1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Copyright (C) 2011 Fujitsu.  All rights reserved.
   4 * Written by Miao Xie <miaox@cn.fujitsu.com>
   5 */
   6
   7#include <linux/slab.h>
   8#include <linux/iversion.h>
   9#include "ctree.h"
  10#include "fs.h"
  11#include "messages.h"
  12#include "misc.h"
  13#include "delayed-inode.h"
  14#include "disk-io.h"
  15#include "transaction.h"
  16#include "qgroup.h"
  17#include "locking.h"
  18#include "inode-item.h"
  19#include "space-info.h"
  20#include "accessors.h"
  21#include "file-item.h"
  22
  23#define BTRFS_DELAYED_WRITEBACK		512
  24#define BTRFS_DELAYED_BACKGROUND	128
  25#define BTRFS_DELAYED_BATCH		16
  26
  27static struct kmem_cache *delayed_node_cache;
  28
  29int __init btrfs_delayed_inode_init(void)
  30{
  31	delayed_node_cache = kmem_cache_create("btrfs_delayed_node",
  32					sizeof(struct btrfs_delayed_node),
  33					0,
  34					SLAB_MEM_SPREAD,
  35					NULL);
  36	if (!delayed_node_cache)
  37		return -ENOMEM;
  38	return 0;
  39}
  40
  41void __cold btrfs_delayed_inode_exit(void)
  42{
  43	kmem_cache_destroy(delayed_node_cache);
  44}
  45
 
 
 
 
 
 
 
 
 
 
 
  46static inline void btrfs_init_delayed_node(
  47				struct btrfs_delayed_node *delayed_node,
  48				struct btrfs_root *root, u64 inode_id)
  49{
  50	delayed_node->root = root;
  51	delayed_node->inode_id = inode_id;
  52	refcount_set(&delayed_node->refs, 0);
  53	delayed_node->ins_root = RB_ROOT_CACHED;
  54	delayed_node->del_root = RB_ROOT_CACHED;
  55	mutex_init(&delayed_node->mutex);
  56	INIT_LIST_HEAD(&delayed_node->n_list);
  57	INIT_LIST_HEAD(&delayed_node->p_list);
  58}
  59
  60static struct btrfs_delayed_node *btrfs_get_delayed_node(
  61		struct btrfs_inode *btrfs_inode)
  62{
  63	struct btrfs_root *root = btrfs_inode->root;
  64	u64 ino = btrfs_ino(btrfs_inode);
  65	struct btrfs_delayed_node *node;
  66
  67	node = READ_ONCE(btrfs_inode->delayed_node);
  68	if (node) {
  69		refcount_inc(&node->refs);
  70		return node;
  71	}
  72
  73	spin_lock(&root->inode_lock);
  74	node = xa_load(&root->delayed_nodes, ino);
  75
  76	if (node) {
  77		if (btrfs_inode->delayed_node) {
  78			refcount_inc(&node->refs);	/* can be accessed */
  79			BUG_ON(btrfs_inode->delayed_node != node);
  80			spin_unlock(&root->inode_lock);
  81			return node;
  82		}
  83
  84		/*
  85		 * It's possible that we're racing into the middle of removing
  86		 * this node from the xarray.  In this case, the refcount
  87		 * was zero and it should never go back to one.  Just return
  88		 * NULL like it was never in the xarray at all; our release
  89		 * function is in the process of removing it.
  90		 *
  91		 * Some implementations of refcount_inc refuse to bump the
  92		 * refcount once it has hit zero.  If we don't do this dance
  93		 * here, refcount_inc() may decide to just WARN_ONCE() instead
  94		 * of actually bumping the refcount.
  95		 *
  96		 * If this node is properly in the xarray, we want to bump the
  97		 * refcount twice, once for the inode and once for this get
  98		 * operation.
  99		 */
 100		if (refcount_inc_not_zero(&node->refs)) {
 101			refcount_inc(&node->refs);
 102			btrfs_inode->delayed_node = node;
 103		} else {
 104			node = NULL;
 105		}
 106
 107		spin_unlock(&root->inode_lock);
 108		return node;
 109	}
 110	spin_unlock(&root->inode_lock);
 111
 112	return NULL;
 113}
 114
 115/* Will return either the node or PTR_ERR(-ENOMEM) */
 116static struct btrfs_delayed_node *btrfs_get_or_create_delayed_node(
 117		struct btrfs_inode *btrfs_inode)
 118{
 119	struct btrfs_delayed_node *node;
 120	struct btrfs_root *root = btrfs_inode->root;
 121	u64 ino = btrfs_ino(btrfs_inode);
 122	int ret;
 123	void *ptr;
 124
 125again:
 126	node = btrfs_get_delayed_node(btrfs_inode);
 127	if (node)
 128		return node;
 129
 130	node = kmem_cache_zalloc(delayed_node_cache, GFP_NOFS);
 131	if (!node)
 132		return ERR_PTR(-ENOMEM);
 133	btrfs_init_delayed_node(node, root, ino);
 134
 135	/* Cached in the inode and can be accessed. */
 136	refcount_set(&node->refs, 2);
 137
 138	/* Allocate and reserve the slot, from now it can return a NULL from xa_load(). */
 139	ret = xa_reserve(&root->delayed_nodes, ino, GFP_NOFS);
 140	if (ret == -ENOMEM) {
 141		kmem_cache_free(delayed_node_cache, node);
 142		return ERR_PTR(-ENOMEM);
 143	}
 144	spin_lock(&root->inode_lock);
 145	ptr = xa_load(&root->delayed_nodes, ino);
 146	if (ptr) {
 147		/* Somebody inserted it, go back and read it. */
 148		spin_unlock(&root->inode_lock);
 149		kmem_cache_free(delayed_node_cache, node);
 150		node = NULL;
 151		goto again;
 152	}
 153	ptr = xa_store(&root->delayed_nodes, ino, node, GFP_ATOMIC);
 154	ASSERT(xa_err(ptr) != -EINVAL);
 155	ASSERT(xa_err(ptr) != -ENOMEM);
 156	ASSERT(ptr == NULL);
 157	btrfs_inode->delayed_node = node;
 158	spin_unlock(&root->inode_lock);
 159
 160	return node;
 161}
 162
 163/*
 164 * Call it when holding delayed_node->mutex
 165 *
 166 * If mod = 1, add this node into the prepared list.
 167 */
 168static void btrfs_queue_delayed_node(struct btrfs_delayed_root *root,
 169				     struct btrfs_delayed_node *node,
 170				     int mod)
 171{
 172	spin_lock(&root->lock);
 173	if (test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
 174		if (!list_empty(&node->p_list))
 175			list_move_tail(&node->p_list, &root->prepare_list);
 176		else if (mod)
 177			list_add_tail(&node->p_list, &root->prepare_list);
 178	} else {
 179		list_add_tail(&node->n_list, &root->node_list);
 180		list_add_tail(&node->p_list, &root->prepare_list);
 181		refcount_inc(&node->refs);	/* inserted into list */
 182		root->nodes++;
 183		set_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags);
 184	}
 185	spin_unlock(&root->lock);
 186}
 187
 188/* Call it when holding delayed_node->mutex */
 189static void btrfs_dequeue_delayed_node(struct btrfs_delayed_root *root,
 190				       struct btrfs_delayed_node *node)
 191{
 192	spin_lock(&root->lock);
 193	if (test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
 194		root->nodes--;
 195		refcount_dec(&node->refs);	/* not in the list */
 196		list_del_init(&node->n_list);
 197		if (!list_empty(&node->p_list))
 198			list_del_init(&node->p_list);
 199		clear_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags);
 200	}
 201	spin_unlock(&root->lock);
 202}
 203
 204static struct btrfs_delayed_node *btrfs_first_delayed_node(
 205			struct btrfs_delayed_root *delayed_root)
 206{
 207	struct list_head *p;
 208	struct btrfs_delayed_node *node = NULL;
 209
 210	spin_lock(&delayed_root->lock);
 211	if (list_empty(&delayed_root->node_list))
 212		goto out;
 213
 214	p = delayed_root->node_list.next;
 215	node = list_entry(p, struct btrfs_delayed_node, n_list);
 216	refcount_inc(&node->refs);
 217out:
 218	spin_unlock(&delayed_root->lock);
 219
 220	return node;
 221}
 222
 223static struct btrfs_delayed_node *btrfs_next_delayed_node(
 224						struct btrfs_delayed_node *node)
 225{
 226	struct btrfs_delayed_root *delayed_root;
 227	struct list_head *p;
 228	struct btrfs_delayed_node *next = NULL;
 229
 230	delayed_root = node->root->fs_info->delayed_root;
 231	spin_lock(&delayed_root->lock);
 232	if (!test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
 233		/* not in the list */
 234		if (list_empty(&delayed_root->node_list))
 235			goto out;
 236		p = delayed_root->node_list.next;
 237	} else if (list_is_last(&node->n_list, &delayed_root->node_list))
 238		goto out;
 239	else
 240		p = node->n_list.next;
 241
 242	next = list_entry(p, struct btrfs_delayed_node, n_list);
 243	refcount_inc(&next->refs);
 244out:
 245	spin_unlock(&delayed_root->lock);
 246
 247	return next;
 248}
 249
 250static void __btrfs_release_delayed_node(
 251				struct btrfs_delayed_node *delayed_node,
 252				int mod)
 253{
 254	struct btrfs_delayed_root *delayed_root;
 255
 256	if (!delayed_node)
 257		return;
 258
 259	delayed_root = delayed_node->root->fs_info->delayed_root;
 260
 261	mutex_lock(&delayed_node->mutex);
 262	if (delayed_node->count)
 263		btrfs_queue_delayed_node(delayed_root, delayed_node, mod);
 264	else
 265		btrfs_dequeue_delayed_node(delayed_root, delayed_node);
 266	mutex_unlock(&delayed_node->mutex);
 267
 268	if (refcount_dec_and_test(&delayed_node->refs)) {
 269		struct btrfs_root *root = delayed_node->root;
 270
 271		spin_lock(&root->inode_lock);
 272		/*
 273		 * Once our refcount goes to zero, nobody is allowed to bump it
 274		 * back up.  We can delete it now.
