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 "misc.h"
  10 #include "delayed-inode.h"
  11 #include "disk-io.h"
  12 #include "transaction.h"
  13 #include "ctree.h"
  14 #include "qgroup.h"
  15 #include "locking.h"
  16
  17 #define BTRFS_DELAYED_WRITEBACK         512
  18 #define BTRFS_DELAYED_BACKGROUND        128
  19 #define BTRFS_DELAYED_BATCH             16
  20
  21 static struct kmem_cache *delayed_node_cache;
  22
  23 int __init btrfs_delayed_inode_init(void)
  24 {
  25         delayed_node_cache = kmem_cache_create("btrfs_delayed_node",
  26                                         sizeof(struct btrfs_delayed_node),
  27                                         0,
  28                                         SLAB_MEM_SPREAD,
  29                                         NULL);
  30         if (!delayed_node_cache)
  31                 return -ENOMEM;
  32         return 0;
  33 }
  34
  35 void __cold btrfs_delayed_inode_exit(void)
  36 {
  37         kmem_cache_destroy(delayed_node_cache);
  38 }
  39
  40 static inline void btrfs_init_delayed_node(
  41                                 struct btrfs_delayed_node *delayed_node,
  42                                 struct btrfs_root *root, u64 inode_id)
  43 {
  44         delayed_node->root = root;
  45         delayed_node->inode_id = inode_id;
  46         refcount_set(&delayed_node->refs, 0);
  47         delayed_node->ins_root = RB_ROOT_CACHED;
  48         delayed_node->del_root = RB_ROOT_CACHED;
  49         mutex_init(&delayed_node->mutex);
  50         INIT_LIST_HEAD(&delayed_node->n_list);
  51         INIT_LIST_HEAD(&delayed_node->p_list);
  52 }
  53
  54 static inline int btrfs_is_continuous_delayed_item(
  55                                         struct btrfs_delayed_item *item1,
  56                                         struct btrfs_delayed_item *item2)
  57 {
  58         if (item1->key.type == BTRFS_DIR_INDEX_KEY &&
  59             item1->key.objectid == item2->key.objectid &&
  60             item1->key.type == item2->key.type &&
  61             item1->key.offset + 1 == item2->key.offset)
  62                 return 1;
  63         return 0;
  64 }
  65
  66 static struct btrfs_delayed_node *btrfs_get_delayed_node(
  67                 struct btrfs_inode *btrfs_inode)
  68 {
  69         struct btrfs_root *root = btrfs_inode->root;
  70         u64 ino = btrfs_ino(btrfs_inode);
  71         struct btrfs_delayed_node *node;
  72
  73         node = READ_ONCE(btrfs_inode->delayed_node);
  74         if (node) {
  75                 refcount_inc(&node->refs);
  76                 return node;
  77         }
  78
  79         spin_lock(&root->inode_lock);
  80         node = radix_tree_lookup(&root->delayed_nodes_tree, ino);
  81
  82         if (node) {
  83                 if (btrfs_inode->delayed_node) {
  84                         refcount_inc(&node->refs);      /* can be accessed */
  85                         BUG_ON(btrfs_inode->delayed_node != node);
  86                         spin_unlock(&root->inode_lock);
  87                         return node;
  88                 }
  89
  90                 /*
  91                  * It's possible that we're racing into the middle of removing
  92                  * this node from the radix tree.  In this case, the refcount
  93                  * was zero and it should never go back to one.  Just return
  94                  * NULL like it was never in the radix at all; our release
  95                  * function is in the process of removing it.
  96                  *
  97                  * Some implementations of refcount_inc refuse to bump the
  98                  * refcount once it has hit zero.  If we don't do this dance
  99                  * here, refcount_inc() may decide to just WARN_ONCE() instead
 100                  * of actually bumping the refcount.
 101                  *
 102                  * If this node is properly in the radix, we want to bump the
 103                  * refcount twice, once for the inode and once for this get
 104                  * operation.
 105                  */
 106                 if (refcount_inc_not_zero(&node->refs)) {
 107                         refcount_inc(&node->refs);
 108                         btrfs_inode->delayed_node = node;
 109                 } else {
 110                         node = NULL;
 111                 }
 112
 113                 spin_unlock(&root->inode_lock);
 114                 return node;
 115         }
 116         spin_unlock(&root->inode_lock);
 117
 118         return NULL;
 119 }
 120
 121 /* Will return either the node or PTR_ERR(-ENOMEM) */
 122 static struct btrfs_delayed_node *btrfs_get_or_create_delayed_node(
 123                 struct btrfs_inode *btrfs_inode)
 124 {
 125         struct btrfs_delayed_node *node;
 126         struct btrfs_root *root = btrfs_inode->root;
 127         u64 ino = btrfs_ino(btrfs_inode);
 128         int ret;
 129
 130 again:
 131         node = btrfs_get_delayed_node(btrfs_inode);
 132         if (node)
 133                 return node;
 134
 135         node = kmem_cache_zalloc(delayed_node_cache, GFP_NOFS);
 136         if (!node)
 137                 return ERR_PTR(-ENOMEM);
 138         btrfs_init_delayed_node(node, root, ino);
 139
 140         /* cached in the btrfs inode and can be accessed */
 141         refcount_set(&node->refs, 2);
 142
 143         ret = radix_tree_preload(GFP_NOFS);
 144         if (ret) {
 145                 kmem_cache_free(delayed_node_cache, node);
 146                 return ERR_PTR(ret);
 147         }
 148
 149         spin_lock(&root->inode_lock);
 150         ret = radix_tree_insert(&root->delayed_nodes_tree, ino, node);
 151         if (ret == -EEXIST) {
 152                 spin_unlock(&root->inode_lock);
 153                 kmem_cache_free(delayed_node_cache, node);
 154                 radix_tree_preload_end();
 155                 goto again;
 156         }
 157         btrfs_inode->delayed_node = node;
 158         spin_unlock(&root->inode_lock);
 159         radix_tree_preload_end();
 160
 161         return node;
 162 }
 163
 164 /*
 165  * Call it when holding delayed_node->mutex
 166  *
 167  * If mod = 1, add this node into the prepared list.
 168  */
 169 static void btrfs_queue_delayed_node(struct btrfs_delayed_root *root,
 170                                      struct btrfs_delayed_node *node,
 171                                      int mod)
 172 {
 173         spin_lock(&root->lock);
 174         if (test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
 175                 if (!list_empty(&node->p_list))
 176                         list_move_tail(&node->p_list, &root->prepare_list);
 177                 else if (mod)
 178                         list_add_tail(&node->p_list, &root->prepare_list);
 179         } else {
 180                 list_add_tail(&node->n_list, &root->node_list);
 181                 list_add_tail(&node->p_list, &root->prepare_list);
 182                 refcount_inc(&node->refs);      /* inserted into list */
 183                 root->nodes++;
 184                 set_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags);
 185         }
 186         spin_unlock(&root->lock);
 187 }
 188
 189 /* Call it when holding delayed_node->mutex */
 190 static void btrfs_dequeue_delayed_node(struct btrfs_delayed_root *root,
 191                                        struct btrfs_delayed_node *node)
 192 {
 193         spin_lock(&root->lock);
 194         if (test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
 195                 root->nodes--;
 196                 refcount_dec(&node->refs);      /* not in the list */
 197                 list_del_init(&node->n_list);
 198                 if (!list_empty(&node->p_list))
 199                         list_del_init(&node->p_list);
 200                 clear_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags);
 201         }
 202         spin_unlock(&root->lock);
 203 }
 204
 205 static struct btrfs_delayed_node *btrfs_first_delayed_node(
 206                         struct btrfs_delayed_root *delayed_root)
 207 {
 208         struct list_head *p;
 209         struct btrfs_delayed_node *node = NULL;
 210
 211         spin_lock(&delayed_root->lock);
 212         if (list_empty(&delayed_root->node_list))
 213                 goto out;
 214
 215         p = delayed_root->node_list.next;
 216         node = list_entry(p, struct btrfs_delayed_node, n_list);
 217         refcount_inc(&node->refs);
 218 out:
 219         spin_unlock(&delayed_root->lock);
 220
 221         return node;
 222 }
 223
 224 static struct btrfs_delayed_node *btrfs_next_delayed_node(
 225                                                 struct btrfs_delayed_node *node)
 226 {
 227         struct btrfs_delayed_root *delayed_root;
 228         struct list_head *p;
 229         struct btrfs_delayed_node *next = NULL;
 230
 231         delayed_root = node->root->fs_info->delayed_root;
 232         spin_lock(&delayed_root->lock);
 233         if (!test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
 234                 /* not in the list */
 235                 if (list_empty(&delayed_root->node_list))
 236                         goto out;
 237                 p = delayed_root->node_list.next;
 238         } else if (list_is_last(&node->n_list, &delayed_root->node_list))
 239                 goto out;
 240         else
 241                 p = node->n_list.next;
 242
 243         next = list_entry(p, struct btrfs_delayed_node, n_list);
 244         refcount_inc(&next->refs);
 245 out:
 246         spin_unlock(&delayed_root->lock);
 247
 248         return next;
 249 }
 250
 251 static void __btrfs_release_delayed_node(
 252                                 struct btrfs_delayed_node *delayed_node,
 253                                 int mod)
 254 {
 255         struct btrfs_delayed_root *delayed_root;
 256
 257         if (!delayed_node)
 258                 return;
 259
 260         delayed_root = delayed_node->root->fs_info->delayed_root;
 261
 262         mutex_lock(&delayed_node->mutex);
 263         if (delayed_node->count)
 264                 btrfs_queue_delayed_node(delayed_root, delayed_node, mod);
 265         else
 266                 btrfs_dequeue_delayed_node(delayed_root, delayed_node);
 267         mutex_unlock(&delayed_node->mutex);
 268
 269         if (refcount_dec_and_test(&delayed_node->refs)) {
 270                 struct btrfs_root *root = delayed_node->root;
 271
 272                 spin_lock(&root->inode_lock);
 273                 /*
 274                  * Once our refcount goes to zero, nobody is allowed to bump it
 275                  * back up.  We can delete it now.
