fs/btrfs/delayed-ref.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2009 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/slab.h>
   8 #include <linux/sort.h>
   9 #include "ctree.h"
  10 #include "delayed-ref.h"
  11 #include "transaction.h"
  12 #include "qgroup.h"
  13 #include "space-info.h"
  14
  15 struct kmem_cache *btrfs_delayed_ref_head_cachep;
  16 struct kmem_cache *btrfs_delayed_tree_ref_cachep;
  17 struct kmem_cache *btrfs_delayed_data_ref_cachep;
  18 struct kmem_cache *btrfs_delayed_extent_op_cachep;
  19 /*
  20  * delayed back reference update tracking.  For subvolume trees
  21  * we queue up extent allocations and backref maintenance for
  22  * delayed processing.   This avoids deep call chains where we
  23  * add extents in the middle of btrfs_search_slot, and it allows
  24  * us to buffer up frequently modified backrefs in an rb tree instead
  25  * of hammering updates on the extent allocation tree.
  26  */
  27
  28 bool btrfs_check_space_for_delayed_refs(struct btrfs_fs_info *fs_info)
  29 {
  30         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
  31         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
  32         bool ret = false;
  33         u64 reserved;
  34
  35         spin_lock(&global_rsv->lock);
  36         reserved = global_rsv->reserved;
  37         spin_unlock(&global_rsv->lock);
  38
  39         /*
  40          * Since the global reserve is just kind of magic we don't really want
  41          * to rely on it to save our bacon, so if our size is more than the
  42          * delayed_refs_rsv and the global rsv then it's time to think about
  43          * bailing.
  44          */
  45         spin_lock(&delayed_refs_rsv->lock);
  46         reserved += delayed_refs_rsv->reserved;
  47         if (delayed_refs_rsv->size >= reserved)
  48                 ret = true;
  49         spin_unlock(&delayed_refs_rsv->lock);
  50         return ret;
  51 }
  52
  53 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans)
  54 {
  55         u64 num_entries =
  56                 atomic_read(&trans->transaction->delayed_refs.num_entries);
  57         u64 avg_runtime;
  58         u64 val;
  59
  60         smp_mb();
  61         avg_runtime = trans->fs_info->avg_delayed_ref_runtime;
  62         val = num_entries * avg_runtime;
  63         if (val >= NSEC_PER_SEC)
  64                 return 1;
  65         if (val >= NSEC_PER_SEC / 2)
  66                 return 2;
  67
  68         return btrfs_check_space_for_delayed_refs(trans->fs_info);
  69 }
  70
  71 /**
  72  * btrfs_delayed_refs_rsv_release - release a ref head's reservation.
  73  * @fs_info - the fs_info for our fs.
  74  * @nr - the number of items to drop.
  75  *
  76  * This drops the delayed ref head's count from the delayed refs rsv and frees
  77  * any excess reservation we had.
  78  */
  79 void btrfs_delayed_refs_rsv_release(struct btrfs_fs_info *fs_info, int nr)
  80 {
  81         struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
  82         u64 num_bytes = btrfs_calc_insert_metadata_size(fs_info, nr);
  83         u64 released = 0;
  84
  85         released = btrfs_block_rsv_release(fs_info, block_rsv, num_bytes, NULL);
  86         if (released)
  87                 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
  88                                               0, released, 0);
  89 }
  90
  91 /*
  92  * btrfs_update_delayed_refs_rsv - adjust the size of the delayed refs rsv
  93  * @trans - the trans that may have generated delayed refs
  94  *
  95  * This is to be called anytime we may have adjusted trans->delayed_ref_updates,
  96  * it'll calculate the additional size and add it to the delayed_refs_rsv.
  97  */
  98 void btrfs_update_delayed_refs_rsv(struct btrfs_trans_handle *trans)
  99 {
 100         struct btrfs_fs_info *fs_info = trans->fs_info;
 101         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_refs_rsv;
 102         u64 num_bytes;
 103
 104         if (!trans->delayed_ref_updates)
 105                 return;
 106
 107         num_bytes = btrfs_calc_insert_metadata_size(fs_info,
 108                                                     trans->delayed_ref_updates);
 109         spin_lock(&delayed_rsv->lock);
 110         delayed_rsv->size += num_bytes;
 111         delayed_rsv->full = 0;
 112         spin_unlock(&delayed_rsv->lock);
 113         trans->delayed_ref_updates = 0;
 114 }
 115
 116 /**
 117  * btrfs_migrate_to_delayed_refs_rsv - transfer bytes to our delayed refs rsv.
 118  * @fs_info - the fs info for our fs.
 119  * @src - the source block rsv to transfer from.
 120  * @num_bytes - the number of bytes to transfer.
 121  *
 122  * This transfers up to the num_bytes amount from the src rsv to the
 123  * delayed_refs_rsv.  Any extra bytes are returned to the space info.
 124  */
 125 void btrfs_migrate_to_delayed_refs_rsv(struct btrfs_fs_info *fs_info,
 126                                        struct btrfs_block_rsv *src,
 127                                        u64 num_bytes)
 128 {
 129         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
 130         u64 to_free = 0;
 131
 132         spin_lock(&src->lock);
 133         src->reserved -= num_bytes;
 134         src->size -= num_bytes;
 135         spin_unlock(&src->lock);
 136
 137         spin_lock(&delayed_refs_rsv->lock);
 138         if (delayed_refs_rsv->size > delayed_refs_rsv->reserved) {
 139                 u64 delta = delayed_refs_rsv->size -
 140                         delayed_refs_rsv->reserved;
 141                 if (num_bytes > delta) {
 142                         to_free = num_bytes - delta;
 143                         num_bytes = delta;
 144                 }
 145         } else {
 146                 to_free = num_bytes;
 147                 num_bytes = 0;
 148         }
 149
 150         if (num_bytes)
 151                 delayed_refs_rsv->reserved += num_bytes;
 152         if (delayed_refs_rsv->reserved >= delayed_refs_rsv->size)
 153                 delayed_refs_rsv->full = 1;
 154         spin_unlock(&delayed_refs_rsv->lock);
 155
 156         if (num_bytes)
 157                 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
 158                                               0, num_bytes, 1);
 159         if (to_free)
 160                 btrfs_space_info_free_bytes_may_use(fs_info,
 161                                 delayed_refs_rsv->space_info, to_free);
 162 }
 163
 164 /**
 165  * btrfs_delayed_refs_rsv_refill - refill based on our delayed refs usage.