 275		 */
 276		ASSERT(refcount_read(&delayed_node->refs) == 0);
 277		xa_erase(&root->delayed_nodes, delayed_node->inode_id);
 278		spin_unlock(&root->inode_lock);
 279		kmem_cache_free(delayed_node_cache, delayed_node);
 280	}
 281}
 282
 283static inline void btrfs_release_delayed_node(struct btrfs_delayed_node *node)
 284{
 285	__btrfs_release_delayed_node(node, 0);
 286}
 287
 288static struct btrfs_delayed_node *btrfs_first_prepared_delayed_node(
 289					struct btrfs_delayed_root *delayed_root)
 290{
 291	struct list_head *p;
 292	struct btrfs_delayed_node *node = NULL;
 293
 294	spin_lock(&delayed_root->lock);
 295	if (list_empty(&delayed_root->prepare_list))
 296		goto out;
 297
 298	p = delayed_root->prepare_list.next;
 299	list_del_init(p);
 300	node = list_entry(p, struct btrfs_delayed_node, p_list);
 301	refcount_inc(&node->refs);
 302out:
 303	spin_unlock(&delayed_root->lock);
 304
 305	return node;
 306}
 307
 308static inline void btrfs_release_prepared_delayed_node(
 309					struct btrfs_delayed_node *node)
 310{
 311	__btrfs_release_delayed_node(node, 1);
 312}
 313
 314static struct btrfs_delayed_item *btrfs_alloc_delayed_item(u16 data_len,
 315					   struct btrfs_delayed_node *node,
 316					   enum btrfs_delayed_item_type type)
 317{
 318	struct btrfs_delayed_item *item;
 319
 320	item = kmalloc(struct_size(item, data, data_len), GFP_NOFS);
 321	if (item) {
 322		item->data_len = data_len;
 323		item->type = type;
 324		item->bytes_reserved = 0;
 325		item->delayed_node = node;
 326		RB_CLEAR_NODE(&item->rb_node);
 327		INIT_LIST_HEAD(&item->log_list);
 328		item->logged = false;
 329		refcount_set(&item->refs, 1);
 330	}
 331	return item;
 332}
 333
 334/*
 335 * Look up the delayed item by key.
 336 *
 337 * @delayed_node: pointer to the delayed node
 338 * @index:	  the dir index value to lookup (offset of a dir index key)
 339 *
 340 * Note: if we don't find the right item, we will return the prev item and
 341 * the next item.
 342 */
 343static struct btrfs_delayed_item *__btrfs_lookup_delayed_item(
 344				struct rb_root *root,
 345				u64 index)
 346{
 347	struct rb_node *node = root->rb_node;
 348	struct btrfs_delayed_item *delayed_item = NULL;
 349
 350	while (node) {
 351		delayed_item = rb_entry(node, struct btrfs_delayed_item,
 352					rb_node);
 353		if (delayed_item->index < index)
 354			node = node->rb_right;
 355		else if (delayed_item->index > index)
 356			node = node->rb_left;
 357		else
 358			return delayed_item;
 359	}
 360
 361	return NULL;
 362}
 363
 364static int __btrfs_add_delayed_item(struct btrfs_delayed_node *delayed_node,
 365				    struct btrfs_delayed_item *ins)
 366{
 367	struct rb_node **p, *node;
 368	struct rb_node *parent_node = NULL;
 369	struct rb_root_cached *root;
 370	struct btrfs_delayed_item *item;
 371	bool leftmost = true;
 372
 373	if (ins->type == BTRFS_DELAYED_INSERTION_ITEM)
 374		root = &delayed_node->ins_root;
 375	else
 376		root = &delayed_node->del_root;
 377
 378	p = &root->rb_root.rb_node;
 379	node = &ins->rb_node;
 380
 381	while (*p) {
 382		parent_node = *p;
 383		item = rb_entry(parent_node, struct btrfs_delayed_item,
 384				 rb_node);
 385
 386		if (item->index < ins->index) {
 387			p = &(*p)->rb_right;
 388			leftmost = false;
 389		} else if (item->index > ins->index) {
 390			p = &(*p)->rb_left;
 391		} else {
 392			return -EEXIST;
 393		}
 394	}
 395
 396	rb_link_node(node, parent_node, p);
 397	rb_insert_color_cached(node, root, leftmost);
 398
 399	if (ins->type == BTRFS_DELAYED_INSERTION_ITEM &&
 400	    ins->index >= delayed_node->index_cnt)
 401		delayed_node->index_cnt = ins->index + 1;
 402
 403	delayed_node->count++;
 404	atomic_inc(&delayed_node->root->fs_info->delayed_root->items);
 405	return 0;
 406}
 407
 408static void finish_one_item(struct btrfs_delayed_root *delayed_root)
 409{
 410	int seq = atomic_inc_return(&delayed_root->items_seq);
 411
 412	/* atomic_dec_return implies a barrier */
 413	if ((atomic_dec_return(&delayed_root->items) <
 414	    BTRFS_DELAYED_BACKGROUND || seq % BTRFS_DELAYED_BATCH == 0))
 415		cond_wake_up_nomb(&delayed_root->wait);
 416}
 417
 418static void __btrfs_remove_delayed_item(struct btrfs_delayed_item *delayed_item)
 419{
 420	struct btrfs_delayed_node *delayed_node = delayed_item->delayed_node;
 421	struct rb_root_cached *root;
 422	struct btrfs_delayed_root *delayed_root;
 423
 424	/* Not inserted, ignore it. */
 425	if (RB_EMPTY_NODE(&delayed_item->rb_node))
 426		return;
 427
 428	/* If it's in a rbtree, then we need to have delayed node locked. */
 429	lockdep_assert_held(&delayed_node->mutex);
 430
 431	delayed_root = delayed_node->root->fs_info->delayed_root;
 432
 433	BUG_ON(!delayed_root);
 434
 435	if (delayed_item->type == BTRFS_DELAYED_INSERTION_ITEM)
 436		root = &delayed_node->ins_root;
 437	else
 438		root = &delayed_node->del_root;
 439
 440	rb_erase_cached(&delayed_item->rb_node, root);
 441	RB_CLEAR_NODE(&delayed_item->rb_node);
 442	delayed_node->count--;
 443
 444	finish_one_item(delayed_root);
 445}
 446
 447static void btrfs_release_delayed_item(struct btrfs_delayed_item *item)
 448{
 449	if (item) {
 450		__btrfs_remove_delayed_item(item);
 451		if (refcount_dec_and_test(&item->refs))
 452			kfree(item);
 453	}
 454}
 455
 456static struct btrfs_delayed_item *__btrfs_first_delayed_insertion_item(
 457					struct btrfs_delayed_node *delayed_node)
 458{
 459	struct rb_node *p;
 460	struct btrfs_delayed_item *item = NULL;
 461
 462	p = rb_first_cached(&delayed_node->ins_root);
 463	if (p)
 464		item = rb_entry(p, struct btrfs_delayed_item, rb_node);
 465
 466	return item;
 467}
 468
 469static struct btrfs_delayed_item *__btrfs_first_delayed_deletion_item(
 470					struct btrfs_delayed_node *delayed_node)
 471{
 472	struct rb_node *p;
 473	struct btrfs_delayed_item *item = NULL;
 474
 475	p = rb_first_cached(&delayed_node->del_root);
 476	if (p)
 477		item = rb_entry(p, struct btrfs_delayed_item, rb_node);
 478
 479	return item;
 480}
 481
 482static struct btrfs_delayed_item *__btrfs_next_delayed_item(
 483						struct btrfs_delayed_item *item)
 484{
 485	struct rb_node *p;
 486	struct btrfs_delayed_item *next = NULL;
 487
 488	p = rb_next(&item->rb_node);
 489	if (p)
 490		next = rb_entry(p, struct btrfs_delayed_item, rb_node);
 491
 492	return next;
 493}
 494
 495static int btrfs_delayed_item_reserve_metadata(struct btrfs_trans_handle *trans,
 496					       struct btrfs_delayed_item *item)
 497{
 498	struct btrfs_block_rsv *src_rsv;
 499	struct btrfs_block_rsv *dst_rsv;
 500	struct btrfs_fs_info *fs_info = trans->fs_info;
 501	u64 num_bytes;
 502	int ret;
 503
 504	if (!trans->bytes_reserved)
 505		return 0;
 506
 507	src_rsv = trans->block_rsv;
 508	dst_rsv = &fs_info->delayed_block_rsv;
 509
 510	num_bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
 511
 512	/*
 513	 * Here we migrate space rsv from transaction rsv, since have already
 514	 * reserved space when starting a transaction.  So no need to reserve
 515	 * qgroup space here.
 516	 */
 517	ret = btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, true);
 518	if (!ret) {
 519		trace_btrfs_space_reservation(fs_info, "delayed_item",
 520					      item->delayed_node->inode_id,
 521					      num_bytes, 1);
 522		/*
 523		 * For insertions we track reserved metadata space by accounting
 524		 * for the number of leaves that will be used, based on the delayed
 525		 * node's curr_index_batch_size and index_item_leaves fields.
 526		 */
 527		if (item->type == BTRFS_DELAYED_DELETION_ITEM)
 528			item->bytes_reserved = num_bytes;
 529	}
 530
 531	return ret;
 532}
 533
 534static void btrfs_delayed_item_release_metadata(struct btrfs_root *root,
 535						struct btrfs_delayed_item *item)
 536{
 537	struct btrfs_block_rsv *rsv;
 538	struct btrfs_fs_info *fs_info = root->fs_info;
 539
 540	if (!item->bytes_reserved)
 541		return;
 542
 543	rsv = &fs_info->delayed_block_rsv;
 544	/*
 545	 * Check btrfs_delayed_item_reserve_metadata() to see why we don't need
 546	 * to release/reserve qgroup space.
 547	 */
 548	trace_btrfs_space_reservation(fs_info, "delayed_item",
 549				      item->delayed_node->inode_id,
 550				      item->bytes_reserved, 0);
 551	btrfs_block_rsv_release(fs_info, rsv, item->bytes_reserved, NULL);
 552}
 553
 554static void btrfs_delayed_item_release_leaves(struct btrfs_delayed_node *node,
 555					      unsigned int num_leaves)
 556{
 557	struct btrfs_fs_info *fs_info = node->root->fs_info;
 558	const u64 bytes = btrfs_calc_insert_metadata_size(fs_info, num_leaves);
 559
 560	/* There are no space reservations during log replay, bail out. */
 561	if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
 562		return;
 563
 564	trace_btrfs_space_reservation(fs_info, "delayed_item", node->inode_id,
 565				      bytes, 0);
 566	btrfs_block_rsv_release(fs_info, &fs_info->delayed_block_rsv, bytes, NULL);
 567}
 568
 569static int btrfs_delayed_inode_reserve_metadata(
 570					struct btrfs_trans_handle *trans,
 571					struct btrfs_root *root,
 572					struct btrfs_delayed_node *node)
 573{
 574	struct btrfs_fs_info *fs_info = root->fs_info;
 575	struct btrfs_block_rsv *src_rsv;
 576	struct btrfs_block_rsv *dst_rsv;
 577	u64 num_bytes;
 578	int ret;
 579
 580	src_rsv = trans->block_rsv;
 581	dst_rsv = &fs_info->delayed_block_rsv;
 582
 583	num_bytes = btrfs_calc_metadata_size(fs_info, 1);
 584
 585	/*
 586	 * btrfs_dirty_inode will update the inode under btrfs_join_transaction
 587	 * which doesn't reserve space for speed.  This is a problem since we
 588	 * still need to reserve space for this update, so try to reserve the
 589	 * space.