 276                  */
 277                 ASSERT(refcount_read(&delayed_node->refs) == 0);
 278                 radix_tree_delete(&root->delayed_nodes_tree,
 279                                   delayed_node->inode_id);
 280                 spin_unlock(&root->inode_lock);
 281                 kmem_cache_free(delayed_node_cache, delayed_node);
 282         }
 283 }
 284
 285 static inline void btrfs_release_delayed_node(struct btrfs_delayed_node *node)
 286 {
 287         __btrfs_release_delayed_node(node, 0);
 288 }
 289
 290 static struct btrfs_delayed_node *btrfs_first_prepared_delayed_node(
 291                                         struct btrfs_delayed_root *delayed_root)
 292 {
 293         struct list_head *p;
 294         struct btrfs_delayed_node *node = NULL;
 295
 296         spin_lock(&delayed_root->lock);
 297         if (list_empty(&delayed_root->prepare_list))
 298                 goto out;
 299
 300         p = delayed_root->prepare_list.next;
 301         list_del_init(p);
 302         node = list_entry(p, struct btrfs_delayed_node, p_list);
 303         refcount_inc(&node->refs);
 304 out:
 305         spin_unlock(&delayed_root->lock);
 306
 307         return node;
 308 }
 309
 310 static inline void btrfs_release_prepared_delayed_node(
 311                                         struct btrfs_delayed_node *node)
 312 {
 313         __btrfs_release_delayed_node(node, 1);
 314 }
 315
 316 static struct btrfs_delayed_item *btrfs_alloc_delayed_item(u32 data_len)
 317 {
 318         struct btrfs_delayed_item *item;
 319         item = kmalloc(sizeof(*item) + data_len, GFP_NOFS);
 320         if (item) {
 321                 item->data_len = data_len;
 322                 item->ins_or_del = 0;
 323                 item->bytes_reserved = 0;
 324                 item->delayed_node = NULL;
 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  * @key:          the key to look up
 334  * @prev:         used to store the prev item if the right item isn't found
 335  * @next:         used to store the next item if the right item isn't found
 336  *
 337  * Note: if we don't find the right item, we will return the prev item and
 338  * the next item.
 339  */
 340 static struct btrfs_delayed_item *__btrfs_lookup_delayed_item(
 341                                 struct rb_root *root,
 342                                 struct btrfs_key *key,
 343                                 struct btrfs_delayed_item **prev,
 344                                 struct btrfs_delayed_item **next)
 345 {
 346         struct rb_node *node, *prev_node = NULL;
 347         struct btrfs_delayed_item *delayed_item = NULL;
 348         int ret = 0;
 349
 350         node = root->rb_node;
 351
 352         while (node) {
 353                 delayed_item = rb_entry(node, struct btrfs_delayed_item,
 354                                         rb_node);
 355                 prev_node = node;
 356                 ret = btrfs_comp_cpu_keys(&delayed_item->key, key);
 357                 if (ret < 0)
 358                         node = node->rb_right;
 359                 else if (ret > 0)
 360                         node = node->rb_left;
 361                 else
 362                         return delayed_item;
 363         }
 364
 365         if (prev) {
 366                 if (!prev_node)
 367                         *prev = NULL;
 368                 else if (ret < 0)
 369                         *prev = delayed_item;
 370                 else if ((node = rb_prev(prev_node)) != NULL) {
 371                         *prev = rb_entry(node, struct btrfs_delayed_item,
 372                                          rb_node);
 373                 } else
 374                         *prev = NULL;
 375         }
 376
 377         if (next) {
 378                 if (!prev_node)
 379                         *next = NULL;
 380                 else if (ret > 0)
 381                         *next = delayed_item;
 382                 else if ((node = rb_next(prev_node)) != NULL) {
 383                         *next = rb_entry(node, struct btrfs_delayed_item,
 384                                          rb_node);
 385                 } else
 386                         *next = NULL;
 387         }
 388         return NULL;
 389 }
 390
 391 static struct btrfs_delayed_item *__btrfs_lookup_delayed_insertion_item(
 392                                         struct btrfs_delayed_node *delayed_node,
 393                                         struct btrfs_key *key)
 394 {
 395         return __btrfs_lookup_delayed_item(&delayed_node->ins_root.rb_root, key,
 396                                            NULL, NULL);
 397 }
 398
 399 static int __btrfs_add_delayed_item(struct btrfs_delayed_node *delayed_node,
 400                                     struct btrfs_delayed_item *ins,
 401                                     int action)
 402 {
 403         struct rb_node **p, *node;
 404         struct rb_node *parent_node = NULL;
 405         struct rb_root_cached *root;
 406         struct btrfs_delayed_item *item;
 407         int cmp;
 408         bool leftmost = true;
 409
 410         if (action == BTRFS_DELAYED_INSERTION_ITEM)
 411                 root = &delayed_node->ins_root;
 412         else if (action == BTRFS_DELAYED_DELETION_ITEM)
 413                 root = &delayed_node->del_root;
 414         else
 415                 BUG();
 416         p = &root->rb_root.rb_node;
 417         node = &ins->rb_node;
 418
 419         while (*p) {
 420                 parent_node = *p;
 421                 item = rb_entry(parent_node, struct btrfs_delayed_item,
 422                                  rb_node);
 423
 424                 cmp = btrfs_comp_cpu_keys(&item->key, &ins->key);
 425                 if (cmp < 0) {
 426                         p = &(*p)->rb_right;
 427                         leftmost = false;
 428                 } else if (cmp > 0) {
 429                         p = &(*p)->rb_left;
 430                 } else {
 431                         return -EEXIST;
 432                 }
 433         }
 434
 435         rb_link_node(node, parent_node, p);
 436         rb_insert_color_cached(node, root, leftmost);
 437         ins->delayed_node = delayed_node;
 438         ins->ins_or_del = action;
 439
 440         if (ins->key.type == BTRFS_DIR_INDEX_KEY &&
 441             action == BTRFS_DELAYED_INSERTION_ITEM &&
 442             ins->key.offset >= delayed_node->index_cnt)
 443                         delayed_node->index_cnt = ins->key.offset + 1;
 444
 445         delayed_node->count++;
 446         atomic_inc(&delayed_node->root->fs_info->delayed_root->items);
 447         return 0;
 448 }
 449
 450 static int __btrfs_add_delayed_insertion_item(struct btrfs_delayed_node *node,
 451                                               struct btrfs_delayed_item *item)
 452 {
 453         return __btrfs_add_delayed_item(node, item,
 454                                         BTRFS_DELAYED_INSERTION_ITEM);
 455 }
 456
 457 static int __btrfs_add_delayed_deletion_item(struct btrfs_delayed_node *node,
 458                                              struct btrfs_delayed_item *item)
 459 {
 460         return __btrfs_add_delayed_item(node, item,
 461                                         BTRFS_DELAYED_DELETION_ITEM);
 462 }
 463
 464 static void finish_one_item(struct btrfs_delayed_root *delayed_root)
 465 {
 466         int seq = atomic_inc_return(&delayed_root->items_seq);
 467
 468         /* atomic_dec_return implies a barrier */
 469         if ((atomic_dec_return(&delayed_root->items) <
 470             BTRFS_DELAYED_BACKGROUND || seq % BTRFS_DELAYED_BATCH == 0))
 471                 cond_wake_up_nomb(&delayed_root->wait);
 472 }
 473
 474 static void __btrfs_remove_delayed_item(struct btrfs_delayed_item *delayed_item)
 475 {
 476         struct rb_root_cached *root;
 477         struct btrfs_delayed_root *delayed_root;
 478
 479         /* Not associated with any delayed_node */
 480         if (!delayed_item->delayed_node)
 481                 return;
 482         delayed_root = delayed_item->delayed_node->root->fs_info->delayed_root;
 483
 484         BUG_ON(!delayed_root);
 485         BUG_ON(delayed_item->ins_or_del != BTRFS_DELAYED_DELETION_ITEM &&
 486                delayed_item->ins_or_del != BTRFS_DELAYED_INSERTION_ITEM);
 487
 488         if (delayed_item->ins_or_del == BTRFS_DELAYED_INSERTION_ITEM)
 489                 root = &delayed_item->delayed_node->ins_root;
 490         else
 491                 root = &delayed_item->delayed_node->del_root;
 492
 493         rb_erase_cached(&delayed_item->rb_node, root);
 494         delayed_item->delayed_node->count--;
 495
 496         finish_one_item(delayed_root);
 497 }
 498
 499 static void btrfs_release_delayed_item(struct btrfs_delayed_item *item)
 500 {
 501         if (item) {
 502                 __btrfs_remove_delayed_item(item);
 503                 if (refcount_dec_and_test(&item->refs))
 504                         kfree(item);
 505         }
 506 }
 507
 508 static struct btrfs_delayed_item *__btrfs_first_delayed_insertion_item(
 509                                         struct btrfs_delayed_node *delayed_node)
 510 {
 511         struct rb_node *p;
 512         struct btrfs_delayed_item *item = NULL;
 513
 514         p = rb_first_cached(&delayed_node->ins_root);
 515         if (p)
 516                 item = rb_entry(p, struct btrfs_delayed_item, rb_node);
 517
 518         return item;
 519 }
 520
