
2178 * Add a pending snapshot associated with the given transaction handle to the
2179 * respective handle. This must be called after the transaction commit started
2180 * and while holding fs_info->trans_lock.
2181 * This serves to guarantee a caller of btrfs_commit_transaction() that it can
2182 * safely free the pending snapshot pointer in case btrfs_commit_transaction()
2183 * returns an error.
2184 */
2185static void add_pending_snapshot(struct btrfs_trans_handle *trans)
2186{
2187	struct btrfs_transaction *cur_trans = trans->transaction;
2188
2189	if (!trans->pending_snapshot)
2190		return;
2191
2192	lockdep_assert_held(&trans->fs_info->trans_lock);
2193	ASSERT(cur_trans->state >= TRANS_STATE_COMMIT_PREP,
2194	       "cur_trans->state=%d", cur_trans->state);
2195
2196	list_add(&trans->pending_snapshot->list, &cur_trans->pending_snapshots);
2197}
2198
2199static void update_commit_stats(struct btrfs_fs_info *fs_info)
2200{
2201	ktime_t now = ktime_get_ns();
2202	ktime_t interval = now - fs_info->commit_stats.critical_section_start_time;
2203
2204	ASSERT(fs_info->commit_stats.critical_section_start_time);
2205
2206	fs_info->commit_stats.commit_count++;
2207	fs_info->commit_stats.last_commit_dur = interval;
2208	fs_info->commit_stats.max_commit_dur =
2209			max_t(u64, fs_info->commit_stats.max_commit_dur, interval);
2210	fs_info->commit_stats.total_commit_dur += interval;
2211	fs_info->commit_stats.critical_section_start_time = 0;
2212}
2213
2214int btrfs_commit_transaction(struct btrfs_trans_handle *trans)
2215{
2216	struct btrfs_fs_info *fs_info = trans->fs_info;
2217	struct btrfs_transaction *cur_trans = trans->transaction;
2218	struct btrfs_transaction *prev_trans = NULL;
2219	int ret;
2220
2221	ASSERT(refcount_read(&trans->use_count) == 1,
2222	       "refcount_read(&trans->use_count)=%d", refcount_read(&trans->use_count));
2223	btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_PREP);
2224
2225	clear_bit(BTRFS_FS_NEED_TRANS_COMMIT, &fs_info->flags);
2226
2227	/* Stop the commit early if ->aborted is set */
2228	if (TRANS_ABORTED(cur_trans)) {
2229		ret = cur_trans->aborted;
2230		goto lockdep_trans_commit_start_release;
2231	}
2232
2233	btrfs_trans_release_metadata(trans);
2234	trans->block_rsv = NULL;
2235
2236	/*
2237	 * We only want one transaction commit doing the flushing so we do not
2238	 * waste a bunch of time on lock contention on the extent root node.
2239	 */
2240	if (!test_and_set_bit(BTRFS_DELAYED_REFS_FLUSHING,
2241			      &cur_trans->delayed_refs.flags)) {
2242		/*
2243		 * Make a pass through all the delayed refs we have so far.
2244		 * Any running threads may add more while we are here.
2245		 */
2246		ret = btrfs_run_delayed_refs(trans, 0);
2247		if (ret)
2248			goto lockdep_trans_commit_start_release;
2249	}
2250
2251	btrfs_create_pending_block_groups(trans);
2252
2253	if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) {
2254		int run_it = 0;
2255
2256		/* this mutex is also taken before trying to set
2257		 * block groups readonly.  We need to make sure
2258		 * that nobody has set a block group readonly
2259		 * after a extents from that block group have been
2260		 * allocated for cache files.  btrfs_set_block_group_ro
2261		 * will wait for the transaction to commit if it
2262		 * finds BTRFS_TRANS_DIRTY_BG_RUN set.
2263		 *
2264		 * The BTRFS_TRANS_DIRTY_BG_RUN flag is also used to make sure
2265		 * only one process starts all the block group IO.  It wouldn't
2266		 * hurt to have more than one go through, but there's no
2267		 * real advantage to it either.
