delayed-inode.c

   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(btrfs_delayed_node, 0);
  32	if (!delayed_node_cache)
  33		return -ENOMEM;
  34	return 0;
  35}
  36
  37void __cold btrfs_delayed_inode_exit(void)
  38{
  39	kmem_cache_destroy(delayed_node_cache);
  40}
  41
  42void btrfs_init_delayed_root(struct btrfs_delayed_root *delayed_root)
  43{
  44	atomic_set(&delayed_root->items, 0);
  45	atomic_set(&delayed_root->items_seq, 0);
  46	delayed_root->nodes = 0;
  47	spin_lock_init(&delayed_root->lock);
  48	init_waitqueue_head(&delayed_root->wait);
  49	INIT_LIST_HEAD(&delayed_root->node_list);
  50	INIT_LIST_HEAD(&delayed_root->prepare_list);
  51}
  52
  53static inline void btrfs_init_delayed_node(
  54				struct btrfs_delayed_node *delayed_node,
  55				struct btrfs_root *root, u64 inode_id)
  56{
  57	delayed_node->root = root;
  58	delayed_node->inode_id = inode_id;
  59	refcount_set(&delayed_node->refs, 0);
  60	btrfs_delayed_node_ref_tracker_dir_init(delayed_node);
  61	delayed_node->ins_root = RB_ROOT_CACHED;
  62	delayed_node->del_root = RB_ROOT_CACHED;
  63	mutex_init(&delayed_node->mutex);
  64	INIT_LIST_HEAD(&delayed_node->n_list);
  65	INIT_LIST_HEAD(&delayed_node->p_list);
  66}
  67
  68static struct btrfs_delayed_node *btrfs_get_delayed_node(
  69		struct btrfs_inode *btrfs_inode,
  70		struct btrfs_ref_tracker *tracker)
  71{
  72	struct btrfs_root *root = btrfs_inode->root;
  73	u64 ino = btrfs_ino(btrfs_inode);
  74	struct btrfs_delayed_node *node;
  75
  76	node = READ_ONCE(btrfs_inode->delayed_node);
  77	if (node) {
  78		refcount_inc(&node->refs);
  79		btrfs_delayed_node_ref_tracker_alloc(node, tracker, GFP_NOFS);
  80		return node;
  81	}
  82
  83	xa_lock(&root->delayed_nodes);
  84	node = xa_load(&root->delayed_nodes, ino);
  85
  86	if (node) {
  87		if (btrfs_inode->delayed_node) {
  88			refcount_inc(&node->refs);	/* can be accessed */
  89			btrfs_delayed_node_ref_tracker_alloc(node, tracker, GFP_ATOMIC);
  90			BUG_ON(btrfs_inode->delayed_node != node);
  91			xa_unlock(&root->delayed_nodes);
  92			return node;
  93		}
  94
  95		/*
  96		 * It's possible that we're racing into the middle of removing
  97		 * this node from the xarray.  In this case, the refcount
  98		 * was zero and it should never go back to one.  Just return
  99		 * NULL like it was never in the xarray at all; our release
 100		 * function is in the process of removing it.
 101		 *
 102		 * Some implementations of refcount_inc refuse to bump the
 103		 * refcount once it has hit zero.  If we don't do this dance
 104		 * here, refcount_inc() may decide to just WARN_ONCE() instead
 105		 * of actually bumping the refcount.
 106		 *
 107		 * If this node is properly in the xarray, we want to bump the
 108		 * refcount twice, once for the inode and once for this get
 109		 * operation.
 110		 */
 111		if (refcount_inc_not_zero(&node->refs)) {
 112			refcount_inc(&node->refs);
 113			btrfs_delayed_node_ref_tracker_alloc(node, tracker, GFP_ATOMIC);
 114			btrfs_delayed_node_ref_tracker_alloc(node, &node->inode_cache_tracker,
 115							     GFP_ATOMIC);
 116			btrfs_inode->delayed_node = node;
 117		} else {
 118			node = NULL;
 119		}
 120
 121		xa_unlock(&root->delayed_nodes);
 122		return node;
 123	}
 124	xa_unlock(&root->delayed_nodes);
 125
 126	return NULL;
 127}
 128
 129/*
 130 * Look up an existing delayed node associated with @btrfs_inode or create a new
 131 * one and insert it to the delayed nodes of the root.
 132 *
 133 * Return the delayed node, or error pointer on failure.
 134 */
 135static struct btrfs_delayed_node *btrfs_get_or_create_delayed_node(
 136		struct btrfs_inode *btrfs_inode,
 137		struct btrfs_ref_tracker *tracker)
 138{
 139	struct btrfs_delayed_node *node;
 140	struct btrfs_root *root = btrfs_inode->root;
 141	u64 ino = btrfs_ino(btrfs_inode);
 142	int ret;
 143	void *ptr;
 144
 145again:
 146	node = btrfs_get_delayed_node(btrfs_inode, tracker);
 147	if (node)
 148		return node;
 149
 150	node = kmem_cache_zalloc(delayed_node_cache, GFP_NOFS);
 151	if (!node)
 152		return ERR_PTR(-ENOMEM);
 153	btrfs_init_delayed_node(node, root, ino);
 154
 155	/* Cached in the inode and can be accessed. */
 156	refcount_set(&node->refs, 2);
 157	btrfs_delayed_node_ref_tracker_alloc(node, tracker, GFP_NOFS);
 158	btrfs_delayed_node_ref_tracker_alloc(node, &node->inode_cache_tracker, GFP_NOFS);
 159
 160	/* Allocate and reserve the slot, from now it can return a NULL from xa_load(). */
 161	ret = xa_reserve(&root->delayed_nodes, ino, GFP_NOFS);
 162	if (ret == -ENOMEM)
 163		goto cleanup;
 164
 165	xa_lock(&root->delayed_nodes);
 166	ptr = xa_load(&root->delayed_nodes, ino);
 167	if (ptr) {
 168		/* Somebody inserted it, go back and read it. */
 169		xa_unlock(&root->delayed_nodes);
 170		goto cleanup;
 171	}
 172	ptr = __xa_store(&root->delayed_nodes, ino, node, GFP_ATOMIC);
 173	ASSERT(xa_err(ptr) != -EINVAL);
 174	ASSERT(xa_err(ptr) != -ENOMEM);
 175	ASSERT(ptr == NULL);
 176	btrfs_inode->delayed_node = node;
 177	xa_unlock(&root->delayed_nodes);
 178
 179	return node;
 180cleanup:
 181	btrfs_delayed_node_ref_tracker_free(node, tracker);
 182	btrfs_delayed_node_ref_tracker_free(node, &node->inode_cache_tracker);
 183	btrfs_delayed_node_ref_tracker_dir_exit(node);
 184	kmem_cache_free(delayed_node_cache, node);
 185	if (ret)
 186		return ERR_PTR(ret);
 187	goto again;
 188}
 189
 190/*
 191 * Call it when holding delayed_node->mutex
 192 *
 193 * If mod = 1, add this node into the prepared list.
 194 */
 195static void btrfs_queue_delayed_node(struct btrfs_delayed_root *root,
 196				     struct btrfs_delayed_node *node,
 197				     int mod)
 198{
 199	spin_lock(&root->lock);
 200	if (test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
 201		if (!list_empty(&node->p_list))
 202			list_move_tail(&node->p_list, &root->prepare_list);
 203		else if (mod)
 204			list_add_tail(&node->p_list, &root->prepare_list);
 205	} else {
 206		list_add_tail(&node->n_list, &root->node_list);
 207		list_add_tail(&node->p_list, &root->prepare_list);
 208		refcount_inc(&node->refs);	/* inserted into list */
 209		btrfs_delayed_node_ref_tracker_alloc(node, &node->node_list_tracker,
 210						     GFP_ATOMIC);
 211		root->nodes++;
 212		set_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags);
 213	}
 214	spin_unlock(&root->lock);
 215}
 216
 217/* Call it when holding delayed_node->mutex */
 218static void btrfs_dequeue_delayed_node(struct btrfs_delayed_root *root,
 219				       struct btrfs_delayed_node *node)
 220{
 221	spin_lock(&root->lock);
 222	if (test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
 223		root->nodes--;
 224		btrfs_delayed_node_ref_tracker_free(node, &node->node_list_tracker);
 225		refcount_dec(&node->refs);	/* not in the list */
 226		list_del_init(&node->n_list);
 227		if (!list_empty(&node->p_list))
 228			list_del_init(&node->p_list);
 229		clear_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags);
 230	}
 231	spin_unlock(&root->lock);
 232}
 233
 234static struct btrfs_delayed_node *btrfs_first_delayed_node(
 235			struct btrfs_delayed_root *delayed_root,
 236			struct btrfs_ref_tracker *tracker)
 237{
 238	struct btrfs_delayed_node *node;
 239
 240	spin_lock(&delayed_root->lock);
 241	node = list_first_entry_or_null(&delayed_root->node_list,
 242					struct btrfs_delayed_node, n_list);
 243	if (node) {
 244		refcount_inc(&node->refs);
 245		btrfs_delayed_node_ref_tracker_alloc(node, tracker, GFP_ATOMIC);
 246	}
 247	spin_unlock(&delayed_root->lock);
 248
 249	return node;
 250}
 251
 252static struct btrfs_delayed_node *btrfs_next_delayed_node(
 253						struct btrfs_delayed_node *node,
 254						struct btrfs_ref_tracker *tracker)
 255{
 256	struct btrfs_delayed_root *delayed_root;
 257	struct list_head *p;
 258	struct btrfs_delayed_node *next = NULL;
 259
 260	delayed_root = node->root->fs_info->delayed_root;
 261	spin_lock(&delayed_root->lock);
 262	if (!test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
 263		/* not in the list */
 264		if (list_empty(&delayed_root->node_list))
 265			goto out;
 266		p = delayed_root->node_list.next;
 267	} else if (list_is_last(&node->n_list, &delayed_root->node_list))
 268		goto out;
 269	else
 270		p = node->n_list.next;
 271
 272	next = list_entry(p, struct btrfs_delayed_node, n_list);
 273	refcount_inc(&next->refs);
 274	btrfs_delayed_node_ref_tracker_alloc(next, tracker, GFP_ATOMIC);
 275out:
 276	spin_unlock(&delayed_root->lock);
 277
 278	return next;
 279}
 280
 281static void __btrfs_release_delayed_node(
 282				struct btrfs_delayed_node *delayed_node,
 283				int mod, struct btrfs_ref_tracker *tracker)
 284{
 285	struct btrfs_delayed_root *delayed_root;
 286
 287	if (!delayed_node)
 288		return;
 289
 290	delayed_root = delayed_node->root->fs_info->delayed_root;
 291
 292	mutex_lock(&delayed_node->mutex);
 293	if (delayed_node->count)
 294		btrfs_queue_delayed_node(delayed_root, delayed_node, mod);
 295	else
 296		btrfs_dequeue_delayed_node(delayed_root, delayed_node);
 297	mutex_unlock(&delayed_node->mutex);
 298
 299	btrfs_delayed_node_ref_tracker_free(delayed_node, tracker);
 300	if (refcount_dec_and_test(&delayed_node->refs)) {
 301		struct btrfs_root *root = delayed_node->root;
 302
 303		xa_erase(&root->delayed_nodes, delayed_node->inode_id);
 304		/*
 305		 * Once our refcount goes to zero, nobody is allowed to bump it
 306		 * back up.  We can delete it now.