 521 static struct btrfs_delayed_item *__btrfs_first_delayed_deletion_item(
 522                                         struct btrfs_delayed_node *delayed_node)
 523 {
 524         struct rb_node *p;
 525         struct btrfs_delayed_item *item = NULL;
 526
 527         p = rb_first_cached(&delayed_node->del_root);
 528         if (p)
 529                 item = rb_entry(p, struct btrfs_delayed_item, rb_node);
 530
 531         return item;
 532 }
 533
 534 static struct btrfs_delayed_item *__btrfs_next_delayed_item(
 535                                                 struct btrfs_delayed_item *item)
 536 {
 537         struct rb_node *p;
 538         struct btrfs_delayed_item *next = NULL;
 539
 540         p = rb_next(&item->rb_node);
 541         if (p)
 542                 next = rb_entry(p, struct btrfs_delayed_item, rb_node);
 543
 544         return next;
 545 }
 546
 547 static int btrfs_delayed_item_reserve_metadata(struct btrfs_trans_handle *trans,
 548                                                struct btrfs_root *root,
 549                                                struct btrfs_delayed_item *item)
 550 {
 551         struct btrfs_block_rsv *src_rsv;
 552         struct btrfs_block_rsv *dst_rsv;
 553         struct btrfs_fs_info *fs_info = root->fs_info;
 554         u64 num_bytes;
 555         int ret;
 556
 557         if (!trans->bytes_reserved)
 558                 return 0;
 559
 560         src_rsv = trans->block_rsv;
 561         dst_rsv = &fs_info->delayed_block_rsv;
 562
 563         num_bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
 564
 565         /*
 566          * Here we migrate space rsv from transaction rsv, since have already
 567          * reserved space when starting a transaction.  So no need to reserve
 568          * qgroup space here.
 569          */
 570         ret = btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, true);
 571         if (!ret) {
 572                 trace_btrfs_space_reservation(fs_info, "delayed_item",
 573                                               item->key.objectid,
 574                                               num_bytes, 1);
 575                 item->bytes_reserved = num_bytes;
 576         }
 577
 578         return ret;
 579 }
 580
 581 static void btrfs_delayed_item_release_metadata(struct btrfs_root *root,
 582                                                 struct btrfs_delayed_item *item)
 583 {
 584         struct btrfs_block_rsv *rsv;
 585         struct btrfs_fs_info *fs_info = root->fs_info;
 586
 587         if (!item->bytes_reserved)
 588                 return;
 589
 590         rsv = &fs_info->delayed_block_rsv;
 591         /*
 592          * Check btrfs_delayed_item_reserve_metadata() to see why we don't need
 593          * to release/reserve qgroup space.
 594          */
 595         trace_btrfs_space_reservation(fs_info, "delayed_item",
 596                                       item->key.objectid, item->bytes_reserved,
 597                                       0);
 598         btrfs_block_rsv_release(fs_info, rsv, item->bytes_reserved, NULL);
 599 }
 600
 601 static int btrfs_delayed_inode_reserve_metadata(
 602                                         struct btrfs_trans_handle *trans,
 603                                         struct btrfs_root *root,
 604                                         struct btrfs_delayed_node *node)
 605 {
 606         struct btrfs_fs_info *fs_info = root->fs_info;
 607         struct btrfs_block_rsv *src_rsv;
 608         struct btrfs_block_rsv *dst_rsv;
 609         u64 num_bytes;
 610         int ret;
 611
 612         src_rsv = trans->block_rsv;
 613         dst_rsv = &fs_info->delayed_block_rsv;
 614
 615         num_bytes = btrfs_calc_metadata_size(fs_info, 1);
 616
 617         /*
 618          * btrfs_dirty_inode will update the inode under btrfs_join_transaction
 619          * which doesn't reserve space for speed.  This is a problem since we
 620          * still need to reserve space for this update, so try to reserve the
 621          * space.
 622          *
 623          * Now if src_rsv == delalloc_block_rsv we'll let it just steal since
 624          * we always reserve enough to update the inode item.
 625          */
 626         if (!src_rsv || (!trans->bytes_reserved &&
 627                          src_rsv->type != BTRFS_BLOCK_RSV_DELALLOC)) {
 628                 ret = btrfs_qgroup_reserve_meta(root, num_bytes,
 629                                           BTRFS_QGROUP_RSV_META_PREALLOC, true);
 630                 if (ret < 0)
 631                         return ret;
 632                 ret = btrfs_block_rsv_add(root, dst_rsv, num_bytes,
 633                                           BTRFS_RESERVE_NO_FLUSH);
 634                 /* NO_FLUSH could only fail with -ENOSPC */
 635                 ASSERT(ret == 0 || ret == -ENOSPC);
 636                 if (ret)
 637                         btrfs_qgroup_free_meta_prealloc(root, num_bytes);
 638         } else {
 639                 ret = btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, true);
 640         }
 641
 642         if (!ret) {
 643                 trace_btrfs_space_reservation(fs_info, "delayed_inode",
 644                                               node->inode_id, num_bytes, 1);
 645                 node->bytes_reserved = num_bytes;
 646         }
 647
 648         return ret;
 649 }
 650
 651 static void btrfs_delayed_inode_release_metadata(struct btrfs_fs_info *fs_info,
 652                                                 struct btrfs_delayed_node *node,
 653                                                 bool qgroup_free)
 654 {
 655         struct btrfs_block_rsv *rsv;
 656
 657         if (!node->bytes_reserved)
 658                 return;
 659
 660         rsv = &fs_info->delayed_block_rsv;
 661         trace_btrfs_space_reservation(fs_info, "delayed_inode",
 662                                       node->inode_id, node->bytes_reserved, 0);
 663         btrfs_block_rsv_release(fs_info, rsv, node->bytes_reserved, NULL);
 664         if (qgroup_free)
 665                 btrfs_qgroup_free_meta_prealloc(node->root,
 666                                 node->bytes_reserved);
 667         else
 668                 btrfs_qgroup_convert_reserved_meta(node->root,
 669                                 node->bytes_reserved);
 670         node->bytes_reserved = 0;
 671 }
 672
 673 /*
 674  * Insert a single delayed item or a batch of delayed items that have consecutive
 675  * keys if they exist.
 676  */
 677 static int btrfs_insert_delayed_item(struct btrfs_trans_handle *trans,
 678                                      struct btrfs_root *root,
 679                                      struct btrfs_path *path,
 680                                      struct btrfs_delayed_item *first_item)
 681 {
 682         LIST_HEAD(batch);
 683         struct btrfs_delayed_item *curr;
 684         struct btrfs_delayed_item *next;
 685         const int max_size = BTRFS_LEAF_DATA_SIZE(root->fs_info);
 686         int total_size;
 687         int nitems;
 688         char *ins_data = NULL;
 689         struct btrfs_key *ins_keys;
 690         u32 *ins_sizes;
 691         int ret;
 692
 693         list_add_tail(&first_item->tree_list, &batch);
 694         nitems = 1;
 695         total_size = first_item->data_len + sizeof(struct btrfs_item);
 696         curr = first_item;
 697
 698         while (true) {
 699                 int next_size;
 700
 701                 next = __btrfs_next_delayed_item(curr);
 702                 if (!next || !btrfs_is_continuous_delayed_item(curr, next))
 703                         break;
 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, &batch);
 710                 nitems++;
 711                 total_size += next_size;
 712                 curr = next;
 713         }
 714
 715         if (nitems == 1) {
 716                 ins_keys = &first_item->key;
 717                 ins_sizes = &first_item->data_len;
 718         } else {
 719                 int i = 0;
 720
 721                 ins_data = kmalloc(nitems * sizeof(u32) +
 722                                    nitems * sizeof(struct btrfs_key), GFP_NOFS);
 723                 if (!ins_data) {
 724                         ret = -ENOMEM;
 725                         goto out;
 726                 }
 727                 ins_sizes = (u32 *)ins_data;
 728                 ins_keys = (struct btrfs_key *)(ins_data + nitems * sizeof(u32));
 729                 list_for_each_entry(curr, &batch, tree_list) {
 730                         ins_keys[i] = curr->key;
 731                         ins_sizes[i] = curr->data_len;
 732                         i++;
 733                 }
 734         }
 735
 736         ret = btrfs_insert_empty_items(trans, root, path, ins_keys, ins_sizes,
 737                                        nitems);
 738         if (ret)
 739                 goto out;
 740
 741         list_for_each_entry(curr, &batch, tree_list) {
 742                 char *data_ptr;
 743
 744                 data_ptr = btrfs_item_ptr(path->nodes[0], path->slots[0], char);
 745                 write_extent_buffer(path->nodes[0], &curr->data,
 746                                     (unsigned long)data_ptr, curr->data_len);
 747                 path->slots[0]++;
 748         }
 749
 750         /*
 751          * Now release our path before releasing the delayed items and their
 752          * metadata reservations, so that we don't block other tasks for more
 753          * time than needed.