 166  * @fs_info - the fs_info for our fs.
 167  * @flush - control how we can flush for this reservation.
 168  *
 169  * This will refill the delayed block_rsv up to 1 items size worth of space and
 170  * will return -ENOSPC if we can't make the reservation.
 171  */
 172 int btrfs_delayed_refs_rsv_refill(struct btrfs_fs_info *fs_info,
 173                                   enum btrfs_reserve_flush_enum flush)
 174 {
 175         struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
 176         u64 limit = btrfs_calc_insert_metadata_size(fs_info, 1);
 177         u64 num_bytes = 0;
 178         int ret = -ENOSPC;
 179
 180         spin_lock(&block_rsv->lock);
 181         if (block_rsv->reserved < block_rsv->size) {
 182                 num_bytes = block_rsv->size - block_rsv->reserved;
 183                 num_bytes = min(num_bytes, limit);
 184         }
 185         spin_unlock(&block_rsv->lock);
 186
 187         if (!num_bytes)
 188                 return 0;
 189
 190         ret = btrfs_reserve_metadata_bytes(fs_info->extent_root, block_rsv,
 191                                            num_bytes, flush);
 192         if (ret)
 193                 return ret;
 194         btrfs_block_rsv_add_bytes(block_rsv, num_bytes, 0);
 195         trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
 196                                       0, num_bytes, 1);
 197         return 0;
 198 }
 199
 200 /*
 201  * compare two delayed tree backrefs with same bytenr and type
 202  */
 203 static int comp_tree_refs(struct btrfs_delayed_tree_ref *ref1,
 204                           struct btrfs_delayed_tree_ref *ref2)
 205 {
 206         if (ref1->node.type == BTRFS_TREE_BLOCK_REF_KEY) {
 207                 if (ref1->root < ref2->root)
 208                         return -1;
 209                 if (ref1->root > ref2->root)
 210                         return 1;
 211         } else {
 212                 if (ref1->parent < ref2->parent)
 213                         return -1;
 214                 if (ref1->parent > ref2->parent)
 215                         return 1;
 216         }
 217         return 0;
 218 }
 219
 220 /*
 221  * compare two delayed data backrefs with same bytenr and type
 222  */
 223 static int comp_data_refs(struct btrfs_delayed_data_ref *ref1,
 224                           struct btrfs_delayed_data_ref *ref2)
 225 {
 226         if (ref1->node.type == BTRFS_EXTENT_DATA_REF_KEY) {
 227                 if (ref1->root < ref2->root)
 228                         return -1;
 229                 if (ref1->root > ref2->root)
 230                         return 1;
 231                 if (ref1->objectid < ref2->objectid)
 232                         return -1;
 233                 if (ref1->objectid > ref2->objectid)
 234                         return 1;
 235                 if (ref1->offset < ref2->offset)
 236                         return -1;
 237                 if (ref1->offset > ref2->offset)
 238                         return 1;
 239         } else {
 240                 if (ref1->parent < ref2->parent)
 241                         return -1;
 242                 if (ref1->parent > ref2->parent)
 243                         return 1;
 244         }
 245         return 0;
 246 }
 247
 248 static int comp_refs(struct btrfs_delayed_ref_node *ref1,
 249                      struct btrfs_delayed_ref_node *ref2,
 250                      bool check_seq)
 251 {
 252         int ret = 0;
 253
 254         if (ref1->type < ref2->type)
 255                 return -1;
 256         if (ref1->type > ref2->type)
 257                 return 1;
 258         if (ref1->type == BTRFS_TREE_BLOCK_REF_KEY ||
 259             ref1->type == BTRFS_SHARED_BLOCK_REF_KEY)
 260                 ret = comp_tree_refs(btrfs_delayed_node_to_tree_ref(ref1),
 261                                      btrfs_delayed_node_to_tree_ref(ref2));
 262         else
 263                 ret = comp_data_refs(btrfs_delayed_node_to_data_ref(ref1),
 264                                      btrfs_delayed_node_to_data_ref(ref2));
 265         if (ret)
 266                 return ret;
 267         if (check_seq) {
 268                 if (ref1->seq < ref2->seq)
 269                         return -1;
 270                 if (ref1->seq > ref2->seq)
 271                         return 1;
 272         }
 273         return 0;
 274 }
 275
 276 /* insert a new ref to head ref rbtree */
 277 static struct btrfs_delayed_ref_head *htree_insert(struct rb_root_cached *root,
 278                                                    struct rb_node *node)
 279 {
 280         struct rb_node **p = &root->rb_root.rb_node;
 281         struct rb_node *parent_node = NULL;
 282         struct btrfs_delayed_ref_head *entry;
 283         struct btrfs_delayed_ref_head *ins;
 284         u64 bytenr;
 285         bool leftmost = true;
 286
 287         ins = rb_entry(node, struct btrfs_delayed_ref_head, href_node);
 288         bytenr = ins->bytenr;
 289         while (*p) {
 290                 parent_node = *p;
 291                 entry = rb_entry(parent_node, struct btrfs_delayed_ref_head,
 292                                  href_node);
 293
 294                 if (bytenr < entry->bytenr) {
 295                         p = &(*p)->rb_left;
 296                 } else if (bytenr > entry->bytenr) {
 297                         p = &(*p)->rb_right;
 298                         leftmost = false;
 299                 } else {
 300                         return entry;
 301                 }
 302         }
 303
 304         rb_link_node(node, parent_node, p);
 305         rb_insert_color_cached(node, root, leftmost);
 306         return NULL;
 307 }
 308
 309 static struct btrfs_delayed_ref_node* tree_insert(struct rb_root_cached *root,
 310                 struct btrfs_delayed_ref_node *ins)
 311 {
 312         struct rb_node **p = &root->rb_root.rb_node;
 313         struct rb_node *node = &ins->ref_node;
 314         struct rb_node *parent_node = NULL;