 307		 */
 308		ASSERT(refcount_read(&delayed_node->refs) == 0);
 309		btrfs_delayed_node_ref_tracker_dir_exit(delayed_node);
 310		kmem_cache_free(delayed_node_cache, delayed_node);
 311	}
 312}
 313
 314static inline void btrfs_release_delayed_node(struct btrfs_delayed_node *node,
 315					      struct btrfs_ref_tracker *tracker)
 316{
 317	__btrfs_release_delayed_node(node, 0, tracker);
 318}
 319
 320static struct btrfs_delayed_node *btrfs_first_prepared_delayed_node(
 321					struct btrfs_delayed_root *delayed_root,
 322					struct btrfs_ref_tracker *tracker)
 323{
 324	struct btrfs_delayed_node *node;
 325
 326	spin_lock(&delayed_root->lock);
 327	node = list_first_entry_or_null(&delayed_root->prepare_list,
 328					struct btrfs_delayed_node, p_list);
 329	if (node) {
 330		list_del_init(&node->p_list);
 331		refcount_inc(&node->refs);
 332		btrfs_delayed_node_ref_tracker_alloc(node, tracker, GFP_ATOMIC);
 333	}
 334	spin_unlock(&delayed_root->lock);
 335
 336	return node;
 337}
 338
 339static inline void btrfs_release_prepared_delayed_node(
 340					struct btrfs_delayed_node *node,
 341					struct btrfs_ref_tracker *tracker)
 342{
 343	__btrfs_release_delayed_node(node, 1, tracker);
 344}
 345
 346static struct btrfs_delayed_item *btrfs_alloc_delayed_item(u16 data_len,
 347					   struct btrfs_delayed_node *node,
 348					   enum btrfs_delayed_item_type type)
 349{
 350	struct btrfs_delayed_item *item;
 351
 352	item = kmalloc(struct_size(item, data, data_len), GFP_NOFS);
 353	if (item) {
 354		item->data_len = data_len;
 355		item->type = type;
 356		item->bytes_reserved = 0;
 357		item->delayed_node = node;
 358		RB_CLEAR_NODE(&item->rb_node);
 359		INIT_LIST_HEAD(&item->log_list);
 360		item->logged = false;
 361		refcount_set(&item->refs, 1);
 362	}
 363	return item;
 364}
 365
 366static int delayed_item_index_cmp(const void *key, const struct rb_node *node)
 367{
 368	const u64 *index = key;
 369	const struct btrfs_delayed_item *delayed_item = rb_entry(node,
 370						 struct btrfs_delayed_item, rb_node);
 371
 372	if (delayed_item->index < *index)
 373		return 1;
 374	else if (delayed_item->index > *index)
 375		return -1;
 376
 377	return 0;
 378}
 379
 380/*
 381 * Look up the delayed item by key.
 382 *
 383 * @delayed_node: pointer to the delayed node
 384 * @index:	  the dir index value to lookup (offset of a dir index key)
 385 *
 386 * Note: if we don't find the right item, we will return the prev item and
 387 * the next item.
 388 */
 389static struct btrfs_delayed_item *__btrfs_lookup_delayed_item(
 390				struct rb_root *root,
 391				u64 index)
 392{
 393	struct rb_node *node;
 394
 395	node = rb_find(&index, root, delayed_item_index_cmp);
 396	return rb_entry_safe(node, struct btrfs_delayed_item, rb_node);
 397}
 398
 399static int btrfs_delayed_item_cmp(const struct rb_node *new,
 400				  const struct rb_node *exist)
 401{
 402	const struct btrfs_delayed_item *new_item =
 403		rb_entry(new, struct btrfs_delayed_item, rb_node);
 404
 405	return delayed_item_index_cmp(&new_item->index, exist);
 406}
 407
 408static int __btrfs_add_delayed_item(struct btrfs_delayed_node *delayed_node,
 409				    struct btrfs_delayed_item *ins)
 410{
 411	struct rb_root_cached *root;
 412	struct rb_node *exist;
 413
 414	if (ins->type == BTRFS_DELAYED_INSERTION_ITEM)
 415		root = &delayed_node->ins_root;
 416	else
 417		root = &delayed_node->del_root;
 418
 419	exist = rb_find_add_cached(&ins->rb_node, root, btrfs_delayed_item_cmp);
 420	if (exist)
 421		return -EEXIST;
 422
 423	if (ins->type == BTRFS_DELAYED_INSERTION_ITEM &&
 424	    ins->index >= delayed_node->index_cnt)
 425		delayed_node->index_cnt = ins->index + 1;
 426
 427	delayed_node->count++;
 428	atomic_inc(&delayed_node->root->fs_info->delayed_root->items);
 429	return 0;
 430}
 431
 432static void finish_one_item(struct btrfs_delayed_root *delayed_root)
 433{
 434	int seq = atomic_inc_return(&delayed_root->items_seq);
 435
 436	/* atomic_dec_return implies a barrier */
 437	if ((atomic_dec_return(&delayed_root->items) <
 438	    BTRFS_DELAYED_BACKGROUND || seq % BTRFS_DELAYED_BATCH == 0))
 439		cond_wake_up_nomb(&delayed_root->wait);
 440}
 441
 442static void __btrfs_remove_delayed_item(struct btrfs_delayed_item *delayed_item)
 443{
 444	struct btrfs_delayed_node *delayed_node = delayed_item->delayed_node;
 445	struct rb_root_cached *root;
 446	struct btrfs_delayed_root *delayed_root;
 447
 448	/* Not inserted, ignore it. */
 449	if (RB_EMPTY_NODE(&delayed_item->rb_node))
 450		return;
 451
 452	/* If it's in a rbtree, then we need to have delayed node locked. */
 453	lockdep_assert_held(&delayed_node->mutex);
 454
 455	delayed_root = delayed_node->root->fs_info->delayed_root;
 456
 457	if (delayed_item->type == BTRFS_DELAYED_INSERTION_ITEM)
 458		root = &delayed_node->ins_root;
 459	else
 460		root = &delayed_node->del_root;
 461
 462	rb_erase_cached(&delayed_item->rb_node, root);
 463	RB_CLEAR_NODE(&delayed_item->rb_node);
 464	delayed_node->count--;
 465
 466	finish_one_item(delayed_root);
 467}
 468
 469static void btrfs_release_delayed_item(struct btrfs_delayed_item *item)
 470{
 471	if (item) {
 472		__btrfs_remove_delayed_item(item);
 473		if (refcount_dec_and_test(&item->refs))
 474			kfree(item);
 475	}
 476}
 477
 478static struct btrfs_delayed_item *__btrfs_first_delayed_insertion_item(
 479					struct btrfs_delayed_node *delayed_node)
 480{
 481	struct rb_node *p = rb_first_cached(&delayed_node->ins_root);
 482
 483	return rb_entry_safe(p, struct btrfs_delayed_item, rb_node);
 484}
 485
 486static struct btrfs_delayed_item *__btrfs_first_delayed_deletion_item(
 487					struct btrfs_delayed_node *delayed_node)
 488{
 489	struct rb_node *p = rb_first_cached(&delayed_node->del_root);
 490
 491	return rb_entry_safe(p, struct btrfs_delayed_item, rb_node);
 492}
 493
 494static struct btrfs_delayed_item *__btrfs_next_delayed_item(
 495						struct btrfs_delayed_item *item)
 496{
 497	struct rb_node *p = rb_next(&item->rb_node);
 498
 499	return rb_entry_safe(p, struct btrfs_delayed_item, rb_node);
 500}
 501
 502static int btrfs_delayed_item_reserve_metadata(struct btrfs_trans_handle *trans,
 503					       struct btrfs_delayed_item *item)
 504{
 505	struct btrfs_block_rsv *src_rsv;
 506	struct btrfs_block_rsv *dst_rsv;
 507	struct btrfs_fs_info *fs_info = trans->fs_info;
 508	u64 num_bytes;
 509	int ret;
 510
 511	if (!trans->bytes_reserved)
 512		return 0;
 513
 514	src_rsv = trans->block_rsv;
 515	dst_rsv = &fs_info->delayed_block_rsv;
 516
 517	num_bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
 518
 519	/*
 520	 * Here we migrate space rsv from transaction rsv, since have already
 521	 * reserved space when starting a transaction.  So no need to reserve
 522	 * qgroup space here.
 523	 */
 524	ret = btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, true);
 525	if (!ret) {
 526		trace_btrfs_space_reservation(fs_info, "delayed_item",
 527					      item->delayed_node->inode_id,
 528					      num_bytes, 1);
 529		/*
 530		 * For insertions we track reserved metadata space by accounting
 531		 * for the number of leaves that will be used, based on the delayed
 532		 * node's curr_index_batch_size and index_item_leaves fields.
 533		 */
 534		if (item->type == BTRFS_DELAYED_DELETION_ITEM)
 535			item->bytes_reserved = num_bytes;
 536	}
 537
 538	return ret;
 539}
 540
 541static void btrfs_delayed_item_release_metadata(struct btrfs_root *root,
 542						struct btrfs_delayed_item *item)
 543{
 544	struct btrfs_block_rsv *rsv;
 545	struct btrfs_fs_info *fs_info = root->fs_info;
 546
 547	if (!item->bytes_reserved)
 548		return;
 549
 550	rsv = &fs_info->delayed_block_rsv;
 551	/*
 552	 * Check btrfs_delayed_item_reserve_metadata() to see why we don't need
 553	 * to release/reserve qgroup space.
 554	 */
 555	trace_btrfs_space_reservation(fs_info, "delayed_item",
 556				      item->delayed_node->inode_id,
 557				      item->bytes_reserved, 0);
 558	btrfs_block_rsv_release(fs_info, rsv, item->bytes_reserved, NULL);
 559}
 560
 561static void btrfs_delayed_item_release_leaves(struct btrfs_delayed_node *node,
 562					      unsigned int num_leaves)
 563{
 564	struct btrfs_fs_info *fs_info = node->root->fs_info;
 565	const u64 bytes = btrfs_calc_insert_metadata_size(fs_info, num_leaves);
 566
 567	/* There are no space reservations during log replay, bail out. */
 568	if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
 569		return;
 570
 571	trace_btrfs_space_reservation(fs_info, "delayed_item", node->inode_id,
 572				      bytes, 0);
 573	btrfs_block_rsv_release(fs_info, &fs_info->delayed_block_rsv, bytes, NULL);
 574}
 575
 576static int btrfs_delayed_inode_reserve_metadata(
 577					struct btrfs_trans_handle *trans,
 578					struct btrfs_root *root,
 579					struct btrfs_delayed_node *node)
 580{
 581	struct btrfs_fs_info *fs_info = root->fs_info;
 582	struct btrfs_block_rsv *src_rsv;
 583	struct btrfs_block_rsv *dst_rsv;
 584	u64 num_bytes;
 585	int ret;
 586
 587	src_rsv = trans->block_rsv;
 588	dst_rsv = &fs_info->delayed_block_rsv;
 589
 590	num_bytes = btrfs_calc_metadata_size(fs_info, 1);
 591
 592	/*
 593	 * btrfs_dirty_inode will update the inode under btrfs_join_transaction
 594	 * which doesn't reserve space for speed.  This is a problem since we
 595	 * still need to reserve space for this update, so try to reserve the
 596	 * space.