 754          */
 755         btrfs_release_path(path);
 756
 757         list_for_each_entry_safe(curr, next, &batch, tree_list) {
 758                 list_del(&curr->tree_list);
 759                 btrfs_delayed_item_release_metadata(root, curr);
 760                 btrfs_release_delayed_item(curr);
 761         }
 762 out:
 763         kfree(ins_data);
 764         return ret;
 765 }
 766
 767 static int btrfs_insert_delayed_items(struct btrfs_trans_handle *trans,
 768                                       struct btrfs_path *path,
 769                                       struct btrfs_root *root,
 770                                       struct btrfs_delayed_node *node)
 771 {
 772         int ret = 0;
 773
 774         while (ret == 0) {
 775                 struct btrfs_delayed_item *curr;
 776
 777                 mutex_lock(&node->mutex);
 778                 curr = __btrfs_first_delayed_insertion_item(node);
 779                 if (!curr) {
 780                         mutex_unlock(&node->mutex);
 781                         break;
 782                 }
 783                 ret = btrfs_insert_delayed_item(trans, root, path, curr);
 784                 mutex_unlock(&node->mutex);
 785         }
 786
 787         return ret;
 788 }
 789
 790 static int btrfs_batch_delete_items(struct btrfs_trans_handle *trans,
 791                                     struct btrfs_root *root,
 792                                     struct btrfs_path *path,
 793                                     struct btrfs_delayed_item *item)
 794 {
 795         struct btrfs_delayed_item *curr, *next;
 796         struct extent_buffer *leaf;
 797         struct btrfs_key key;
 798         struct list_head head;
 799         int nitems, i, last_item;
 800         int ret = 0;
 801
 802         BUG_ON(!path->nodes[0]);
 803
 804         leaf = path->nodes[0];
 805
 806         i = path->slots[0];
 807         last_item = btrfs_header_nritems(leaf) - 1;
 808         if (i > last_item)
 809                 return -ENOENT; /* FIXME: Is errno suitable? */
 810
 811         next = item;
 812         INIT_LIST_HEAD(&head);
 813         btrfs_item_key_to_cpu(leaf, &key, i);
 814         nitems = 0;
 815         /*
 816          * count the number of the dir index items that we can delete in batch
 817          */
 818         while (btrfs_comp_cpu_keys(&next->key, &key) == 0) {
 819                 list_add_tail(&next->tree_list, &head);
 820                 nitems++;
 821
 822                 curr = next;
 823                 next = __btrfs_next_delayed_item(curr);
 824                 if (!next)
 825                         break;
 826
 827                 if (!btrfs_is_continuous_delayed_item(curr, next))
 828                         break;
 829
 830                 i++;
 831                 if (i > last_item)
 832                         break;
 833                 btrfs_item_key_to_cpu(leaf, &key, i);
 834         }
 835
 836         if (!nitems)
 837                 return 0;
 838
 839         ret = btrfs_del_items(trans, root, path, path->slots[0], nitems);
 840         if (ret)
 841                 goto out;
 842
 843         list_for_each_entry_safe(curr, next, &head, tree_list) {
 844                 btrfs_delayed_item_release_metadata(root, curr);
 845                 list_del(&curr->tree_list);
 846                 btrfs_release_delayed_item(curr);
 847         }
 848
 849 out:
 850         return ret;
 851 }
 852
 853 static int btrfs_delete_delayed_items(struct btrfs_trans_handle *trans,
 854                                       struct btrfs_path *path,
 855                                       struct btrfs_root *root,
 856                                       struct btrfs_delayed_node *node)
 857 {
 858         struct btrfs_delayed_item *curr, *prev;
 859         int ret = 0;
 860
 861 do_again:
 862         mutex_lock(&node->mutex);
 863         curr = __btrfs_first_delayed_deletion_item(node);
 864         if (!curr)
 865                 goto delete_fail;
 866
 867         ret = btrfs_search_slot(trans, root, &curr->key, path, -1, 1);
 868         if (ret < 0)
 869                 goto delete_fail;
 870         else if (ret > 0) {
 871                 /*
 872                  * can't find the item which the node points to, so this node
 873                  * is invalid, just drop it.
 874                  */
 875                 prev = curr;
 876                 curr = __btrfs_next_delayed_item(prev);
 877                 btrfs_release_delayed_item(prev);
 878                 ret = 0;
 879                 btrfs_release_path(path);
 880                 if (curr) {
 881                         mutex_unlock(&node->mutex);
 882                         goto do_again;
 883                 } else
 884                         goto delete_fail;
 885         }
 886
 887         btrfs_batch_delete_items(trans, root, path, curr);
 888         btrfs_release_path(path);
 889         mutex_unlock(&node->mutex);
 890         goto do_again;
 891
 892 delete_fail:
 893         btrfs_release_path(path);
 894         mutex_unlock(&node->mutex);
 895         return ret;
 896 }
 897
 898 static void btrfs_release_delayed_inode(struct btrfs_delayed_node *delayed_node)
 899 {
 900         struct btrfs_delayed_root *delayed_root;
 901
 902         if (delayed_node &&
 903             test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
 904                 BUG_ON(!delayed_node->root);
 905                 clear_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags);
 906                 delayed_node->count--;
 907
 908                 delayed_root = delayed_node->root->fs_info->delayed_root;
 909                 finish_one_item(delayed_root);
 910         }
 911 }
 912
 913 static void btrfs_release_delayed_iref(struct btrfs_delayed_node *delayed_node)
 914 {
 915
 916         if (test_and_clear_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags)) {
 917                 struct btrfs_delayed_root *delayed_root;
 918
 919                 ASSERT(delayed_node->root);
 920                 delayed_node->count--;
 921
 922                 delayed_root = delayed_node->root->fs_info->delayed_root;
 923                 finish_one_item(delayed_root);
 924         }
 925 }
 926
 927 static int __btrfs_update_delayed_inode(struct btrfs_trans_handle *trans,
 928                                         struct btrfs_root *root,
 929                                         struct btrfs_path *path,
 930                                         struct btrfs_delayed_node *node)
 931 {
 932         struct btrfs_fs_info *fs_info = root->fs_info;
 933         struct btrfs_key key;
 934         struct btrfs_inode_item *inode_item;
 935         struct extent_buffer *leaf;
 936         int mod;
 937         int ret;
 938
 939         key.objectid = node->inode_id;
 940         key.type = BTRFS_INODE_ITEM_KEY;
 941         key.offset = 0;
 942
 943         if (test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &node->flags))
 944                 mod = -1;
 945         else
 946                 mod = 1;
 947
 948         ret = btrfs_lookup_inode(trans, root, path, &key, mod);
 949         if (ret > 0)
 950                 ret = -ENOENT;
 951         if (ret < 0)
 952                 goto out;
 953
 954         leaf = path->nodes[0];
 955         inode_item = btrfs_item_ptr(leaf, path->slots[0],
 956                                     struct btrfs_inode_item);
 957         write_extent_buffer(leaf, &node->inode_item, (unsigned long)inode_item,
 958                             sizeof(struct btrfs_inode_item));
 959         btrfs_mark_buffer_dirty(leaf);
 960
 961         if (!test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &node->flags))
 962                 goto out;
 963
 964         path->slots[0]++;
 965         if (path->slots[0] >= btrfs_header_nritems(leaf))
 966                 goto search;
 967 again:
 968         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 969         if (key.objectid != node->inode_id)
 970                 goto out;
 971
 972         if (key.type != BTRFS_INODE_REF_KEY &&
 973             key.type != BTRFS_INODE_EXTREF_KEY)
 974                 goto out;
 975
 976         /*
 977          * Delayed iref deletion is for the inode who has only one link,
 978          * so there is only one iref. The case that several irefs are
 979          * in the same item doesn't exist.