 590	 *
 591	 * Now if src_rsv == delalloc_block_rsv we'll let it just steal since
 592	 * we always reserve enough to update the inode item.
 593	 */
 594	if (!src_rsv || (!trans->bytes_reserved &&
 595			 src_rsv->type != BTRFS_BLOCK_RSV_DELALLOC)) {
 596		ret = btrfs_qgroup_reserve_meta(root, num_bytes,
 597					  BTRFS_QGROUP_RSV_META_PREALLOC, true);
 598		if (ret < 0)
 599			return ret;
 600		ret = btrfs_block_rsv_add(fs_info, dst_rsv, num_bytes,
 601					  BTRFS_RESERVE_NO_FLUSH);
 602		/* NO_FLUSH could only fail with -ENOSPC */
 603		ASSERT(ret == 0 || ret == -ENOSPC);
 604		if (ret)
 605			btrfs_qgroup_free_meta_prealloc(root, num_bytes);
 606	} else {
 607		ret = btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, true);
 608	}
 609
 610	if (!ret) {
 611		trace_btrfs_space_reservation(fs_info, "delayed_inode",
 612					      node->inode_id, num_bytes, 1);
 613		node->bytes_reserved = num_bytes;
 614	}
 615
 616	return ret;
 617}
 618
 619static void btrfs_delayed_inode_release_metadata(struct btrfs_fs_info *fs_info,
 620						struct btrfs_delayed_node *node,
 621						bool qgroup_free)
 622{
 623	struct btrfs_block_rsv *rsv;
 624
 625	if (!node->bytes_reserved)
 626		return;
 627
 628	rsv = &fs_info->delayed_block_rsv;
 629	trace_btrfs_space_reservation(fs_info, "delayed_inode",
 630				      node->inode_id, node->bytes_reserved, 0);
 631	btrfs_block_rsv_release(fs_info, rsv, node->bytes_reserved, NULL);
 632	if (qgroup_free)
 633		btrfs_qgroup_free_meta_prealloc(node->root,
 634				node->bytes_reserved);
 635	else
 636		btrfs_qgroup_convert_reserved_meta(node->root,
 637				node->bytes_reserved);
 638	node->bytes_reserved = 0;
 639}
 640
 641/*
 642 * Insert a single delayed item or a batch of delayed items, as many as possible
 643 * that fit in a leaf. The delayed items (dir index keys) are sorted by their key
 644 * in the rbtree, and if there's a gap between two consecutive dir index items,
 645 * then it means at some point we had delayed dir indexes to add but they got
 646 * removed (by btrfs_delete_delayed_dir_index()) before we attempted to flush them
 647 * into the subvolume tree. Dir index keys also have their offsets coming from a
 648 * monotonically increasing counter, so we can't get new keys with an offset that
 649 * fits within a gap between delayed dir index items.
 650 */
 651static int btrfs_insert_delayed_item(struct btrfs_trans_handle *trans,
 652				     struct btrfs_root *root,
 653				     struct btrfs_path *path,
 654				     struct btrfs_delayed_item *first_item)
 655{
 656	struct btrfs_fs_info *fs_info = root->fs_info;
 657	struct btrfs_delayed_node *node = first_item->delayed_node;
 658	LIST_HEAD(item_list);
 659	struct btrfs_delayed_item *curr;
 660	struct btrfs_delayed_item *next;
 661	const int max_size = BTRFS_LEAF_DATA_SIZE(fs_info);
 662	struct btrfs_item_batch batch;
 663	struct btrfs_key first_key;
 664	const u32 first_data_size = first_item->data_len;
 665	int total_size;
 666	char *ins_data = NULL;
 667	int ret;
 668	bool continuous_keys_only = false;
 669
 670	lockdep_assert_held(&node->mutex);
 671
 672	/*
 673	 * During normal operation the delayed index offset is continuously
 674	 * increasing, so we can batch insert all items as there will not be any
 675	 * overlapping keys in the tree.
 676	 *
 677	 * The exception to this is log replay, where we may have interleaved
 678	 * offsets in the tree, so our batch needs to be continuous keys only in
 679	 * order to ensure we do not end up with out of order items in our leaf.
 680	 */
 681	if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
 682		continuous_keys_only = true;
 683
 684	/*
 685	 * For delayed items to insert, we track reserved metadata bytes based
 686	 * on the number of leaves that we will use.
 687	 * See btrfs_insert_delayed_dir_index() and
 688	 * btrfs_delayed_item_reserve_metadata()).
 689	 */
 690	ASSERT(first_item->bytes_reserved == 0);
 691
 692	list_add_tail(&first_item->tree_list, &item_list);
 693	batch.total_data_size = first_data_size;
 694	batch.nr = 1;
 695	total_size = first_data_size + sizeof(struct btrfs_item);
 696	curr = first_item;
 697
 698	while (true) {
 699		int next_size;
 700
 701		next = __btrfs_next_delayed_item(curr);
 702		if (!next)
 703			break;
 704
 705		/*
 706		 * We cannot allow gaps in the key space if we're doing log
 707		 * replay.
 708		 */
 709		if (continuous_keys_only && (next->index != curr->index + 1))
 710			break;
 711
 712		ASSERT(next->bytes_reserved == 0);
 713
 714		next_size = next->data_len + sizeof(struct btrfs_item);
 715		if (total_size + next_size > max_size)
 716			break;
 717
 718		list_add_tail(&next->tree_list, &item_list);
 719		batch.nr++;
 720		total_size += next_size;
 721		batch.total_data_size += next->data_len;
 722		curr = next;
 723	}
 724
 725	if (batch.nr == 1) {
 726		first_key.objectid = node->inode_id;
 727		first_key.type = BTRFS_DIR_INDEX_KEY;
 728		first_key.offset = first_item->index;
 729		batch.keys = &first_key;
 730		batch.data_sizes = &first_data_size;
 731	} else {
 732		struct btrfs_key *ins_keys;
 733		u32 *ins_sizes;
 734		int i = 0;
 735
 736		ins_data = kmalloc(batch.nr * sizeof(u32) +
 737				   batch.nr * sizeof(struct btrfs_key), GFP_NOFS);
 738		if (!ins_data) {
 739			ret = -ENOMEM;
 740			goto out;
 741		}
 742		ins_sizes = (u32 *)ins_data;
 743		ins_keys = (struct btrfs_key *)(ins_data + batch.nr * sizeof(u32));
 744		batch.keys = ins_keys;
 745		batch.data_sizes = ins_sizes;
 746		list_for_each_entry(curr, &item_list, tree_list) {
 747			ins_keys[i].objectid = node->inode_id;
 748			ins_keys[i].type = BTRFS_DIR_INDEX_KEY;
 749			ins_keys[i].offset = curr->index;
 750			ins_sizes[i] = curr->data_len;
 751			i++;
 752		}
 753	}
 754
 755	ret = btrfs_insert_empty_items(trans, root, path, &batch);
 756	if (ret)
 757		goto out;
 758
 759	list_for_each_entry(curr, &item_list, tree_list) {
 760		char *data_ptr;
 761
 762		data_ptr = btrfs_item_ptr(path->nodes[0], path->slots[0], char);
 763		write_extent_buffer(path->nodes[0], &curr->data,
 764				    (unsigned long)data_ptr, curr->data_len);
 765		path->slots[0]++;
 766	}
 767
 768	/*
 769	 * Now release our path before releasing the delayed items and their
 770	 * metadata reservations, so that we don't block other tasks for more
 771	 * time than needed.
 772	 */
 773	btrfs_release_path(path);
 774
 775	ASSERT(node->index_item_leaves > 0);
 776
 777	/*
 778	 * For normal operations we will batch an entire leaf's worth of delayed
 779	 * items, so if there are more items to process we can decrement
 780	 * index_item_leaves by 1 as we inserted 1 leaf's worth of items.
 781	 *
 782	 * However for log replay we may not have inserted an entire leaf's
 783	 * worth of items, we may have not had continuous items, so decrementing
 784	 * here would mess up the index_item_leaves accounting.  For this case
 785	 * only clean up the accounting when there are no items left.
 786	 */
 787	if (next && !continuous_keys_only) {
 788		/*
 789		 * We inserted one batch of items into a leaf a there are more
 790		 * items to flush in a future batch, now release one unit of
 791		 * metadata space from the delayed block reserve, corresponding
 792		 * the leaf we just flushed to.
 793		 */
 794		btrfs_delayed_item_release_leaves(node, 1);
 795		node->index_item_leaves--;
 796	} else if (!next) {
 797		/*
 798		 * There are no more items to insert. We can have a number of
 799		 * reserved leaves > 1 here - this happens when many dir index
 800		 * items are added and then removed before they are flushed (file
 801		 * names with a very short life, never span a transaction). So
 802		 * release all remaining leaves.