 597	 *
 598	 * Now if src_rsv == delalloc_block_rsv we'll let it just steal since
 599	 * we always reserve enough to update the inode item.
 600	 */
 601	if (!src_rsv || (!trans->bytes_reserved &&
 602			 src_rsv->type != BTRFS_BLOCK_RSV_DELALLOC)) {
 603		ret = btrfs_qgroup_reserve_meta(root, num_bytes,
 604					  BTRFS_QGROUP_RSV_META_PREALLOC, true);
 605		if (ret < 0)
 606			return ret;
 607		ret = btrfs_block_rsv_add(fs_info, dst_rsv, num_bytes,
 608					  BTRFS_RESERVE_NO_FLUSH);
 609		/* NO_FLUSH could only fail with -ENOSPC */
 610		ASSERT(ret == 0 || ret == -ENOSPC);
 611		if (ret)
 612			btrfs_qgroup_free_meta_prealloc(root, num_bytes);
 613	} else {
 614		ret = btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, true);
 615	}
 616
 617	if (!ret) {
 618		trace_btrfs_space_reservation(fs_info, "delayed_inode",
 619					      node->inode_id, num_bytes, 1);
 620		node->bytes_reserved = num_bytes;
 621	}
 622
 623	return ret;
 624}
 625
 626static void btrfs_delayed_inode_release_metadata(struct btrfs_fs_info *fs_info,
 627						struct btrfs_delayed_node *node,
 628						bool qgroup_free)
 629{
 630	struct btrfs_block_rsv *rsv;
 631
 632	if (!node->bytes_reserved)
 633		return;
 634
 635	rsv = &fs_info->delayed_block_rsv;
 636	trace_btrfs_space_reservation(fs_info, "delayed_inode",
 637				      node->inode_id, node->bytes_reserved, 0);
 638	btrfs_block_rsv_release(fs_info, rsv, node->bytes_reserved, NULL);
 639	if (qgroup_free)
 640		btrfs_qgroup_free_meta_prealloc(node->root,
 641				node->bytes_reserved);
 642	else
 643		btrfs_qgroup_convert_reserved_meta(node->root,
 644				node->bytes_reserved);
 645	node->bytes_reserved = 0;
 646}
 647
 648/*
 649 * Insert a single delayed item or a batch of delayed items, as many as possible
 650 * that fit in a leaf. The delayed items (dir index keys) are sorted by their key
 651 * in the rbtree, and if there's a gap between two consecutive dir index items,
 652 * then it means at some point we had delayed dir indexes to add but they got
 653 * removed (by btrfs_delete_delayed_dir_index()) before we attempted to flush them
 654 * into the subvolume tree. Dir index keys also have their offsets coming from a
 655 * monotonically increasing counter, so we can't get new keys with an offset that
 656 * fits within a gap between delayed dir index items.
 657 */
 658static int btrfs_insert_delayed_item(struct btrfs_trans_handle *trans,
 659				     struct btrfs_root *root,
 660				     struct btrfs_path *path,
 661				     struct btrfs_delayed_item *first_item)
 662{
 663	struct btrfs_fs_info *fs_info = root->fs_info;
 664	struct btrfs_delayed_node *node = first_item->delayed_node;
 665	LIST_HEAD(item_list);
 666	struct btrfs_delayed_item *curr;
 667	struct btrfs_delayed_item *next;
 668	const int max_size = BTRFS_LEAF_DATA_SIZE(fs_info);
 669	struct btrfs_item_batch batch;
 670	struct btrfs_key first_key;
 671	const u32 first_data_size = first_item->data_len;
 672	int total_size;
 673	char AUTO_KFREE(ins_data);
 674	int ret;
 675	bool continuous_keys_only = false;
 676
 677	lockdep_assert_held(&node->mutex);
 678
 679	/*
 680	 * During normal operation the delayed index offset is continuously
 681	 * increasing, so we can batch insert all items as there will not be any
 682	 * overlapping keys in the tree.
 683	 *
 684	 * The exception to this is log replay, where we may have interleaved
 685	 * offsets in the tree, so our batch needs to be continuous keys only in
 686	 * order to ensure we do not end up with out of order items in our leaf.
 687	 */
 688	if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
 689		continuous_keys_only = true;
 690
 691	/*
 692	 * For delayed items to insert, we track reserved metadata bytes based
 693	 * on the number of leaves that we will use.
 694	 * See btrfs_insert_delayed_dir_index() and
 695	 * btrfs_delayed_item_reserve_metadata()).
 696	 */
 697	ASSERT(first_item->bytes_reserved == 0);
 698
 699	list_add_tail(&first_item->tree_list, &item_list);
 700	batch.total_data_size = first_data_size;
 701	batch.nr = 1;
 702	total_size = first_data_size + sizeof(struct btrfs_item);
 703	curr = first_item;
 704
 705	while (true) {
 706		int next_size;
 707
 708		next = __btrfs_next_delayed_item(curr);
 709		if (!next)
 710			break;
 711
 712		/*
 713		 * We cannot allow gaps in the key space if we're doing log
 714		 * replay.
 715		 */
 716		if (continuous_keys_only && (next->index != curr->index + 1))
 717			break;
 718
 719		ASSERT(next->bytes_reserved == 0);
 720
 721		next_size = next->data_len + sizeof(struct btrfs_item);
 722		if (total_size + next_size > max_size)
 723			break;
 724
 725		list_add_tail(&next->tree_list, &item_list);
 726		batch.nr++;
 727		total_size += next_size;
 728		batch.total_data_size += next->data_len;
 729		curr = next;
 730	}
 731
 732	if (batch.nr == 1) {
 733		first_key.objectid = node->inode_id;
 734		first_key.type = BTRFS_DIR_INDEX_KEY;
 735		first_key.offset = first_item->index;
 736		batch.keys = &first_key;
 737		batch.data_sizes = &first_data_size;
 738	} else {
 739		struct btrfs_key *ins_keys;
 740		u32 *ins_sizes;
 741		int i = 0;
 742
 743		ins_data = kmalloc_array(batch.nr,
 744					 sizeof(u32) + sizeof(struct btrfs_key), GFP_NOFS);
 745		if (!ins_data)
 746			return -ENOMEM;
 747		ins_sizes = (u32 *)ins_data;
 748		ins_keys = (struct btrfs_key *)(ins_data + batch.nr * sizeof(u32));
 749		batch.keys = ins_keys;
 750		batch.data_sizes = ins_sizes;
 751		list_for_each_entry(curr, &item_list, tree_list) {
 752			ins_keys[i].objectid = node->inode_id;
 753			ins_keys[i].type = BTRFS_DIR_INDEX_KEY;
 754			ins_keys[i].offset = curr->index;
 755			ins_sizes[i] = curr->data_len;
 756			i++;
 757		}
 758	}
 759
 760	ret = btrfs_insert_empty_items(trans, root, path, &batch);
 761	if (ret)
 762		return ret;
 763
 764	list_for_each_entry(curr, &item_list, tree_list) {
 765		char *data_ptr;
 766
 767		data_ptr = btrfs_item_ptr(path->nodes[0], path->slots[0], char);
 768		write_extent_buffer(path->nodes[0], &curr->data,
 769				    (unsigned long)data_ptr, curr->data_len);
 770		path->slots[0]++;
 771	}
 772
 773	/*
 774	 * Now release our path before releasing the delayed items and their
 775	 * metadata reservations, so that we don't block other tasks for more
 776	 * time than needed.
 777	 */
 778	btrfs_release_path(path);
 779
 780	ASSERT(node->index_item_leaves > 0);
 781
 782	/*
 783	 * For normal operations we will batch an entire leaf's worth of delayed
 784	 * items, so if there are more items to process we can decrement
 785	 * index_item_leaves by 1 as we inserted 1 leaf's worth of items.
 786	 *
 787	 * However for log replay we may not have inserted an entire leaf's
 788	 * worth of items, we may have not had continuous items, so decrementing
 789	 * here would mess up the index_item_leaves accounting.  For this case
 790	 * only clean up the accounting when there are no items left.
 791	 */
 792	if (next && !continuous_keys_only) {
 793		/*
 794		 * We inserted one batch of items into a leaf a there are more
 795		 * items to flush in a future batch, now release one unit of
 796		 * metadata space from the delayed block reserve, corresponding
 797		 * the leaf we just flushed to.
 798		 */
 799		btrfs_delayed_item_release_leaves(node, 1);
 800		node->index_item_leaves--;
 801	} else if (!next) {
 802		/*
 803		 * There are no more items to insert. We can have a number of
 804		 * reserved leaves > 1 here - this happens when many dir index
 805		 * items are added and then removed before they are flushed (file
 806		 * names with a very short life, never span a transaction). So
 807		 * release all remaining leaves.
 808		 */
 809		btrfs_delayed_item_release_leaves(node, node->index_item_leaves);
 810		node->index_item_leaves = 0;
 811	}
 812
 813	list_for_each_entry_safe(curr, next, &item_list, tree_list) {
 814		list_del(&curr->tree_list);
 815		btrfs_release_delayed_item(curr);
 816	}
 817
 818	return 0;
 819}
 820
 821static int btrfs_insert_delayed_items(struct btrfs_trans_handle *trans,
 822				      struct btrfs_path *path,
 823				      struct btrfs_root *root,
 824				      struct btrfs_delayed_node *node)
 825{
 826	int ret = 0;
 827
 828	while (ret == 0) {
 829		struct btrfs_delayed_item *curr;
 830
 831		mutex_lock(&node->mutex);
 832		curr = __btrfs_first_delayed_insertion_item(node);
 833		if (!curr) {
 834			mutex_unlock(&node->mutex);
 835			break;
 836		}
 837		ret = btrfs_insert_delayed_item(trans, root, path, curr);
 838		mutex_unlock(&node->mutex);
 839	}
 840
 841	return ret;
 842}
 843
 844static int btrfs_batch_delete_items(struct btrfs_trans_handle *trans,
 845				    struct btrfs_root *root,
 846				    struct btrfs_path *path,
 847				    struct btrfs_delayed_item *item)
 848{
 849	const u64 ino = item->delayed_node->inode_id;
 850	struct btrfs_fs_info *fs_info = root->fs_info;
 851	struct btrfs_delayed_item *curr, *next;
 852	struct extent_buffer *leaf = path->nodes[0];
 853	LIST_HEAD(batch_list);
 854	int nitems, slot, last_slot;
 855	int ret;
 856	u64 total_reserved_size = item->bytes_reserved;
 857
 858	ASSERT(leaf != NULL);
 859
 860	slot = path->slots[0];
 861	last_slot = btrfs_header_nritems(leaf) - 1;
 862	/*
 863	 * Our caller always gives us a path pointing to an existing item, so
 864	 * this can not happen.
