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 #include "tree-mod-log.h"
  15
  16 struct kmem_cache *btrfs_delayed_ref_head_cachep;
  17 struct kmem_cache *btrfs_delayed_tree_ref_cachep;
  18 struct kmem_cache *btrfs_delayed_data_ref_cachep;
  19 struct kmem_cache *btrfs_delayed_extent_op_cachep;
  20 /*
  21  * delayed back reference update tracking.  For subvolume trees
  22  * we queue up extent allocations and backref maintenance for
  23  * delayed processing.   This avoids deep call chains where we
  24  * add extents in the middle of btrfs_search_slot, and it allows
  25  * us to buffer up frequently modified backrefs in an rb tree instead
  26  * of hammering updates on the extent allocation tree.
  27  */
  28
  29 bool btrfs_check_space_for_delayed_refs(struct btrfs_fs_info *fs_info)
  30 {
  31         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
  32         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
  33         bool ret = false;
  34         u64 reserved;
  35
  36         spin_lock(&global_rsv->lock);
  37         reserved = global_rsv->reserved;
  38         spin_unlock(&global_rsv->lock);
  39
  40         /*
  41          * Since the global reserve is just kind of magic we don't really want
  42          * to rely on it to save our bacon, so if our size is more than the
  43          * delayed_refs_rsv and the global rsv then it's time to think about
  44          * bailing.
  45          */
  46         spin_lock(&delayed_refs_rsv->lock);
  47         reserved += delayed_refs_rsv->reserved;
  48         if (delayed_refs_rsv->size >= reserved)
  49                 ret = true;
  50         spin_unlock(&delayed_refs_rsv->lock);
  51         return ret;
  52 }
  53
  54 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans)
  55 {
  56         u64 num_entries =
  57                 atomic_read(&trans->transaction->delayed_refs.num_entries);
  58         u64 avg_runtime;
  59         u64 val;
  60
  61         smp_mb();
  62         avg_runtime = trans->fs_info->avg_delayed_ref_runtime;
  63         val = num_entries * avg_runtime;
  64         if (val >= NSEC_PER_SEC)
  65                 return 1;
  66         if (val >= NSEC_PER_SEC / 2)
  67                 return 2;
  68
  69         return btrfs_check_space_for_delayed_refs(trans->fs_info);
  70 }
  71
  72 /**
  73  * Release a ref head's reservation
  74  *
  75  * @fs_info:  the filesystem
  76  * @nr:       number of items to drop
  77  *
  78  * This drops the delayed ref head's count from the delayed refs rsv and frees
  79  * any excess reservation we had.
  80  */
  81 void btrfs_delayed_refs_rsv_release(struct btrfs_fs_info *fs_info, int nr)
  82 {
  83         struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
  84         u64 num_bytes = btrfs_calc_insert_metadata_size(fs_info, nr);
  85         u64 released = 0;
  86
  87         released = btrfs_block_rsv_release(fs_info, block_rsv, num_bytes, NULL);
  88         if (released)
  89                 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
  90                                               0, released, 0);
  91 }
  92
  93 /*
  94  * btrfs_update_delayed_refs_rsv - adjust the size of the delayed refs rsv
  95  * @trans - the trans that may have generated delayed refs
  96  *
  97  * This is to be called anytime we may have adjusted trans->delayed_ref_updates,
  98  * it'll calculate the additional size and add it to the delayed_refs_rsv.
  99  */
 100 void btrfs_update_delayed_refs_rsv(struct btrfs_trans_handle *trans)
 101 {
 102         struct btrfs_fs_info *fs_info = trans->fs_info;
 103         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_refs_rsv;
 104         u64 num_bytes;
 105
 106         if (!trans->delayed_ref_updates)
 107                 return;
 108
 109         num_bytes = btrfs_calc_insert_metadata_size(fs_info,
 110                                                     trans->delayed_ref_updates);
 111         spin_lock(&delayed_rsv->lock);
 112         delayed_rsv->size += num_bytes;
 113         delayed_rsv->full = 0;
 114         spin_unlock(&delayed_rsv->lock);
 115         trans->delayed_ref_updates = 0;
 116 }
 117
 118 /**
 119  * Transfer bytes to our delayed refs rsv
 120  *
 121  * @fs_info:   the filesystem
 122  * @src:       source block rsv to transfer from
 123  * @num_bytes: number of bytes to transfer
 124  *
 125  * This transfers up to the num_bytes amount from the src rsv to the
 126  * delayed_refs_rsv.  Any extra bytes are returned to the space info.