 315         struct btrfs_delayed_ref_node *entry;
 316         bool leftmost = true;
 317
 318         while (*p) {
 319                 int comp;
 320
 321                 parent_node = *p;
 322                 entry = rb_entry(parent_node, struct btrfs_delayed_ref_node,
 323                                  ref_node);
 324                 comp = comp_refs(ins, entry, true);
 325                 if (comp < 0) {
 326                         p = &(*p)->rb_left;
 327                 } else if (comp > 0) {
 328                         p = &(*p)->rb_right;
 329                         leftmost = false;
 330                 } else {
 331                         return entry;
 332                 }
 333         }
 334
 335         rb_link_node(node, parent_node, p);
 336         rb_insert_color_cached(node, root, leftmost);
 337         return NULL;
 338 }
 339
 340 static struct btrfs_delayed_ref_head *find_first_ref_head(
 341                 struct btrfs_delayed_ref_root *dr)
 342 {
 343         struct rb_node *n;
 344         struct btrfs_delayed_ref_head *entry;
 345
 346         n = rb_first_cached(&dr->href_root);
 347         if (!n)
 348                 return NULL;
 349
 350         entry = rb_entry(n, struct btrfs_delayed_ref_head, href_node);
 351
 352         return entry;
 353 }
 354
 355 /*
 356  * Find a head entry based on bytenr. This returns the delayed ref head if it
 357  * was able to find one, or NULL if nothing was in that spot.  If return_bigger
 358  * is given, the next bigger entry is returned if no exact match is found.
 359  */
 360 static struct btrfs_delayed_ref_head *find_ref_head(
 361                 struct btrfs_delayed_ref_root *dr, u64 bytenr,
 362                 bool return_bigger)
 363 {
 364         struct rb_root *root = &dr->href_root.rb_root;
 365         struct rb_node *n;
 366         struct btrfs_delayed_ref_head *entry;
 367
 368         n = root->rb_node;
 369         entry = NULL;
 370         while (n) {
 371                 entry = rb_entry(n, struct btrfs_delayed_ref_head, href_node);
 372
 373                 if (bytenr < entry->bytenr)
 374                         n = n->rb_left;
 375                 else if (bytenr > entry->bytenr)
 376                         n = n->rb_right;
 377                 else
 378                         return entry;
 379         }
 380         if (entry && return_bigger) {
 381                 if (bytenr > entry->bytenr) {
 382                         n = rb_next(&entry->href_node);
 383                         if (!n)
 384                                 return NULL;
 385                         entry = rb_entry(n, struct btrfs_delayed_ref_head,
 386                                          href_node);
 387                 }
 388                 return entry;
 389         }
 390         return NULL;
 391 }
 392
 393 int btrfs_delayed_ref_lock(struct btrfs_delayed_ref_root *delayed_refs,
 394                            struct btrfs_delayed_ref_head *head)
 395 {
 396         lockdep_assert_held(&delayed_refs->lock);
 397         if (mutex_trylock(&head->mutex))
 398                 return 0;
 399
 400         refcount_inc(&head->refs);
 401         spin_unlock(&delayed_refs->lock);
 402
 403         mutex_lock(&head->mutex);
 404         spin_lock(&delayed_refs->lock);
 405         if (RB_EMPTY_NODE(&head->href_node)) {
 406                 mutex_unlock(&head->mutex);
 407                 btrfs_put_delayed_ref_head(head);
 408                 return -EAGAIN;
 409         }
 410         btrfs_put_delayed_ref_head(head);
 411         return 0;
 412 }
 413
 414 static inline void drop_delayed_ref(struct btrfs_trans_handle *trans,
 415                                     struct btrfs_delayed_ref_root *delayed_refs,
 416                                     struct btrfs_delayed_ref_head *head,
 417                                     struct btrfs_delayed_ref_node *ref)
 418 {
 419         lockdep_assert_held(&head->lock);
 420         rb_erase_cached(&ref->ref_node, &head->ref_tree);
 421         RB_CLEAR_NODE(&ref->ref_node);
 422         if (!list_empty(&ref->add_list))
 423                 list_del(&ref->add_list);
 424         ref->in_tree = 0;
 425         btrfs_put_delayed_ref(ref);
 426         atomic_dec(&delayed_refs->num_entries);
 427 }
 428
 429 static bool merge_ref(struct btrfs_trans_handle *trans,
 430                       struct btrfs_delayed_ref_root *delayed_refs,
 431                       struct btrfs_delayed_ref_head *head,
 432                       struct btrfs_delayed_ref_node *ref,
 433                       u64 seq)
 434 {
 435         struct btrfs_delayed_ref_node *next;
 436         struct rb_node *node = rb_next(&ref->ref_node);
 437         bool done = false;
 438
 439         while (!done && node) {
 440                 int mod;
 441
 442                 next = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
 443                 node = rb_next(node);
 444                 if (seq && next->seq >= seq)
 445                         break;
 446                 if (comp_refs(ref, next, false))
 447                         break;
 448
 449                 if (ref->action == next->action) {
 450                         mod = next->ref_mod;
 451                 } else {
 452                         if (ref->ref_mod < next->ref_mod) {
 453                                 swap(ref, next);
 454                                 done = true;
 455                         }
 456                         mod = -next->ref_mod;
 457                 }
 458
 459                 drop_delayed_ref(trans, delayed_refs, head, next);
 460                 ref->ref_mod += mod;
 461                 if (ref->ref_mod == 0) {
 462                         drop_delayed_ref(trans, delayed_refs, head, ref);
 463                         done = true;
 464                 } else {
 465                         /*
 466                          * Can't have multiples of the same ref on a tree block.
