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 inline int btrfs_is_continuous_delayed_item(
  56                                         struct btrfs_delayed_item *item1,
  57                                         struct btrfs_delayed_item *item2)
  58 {
  59         if (item1->key.type == BTRFS_DIR_INDEX_KEY &&
  60             item1->key.objectid == item2->key.objectid &&
  61             item1->key.type == item2->key.type &&
  62             item1->key.offset + 1 == item2->key.offset)
  63                 return 1;
  64         return 0;
  65 }
  66
  67 static struct btrfs_delayed_node *btrfs_get_delayed_node(
  68                 struct btrfs_inode *btrfs_inode)
  69 {
  70         struct btrfs_root *root = btrfs_inode->root;
  71         u64 ino = btrfs_ino(btrfs_inode);
  72         struct btrfs_delayed_node *node;
  73
  74         node = READ_ONCE(btrfs_inode->delayed_node);
  75         if (node) {
  76                 refcount_inc(&node->refs);
  77                 return node;
  78         }
  79
  80         spin_lock(&root->inode_lock);
  81         node = xa_load(&root->delayed_nodes, ino);
  82
  83         if (node) {
  84                 if (btrfs_inode->delayed_node) {
  85                         refcount_inc(&node->refs);      /* can be accessed */
  86                         BUG_ON(btrfs_inode->delayed_node != node);
  87                         spin_unlock(&root->inode_lock);
  88                         return node;
  89                 }
  90
  91                 /*
  92                  * It's possible that we're racing into the middle of removing
  93                  * this node from the xarray.  In this case, the refcount
  94                  * was zero and it should never go back to one.  Just return
  95                  * NULL like it was never in the xarray at all; our release
  96                  * function is in the process of removing it.
  97                  *
  98                  * Some implementations of refcount_inc refuse to bump the
  99                  * refcount once it has hit zero.  If we don't do this dance
 100                  * here, refcount_inc() may decide to just WARN_ONCE() instead
 101                  * of actually bumping the refcount.
 102                  *
 103                  * If this node is properly in the xarray, we want to bump the
 104                  * refcount twice, once for the inode and once for this get
 105                  * operation.
 106                  */
 107                 if (refcount_inc_not_zero(&node->refs)) {
 108                         refcount_inc(&node->refs);
 109                         btrfs_inode->delayed_node = node;
 110                 } else {
 111                         node = NULL;
 112                 }
 113
 114                 spin_unlock(&root->inode_lock);
 115                 return node;
 116         }
 117         spin_unlock(&root->inode_lock);
 118
 119         return NULL;
 120 }
 121
 122 /* Will return either the node or PTR_ERR(-ENOMEM) */
 123 static struct btrfs_delayed_node *btrfs_get_or_create_delayed_node(
 124                 struct btrfs_inode *btrfs_inode)
 125 {
 126         struct btrfs_delayed_node *node;
 127         struct btrfs_root *root = btrfs_inode->root;
 128         u64 ino = btrfs_ino(btrfs_inode);
 129         int ret;
 130
 131         do {
 132                 node = btrfs_get_delayed_node(btrfs_inode);
 133                 if (node)
 134                         return node;
 135
 136                 node = kmem_cache_zalloc(delayed_node_cache, GFP_NOFS);
 137                 if (!node)
 138                         return ERR_PTR(-ENOMEM);
 139                 btrfs_init_delayed_node(node, root, ino);
 140
 141                 /* Cached in the inode and can be accessed */
 142                 refcount_set(&node->refs, 2);
 143
 144                 spin_lock(&root->inode_lock);
 145                 ret = xa_insert(&root->delayed_nodes, ino, node, GFP_NOFS);
 146                 if (ret) {
 147                         spin_unlock(&root->inode_lock);
 148                         kmem_cache_free(delayed_node_cache, node);
 149                         if (ret != -EBUSY)
 150                                 return ERR_PTR(ret);
 151                 }
 152         } while (ret);
 153         btrfs_inode->delayed_node = node;
 154         spin_unlock(&root->inode_lock);
 155
 156         return node;
 157 }
 158
 159 /*
 160  * Call it when holding delayed_node->mutex
 161  *
 162  * If mod = 1, add this node into the prepared list.
 163  */
 164 static void btrfs_queue_delayed_node(struct btrfs_delayed_root *root,
 165                                      struct btrfs_delayed_node *node,
 166                                      int mod)
 167 {
 168         spin_lock(&root->lock);
 169         if (test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
 170                 if (!list_empty(&node->p_list))
 171                         list_move_tail(&node->p_list, &root->prepare_list);
 172                 else if (mod)
 173                         list_add_tail(&node->p_list, &root->prepare_list);
 174         } else {
 175                 list_add_tail(&node->n_list, &root->node_list);
 176                 list_add_tail(&node->p_list, &root->prepare_list);
 177                 refcount_inc(&node->refs);      /* inserted into list */
 178                 root->nodes++;
 179                 set_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags);
 180         }
 181         spin_unlock(&root->lock);
 182 }
 183
 184 /* Call it when holding delayed_node->mutex */
 185 static void btrfs_dequeue_delayed_node(struct btrfs_delayed_root *root,
 186                                        struct btrfs_delayed_node *node)
 187 {
 188         spin_lock(&root->lock);
 189         if (test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
 190                 root->nodes--;
 191                 refcount_dec(&node->refs);      /* not in the list */
 192                 list_del_init(&node->n_list);
 193                 if (!list_empty(&node->p_list))
 194                         list_del_init(&node->p_list);
 195                 clear_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags);
 196         }
 197         spin_unlock(&root->lock);
 198 }
 199
 200 static struct btrfs_delayed_node *btrfs_first_delayed_node(
 201                         struct btrfs_delayed_root *delayed_root)
 202 {
 203         struct list_head *p;
 204         struct btrfs_delayed_node *node = NULL;
 205
 206         spin_lock(&delayed_root->lock);
 207         if (list_empty(&delayed_root->node_list))
 208                 goto out;
 209
 210         p = delayed_root->node_list.next;
 211         node = list_entry(p, struct btrfs_delayed_node, n_list);
 212         refcount_inc(&node->refs);
 213 out:
 214         spin_unlock(&delayed_root->lock);
 215
 216         return node;
 217 }
 218
 219 static struct btrfs_delayed_node *btrfs_next_delayed_node(
 220                                                 struct btrfs_delayed_node *node)
 221 {
 222         struct btrfs_delayed_root *delayed_root;
 223         struct list_head *p;
 224         struct btrfs_delayed_node *next = NULL;
 225
 226         delayed_root = node->root->fs_info->delayed_root;
 227         spin_lock(&delayed_root->lock);
 228         if (!test_bit(BTRFS_DELAYED_NODE_IN_LIST, &node->flags)) {
 229                 /* not in the list */
 230                 if (list_empty(&delayed_root->node_list))
 231                         goto out;
 232                 p = delayed_root->node_list.next;
 233         } else if (list_is_last(&node->n_list, &delayed_root->node_list))
 234                 goto out;
 235         else
 236                 p = node->n_list.next;
 237
 238         next = list_entry(p, struct btrfs_delayed_node, n_list);
 239         refcount_inc(&next->refs);
 240 out:
 241         spin_unlock(&delayed_root->lock);
 242
 243         return next;
 244 }
 245
 246 static void __btrfs_release_delayed_node(
 247                                 struct btrfs_delayed_node *delayed_node,
 248                                 int mod)
 249 {
 250         struct btrfs_delayed_root *delayed_root;
 251
 252         if (!delayed_node)
 253                 return;
 254
 255         delayed_root = delayed_node->root->fs_info->delayed_root;
 256
 257         mutex_lock(&delayed_node->mutex);
 258         if (delayed_node->count)
 259                 btrfs_queue_delayed_node(delayed_root, delayed_node, mod);
 260         else
 261                 btrfs_dequeue_delayed_node(delayed_root, delayed_node);
 262         mutex_unlock(&delayed_node->mutex);
 263
 264         if (refcount_dec_and_test(&delayed_node->refs)) {
 265                 struct btrfs_root *root = delayed_node->root;
 266
 267                 spin_lock(&root->inode_lock);
 268                 /*
 269                  * Once our refcount goes to zero, nobody is allowed to bump it
 270                  * back up.  We can delete it now.
 271                  */
 272                 ASSERT(refcount_read(&delayed_node->refs) == 0);
 273                 xa_erase(&root->delayed_nodes, 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(u32 data_len)
 311 {
 312         struct btrfs_delayed_item *item;
 313         item = kmalloc(sizeof(*item) + data_len, GFP_NOFS);
 314         if (item) {
 315                 item->data_len = data_len;
 316                 item->ins_or_del = 0;
 317                 item->bytes_reserved = 0;
 318                 item->delayed_node = NULL;
 319                 refcount_set(&item->refs, 1);
 320         }
 321         return item;
 322 }
 323
 324 /*
 325  * __btrfs_lookup_delayed_item - look up the delayed item by key
 326  * @delayed_node: pointer to the delayed node
 327  * @key:          the key to look up
 328  * @prev:         used to store the prev item if the right item isn't found
 329  * @next:         used to store the next item if the right item isn't found
 330  *
 331  * Note: if we don't find the right item, we will return the prev item and
 332  * the next item.
