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