 467                          */
 468                         WARN_ON(ref->type == BTRFS_TREE_BLOCK_REF_KEY ||
 469                                 ref->type == BTRFS_SHARED_BLOCK_REF_KEY);
 470                 }
 471         }
 472
 473         return done;
 474 }
 475
 476 void btrfs_merge_delayed_refs(struct btrfs_trans_handle *trans,
 477                               struct btrfs_delayed_ref_root *delayed_refs,
 478                               struct btrfs_delayed_ref_head *head)
 479 {
 480         struct btrfs_fs_info *fs_info = trans->fs_info;
 481         struct btrfs_delayed_ref_node *ref;
 482         struct rb_node *node;
 483         u64 seq = 0;
 484
 485         lockdep_assert_held(&head->lock);
 486
 487         if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
 488                 return;
 489
 490         /* We don't have too many refs to merge for data. */
 491         if (head->is_data)
 492                 return;
 493
 494         read_lock(&fs_info->tree_mod_log_lock);
 495         if (!list_empty(&fs_info->tree_mod_seq_list)) {
 496                 struct seq_list *elem;
 497
 498                 elem = list_first_entry(&fs_info->tree_mod_seq_list,
 499                                         struct seq_list, list);
 500                 seq = elem->seq;
 501         }
 502         read_unlock(&fs_info->tree_mod_log_lock);
 503
 504 again:
 505         for (node = rb_first_cached(&head->ref_tree); node;
 506              node = rb_next(node)) {
 507                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
 508                 if (seq && ref->seq >= seq)
 509                         continue;
 510                 if (merge_ref(trans, delayed_refs, head, ref, seq))
 511                         goto again;
 512         }
 513 }
 514
 515 int btrfs_check_delayed_seq(struct btrfs_fs_info *fs_info, u64 seq)
 516 {
 517         struct seq_list *elem;
 518         int ret = 0;
 519
 520         read_lock(&fs_info->tree_mod_log_lock);
 521         if (!list_empty(&fs_info->tree_mod_seq_list)) {
 522                 elem = list_first_entry(&fs_info->tree_mod_seq_list,
 523                                         struct seq_list, list);
 524                 if (seq >= elem->seq) {
 525                         btrfs_debug(fs_info,
 526                                 "holding back delayed_ref %#x.%x, lowest is %#x.%x",
 527                                 (u32)(seq >> 32), (u32)seq,
 528                                 (u32)(elem->seq >> 32), (u32)elem->seq);
 529                         ret = 1;
 530                 }
 531         }
 532
 533         read_unlock(&fs_info->tree_mod_log_lock);
 534         return ret;
 535 }
 536
 537 struct btrfs_delayed_ref_head *btrfs_select_ref_head(
 538                 struct btrfs_delayed_ref_root *delayed_refs)
 539 {
 540         struct btrfs_delayed_ref_head *head;
 541
 542 again:
 543         head = find_ref_head(delayed_refs, delayed_refs->run_delayed_start,
 544                              true);
 545         if (!head && delayed_refs->run_delayed_start != 0) {
 546                 delayed_refs->run_delayed_start = 0;
 547                 head = find_first_ref_head(delayed_refs);
 548         }
 549         if (!head)
 550                 return NULL;
 551
 552         while (head->processing) {
 553                 struct rb_node *node;
 554
 555                 node = rb_next(&head->href_node);
 556                 if (!node) {
 557                         if (delayed_refs->run_delayed_start == 0)
 558                                 return NULL;
 559                         delayed_refs->run_delayed_start = 0;
 560                         goto again;
 561                 }
 562                 head = rb_entry(node, struct btrfs_delayed_ref_head,
 563                                 href_node);
 564         }
 565
 566         head->processing = 1;
 567         WARN_ON(delayed_refs->num_heads_ready == 0);
 568         delayed_refs->num_heads_ready--;
 569         delayed_refs->run_delayed_start = head->bytenr +
 570                 head->num_bytes;
 571         return head;
 572 }
 573
 574 void btrfs_delete_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
 575                            struct btrfs_delayed_ref_head *head)
 576 {
 577         lockdep_assert_held(&delayed_refs->lock);
 578         lockdep_assert_held(&head->lock);
 579
 580         rb_erase_cached(&head->href_node, &delayed_refs->href_root);
 581         RB_CLEAR_NODE(&head->href_node);
 582         atomic_dec(&delayed_refs->num_entries);
 583         delayed_refs->num_heads--;
 584         if (head->processing == 0)
 585                 delayed_refs->num_heads_ready--;
 586 }
 587
 588 /*
 589  * Helper to insert the ref_node to the tail or merge with tail.
 590  *
 591  * Return 0 for insert.
 592  * Return >0 for merge.