 980          */
 981         btrfs_del_item(trans, root, path);
 982 out:
 983         btrfs_release_delayed_iref(node);
 984         btrfs_release_path(path);
 985 err_out:
 986         btrfs_delayed_inode_release_metadata(fs_info, node, (ret < 0));
 987         btrfs_release_delayed_inode(node);
 988
 989         /*
 990          * If we fail to update the delayed inode we need to abort the
 991          * transaction, because we could leave the inode with the improper
 992          * counts behind.
 993          */
 994         if (ret && ret != -ENOENT)
 995                 btrfs_abort_transaction(trans, ret);
 996
 997         return ret;
 998
 999 search:
1000         btrfs_release_path(path);
1001
1002         key.type = BTRFS_INODE_EXTREF_KEY;
1003         key.offset = -1;
1004
1005         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1006         if (ret < 0)
1007                 goto err_out;
1008         ASSERT(ret);
1009
1010         ret = 0;
1011         leaf = path->nodes[0];
1012         path->slots[0]--;
1013         goto again;
1014 }
1015
1016 static inline int btrfs_update_delayed_inode(struct btrfs_trans_handle *trans,
1017                                              struct btrfs_root *root,
1018                                              struct btrfs_path *path,
1019                                              struct btrfs_delayed_node *node)
1020 {
1021         int ret;
1022
1023         mutex_lock(&node->mutex);
1024         if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &node->flags)) {
1025                 mutex_unlock(&node->mutex);
1026                 return 0;
1027         }
1028
1029         ret = __btrfs_update_delayed_inode(trans, root, path, node);
1030         mutex_unlock(&node->mutex);
1031         return ret;
1032 }
1033
1034 static inline int
1035 __btrfs_commit_inode_delayed_items(struct btrfs_trans_handle *trans,
1036                                    struct btrfs_path *path,
1037                                    struct btrfs_delayed_node *node)
1038 {
1039         int ret;
1040
1041         ret = btrfs_insert_delayed_items(trans, path, node->root, node);
1042         if (ret)
1043                 return ret;
1044
1045         ret = btrfs_delete_delayed_items(trans, path, node->root, node);
1046         if (ret)
1047                 return ret;
1048
1049         ret = btrfs_update_delayed_inode(trans, node->root, path, node);
1050         return ret;
1051 }
1052
1053 /*
1054  * Called when committing the transaction.
1055  * Returns 0 on success.
1056  * Returns < 0 on error and returns with an aborted transaction with any
1057  * outstanding delayed items cleaned up.
1058  */
1059 static int __btrfs_run_delayed_items(struct btrfs_trans_handle *trans, int nr)
1060 {
1061         struct btrfs_fs_info *fs_info = trans->fs_info;
1062         struct btrfs_delayed_root *delayed_root;
1063         struct btrfs_delayed_node *curr_node, *prev_node;
1064         struct btrfs_path *path;
1065         struct btrfs_block_rsv *block_rsv;
1066         int ret = 0;
1067         bool count = (nr > 0);
1068
1069         if (TRANS_ABORTED(trans))
1070                 return -EIO;
1071
1072         path = btrfs_alloc_path();
1073         if (!path)
1074                 return -ENOMEM;
1075
1076         block_rsv = trans->block_rsv;
1077         trans->block_rsv = &fs_info->delayed_block_rsv;
1078
1079         delayed_root = fs_info->delayed_root;
1080
1081         curr_node = btrfs_first_delayed_node(delayed_root);
1082         while (curr_node && (!count || nr--)) {
1083                 ret = __btrfs_commit_inode_delayed_items(trans, path,
1084                                                          curr_node);
1085                 if (ret) {
1086                         btrfs_release_delayed_node(curr_node);
1087                         curr_node = NULL;
1088                         btrfs_abort_transaction(trans, ret);
1089                         break;
1090                 }
1091
1092                 prev_node = curr_node;
1093                 curr_node = btrfs_next_delayed_node(curr_node);
1094                 btrfs_release_delayed_node(prev_node);
1095         }
1096
1097         if (curr_node)
1098                 btrfs_release_delayed_node(curr_node);
1099         btrfs_free_path(path);
1100         trans->block_rsv = block_rsv;
1101
1102         return ret;
1103 }
1104
1105 int btrfs_run_delayed_items(struct btrfs_trans_handle *trans)
1106 {
1107         return __btrfs_run_delayed_items(trans, -1);
1108 }
1109
1110 int btrfs_run_delayed_items_nr(struct btrfs_trans_handle *trans, int nr)
1111 {
1112         return __btrfs_run_delayed_items(trans, nr);
1113 }
1114
1115 int btrfs_commit_inode_delayed_items(struct btrfs_trans_handle *trans,
1116                                      struct btrfs_inode *inode)
1117 {
1118         struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1119         struct btrfs_path *path;
1120         struct btrfs_block_rsv *block_rsv;
1121         int ret;
1122
1123         if (!delayed_node)
1124                 return 0;
1125
1126         mutex_lock(&delayed_node->mutex);
1127         if (!delayed_node->count) {
1128                 mutex_unlock(&delayed_node->mutex);
1129                 btrfs_release_delayed_node(delayed_node);
1130                 return 0;
1131         }
1132         mutex_unlock(&delayed_node->mutex);
1133
1134         path = btrfs_alloc_path();
1135         if (!path) {
1136                 btrfs_release_delayed_node(delayed_node);
1137                 return -ENOMEM;
1138         }
1139
1140         block_rsv = trans->block_rsv;
1141         trans->block_rsv = &delayed_node->root->fs_info->delayed_block_rsv;
1142
1143         ret = __btrfs_commit_inode_delayed_items(trans, path, delayed_node);
1144
1145         btrfs_release_delayed_node(delayed_node);
1146         btrfs_free_path(path);
1147         trans->block_rsv = block_rsv;
1148
1149         return ret;
1150 }
1151
1152 int btrfs_commit_inode_delayed_inode(struct btrfs_inode *inode)
1153 {
1154         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1155         struct btrfs_trans_handle *trans;
1156         struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1157         struct btrfs_path *path;
1158         struct btrfs_block_rsv *block_rsv;
1159         int ret;
1160
1161         if (!delayed_node)
1162                 return 0;
1163
1164         mutex_lock(&delayed_node->mutex);
1165         if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1166                 mutex_unlock(&delayed_node->mutex);
1167                 btrfs_release_delayed_node(delayed_node);
1168                 return 0;
1169         }
1170         mutex_unlock(&delayed_node->mutex);
1171
1172         trans = btrfs_join_transaction(delayed_node->root);
1173         if (IS_ERR(trans)) {
1174                 ret = PTR_ERR(trans);
1175                 goto out;
1176         }
1177
1178         path = btrfs_alloc_path();
1179         if (!path) {
1180                 ret = -ENOMEM;
1181                 goto trans_out;
1182         }
1183
1184         block_rsv = trans->block_rsv;
1185         trans->block_rsv = &fs_info->delayed_block_rsv;
1186
1187         mutex_lock(&delayed_node->mutex);
1188         if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags))
1189                 ret = __btrfs_update_delayed_inode(trans, delayed_node->root,
1190                                                    path, delayed_node);
1191         else
1192                 ret = 0;
1193         mutex_unlock(&delayed_node->mutex);
1194
1195         btrfs_free_path(path);
1196         trans->block_rsv = block_rsv;
1197 trans_out:
1198         btrfs_end_transaction(trans);
1199         btrfs_btree_balance_dirty(fs_info);
1200 out:
1201         btrfs_release_delayed_node(delayed_node);
1202
1203         return ret;
1204 }
1205
1206 void btrfs_remove_delayed_node(struct btrfs_inode *inode)
1207 {
1208         struct btrfs_delayed_node *delayed_node;
1209
1210         delayed_node = READ_ONCE(inode->delayed_node);
1211         if (!delayed_node)
1212                 return;
1213
1214         inode->delayed_node = NULL;
1215         btrfs_release_delayed_node(delayed_node);
1216 }
1217
1218 struct btrfs_async_delayed_work {
1219         struct btrfs_delayed_root *delayed_root;
1220         int nr;
1221         struct btrfs_work work;
1222 };
1223
1224 static void btrfs_async_run_delayed_root(struct btrfs_work *work)
1225 {
1226         struct btrfs_async_delayed_work *async_work;
1227         struct btrfs_delayed_root *delayed_root;
1228         struct btrfs_trans_handle *trans;
1229         struct btrfs_path *path;
1230         struct btrfs_delayed_node *delayed_node = NULL;
1231         struct btrfs_root *root;
1232         struct btrfs_block_rsv *block_rsv;
1233         int total_done = 0;
1234
1235         async_work = container_of(work, struct btrfs_async_delayed_work, work);
1236         delayed_root = async_work->delayed_root;
1237
1238         path = btrfs_alloc_path();
1239         if (!path)
1240                 goto out;
1241
1242         do {
1243                 if (atomic_read(&delayed_root->items) <
1244                     BTRFS_DELAYED_BACKGROUND / 2)
1245                         break;
1246
1247                 delayed_node = btrfs_first_prepared_delayed_node(delayed_root);
1248                 if (!delayed_node)
1249                         break;
1250
1251                 root = delayed_node->root;
