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