 593  */
 594 static int insert_delayed_ref(struct btrfs_trans_handle *trans,
 595                               struct btrfs_delayed_ref_root *root,
 596                               struct btrfs_delayed_ref_head *href,
 597                               struct btrfs_delayed_ref_node *ref)
 598 {
 599         struct btrfs_delayed_ref_node *exist;
 600         int mod;
 601         int ret = 0;
 602
 603         spin_lock(&href->lock);
 604         exist = tree_insert(&href->ref_tree, ref);
 605         if (!exist)
 606                 goto inserted;
 607
 608         /* Now we are sure we can merge */
 609         ret = 1;
 610         if (exist->action == ref->action) {
 611                 mod = ref->ref_mod;
 612         } else {
 613                 /* Need to change action */
 614                 if (exist->ref_mod < ref->ref_mod) {
 615                         exist->action = ref->action;
 616                         mod = -exist->ref_mod;
 617                         exist->ref_mod = ref->ref_mod;
 618                         if (ref->action == BTRFS_ADD_DELAYED_REF)
 619                                 list_add_tail(&exist->add_list,
 620                                               &href->ref_add_list);
 621                         else if (ref->action == BTRFS_DROP_DELAYED_REF) {
 622                                 ASSERT(!list_empty(&exist->add_list));
 623                                 list_del(&exist->add_list);
 624                         } else {
 625                                 ASSERT(0);
 626                         }
 627                 } else
 628                         mod = -ref->ref_mod;
 629         }
 630         exist->ref_mod += mod;
 631
 632         /* remove existing tail if its ref_mod is zero */
 633         if (exist->ref_mod == 0)
 634                 drop_delayed_ref(trans, root, href, exist);
 635         spin_unlock(&href->lock);
 636         return ret;
 637 inserted:
 638         if (ref->action == BTRFS_ADD_DELAYED_REF)
 639                 list_add_tail(&ref->add_list, &href->ref_add_list);
 640         atomic_inc(&root->num_entries);
 641         spin_unlock(&href->lock);
 642         return ret;
 643 }
 644
 645 /*
 646  * helper function to update the accounting in the head ref
 647  * existing and update must have the same bytenr
 648  */
 649 static noinline void update_existing_head_ref(struct btrfs_trans_handle *trans,
 650                          struct btrfs_delayed_ref_head *existing,
 651                          struct btrfs_delayed_ref_head *update,
 652                          int *old_ref_mod_ret)
 653 {
 654         struct btrfs_delayed_ref_root *delayed_refs =
 655                 &trans->transaction->delayed_refs;
 656         struct btrfs_fs_info *fs_info = trans->fs_info;
 657         int old_ref_mod;
 658
 659         BUG_ON(existing->is_data != update->is_data);
 660
 661         spin_lock(&existing->lock);
 662         if (update->must_insert_reserved) {
 663                 /* if the extent was freed and then
 664                  * reallocated before the delayed ref
 665                  * entries were processed, we can end up
 666                  * with an existing head ref without
 667                  * the must_insert_reserved flag set.
 668                  * Set it again here
 669                  */
 670                 existing->must_insert_reserved = update->must_insert_reserved;
 671
 672                 /*
 673                  * update the num_bytes so we make sure the accounting
 674                  * is done correctly
 675                  */
 676                 existing->num_bytes = update->num_bytes;
 677
 678         }
 679
 680         if (update->extent_op) {
 681                 if (!existing->extent_op) {
 682                         existing->extent_op = update->extent_op;
 683                 } else {
 684                         if (update->extent_op->update_key) {
 685                                 memcpy(&existing->extent_op->key,
 686                                        &update->extent_op->key,
 687                                        sizeof(update->extent_op->key));
 688                                 existing->extent_op->update_key = true;
 689                         }
 690                         if (update->extent_op->update_flags) {
 691                                 existing->extent_op->flags_to_set |=
 692                                         update->extent_op->flags_to_set;
 693                                 existing->extent_op->update_flags = true;
 694                         }
 695                         btrfs_free_delayed_extent_op(update->extent_op);
 696                 }
 697         }
 698         /*
 699          * update the reference mod on the head to reflect this new operation,
 700          * only need the lock for this case cause we could be processing it
 701          * currently, for refs we just added we know we're a-ok.
 702          */
 703         old_ref_mod = existing->total_ref_mod;
 704         if (old_ref_mod_ret)
 705                 *old_ref_mod_ret = old_ref_mod;
 706         existing->ref_mod += update->ref_mod;
 707         existing->total_ref_mod += update->ref_mod;
 708
 709         /*
 710          * If we are going to from a positive ref mod to a negative or vice
 711          * versa we need to make sure to adjust pending_csums accordingly.
 712          */
 713         if (existing->is_data) {
 714                 u64 csum_leaves =
 715                         btrfs_csum_bytes_to_leaves(fs_info,
 716                                                    existing->num_bytes);
 717
 718                 if (existing->total_ref_mod >= 0 && old_ref_mod < 0) {
 719                         delayed_refs->pending_csums -= existing->num_bytes;
 720                         btrfs_delayed_refs_rsv_release(fs_info, csum_leaves);
 721                 }
 722                 if (existing->total_ref_mod < 0 && old_ref_mod >= 0) {
 723                         delayed_refs->pending_csums += existing->num_bytes;
 724                         trans->delayed_ref_updates += csum_leaves;
 725                 }
 726         }
 727         spin_unlock(&existing->lock);
 728 }
 729
 730 static void init_delayed_ref_head(struct btrfs_delayed_ref_head *head_ref,
 731                                   struct btrfs_qgroup_extent_record *qrecord,
 732                                   u64 bytenr, u64 num_bytes, u64 ref_root,
 733                                   u64 reserved, int action, bool is_data,
 734                                   bool is_system)
 735 {
 736         int count_mod = 1;
 737         int must_insert_reserved = 0;
 738
 739         /* If reserved is provided, it must be a data extent. */
 740         BUG_ON(!is_data && reserved);
 741
 742         /*
 743          * The head node stores the sum of all the mods, so dropping a ref
 744          * should drop the sum in the head node by one.
 745          */
 746         if (action == BTRFS_UPDATE_DELAYED_HEAD)
 747                 count_mod = 0;
 748         else if (action == BTRFS_DROP_DELAYED_REF)
 749                 count_mod = -1;
 750
 751         /*
 752          * BTRFS_ADD_DELAYED_EXTENT means that we need to update the reserved
 753          * accounting when the extent is finally added, or if a later
 754          * modification deletes the delayed ref without ever inserting the
 755          * extent into the extent allocation tree.  ref->must_insert_reserved
 756          * is the flag used to record that accounting mods are required.
 757          *
 758          * Once we record must_insert_reserved, switch the action to
 759          * BTRFS_ADD_DELAYED_REF because other special casing is not required.