2268		 */
2269		mutex_lock(&fs_info->ro_block_group_mutex);
2270		if (!test_and_set_bit(BTRFS_TRANS_DIRTY_BG_RUN,
2271				      &cur_trans->flags))
2272			run_it = 1;
2273		mutex_unlock(&fs_info->ro_block_group_mutex);
2274
2275		if (run_it) {
2276			ret = btrfs_start_dirty_block_groups(trans);
2277			if (ret)
2278				goto lockdep_trans_commit_start_release;
2279		}
2280	}
2281
2282	spin_lock(&fs_info->trans_lock);
2283	if (cur_trans->state >= TRANS_STATE_COMMIT_PREP) {
2284		enum btrfs_trans_state want_state = TRANS_STATE_COMPLETED;
2285
2286		add_pending_snapshot(trans);
2287
2288		spin_unlock(&fs_info->trans_lock);
2289		refcount_inc(&cur_trans->use_count);
2290
2291		if (trans->in_fsync)
2292			want_state = TRANS_STATE_SUPER_COMMITTED;
2293
2294		btrfs_trans_state_lockdep_release(fs_info,
2295						  BTRFS_LOCKDEP_TRANS_COMMIT_PREP);
2296		ret = btrfs_end_transaction(trans);
2297		wait_for_commit(cur_trans, want_state);
2298
2299		if (TRANS_ABORTED(cur_trans))
2300			ret = cur_trans->aborted;
2301
2302		btrfs_put_transaction(cur_trans);
2303
2304		return ret;
2305	}
2306
2307	cur_trans->state = TRANS_STATE_COMMIT_PREP;
2308	wake_up(&fs_info->transaction_blocked_wait);
2309	btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_PREP);
2310
2311	if (!list_is_first(&cur_trans->list, &fs_info->trans_list)) {
2312		enum btrfs_trans_state want_state = TRANS_STATE_COMPLETED;
2313
2314		if (trans->in_fsync)
2315			want_state = TRANS_STATE_SUPER_COMMITTED;
2316
2317		prev_trans = list_prev_entry(cur_trans, list);
2318		if (prev_trans->state < want_state) {
2319			refcount_inc(&prev_trans->use_count);
2320			spin_unlock(&fs_info->trans_lock);
2321
2322			wait_for_commit(prev_trans, want_state);
2323
2324			ret = READ_ONCE(prev_trans->aborted);
2325
2326			btrfs_put_transaction(prev_trans);
2327			if (ret)
2328				goto lockdep_release;
2329			spin_lock(&fs_info->trans_lock);
2330		}
2331	} else {
2332		/*
2333		 * The previous transaction was aborted and was already removed
2334		 * from the list of transactions at fs_info->trans_list. So we
2335		 * abort to prevent writing a new superblock that reflects a
2336		 * corrupt state (pointing to trees with unwritten nodes/leafs).
2337		 */
2338		if (BTRFS_FS_ERROR(fs_info)) {
2339			spin_unlock(&fs_info->trans_lock);
2340			ret = -EROFS;
2341			goto lockdep_release;
2342		}
2343	}
2344
2345	cur_trans->state = TRANS_STATE_COMMIT_START;
2346	wake_up(&fs_info->transaction_blocked_wait);
2347	spin_unlock(&fs_info->trans_lock);
2348
2349	/*
2350	 * Get the time spent on the work done by the commit thread and not
2351	 * the time spent waiting on a previous commit
2352	 */
2353	fs_info->commit_stats.critical_section_start_time = ktime_get_ns();
2354	extwriter_counter_dec(cur_trans, trans->type);
2355
2356	ret = btrfs_start_delalloc_flush(fs_info);
2357	if (ret)
2358		goto lockdep_release;
2359
2360	ret = btrfs_run_delayed_items(trans);
2361	if (ret)
2362		goto lockdep_release;
2363
2364	/*
2365	 * The thread has started/joined the transaction thus it holds the
2366	 * lockdep map as a reader. It has to release it before acquiring the
2367	 * lockdep map as a writer.
2368	 */
2369	btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters);
2370	btrfs_might_wait_for_event(fs_info, btrfs_trans_num_extwriters);
2371	wait_event(cur_trans->writer_wait,
2372		   extwriter_counter_read(cur_trans) == 0);
2373
2374	/* some pending stuffs might be added after the previous flush. */
2375	ret = btrfs_run_delayed_items(trans);
2376	if (ret) {
2377		btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
2378		goto cleanup_transaction;
2379	}
2380
2381	btrfs_wait_delalloc_flush(fs_info);
2382
2383	/*
2384	 * Wait for all ordered extents started by a fast fsync that joined this
2385	 * transaction. Otherwise if this transaction commits before the ordered
2386	 * extents complete we lose logged data after a power failure.
2387	 */
2388	btrfs_might_wait_for_event(fs_info, btrfs_trans_pending_ordered);
2389	wait_event(cur_trans->pending_wait,
2390		   atomic_read(&cur_trans->pending_ordered) == 0);
2391
2392	btrfs_scrub_pause(fs_info);
2393	/*
2394	 * Ok now we need to make sure to block out any other joins while we
2395	 * commit the transaction.  We could have started a join before setting
2396	 * COMMIT_DOING so make sure to wait for num_writers to == 1 again.