 803		 */
 804		btrfs_delayed_item_release_leaves(node, node->index_item_leaves);
 805		node->index_item_leaves = 0;
 806	}
 807
 808	list_for_each_entry_safe(curr, next, &item_list, tree_list) {
 809		list_del(&curr->tree_list);
 810		btrfs_release_delayed_item(curr);
 811	}
 812out:
 813	kfree(ins_data);
 814	return ret;
 815}
 816
 817static int btrfs_insert_delayed_items(struct btrfs_trans_handle *trans,
 818				      struct btrfs_path *path,
 819				      struct btrfs_root *root,
 820				      struct btrfs_delayed_node *node)
 821{
 822	int ret = 0;
 823
 824	while (ret == 0) {
 825		struct btrfs_delayed_item *curr;
 826
 827		mutex_lock(&node->mutex);
 828		curr = __btrfs_first_delayed_insertion_item(node);
 829		if (!curr) {
 830			mutex_unlock(&node->mutex);
 831			break;
 832		}
 833		ret = btrfs_insert_delayed_item(trans, root, path, curr);
 834		mutex_unlock(&node->mutex);
 835	}
 836
 837	return ret;
 838}
 839
 840static int btrfs_batch_delete_items(struct btrfs_trans_handle *trans,
 841				    struct btrfs_root *root,
 842				    struct btrfs_path *path,
 843				    struct btrfs_delayed_item *item)
 844{
 845	const u64 ino = item->delayed_node->inode_id;
 846	struct btrfs_fs_info *fs_info = root->fs_info;
 847	struct btrfs_delayed_item *curr, *next;
 848	struct extent_buffer *leaf = path->nodes[0];
 849	LIST_HEAD(batch_list);
 850	int nitems, slot, last_slot;
 851	int ret;
 852	u64 total_reserved_size = item->bytes_reserved;
 853
 854	ASSERT(leaf != NULL);
 855
 856	slot = path->slots[0];
 857	last_slot = btrfs_header_nritems(leaf) - 1;
 858	/*
 859	 * Our caller always gives us a path pointing to an existing item, so
 860	 * this can not happen.
 861	 */
 862	ASSERT(slot <= last_slot);
 863	if (WARN_ON(slot > last_slot))
 864		return -ENOENT;
 865
 866	nitems = 1;
 867	curr = item;
 868	list_add_tail(&curr->tree_list, &batch_list);
 869
 870	/*
 871	 * Keep checking if the next delayed item matches the next item in the
 872	 * leaf - if so, we can add it to the batch of items to delete from the
 873	 * leaf.
 874	 */
 875	while (slot < last_slot) {
 876		struct btrfs_key key;
 877
 878		next = __btrfs_next_delayed_item(curr);
 879		if (!next)
 880			break;
 881
 882		slot++;
 883		btrfs_item_key_to_cpu(leaf, &key, slot);
 884		if (key.objectid != ino ||
 885		    key.type != BTRFS_DIR_INDEX_KEY ||
 886		    key.offset != next->index)
 887			break;
 888		nitems++;
 889		curr = next;
 890		list_add_tail(&curr->tree_list, &batch_list);
 891		total_reserved_size += curr->bytes_reserved;
 892	}
 893
 894	ret = btrfs_del_items(trans, root, path, path->slots[0], nitems);
 895	if (ret)
 896		return ret;
 897
 898	/* In case of BTRFS_FS_LOG_RECOVERING items won't have reserved space */
 899	if (total_reserved_size > 0) {
 900		/*
 901		 * Check btrfs_delayed_item_reserve_metadata() to see why we
 902		 * don't need to release/reserve qgroup space.
 903		 */
 904		trace_btrfs_space_reservation(fs_info, "delayed_item", ino,
 905					      total_reserved_size, 0);
 906		btrfs_block_rsv_release(fs_info, &fs_info->delayed_block_rsv,
 907					total_reserved_size, NULL);
 908	}
 909
 910	list_for_each_entry_safe(curr, next, &batch_list, tree_list) {
 911		list_del(&curr->tree_list);
 912		btrfs_release_delayed_item(curr);
 913	}
 914
 915	return 0;
 916}
 917
 918static int btrfs_delete_delayed_items(struct btrfs_trans_handle *trans,
 919				      struct btrfs_path *path,
 920				      struct btrfs_root *root,
 921				      struct btrfs_delayed_node *node)
 922{
 923	struct btrfs_key key;
 924	int ret = 0;
 925
 926	key.objectid = node->inode_id;
 927	key.type = BTRFS_DIR_INDEX_KEY;
 928
 929	while (ret == 0) {
 930		struct btrfs_delayed_item *item;
 931
 932		mutex_lock(&node->mutex);
 933		item = __btrfs_first_delayed_deletion_item(node);
 934		if (!item) {
 935			mutex_unlock(&node->mutex);
 936			break;
 937		}
 938
 939		key.offset = item->index;
 940		ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 941		if (ret > 0) {
 942			/*
 943			 * There's no matching item in the leaf. This means we
 944			 * have already deleted this item in a past run of the
 945			 * delayed items. We ignore errors when running delayed
 946			 * items from an async context, through a work queue job
 947			 * running btrfs_async_run_delayed_root(), and don't
 948			 * release delayed items that failed to complete. This
 949			 * is because we will retry later, and at transaction
 950			 * commit time we always run delayed items and will
 951			 * then deal with errors if they fail to run again.
 952			 *
 953			 * So just release delayed items for which we can't find
 954			 * an item in the tree, and move to the next item.
 955			 */
 956			btrfs_release_path(path);
 957			btrfs_release_delayed_item(item);
 958			ret = 0;
 959		} else if (ret == 0) {
 960			ret = btrfs_batch_delete_items(trans, root, path, item);
 961			btrfs_release_path(path);
 962		}
 963
 964		/*
 965		 * We unlock and relock on each iteration, this is to prevent
 966		 * blocking other tasks for too long while we are being run from
 967		 * the async context (work queue job). Those tasks are typically
 968		 * running system calls like creat/mkdir/rename/unlink/etc which
 969		 * need to add delayed items to this delayed node.
 970		 */
 971		mutex_unlock(&node->mutex);
 972	}
 973
 974	return ret;
 975}
 976
 977static void btrfs_release_delayed_inode(struct btrfs_delayed_node *delayed_node)
 978{
 979	struct btrfs_delayed_root *delayed_root;
 980
 981	if (delayed_node &&
 982	    test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
 983		BUG_ON(!delayed_node->root);
 984		clear_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags);
 985		delayed_node->count--;
 986
 987		delayed_root = delayed_node->root->fs_info->delayed_root;
 988		finish_one_item(delayed_root);
 989	}
 990}
 991
 992static void btrfs_release_delayed_iref(struct btrfs_delayed_node *delayed_node)
 993{
 994
 995	if (test_and_clear_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags)) {
 996		struct btrfs_delayed_root *delayed_root;
 997
 998		ASSERT(delayed_node->root);
 999		delayed_node->count--;
1000
1001		delayed_root = delayed_node->root->fs_info->delayed_root;
1002		finish_one_item(delayed_root);
1003	}
1004}
1005
1006static int __btrfs_update_delayed_inode(struct btrfs_trans_handle *trans,
1007					struct btrfs_root *root,
1008					struct btrfs_path *path,
1009					struct btrfs_delayed_node *node)
1010{
1011	struct btrfs_fs_info *fs_info = root->fs_info;
1012	struct btrfs_key key;
1013	struct btrfs_inode_item *inode_item;
1014	struct extent_buffer *leaf;
1015	int mod;
1016	int ret;
1017
1018	key.objectid = node->inode_id;
1019	key.type = BTRFS_INODE_ITEM_KEY;
1020	key.offset = 0;
1021
1022	if (test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &node->flags))
1023		mod = -1;
1024	else
1025		mod = 1;
1026
1027	ret = btrfs_lookup_inode(trans, root, path, &key, mod);
1028	if (ret > 0)
1029		ret = -ENOENT;
1030	if (ret < 0)
1031		goto out;
1032
1033	leaf = path->nodes[0];
1034	inode_item = btrfs_item_ptr(leaf, path->slots[0],
1035				    struct btrfs_inode_item);
1036	write_extent_buffer(leaf, &node->inode_item, (unsigned long)inode_item,
1037			    sizeof(struct btrfs_inode_item));
1038	btrfs_mark_buffer_dirty(trans, leaf);
1039
1040	if (!test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &node->flags))
1041		goto out;
1042
1043	/*
1044	 * Now we're going to delete the INODE_REF/EXTREF, which should be the
1045	 * only one ref left.  Check if the next item is an INODE_REF/EXTREF.
1046	 *
1047	 * But if we're the last item already, release and search for the last
1048	 * INODE_REF/EXTREF.
1049	 */
1050	if (path->slots[0] + 1 >= btrfs_header_nritems(leaf)) {
1051		key.objectid = node->inode_id;
1052		key.type = BTRFS_INODE_EXTREF_KEY;
1053		key.offset = (u64)-1;
1054
1055		btrfs_release_path(path);
1056		ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1057		if (ret < 0)
1058			goto err_out;
1059		ASSERT(ret > 0);
1060		ASSERT(path->slots[0] > 0);
1061		ret = 0;
1062		path->slots[0]--;
1063		leaf = path->nodes[0];
1064	} else {
1065		path->slots[0]++;
1066	}
1067	btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1068	if (key.objectid != node->inode_id)
1069		goto out;
1070	if (key.type != BTRFS_INODE_REF_KEY &&
1071	    key.type != BTRFS_INODE_EXTREF_KEY)
1072		goto out;
1073
1074	/*
1075	 * Delayed iref deletion is for the inode who has only one link,
1076	 * so there is only one iref. The case that several irefs are
1077	 * in the same item doesn't exist.
1078	 */
1079	ret = btrfs_del_item(trans, root, path);
1080out:
1081	btrfs_release_delayed_iref(node);
1082	btrfs_release_path(path);
1083err_out:
1084	btrfs_delayed_inode_release_metadata(fs_info, node, (ret < 0));
1085	btrfs_release_delayed_inode(node);
1086
1087	/*
1088	 * If we fail to update the delayed inode we need to abort the
1089	 * transaction, because we could leave the inode with the improper
1090	 * counts behind.
1091	 */
1092	if (ret && ret != -ENOENT)
1093		btrfs_abort_transaction(trans, ret);
1094
1095	return ret;
1096}
1097
1098static inline int btrfs_update_delayed_inode(struct btrfs_trans_handle *trans,
1099					     struct btrfs_root *root,
1100					     struct btrfs_path *path,
1101					     struct btrfs_delayed_node *node)
1102{
1103	int ret;
1104
1105	mutex_lock(&node->mutex);
1106	if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &node->flags)) {
1107		mutex_unlock(&node->mutex);
1108		return 0;
1109	}
1110
1111	ret = __btrfs_update_delayed_inode(trans, root, path, node);
1112	mutex_unlock(&node->mutex);
1113	return ret;
1114}
1115
1116static inline int
1117__btrfs_commit_inode_delayed_items(struct btrfs_trans_handle *trans,
1118				   struct btrfs_path *path,
1119				   struct btrfs_delayed_node *node)
1120{
1121	int ret;
1122
1123	ret = btrfs_insert_delayed_items(trans, path, node->root, node);
1124	if (ret)
1125		return ret;
1126
1127	ret = btrfs_delete_delayed_items(trans, path, node->root, node);
1128	if (ret)
1129		return ret;
1130
 
 
 
1131	ret = btrfs_update_delayed_inode(trans, node->root, path, node);
1132	return ret;
1133}
1134
1135/*
1136 * Called when committing the transaction.