 333  */
 334 static struct btrfs_delayed_item *__btrfs_lookup_delayed_item(
 335                                 struct rb_root *root,
 336                                 struct btrfs_key *key,
 337                                 struct btrfs_delayed_item **prev,
 338                                 struct btrfs_delayed_item **next)
 339 {
 340         struct rb_node *node, *prev_node = NULL;
 341         struct btrfs_delayed_item *delayed_item = NULL;
 342         int ret = 0;
 343
 344         node = root->rb_node;
 345
 346         while (node) {
 347                 delayed_item = rb_entry(node, struct btrfs_delayed_item,
 348                                         rb_node);
 349                 prev_node = node;
 350                 ret = btrfs_comp_cpu_keys(&delayed_item->key, key);
 351                 if (ret < 0)
 352                         node = node->rb_right;
 353                 else if (ret > 0)
 354                         node = node->rb_left;
 355                 else
 356                         return delayed_item;
 357         }
 358
 359         if (prev) {
 360                 if (!prev_node)
 361                         *prev = NULL;
 362                 else if (ret < 0)
 363                         *prev = delayed_item;
 364                 else if ((node = rb_prev(prev_node)) != NULL) {
 365                         *prev = rb_entry(node, struct btrfs_delayed_item,
 366                                          rb_node);
 367                 } else
 368                         *prev = NULL;
 369         }
 370
 371         if (next) {
 372                 if (!prev_node)
 373                         *next = NULL;
 374                 else if (ret > 0)
 375                         *next = delayed_item;
 376                 else if ((node = rb_next(prev_node)) != NULL) {
 377                         *next = rb_entry(node, struct btrfs_delayed_item,
 378                                          rb_node);
 379                 } else
 380                         *next = NULL;
 381         }
 382         return NULL;
 383 }
 384
 385 static struct btrfs_delayed_item *__btrfs_lookup_delayed_insertion_item(
 386                                         struct btrfs_delayed_node *delayed_node,
 387                                         struct btrfs_key *key)
 388 {
 389         return __btrfs_lookup_delayed_item(&delayed_node->ins_root.rb_root, key,
 390                                            NULL, NULL);
 391 }
 392
 393 static int __btrfs_add_delayed_item(struct btrfs_delayed_node *delayed_node,
 394                                     struct btrfs_delayed_item *ins,
 395                                     int action)
 396 {
 397         struct rb_node **p, *node;
 398         struct rb_node *parent_node = NULL;
 399         struct rb_root_cached *root;
 400         struct btrfs_delayed_item *item;
 401         int cmp;
 402         bool leftmost = true;
 403
 404         if (action == BTRFS_DELAYED_INSERTION_ITEM)
 405                 root = &delayed_node->ins_root;
 406         else if (action == BTRFS_DELAYED_DELETION_ITEM)
 407                 root = &delayed_node->del_root;
 408         else
 409                 BUG();
 410         p = &root->rb_root.rb_node;
 411         node = &ins->rb_node;
 412
 413         while (*p) {
 414                 parent_node = *p;
 415                 item = rb_entry(parent_node, struct btrfs_delayed_item,
 416                                  rb_node);
 417
 418                 cmp = btrfs_comp_cpu_keys(&item->key, &ins->key);
 419                 if (cmp < 0) {
 420                         p = &(*p)->rb_right;
 421                         leftmost = false;
 422                 } else if (cmp > 0) {
 423                         p = &(*p)->rb_left;
 424                 } else {
 425                         return -EEXIST;
 426                 }
 427         }
 428
 429         rb_link_node(node, parent_node, p);
 430         rb_insert_color_cached(node, root, leftmost);
 431         ins->delayed_node = delayed_node;
 432         ins->ins_or_del = action;
 433
 434         if (ins->key.type == BTRFS_DIR_INDEX_KEY &&
 435             action == BTRFS_DELAYED_INSERTION_ITEM &&
 436             ins->key.offset >= delayed_node->index_cnt)
 437                         delayed_node->index_cnt = ins->key.offset + 1;
 438
 439         delayed_node->count++;
 440         atomic_inc(&delayed_node->root->fs_info->delayed_root->items);
 441         return 0;
 442 }
 443
 444 static int __btrfs_add_delayed_insertion_item(struct btrfs_delayed_node *node,
 445                                               struct btrfs_delayed_item *item)
 446 {
 447         return __btrfs_add_delayed_item(node, item,
 448                                         BTRFS_DELAYED_INSERTION_ITEM);
 449 }
 450
 451 static int __btrfs_add_delayed_deletion_item(struct btrfs_delayed_node *node,
 452                                              struct btrfs_delayed_item *item)
 453 {
 454         return __btrfs_add_delayed_item(node, item,
 455                                         BTRFS_DELAYED_DELETION_ITEM);
 456 }
 457
 458 static void finish_one_item(struct btrfs_delayed_root *delayed_root)
 459 {
 460         int seq = atomic_inc_return(&delayed_root->items_seq);
 461
 462         /* atomic_dec_return implies a barrier */
 463         if ((atomic_dec_return(&delayed_root->items) <
 464             BTRFS_DELAYED_BACKGROUND || seq % BTRFS_DELAYED_BATCH == 0))
 465                 cond_wake_up_nomb(&delayed_root->wait);
 466 }
 467
 468 static void __btrfs_remove_delayed_item(struct btrfs_delayed_item *delayed_item)
 469 {
 470         struct rb_root_cached *root;
 471         struct btrfs_delayed_root *delayed_root;
 472
 473         /* Not associated with any delayed_node */
 474         if (!delayed_item->delayed_node)
 475                 return;
 476         delayed_root = delayed_item->delayed_node->root->fs_info->delayed_root;
 477
 478         BUG_ON(!delayed_root);
 479         BUG_ON(delayed_item->ins_or_del != BTRFS_DELAYED_DELETION_ITEM &&
 480                delayed_item->ins_or_del != BTRFS_DELAYED_INSERTION_ITEM);
 481
 482         if (delayed_item->ins_or_del == BTRFS_DELAYED_INSERTION_ITEM)
 483                 root = &delayed_item->delayed_node->ins_root;
 484         else
 485                 root = &delayed_item->delayed_node->del_root;
 486
 487         rb_erase_cached(&delayed_item->rb_node, root);
 488         delayed_item->delayed_node->count--;
 489
 490         finish_one_item(delayed_root);
 491 }
 492
 493 static void btrfs_release_delayed_item(struct btrfs_delayed_item *item)
 494 {
 495         if (item) {
 496                 __btrfs_remove_delayed_item(item);
 497                 if (refcount_dec_and_test(&item->refs))
 498                         kfree(item);
 499         }
 500 }
 501
 502 static struct btrfs_delayed_item *__btrfs_first_delayed_insertion_item(
 503                                         struct btrfs_delayed_node *delayed_node)
 504 {
 505         struct rb_node *p;
 506         struct btrfs_delayed_item *item = NULL;
 507
 508         p = rb_first_cached(&delayed_node->ins_root);
 509         if (p)
 510                 item = rb_entry(p, struct btrfs_delayed_item, rb_node);
 511
 512         return item;
 513 }
 514
 515 static struct btrfs_delayed_item *__btrfs_first_delayed_deletion_item(
 516                                         struct btrfs_delayed_node *delayed_node)
 517 {
 518         struct rb_node *p;
 519         struct btrfs_delayed_item *item = NULL;
 520
 521         p = rb_first_cached(&delayed_node->del_root);
 522         if (p)
 523                 item = rb_entry(p, struct btrfs_delayed_item, rb_node);
 524
 525         return item;
 526 }
 527
 528 static struct btrfs_delayed_item *__btrfs_next_delayed_item(
 529                                                 struct btrfs_delayed_item *item)
 530 {
 531         struct rb_node *p;
 532         struct btrfs_delayed_item *next = NULL;
 533
 534         p = rb_next(&item->rb_node);
 535         if (p)
 536                 next = rb_entry(p, struct btrfs_delayed_item, rb_node);
 537
 538         return next;
 539 }
 540
 541 static int btrfs_delayed_item_reserve_metadata(struct btrfs_trans_handle *trans,
 542                                                struct btrfs_root *root,
 543                                                struct btrfs_delayed_item *item)
 544 {
 545         struct btrfs_block_rsv *src_rsv;
 546         struct btrfs_block_rsv *dst_rsv;
 547         struct btrfs_fs_info *fs_info = root->fs_info;
 548         u64 num_bytes;
 549         int ret;
 550
 551         if (!trans->bytes_reserved)
 552                 return 0;
 553
 554         src_rsv = trans->block_rsv;
 555         dst_rsv = &fs_info->delayed_block_rsv;
 556
 557         num_bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
 558
 559         /*
 560          * Here we migrate space rsv from transaction rsv, since have already
 561          * reserved space when starting a transaction.  So no need to reserve
 562          * qgroup space here.