2397	 */
2398	spin_lock(&fs_info->trans_lock);
2399	add_pending_snapshot(trans);
2400	cur_trans->state = TRANS_STATE_COMMIT_DOING;
2401	spin_unlock(&fs_info->trans_lock);
2402
2403	/*
2404	 * The thread has started/joined the transaction thus it holds the
2405	 * lockdep map as a reader. It has to release it before acquiring the
2406	 * lockdep map as a writer.
2407	 */
2408	btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
2409	btrfs_might_wait_for_event(fs_info, btrfs_trans_num_writers);
2410	wait_event(cur_trans->writer_wait,
2411		   atomic_read(&cur_trans->num_writers) == 1);
2412
2413	/*
2414	 * Make lockdep happy by acquiring the state locks after
2415	 * btrfs_trans_num_writers is released. If we acquired the state locks
2416	 * before releasing the btrfs_trans_num_writers lock then lockdep would
2417	 * complain because we did not follow the reverse order unlocking rule.
2418	 */
2419	btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED);
2420	btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED);
2421	btrfs_trans_state_lockdep_acquire(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
2422
2423	/*
2424	 * We've started the commit, clear the flag in case we were triggered to
2425	 * do an async commit but somebody else started before the transaction
2426	 * kthread could do the work.
2427	 */
2428	clear_bit(BTRFS_FS_COMMIT_TRANS, &fs_info->flags);
2429
2430	if (TRANS_ABORTED(cur_trans)) {
2431		ret = cur_trans->aborted;
2432		btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
2433		goto scrub_continue;
2434	}
2435	/*
2436	 * the reloc mutex makes sure that we stop
2437	 * the balancing code from coming in and moving
2438	 * extents around in the middle of the commit
2439	 */
2440	mutex_lock(&fs_info->reloc_mutex);
2441
2442	/*
2443	 * We needn't worry about the delayed items because we will
2444	 * deal with them in create_pending_snapshot(), which is the
2445	 * core function of the snapshot creation.
2446	 */
2447	ret = create_pending_snapshots(trans);
2448	if (ret)
2449		goto unlock_reloc;
2450
2451	/*
2452	 * We insert the dir indexes of the snapshots and update the inode
2453	 * of the snapshots' parents after the snapshot creation, so there
2454	 * are some delayed items which are not dealt with. Now deal with
2455	 * them.
2456	 *
2457	 * We needn't worry that this operation will corrupt the snapshots,
2458	 * because all the tree which are snapshotted will be forced to COW
2459	 * the nodes and leaves.
2460	 */
2461	ret = btrfs_run_delayed_items(trans);
2462	if (ret)
2463		goto unlock_reloc;
2464
2465	ret = btrfs_run_delayed_refs(trans, U64_MAX);
2466	if (ret)
2467		goto unlock_reloc;
2468
2469	/*
2470	 * make sure none of the code above managed to slip in a
2471	 * delayed item
2472	 */
2473	btrfs_assert_delayed_root_empty(fs_info);
2474
2475	WARN_ON(cur_trans != trans->transaction);
2476
2477	ret = commit_fs_roots(trans);
2478	if (ret)
2479		goto unlock_reloc;
2480
2481	/* commit_fs_roots gets rid of all the tree log roots, it is now
2482	 * safe to free the root of tree log roots
2483	 */
2484	btrfs_free_log_root_tree(trans, fs_info);
2485
2486	/*
2487	 * Since fs roots are all committed, we can get a quite accurate
2488	 * new_roots. So let's do quota accounting.
2489	 */
2490	ret = btrfs_qgroup_account_extents(trans);
2491	if (ret < 0)
2492		goto unlock_reloc;
2493
2494	ret = commit_cowonly_roots(trans);
2495	if (ret)
2496		goto unlock_reloc;
2497
2498	/*
2499	 * The tasks which save the space cache and inode cache may also
2500	 * update ->aborted, check it.