1137 * Returns 0 on success.
1138 * Returns < 0 on error and returns with an aborted transaction with any
1139 * outstanding delayed items cleaned up.
1140 */
1141static int __btrfs_run_delayed_items(struct btrfs_trans_handle *trans, int nr)
1142{
1143	struct btrfs_fs_info *fs_info = trans->fs_info;
1144	struct btrfs_delayed_root *delayed_root;
1145	struct btrfs_delayed_node *curr_node, *prev_node;
1146	struct btrfs_path *path;
1147	struct btrfs_block_rsv *block_rsv;
1148	int ret = 0;
1149	bool count = (nr > 0);
1150
1151	if (TRANS_ABORTED(trans))
1152		return -EIO;
1153
1154	path = btrfs_alloc_path();
1155	if (!path)
1156		return -ENOMEM;
1157
1158	block_rsv = trans->block_rsv;
1159	trans->block_rsv = &fs_info->delayed_block_rsv;
1160
1161	delayed_root = fs_info->delayed_root;
1162
1163	curr_node = btrfs_first_delayed_node(delayed_root);
1164	while (curr_node && (!count || nr--)) {
1165		ret = __btrfs_commit_inode_delayed_items(trans, path,
1166							 curr_node);
1167		if (ret) {
1168			btrfs_abort_transaction(trans, ret);
1169			break;
1170		}
1171
1172		prev_node = curr_node;
1173		curr_node = btrfs_next_delayed_node(curr_node);
1174		/*
1175		 * See the comment below about releasing path before releasing
1176		 * node. If the commit of delayed items was successful the path
1177		 * should always be released, but in case of an error, it may
1178		 * point to locked extent buffers (a leaf at the very least).
1179		 */
1180		ASSERT(path->nodes[0] == NULL);
1181		btrfs_release_delayed_node(prev_node);
1182	}
1183
1184	/*
1185	 * Release the path to avoid a potential deadlock and lockdep splat when
1186	 * releasing the delayed node, as that requires taking the delayed node's
1187	 * mutex. If another task starts running delayed items before we take
1188	 * the mutex, it will first lock the mutex and then it may try to lock
1189	 * the same btree path (leaf).
1190	 */
1191	btrfs_free_path(path);
1192
1193	if (curr_node)
1194		btrfs_release_delayed_node(curr_node);
1195	trans->block_rsv = block_rsv;
1196
1197	return ret;
1198}
1199
1200int btrfs_run_delayed_items(struct btrfs_trans_handle *trans)
1201{
1202	return __btrfs_run_delayed_items(trans, -1);
1203}
1204
1205int btrfs_run_delayed_items_nr(struct btrfs_trans_handle *trans, int nr)
1206{
1207	return __btrfs_run_delayed_items(trans, nr);
1208}
1209
1210int btrfs_commit_inode_delayed_items(struct btrfs_trans_handle *trans,
1211				     struct btrfs_inode *inode)
1212{
1213	struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1214	struct btrfs_path *path;
1215	struct btrfs_block_rsv *block_rsv;
1216	int ret;
1217
1218	if (!delayed_node)
1219		return 0;
1220
1221	mutex_lock(&delayed_node->mutex);
1222	if (!delayed_node->count) {
1223		mutex_unlock(&delayed_node->mutex);
1224		btrfs_release_delayed_node(delayed_node);
1225		return 0;
1226	}
1227	mutex_unlock(&delayed_node->mutex);
1228
1229	path = btrfs_alloc_path();
1230	if (!path) {
1231		btrfs_release_delayed_node(delayed_node);
1232		return -ENOMEM;
1233	}
1234
1235	block_rsv = trans->block_rsv;
1236	trans->block_rsv = &delayed_node->root->fs_info->delayed_block_rsv;
1237
1238	ret = __btrfs_commit_inode_delayed_items(trans, path, delayed_node);
1239
1240	btrfs_release_delayed_node(delayed_node);
1241	btrfs_free_path(path);
1242	trans->block_rsv = block_rsv;
1243
1244	return ret;
1245}
1246
1247int btrfs_commit_inode_delayed_inode(struct btrfs_inode *inode)
1248{
1249	struct btrfs_fs_info *fs_info = inode->root->fs_info;
1250	struct btrfs_trans_handle *trans;
1251	struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1252	struct btrfs_path *path;
1253	struct btrfs_block_rsv *block_rsv;
1254	int ret;
1255
1256	if (!delayed_node)
1257		return 0;
1258
1259	mutex_lock(&delayed_node->mutex);
1260	if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1261		mutex_unlock(&delayed_node->mutex);
1262		btrfs_release_delayed_node(delayed_node);
1263		return 0;
1264	}
1265	mutex_unlock(&delayed_node->mutex);
1266
1267	trans = btrfs_join_transaction(delayed_node->root);
1268	if (IS_ERR(trans)) {
1269		ret = PTR_ERR(trans);
1270		goto out;
1271	}
1272
1273	path = btrfs_alloc_path();
1274	if (!path) {
1275		ret = -ENOMEM;
1276		goto trans_out;
1277	}
1278
1279	block_rsv = trans->block_rsv;
1280	trans->block_rsv = &fs_info->delayed_block_rsv;
1281
1282	mutex_lock(&delayed_node->mutex);
1283	if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags))
1284		ret = __btrfs_update_delayed_inode(trans, delayed_node->root,
1285						   path, delayed_node);
1286	else
1287		ret = 0;
1288	mutex_unlock(&delayed_node->mutex);
1289
1290	btrfs_free_path(path);
1291	trans->block_rsv = block_rsv;
1292trans_out:
1293	btrfs_end_transaction(trans);
1294	btrfs_btree_balance_dirty(fs_info);
1295out:
1296	btrfs_release_delayed_node(delayed_node);
1297
1298	return ret;
1299}
1300
1301void btrfs_remove_delayed_node(struct btrfs_inode *inode)
1302{
1303	struct btrfs_delayed_node *delayed_node;
1304
1305	delayed_node = READ_ONCE(inode->delayed_node);
1306	if (!delayed_node)
1307		return;
1308
1309	inode->delayed_node = NULL;
1310	btrfs_release_delayed_node(delayed_node);
1311}
1312
1313struct btrfs_async_delayed_work {
1314	struct btrfs_delayed_root *delayed_root;
1315	int nr;
1316	struct btrfs_work work;
1317};
1318
1319static void btrfs_async_run_delayed_root(struct btrfs_work *work)
1320{
1321	struct btrfs_async_delayed_work *async_work;
1322	struct btrfs_delayed_root *delayed_root;
1323	struct btrfs_trans_handle *trans;
1324	struct btrfs_path *path;
1325	struct btrfs_delayed_node *delayed_node = NULL;
1326	struct btrfs_root *root;
1327	struct btrfs_block_rsv *block_rsv;
1328	int total_done = 0;
1329
1330	async_work = container_of(work, struct btrfs_async_delayed_work, work);
1331	delayed_root = async_work->delayed_root;
1332
1333	path = btrfs_alloc_path();
1334	if (!path)
1335		goto out;
1336
1337	do {
1338		if (atomic_read(&delayed_root->items) <
1339		    BTRFS_DELAYED_BACKGROUND / 2)
1340			break;
1341
1342		delayed_node = btrfs_first_prepared_delayed_node(delayed_root);
1343		if (!delayed_node)
1344			break;
1345
1346		root = delayed_node->root;
1347
1348		trans = btrfs_join_transaction(root);
1349		if (IS_ERR(trans)) {
1350			btrfs_release_path(path);
1351			btrfs_release_prepared_delayed_node(delayed_node);
1352			total_done++;
1353			continue;
1354		}
1355
1356		block_rsv = trans->block_rsv;
1357		trans->block_rsv = &root->fs_info->delayed_block_rsv;
1358
1359		__btrfs_commit_inode_delayed_items(trans, path, delayed_node);
1360
1361		trans->block_rsv = block_rsv;
1362		btrfs_end_transaction(trans);
1363		btrfs_btree_balance_dirty_nodelay(root->fs_info);
1364
1365		btrfs_release_path(path);
1366		btrfs_release_prepared_delayed_node(delayed_node);
1367		total_done++;
1368
1369	} while ((async_work->nr == 0 && total_done < BTRFS_DELAYED_WRITEBACK)
1370		 || total_done < async_work->nr);
1371
1372	btrfs_free_path(path);
1373out:
1374	wake_up(&delayed_root->wait);
1375	kfree(async_work);
1376}
1377
1378
1379static int btrfs_wq_run_delayed_node(struct btrfs_delayed_root *delayed_root,
1380				     struct btrfs_fs_info *fs_info, int nr)
1381{
1382	struct btrfs_async_delayed_work *async_work;
1383
1384	async_work = kmalloc(sizeof(*async_work), GFP_NOFS);
1385	if (!async_work)
1386		return -ENOMEM;
1387
1388	async_work->delayed_root = delayed_root;
1389	btrfs_init_work(&async_work->work, btrfs_async_run_delayed_root, NULL);
1390	async_work->nr = nr;
1391
1392	btrfs_queue_work(fs_info->delayed_workers, &async_work->work);
1393	return 0;
1394}
1395
1396void btrfs_assert_delayed_root_empty(struct btrfs_fs_info *fs_info)
1397{
1398	WARN_ON(btrfs_first_delayed_node(fs_info->delayed_root));
1399}
1400
1401static int could_end_wait(struct btrfs_delayed_root *delayed_root, int seq)
1402{
1403	int val = atomic_read(&delayed_root->items_seq);
1404
1405	if (val < seq || val >= seq + BTRFS_DELAYED_BATCH)
1406		return 1;
1407
1408	if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND)
1409		return 1;
1410
1411	return 0;
1412}
1413
1414void btrfs_balance_delayed_items(struct btrfs_fs_info *fs_info)
1415{
1416	struct btrfs_delayed_root *delayed_root = fs_info->delayed_root;
1417
1418	if ((atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND) ||
1419		btrfs_workqueue_normal_congested(fs_info->delayed_workers))
1420		return;
1421
1422	if (atomic_read(&delayed_root->items) >= BTRFS_DELAYED_WRITEBACK) {
1423		int seq;
1424		int ret;
1425
1426		seq = atomic_read(&delayed_root->items_seq);
1427
1428		ret = btrfs_wq_run_delayed_node(delayed_root, fs_info, 0);
1429		if (ret)
1430			return;
1431
1432		wait_event_interruptible(delayed_root->wait,
1433					 could_end_wait(delayed_root, seq));
1434		return;
1435	}
1436
1437	btrfs_wq_run_delayed_node(delayed_root, fs_info, BTRFS_DELAYED_BATCH);
1438}
1439
1440static void btrfs_release_dir_index_item_space(struct btrfs_trans_handle *trans)
1441{
1442	struct btrfs_fs_info *fs_info = trans->fs_info;
1443	const u64 bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
1444
1445	if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
1446		return;
1447
1448	/*
1449	 * Adding the new dir index item does not require touching another
1450	 * leaf, so we can release 1 unit of metadata that was previously
1451	 * reserved when starting the transaction. This applies only to
1452	 * the case where we had a transaction start and excludes the
1453	 * transaction join case (when replaying log trees).