 127  */
 128 void btrfs_migrate_to_delayed_refs_rsv(struct btrfs_fs_info *fs_info,
 129                                        struct btrfs_block_rsv *src,
 130                                        u64 num_bytes)
 131 {
 132         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
 133         u64 to_free = 0;
 134
 135         spin_lock(&src->lock);
 136         src->reserved -= num_bytes;
 137         src->size -= num_bytes;
 138         spin_unlock(&src->lock);
 139
 140         spin_lock(&delayed_refs_rsv->lock);
 141         if (delayed_refs_rsv->size > delayed_refs_rsv->reserved) {
 142                 u64 delta = delayed_refs_rsv->size -
 143                         delayed_refs_rsv->reserved;
 144                 if (num_bytes > delta) {
 145                         to_free = num_bytes - delta;
 146                         num_bytes = delta;
 147                 }
 148         } else {
 149                 to_free = num_bytes;
 150                 num_bytes = 0;
 151         }
 152
 153         if (num_bytes)
 154                 delayed_refs_rsv->reserved += num_bytes;
 155         if (delayed_refs_rsv->reserved >= delayed_refs_rsv->size)
 156                 delayed_refs_rsv->full = 1;
 157         spin_unlock(&delayed_refs_rsv->lock);
 158
 159         if (num_bytes)
 160                 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
 161                                               0, num_bytes, 1);
 162         if (to_free)
 163                 btrfs_space_info_free_bytes_may_use(fs_info,
 164                                 delayed_refs_rsv->space_info, to_free);
 165 }
 166
 167 /**
 168  * Refill based on our delayed refs usage
 169  *
 170  * @fs_info: the filesystem
 171  * @flush:   control how we can flush for this reservation.
 172  *
 173  * This will refill the delayed block_rsv up to 1 items size worth of space and
 174  * will return -ENOSPC if we can't make the reservation.
 175  */
 176 int btrfs_delayed_refs_rsv_refill(struct btrfs_fs_info *fs_info,
 177                                   enum btrfs_reserve_flush_enum flush)
 178 {
 179         struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
 180         u64 limit = btrfs_calc_insert_metadata_size(fs_info, 1);
 181         u64 num_bytes = 0;
 182         int ret = -ENOSPC;
 183
 184         spin_lock(&block_rsv->lock);
 185         if (block_rsv->reserved < block_rsv->size) {
 186                 num_bytes = block_rsv->size - block_rsv->reserved;
 187                 num_bytes = min(num_bytes, limit);
 188         }
 189         spin_unlock(&block_rsv->lock);
 190
 191         if (!num_bytes)
 192                 return 0;
 193
 194         ret = btrfs_reserve_metadata_bytes(fs_info->extent_root, block_rsv,
 195                                            num_bytes, flush);
 196         if (ret)
 197                 return ret;
 198         btrfs_block_rsv_add_bytes(block_rsv, num_bytes, 0);
 199         trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
 200                                       0, num_bytes, 1);
 201         return 0;
 202 }
 203
 204 /*
 205  * compare two delayed tree backrefs with same bytenr and type
 206  */
 207 static int comp_tree_refs(struct btrfs_delayed_tree_ref *ref1,
 208                           struct btrfs_delayed_tree_ref *ref2)
 209 {
 210         if (ref1->node.type == BTRFS_TREE_BLOCK_REF_KEY) {
 211                 if (ref1->root < ref2->root)
 212                         return -1;
 213                 if (ref1->root > ref2->root)
 214                         return 1;
 215         } else {
 216                 if (ref1->parent < ref2->parent)
 217                         return -1;
 218                 if (ref1->parent > ref2->parent)
 219                         return 1;
 220         }
 221         return 0;
 222 }
 223
 224 /*
 225  * compare two delayed data backrefs with same bytenr and type
 226  */
 227 static int comp_data_refs(struct btrfs_delayed_data_ref *ref1,
 228                           struct btrfs_delayed_data_ref *ref2)
 229 {
 230         if (ref1->node.type == BTRFS_EXTENT_DATA_REF_KEY) {
 231                 if (ref1->root < ref2->root)
 232                         return -1;
 233                 if (ref1->root > ref2->root)
 234                         return 1;
 235                 if (ref1->objectid < ref2->objectid)
 236                         return -1;
 237                 if (ref1->objectid > ref2->objectid)
 238                         return 1;
 239                 if (ref1->offset < ref2->offset)
 240                         return -1;
 241                 if (ref1->offset > ref2->offset)
 242                         return 1;
 243         } else {
 244                 if (ref1->parent < ref2->parent)
 245                         return -1;
 246                 if (ref1->parent > ref2->parent)
 247                         return 1;
 248         }
 249         return 0;
 250 }
 251
 252 static int comp_refs(struct btrfs_delayed_ref_node *ref1,
 253                      struct btrfs_delayed_ref_node *ref2,
 254                      bool check_seq)
 255 {
 256         int ret = 0;
 257
 258         if (ref1->type < ref2->type)
 259                 return -1;
 260         if (ref1->type > ref2->type)
 261                 return 1;
 262         if (ref1->type == BTRFS_TREE_BLOCK_REF_KEY ||
 263             ref1->type == BTRFS_SHARED_BLOCK_REF_KEY)