2501	 */
2502	if (TRANS_ABORTED(cur_trans)) {
2503		ret = cur_trans->aborted;
2504		goto unlock_reloc;
2505	}
2506
2507	cur_trans = fs_info->running_transaction;
2508
2509	btrfs_set_root_node(&fs_info->tree_root->root_item,
2510			    fs_info->tree_root->node);
2511	list_add_tail(&fs_info->tree_root->dirty_list,
2512		      &cur_trans->switch_commits);
2513
2514	btrfs_set_root_node(&fs_info->chunk_root->root_item,
2515			    fs_info->chunk_root->node);
2516	list_add_tail(&fs_info->chunk_root->dirty_list,
2517		      &cur_trans->switch_commits);
2518
2519	switch_commit_roots(trans);
2520
2521	ASSERT(list_empty(&cur_trans->dirty_bgs));
2522	ASSERT(list_empty(&cur_trans->io_bgs));
2523	update_super_roots(fs_info);
2524
2525	btrfs_set_super_log_root(fs_info->super_copy, 0);
2526	btrfs_set_super_log_root_level(fs_info->super_copy, 0);
2527	memcpy(fs_info->super_for_commit, fs_info->super_copy,
2528	       sizeof(*fs_info->super_copy));
2529
2530	btrfs_commit_device_sizes(cur_trans);
2531
2532	clear_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
2533	clear_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
2534
2535	btrfs_trans_release_chunk_metadata(trans);
2536
2537	/*
2538	 * Before changing the transaction state to TRANS_STATE_UNBLOCKED and
2539	 * setting fs_info->running_transaction to NULL, lock tree_log_mutex to
2540	 * make sure that before we commit our superblock, no other task can
2541	 * start a new transaction and commit a log tree before we commit our
2542	 * superblock. Anyone trying to commit a log tree locks this mutex before
2543	 * writing its superblock.
2544	 */
2545	mutex_lock(&fs_info->tree_log_mutex);
2546
2547	spin_lock(&fs_info->trans_lock);
2548	cur_trans->state = TRANS_STATE_UNBLOCKED;
2549	fs_info->running_transaction = NULL;
2550	spin_unlock(&fs_info->trans_lock);
2551	mutex_unlock(&fs_info->reloc_mutex);
2552
2553	wake_up(&fs_info->transaction_wait);
2554	btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
2555
2556	/* If we have features changed, wake up the cleaner to update sysfs. */
2557	if (test_bit(BTRFS_FS_FEATURE_CHANGED, &fs_info->flags) &&
2558	    fs_info->cleaner_kthread)
2559		wake_up_process(fs_info->cleaner_kthread);
2560
2561	ret = btrfs_write_and_wait_transaction(trans);
2562	if (ret) {
2563		btrfs_handle_fs_error(fs_info, ret,
2564				      "Error while writing out transaction");
2565		mutex_unlock(&fs_info->tree_log_mutex);
2566		goto scrub_continue;
2567	}
2568
2569	ret = write_all_supers(fs_info, 0);
2570	/*
2571	 * the super is written, we can safely allow the tree-loggers
2572	 * to go about their business
2573	 */
2574	mutex_unlock(&fs_info->tree_log_mutex);
2575	if (ret)
2576		goto scrub_continue;
2577
2578	update_commit_stats(fs_info);
2579	/*
2580	 * We needn't acquire the lock here because there is no other task
2581	 * which can change it.
2582	 */
2583	cur_trans->state = TRANS_STATE_SUPER_COMMITTED;
2584	wake_up(&cur_trans->commit_wait);
2585	btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED);
2586
2587	ret = btrfs_finish_extent_commit(trans);
2588	if (ret)
2589		goto scrub_continue;
2590
2591	if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &cur_trans->flags))
2592		btrfs_clear_space_info_full(fs_info);
2593
2594	btrfs_set_last_trans_committed(fs_info, cur_trans->transid);
2595	/*
2596	 * We needn't acquire the lock here because there is no other task
2597	 * which can change it.