1454	 */
1455	trace_btrfs_space_reservation(fs_info, "transaction",
1456				      trans->transid, bytes, 0);
1457	btrfs_block_rsv_release(fs_info, trans->block_rsv, bytes, NULL);
1458	ASSERT(trans->bytes_reserved >= bytes);
1459	trans->bytes_reserved -= bytes;
1460}
1461
1462/* Will return 0, -ENOMEM or -EEXIST (index number collision, unexpected). */
1463int btrfs_insert_delayed_dir_index(struct btrfs_trans_handle *trans,
1464				   const char *name, int name_len,
1465				   struct btrfs_inode *dir,
1466				   struct btrfs_disk_key *disk_key, u8 flags,
1467				   u64 index)
1468{
1469	struct btrfs_fs_info *fs_info = trans->fs_info;
1470	const unsigned int leaf_data_size = BTRFS_LEAF_DATA_SIZE(fs_info);
1471	struct btrfs_delayed_node *delayed_node;
1472	struct btrfs_delayed_item *delayed_item;
1473	struct btrfs_dir_item *dir_item;
1474	bool reserve_leaf_space;
1475	u32 data_len;
1476	int ret;
1477
1478	delayed_node = btrfs_get_or_create_delayed_node(dir);
1479	if (IS_ERR(delayed_node))
1480		return PTR_ERR(delayed_node);
1481
1482	delayed_item = btrfs_alloc_delayed_item(sizeof(*dir_item) + name_len,
1483						delayed_node,
1484						BTRFS_DELAYED_INSERTION_ITEM);
1485	if (!delayed_item) {
1486		ret = -ENOMEM;
1487		goto release_node;
1488	}
1489
1490	delayed_item->index = index;
1491
1492	dir_item = (struct btrfs_dir_item *)delayed_item->data;
1493	dir_item->location = *disk_key;
1494	btrfs_set_stack_dir_transid(dir_item, trans->transid);
1495	btrfs_set_stack_dir_data_len(dir_item, 0);
1496	btrfs_set_stack_dir_name_len(dir_item, name_len);
1497	btrfs_set_stack_dir_flags(dir_item, flags);
1498	memcpy((char *)(dir_item + 1), name, name_len);
1499
1500	data_len = delayed_item->data_len + sizeof(struct btrfs_item);
1501
1502	mutex_lock(&delayed_node->mutex);
1503
1504	/*
1505	 * First attempt to insert the delayed item. This is to make the error
1506	 * handling path simpler in case we fail (-EEXIST). There's no risk of
1507	 * any other task coming in and running the delayed item before we do
1508	 * the metadata space reservation below, because we are holding the
1509	 * delayed node's mutex and that mutex must also be locked before the
1510	 * node's delayed items can be run.
1511	 */
1512	ret = __btrfs_add_delayed_item(delayed_node, delayed_item);
1513	if (unlikely(ret)) {
1514		btrfs_err(trans->fs_info,
1515"error adding delayed dir index item, name: %.*s, index: %llu, root: %llu, dir: %llu, dir->index_cnt: %llu, delayed_node->index_cnt: %llu, error: %d",
1516			  name_len, name, index, btrfs_root_id(delayed_node->root),
1517			  delayed_node->inode_id, dir->index_cnt,
1518			  delayed_node->index_cnt, ret);
1519		btrfs_release_delayed_item(delayed_item);
1520		btrfs_release_dir_index_item_space(trans);
1521		mutex_unlock(&delayed_node->mutex);
1522		goto release_node;
1523	}
1524
1525	if (delayed_node->index_item_leaves == 0 ||
1526	    delayed_node->curr_index_batch_size + data_len > leaf_data_size) {
1527		delayed_node->curr_index_batch_size = data_len;
1528		reserve_leaf_space = true;
1529	} else {
1530		delayed_node->curr_index_batch_size += data_len;
1531		reserve_leaf_space = false;
1532	}
1533
1534	if (reserve_leaf_space) {
1535		ret = btrfs_delayed_item_reserve_metadata(trans, delayed_item);
1536		/*
1537		 * Space was reserved for a dir index item insertion when we
1538		 * started the transaction, so getting a failure here should be
1539		 * impossible.
1540		 */
1541		if (WARN_ON(ret)) {
1542			btrfs_release_delayed_item(delayed_item);
1543			mutex_unlock(&delayed_node->mutex);
1544			goto release_node;
1545		}
1546
1547		delayed_node->index_item_leaves++;
1548	} else {
1549		btrfs_release_dir_index_item_space(trans);
1550	}
1551	mutex_unlock(&delayed_node->mutex);
1552
1553release_node:
1554	btrfs_release_delayed_node(delayed_node);
1555	return ret;
1556}
1557
1558static int btrfs_delete_delayed_insertion_item(struct btrfs_fs_info *fs_info,
1559					       struct btrfs_delayed_node *node,
1560					       u64 index)
1561{
1562	struct btrfs_delayed_item *item;
1563
1564	mutex_lock(&node->mutex);
1565	item = __btrfs_lookup_delayed_item(&node->ins_root.rb_root, index);
1566	if (!item) {
1567		mutex_unlock(&node->mutex);
1568		return 1;
1569	}
1570
1571	/*
1572	 * For delayed items to insert, we track reserved metadata bytes based
1573	 * on the number of leaves that we will use.
1574	 * See btrfs_insert_delayed_dir_index() and
1575	 * btrfs_delayed_item_reserve_metadata()).
1576	 */
1577	ASSERT(item->bytes_reserved == 0);
1578	ASSERT(node->index_item_leaves > 0);
1579
1580	/*
1581	 * If there's only one leaf reserved, we can decrement this item from the
1582	 * current batch, otherwise we can not because we don't know which leaf
1583	 * it belongs to. With the current limit on delayed items, we rarely
1584	 * accumulate enough dir index items to fill more than one leaf (even
1585	 * when using a leaf size of 4K).
1586	 */
1587	if (node->index_item_leaves == 1) {
1588		const u32 data_len = item->data_len + sizeof(struct btrfs_item);
1589
1590		ASSERT(node->curr_index_batch_size >= data_len);
1591		node->curr_index_batch_size -= data_len;
1592	}
1593
1594	btrfs_release_delayed_item(item);
1595
1596	/* If we now have no more dir index items, we can release all leaves. */
1597	if (RB_EMPTY_ROOT(&node->ins_root.rb_root)) {
1598		btrfs_delayed_item_release_leaves(node, node->index_item_leaves);
1599		node->index_item_leaves = 0;
1600	}
1601
1602	mutex_unlock(&node->mutex);
1603	return 0;
1604}
1605
1606int btrfs_delete_delayed_dir_index(struct btrfs_trans_handle *trans,
1607				   struct btrfs_inode *dir, u64 index)
1608{
1609	struct btrfs_delayed_node *node;
1610	struct btrfs_delayed_item *item;
1611	int ret;
1612
1613	node = btrfs_get_or_create_delayed_node(dir);
1614	if (IS_ERR(node))
1615		return PTR_ERR(node);
1616
1617	ret = btrfs_delete_delayed_insertion_item(trans->fs_info, node, index);
1618	if (!ret)
1619		goto end;
1620
1621	item = btrfs_alloc_delayed_item(0, node, BTRFS_DELAYED_DELETION_ITEM);
1622	if (!item) {
1623		ret = -ENOMEM;
1624		goto end;
1625	}
1626
1627	item->index = index;
1628
1629	ret = btrfs_delayed_item_reserve_metadata(trans, item);
1630	/*
1631	 * we have reserved enough space when we start a new transaction,
1632	 * so reserving metadata failure is impossible.
1633	 */
1634	if (ret < 0) {
1635		btrfs_err(trans->fs_info,
1636"metadata reservation failed for delayed dir item deltiona, should have been reserved");
1637		btrfs_release_delayed_item(item);
1638		goto end;
1639	}
1640
1641	mutex_lock(&node->mutex);
1642	ret = __btrfs_add_delayed_item(node, item);
1643	if (unlikely(ret)) {
1644		btrfs_err(trans->fs_info,
1645			  "err add delayed dir index item(index: %llu) into the deletion tree of the delayed node(root id: %llu, inode id: %llu, errno: %d)",
1646			  index, node->root->root_key.objectid,
1647			  node->inode_id, ret);
1648		btrfs_delayed_item_release_metadata(dir->root, item);
1649		btrfs_release_delayed_item(item);
1650	}
1651	mutex_unlock(&node->mutex);
1652end:
1653	btrfs_release_delayed_node(node);
1654	return ret;
1655}
1656
1657int btrfs_inode_delayed_dir_index_count(struct btrfs_inode *inode)
1658{
1659	struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1660
1661	if (!delayed_node)
1662		return -ENOENT;
1663
1664	/*
1665	 * Since we have held i_mutex of this directory, it is impossible that
1666	 * a new directory index is added into the delayed node and index_cnt
1667	 * is updated now. So we needn't lock the delayed node.
1668	 */
1669	if (!delayed_node->index_cnt) {
1670		btrfs_release_delayed_node(delayed_node);
1671		return -EINVAL;
1672	}
1673
1674	inode->index_cnt = delayed_node->index_cnt;
1675	btrfs_release_delayed_node(delayed_node);
1676	return 0;
1677}
1678
1679bool btrfs_readdir_get_delayed_items(struct inode *inode,
1680				     u64 last_index,
1681				     struct list_head *ins_list,
1682				     struct list_head *del_list)
1683{
1684	struct btrfs_delayed_node *delayed_node;
1685	struct btrfs_delayed_item *item;
1686
1687	delayed_node = btrfs_get_delayed_node(BTRFS_I(inode));
1688	if (!delayed_node)
1689		return false;
1690
1691	/*
1692	 * We can only do one readdir with delayed items at a time because of
1693	 * item->readdir_list.