 760          */
 761         if (action == BTRFS_ADD_DELAYED_EXTENT)
 762                 must_insert_reserved = 1;
 763         else
 764                 must_insert_reserved = 0;
 765
 766         refcount_set(&head_ref->refs, 1);
 767         head_ref->bytenr = bytenr;
 768         head_ref->num_bytes = num_bytes;
 769         head_ref->ref_mod = count_mod;
 770         head_ref->must_insert_reserved = must_insert_reserved;
 771         head_ref->is_data = is_data;
 772         head_ref->is_system = is_system;
 773         head_ref->ref_tree = RB_ROOT_CACHED;
 774         INIT_LIST_HEAD(&head_ref->ref_add_list);
 775         RB_CLEAR_NODE(&head_ref->href_node);
 776         head_ref->processing = 0;
 777         head_ref->total_ref_mod = count_mod;
 778         spin_lock_init(&head_ref->lock);
 779         mutex_init(&head_ref->mutex);
 780
 781         if (qrecord) {
 782                 if (ref_root && reserved) {
 783                         qrecord->data_rsv = reserved;
 784                         qrecord->data_rsv_refroot = ref_root;
 785                 }
 786                 qrecord->bytenr = bytenr;
 787                 qrecord->num_bytes = num_bytes;
 788                 qrecord->old_roots = NULL;
 789         }
 790 }
 791
 792 /*
 793  * helper function to actually insert a head node into the rbtree.
 794  * this does all the dirty work in terms of maintaining the correct
 795  * overall modification count.
 796  */
 797 static noinline struct btrfs_delayed_ref_head *
 798 add_delayed_ref_head(struct btrfs_trans_handle *trans,
 799                      struct btrfs_delayed_ref_head *head_ref,
 800                      struct btrfs_qgroup_extent_record *qrecord,
 801                      int action, int *qrecord_inserted_ret,
 802                      int *old_ref_mod, int *new_ref_mod)
 803 {
 804         struct btrfs_delayed_ref_head *existing;
 805         struct btrfs_delayed_ref_root *delayed_refs;
 806         int qrecord_inserted = 0;
 807
 808         delayed_refs = &trans->transaction->delayed_refs;
 809
 810         /* Record qgroup extent info if provided */
 811         if (qrecord) {
 812                 if (btrfs_qgroup_trace_extent_nolock(trans->fs_info,
 813                                         delayed_refs, qrecord))
 814                         kfree(qrecord);
 815                 else
 816                         qrecord_inserted = 1;
 817         }
 818
 819         trace_add_delayed_ref_head(trans->fs_info, head_ref, action);
 820
 821         existing = htree_insert(&delayed_refs->href_root,
 822                                 &head_ref->href_node);
 823         if (existing) {
 824                 update_existing_head_ref(trans, existing, head_ref,
 825                                          old_ref_mod);
 826                 /*
 827                  * we've updated the existing ref, free the newly
 828                  * allocated ref
 829                  */
 830                 kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
 831                 head_ref = existing;
 832         } else {
 833                 if (old_ref_mod)
 834                         *old_ref_mod = 0;
 835                 if (head_ref->is_data && head_ref->ref_mod < 0) {
 836                         delayed_refs->pending_csums += head_ref->num_bytes;
 837                         trans->delayed_ref_updates +=
 838                                 btrfs_csum_bytes_to_leaves(trans->fs_info,
 839                                                            head_ref->num_bytes);
 840                 }
 841                 delayed_refs->num_heads++;
 842                 delayed_refs->num_heads_ready++;
 843                 atomic_inc(&delayed_refs->num_entries);
 844                 trans->delayed_ref_updates++;
 845         }
 846         if (qrecord_inserted_ret)
 847                 *qrecord_inserted_ret = qrecord_inserted;
 848         if (new_ref_mod)
 849                 *new_ref_mod = head_ref->total_ref_mod;
 850
 851         return head_ref;
 852 }
 853
 854 /*
 855  * init_delayed_ref_common - Initialize the structure which represents a
 856  *                           modification to a an extent.
 857  *
 858  * @fs_info:    Internal to the mounted filesystem mount structure.
 859  *
 860  * @ref:        The structure which is going to be initialized.
 861  *
 862  * @bytenr:     The logical address of the extent for which a modification is
 863  *              going to be recorded.
 864  *
 865  * @num_bytes:  Size of the extent whose modification is being recorded.
 866  *
 867  * @ref_root:   The id of the root where this modification has originated, this
 868  *              can be either one of the well-known metadata trees or the
 869  *              subvolume id which references this extent.
 870  *
 871  * @action:     Can be one of BTRFS_ADD_DELAYED_REF/BTRFS_DROP_DELAYED_REF or
 872  *              BTRFS_ADD_DELAYED_EXTENT
 873  *
 874  * @ref_type:   Holds the type of the extent which is being recorded, can be
 875  *              one of BTRFS_SHARED_BLOCK_REF_KEY/BTRFS_TREE_BLOCK_REF_KEY
 876  *              when recording a metadata extent or BTRFS_SHARED_DATA_REF_KEY/
 877  *              BTRFS_EXTENT_DATA_REF_KEY when recording data extent
 878  */
 879 static void init_delayed_ref_common(struct btrfs_fs_info *fs_info,
 880                                     struct btrfs_delayed_ref_node *ref,
 881                                     u64 bytenr, u64 num_bytes, u64 ref_root,
 882                                     int action, u8 ref_type)
 883 {
 884         u64 seq = 0;
 885
 886         if (action == BTRFS_ADD_DELAYED_EXTENT)
 887                 action = BTRFS_ADD_DELAYED_REF;
 888
 889         if (is_fstree(ref_root))
 890                 seq = atomic64_read(&fs_info->tree_mod_seq);
 891
 892         refcount_set(&ref->refs, 1);
 893         ref->bytenr = bytenr;
 894         ref->num_bytes = num_bytes;
 895         ref->ref_mod = 1;
 896         ref->action = action;
 897         ref->is_head = 0;
 898         ref->in_tree = 1;
 899         ref->seq = seq;
 900         ref->type = ref_type;
 901         RB_CLEAR_NODE(&ref->ref_node);
 902         INIT_LIST_HEAD(&ref->add_list);
 903 }
 904
 905 /*
 906  * add a delayed tree ref.  This does all of the accounting required
 907  * to make sure the delayed ref is eventually processed before this
 908  * transaction commits.