1252
1253                 trans = btrfs_join_transaction(root);
1254                 if (IS_ERR(trans)) {
1255                         btrfs_release_path(path);
1256                         btrfs_release_prepared_delayed_node(delayed_node);
1257                         total_done++;
1258                         continue;
1259                 }
1260
1261                 block_rsv = trans->block_rsv;
1262                 trans->block_rsv = &root->fs_info->delayed_block_rsv;
1263
1264                 __btrfs_commit_inode_delayed_items(trans, path, delayed_node);
1265
1266                 trans->block_rsv = block_rsv;
1267                 btrfs_end_transaction(trans);
1268                 btrfs_btree_balance_dirty_nodelay(root->fs_info);
1269
1270                 btrfs_release_path(path);
1271                 btrfs_release_prepared_delayed_node(delayed_node);
1272                 total_done++;
1273
1274         } while ((async_work->nr == 0 && total_done < BTRFS_DELAYED_WRITEBACK)
1275                  || total_done < async_work->nr);
1276
1277         btrfs_free_path(path);
1278 out:
1279         wake_up(&delayed_root->wait);
1280         kfree(async_work);
1281 }
1282
1283
1284 static int btrfs_wq_run_delayed_node(struct btrfs_delayed_root *delayed_root,
1285                                      struct btrfs_fs_info *fs_info, int nr)
1286 {
1287         struct btrfs_async_delayed_work *async_work;
1288
1289         async_work = kmalloc(sizeof(*async_work), GFP_NOFS);
1290         if (!async_work)
1291                 return -ENOMEM;
1292
1293         async_work->delayed_root = delayed_root;
1294         btrfs_init_work(&async_work->work, btrfs_async_run_delayed_root, NULL,
1295                         NULL);
1296         async_work->nr = nr;
1297
1298         btrfs_queue_work(fs_info->delayed_workers, &async_work->work);
1299         return 0;
1300 }
1301
1302 void btrfs_assert_delayed_root_empty(struct btrfs_fs_info *fs_info)
1303 {
1304         WARN_ON(btrfs_first_delayed_node(fs_info->delayed_root));
1305 }
1306
1307 static int could_end_wait(struct btrfs_delayed_root *delayed_root, int seq)
1308 {
1309         int val = atomic_read(&delayed_root->items_seq);
1310
1311         if (val < seq || val >= seq + BTRFS_DELAYED_BATCH)
1312                 return 1;
1313
1314         if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND)
1315                 return 1;
1316
1317         return 0;
1318 }
1319
1320 void btrfs_balance_delayed_items(struct btrfs_fs_info *fs_info)
1321 {
1322         struct btrfs_delayed_root *delayed_root = fs_info->delayed_root;
1323
1324         if ((atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND) ||
1325                 btrfs_workqueue_normal_congested(fs_info->delayed_workers))
1326                 return;
1327
1328         if (atomic_read(&delayed_root->items) >= BTRFS_DELAYED_WRITEBACK) {
1329                 int seq;
1330                 int ret;
1331
1332                 seq = atomic_read(&delayed_root->items_seq);
1333
1334                 ret = btrfs_wq_run_delayed_node(delayed_root, fs_info, 0);
1335                 if (ret)
1336                         return;
1337
1338                 wait_event_interruptible(delayed_root->wait,
1339                                          could_end_wait(delayed_root, seq));
1340                 return;
1341         }
1342
1343         btrfs_wq_run_delayed_node(delayed_root, fs_info, BTRFS_DELAYED_BATCH);
1344 }
1345
1346 /* Will return 0 or -ENOMEM */
1347 int btrfs_insert_delayed_dir_index(struct btrfs_trans_handle *trans,
1348                                    const char *name, int name_len,
1349                                    struct btrfs_inode *dir,
1350                                    struct btrfs_disk_key *disk_key, u8 type,
1351                                    u64 index)
1352 {
1353         struct btrfs_delayed_node *delayed_node;
1354         struct btrfs_delayed_item *delayed_item;
1355         struct btrfs_dir_item *dir_item;
1356         int ret;
1357
1358         delayed_node = btrfs_get_or_create_delayed_node(dir);
1359         if (IS_ERR(delayed_node))
1360                 return PTR_ERR(delayed_node);
1361
1362         delayed_item = btrfs_alloc_delayed_item(sizeof(*dir_item) + name_len);
1363         if (!delayed_item) {
1364                 ret = -ENOMEM;
1365                 goto release_node;
1366         }
1367
1368         delayed_item->key.objectid = btrfs_ino(dir);
1369         delayed_item->key.type = BTRFS_DIR_INDEX_KEY;
1370         delayed_item->key.offset = index;
1371
1372         dir_item = (struct btrfs_dir_item *)delayed_item->data;
1373         dir_item->location = *disk_key;
1374         btrfs_set_stack_dir_transid(dir_item, trans->transid);
1375         btrfs_set_stack_dir_data_len(dir_item, 0);
1376         btrfs_set_stack_dir_name_len(dir_item, name_len);
1377         btrfs_set_stack_dir_type(dir_item, type);
1378         memcpy((char *)(dir_item + 1), name, name_len);
1379
1380         ret = btrfs_delayed_item_reserve_metadata(trans, dir->root, delayed_item);
1381         /*
1382          * we have reserved enough space when we start a new transaction,
1383          * so reserving metadata failure is impossible
1384          */
1385         BUG_ON(ret);
1386
1387         mutex_lock(&delayed_node->mutex);
1388         ret = __btrfs_add_delayed_insertion_item(delayed_node, delayed_item);
1389         if (unlikely(ret)) {
1390                 btrfs_err(trans->fs_info,
1391                           "err add delayed dir index item(name: %.*s) into the insertion tree of the delayed node(root id: %llu, inode id: %llu, errno: %d)",
1392                           name_len, name, delayed_node->root->root_key.objectid,
1393                           delayed_node->inode_id, ret);
1394                 BUG();
1395         }
1396         mutex_unlock(&delayed_node->mutex);
1397
1398 release_node:
1399         btrfs_release_delayed_node(delayed_node);
1400         return ret;
1401 }
1402
1403 static int btrfs_delete_delayed_insertion_item(struct btrfs_fs_info *fs_info,
1404                                                struct btrfs_delayed_node *node,
1405                                                struct btrfs_key *key)
1406 {
1407         struct btrfs_delayed_item *item;
1408
1409         mutex_lock(&node->mutex);
1410         item = __btrfs_lookup_delayed_insertion_item(node, key);
1411         if (!item) {
1412                 mutex_unlock(&node->mutex);
1413                 return 1;
1414         }
1415
1416         btrfs_delayed_item_release_metadata(node->root, item);
1417         btrfs_release_delayed_item(item);
1418         mutex_unlock(&node->mutex);
1419         return 0;
1420 }
1421
1422 int btrfs_delete_delayed_dir_index(struct btrfs_trans_handle *trans,
1423                                    struct btrfs_inode *dir, u64 index)
1424 {
1425         struct btrfs_delayed_node *node;
1426         struct btrfs_delayed_item *item;
1427         struct btrfs_key item_key;
1428         int ret;
1429
1430         node = btrfs_get_or_create_delayed_node(dir);
1431         if (IS_ERR(node))
1432                 return PTR_ERR(node);
1433
1434         item_key.objectid = btrfs_ino(dir);
1435         item_key.type = BTRFS_DIR_INDEX_KEY;
1436         item_key.offset = index;
1437
1438         ret = btrfs_delete_delayed_insertion_item(trans->fs_info, node,
1439                                                   &item_key);
1440         if (!ret)
1441                 goto end;
1442
1443         item = btrfs_alloc_delayed_item(0);
1444         if (!item) {
1445                 ret = -ENOMEM;
1446                 goto end;
1447         }
1448
1449         item->key = item_key;
1450
1451         ret = btrfs_delayed_item_reserve_metadata(trans, dir->root, item);
1452         /*
1453          * we have reserved enough space when we start a new transaction,
1454          * so reserving metadata failure is impossible.
1455          */
1456         if (ret < 0) {
1457                 btrfs_err(trans->fs_info,
1458 "metadata reservation failed for delayed dir item deltiona, should have been reserved");
1459                 btrfs_release_delayed_item(item);
1460                 goto end;
1461         }
1462
1463         mutex_lock(&node->mutex);
1464         ret = __btrfs_add_delayed_deletion_item(node, item);
1465         if (unlikely(ret)) {
1466                 btrfs_err(trans->fs_info,
1467                           "err add delayed dir index item(index: %llu) into the deletion tree of the delayed node(root id: %llu, inode id: %llu, errno: %d)",
1468                           index, node->root->root_key.objectid,
1469                           node->inode_id, ret);
1470                 btrfs_delayed_item_release_metadata(dir->root, item);
1471                 btrfs_release_delayed_item(item);
1472         }
1473         mutex_unlock(&node->mutex);
1474 end:
1475         btrfs_release_delayed_node(node);
1476         return ret;
1477 }
1478
1479 int btrfs_inode_delayed_dir_index_count(struct btrfs_inode *inode)
1480 {
1481         struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1482
1483         if (!delayed_node)
1484                 return -ENOENT;
1485
1486         /*
1487          * Since we have held i_mutex of this directory, it is impossible that
1488          * a new directory index is added into the delayed node and index_cnt
1489          * is updated now. So we needn't lock the delayed node.