 563          */
 564         ret = btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, true);
 565         if (!ret) {
 566                 trace_btrfs_space_reservation(fs_info, "delayed_item",
 567                                               item->key.objectid,
 568                                               num_bytes, 1);
 569                 item->bytes_reserved = num_bytes;
 570         }
 571
 572         return ret;
 573 }
 574
 575 static void btrfs_delayed_item_release_metadata(struct btrfs_root *root,
 576                                                 struct btrfs_delayed_item *item)
 577 {
 578         struct btrfs_block_rsv *rsv;
 579         struct btrfs_fs_info *fs_info = root->fs_info;
 580
 581         if (!item->bytes_reserved)
 582                 return;
 583
 584         rsv = &fs_info->delayed_block_rsv;
 585         /*
 586          * Check btrfs_delayed_item_reserve_metadata() to see why we don't need
 587          * to release/reserve qgroup space.
 588          */
 589         trace_btrfs_space_reservation(fs_info, "delayed_item",
 590                                       item->key.objectid, item->bytes_reserved,
 591                                       0);
 592         btrfs_block_rsv_release(fs_info, rsv, item->bytes_reserved, NULL);
 593 }
 594
 595 static int btrfs_delayed_inode_reserve_metadata(
 596                                         struct btrfs_trans_handle *trans,
 597                                         struct btrfs_root *root,
 598                                         struct btrfs_delayed_node *node)
 599 {
 600         struct btrfs_fs_info *fs_info = root->fs_info;
 601         struct btrfs_block_rsv *src_rsv;
 602         struct btrfs_block_rsv *dst_rsv;
 603         u64 num_bytes;
 604         int ret;
 605
 606         src_rsv = trans->block_rsv;
 607         dst_rsv = &fs_info->delayed_block_rsv;
 608
 609         num_bytes = btrfs_calc_metadata_size(fs_info, 1);
 610
 611         /*
 612          * btrfs_dirty_inode will update the inode under btrfs_join_transaction
 613          * which doesn't reserve space for speed.  This is a problem since we
 614          * still need to reserve space for this update, so try to reserve the
 615          * space.
 616          *
 617          * Now if src_rsv == delalloc_block_rsv we'll let it just steal since
 618          * we always reserve enough to update the inode item.
 619          */
 620         if (!src_rsv || (!trans->bytes_reserved &&
 621                          src_rsv->type != BTRFS_BLOCK_RSV_DELALLOC)) {
 622                 ret = btrfs_qgroup_reserve_meta(root, num_bytes,
 623                                           BTRFS_QGROUP_RSV_META_PREALLOC, true);
 624                 if (ret < 0)
 625                         return ret;
 626                 ret = btrfs_block_rsv_add(fs_info, dst_rsv, num_bytes,
 627                                           BTRFS_RESERVE_NO_FLUSH);
 628                 /* NO_FLUSH could only fail with -ENOSPC */
 629                 ASSERT(ret == 0 || ret == -ENOSPC);
 630                 if (ret)
 631                         btrfs_qgroup_free_meta_prealloc(root, num_bytes);
 632         } else {
 633                 ret = btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, true);
 634         }
 635
 636         if (!ret) {
 637                 trace_btrfs_space_reservation(fs_info, "delayed_inode",
 638                                               node->inode_id, num_bytes, 1);
 639                 node->bytes_reserved = num_bytes;
 640         }
 641
 642         return ret;
 643 }
 644
 645 static void btrfs_delayed_inode_release_metadata(struct btrfs_fs_info *fs_info,
 646                                                 struct btrfs_delayed_node *node,
 647                                                 bool qgroup_free)
 648 {
 649         struct btrfs_block_rsv *rsv;
 650
 651         if (!node->bytes_reserved)
 652                 return;
 653
 654         rsv = &fs_info->delayed_block_rsv;
 655         trace_btrfs_space_reservation(fs_info, "delayed_inode",
 656                                       node->inode_id, node->bytes_reserved, 0);
 657         btrfs_block_rsv_release(fs_info, rsv, node->bytes_reserved, NULL);
 658         if (qgroup_free)
 659                 btrfs_qgroup_free_meta_prealloc(node->root,
 660                                 node->bytes_reserved);
 661         else
 662                 btrfs_qgroup_convert_reserved_meta(node->root,
 663                                 node->bytes_reserved);
 664         node->bytes_reserved = 0;
 665 }
 666
 667 /*
 668  * Insert a single delayed item or a batch of delayed items that have consecutive
 669  * keys if they exist.
 670  */
 671 static int btrfs_insert_delayed_item(struct btrfs_trans_handle *trans,
 672                                      struct btrfs_root *root,
 673                                      struct btrfs_path *path,
 674                                      struct btrfs_delayed_item *first_item)
 675 {
 676         LIST_HEAD(item_list);
 677         struct btrfs_delayed_item *curr;
 678         struct btrfs_delayed_item *next;
 679         const int max_size = BTRFS_LEAF_DATA_SIZE(root->fs_info);
 680         struct btrfs_item_batch batch;
 681         int total_size;
 682         char *ins_data = NULL;
 683         int ret;
 684
 685         list_add_tail(&first_item->tree_list, &item_list);
 686         batch.total_data_size = first_item->data_len;
 687         batch.nr = 1;
 688         total_size = first_item->data_len + sizeof(struct btrfs_item);
 689         curr = first_item;
 690
 691         while (true) {
 692                 int next_size;
 693
 694                 next = __btrfs_next_delayed_item(curr);
 695                 if (!next || !btrfs_is_continuous_delayed_item(curr, next))
 696                         break;
 697
 698                 next_size = next->data_len + sizeof(struct btrfs_item);
 699                 if (total_size + next_size > max_size)
 700                         break;
 701
 702                 list_add_tail(&next->tree_list, &item_list);
 703                 batch.nr++;
 704                 total_size += next_size;
 705                 batch.total_data_size += next->data_len;
 706                 curr = next;
 707         }
 708
 709         if (batch.nr == 1) {
 710                 batch.keys = &first_item->key;
 711                 batch.data_sizes = &first_item->data_len;
 712         } else {
 713                 struct btrfs_key *ins_keys;
 714                 u32 *ins_sizes;
 715                 int i = 0;
 716
 717                 ins_data = kmalloc(batch.nr * sizeof(u32) +
 718                                    batch.nr * sizeof(struct btrfs_key), GFP_NOFS);
 719                 if (!ins_data) {
 720                         ret = -ENOMEM;
 721                         goto out;
 722                 }
 723                 ins_sizes = (u32 *)ins_data;
 724                 ins_keys = (struct btrfs_key *)(ins_data + batch.nr * sizeof(u32));
 725                 batch.keys = ins_keys;
 726                 batch.data_sizes = ins_sizes;
 727                 list_for_each_entry(curr, &item_list, tree_list) {
 728                         ins_keys[i] = curr->key;
 729                         ins_sizes[i] = curr->data_len;
 730                         i++;
 731                 }
 732         }
 733
 734         ret = btrfs_insert_empty_items(trans, root, path, &batch);
 735         if (ret)
 736                 goto out;
 737
 738         list_for_each_entry(curr, &item_list, tree_list) {
 739                 char *data_ptr;
 740
 741                 data_ptr = btrfs_item_ptr(path->nodes[0], path->slots[0], char);
 742                 write_extent_buffer(path->nodes[0], &curr->data,
 743                                     (unsigned long)data_ptr, curr->data_len);
 744                 path->slots[0]++;
 745         }
 746
 747         /*
 748          * Now release our path before releasing the delayed items and their
 749          * metadata reservations, so that we don't block other tasks for more
 750          * time than needed.