 865	 */
 866	ASSERT(slot <= last_slot);
 867	if (WARN_ON(slot > last_slot))
 868		return -ENOENT;
 869
 870	nitems = 1;
 871	curr = item;
 872	list_add_tail(&curr->tree_list, &batch_list);
 873
 874	/*
 875	 * Keep checking if the next delayed item matches the next item in the
 876	 * leaf - if so, we can add it to the batch of items to delete from the
 877	 * leaf.
 878	 */
 879	while (slot < last_slot) {
 880		struct btrfs_key key;
 881
 882		next = __btrfs_next_delayed_item(curr);
 883		if (!next)
 884			break;
 885
 886		slot++;
 887		btrfs_item_key_to_cpu(leaf, &key, slot);
 888		if (key.objectid != ino ||
 889		    key.type != BTRFS_DIR_INDEX_KEY ||
 890		    key.offset != next->index)
 891			break;
 892		nitems++;
 893		curr = next;
 894		list_add_tail(&curr->tree_list, &batch_list);
 895		total_reserved_size += curr->bytes_reserved;
 896	}
 897
 898	ret = btrfs_del_items(trans, root, path, path->slots[0], nitems);
 899	if (ret)
 900		return ret;
 901
 902	/* In case of BTRFS_FS_LOG_RECOVERING items won't have reserved space */
 903	if (total_reserved_size > 0) {
 904		/*
 905		 * Check btrfs_delayed_item_reserve_metadata() to see why we
 906		 * don't need to release/reserve qgroup space.
 907		 */
 908		trace_btrfs_space_reservation(fs_info, "delayed_item", ino,
 909					      total_reserved_size, 0);
 910		btrfs_block_rsv_release(fs_info, &fs_info->delayed_block_rsv,
 911					total_reserved_size, NULL);
 912	}
 913
 914	list_for_each_entry_safe(curr, next, &batch_list, tree_list) {
 915		list_del(&curr->tree_list);
 916		btrfs_release_delayed_item(curr);
 917	}
 918
 919	return 0;
 920}
 921
 922static int btrfs_delete_delayed_items(struct btrfs_trans_handle *trans,
 923				      struct btrfs_path *path,
 924				      struct btrfs_root *root,
 925				      struct btrfs_delayed_node *node)
 926{
 927	struct btrfs_key key;
 928	int ret = 0;
 929
 930	key.objectid = node->inode_id;
 931	key.type = BTRFS_DIR_INDEX_KEY;
 932
 933	while (ret == 0) {
 934		struct btrfs_delayed_item *item;
 935
 936		mutex_lock(&node->mutex);
 937		item = __btrfs_first_delayed_deletion_item(node);
 938		if (!item) {
 939			mutex_unlock(&node->mutex);
 940			break;
 941		}
 942
 943		key.offset = item->index;
 944		ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 945		if (ret > 0) {
 946			/*
 947			 * There's no matching item in the leaf. This means we
 948			 * have already deleted this item in a past run of the
 949			 * delayed items. We ignore errors when running delayed
 950			 * items from an async context, through a work queue job
 951			 * running btrfs_async_run_delayed_root(), and don't
 952			 * release delayed items that failed to complete. This
 953			 * is because we will retry later, and at transaction
 954			 * commit time we always run delayed items and will
 955			 * then deal with errors if they fail to run again.
 956			 *
 957			 * So just release delayed items for which we can't find
 958			 * an item in the tree, and move to the next item.
 959			 */
 960			btrfs_release_path(path);
 961			btrfs_release_delayed_item(item);
 962			ret = 0;
 963		} else if (ret == 0) {
 964			ret = btrfs_batch_delete_items(trans, root, path, item);
 965			btrfs_release_path(path);
 966		}
 967
 968		/*
 969		 * We unlock and relock on each iteration, this is to prevent
 970		 * blocking other tasks for too long while we are being run from
 971		 * the async context (work queue job). Those tasks are typically
 972		 * running system calls like creat/mkdir/rename/unlink/etc which
 973		 * need to add delayed items to this delayed node.
 974		 */
 975		mutex_unlock(&node->mutex);
 976	}
 977
 978	return ret;
 979}
 980
 981static void btrfs_release_delayed_inode(struct btrfs_delayed_node *delayed_node)
 982{
 983	struct btrfs_delayed_root *delayed_root;
 984
 985	if (delayed_node &&
 986	    test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
 987		ASSERT(delayed_node->root);
 988		clear_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags);
 989		delayed_node->count--;
 990
 991		delayed_root = delayed_node->root->fs_info->delayed_root;
 992		finish_one_item(delayed_root);
 993	}
 994}
 995
 996static void btrfs_release_delayed_iref(struct btrfs_delayed_node *delayed_node)
 997{
 998
 999	if (test_and_clear_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags)) {
1000		struct btrfs_delayed_root *delayed_root;
1001
1002		ASSERT(delayed_node->root);
1003		delayed_node->count--;
1004
1005		delayed_root = delayed_node->root->fs_info->delayed_root;
1006		finish_one_item(delayed_root);
1007	}
1008}
1009
1010static int __btrfs_update_delayed_inode(struct btrfs_trans_handle *trans,
1011					struct btrfs_root *root,
1012					struct btrfs_path *path,
1013					struct btrfs_delayed_node *node)
1014{
1015	struct btrfs_fs_info *fs_info = root->fs_info;
1016	struct btrfs_key key;
1017	struct btrfs_inode_item *inode_item;
1018	struct extent_buffer *leaf;
1019	int mod;
1020	int ret;
1021
1022	key.objectid = node->inode_id;
1023	key.type = BTRFS_INODE_ITEM_KEY;
1024	key.offset = 0;
1025
1026	if (test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &node->flags))
1027		mod = -1;
1028	else
1029		mod = 1;
1030
1031	ret = btrfs_lookup_inode(trans, root, path, &key, mod);
1032	if (ret > 0)
1033		ret = -ENOENT;
1034	if (ret < 0) {
1035		/*
1036		 * If we fail to update the delayed inode we need to abort the
1037		 * transaction, because we could leave the inode with the
1038		 * improper counts behind.
1039		 */
1040		if (unlikely(ret != -ENOENT))
1041			btrfs_abort_transaction(trans, ret);
1042		goto out;
1043	}
1044
1045	leaf = path->nodes[0];
1046	inode_item = btrfs_item_ptr(leaf, path->slots[0],
1047				    struct btrfs_inode_item);
1048	write_extent_buffer(leaf, &node->inode_item, (unsigned long)inode_item,
1049			    sizeof(struct btrfs_inode_item));
1050
1051	if (!test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &node->flags))
1052		goto out;
1053
1054	/*
1055	 * Now we're going to delete the INODE_REF/EXTREF, which should be the
1056	 * only one ref left.  Check if the next item is an INODE_REF/EXTREF.
1057	 *
1058	 * But if we're the last item already, release and search for the last
1059	 * INODE_REF/EXTREF.
1060	 */
1061	if (path->slots[0] + 1 >= btrfs_header_nritems(leaf)) {
1062		key.objectid = node->inode_id;
1063		key.type = BTRFS_INODE_EXTREF_KEY;
1064		key.offset = (u64)-1;
1065
1066		btrfs_release_path(path);
1067		ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1068		if (unlikely(ret < 0)) {
1069			btrfs_abort_transaction(trans, ret);
1070			goto err_out;
1071		}
1072		ASSERT(ret > 0);
1073		ASSERT(path->slots[0] > 0);
1074		ret = 0;
1075		path->slots[0]--;
1076		leaf = path->nodes[0];
1077	} else {
1078		path->slots[0]++;
1079	}
1080	btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1081	if (key.objectid != node->inode_id)
1082		goto out;
1083	if (key.type != BTRFS_INODE_REF_KEY &&
1084	    key.type != BTRFS_INODE_EXTREF_KEY)
1085		goto out;
1086
1087	/*
1088	 * Delayed iref deletion is for the inode who has only one link,
1089	 * so there is only one iref. The case that several irefs are
1090	 * in the same item doesn't exist.
1091	 */
1092	ret = btrfs_del_item(trans, root, path);
1093	if (ret < 0)
1094		btrfs_abort_transaction(trans, ret);
1095out:
1096	btrfs_release_delayed_iref(node);
1097	btrfs_release_path(path);
1098err_out:
1099	btrfs_delayed_inode_release_metadata(fs_info, node, (ret < 0));
1100	btrfs_release_delayed_inode(node);
1101	return ret;
1102}
1103
1104static inline int btrfs_update_delayed_inode(struct btrfs_trans_handle *trans,
1105					     struct btrfs_root *root,
1106					     struct btrfs_path *path,
1107					     struct btrfs_delayed_node *node)
1108{
1109	int ret;
1110
1111	mutex_lock(&node->mutex);
1112	if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &node->flags)) {
1113		mutex_unlock(&node->mutex);
1114		return 0;
1115	}
1116
1117	ret = __btrfs_update_delayed_inode(trans, root, path, node);
1118	mutex_unlock(&node->mutex);
1119	return ret;
1120}
1121
1122static inline int
1123__btrfs_commit_inode_delayed_items(struct btrfs_trans_handle *trans,
1124				   struct btrfs_path *path,
1125				   struct btrfs_delayed_node *node)
1126{
1127	int ret;
1128
1129	ret = btrfs_insert_delayed_items(trans, path, node->root, node);
1130	if (ret)
1131		return ret;
1132
1133	ret = btrfs_delete_delayed_items(trans, path, node->root, node);
1134	if (ret)
1135		return ret;
1136
1137	ret = btrfs_record_root_in_trans(trans, node->root);
1138	if (ret)
1139		return ret;
1140	ret = btrfs_update_delayed_inode(trans, node->root, path, node);
1141	return ret;
1142}
1143
1144/*
1145 * Called when committing the transaction.
1146 * Returns 0 on success.
1147 * Returns < 0 on error and returns with an aborted transaction with any
1148 * outstanding delayed items cleaned up.