1490          */
1491         if (!delayed_node->index_cnt) {
1492                 btrfs_release_delayed_node(delayed_node);
1493                 return -EINVAL;
1494         }
1495
1496         inode->index_cnt = delayed_node->index_cnt;
1497         btrfs_release_delayed_node(delayed_node);
1498         return 0;
1499 }
1500
1501 bool btrfs_readdir_get_delayed_items(struct inode *inode,
1502                                      struct list_head *ins_list,
1503                                      struct list_head *del_list)
1504 {
1505         struct btrfs_delayed_node *delayed_node;
1506         struct btrfs_delayed_item *item;
1507
1508         delayed_node = btrfs_get_delayed_node(BTRFS_I(inode));
1509         if (!delayed_node)
1510                 return false;
1511
1512         /*
1513          * We can only do one readdir with delayed items at a time because of
1514          * item->readdir_list.
1515          */
1516         btrfs_inode_unlock(inode, BTRFS_ILOCK_SHARED);
1517         btrfs_inode_lock(inode, 0);
1518
1519         mutex_lock(&delayed_node->mutex);
1520         item = __btrfs_first_delayed_insertion_item(delayed_node);
1521         while (item) {
1522                 refcount_inc(&item->refs);
1523                 list_add_tail(&item->readdir_list, ins_list);
1524                 item = __btrfs_next_delayed_item(item);
1525         }
1526
1527         item = __btrfs_first_delayed_deletion_item(delayed_node);
1528         while (item) {
1529                 refcount_inc(&item->refs);
1530                 list_add_tail(&item->readdir_list, del_list);
1531                 item = __btrfs_next_delayed_item(item);
1532         }
1533         mutex_unlock(&delayed_node->mutex);
1534         /*
1535          * This delayed node is still cached in the btrfs inode, so refs
1536          * must be > 1 now, and we needn't check it is going to be freed
1537          * or not.
1538          *
1539          * Besides that, this function is used to read dir, we do not
1540          * insert/delete delayed items in this period. So we also needn't
1541          * requeue or dequeue this delayed node.
1542          */
1543         refcount_dec(&delayed_node->refs);
1544
1545         return true;
1546 }
1547
1548 void btrfs_readdir_put_delayed_items(struct inode *inode,
1549                                      struct list_head *ins_list,
1550                                      struct list_head *del_list)
1551 {
1552         struct btrfs_delayed_item *curr, *next;
1553
1554         list_for_each_entry_safe(curr, next, ins_list, readdir_list) {
1555                 list_del(&curr->readdir_list);
1556                 if (refcount_dec_and_test(&curr->refs))
1557                         kfree(curr);
1558         }
1559
1560         list_for_each_entry_safe(curr, next, del_list, readdir_list) {
1561                 list_del(&curr->readdir_list);
1562                 if (refcount_dec_and_test(&curr->refs))
1563                         kfree(curr);
1564         }
1565
1566         /*
1567          * The VFS is going to do up_read(), so we need to downgrade back to a
1568          * read lock.
1569          */
1570         downgrade_write(&inode->i_rwsem);
1571 }
1572
1573 int btrfs_should_delete_dir_index(struct list_head *del_list,
1574                                   u64 index)
1575 {
1576         struct btrfs_delayed_item *curr;
1577         int ret = 0;
1578
1579         list_for_each_entry(curr, del_list, readdir_list) {
1580                 if (curr->key.offset > index)
1581                         break;
1582                 if (curr->key.offset == index) {
1583                         ret = 1;
1584                         break;
1585                 }
1586         }
1587         return ret;
1588 }
1589
1590 /*
1591  * btrfs_readdir_delayed_dir_index - read dir info stored in the delayed tree
1592  *
1593  */
1594 int btrfs_readdir_delayed_dir_index(struct dir_context *ctx,
1595                                     struct list_head *ins_list)
1596 {
1597         struct btrfs_dir_item *di;
1598         struct btrfs_delayed_item *curr, *next;
1599         struct btrfs_key location;
1600         char *name;
1601         int name_len;
1602         int over = 0;
1603         unsigned char d_type;
1604
1605         if (list_empty(ins_list))
1606                 return 0;
1607
1608         /*
1609          * Changing the data of the delayed item is impossible. So
1610          * we needn't lock them. And we have held i_mutex of the
1611          * directory, nobody can delete any directory indexes now.
1612          */
1613         list_for_each_entry_safe(curr, next, ins_list, readdir_list) {
1614                 list_del(&curr->readdir_list);
1615
1616                 if (curr->key.offset < ctx->pos) {
1617                         if (refcount_dec_and_test(&curr->refs))
1618                                 kfree(curr);
1619                         continue;
1620                 }
1621
1622                 ctx->pos = curr->key.offset;
1623
1624                 di = (struct btrfs_dir_item *)curr->data;
1625                 name = (char *)(di + 1);
1626                 name_len = btrfs_stack_dir_name_len(di);
1627
1628                 d_type = fs_ftype_to_dtype(di->type);
1629                 btrfs_disk_key_to_cpu(&location, &di->location);
1630
1631                 over = !dir_emit(ctx, name, name_len,
1632                                location.objectid, d_type);
1633
1634                 if (refcount_dec_and_test(&curr->refs))
1635                         kfree(curr);
1636
1637                 if (over)
1638                         return 1;
1639                 ctx->pos++;
1640         }
1641         return 0;
1642 }
1643
1644 static void fill_stack_inode_item(struct btrfs_trans_handle *trans,
1645                                   struct btrfs_inode_item *inode_item,
1646                                   struct inode *inode)
1647 {
1648         u64 flags;
1649
1650         btrfs_set_stack_inode_uid(inode_item, i_uid_read(inode));
1651         btrfs_set_stack_inode_gid(inode_item, i_gid_read(inode));
1652         btrfs_set_stack_inode_size(inode_item, BTRFS_I(inode)->disk_i_size);
1653         btrfs_set_stack_inode_mode(inode_item, inode->i_mode);
1654         btrfs_set_stack_inode_nlink(inode_item, inode->i_nlink);
1655         btrfs_set_stack_inode_nbytes(inode_item, inode_get_bytes(inode));
1656         btrfs_set_stack_inode_generation(inode_item,
1657                                          BTRFS_I(inode)->generation);
1658         btrfs_set_stack_inode_sequence(inode_item,
1659                                        inode_peek_iversion(inode));
1660         btrfs_set_stack_inode_transid(inode_item, trans->transid);
1661         btrfs_set_stack_inode_rdev(inode_item, inode->i_rdev);
1662         flags = btrfs_inode_combine_flags(BTRFS_I(inode)->flags,
1663                                           BTRFS_I(inode)->ro_flags);
1664         btrfs_set_stack_inode_flags(inode_item, flags);
1665         btrfs_set_stack_inode_block_group(inode_item, 0);
1666
1667         btrfs_set_stack_timespec_sec(&inode_item->atime,
1668                                      inode->i_atime.tv_sec);
1669         btrfs_set_stack_timespec_nsec(&inode_item->atime,
1670                                       inode->i_atime.tv_nsec);
1671
1672         btrfs_set_stack_timespec_sec(&inode_item->mtime,
1673                                      inode->i_mtime.tv_sec);
1674         btrfs_set_stack_timespec_nsec(&inode_item->mtime,
1675                                       inode->i_mtime.tv_nsec);
1676
1677         btrfs_set_stack_timespec_sec(&inode_item->ctime,
1678                                      inode->i_ctime.tv_sec);
1679         btrfs_set_stack_timespec_nsec(&inode_item->ctime,
1680                                       inode->i_ctime.tv_nsec);
1681
1682         btrfs_set_stack_timespec_sec(&inode_item->otime,
1683                                      BTRFS_I(inode)->i_otime.tv_sec);
1684         btrfs_set_stack_timespec_nsec(&inode_item->otime,
1685                                      BTRFS_I(inode)->i_otime.tv_nsec);
1686 }
1687
1688 int btrfs_fill_inode(struct inode *inode, u32 *rdev)
1689 {
1690         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
1691         struct btrfs_delayed_node *delayed_node;
1692         struct btrfs_inode_item *inode_item;
1693
1694         delayed_node = btrfs_get_delayed_node(BTRFS_I(inode));
1695         if (!delayed_node)
1696                 return -ENOENT;
1697
1698         mutex_lock(&delayed_node->mutex);
1699         if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1700                 mutex_unlock(&delayed_node->mutex);
1701                 btrfs_release_delayed_node(delayed_node);
1702                 return -ENOENT;
1703         }
1704
1705         inode_item = &delayed_node->inode_item;
1706
1707         i_uid_write(inode, btrfs_stack_inode_uid(inode_item));
1708         i_gid_write(inode, btrfs_stack_inode_gid(inode_item));
1709         btrfs_i_size_write(BTRFS_I(inode), btrfs_stack_inode_size(inode_item));
1710         btrfs_inode_set_file_extent_range(BTRFS_I(inode), 0,