 751          */
 752         btrfs_release_path(path);
 753
 754         list_for_each_entry_safe(curr, next, &item_list, tree_list) {
 755                 list_del(&curr->tree_list);
 756                 btrfs_delayed_item_release_metadata(root, curr);
 757                 btrfs_release_delayed_item(curr);
 758         }
 759 out:
 760         kfree(ins_data);
 761         return ret;
 762 }
 763
 764 static int btrfs_insert_delayed_items(struct btrfs_trans_handle *trans,
 765                                       struct btrfs_path *path,
 766                                       struct btrfs_root *root,
 767                                       struct btrfs_delayed_node *node)
 768 {
 769         int ret = 0;
 770
 771         while (ret == 0) {
 772                 struct btrfs_delayed_item *curr;
 773
 774                 mutex_lock(&node->mutex);
 775                 curr = __btrfs_first_delayed_insertion_item(node);
 776                 if (!curr) {
 777                         mutex_unlock(&node->mutex);
 778                         break;
 779                 }
 780                 ret = btrfs_insert_delayed_item(trans, root, path, curr);
 781                 mutex_unlock(&node->mutex);
 782         }
 783
 784         return ret;
 785 }
 786
 787 static int btrfs_batch_delete_items(struct btrfs_trans_handle *trans,
 788                                     struct btrfs_root *root,
 789                                     struct btrfs_path *path,
 790                                     struct btrfs_delayed_item *item)
 791 {
 792         struct btrfs_delayed_item *curr, *next;
 793         struct extent_buffer *leaf;
 794         struct btrfs_key key;
 795         struct list_head head;
 796         int nitems, i, last_item;
 797         int ret = 0;
 798
 799         BUG_ON(!path->nodes[0]);
 800
 801         leaf = path->nodes[0];
 802
 803         i = path->slots[0];
 804         last_item = btrfs_header_nritems(leaf) - 1;
 805         if (i > last_item)
 806                 return -ENOENT; /* FIXME: Is errno suitable? */
 807
 808         next = item;
 809         INIT_LIST_HEAD(&head);
 810         btrfs_item_key_to_cpu(leaf, &key, i);
 811         nitems = 0;
 812         /*
 813          * count the number of the dir index items that we can delete in batch
 814          */
 815         while (btrfs_comp_cpu_keys(&next->key, &key) == 0) {
 816                 list_add_tail(&next->tree_list, &head);
 817                 nitems++;
 818
 819                 curr = next;
 820                 next = __btrfs_next_delayed_item(curr);
 821                 if (!next)
 822                         break;
 823
 824                 if (!btrfs_is_continuous_delayed_item(curr, next))
 825                         break;
 826
 827                 i++;
 828                 if (i > last_item)
 829                         break;
 830                 btrfs_item_key_to_cpu(leaf, &key, i);
 831         }
 832
 833         if (!nitems)
 834                 return 0;
 835
 836         ret = btrfs_del_items(trans, root, path, path->slots[0], nitems);
 837         if (ret)
 838                 goto out;
 839
 840         list_for_each_entry_safe(curr, next, &head, tree_list) {
 841                 btrfs_delayed_item_release_metadata(root, curr);
 842                 list_del(&curr->tree_list);
 843                 btrfs_release_delayed_item(curr);
 844         }
 845
 846 out:
 847         return ret;
 848 }
 849
 850 static int btrfs_delete_delayed_items(struct btrfs_trans_handle *trans,
 851                                       struct btrfs_path *path,
 852                                       struct btrfs_root *root,
 853                                       struct btrfs_delayed_node *node)
 854 {
 855         struct btrfs_delayed_item *curr, *prev;
 856         int ret = 0;
 857
 858 do_again:
 859         mutex_lock(&node->mutex);
 860         curr = __btrfs_first_delayed_deletion_item(node);
 861         if (!curr)
 862                 goto delete_fail;
 863
 864         ret = btrfs_search_slot(trans, root, &curr->key, path, -1, 1);
 865         if (ret < 0)
 866                 goto delete_fail;
 867         else if (ret > 0) {
 868                 /*
 869                  * can't find the item which the node points to, so this node
 870                  * is invalid, just drop it.
 871                  */
 872                 prev = curr;
 873                 curr = __btrfs_next_delayed_item(prev);
 874                 btrfs_release_delayed_item(prev);
 875                 ret = 0;
 876                 btrfs_release_path(path);
 877                 if (curr) {
 878                         mutex_unlock(&node->mutex);
 879                         goto do_again;
 880                 } else
 881                         goto delete_fail;
 882         }
 883
 884         btrfs_batch_delete_items(trans, root, path, curr);
 885         btrfs_release_path(path);
 886         mutex_unlock(&node->mutex);
 887         goto do_again;
 888
 889 delete_fail:
 890         btrfs_release_path(path);
 891         mutex_unlock(&node->mutex);
 892         return ret;
 893 }
 894
 895 static void btrfs_release_delayed_inode(struct btrfs_delayed_node *delayed_node)
 896 {
 897         struct btrfs_delayed_root *delayed_root;
 898
 899         if (delayed_node &&
 900             test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
 901                 BUG_ON(!delayed_node->root);
 902                 clear_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags);
 903                 delayed_node->count--;
 904
 905                 delayed_root = delayed_node->root->fs_info->delayed_root;
 906                 finish_one_item(delayed_root);
 907         }
 908 }
 909
 910 static void btrfs_release_delayed_iref(struct btrfs_delayed_node *delayed_node)
 911 {
 912
 913         if (test_and_clear_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags)) {
 914                 struct btrfs_delayed_root *delayed_root;
 915
 916                 ASSERT(delayed_node->root);
 917                 delayed_node->count--;
 918
 919                 delayed_root = delayed_node->root->fs_info->delayed_root;
 920                 finish_one_item(delayed_root);
 921         }
 922 }
 923
 924 static int __btrfs_update_delayed_inode(struct btrfs_trans_handle *trans,
 925                                         struct btrfs_root *root,
 926                                         struct btrfs_path *path,
 927                                         struct btrfs_delayed_node *node)
 928 {
 929         struct btrfs_fs_info *fs_info = root->fs_info;
 930         struct btrfs_key key;
 931         struct btrfs_inode_item *inode_item;
 932         struct extent_buffer *leaf;
 933         int mod;
 934         int ret;
 935
 936         key.objectid = node->inode_id;
 937         key.type = BTRFS_INODE_ITEM_KEY;
 938         key.offset = 0;
 939
 940         if (test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &node->flags))
 941                 mod = -1;
 942         else
 943                 mod = 1;
 944
 945         ret = btrfs_lookup_inode(trans, root, path, &key, mod);
 946         if (ret > 0)
 947                 ret = -ENOENT;
 948         if (ret < 0)
 949                 goto out;
 950
 951         leaf = path->nodes[0];
 952         inode_item = btrfs_item_ptr(leaf, path->slots[0],
 953                                     struct btrfs_inode_item);
 954         write_extent_buffer(leaf, &node->inode_item, (unsigned long)inode_item,
 955                             sizeof(struct btrfs_inode_item));
 956         btrfs_mark_buffer_dirty(leaf);
 957
 958         if (!test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &node->flags))
 959                 goto out;
 960
 961         path->slots[0]++;
 962         if (path->slots[0] >= btrfs_header_nritems(leaf))
 963                 goto search;
 964 again:
 965         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 966         if (key.objectid != node->inode_id)
 967                 goto out;
 968
 969         if (key.type != BTRFS_INODE_REF_KEY &&
 970             key.type != BTRFS_INODE_EXTREF_KEY)
 971                 goto out;
 972
 973         /*
 974          * Delayed iref deletion is for the inode who has only one link,
 975          * so there is only one iref. The case that several irefs are
 976          * in the same item doesn't exist.
 977          */
 978         btrfs_del_item(trans, root, path);
 979 out:
 980         btrfs_release_delayed_iref(node);
 981         btrfs_release_path(path);
 982 err_out:
 983         btrfs_delayed_inode_release_metadata(fs_info, node, (ret < 0));
 984         btrfs_release_delayed_inode(node);
 985
 986         /*
 987          * If we fail to update the delayed inode we need to abort the
 988          * transaction, because we could leave the inode with the improper
 989          * counts behind.
 990          */
 991         if (ret && ret != -ENOENT)
 992                 btrfs_abort_transaction(trans, ret);
 993
 994         return ret;
 995
 996 search:
 997         btrfs_release_path(path);
 998
 999         key.type = BTRFS_INODE_EXTREF_KEY;
1000         key.offset = -1;
1001
1002         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1003         if (ret < 0)
1004                 goto err_out;
1005         ASSERT(ret);
1006
1007         ret = 0;
1008         leaf = path->nodes[0];
1009         path->slots[0]--;
1010         goto again;
1011 }
1012
1013 static inline int btrfs_update_delayed_inode(struct btrfs_trans_handle *trans,
1014                                              struct btrfs_root *root,
1015                                              struct btrfs_path *path,
1016                                              struct btrfs_delayed_node *node)
1017 {
1018         int ret;
1019
1020         mutex_lock(&node->mutex);
1021         if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &node->flags)) {
1022                 mutex_unlock(&node->mutex);
1023                 return 0;
1024         }
1025
1026         ret = __btrfs_update_delayed_inode(trans, root, path, node);
1027         mutex_unlock(&node->mutex);
1028         return ret;
1029 }
1030
1031 static inline int
1032 __btrfs_commit_inode_delayed_items(struct btrfs_trans_handle *trans,
1033                                    struct btrfs_path *path,
1034                                    struct btrfs_delayed_node *node)
1035 {
1036         int ret;
1037
1038         ret = btrfs_insert_delayed_items(trans, path, node->root, node);
1039         if (ret)
1040                 return ret;
1041
1042         ret = btrfs_delete_delayed_items(trans, path, node->root, node);
1043         if (ret)
1044                 return ret;
1045
1046         ret = btrfs_update_delayed_inode(trans, node->root, path, node);
1047         return ret;
1048 }
1049
1050 /*
1051  * Called when committing the transaction.