1149 */
1150static int __btrfs_run_delayed_items(struct btrfs_trans_handle *trans, int nr)
1151{
1152	struct btrfs_fs_info *fs_info = trans->fs_info;
1153	struct btrfs_delayed_root *delayed_root;
1154	struct btrfs_delayed_node *curr_node, *prev_node;
1155	struct btrfs_ref_tracker curr_delayed_node_tracker, prev_delayed_node_tracker;
1156	struct btrfs_path *path;
1157	struct btrfs_block_rsv *block_rsv;
1158	int ret = 0;
1159	bool count = (nr > 0);
1160
1161	if (TRANS_ABORTED(trans))
1162		return -EIO;
1163
1164	path = btrfs_alloc_path();
1165	if (!path)
1166		return -ENOMEM;
1167
1168	block_rsv = trans->block_rsv;
1169	trans->block_rsv = &fs_info->delayed_block_rsv;
1170
1171	delayed_root = fs_info->delayed_root;
1172
1173	curr_node = btrfs_first_delayed_node(delayed_root, &curr_delayed_node_tracker);
1174	while (curr_node && (!count || nr--)) {
1175		ret = __btrfs_commit_inode_delayed_items(trans, path,
1176							 curr_node);
1177		if (unlikely(ret)) {
1178			btrfs_abort_transaction(trans, ret);
1179			break;
1180		}
1181
1182		prev_node = curr_node;
1183		prev_delayed_node_tracker = curr_delayed_node_tracker;
1184		curr_node = btrfs_next_delayed_node(curr_node, &curr_delayed_node_tracker);
1185		/*
1186		 * See the comment below about releasing path before releasing
1187		 * node. If the commit of delayed items was successful the path
1188		 * should always be released, but in case of an error, it may
1189		 * point to locked extent buffers (a leaf at the very least).
1190		 */
1191		ASSERT(path->nodes[0] == NULL);
1192		btrfs_release_delayed_node(prev_node, &prev_delayed_node_tracker);
1193	}
1194
1195	/*
1196	 * Release the path to avoid a potential deadlock and lockdep splat when
1197	 * releasing the delayed node, as that requires taking the delayed node's
1198	 * mutex. If another task starts running delayed items before we take
1199	 * the mutex, it will first lock the mutex and then it may try to lock
1200	 * the same btree path (leaf).
1201	 */
1202	btrfs_free_path(path);
1203
1204	if (curr_node)
1205		btrfs_release_delayed_node(curr_node, &curr_delayed_node_tracker);
1206	trans->block_rsv = block_rsv;
1207
1208	return ret;
1209}
1210
1211int btrfs_run_delayed_items(struct btrfs_trans_handle *trans)
1212{
1213	return __btrfs_run_delayed_items(trans, -1);
1214}
1215
1216int btrfs_run_delayed_items_nr(struct btrfs_trans_handle *trans, int nr)
1217{
1218	return __btrfs_run_delayed_items(trans, nr);
1219}
1220
1221int btrfs_commit_inode_delayed_items(struct btrfs_trans_handle *trans,
1222				     struct btrfs_inode *inode)
1223{
1224	struct btrfs_ref_tracker delayed_node_tracker;
1225	struct btrfs_delayed_node *delayed_node =
1226		btrfs_get_delayed_node(inode, &delayed_node_tracker);
1227	BTRFS_PATH_AUTO_FREE(path);
1228	struct btrfs_block_rsv *block_rsv;
1229	int ret;
1230
1231	if (!delayed_node)
1232		return 0;
1233
1234	mutex_lock(&delayed_node->mutex);
1235	if (!delayed_node->count) {
1236		mutex_unlock(&delayed_node->mutex);
1237		btrfs_release_delayed_node(delayed_node, &delayed_node_tracker);
1238		return 0;
1239	}
1240	mutex_unlock(&delayed_node->mutex);
1241
1242	path = btrfs_alloc_path();
1243	if (!path) {
1244		btrfs_release_delayed_node(delayed_node, &delayed_node_tracker);
1245		return -ENOMEM;
1246	}
1247
1248	block_rsv = trans->block_rsv;
1249	trans->block_rsv = &delayed_node->root->fs_info->delayed_block_rsv;
1250
1251	ret = __btrfs_commit_inode_delayed_items(trans, path, delayed_node);
1252
1253	btrfs_release_delayed_node(delayed_node, &delayed_node_tracker);
1254	trans->block_rsv = block_rsv;
1255
1256	return ret;
1257}
1258
1259int btrfs_commit_inode_delayed_inode(struct btrfs_inode *inode)
1260{
1261	struct btrfs_fs_info *fs_info = inode->root->fs_info;
1262	struct btrfs_trans_handle *trans;
1263	struct btrfs_ref_tracker delayed_node_tracker;
1264	struct btrfs_delayed_node *delayed_node;
1265	struct btrfs_path *path;
1266	struct btrfs_block_rsv *block_rsv;
1267	int ret;
1268
1269	delayed_node = btrfs_get_delayed_node(inode, &delayed_node_tracker);
1270	if (!delayed_node)
1271		return 0;
1272
1273	mutex_lock(&delayed_node->mutex);
1274	if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1275		mutex_unlock(&delayed_node->mutex);
1276		btrfs_release_delayed_node(delayed_node, &delayed_node_tracker);
1277		return 0;
1278	}
1279	mutex_unlock(&delayed_node->mutex);
1280
1281	trans = btrfs_join_transaction(delayed_node->root);
1282	if (IS_ERR(trans)) {
1283		ret = PTR_ERR(trans);
1284		goto out;
1285	}
1286
1287	path = btrfs_alloc_path();
1288	if (!path) {
1289		ret = -ENOMEM;
1290		goto trans_out;
1291	}
1292
1293	block_rsv = trans->block_rsv;
1294	trans->block_rsv = &fs_info->delayed_block_rsv;
1295
1296	mutex_lock(&delayed_node->mutex);
1297	if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags))
1298		ret = __btrfs_update_delayed_inode(trans, delayed_node->root,
1299						   path, delayed_node);
1300	else
1301		ret = 0;
1302	mutex_unlock(&delayed_node->mutex);
1303
1304	btrfs_free_path(path);
1305	trans->block_rsv = block_rsv;
1306trans_out:
1307	btrfs_end_transaction(trans);
1308	btrfs_btree_balance_dirty(fs_info);
1309out:
1310	btrfs_release_delayed_node(delayed_node, &delayed_node_tracker);
1311
1312	return ret;
1313}
1314
1315void btrfs_remove_delayed_node(struct btrfs_inode *inode)
1316{
1317	struct btrfs_delayed_node *delayed_node;
1318
1319	delayed_node = READ_ONCE(inode->delayed_node);
1320	if (!delayed_node)
1321		return;
1322
1323	inode->delayed_node = NULL;
1324
1325	btrfs_release_delayed_node(delayed_node, &delayed_node->inode_cache_tracker);
1326}
1327
1328struct btrfs_async_delayed_work {
1329	struct btrfs_delayed_root *delayed_root;
1330	int nr;
1331	struct btrfs_work work;
1332};
1333
1334static void btrfs_async_run_delayed_root(struct btrfs_work *work)
1335{
1336	struct btrfs_async_delayed_work *async_work;
1337	struct btrfs_delayed_root *delayed_root;
1338	struct btrfs_trans_handle *trans;
1339	struct btrfs_path *path;
1340	struct btrfs_delayed_node *delayed_node = NULL;
1341	struct btrfs_ref_tracker delayed_node_tracker;
1342	struct btrfs_root *root;
1343	struct btrfs_block_rsv *block_rsv;
1344	int total_done = 0;
1345
1346	async_work = container_of(work, struct btrfs_async_delayed_work, work);
1347	delayed_root = async_work->delayed_root;
1348
1349	path = btrfs_alloc_path();
1350	if (!path)
1351		goto out;
1352
1353	do {
1354		if (atomic_read(&delayed_root->items) <
1355		    BTRFS_DELAYED_BACKGROUND / 2)
1356			break;
1357
1358		delayed_node = btrfs_first_prepared_delayed_node(delayed_root,
1359								 &delayed_node_tracker);
1360		if (!delayed_node)
1361			break;
1362
1363		root = delayed_node->root;
1364
1365		trans = btrfs_join_transaction(root);
1366		if (IS_ERR(trans)) {
1367			btrfs_release_path(path);
1368			btrfs_release_prepared_delayed_node(delayed_node,
1369							    &delayed_node_tracker);
1370			total_done++;
1371			continue;
1372		}
1373
1374		block_rsv = trans->block_rsv;
1375		trans->block_rsv = &root->fs_info->delayed_block_rsv;
1376
1377		__btrfs_commit_inode_delayed_items(trans, path, delayed_node);
1378
1379		trans->block_rsv = block_rsv;
1380		btrfs_end_transaction(trans);
1381		btrfs_btree_balance_dirty_nodelay(root->fs_info);
1382
1383		btrfs_release_path(path);
1384		btrfs_release_prepared_delayed_node(delayed_node,
1385						    &delayed_node_tracker);
1386		total_done++;
1387
1388	} while ((async_work->nr == 0 && total_done < BTRFS_DELAYED_WRITEBACK)
1389		 || total_done < async_work->nr);
1390
1391	btrfs_free_path(path);
1392out:
1393	wake_up(&delayed_root->wait);
1394	kfree(async_work);
1395}
1396
1397
1398static int btrfs_wq_run_delayed_node(struct btrfs_delayed_root *delayed_root,
1399				     struct btrfs_fs_info *fs_info, int nr)
1400{
1401	struct btrfs_async_delayed_work *async_work;
1402
1403	async_work = kmalloc(sizeof(*async_work), GFP_NOFS);
1404	if (!async_work)
1405		return -ENOMEM;
1406
1407	async_work->delayed_root = delayed_root;
1408	btrfs_init_work(&async_work->work, btrfs_async_run_delayed_root, NULL);
1409	async_work->nr = nr;
1410
1411	btrfs_queue_work(fs_info->delayed_workers, &async_work->work);
1412	return 0;
1413}
1414
1415void btrfs_assert_delayed_root_empty(struct btrfs_fs_info *fs_info)
1416{
1417	struct btrfs_ref_tracker delayed_node_tracker;
1418	struct btrfs_delayed_node *node;
1419
1420	node = btrfs_first_delayed_node( fs_info->delayed_root, &delayed_node_tracker);
1421	if (WARN_ON(node)) {
1422		btrfs_delayed_node_ref_tracker_free(node,
1423						    &delayed_node_tracker);
1424		refcount_dec(&node->refs);
1425	}
1426}
1427
1428static bool could_end_wait(struct btrfs_delayed_root *delayed_root, int seq)
1429{
1430	int val = atomic_read(&delayed_root->items_seq);
1431
1432	if (val < seq || val >= seq + BTRFS_DELAYED_BATCH)
1433		return true;
1434
1435	if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND)
1436		return true;
1437
1438	return false;
1439}
1440
1441void btrfs_balance_delayed_items(struct btrfs_fs_info *fs_info)
1442{
1443	struct btrfs_delayed_root *delayed_root = fs_info->delayed_root;
1444
1445	if ((atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND) ||
1446		btrfs_workqueue_normal_congested(fs_info->delayed_workers))
1447		return;
1448
1449	if (atomic_read(&delayed_root->items) >= BTRFS_DELAYED_WRITEBACK) {
1450		int seq;
1451		int ret;
1452
1453		seq = atomic_read(&delayed_root->items_seq);
1454
1455		ret = btrfs_wq_run_delayed_node(delayed_root, fs_info, 0);
1456		if (ret)
1457			return;
1458
1459		wait_event_interruptible(delayed_root->wait,
1460					 could_end_wait(delayed_root, seq));
1461		return;
1462	}
1463
1464	btrfs_wq_run_delayed_node(delayed_root, fs_info, BTRFS_DELAYED_BATCH);
1465}
1466
1467static void btrfs_release_dir_index_item_space(struct btrfs_trans_handle *trans)
1468{
1469	struct btrfs_fs_info *fs_info = trans->fs_info;
1470	const u64 bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
1471
1472	if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
1473		return;
1474
1475	/*
1476	 * Adding the new dir index item does not require touching another
1477	 * leaf, so we can release 1 unit of metadata that was previously
1478	 * reserved when starting the transaction. This applies only to
1479	 * the case where we had a transaction start and excludes the
1480	 * transaction join case (when replaying log trees).
