fs/btrfs/delayed-inode.c

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