1052  * Returns 0 on success.
1053  * Returns < 0 on error and returns with an aborted transaction with any
1054  * outstanding delayed items cleaned up.
1055  */
1056 static int __btrfs_run_delayed_items(struct btrfs_trans_handle *trans, int nr)
1057 {
1058         struct btrfs_fs_info *fs_info = trans->fs_info;
1059         struct btrfs_delayed_root *delayed_root;
1060         struct btrfs_delayed_node *curr_node, *prev_node;
1061         struct btrfs_path *path;
1062         struct btrfs_block_rsv *block_rsv;
1063         int ret = 0;
1064         bool count = (nr > 0);
1065
1066         if (TRANS_ABORTED(trans))
1067                 return -EIO;
1068
1069         path = btrfs_alloc_path();
1070         if (!path)
1071                 return -ENOMEM;
1072
1073         block_rsv = trans->block_rsv;
1074         trans->block_rsv = &fs_info->delayed_block_rsv;
1075
1076         delayed_root = fs_info->delayed_root;
1077
1078         curr_node = btrfs_first_delayed_node(delayed_root);
1079         while (curr_node && (!count || nr--)) {
1080                 ret = __btrfs_commit_inode_delayed_items(trans, path,
1081                                                          curr_node);
1082                 if (ret) {
1083                         btrfs_release_delayed_node(curr_node);
1084                         curr_node = NULL;
1085                         btrfs_abort_transaction(trans, ret);
1086                         break;
1087                 }
1088
1089                 prev_node = curr_node;
1090                 curr_node = btrfs_next_delayed_node(curr_node);
1091                 btrfs_release_delayed_node(prev_node);
1092         }
1093
1094         if (curr_node)
1095                 btrfs_release_delayed_node(curr_node);
1096         btrfs_free_path(path);
1097         trans->block_rsv = block_rsv;
1098
1099         return ret;
1100 }
1101
1102 int btrfs_run_delayed_items(struct btrfs_trans_handle *trans)
1103 {
1104         return __btrfs_run_delayed_items(trans, -1);
1105 }
1106
1107 int btrfs_run_delayed_items_nr(struct btrfs_trans_handle *trans, int nr)
1108 {
1109         return __btrfs_run_delayed_items(trans, nr);
1110 }
1111
1112 int btrfs_commit_inode_delayed_items(struct btrfs_trans_handle *trans,
1113                                      struct btrfs_inode *inode)
1114 {
1115         struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1116         struct btrfs_path *path;
1117         struct btrfs_block_rsv *block_rsv;
1118         int ret;
1119
1120         if (!delayed_node)
1121                 return 0;
1122
1123         mutex_lock(&delayed_node->mutex);
1124         if (!delayed_node->count) {
1125                 mutex_unlock(&delayed_node->mutex);
1126                 btrfs_release_delayed_node(delayed_node);
1127                 return 0;
1128         }
1129         mutex_unlock(&delayed_node->mutex);
1130
1131         path = btrfs_alloc_path();
1132         if (!path) {
1133                 btrfs_release_delayed_node(delayed_node);
1134                 return -ENOMEM;
1135         }
1136
1137         block_rsv = trans->block_rsv;
1138         trans->block_rsv = &delayed_node->root->fs_info->delayed_block_rsv;
1139
1140         ret = __btrfs_commit_inode_delayed_items(trans, path, delayed_node);
1141
1142         btrfs_release_delayed_node(delayed_node);
1143         btrfs_free_path(path);
1144         trans->block_rsv = block_rsv;
1145
1146         return ret;
1147 }
1148
1149 int btrfs_commit_inode_delayed_inode(struct btrfs_inode *inode)
1150 {
1151         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1152         struct btrfs_trans_handle *trans;
1153         struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1154         struct btrfs_path *path;
1155         struct btrfs_block_rsv *block_rsv;
1156         int ret;
1157
1158         if (!delayed_node)
1159                 return 0;
1160
1161         mutex_lock(&delayed_node->mutex);
1162         if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1163                 mutex_unlock(&delayed_node->mutex);
1164                 btrfs_release_delayed_node(delayed_node);
1165                 return 0;
1166         }
1167         mutex_unlock(&delayed_node->mutex);
1168
1169         trans = btrfs_join_transaction(delayed_node->root);
1170         if (IS_ERR(trans)) {
1171                 ret = PTR_ERR(trans);
1172                 goto out;
1173         }
1174
1175         path = btrfs_alloc_path();
1176         if (!path) {
1177                 ret = -ENOMEM;
1178                 goto trans_out;
1179         }
1180
1181         block_rsv = trans->block_rsv;
1182         trans->block_rsv = &fs_info->delayed_block_rsv;
1183
1184         mutex_lock(&delayed_node->mutex);
1185         if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags))
1186                 ret = __btrfs_update_delayed_inode(trans, delayed_node->root,
1187                                                    path, delayed_node);
1188         else
1189                 ret = 0;
1190         mutex_unlock(&delayed_node->mutex);
1191
1192         btrfs_free_path(path);
1193         trans->block_rsv = block_rsv;
1194 trans_out:
1195         btrfs_end_transaction(trans);
1196         btrfs_btree_balance_dirty(fs_info);
1197 out:
1198         btrfs_release_delayed_node(delayed_node);
1199
1200         return ret;
1201 }
1202
1203 void btrfs_remove_delayed_node(struct btrfs_inode *inode)
1204 {
1205         struct btrfs_delayed_node *delayed_node;
1206
1207         delayed_node = READ_ONCE(inode->delayed_node);
1208         if (!delayed_node)
1209                 return;
1210
1211         inode->delayed_node = NULL;
1212         btrfs_release_delayed_node(delayed_node);
1213 }
1214
1215 struct btrfs_async_delayed_work {
1216         struct btrfs_delayed_root *delayed_root;
1217         int nr;
1218         struct btrfs_work work;
1219 };
1220
1221 static void btrfs_async_run_delayed_root(struct btrfs_work *work)
1222 {
1223         struct btrfs_async_delayed_work *async_work;
1224         struct btrfs_delayed_root *delayed_root;
1225         struct btrfs_trans_handle *trans;
1226         struct btrfs_path *path;
1227         struct btrfs_delayed_node *delayed_node = NULL;
1228         struct btrfs_root *root;
1229         struct btrfs_block_rsv *block_rsv;
1230         int total_done = 0;
1231
1232         async_work = container_of(work, struct btrfs_async_delayed_work, work);
1233         delayed_root = async_work->delayed_root;
1234
1235         path = btrfs_alloc_path();
1236         if (!path)
1237                 goto out;
1238
1239         do {
1240                 if (atomic_read(&delayed_root->items) <
1241                     BTRFS_DELAYED_BACKGROUND / 2)
1242                         break;
1243
1244                 delayed_node = btrfs_first_prepared_delayed_node(delayed_root);
1245                 if (!delayed_node)
1246                         break;
1247
1248                 root = delayed_node->root;
1249
1250                 trans = btrfs_join_transaction(root);
1251                 if (IS_ERR(trans)) {
1252                         btrfs_release_path(path);
1253                         btrfs_release_prepared_delayed_node(delayed_node);
1254                         total_done++;
1255                         continue;
1256                 }
1257
1258                 block_rsv = trans->block_rsv;
1259                 trans->block_rsv = &root->fs_info->delayed_block_rsv;
1260
1261                 __btrfs_commit_inode_delayed_items(trans, path, delayed_node);
1262
1263                 trans->block_rsv = block_rsv;
1264                 btrfs_end_transaction(trans);
1265                 btrfs_btree_balance_dirty_nodelay(root->fs_info);
1266
1267                 btrfs_release_path(path);
1268                 btrfs_release_prepared_delayed_node(delayed_node);
1269                 total_done++;
1270
1271         } while ((async_work->nr == 0 && total_done < BTRFS_DELAYED_WRITEBACK)
1272                  || total_done < async_work->nr);
1273
1274         btrfs_free_path(path);
1275 out:
1276         wake_up(&delayed_root->wait);
1277         kfree(async_work);
1278 }
1279
1280
1281 static int btrfs_wq_run_delayed_node(struct btrfs_delayed_root *delayed_root,
1282                                      struct btrfs_fs_info *fs_info, int nr)
1283 {
1284         struct btrfs_async_delayed_work *async_work;
1285
1286         async_work = kmalloc(sizeof(*async_work), GFP_NOFS);
1287         if (!async_work)
1288                 return -ENOMEM;
1289
1290         async_work->delayed_root = delayed_root;
1291         btrfs_init_work(&async_work->work, btrfs_async_run_delayed_root, NULL,
1292                         NULL);