1481	 */
1482	trace_btrfs_space_reservation(fs_info, "transaction",
1483				      trans->transid, bytes, 0);
1484	btrfs_block_rsv_release(fs_info, trans->block_rsv, bytes, NULL);
1485	ASSERT(trans->bytes_reserved >= bytes);
1486	trans->bytes_reserved -= bytes;
1487}
1488
1489/* Will return 0, -ENOMEM or -EEXIST (index number collision, unexpected). */
1490int btrfs_insert_delayed_dir_index(struct btrfs_trans_handle *trans,
1491				   const char *name, int name_len,
1492				   struct btrfs_inode *dir,
1493				   const struct btrfs_disk_key *disk_key, u8 flags,
1494				   u64 index)
1495{
1496	struct btrfs_fs_info *fs_info = trans->fs_info;
1497	const unsigned int leaf_data_size = BTRFS_LEAF_DATA_SIZE(fs_info);
1498	struct btrfs_delayed_node *delayed_node;
1499	struct btrfs_ref_tracker delayed_node_tracker;
1500	struct btrfs_delayed_item *delayed_item;
1501	struct btrfs_dir_item *dir_item;
1502	bool reserve_leaf_space;
1503	u32 data_len;
1504	int ret;
1505
1506	delayed_node = btrfs_get_or_create_delayed_node(dir, &delayed_node_tracker);
1507	if (IS_ERR(delayed_node))
1508		return PTR_ERR(delayed_node);
1509
1510	delayed_item = btrfs_alloc_delayed_item(sizeof(*dir_item) + name_len,
1511						delayed_node,
1512						BTRFS_DELAYED_INSERTION_ITEM);
1513	if (!delayed_item) {
1514		ret = -ENOMEM;
1515		goto release_node;
1516	}
1517
1518	delayed_item->index = index;
1519
1520	dir_item = (struct btrfs_dir_item *)delayed_item->data;
1521	dir_item->location = *disk_key;
1522	btrfs_set_stack_dir_transid(dir_item, trans->transid);
1523	btrfs_set_stack_dir_data_len(dir_item, 0);
1524	btrfs_set_stack_dir_name_len(dir_item, name_len);
1525	btrfs_set_stack_dir_flags(dir_item, flags);
1526	memcpy((char *)(dir_item + 1), name, name_len);
1527
1528	data_len = delayed_item->data_len + sizeof(struct btrfs_item);
1529
1530	mutex_lock(&delayed_node->mutex);
1531
1532	/*
1533	 * First attempt to insert the delayed item. This is to make the error
1534	 * handling path simpler in case we fail (-EEXIST). There's no risk of
1535	 * any other task coming in and running the delayed item before we do
1536	 * the metadata space reservation below, because we are holding the
1537	 * delayed node's mutex and that mutex must also be locked before the
1538	 * node's delayed items can be run.
1539	 */
1540	ret = __btrfs_add_delayed_item(delayed_node, delayed_item);
1541	if (unlikely(ret)) {
1542		btrfs_err(trans->fs_info,
1543"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",
1544			  name_len, name, index, btrfs_root_id(delayed_node->root),
1545			  delayed_node->inode_id, dir->index_cnt,
1546			  delayed_node->index_cnt, ret);
1547		btrfs_release_delayed_item(delayed_item);
1548		btrfs_release_dir_index_item_space(trans);
1549		mutex_unlock(&delayed_node->mutex);
1550		goto release_node;
1551	}
1552
1553	if (delayed_node->index_item_leaves == 0 ||
1554	    delayed_node->curr_index_batch_size + data_len > leaf_data_size) {
1555		delayed_node->curr_index_batch_size = data_len;
1556		reserve_leaf_space = true;
1557	} else {
1558		delayed_node->curr_index_batch_size += data_len;
1559		reserve_leaf_space = false;
1560	}
1561
1562	if (reserve_leaf_space) {
1563		ret = btrfs_delayed_item_reserve_metadata(trans, delayed_item);
1564		/*
1565		 * Space was reserved for a dir index item insertion when we
1566		 * started the transaction, so getting a failure here should be
1567		 * impossible.
1568		 */
1569		if (WARN_ON(ret)) {
1570			btrfs_release_delayed_item(delayed_item);
1571			mutex_unlock(&delayed_node->mutex);
1572			goto release_node;
1573		}
1574
1575		delayed_node->index_item_leaves++;
1576	} else {
1577		btrfs_release_dir_index_item_space(trans);
1578	}
1579	mutex_unlock(&delayed_node->mutex);
1580
1581release_node:
1582	btrfs_release_delayed_node(delayed_node, &delayed_node_tracker);
1583	return ret;
1584}
1585
1586static bool btrfs_delete_delayed_insertion_item(struct btrfs_delayed_node *node,
1587						u64 index)
1588{
1589	struct btrfs_delayed_item *item;
1590
1591	mutex_lock(&node->mutex);
1592	item = __btrfs_lookup_delayed_item(&node->ins_root.rb_root, index);
1593	if (!item) {
1594		mutex_unlock(&node->mutex);
1595		return false;
1596	}
1597
1598	/*
1599	 * For delayed items to insert, we track reserved metadata bytes based
1600	 * on the number of leaves that we will use.
1601	 * See btrfs_insert_delayed_dir_index() and
1602	 * btrfs_delayed_item_reserve_metadata()).
1603	 */
1604	ASSERT(item->bytes_reserved == 0);
1605	ASSERT(node->index_item_leaves > 0);
1606
1607	/*
1608	 * If there's only one leaf reserved, we can decrement this item from the
1609	 * current batch, otherwise we can not because we don't know which leaf
1610	 * it belongs to. With the current limit on delayed items, we rarely
1611	 * accumulate enough dir index items to fill more than one leaf (even
1612	 * when using a leaf size of 4K).
1613	 */
1614	if (node->index_item_leaves == 1) {
1615		const u32 data_len = item->data_len + sizeof(struct btrfs_item);
1616
1617		ASSERT(node->curr_index_batch_size >= data_len);
1618		node->curr_index_batch_size -= data_len;
1619	}
1620
1621	btrfs_release_delayed_item(item);
1622
1623	/* If we now have no more dir index items, we can release all leaves. */
1624	if (RB_EMPTY_ROOT(&node->ins_root.rb_root)) {
1625		btrfs_delayed_item_release_leaves(node, node->index_item_leaves);
1626		node->index_item_leaves = 0;
1627	}
1628
1629	mutex_unlock(&node->mutex);
1630	return true;
1631}
1632
1633int btrfs_delete_delayed_dir_index(struct btrfs_trans_handle *trans,
1634				   struct btrfs_inode *dir, u64 index)
1635{
1636	struct btrfs_delayed_node *node;
1637	struct btrfs_ref_tracker delayed_node_tracker;
1638	struct btrfs_delayed_item *item;
1639	int ret;
1640
1641	node = btrfs_get_or_create_delayed_node(dir, &delayed_node_tracker);
1642	if (IS_ERR(node))
1643		return PTR_ERR(node);
1644
1645	if (btrfs_delete_delayed_insertion_item(node, index)) {
1646		ret = 0;
1647		goto end;
1648	}
1649
1650	item = btrfs_alloc_delayed_item(0, node, BTRFS_DELAYED_DELETION_ITEM);
1651	if (!item) {
1652		ret = -ENOMEM;
1653		goto end;
1654	}
1655
1656	item->index = index;
1657
1658	ret = btrfs_delayed_item_reserve_metadata(trans, item);
1659	/*
1660	 * we have reserved enough space when we start a new transaction,
1661	 * so reserving metadata failure is impossible.
1662	 */
1663	if (ret < 0) {
1664		btrfs_err(trans->fs_info,
1665"metadata reservation failed for delayed dir item deletion, index: %llu, root: %llu, inode: %llu, error: %d",
1666			  index, btrfs_root_id(node->root), node->inode_id, ret);
1667		btrfs_release_delayed_item(item);
1668		goto end;
1669	}
1670
1671	mutex_lock(&node->mutex);
1672	ret = __btrfs_add_delayed_item(node, item);
1673	if (unlikely(ret)) {
1674		btrfs_err(trans->fs_info,
1675"failed to add delayed dir index item, root: %llu, inode: %llu, index: %llu, error: %d",
1676			  btrfs_root_id(node->root), node->inode_id, index, ret);
1677		btrfs_delayed_item_release_metadata(dir->root, item);
1678		btrfs_release_delayed_item(item);
1679	}
1680	mutex_unlock(&node->mutex);
1681end:
1682	btrfs_release_delayed_node(node, &delayed_node_tracker);
1683	return ret;
1684}
1685
1686int btrfs_inode_delayed_dir_index_count(struct btrfs_inode *inode)
1687{
1688	struct btrfs_ref_tracker delayed_node_tracker;
1689	struct btrfs_delayed_node *delayed_node;
1690
1691	delayed_node = btrfs_get_delayed_node(inode, &delayed_node_tracker);
1692	if (!delayed_node)
1693		return -ENOENT;
1694
1695	/*
1696	 * Since we have held i_mutex of this directory, it is impossible that
1697	 * a new directory index is added into the delayed node and index_cnt
1698	 * is updated now. So we needn't lock the delayed node.
1699	 */
1700	if (!delayed_node->index_cnt) {
1701		btrfs_release_delayed_node(delayed_node, &delayed_node_tracker);
1702		return -EINVAL;
1703	}
1704
1705	inode->index_cnt = delayed_node->index_cnt;
1706	btrfs_release_delayed_node(delayed_node, &delayed_node_tracker);
1707	return 0;
1708}
1709
1710bool btrfs_readdir_get_delayed_items(struct btrfs_inode *inode,
1711				     u64 last_index,
1712				     struct list_head *ins_list,
1713				     struct list_head *del_list)
1714{
1715	struct btrfs_delayed_node *delayed_node;
1716	struct btrfs_delayed_item *item;
1717	struct btrfs_ref_tracker delayed_node_tracker;
1718
1719	delayed_node = btrfs_get_delayed_node(inode, &delayed_node_tracker);
1720	if (!delayed_node)
1721		return false;
1722
1723	/*
1724	 * We can only do one readdir with delayed items at a time because of
1725	 * item->readdir_list.