1694	 */
1695	btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_SHARED);
1696	btrfs_inode_lock(BTRFS_I(inode), 0);
1697
1698	mutex_lock(&delayed_node->mutex);
1699	item = __btrfs_first_delayed_insertion_item(delayed_node);
1700	while (item && item->index <= last_index) {
1701		refcount_inc(&item->refs);
1702		list_add_tail(&item->readdir_list, ins_list);
1703		item = __btrfs_next_delayed_item(item);
1704	}
1705
1706	item = __btrfs_first_delayed_deletion_item(delayed_node);
1707	while (item && item->index <= last_index) {
1708		refcount_inc(&item->refs);
1709		list_add_tail(&item->readdir_list, del_list);
1710		item = __btrfs_next_delayed_item(item);
1711	}
1712	mutex_unlock(&delayed_node->mutex);
1713	/*
1714	 * This delayed node is still cached in the btrfs inode, so refs
1715	 * must be > 1 now, and we needn't check it is going to be freed
1716	 * or not.
1717	 *
1718	 * Besides that, this function is used to read dir, we do not
1719	 * insert/delete delayed items in this period. So we also needn't
1720	 * requeue or dequeue this delayed node.
1721	 */
1722	refcount_dec(&delayed_node->refs);
1723
1724	return true;
1725}
1726
1727void btrfs_readdir_put_delayed_items(struct inode *inode,
1728				     struct list_head *ins_list,
1729				     struct list_head *del_list)
1730{
1731	struct btrfs_delayed_item *curr, *next;
1732
1733	list_for_each_entry_safe(curr, next, ins_list, readdir_list) {
1734		list_del(&curr->readdir_list);
1735		if (refcount_dec_and_test(&curr->refs))
1736			kfree(curr);
1737	}
1738
1739	list_for_each_entry_safe(curr, next, del_list, readdir_list) {
1740		list_del(&curr->readdir_list);
1741		if (refcount_dec_and_test(&curr->refs))
1742			kfree(curr);
1743	}
1744
1745	/*
1746	 * The VFS is going to do up_read(), so we need to downgrade back to a
1747	 * read lock.
1748	 */
1749	downgrade_write(&inode->i_rwsem);
1750}
1751
1752int btrfs_should_delete_dir_index(struct list_head *del_list,
1753				  u64 index)
1754{
1755	struct btrfs_delayed_item *curr;
1756	int ret = 0;
1757
1758	list_for_each_entry(curr, del_list, readdir_list) {
1759		if (curr->index > index)
1760			break;
1761		if (curr->index == index) {
1762			ret = 1;
1763			break;
1764		}
1765	}
1766	return ret;
1767}
1768
1769/*
1770 * Read dir info stored in the delayed tree.
1771 */
1772int btrfs_readdir_delayed_dir_index(struct dir_context *ctx,
1773				    struct list_head *ins_list)
1774{
1775	struct btrfs_dir_item *di;
1776	struct btrfs_delayed_item *curr, *next;
1777	struct btrfs_key location;
1778	char *name;
1779	int name_len;
1780	int over = 0;
1781	unsigned char d_type;
1782
1783	/*
1784	 * Changing the data of the delayed item is impossible. So
1785	 * we needn't lock them. And we have held i_mutex of the
1786	 * directory, nobody can delete any directory indexes now.
1787	 */
1788	list_for_each_entry_safe(curr, next, ins_list, readdir_list) {
1789		list_del(&curr->readdir_list);
1790
1791		if (curr->index < ctx->pos) {
1792			if (refcount_dec_and_test(&curr->refs))
1793				kfree(curr);
1794			continue;
1795		}
1796
1797		ctx->pos = curr->index;
1798
1799		di = (struct btrfs_dir_item *)curr->data;
1800		name = (char *)(di + 1);
1801		name_len = btrfs_stack_dir_name_len(di);
1802
1803		d_type = fs_ftype_to_dtype(btrfs_dir_flags_to_ftype(di->type));
1804		btrfs_disk_key_to_cpu(&location, &di->location);
1805
1806		over = !dir_emit(ctx, name, name_len,
1807			       location.objectid, d_type);
1808
1809		if (refcount_dec_and_test(&curr->refs))
1810			kfree(curr);
1811
1812		if (over)
1813			return 1;
1814		ctx->pos++;
1815	}
1816	return 0;
1817}
1818
1819static void fill_stack_inode_item(struct btrfs_trans_handle *trans,
1820				  struct btrfs_inode_item *inode_item,
1821				  struct inode *inode)
1822{
1823	u64 flags;
1824
1825	btrfs_set_stack_inode_uid(inode_item, i_uid_read(inode));
1826	btrfs_set_stack_inode_gid(inode_item, i_gid_read(inode));
1827	btrfs_set_stack_inode_size(inode_item, BTRFS_I(inode)->disk_i_size);
1828	btrfs_set_stack_inode_mode(inode_item, inode->i_mode);
1829	btrfs_set_stack_inode_nlink(inode_item, inode->i_nlink);
1830	btrfs_set_stack_inode_nbytes(inode_item, inode_get_bytes(inode));
1831	btrfs_set_stack_inode_generation(inode_item,
1832					 BTRFS_I(inode)->generation);
1833	btrfs_set_stack_inode_sequence(inode_item,
1834				       inode_peek_iversion(inode));
1835	btrfs_set_stack_inode_transid(inode_item, trans->transid);
1836	btrfs_set_stack_inode_rdev(inode_item, inode->i_rdev);
1837	flags = btrfs_inode_combine_flags(BTRFS_I(inode)->flags,
1838					  BTRFS_I(inode)->ro_flags);
1839	btrfs_set_stack_inode_flags(inode_item, flags);
1840	btrfs_set_stack_inode_block_group(inode_item, 0);
1841
1842	btrfs_set_stack_timespec_sec(&inode_item->atime,
1843				     inode_get_atime_sec(inode));
1844	btrfs_set_stack_timespec_nsec(&inode_item->atime,
1845				      inode_get_atime_nsec(inode));
1846
1847	btrfs_set_stack_timespec_sec(&inode_item->mtime,
1848				     inode_get_mtime_sec(inode));
1849	btrfs_set_stack_timespec_nsec(&inode_item->mtime,
1850				      inode_get_mtime_nsec(inode));
1851
1852	btrfs_set_stack_timespec_sec(&inode_item->ctime,
1853				     inode_get_ctime_sec(inode));
1854	btrfs_set_stack_timespec_nsec(&inode_item->ctime,
1855				      inode_get_ctime_nsec(inode));
1856
1857	btrfs_set_stack_timespec_sec(&inode_item->otime, BTRFS_I(inode)->i_otime_sec);
1858	btrfs_set_stack_timespec_nsec(&inode_item->otime, BTRFS_I(inode)->i_otime_nsec);
1859}
1860
1861int btrfs_fill_inode(struct inode *inode, u32 *rdev)
1862{
1863	struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
1864	struct btrfs_delayed_node *delayed_node;
1865	struct btrfs_inode_item *inode_item;
1866
1867	delayed_node = btrfs_get_delayed_node(BTRFS_I(inode));
1868	if (!delayed_node)
1869		return -ENOENT;
1870
1871	mutex_lock(&delayed_node->mutex);
1872	if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1873		mutex_unlock(&delayed_node->mutex);
1874		btrfs_release_delayed_node(delayed_node);
1875		return -ENOENT;
1876	}
1877
1878	inode_item = &delayed_node->inode_item;
1879
1880	i_uid_write(inode, btrfs_stack_inode_uid(inode_item));
1881	i_gid_write(inode, btrfs_stack_inode_gid(inode_item));
1882	btrfs_i_size_write(BTRFS_I(inode), btrfs_stack_inode_size(inode_item));
1883	btrfs_inode_set_file_extent_range(BTRFS_I(inode), 0,
1884			round_up(i_size_read(inode), fs_info->sectorsize));
1885	inode->i_mode = btrfs_stack_inode_mode(inode_item);
1886	set_nlink(inode, btrfs_stack_inode_nlink(inode_item));
1887	inode_set_bytes(inode, btrfs_stack_inode_nbytes(inode_item));
1888	BTRFS_I(inode)->generation = btrfs_stack_inode_generation(inode_item);
1889        BTRFS_I(inode)->last_trans = btrfs_stack_inode_transid(inode_item);
1890
1891	inode_set_iversion_queried(inode,
1892				   btrfs_stack_inode_sequence(inode_item));
1893	inode->i_rdev = 0;
1894	*rdev = btrfs_stack_inode_rdev(inode_item);
1895	btrfs_inode_split_flags(btrfs_stack_inode_flags(inode_item),
1896				&BTRFS_I(inode)->flags, &BTRFS_I(inode)->ro_flags);
1897
1898	inode_set_atime(inode, btrfs_stack_timespec_sec(&inode_item->atime),
1899			btrfs_stack_timespec_nsec(&inode_item->atime));
1900
1901	inode_set_mtime(inode, btrfs_stack_timespec_sec(&inode_item->mtime),
1902			btrfs_stack_timespec_nsec(&inode_item->mtime));
1903
1904	inode_set_ctime(inode, btrfs_stack_timespec_sec(&inode_item->ctime),
1905			btrfs_stack_timespec_nsec(&inode_item->ctime));
1906
1907	BTRFS_I(inode)->i_otime_sec = btrfs_stack_timespec_sec(&inode_item->otime);
1908	BTRFS_I(inode)->i_otime_nsec = btrfs_stack_timespec_nsec(&inode_item->otime);
1909
1910	inode->i_generation = BTRFS_I(inode)->generation;
1911	BTRFS_I(inode)->index_cnt = (u64)-1;
 
1912
1913	mutex_unlock(&delayed_node->mutex);
1914	btrfs_release_delayed_node(delayed_node);
1915	return 0;
1916}
1917
1918int btrfs_delayed_update_inode(struct btrfs_trans_handle *trans,
1919			       struct btrfs_inode *inode)
1920{
1921	struct btrfs_root *root = inode->root;
1922	struct btrfs_delayed_node *delayed_node;
1923	int ret = 0;
1924
1925	delayed_node = btrfs_get_or_create_delayed_node(inode);
1926	if (IS_ERR(delayed_node))
1927		return PTR_ERR(delayed_node);
1928
1929	mutex_lock(&delayed_node->mutex);
1930	if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1931		fill_stack_inode_item(trans, &delayed_node->inode_item,
1932				      &inode->vfs_inode);
1933		goto release_node;
1934	}
1935
1936	ret = btrfs_delayed_inode_reserve_metadata(trans, root, delayed_node);
1937	if (ret)
1938		goto release_node;
1939
1940	fill_stack_inode_item(trans, &delayed_node->inode_item, &inode->vfs_inode);
1941	set_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags);
1942	delayed_node->count++;
1943	atomic_inc(&root->fs_info->delayed_root->items);
1944release_node:
1945	mutex_unlock(&delayed_node->mutex);
1946	btrfs_release_delayed_node(delayed_node);
1947	return ret;
1948}
1949
1950int btrfs_delayed_delete_inode_ref(struct btrfs_inode *inode)
1951{
1952	struct btrfs_fs_info *fs_info = inode->root->fs_info;
1953	struct btrfs_delayed_node *delayed_node;
1954
1955	/*
1956	 * we don't do delayed inode updates during log recovery because it
1957	 * leads to enospc problems.  This means we also can't do
1958	 * delayed inode refs
1959	 */
1960	if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
1961		return -EAGAIN;
1962
1963	delayed_node = btrfs_get_or_create_delayed_node(inode);
1964	if (IS_ERR(delayed_node))
1965		return PTR_ERR(delayed_node);
1966
1967	/*
1968	 * We don't reserve space for inode ref deletion is because:
1969	 * - We ONLY do async inode ref deletion for the inode who has only
1970	 *   one link(i_nlink == 1), it means there is only one inode ref.