1293         async_work->nr = nr;
1294
1295         btrfs_queue_work(fs_info->delayed_workers, &async_work->work);
1296         return 0;
1297 }
1298
1299 void btrfs_assert_delayed_root_empty(struct btrfs_fs_info *fs_info)
1300 {
1301         WARN_ON(btrfs_first_delayed_node(fs_info->delayed_root));
1302 }
1303
1304 static int could_end_wait(struct btrfs_delayed_root *delayed_root, int seq)
1305 {
1306         int val = atomic_read(&delayed_root->items_seq);
1307
1308         if (val < seq || val >= seq + BTRFS_DELAYED_BATCH)
1309                 return 1;
1310
1311         if (atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND)
1312                 return 1;
1313
1314         return 0;
1315 }
1316
1317 void btrfs_balance_delayed_items(struct btrfs_fs_info *fs_info)
1318 {
1319         struct btrfs_delayed_root *delayed_root = fs_info->delayed_root;
1320
1321         if ((atomic_read(&delayed_root->items) < BTRFS_DELAYED_BACKGROUND) ||
1322                 btrfs_workqueue_normal_congested(fs_info->delayed_workers))
1323                 return;
1324
1325         if (atomic_read(&delayed_root->items) >= BTRFS_DELAYED_WRITEBACK) {
1326                 int seq;
1327                 int ret;
1328
1329                 seq = atomic_read(&delayed_root->items_seq);
1330
1331                 ret = btrfs_wq_run_delayed_node(delayed_root, fs_info, 0);
1332                 if (ret)
1333                         return;
1334
1335                 wait_event_interruptible(delayed_root->wait,
1336                                          could_end_wait(delayed_root, seq));
1337                 return;
1338         }
1339
1340         btrfs_wq_run_delayed_node(delayed_root, fs_info, BTRFS_DELAYED_BATCH);
1341 }
1342
1343 /* Will return 0 or -ENOMEM */
1344 int btrfs_insert_delayed_dir_index(struct btrfs_trans_handle *trans,
1345                                    const char *name, int name_len,
1346                                    struct btrfs_inode *dir,
1347                                    struct btrfs_disk_key *disk_key, u8 type,
1348                                    u64 index)
1349 {
1350         struct btrfs_delayed_node *delayed_node;
1351         struct btrfs_delayed_item *delayed_item;
1352         struct btrfs_dir_item *dir_item;
1353         int ret;
1354
1355         delayed_node = btrfs_get_or_create_delayed_node(dir);
1356         if (IS_ERR(delayed_node))
1357                 return PTR_ERR(delayed_node);
1358
1359         delayed_item = btrfs_alloc_delayed_item(sizeof(*dir_item) + name_len);
1360         if (!delayed_item) {
1361                 ret = -ENOMEM;
1362                 goto release_node;
1363         }
1364
1365         delayed_item->key.objectid = btrfs_ino(dir);
1366         delayed_item->key.type = BTRFS_DIR_INDEX_KEY;
1367         delayed_item->key.offset = index;
1368
1369         dir_item = (struct btrfs_dir_item *)delayed_item->data;
1370         dir_item->location = *disk_key;
1371         btrfs_set_stack_dir_transid(dir_item, trans->transid);
1372         btrfs_set_stack_dir_data_len(dir_item, 0);
1373         btrfs_set_stack_dir_name_len(dir_item, name_len);
1374         btrfs_set_stack_dir_type(dir_item, type);
1375         memcpy((char *)(dir_item + 1), name, name_len);
1376
1377         ret = btrfs_delayed_item_reserve_metadata(trans, dir->root, delayed_item);
1378         /*
1379          * we have reserved enough space when we start a new transaction,
1380          * so reserving metadata failure is impossible
1381          */
1382         BUG_ON(ret);
1383
1384         mutex_lock(&delayed_node->mutex);
1385         ret = __btrfs_add_delayed_insertion_item(delayed_node, delayed_item);
1386         if (unlikely(ret)) {
1387                 btrfs_err(trans->fs_info,
1388                           "err add delayed dir index item(name: %.*s) into the insertion tree of the delayed node(root id: %llu, inode id: %llu, errno: %d)",
1389                           name_len, name, delayed_node->root->root_key.objectid,
1390                           delayed_node->inode_id, ret);
1391                 BUG();
1392         }
1393         mutex_unlock(&delayed_node->mutex);
1394
1395 release_node:
1396         btrfs_release_delayed_node(delayed_node);
1397         return ret;
1398 }
1399
1400 static int btrfs_delete_delayed_insertion_item(struct btrfs_fs_info *fs_info,
1401                                                struct btrfs_delayed_node *node,
1402                                                struct btrfs_key *key)
1403 {
1404         struct btrfs_delayed_item *item;
1405
1406         mutex_lock(&node->mutex);
1407         item = __btrfs_lookup_delayed_insertion_item(node, key);
1408         if (!item) {
1409                 mutex_unlock(&node->mutex);
1410                 return 1;
1411         }
1412
1413         btrfs_delayed_item_release_metadata(node->root, item);
1414         btrfs_release_delayed_item(item);
1415         mutex_unlock(&node->mutex);
1416         return 0;
1417 }
1418
1419 int btrfs_delete_delayed_dir_index(struct btrfs_trans_handle *trans,
1420                                    struct btrfs_inode *dir, u64 index)
1421 {
1422         struct btrfs_delayed_node *node;
1423         struct btrfs_delayed_item *item;
1424         struct btrfs_key item_key;
1425         int ret;
1426
1427         node = btrfs_get_or_create_delayed_node(dir);
1428         if (IS_ERR(node))
1429                 return PTR_ERR(node);
1430
1431         item_key.objectid = btrfs_ino(dir);
1432         item_key.type = BTRFS_DIR_INDEX_KEY;
1433         item_key.offset = index;
1434
1435         ret = btrfs_delete_delayed_insertion_item(trans->fs_info, node,
1436                                                   &item_key);
1437         if (!ret)
1438                 goto end;
1439
1440         item = btrfs_alloc_delayed_item(0);
1441         if (!item) {
1442                 ret = -ENOMEM;
1443                 goto end;
1444         }
1445
1446         item->key = item_key;
1447
1448         ret = btrfs_delayed_item_reserve_metadata(trans, dir->root, item);
1449         /*
1450          * we have reserved enough space when we start a new transaction,
1451          * so reserving metadata failure is impossible.
1452          */
1453         if (ret < 0) {
1454                 btrfs_err(trans->fs_info,
1455 "metadata reservation failed for delayed dir item deltiona, should have been reserved");
1456                 btrfs_release_delayed_item(item);
1457                 goto end;
1458         }
1459
1460         mutex_lock(&node->mutex);
1461         ret = __btrfs_add_delayed_deletion_item(node, item);
1462         if (unlikely(ret)) {
1463                 btrfs_err(trans->fs_info,
1464                           "err add delayed dir index item(index: %llu) into the deletion tree of the delayed node(root id: %llu, inode id: %llu, errno: %d)",
1465                           index, node->root->root_key.objectid,
1466                           node->inode_id, ret);
1467                 btrfs_delayed_item_release_metadata(dir->root, item);
1468                 btrfs_release_delayed_item(item);
1469         }
1470         mutex_unlock(&node->mutex);
1471 end:
1472         btrfs_release_delayed_node(node);
1473         return ret;
1474 }
1475
1476 int btrfs_inode_delayed_dir_index_count(struct btrfs_inode *inode)
1477 {
1478         struct btrfs_delayed_node *delayed_node = btrfs_get_delayed_node(inode);
1479
1480         if (!delayed_node)
1481                 return -ENOENT;
1482
1483         /*
1484          * Since we have held i_mutex of this directory, it is impossible that
1485          * a new directory index is added into the delayed node and index_cnt
1486          * is updated now. So we needn't lock the delayed node.
1487          */
1488         if (!delayed_node->index_cnt) {
1489                 btrfs_release_delayed_node(delayed_node);
1490                 return -EINVAL;
1491         }
1492
1493         inode->index_cnt = delayed_node->index_cnt;
1494         btrfs_release_delayed_node(delayed_node);
1495         return 0;
1496 }
1497
1498 bool btrfs_readdir_get_delayed_items(struct inode *inode,
1499                                      struct list_head *ins_list,
1500                                      struct list_head *del_list)
1501 {
1502         struct btrfs_delayed_node *delayed_node;
1503         struct btrfs_delayed_item *item;
1504
1505         delayed_node = btrfs_get_delayed_node(BTRFS_I(inode));
1506         if (!delayed_node)
1507                 return false;
1508
1509         /*
1510          * We can only do one readdir with delayed items at a time because of
1511          * item->readdir_list.