1726	 */
1727	btrfs_inode_unlock(inode, BTRFS_ILOCK_SHARED);
1728	btrfs_inode_lock(inode, 0);
1729
1730	mutex_lock(&delayed_node->mutex);
1731	item = __btrfs_first_delayed_insertion_item(delayed_node);
1732	while (item && item->index <= last_index) {
1733		refcount_inc(&item->refs);
1734		list_add_tail(&item->readdir_list, ins_list);
1735		item = __btrfs_next_delayed_item(item);
1736	}
1737
1738	item = __btrfs_first_delayed_deletion_item(delayed_node);
1739	while (item && item->index <= last_index) {
1740		refcount_inc(&item->refs);
1741		list_add_tail(&item->readdir_list, del_list);
1742		item = __btrfs_next_delayed_item(item);
1743	}
1744	mutex_unlock(&delayed_node->mutex);
1745	/*
1746	 * This delayed node is still cached in the btrfs inode, so refs
1747	 * must be > 1 now, and we needn't check it is going to be freed
1748	 * or not.
1749	 *
1750	 * Besides that, this function is used to read dir, we do not
1751	 * insert/delete delayed items in this period. So we also needn't
1752	 * requeue or dequeue this delayed node.
1753	 */
1754	btrfs_delayed_node_ref_tracker_free(delayed_node, &delayed_node_tracker);
1755	refcount_dec(&delayed_node->refs);
1756
1757	return true;
1758}
1759
1760void btrfs_readdir_put_delayed_items(struct btrfs_inode *inode,
1761				     struct list_head *ins_list,
1762				     struct list_head *del_list)
1763{
1764	struct btrfs_delayed_item *curr, *next;
1765
1766	list_for_each_entry_safe(curr, next, ins_list, readdir_list) {
1767		list_del(&curr->readdir_list);
1768		if (refcount_dec_and_test(&curr->refs))
1769			kfree(curr);
1770	}
1771
1772	list_for_each_entry_safe(curr, next, del_list, readdir_list) {
1773		list_del(&curr->readdir_list);
1774		if (refcount_dec_and_test(&curr->refs))
1775			kfree(curr);
1776	}
1777
1778	/*
1779	 * The VFS is going to do up_read(), so we need to downgrade back to a
1780	 * read lock.
1781	 */
1782	downgrade_write(&inode->vfs_inode.i_rwsem);
1783}
1784
1785bool btrfs_should_delete_dir_index(const struct list_head *del_list, u64 index)
1786{
1787	struct btrfs_delayed_item *curr;
1788	bool ret = false;
1789
1790	list_for_each_entry(curr, del_list, readdir_list) {
1791		if (curr->index > index)
1792			break;
1793		if (curr->index == index) {
1794			ret = true;
1795			break;
1796		}
1797	}
1798	return ret;
1799}
1800
1801/*
1802 * Read dir info stored in the delayed tree.
1803 */
1804bool btrfs_readdir_delayed_dir_index(struct dir_context *ctx,
1805				     const struct list_head *ins_list)
1806{
1807	struct btrfs_dir_item *di;
1808	struct btrfs_delayed_item *curr, *next;
1809	struct btrfs_key location;
1810	char *name;
1811	int name_len;
1812	unsigned char d_type;
1813
1814	/*
1815	 * Changing the data of the delayed item is impossible. So
1816	 * we needn't lock them. And we have held i_mutex of the
1817	 * directory, nobody can delete any directory indexes now.
1818	 */
1819	list_for_each_entry_safe(curr, next, ins_list, readdir_list) {
1820		bool over;
1821
1822		list_del(&curr->readdir_list);
1823
1824		if (curr->index < ctx->pos) {
1825			if (refcount_dec_and_test(&curr->refs))
1826				kfree(curr);
1827			continue;
1828		}
1829
1830		ctx->pos = curr->index;
1831
1832		di = (struct btrfs_dir_item *)curr->data;
1833		name = (char *)(di + 1);
1834		name_len = btrfs_stack_dir_name_len(di);
1835
1836		d_type = fs_ftype_to_dtype(btrfs_dir_flags_to_ftype(di->type));
1837		btrfs_disk_key_to_cpu(&location, &di->location);
1838
1839		over = !dir_emit(ctx, name, name_len, location.objectid, d_type);
1840
1841		if (refcount_dec_and_test(&curr->refs))
1842			kfree(curr);
1843
1844		if (over)
1845			return true;
1846		ctx->pos++;
1847	}
1848	return false;
1849}
1850
1851static void fill_stack_inode_item(struct btrfs_trans_handle *trans,
1852				  struct btrfs_inode_item *inode_item,
1853				  struct btrfs_inode *inode)
1854{
1855	struct inode *vfs_inode = &inode->vfs_inode;
1856	u64 flags;
1857
1858	btrfs_set_stack_inode_uid(inode_item, i_uid_read(vfs_inode));
1859	btrfs_set_stack_inode_gid(inode_item, i_gid_read(vfs_inode));
1860	btrfs_set_stack_inode_size(inode_item, inode->disk_i_size);
1861	btrfs_set_stack_inode_mode(inode_item, vfs_inode->i_mode);
1862	btrfs_set_stack_inode_nlink(inode_item, vfs_inode->i_nlink);
1863	btrfs_set_stack_inode_nbytes(inode_item, inode_get_bytes(vfs_inode));
1864	btrfs_set_stack_inode_generation(inode_item, inode->generation);
1865	btrfs_set_stack_inode_sequence(inode_item,
1866				       inode_peek_iversion(vfs_inode));
1867	btrfs_set_stack_inode_transid(inode_item, trans->transid);
1868	btrfs_set_stack_inode_rdev(inode_item, vfs_inode->i_rdev);
1869	flags = btrfs_inode_combine_flags(inode->flags, inode->ro_flags);
1870	btrfs_set_stack_inode_flags(inode_item, flags);
1871	btrfs_set_stack_inode_block_group(inode_item, 0);
1872
1873	btrfs_set_stack_timespec_sec(&inode_item->atime,
1874				     inode_get_atime_sec(vfs_inode));
1875	btrfs_set_stack_timespec_nsec(&inode_item->atime,
1876				      inode_get_atime_nsec(vfs_inode));
1877
1878	btrfs_set_stack_timespec_sec(&inode_item->mtime,
1879				     inode_get_mtime_sec(vfs_inode));
1880	btrfs_set_stack_timespec_nsec(&inode_item->mtime,
1881				      inode_get_mtime_nsec(vfs_inode));
1882
1883	btrfs_set_stack_timespec_sec(&inode_item->ctime,
1884				     inode_get_ctime_sec(vfs_inode));
1885	btrfs_set_stack_timespec_nsec(&inode_item->ctime,
1886				      inode_get_ctime_nsec(vfs_inode));
1887
1888	btrfs_set_stack_timespec_sec(&inode_item->otime, inode->i_otime_sec);
1889	btrfs_set_stack_timespec_nsec(&inode_item->otime, inode->i_otime_nsec);
1890}
1891
1892int btrfs_fill_inode(struct btrfs_inode *inode, u32 *rdev)
1893{
1894	struct btrfs_delayed_node *delayed_node;
1895	struct btrfs_ref_tracker delayed_node_tracker;
1896	struct btrfs_inode_item *inode_item;
1897	struct inode *vfs_inode = &inode->vfs_inode;
1898
1899	delayed_node = btrfs_get_delayed_node(inode, &delayed_node_tracker);
1900	if (!delayed_node)
1901		return -ENOENT;
1902
1903	mutex_lock(&delayed_node->mutex);
1904	if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1905		mutex_unlock(&delayed_node->mutex);
1906		btrfs_release_delayed_node(delayed_node, &delayed_node_tracker);
1907		return -ENOENT;
1908	}
1909
1910	inode_item = &delayed_node->inode_item;
1911
1912	i_uid_write(vfs_inode, btrfs_stack_inode_uid(inode_item));
1913	i_gid_write(vfs_inode, btrfs_stack_inode_gid(inode_item));
1914	btrfs_i_size_write(inode, btrfs_stack_inode_size(inode_item));
1915	vfs_inode->i_mode = btrfs_stack_inode_mode(inode_item);
1916	set_nlink(vfs_inode, btrfs_stack_inode_nlink(inode_item));
1917	inode_set_bytes(vfs_inode, btrfs_stack_inode_nbytes(inode_item));
1918	inode->generation = btrfs_stack_inode_generation(inode_item);
1919	inode->last_trans = btrfs_stack_inode_transid(inode_item);
1920
1921	inode_set_iversion_queried(vfs_inode, btrfs_stack_inode_sequence(inode_item));
1922	vfs_inode->i_rdev = 0;
1923	*rdev = btrfs_stack_inode_rdev(inode_item);
1924	btrfs_inode_split_flags(btrfs_stack_inode_flags(inode_item),
1925				&inode->flags, &inode->ro_flags);
1926
1927	inode_set_atime(vfs_inode, btrfs_stack_timespec_sec(&inode_item->atime),
1928			btrfs_stack_timespec_nsec(&inode_item->atime));
1929
1930	inode_set_mtime(vfs_inode, btrfs_stack_timespec_sec(&inode_item->mtime),
1931			btrfs_stack_timespec_nsec(&inode_item->mtime));
1932
1933	inode_set_ctime(vfs_inode, btrfs_stack_timespec_sec(&inode_item->ctime),
1934			btrfs_stack_timespec_nsec(&inode_item->ctime));
1935
1936	inode->i_otime_sec = btrfs_stack_timespec_sec(&inode_item->otime);
1937	inode->i_otime_nsec = btrfs_stack_timespec_nsec(&inode_item->otime);
1938
1939	vfs_inode->i_generation = inode->generation;
1940	if (S_ISDIR(vfs_inode->i_mode))
1941		inode->index_cnt = (u64)-1;
1942
1943	mutex_unlock(&delayed_node->mutex);
1944	btrfs_release_delayed_node(delayed_node, &delayed_node_tracker);
1945	return 0;
1946}
1947
1948int btrfs_delayed_update_inode(struct btrfs_trans_handle *trans,
1949			       struct btrfs_inode *inode)
1950{
1951	struct btrfs_root *root = inode->root;
1952	struct btrfs_delayed_node *delayed_node;
1953	struct btrfs_ref_tracker delayed_node_tracker;
1954	int ret = 0;
1955
1956	delayed_node = btrfs_get_or_create_delayed_node(inode, &delayed_node_tracker);
1957	if (IS_ERR(delayed_node))
1958		return PTR_ERR(delayed_node);
1959
1960	mutex_lock(&delayed_node->mutex);
1961	if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1962		fill_stack_inode_item(trans, &delayed_node->inode_item, inode);
1963		goto release_node;
1964	}
1965
1966	ret = btrfs_delayed_inode_reserve_metadata(trans, root, delayed_node);
1967	if (ret)
1968		goto release_node;
1969
1970	fill_stack_inode_item(trans, &delayed_node->inode_item, inode);
1971	set_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags);
1972	delayed_node->count++;
1973	atomic_inc(&root->fs_info->delayed_root->items);
1974release_node:
1975	mutex_unlock(&delayed_node->mutex);
1976	btrfs_release_delayed_node(delayed_node, &delayed_node_tracker);
1977	return ret;
1978}
1979
1980int btrfs_delayed_delete_inode_ref(struct btrfs_inode *inode)
1981{
1982	struct btrfs_fs_info *fs_info = inode->root->fs_info;
1983	struct btrfs_delayed_node *delayed_node;
1984	struct btrfs_ref_tracker delayed_node_tracker;
1985
1986	/*
1987	 * we don't do delayed inode updates during log recovery because it
1988	 * leads to enospc problems.  This means we also can't do
1989	 * delayed inode refs
1990	 */
1991	if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
1992		return -EAGAIN;
1993
1994	delayed_node = btrfs_get_or_create_delayed_node(inode, &delayed_node_tracker);
1995	if (IS_ERR(delayed_node))
1996		return PTR_ERR(delayed_node);
1997
1998	/*
1999	 * We don't reserve space for inode ref deletion is because:
2000	 * - We ONLY do async inode ref deletion for the inode who has only
2001	 *   one link(i_nlink == 1), it means there is only one inode ref.