1711                         round_up(i_size_read(inode), fs_info->sectorsize));
1712         inode->i_mode = btrfs_stack_inode_mode(inode_item);
1713         set_nlink(inode, btrfs_stack_inode_nlink(inode_item));
1714         inode_set_bytes(inode, btrfs_stack_inode_nbytes(inode_item));
1715         BTRFS_I(inode)->generation = btrfs_stack_inode_generation(inode_item);
1716         BTRFS_I(inode)->last_trans = btrfs_stack_inode_transid(inode_item);
1717
1718         inode_set_iversion_queried(inode,
1719                                    btrfs_stack_inode_sequence(inode_item));
1720         inode->i_rdev = 0;
1721         *rdev = btrfs_stack_inode_rdev(inode_item);
1722         btrfs_inode_split_flags(btrfs_stack_inode_flags(inode_item),
1723                                 &BTRFS_I(inode)->flags, &BTRFS_I(inode)->ro_flags);
1724
1725         inode->i_atime.tv_sec = btrfs_stack_timespec_sec(&inode_item->atime);
1726         inode->i_atime.tv_nsec = btrfs_stack_timespec_nsec(&inode_item->atime);
1727
1728         inode->i_mtime.tv_sec = btrfs_stack_timespec_sec(&inode_item->mtime);
1729         inode->i_mtime.tv_nsec = btrfs_stack_timespec_nsec(&inode_item->mtime);
1730
1731         inode->i_ctime.tv_sec = btrfs_stack_timespec_sec(&inode_item->ctime);
1732         inode->i_ctime.tv_nsec = btrfs_stack_timespec_nsec(&inode_item->ctime);
1733
1734         BTRFS_I(inode)->i_otime.tv_sec =
1735                 btrfs_stack_timespec_sec(&inode_item->otime);
1736         BTRFS_I(inode)->i_otime.tv_nsec =
1737                 btrfs_stack_timespec_nsec(&inode_item->otime);
1738
1739         inode->i_generation = BTRFS_I(inode)->generation;
1740         BTRFS_I(inode)->index_cnt = (u64)-1;
1741
1742         mutex_unlock(&delayed_node->mutex);
1743         btrfs_release_delayed_node(delayed_node);
1744         return 0;
1745 }
1746
1747 int btrfs_delayed_update_inode(struct btrfs_trans_handle *trans,
1748                                struct btrfs_root *root,
1749                                struct btrfs_inode *inode)
1750 {
1751         struct btrfs_delayed_node *delayed_node;
1752         int ret = 0;
1753
1754         delayed_node = btrfs_get_or_create_delayed_node(inode);
1755         if (IS_ERR(delayed_node))
1756                 return PTR_ERR(delayed_node);
1757
1758         mutex_lock(&delayed_node->mutex);
1759         if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1760                 fill_stack_inode_item(trans, &delayed_node->inode_item,
1761                                       &inode->vfs_inode);
1762                 goto release_node;
1763         }
1764
1765         ret = btrfs_delayed_inode_reserve_metadata(trans, root, delayed_node);
1766         if (ret)
1767                 goto release_node;
1768
1769         fill_stack_inode_item(trans, &delayed_node->inode_item, &inode->vfs_inode);
1770         set_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags);
1771         delayed_node->count++;
1772         atomic_inc(&root->fs_info->delayed_root->items);
1773 release_node:
1774         mutex_unlock(&delayed_node->mutex);
1775         btrfs_release_delayed_node(delayed_node);
1776         return ret;
1777 }
1778
1779 int btrfs_delayed_delete_inode_ref(struct btrfs_inode *inode)
1780 {
1781         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1782         struct btrfs_delayed_node *delayed_node;
1783
1784         /*
1785          * we don't do delayed inode updates during log recovery because it
1786          * leads to enospc problems.  This means we also can't do
1787          * delayed inode refs
1788          */
1789         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
1790                 return -EAGAIN;
1791
1792         delayed_node = btrfs_get_or_create_delayed_node(inode);
1793         if (IS_ERR(delayed_node))
1794                 return PTR_ERR(delayed_node);
1795
1796         /*
1797          * We don't reserve space for inode ref deletion is because:
1798          * - We ONLY do async inode ref deletion for the inode who has only
1799          *   one link(i_nlink == 1), it means there is only one inode ref.
1800          *   And in most case, the inode ref and the inode item are in the
1801          *   same leaf, and we will deal with them at the same time.
1802          *   Since we are sure we will reserve the space for the inode item,
1803          *   it is unnecessary to reserve space for inode ref deletion.
1804          * - If the inode ref and the inode item are not in the same leaf,
1805          *   We also needn't worry about enospc problem, because we reserve
1806          *   much more space for the inode update than it needs.
1807          * - At the worst, we can steal some space from the global reservation.
1808          *   It is very rare.
1809          */
1810         mutex_lock(&delayed_node->mutex);
1811         if (test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags))
1812                 goto release_node;
1813
1814         set_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags);
1815         delayed_node->count++;
1816         atomic_inc(&fs_info->delayed_root->items);
1817 release_node:
1818         mutex_unlock(&delayed_node->mutex);
1819         btrfs_release_delayed_node(delayed_node);
1820         return 0;
1821 }
1822
1823 static void __btrfs_kill_delayed_node(struct btrfs_delayed_node *delayed_node)
1824 {
1825         struct btrfs_root *root = delayed_node->root;
1826         struct btrfs_fs_info *fs_info = root->fs_info;
1827         struct btrfs_delayed_item *curr_item, *prev_item;
1828
1829         mutex_lock(&delayed_node->mutex);
1830         curr_item = __btrfs_first_delayed_insertion_item(delayed_node);
1831         while (curr_item) {
1832                 btrfs_delayed_item_release_metadata(root, curr_item);
1833                 prev_item = curr_item;
1834                 curr_item = __btrfs_next_delayed_item(prev_item);
1835                 btrfs_release_delayed_item(prev_item);
1836         }
1837
1838         curr_item = __btrfs_first_delayed_deletion_item(delayed_node);
1839         while (curr_item) {
1840                 btrfs_delayed_item_release_metadata(root, curr_item);
1841                 prev_item = curr_item;
1842                 curr_item = __btrfs_next_delayed_item(prev_item);
1843                 btrfs_release_delayed_item(prev_item);
1844         }
1845
1846         btrfs_release_delayed_iref(delayed_node);
1847
1848         if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1849                 btrfs_delayed_inode_release_metadata(fs_info, delayed_node, false);
1850                 btrfs_release_delayed_inode(delayed_node);
1851         }
1852         mutex_unlock(&delayed_node->mutex);
1853 }
1854
1855 void btrfs_kill_delayed_inode_items(struct btrfs_inode *inode)
1856 {
1857         struct btrfs_delayed_node *delayed_node;
1858
1859         delayed_node = btrfs_get_delayed_node(inode);
1860         if (!delayed_node)
1861                 return;
1862
1863         __btrfs_kill_delayed_node(delayed_node);
1864         btrfs_release_delayed_node(delayed_node);
1865 }
1866
1867 void btrfs_kill_all_delayed_nodes(struct btrfs_root *root)
1868 {
1869         u64 inode_id = 0;
1870         struct btrfs_delayed_node *delayed_nodes[8];
1871         int i, n;
1872
1873         while (1) {
1874                 spin_lock(&root->inode_lock);
1875                 n = radix_tree_gang_lookup(&root->delayed_nodes_tree,
1876                                            (void **)delayed_nodes, inode_id,
1877                                            ARRAY_SIZE(delayed_nodes));
1878                 if (!n) {
1879                         spin_unlock(&root->inode_lock);
1880                         break;
1881                 }
1882
1883                 inode_id = delayed_nodes[n - 1]->inode_id + 1;
1884                 for (i = 0; i < n; i++) {
1885                         /*
1886                          * Don't increase refs in case the node is dead and
1887                          * about to be removed from the tree in the loop below
1888                          */
1889                         if (!refcount_inc_not_zero(&delayed_nodes[i]->refs))
1890                                 delayed_nodes[i] = NULL;
1891                 }
1892                 spin_unlock(&root->inode_lock);
1893
1894                 for (i = 0; i < n; i++) {
1895                         if (!delayed_nodes[i])
1896                                 continue;
1897                         __btrfs_kill_delayed_node(delayed_nodes[i]);
1898                         btrfs_release_delayed_node(delayed_nodes[i]);
1899                 }
1900         }
1901 }
1902
1903 void btrfs_destroy_delayed_inodes(struct btrfs_fs_info *fs_info)
1904 {
1905         struct btrfs_delayed_node *curr_node, *prev_node;
1906
1907         curr_node = btrfs_first_delayed_node(fs_info->delayed_root);
1908         while (curr_node) {
1909                 __btrfs_kill_delayed_node(curr_node);
1910
1911                 prev_node = curr_node;
1912                 curr_node = btrfs_next_delayed_node(curr_node);
1913                 btrfs_release_delayed_node(prev_node);
1914         }
1915 }
1916