1512          */
1513         btrfs_inode_unlock(inode, BTRFS_ILOCK_SHARED);
1514         btrfs_inode_lock(inode, 0);
1515
1516         mutex_lock(&delayed_node->mutex);
1517         item = __btrfs_first_delayed_insertion_item(delayed_node);
1518         while (item) {
1519                 refcount_inc(&item->refs);
1520                 list_add_tail(&item->readdir_list, ins_list);
1521                 item = __btrfs_next_delayed_item(item);
1522         }
1523
1524         item = __btrfs_first_delayed_deletion_item(delayed_node);
1525         while (item) {
1526                 refcount_inc(&item->refs);
1527                 list_add_tail(&item->readdir_list, del_list);
1528                 item = __btrfs_next_delayed_item(item);
1529         }
1530         mutex_unlock(&delayed_node->mutex);
1531         /*
1532          * This delayed node is still cached in the btrfs inode, so refs
1533          * must be > 1 now, and we needn't check it is going to be freed
1534          * or not.
1535          *
1536          * Besides that, this function is used to read dir, we do not
1537          * insert/delete delayed items in this period. So we also needn't
1538          * requeue or dequeue this delayed node.
1539          */
1540         refcount_dec(&delayed_node->refs);
1541
1542         return true;
1543 }
1544
1545 void btrfs_readdir_put_delayed_items(struct inode *inode,
1546                                      struct list_head *ins_list,
1547                                      struct list_head *del_list)
1548 {
1549         struct btrfs_delayed_item *curr, *next;
1550
1551         list_for_each_entry_safe(curr, next, ins_list, readdir_list) {
1552                 list_del(&curr->readdir_list);
1553                 if (refcount_dec_and_test(&curr->refs))
1554                         kfree(curr);
1555         }
1556
1557         list_for_each_entry_safe(curr, next, del_list, readdir_list) {
1558                 list_del(&curr->readdir_list);
1559                 if (refcount_dec_and_test(&curr->refs))
1560                         kfree(curr);
1561         }
1562
1563         /*
1564          * The VFS is going to do up_read(), so we need to downgrade back to a
1565          * read lock.
1566          */
1567         downgrade_write(&inode->i_rwsem);
1568 }
1569
1570 int btrfs_should_delete_dir_index(struct list_head *del_list,
1571                                   u64 index)
1572 {
1573         struct btrfs_delayed_item *curr;
1574         int ret = 0;
1575
1576         list_for_each_entry(curr, del_list, readdir_list) {
1577                 if (curr->key.offset > index)
1578                         break;
1579                 if (curr->key.offset == index) {
1580                         ret = 1;
1581                         break;
1582                 }
1583         }
1584         return ret;
1585 }
1586
1587 /*
1588  * btrfs_readdir_delayed_dir_index - read dir info stored in the delayed tree
1589  *
1590  */
1591 int btrfs_readdir_delayed_dir_index(struct dir_context *ctx,
1592                                     struct list_head *ins_list)
1593 {
1594         struct btrfs_dir_item *di;
1595         struct btrfs_delayed_item *curr, *next;
1596         struct btrfs_key location;
1597         char *name;
1598         int name_len;
1599         int over = 0;
1600         unsigned char d_type;
1601
1602         if (list_empty(ins_list))
1603                 return 0;
1604
1605         /*
1606          * Changing the data of the delayed item is impossible. So
1607          * we needn't lock them. And we have held i_mutex of the
1608          * directory, nobody can delete any directory indexes now.
1609          */
1610         list_for_each_entry_safe(curr, next, ins_list, readdir_list) {
1611                 list_del(&curr->readdir_list);
1612
1613                 if (curr->key.offset < ctx->pos) {
1614                         if (refcount_dec_and_test(&curr->refs))
1615                                 kfree(curr);
1616                         continue;
1617                 }
1618
1619                 ctx->pos = curr->key.offset;
1620
1621                 di = (struct btrfs_dir_item *)curr->data;
1622                 name = (char *)(di + 1);
1623                 name_len = btrfs_stack_dir_name_len(di);
1624
1625                 d_type = fs_ftype_to_dtype(di->type);
1626                 btrfs_disk_key_to_cpu(&location, &di->location);
1627
1628                 over = !dir_emit(ctx, name, name_len,
1629                                location.objectid, d_type);
1630
1631                 if (refcount_dec_and_test(&curr->refs))
1632                         kfree(curr);
1633
1634                 if (over)
1635                         return 1;
1636                 ctx->pos++;
1637         }
1638         return 0;
1639 }
1640
1641 static void fill_stack_inode_item(struct btrfs_trans_handle *trans,
1642                                   struct btrfs_inode_item *inode_item,
1643                                   struct inode *inode)
1644 {
1645         u64 flags;
1646
1647         btrfs_set_stack_inode_uid(inode_item, i_uid_read(inode));
1648         btrfs_set_stack_inode_gid(inode_item, i_gid_read(inode));
1649         btrfs_set_stack_inode_size(inode_item, BTRFS_I(inode)->disk_i_size);
1650         btrfs_set_stack_inode_mode(inode_item, inode->i_mode);
1651         btrfs_set_stack_inode_nlink(inode_item, inode->i_nlink);
1652         btrfs_set_stack_inode_nbytes(inode_item, inode_get_bytes(inode));
1653         btrfs_set_stack_inode_generation(inode_item,
1654                                          BTRFS_I(inode)->generation);
1655         btrfs_set_stack_inode_sequence(inode_item,
1656                                        inode_peek_iversion(inode));
1657         btrfs_set_stack_inode_transid(inode_item, trans->transid);
1658         btrfs_set_stack_inode_rdev(inode_item, inode->i_rdev);
1659         flags = btrfs_inode_combine_flags(BTRFS_I(inode)->flags,
1660                                           BTRFS_I(inode)->ro_flags);
1661         btrfs_set_stack_inode_flags(inode_item, flags);
1662         btrfs_set_stack_inode_block_group(inode_item, 0);
1663
1664         btrfs_set_stack_timespec_sec(&inode_item->atime,
1665                                      inode->i_atime.tv_sec);
1666         btrfs_set_stack_timespec_nsec(&inode_item->atime,
1667                                       inode->i_atime.tv_nsec);
1668
1669         btrfs_set_stack_timespec_sec(&inode_item->mtime,
1670                                      inode->i_mtime.tv_sec);
1671         btrfs_set_stack_timespec_nsec(&inode_item->mtime,
1672                                       inode->i_mtime.tv_nsec);
1673
1674         btrfs_set_stack_timespec_sec(&inode_item->ctime,
1675                                      inode->i_ctime.tv_sec);
1676         btrfs_set_stack_timespec_nsec(&inode_item->ctime,
1677                                       inode->i_ctime.tv_nsec);
1678
1679         btrfs_set_stack_timespec_sec(&inode_item->otime,
1680                                      BTRFS_I(inode)->i_otime.tv_sec);
1681         btrfs_set_stack_timespec_nsec(&inode_item->otime,
1682                                      BTRFS_I(inode)->i_otime.tv_nsec);
1683 }
1684
1685 int btrfs_fill_inode(struct inode *inode, u32 *rdev)
1686 {
1687         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
1688         struct btrfs_delayed_node *delayed_node;
1689         struct btrfs_inode_item *inode_item;
1690
1691         delayed_node = btrfs_get_delayed_node(BTRFS_I(inode));
1692         if (!delayed_node)
1693                 return -ENOENT;
1694
1695         mutex_lock(&delayed_node->mutex);
1696         if (!test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1697                 mutex_unlock(&delayed_node->mutex);
1698                 btrfs_release_delayed_node(delayed_node);
1699                 return -ENOENT;
1700         }
1701
1702         inode_item = &delayed_node->inode_item;
1703
1704         i_uid_write(inode, btrfs_stack_inode_uid(inode_item));
1705         i_gid_write(inode, btrfs_stack_inode_gid(inode_item));
1706         btrfs_i_size_write(BTRFS_I(inode), btrfs_stack_inode_size(inode_item));
1707         btrfs_inode_set_file_extent_range(BTRFS_I(inode), 0,
1708                         round_up(i_size_read(inode), fs_info->sectorsize));
1709         inode->i_mode = btrfs_stack_inode_mode(inode_item);
1710         set_nlink(inode, btrfs_stack_inode_nlink(inode_item));
1711         inode_set_bytes(inode, btrfs_stack_inode_nbytes(inode_item));
1712         BTRFS_I(inode)->generation = btrfs_stack_inode_generation(inode_item);
1713         BTRFS_I(inode)->last_trans = btrfs_stack_inode_transid(inode_item);
1714