2002	 *   And in most case, the inode ref and the inode item are in the
2003	 *   same leaf, and we will deal with them at the same time.
2004	 *   Since we are sure we will reserve the space for the inode item,
2005	 *   it is unnecessary to reserve space for inode ref deletion.
2006	 * - If the inode ref and the inode item are not in the same leaf,
2007	 *   We also needn't worry about enospc problem, because we reserve
2008	 *   much more space for the inode update than it needs.
2009	 * - At the worst, we can steal some space from the global reservation.
2010	 *   It is very rare.
2011	 */
2012	mutex_lock(&delayed_node->mutex);
2013	if (!test_and_set_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags)) {
2014		delayed_node->count++;
2015		atomic_inc(&fs_info->delayed_root->items);
2016	}
2017	mutex_unlock(&delayed_node->mutex);
2018	btrfs_release_delayed_node(delayed_node, &delayed_node_tracker);
2019	return 0;
2020}
2021
2022static void __btrfs_kill_delayed_node(struct btrfs_delayed_node *delayed_node)
2023{
2024	struct btrfs_root *root = delayed_node->root;
2025	struct btrfs_fs_info *fs_info = root->fs_info;
2026	struct btrfs_delayed_item *curr_item, *prev_item;
2027
2028	mutex_lock(&delayed_node->mutex);
2029	curr_item = __btrfs_first_delayed_insertion_item(delayed_node);
2030	while (curr_item) {
2031		prev_item = curr_item;
2032		curr_item = __btrfs_next_delayed_item(prev_item);
2033		btrfs_release_delayed_item(prev_item);
2034	}
2035
2036	if (delayed_node->index_item_leaves > 0) {
2037		btrfs_delayed_item_release_leaves(delayed_node,
2038					  delayed_node->index_item_leaves);
2039		delayed_node->index_item_leaves = 0;
2040	}
2041
2042	curr_item = __btrfs_first_delayed_deletion_item(delayed_node);
2043	while (curr_item) {
2044		btrfs_delayed_item_release_metadata(root, curr_item);
2045		prev_item = curr_item;
2046		curr_item = __btrfs_next_delayed_item(prev_item);
2047		btrfs_release_delayed_item(prev_item);
2048	}
2049
2050	btrfs_release_delayed_iref(delayed_node);
2051
2052	if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
2053		btrfs_delayed_inode_release_metadata(fs_info, delayed_node, false);
2054		btrfs_release_delayed_inode(delayed_node);
2055	}
2056	mutex_unlock(&delayed_node->mutex);
2057}
2058
2059void btrfs_kill_delayed_inode_items(struct btrfs_inode *inode)
2060{
2061	struct btrfs_delayed_node *delayed_node;
2062	struct btrfs_ref_tracker delayed_node_tracker;
2063
2064	delayed_node = btrfs_get_delayed_node(inode, &delayed_node_tracker);
2065	if (!delayed_node)
2066		return;
2067
2068	__btrfs_kill_delayed_node(delayed_node);
2069	btrfs_release_delayed_node(delayed_node, &delayed_node_tracker);
2070}
2071
2072void btrfs_kill_all_delayed_nodes(struct btrfs_root *root)
2073{
2074	unsigned long index = 0;
2075	struct btrfs_delayed_node *delayed_nodes[8];
2076	struct btrfs_ref_tracker delayed_node_trackers[8];
2077
2078	while (1) {
2079		struct btrfs_delayed_node *node;
2080		int count;
2081
2082		xa_lock(&root->delayed_nodes);
2083		if (xa_empty(&root->delayed_nodes)) {
2084			xa_unlock(&root->delayed_nodes);
2085			return;
2086		}
2087
2088		count = 0;
2089		xa_for_each_start(&root->delayed_nodes, index, node, index) {
2090			/*
2091			 * Don't increase refs in case the node is dead and
2092			 * about to be removed from the tree in the loop below
2093			 */
2094			if (refcount_inc_not_zero(&node->refs)) {
2095				btrfs_delayed_node_ref_tracker_alloc(node,
2096						     &delayed_node_trackers[count],
2097						     GFP_ATOMIC);
2098				delayed_nodes[count] = node;
2099				count++;
2100			}
2101			if (count >= ARRAY_SIZE(delayed_nodes))
2102				break;
2103		}
2104		xa_unlock(&root->delayed_nodes);
2105		index++;
2106
2107		for (int i = 0; i < count; i++) {
2108			__btrfs_kill_delayed_node(delayed_nodes[i]);
2109			btrfs_delayed_node_ref_tracker_dir_print(delayed_nodes[i]);
2110			btrfs_release_delayed_node(delayed_nodes[i],
2111						   &delayed_node_trackers[i]);
2112		}
2113	}
2114}
2115
2116void btrfs_destroy_delayed_inodes(struct btrfs_fs_info *fs_info)
2117{
2118	struct btrfs_delayed_node *curr_node, *prev_node;
2119	struct btrfs_ref_tracker curr_delayed_node_tracker, prev_delayed_node_tracker;
2120
2121	curr_node = btrfs_first_delayed_node(fs_info->delayed_root,
2122					     &curr_delayed_node_tracker);
2123	while (curr_node) {
2124		__btrfs_kill_delayed_node(curr_node);
2125
2126		prev_node = curr_node;
2127		prev_delayed_node_tracker = curr_delayed_node_tracker;
2128		curr_node = btrfs_next_delayed_node(curr_node, &curr_delayed_node_tracker);
2129		btrfs_release_delayed_node(prev_node, &prev_delayed_node_tracker);
2130	}
2131}
2132
2133void btrfs_log_get_delayed_items(struct btrfs_inode *inode,
2134				 struct list_head *ins_list,
2135				 struct list_head *del_list)
2136{
2137	struct btrfs_delayed_node *node;
2138	struct btrfs_delayed_item *item;
2139	struct btrfs_ref_tracker delayed_node_tracker;
2140
2141	node = btrfs_get_delayed_node(inode, &delayed_node_tracker);
2142	if (!node)
2143		return;
2144
2145	mutex_lock(&node->mutex);
2146	item = __btrfs_first_delayed_insertion_item(node);
2147	while (item) {
2148		/*
2149		 * It's possible that the item is already in a log list. This
2150		 * can happen in case two tasks are trying to log the same
2151		 * directory. For example if we have tasks A and task B:
2152		 *
2153		 * Task A collected the delayed items into a log list while
2154		 * under the inode's log_mutex (at btrfs_log_inode()), but it
2155		 * only releases the items after logging the inodes they point
2156		 * to (if they are new inodes), which happens after unlocking
2157		 * the log mutex;
2158		 *
2159		 * Task B enters btrfs_log_inode() and acquires the log_mutex
2160		 * of the same directory inode, before task B releases the
2161		 * delayed items. This can happen for example when logging some
2162		 * inode we need to trigger logging of its parent directory, so
2163		 * logging two files that have the same parent directory can
2164		 * lead to this.
2165		 *
2166		 * If this happens, just ignore delayed items already in a log
2167		 * list. All the tasks logging the directory are under a log
2168		 * transaction and whichever finishes first can not sync the log
2169		 * before the other completes and leaves the log transaction.
2170		 */
2171		if (!item->logged && list_empty(&item->log_list)) {
2172			refcount_inc(&item->refs);
2173			list_add_tail(&item->log_list, ins_list);
2174		}
2175		item = __btrfs_next_delayed_item(item);
2176	}
2177
2178	item = __btrfs_first_delayed_deletion_item(node);
2179	while (item) {
2180		/* It may be non-empty, for the same reason mentioned above. */
2181		if (!item->logged && list_empty(&item->log_list)) {
2182			refcount_inc(&item->refs);
2183			list_add_tail(&item->log_list, del_list);
2184		}
2185		item = __btrfs_next_delayed_item(item);
2186	}
2187	mutex_unlock(&node->mutex);
2188
2189	/*
2190	 * We are called during inode logging, which means the inode is in use
2191	 * and can not be evicted before we finish logging the inode. So we never
2192	 * have the last reference on the delayed inode.
2193	 * Also, we don't use btrfs_release_delayed_node() because that would
2194	 * requeue the delayed inode (change its order in the list of prepared
2195	 * nodes) and we don't want to do such change because we don't create or
2196	 * delete delayed items.
2197	 */
2198	ASSERT(refcount_read(&node->refs) > 1);
2199	btrfs_delayed_node_ref_tracker_free(node, &delayed_node_tracker);
2200	refcount_dec(&node->refs);
2201}
2202
2203void btrfs_log_put_delayed_items(struct btrfs_inode *inode,
2204				 struct list_head *ins_list,
2205				 struct list_head *del_list)
2206{
2207	struct btrfs_delayed_node *node;
2208	struct btrfs_delayed_item *item;
2209	struct btrfs_delayed_item *next;
2210	struct btrfs_ref_tracker delayed_node_tracker;
2211
2212	node = btrfs_get_delayed_node(inode, &delayed_node_tracker);
2213	if (!node)
2214		return;
2215
2216	mutex_lock(&node->mutex);
2217
2218	list_for_each_entry_safe(item, next, ins_list, log_list) {
2219		item->logged = true;
2220		list_del_init(&item->log_list);
2221		if (refcount_dec_and_test(&item->refs))
2222			kfree(item);
2223	}
2224
2225	list_for_each_entry_safe(item, next, del_list, log_list) {
2226		item->logged = true;
2227		list_del_init(&item->log_list);
2228		if (refcount_dec_and_test(&item->refs))
2229			kfree(item);
2230	}
2231
2232	mutex_unlock(&node->mutex);
2233
2234	/*
2235	 * We are called during inode logging, which means the inode is in use
2236	 * and can not be evicted before we finish logging the inode. So we never
2237	 * have the last reference on the delayed inode.
2238	 * Also, we don't use btrfs_release_delayed_node() because that would
2239	 * requeue the delayed inode (change its order in the list of prepared
2240	 * nodes) and we don't want to do such change because we don't create or
2241	 * delete delayed items.
2242	 */
2243	ASSERT(refcount_read(&node->refs) > 1);
2244	btrfs_delayed_node_ref_tracker_free(node, &delayed_node_tracker);
2245	refcount_dec(&node->refs);
2246}