1971	 *   And in most case, the inode ref and the inode item are in the
1972	 *   same leaf, and we will deal with them at the same time.
1973	 *   Since we are sure we will reserve the space for the inode item,
1974	 *   it is unnecessary to reserve space for inode ref deletion.
1975	 * - If the inode ref and the inode item are not in the same leaf,
1976	 *   We also needn't worry about enospc problem, because we reserve
1977	 *   much more space for the inode update than it needs.
1978	 * - At the worst, we can steal some space from the global reservation.
1979	 *   It is very rare.
1980	 */
1981	mutex_lock(&delayed_node->mutex);
1982	if (test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags))
1983		goto release_node;
1984
1985	set_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags);
1986	delayed_node->count++;
1987	atomic_inc(&fs_info->delayed_root->items);
1988release_node:
1989	mutex_unlock(&delayed_node->mutex);
1990	btrfs_release_delayed_node(delayed_node);
1991	return 0;
1992}
1993
1994static void __btrfs_kill_delayed_node(struct btrfs_delayed_node *delayed_node)
1995{
1996	struct btrfs_root *root = delayed_node->root;
1997	struct btrfs_fs_info *fs_info = root->fs_info;
1998	struct btrfs_delayed_item *curr_item, *prev_item;
1999
2000	mutex_lock(&delayed_node->mutex);
2001	curr_item = __btrfs_first_delayed_insertion_item(delayed_node);
2002	while (curr_item) {
2003		prev_item = curr_item;
2004		curr_item = __btrfs_next_delayed_item(prev_item);
2005		btrfs_release_delayed_item(prev_item);
2006	}
2007
2008	if (delayed_node->index_item_leaves > 0) {
2009		btrfs_delayed_item_release_leaves(delayed_node,
2010					  delayed_node->index_item_leaves);
2011		delayed_node->index_item_leaves = 0;
2012	}
2013
2014	curr_item = __btrfs_first_delayed_deletion_item(delayed_node);
2015	while (curr_item) {
2016		btrfs_delayed_item_release_metadata(root, curr_item);
2017		prev_item = curr_item;
2018		curr_item = __btrfs_next_delayed_item(prev_item);
2019		btrfs_release_delayed_item(prev_item);
2020	}
2021
2022	btrfs_release_delayed_iref(delayed_node);
2023
2024	if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
2025		btrfs_delayed_inode_release_metadata(fs_info, delayed_node, false);
2026		btrfs_release_delayed_inode(delayed_node);
2027	}
2028	mutex_unlock(&delayed_node->mutex);
2029}
2030
2031void btrfs_kill_delayed_inode_items(struct btrfs_inode *inode)
2032{
2033	struct btrfs_delayed_node *delayed_node;
2034
2035	delayed_node = btrfs_get_delayed_node(inode);
2036	if (!delayed_node)
2037		return;
2038
2039	__btrfs_kill_delayed_node(delayed_node);
2040	btrfs_release_delayed_node(delayed_node);
2041}
2042
2043void btrfs_kill_all_delayed_nodes(struct btrfs_root *root)
2044{
2045	unsigned long index = 0;
2046	struct btrfs_delayed_node *delayed_nodes[8];
2047
2048	while (1) {
2049		struct btrfs_delayed_node *node;
2050		int count;
2051
2052		spin_lock(&root->inode_lock);
2053		if (xa_empty(&root->delayed_nodes)) {
2054			spin_unlock(&root->inode_lock);
2055			return;
2056		}
2057
2058		count = 0;
2059		xa_for_each_start(&root->delayed_nodes, index, node, index) {
2060			/*
2061			 * Don't increase refs in case the node is dead and
2062			 * about to be removed from the tree in the loop below
2063			 */
2064			if (refcount_inc_not_zero(&node->refs)) {
2065				delayed_nodes[count] = node;
2066				count++;
2067			}
2068			if (count >= ARRAY_SIZE(delayed_nodes))
2069				break;
2070		}
2071		spin_unlock(&root->inode_lock);
2072		index++;
2073
2074		for (int i = 0; i < count; i++) {
2075			__btrfs_kill_delayed_node(delayed_nodes[i]);
2076			btrfs_release_delayed_node(delayed_nodes[i]);
2077		}
2078	}
2079}
2080
2081void btrfs_destroy_delayed_inodes(struct btrfs_fs_info *fs_info)
2082{
2083	struct btrfs_delayed_node *curr_node, *prev_node;
2084
2085	curr_node = btrfs_first_delayed_node(fs_info->delayed_root);
2086	while (curr_node) {
2087		__btrfs_kill_delayed_node(curr_node);
2088
2089		prev_node = curr_node;
2090		curr_node = btrfs_next_delayed_node(curr_node);
2091		btrfs_release_delayed_node(prev_node);
2092	}
2093}
2094
2095void btrfs_log_get_delayed_items(struct btrfs_inode *inode,
2096				 struct list_head *ins_list,
2097				 struct list_head *del_list)
2098{
2099	struct btrfs_delayed_node *node;
2100	struct btrfs_delayed_item *item;
2101
2102	node = btrfs_get_delayed_node(inode);
2103	if (!node)
2104		return;
2105
2106	mutex_lock(&node->mutex);
2107	item = __btrfs_first_delayed_insertion_item(node);
2108	while (item) {
2109		/*
2110		 * It's possible that the item is already in a log list. This
2111		 * can happen in case two tasks are trying to log the same
2112		 * directory. For example if we have tasks A and task B:
2113		 *
2114		 * Task A collected the delayed items into a log list while
2115		 * under the inode's log_mutex (at btrfs_log_inode()), but it
2116		 * only releases the items after logging the inodes they point
2117		 * to (if they are new inodes), which happens after unlocking
2118		 * the log mutex;
2119		 *
2120		 * Task B enters btrfs_log_inode() and acquires the log_mutex
2121		 * of the same directory inode, before task B releases the
2122		 * delayed items. This can happen for example when logging some
2123		 * inode we need to trigger logging of its parent directory, so
2124		 * logging two files that have the same parent directory can
2125		 * lead to this.
2126		 *
2127		 * If this happens, just ignore delayed items already in a log
2128		 * list. All the tasks logging the directory are under a log
2129		 * transaction and whichever finishes first can not sync the log
2130		 * before the other completes and leaves the log transaction.
2131		 */
2132		if (!item->logged && list_empty(&item->log_list)) {
2133			refcount_inc(&item->refs);
2134			list_add_tail(&item->log_list, ins_list);
2135		}
2136		item = __btrfs_next_delayed_item(item);
2137	}
2138
2139	item = __btrfs_first_delayed_deletion_item(node);
2140	while (item) {
2141		/* It may be non-empty, for the same reason mentioned above. */
2142		if (!item->logged && list_empty(&item->log_list)) {
2143			refcount_inc(&item->refs);
2144			list_add_tail(&item->log_list, del_list);
2145		}
2146		item = __btrfs_next_delayed_item(item);
2147	}
2148	mutex_unlock(&node->mutex);
2149
2150	/*
2151	 * We are called during inode logging, which means the inode is in use
2152	 * and can not be evicted before we finish logging the inode. So we never
2153	 * have the last reference on the delayed inode.
2154	 * Also, we don't use btrfs_release_delayed_node() because that would
2155	 * requeue the delayed inode (change its order in the list of prepared
2156	 * nodes) and we don't want to do such change because we don't create or
2157	 * delete delayed items.
2158	 */
2159	ASSERT(refcount_read(&node->refs) > 1);
2160	refcount_dec(&node->refs);
2161}
2162
2163void btrfs_log_put_delayed_items(struct btrfs_inode *inode,
2164				 struct list_head *ins_list,
2165				 struct list_head *del_list)
2166{
2167	struct btrfs_delayed_node *node;
2168	struct btrfs_delayed_item *item;
2169	struct btrfs_delayed_item *next;
2170
2171	node = btrfs_get_delayed_node(inode);
2172	if (!node)
2173		return;
2174
2175	mutex_lock(&node->mutex);
2176
2177	list_for_each_entry_safe(item, next, ins_list, log_list) {
2178		item->logged = true;
2179		list_del_init(&item->log_list);
2180		if (refcount_dec_and_test(&item->refs))
2181			kfree(item);
2182	}
2183
2184	list_for_each_entry_safe(item, next, del_list, log_list) {
2185		item->logged = true;
2186		list_del_init(&item->log_list);
2187		if (refcount_dec_and_test(&item->refs))
2188			kfree(item);
2189	}
2190
2191	mutex_unlock(&node->mutex);
2192
2193	/*
2194	 * We are called during inode logging, which means the inode is in use
2195	 * and can not be evicted before we finish logging the inode. So we never
2196	 * have the last reference on the delayed inode.
2197	 * Also, we don't use btrfs_release_delayed_node() because that would
2198	 * requeue the delayed inode (change its order in the list of prepared
2199	 * nodes) and we don't want to do such change because we don't create or
2200	 * delete delayed items.
2201	 */
2202	ASSERT(refcount_read(&node->refs) > 1);
2203	refcount_dec(&node->refs);
2204}