1715         inode_set_iversion_queried(inode,
1716                                    btrfs_stack_inode_sequence(inode_item));
1717         inode->i_rdev = 0;
1718         *rdev = btrfs_stack_inode_rdev(inode_item);
1719         btrfs_inode_split_flags(btrfs_stack_inode_flags(inode_item),
1720                                 &BTRFS_I(inode)->flags, &BTRFS_I(inode)->ro_flags);
1721
1722         inode->i_atime.tv_sec = btrfs_stack_timespec_sec(&inode_item->atime);
1723         inode->i_atime.tv_nsec = btrfs_stack_timespec_nsec(&inode_item->atime);
1724
1725         inode->i_mtime.tv_sec = btrfs_stack_timespec_sec(&inode_item->mtime);
1726         inode->i_mtime.tv_nsec = btrfs_stack_timespec_nsec(&inode_item->mtime);
1727
1728         inode->i_ctime.tv_sec = btrfs_stack_timespec_sec(&inode_item->ctime);
1729         inode->i_ctime.tv_nsec = btrfs_stack_timespec_nsec(&inode_item->ctime);
1730
1731         BTRFS_I(inode)->i_otime.tv_sec =
1732                 btrfs_stack_timespec_sec(&inode_item->otime);
1733         BTRFS_I(inode)->i_otime.tv_nsec =
1734                 btrfs_stack_timespec_nsec(&inode_item->otime);
1735
1736         inode->i_generation = BTRFS_I(inode)->generation;
1737         BTRFS_I(inode)->index_cnt = (u64)-1;
1738
1739         mutex_unlock(&delayed_node->mutex);
1740         btrfs_release_delayed_node(delayed_node);
1741         return 0;
1742 }
1743
1744 int btrfs_delayed_update_inode(struct btrfs_trans_handle *trans,
1745                                struct btrfs_root *root,
1746                                struct btrfs_inode *inode)
1747 {
1748         struct btrfs_delayed_node *delayed_node;
1749         int ret = 0;
1750
1751         delayed_node = btrfs_get_or_create_delayed_node(inode);
1752         if (IS_ERR(delayed_node))
1753                 return PTR_ERR(delayed_node);
1754
1755         mutex_lock(&delayed_node->mutex);
1756         if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1757                 fill_stack_inode_item(trans, &delayed_node->inode_item,
1758                                       &inode->vfs_inode);
1759                 goto release_node;
1760         }
1761
1762         ret = btrfs_delayed_inode_reserve_metadata(trans, root, delayed_node);
1763         if (ret)
1764                 goto release_node;
1765
1766         fill_stack_inode_item(trans, &delayed_node->inode_item, &inode->vfs_inode);
1767         set_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags);
1768         delayed_node->count++;
1769         atomic_inc(&root->fs_info->delayed_root->items);
1770 release_node:
1771         mutex_unlock(&delayed_node->mutex);
1772         btrfs_release_delayed_node(delayed_node);
1773         return ret;
1774 }
1775
1776 int btrfs_delayed_delete_inode_ref(struct btrfs_inode *inode)
1777 {
1778         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1779         struct btrfs_delayed_node *delayed_node;
1780
1781         /*
1782          * we don't do delayed inode updates during log recovery because it
1783          * leads to enospc problems.  This means we also can't do
1784          * delayed inode refs
1785          */
1786         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
1787                 return -EAGAIN;
1788
1789         delayed_node = btrfs_get_or_create_delayed_node(inode);
1790         if (IS_ERR(delayed_node))
1791                 return PTR_ERR(delayed_node);
1792
1793         /*
1794          * We don't reserve space for inode ref deletion is because:
1795          * - We ONLY do async inode ref deletion for the inode who has only
1796          *   one link(i_nlink == 1), it means there is only one inode ref.
1797          *   And in most case, the inode ref and the inode item are in the
1798          *   same leaf, and we will deal with them at the same time.
1799          *   Since we are sure we will reserve the space for the inode item,
1800          *   it is unnecessary to reserve space for inode ref deletion.
1801          * - If the inode ref and the inode item are not in the same leaf,
1802          *   We also needn't worry about enospc problem, because we reserve
1803          *   much more space for the inode update than it needs.
1804          * - At the worst, we can steal some space from the global reservation.
1805          *   It is very rare.
1806          */
1807         mutex_lock(&delayed_node->mutex);
1808         if (test_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags))
1809                 goto release_node;
1810
1811         set_bit(BTRFS_DELAYED_NODE_DEL_IREF, &delayed_node->flags);
1812         delayed_node->count++;
1813         atomic_inc(&fs_info->delayed_root->items);
1814 release_node:
1815         mutex_unlock(&delayed_node->mutex);
1816         btrfs_release_delayed_node(delayed_node);
1817         return 0;
1818 }
1819
1820 static void __btrfs_kill_delayed_node(struct btrfs_delayed_node *delayed_node)
1821 {
1822         struct btrfs_root *root = delayed_node->root;
1823         struct btrfs_fs_info *fs_info = root->fs_info;
1824         struct btrfs_delayed_item *curr_item, *prev_item;
1825
1826         mutex_lock(&delayed_node->mutex);
1827         curr_item = __btrfs_first_delayed_insertion_item(delayed_node);
1828         while (curr_item) {
1829                 btrfs_delayed_item_release_metadata(root, curr_item);
1830                 prev_item = curr_item;
1831                 curr_item = __btrfs_next_delayed_item(prev_item);
1832                 btrfs_release_delayed_item(prev_item);
1833         }
1834
1835         curr_item = __btrfs_first_delayed_deletion_item(delayed_node);
1836         while (curr_item) {
1837                 btrfs_delayed_item_release_metadata(root, curr_item);
1838                 prev_item = curr_item;
1839                 curr_item = __btrfs_next_delayed_item(prev_item);
1840                 btrfs_release_delayed_item(prev_item);
1841         }
1842
1843         btrfs_release_delayed_iref(delayed_node);
1844
1845         if (test_bit(BTRFS_DELAYED_NODE_INODE_DIRTY, &delayed_node->flags)) {
1846                 btrfs_delayed_inode_release_metadata(fs_info, delayed_node, false);
1847                 btrfs_release_delayed_inode(delayed_node);
1848         }
1849         mutex_unlock(&delayed_node->mutex);
1850 }
1851
1852 void btrfs_kill_delayed_inode_items(struct btrfs_inode *inode)
1853 {
1854         struct btrfs_delayed_node *delayed_node;
1855
1856         delayed_node = btrfs_get_delayed_node(inode);
1857         if (!delayed_node)
1858                 return;
1859
1860         __btrfs_kill_delayed_node(delayed_node);
1861         btrfs_release_delayed_node(delayed_node);
1862 }
1863
1864 void btrfs_kill_all_delayed_nodes(struct btrfs_root *root)
1865 {
1866         unsigned long index = 0;
1867         struct btrfs_delayed_node *delayed_node;
1868         struct btrfs_delayed_node *delayed_nodes[8];
1869
1870         while (1) {
1871                 int n = 0;
1872
1873                 spin_lock(&root->inode_lock);
1874                 if (xa_empty(&root->delayed_nodes)) {
1875                         spin_unlock(&root->inode_lock);
1876                         return;
1877                 }
1878
1879                 xa_for_each_start(&root->delayed_nodes, index, delayed_node, index) {
1880                         /*
1881                          * Don't increase refs in case the node is dead and
1882                          * about to be removed from the tree in the loop below
1883                          */
1884                         if (refcount_inc_not_zero(&delayed_node->refs)) {
1885                                 delayed_nodes[n] = delayed_node;
1886                                 n++;
1887                         }
1888                         if (n >= ARRAY_SIZE(delayed_nodes))
1889                                 break;
1890                 }
1891                 index++;
1892                 spin_unlock(&root->inode_lock);
1893
1894                 for (int i = 0; i < n; i++) {
1895                         __btrfs_kill_delayed_node(delayed_nodes[i]);
1896                         btrfs_release_delayed_node(delayed_nodes[i]);
1897                 }
1898         }
1899 }
1900
1901 void btrfs_destroy_delayed_inodes(struct btrfs_fs_info *fs_info)
1902 {
1903         struct btrfs_delayed_node *curr_node, *prev_node;
1904
1905         curr_node = btrfs_first_delayed_node(fs_info->delayed_root);
1906         while (curr_node) {
1907                 __btrfs_kill_delayed_node(curr_node);
1908
1909                 prev_node = curr_node;
1910                 curr_node = btrfs_next_delayed_node(curr_node);
1911                 btrfs_release_delayed_node(prev_node);
1912         }
1913 }
1914