 909  */
 910 int btrfs_add_delayed_tree_ref(struct btrfs_trans_handle *trans,
 911                                struct btrfs_ref *generic_ref,
 912                                struct btrfs_delayed_extent_op *extent_op,
 913                                int *old_ref_mod, int *new_ref_mod)
 914 {
 915         struct btrfs_fs_info *fs_info = trans->fs_info;
 916         struct btrfs_delayed_tree_ref *ref;
 917         struct btrfs_delayed_ref_head *head_ref;
 918         struct btrfs_delayed_ref_root *delayed_refs;
 919         struct btrfs_qgroup_extent_record *record = NULL;
 920         int qrecord_inserted;
 921         bool is_system;
 922         int action = generic_ref->action;
 923         int level = generic_ref->tree_ref.level;
 924         int ret;
 925         u64 bytenr = generic_ref->bytenr;
 926         u64 num_bytes = generic_ref->len;
 927         u64 parent = generic_ref->parent;
 928         u8 ref_type;
 929
 930         is_system = (generic_ref->real_root == BTRFS_CHUNK_TREE_OBJECTID);
 931
 932         ASSERT(generic_ref->type == BTRFS_REF_METADATA && generic_ref->action);
 933         BUG_ON(extent_op && extent_op->is_data);
 934         ref = kmem_cache_alloc(btrfs_delayed_tree_ref_cachep, GFP_NOFS);
 935         if (!ref)
 936                 return -ENOMEM;
 937
 938         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
 939         if (!head_ref) {
 940                 kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
 941                 return -ENOMEM;
 942         }
 943
 944         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) &&
 945             is_fstree(generic_ref->real_root) &&
 946             is_fstree(generic_ref->tree_ref.root) &&
 947             !generic_ref->skip_qgroup) {
 948                 record = kzalloc(sizeof(*record), GFP_NOFS);
 949                 if (!record) {
 950                         kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
 951                         kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
 952                         return -ENOMEM;
 953                 }
 954         }
 955
 956         if (parent)
 957                 ref_type = BTRFS_SHARED_BLOCK_REF_KEY;
 958         else
 959                 ref_type = BTRFS_TREE_BLOCK_REF_KEY;
 960
 961         init_delayed_ref_common(fs_info, &ref->node, bytenr, num_bytes,
 962                                 generic_ref->tree_ref.root, action, ref_type);
 963         ref->root = generic_ref->tree_ref.root;
 964         ref->parent = parent;
 965         ref->level = level;
 966
 967         init_delayed_ref_head(head_ref, record, bytenr, num_bytes,
 968                               generic_ref->tree_ref.root, 0, action, false,
 969                               is_system);
 970         head_ref->extent_op = extent_op;
 971
 972         delayed_refs = &trans->transaction->delayed_refs;
 973         spin_lock(&delayed_refs->lock);
 974
 975         /*
 976          * insert both the head node and the new ref without dropping
 977          * the spin lock
 978          */
 979         head_ref = add_delayed_ref_head(trans, head_ref, record,
 980                                         action, &qrecord_inserted,
 981                                         old_ref_mod, new_ref_mod);
 982
 983         ret = insert_delayed_ref(trans, delayed_refs, head_ref, &ref->node);
 984         spin_unlock(&delayed_refs->lock);
 985
 986         /*
 987          * Need to update the delayed_refs_rsv with any changes we may have
 988          * made.
 989          */
 990         btrfs_update_delayed_refs_rsv(trans);
 991
 992         trace_add_delayed_tree_ref(fs_info, &ref->node, ref,
 993                                    action == BTRFS_ADD_DELAYED_EXTENT ?
 994                                    BTRFS_ADD_DELAYED_REF : action);
 995         if (ret > 0)
 996                 kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
 997
 998         if (qrecord_inserted)
 999                 btrfs_qgroup_trace_extent_post(fs_info, record);
1000
1001         return 0;
1002 }
1003
1004 /*
1005  * add a delayed data ref. it's similar to btrfs_add_delayed_tree_ref.
1006  */
1007 int btrfs_add_delayed_data_ref(struct btrfs_trans_handle *trans,
1008                                struct btrfs_ref *generic_ref,
1009                                u64 reserved, int *old_ref_mod,
1010                                int *new_ref_mod)
1011 {
1012         struct btrfs_fs_info *fs_info = trans->fs_info;
1013         struct btrfs_delayed_data_ref *ref;
1014         struct btrfs_delayed_ref_head *head_ref;
1015         struct btrfs_delayed_ref_root *delayed_refs;
1016         struct btrfs_qgroup_extent_record *record = NULL;
1017         int qrecord_inserted;
1018         int action = generic_ref->action;
1019         int ret;
1020         u64 bytenr = generic_ref->bytenr;
1021         u64 num_bytes = generic_ref->len;
1022         u64 parent = generic_ref->parent;
1023         u64 ref_root = generic_ref->data_ref.ref_root;
1024         u64 owner = generic_ref->data_ref.ino;
1025         u64 offset = generic_ref->data_ref.offset;
1026         u8 ref_type;
1027
1028         ASSERT(generic_ref->type == BTRFS_REF_DATA && action);
1029         ref = kmem_cache_alloc(btrfs_delayed_data_ref_cachep, GFP_NOFS);
1030         if (!ref)
1031                 return -ENOMEM;
1032
1033         if (parent)
1034                 ref_type = BTRFS_SHARED_DATA_REF_KEY;
1035         else
1036                 ref_type = BTRFS_EXTENT_DATA_REF_KEY;
1037         init_delayed_ref_common(fs_info, &ref->node, bytenr, num_bytes,
1038                                 ref_root, action, ref_type);
1039         ref->root = ref_root;
1040         ref->parent = parent;
1041         ref->objectid = owner;
1042         ref->offset = offset;
1043
1044
1045         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
1046         if (!head_ref) {
1047                 kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
1048                 return -ENOMEM;
1049         }
1050
1051         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) &&
1052             is_fstree(ref_root) &&
1053             is_fstree(generic_ref->real_root) &&
1054             !generic_ref->skip_qgroup) {
1055                 record = kzalloc(sizeof(*record), GFP_NOFS);
1056                 if (!record) {
1057                         kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
1058                         kmem_cache_free(btrfs_delayed_ref_head_cachep,
1059                                         head_ref);
1060                         return -ENOMEM;
1061                 }
1062         }
1063
1064         init_delayed_ref_head(head_ref, record, bytenr, num_bytes, ref_root,
1065                               reserved, action, true, false);
1066         head_ref->extent_op = NULL;
1067
1068         delayed_refs = &trans->transaction->delayed_refs;
1069         spin_lock(&delayed_refs->lock);
1070
1071         /*
1072          * insert both the head node and the new ref without dropping
1073          * the spin lock
1074          */
1075         head_ref = add_delayed_ref_head(trans, head_ref, record,
1076                                         action, &qrecord_inserted,
1077                                         old_ref_mod, new_ref_mod);
1078
1079         ret = insert_delayed_ref(trans, delayed_refs, head_ref, &ref->node);
1080         spin_unlock(&delayed_refs->lock);
1081
1082         /*
1083          * Need to update the delayed_refs_rsv with any changes we may have
1084          * made.