2598	 */
2599	cur_trans->state = TRANS_STATE_COMPLETED;
2600	wake_up(&cur_trans->commit_wait);
2601	btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED);
2602
2603	spin_lock(&fs_info->trans_lock);
2604	list_del_init(&cur_trans->list);
2605	spin_unlock(&fs_info->trans_lock);
2606
2607	btrfs_put_transaction(cur_trans);
2608	btrfs_put_transaction(cur_trans);
2609
2610	if (trans->type & __TRANS_FREEZABLE)
2611		sb_end_intwrite(fs_info->sb);
2612
2613	trace_btrfs_transaction_commit(fs_info);
2614
2615	btrfs_scrub_continue(fs_info);
2616
2617	if (current->journal_info == trans)
2618		current->journal_info = NULL;
2619
2620	kmem_cache_free(btrfs_trans_handle_cachep, trans);
2621
2622	return ret;
2623
2624unlock_reloc:
2625	mutex_unlock(&fs_info->reloc_mutex);
2626	btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_UNBLOCKED);
2627scrub_continue:
2628	btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_SUPER_COMMITTED);
2629	btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMPLETED);
2630	btrfs_scrub_continue(fs_info);
2631cleanup_transaction:
2632	btrfs_trans_release_metadata(trans);
2633	btrfs_cleanup_pending_block_groups(trans);
2634	btrfs_trans_release_chunk_metadata(trans);
2635	trans->block_rsv = NULL;
2636	btrfs_warn(fs_info, "Skipping commit of aborted transaction.");
2637	if (current->journal_info == trans)
2638		current->journal_info = NULL;
2639	cleanup_transaction(trans, ret);
2640
2641	return ret;
2642
2643lockdep_release:
2644	btrfs_lockdep_release(fs_info, btrfs_trans_num_extwriters);
2645	btrfs_lockdep_release(fs_info, btrfs_trans_num_writers);
2646	goto cleanup_transaction;
2647
2648lockdep_trans_commit_start_release:
2649	btrfs_trans_state_lockdep_release(fs_info, BTRFS_LOCKDEP_TRANS_COMMIT_PREP);
2650	btrfs_end_transaction(trans);
2651	return ret;
2652}
2653
2654/*
2655 * return < 0 if error
2656 * 0 if there are no more dead_roots at the time of call
2657 * 1 there are more to be processed, call me again
2658 *
2659 * The return value indicates there are certainly more snapshots to delete, but
2660 * if there comes a new one during processing, it may return 0. We don't mind,
2661 * because btrfs_commit_super will poke cleaner thread and it will process it a
2662 * few seconds later.
2663 */
2664int btrfs_clean_one_deleted_snapshot(struct btrfs_fs_info *fs_info)
2665{
2666	struct btrfs_root *root;
2667	int ret;
2668
2669	spin_lock(&fs_info->trans_lock);
2670	if (list_empty(&fs_info->dead_roots)) {
2671		spin_unlock(&fs_info->trans_lock);
2672		return 0;
2673	}
2674	root = list_first_entry(&fs_info->dead_roots,
2675			struct btrfs_root, root_list);
2676	list_del_init(&root->root_list);
2677	spin_unlock(&fs_info->trans_lock);
2678
2679	btrfs_debug(fs_info, "cleaner removing %llu", btrfs_root_id(root));
2680
2681	btrfs_kill_all_delayed_nodes(root);
2682
2683	if (btrfs_header_backref_rev(root->node) <
2684			BTRFS_MIXED_BACKREF_REV)
2685		ret = btrfs_drop_snapshot(root, false, false);
2686	else
2687		ret = btrfs_drop_snapshot(root, true, false);
2688
2689	btrfs_put_root(root);
2690	return (ret < 0) ? 0 : 1;
2691}
2692
2693/*
2694 * We only mark the transaction aborted and then set the file system read-only.
2695 * This will prevent new transactions from starting or trying to join this
2696 * one.
2697 *
2698 * This means that error recovery at the call site is limited to freeing
2699 * any local memory allocations and passing the error code up without
2700 * further cleanup. The transaction should complete as it normally would
2701 * in the call path but will return -EIO.
2702 *
2703 * We'll complete the cleanup in btrfs_end_transaction and
2704 * btrfs_commit_transaction.
2705 */
2706void __cold __btrfs_abort_transaction(struct btrfs_trans_handle *trans,
2707				      const char *function,
2708				      unsigned int line, int error, bool first_hit)
2709{
2710	struct btrfs_fs_info *fs_info = trans->fs_info;
2711
2712	WRITE_ONCE(trans->aborted, error);
2713	WRITE_ONCE(trans->transaction->aborted, error);
2714	if (first_hit && error == -ENOSPC)
2715		btrfs_dump_space_info_for_trans_abort(fs_info);
2716	/* Wake up anybody who may be waiting on this transaction */
2717	wake_up(&fs_info->transaction_wait);
2718	wake_up(&fs_info->transaction_blocked_wait);
2719	__btrfs_handle_fs_error(fs_info, function, line, error, NULL);
2720}
2721
2722int __init btrfs_transaction_init(void)
2723{
2724	btrfs_trans_handle_cachep = KMEM_CACHE(btrfs_trans_handle, SLAB_TEMPORARY);
2725	if (!btrfs_trans_handle_cachep)
2726		return -ENOMEM;
2727	return 0;
2728}
2729
2730void __cold btrfs_transaction_exit(void)
2731{
2732	kmem_cache_destroy(btrfs_trans_handle_cachep);
2733}