1085          */
1086         btrfs_update_delayed_refs_rsv(trans);
1087
1088         trace_add_delayed_data_ref(trans->fs_info, &ref->node, ref,
1089                                    action == BTRFS_ADD_DELAYED_EXTENT ?
1090                                    BTRFS_ADD_DELAYED_REF : action);
1091         if (ret > 0)
1092                 kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
1093
1094
1095         if (qrecord_inserted)
1096                 return btrfs_qgroup_trace_extent_post(fs_info, record);
1097         return 0;
1098 }
1099
1100 int btrfs_add_delayed_extent_op(struct btrfs_trans_handle *trans,
1101                                 u64 bytenr, u64 num_bytes,
1102                                 struct btrfs_delayed_extent_op *extent_op)
1103 {
1104         struct btrfs_delayed_ref_head *head_ref;
1105         struct btrfs_delayed_ref_root *delayed_refs;
1106
1107         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
1108         if (!head_ref)
1109                 return -ENOMEM;
1110
1111         init_delayed_ref_head(head_ref, NULL, bytenr, num_bytes, 0, 0,
1112                               BTRFS_UPDATE_DELAYED_HEAD, extent_op->is_data,
1113                               false);
1114         head_ref->extent_op = extent_op;
1115
1116         delayed_refs = &trans->transaction->delayed_refs;
1117         spin_lock(&delayed_refs->lock);
1118
1119         add_delayed_ref_head(trans, head_ref, NULL, BTRFS_UPDATE_DELAYED_HEAD,
1120                              NULL, NULL, NULL);
1121
1122         spin_unlock(&delayed_refs->lock);
1123
1124         /*
1125          * Need to update the delayed_refs_rsv with any changes we may have
1126          * made.
1127          */
1128         btrfs_update_delayed_refs_rsv(trans);
1129         return 0;
1130 }
1131
1132 /*
1133  * This does a simple search for the head node for a given extent.  Returns the
1134  * head node if found, or NULL if not.
1135  */
1136 struct btrfs_delayed_ref_head *
1137 btrfs_find_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs, u64 bytenr)
1138 {
1139         lockdep_assert_held(&delayed_refs->lock);
1140
1141         return find_ref_head(delayed_refs, bytenr, false);
1142 }
1143
1144 void __cold btrfs_delayed_ref_exit(void)
1145 {
1146         kmem_cache_destroy(btrfs_delayed_ref_head_cachep);
1147         kmem_cache_destroy(btrfs_delayed_tree_ref_cachep);
1148         kmem_cache_destroy(btrfs_delayed_data_ref_cachep);
1149         kmem_cache_destroy(btrfs_delayed_extent_op_cachep);
1150 }
1151
1152 int __init btrfs_delayed_ref_init(void)
1153 {
1154         btrfs_delayed_ref_head_cachep = kmem_cache_create(
1155                                 "btrfs_delayed_ref_head",
1156                                 sizeof(struct btrfs_delayed_ref_head), 0,
1157                                 SLAB_MEM_SPREAD, NULL);
1158         if (!btrfs_delayed_ref_head_cachep)
1159                 goto fail;
1160
1161         btrfs_delayed_tree_ref_cachep = kmem_cache_create(
1162                                 "btrfs_delayed_tree_ref",
1163                                 sizeof(struct btrfs_delayed_tree_ref), 0,
1164                                 SLAB_MEM_SPREAD, NULL);
1165         if (!btrfs_delayed_tree_ref_cachep)
1166                 goto fail;
1167
1168         btrfs_delayed_data_ref_cachep = kmem_cache_create(
1169                                 "btrfs_delayed_data_ref",
1170                                 sizeof(struct btrfs_delayed_data_ref), 0,
1171                                 SLAB_MEM_SPREAD, NULL);
1172         if (!btrfs_delayed_data_ref_cachep)
1173                 goto fail;
1174
1175         btrfs_delayed_extent_op_cachep = kmem_cache_create(
1176                                 "btrfs_delayed_extent_op",
1177                                 sizeof(struct btrfs_delayed_extent_op), 0,
1178                                 SLAB_MEM_SPREAD, NULL);
1179         if (!btrfs_delayed_extent_op_cachep)
1180                 goto fail;
1181
1182         return 0;
1183 fail:
1184         btrfs_delayed_ref_exit();
1185         return -ENOMEM;
1186 }