 264                 ret = comp_tree_refs(btrfs_delayed_node_to_tree_ref(ref1),
 265                                      btrfs_delayed_node_to_tree_ref(ref2));
 266         else
 267                 ret = comp_data_refs(btrfs_delayed_node_to_data_ref(ref1),
 268                                      btrfs_delayed_node_to_data_ref(ref2));
 269         if (ret)
 270                 return ret;
 271         if (check_seq) {
 272                 if (ref1->seq < ref2->seq)
 273                         return -1;
 274                 if (ref1->seq > ref2->seq)
 275                         return 1;
 276         }
 277         return 0;
 278 }
 279
 280 /* insert a new ref to head ref rbtree */
 281 static struct btrfs_delayed_ref_head *htree_insert(struct rb_root_cached *root,
 282                                                    struct rb_node *node)
 283 {
 284         struct rb_node **p = &root->rb_root.rb_node;
 285         struct rb_node *parent_node = NULL;
 286         struct btrfs_delayed_ref_head *entry;
 287         struct btrfs_delayed_ref_head *ins;
 288         u64 bytenr;
 289         bool leftmost = true;
 290
 291         ins = rb_entry(node, struct btrfs_delayed_ref_head, href_node);
 292         bytenr = ins->bytenr;
 293         while (*p) {
 294                 parent_node = *p;
 295                 entry = rb_entry(parent_node, struct btrfs_delayed_ref_head,
 296                                  href_node);
 297
 298                 if (bytenr < entry->bytenr) {
 299                         p = &(*p)->rb_left;
 300                 } else if (bytenr > entry->bytenr) {
 301                         p = &(*p)->rb_right;
 302                         leftmost = false;
 303                 } else {
 304                         return entry;
 305                 }
 306         }
 307
 308         rb_link_node(node, parent_node, p);
 309         rb_insert_color_cached(node, root, leftmost);
 310         return NULL;
 311 }
 312
 313 static struct btrfs_delayed_ref_node* tree_insert(struct rb_root_cached *root,
 314                 struct btrfs_delayed_ref_node *ins)
 315 {
 316         struct rb_node **p = &root->rb_root.rb_node;
 317         struct rb_node *node = &ins->ref_node;
 318         struct rb_node *parent_node = NULL;
 319         struct btrfs_delayed_ref_node *entry;
 320         bool leftmost = true;
 321
 322         while (*p) {
 323                 int comp;
 324
 325                 parent_node = *p;
 326                 entry = rb_entry(parent_node, struct btrfs_delayed_ref_node,
 327                                  ref_node);
 328                 comp = comp_refs(ins, entry, true);
 329                 if (comp < 0) {
 330                         p = &(*p)->rb_left;
 331                 } else if (comp > 0) {
 332                         p = &(*p)->rb_right;
 333                         leftmost = false;
 334                 } else {
 335                         return entry;
 336                 }
 337         }
 338
 339         rb_link_node(node, parent_node, p);
 340         rb_insert_color_cached(node, root, leftmost);
 341         return NULL;
 342 }
 343
 344 static struct btrfs_delayed_ref_head *find_first_ref_head(
 345                 struct btrfs_delayed_ref_root *dr)
 346 {
 347         struct rb_node *n;
 348         struct btrfs_delayed_ref_head *entry;
 349
 350         n = rb_first_cached(&dr->href_root);
 351         if (!n)
 352                 return NULL;
 353
 354         entry = rb_entry(n, struct btrfs_delayed_ref_head, href_node);
 355
 356         return entry;
 357 }
 358
 359 /*
 360  * Find a head entry based on bytenr. This returns the delayed ref head if it
 361  * was able to find one, or NULL if nothing was in that spot.  If return_bigger
 362  * is given, the next bigger entry is returned if no exact match is found.
 363  */
 364 static struct btrfs_delayed_ref_head *find_ref_head(
 365                 struct btrfs_delayed_ref_root *dr, u64 bytenr,
 366                 bool return_bigger)
 367 {
 368         struct rb_root *root = &dr->href_root.rb_root;
 369         struct rb_node *n;
 370         struct btrfs_delayed_ref_head *entry;
 371
 372         n = root->rb_node;
 373         entry = NULL;
 374         while (n) {
 375                 entry = rb_entry(n, struct btrfs_delayed_ref_head, href_node);
 376
 377                 if (bytenr < entry->bytenr)
 378                         n = n->rb_left;
 379                 else if (bytenr > entry->bytenr)
 380                         n = n->rb_right;
 381                 else
 382                         return entry;
 383         }
 384         if (entry && return_bigger) {
 385                 if (bytenr > entry->bytenr) {
 386                         n = rb_next(&entry->href_node);
 387                         if (!n)
 388                                 return NULL;
 389                         entry = rb_entry(n, struct btrfs_delayed_ref_head,
 390                                          href_node);
 391                 }
 392                 return entry;
 393         }
 394         return NULL;
 395 }
 396
 397 int btrfs_delayed_ref_lock(struct btrfs_delayed_ref_root *delayed_refs,
 398                            struct btrfs_delayed_ref_head *head)
 399 {
 400         lockdep_assert_held(&delayed_refs->lock);
 401         if (mutex_trylock(&head->mutex))
 402                 return 0;
 403
 404         refcount_inc(&head->refs);
 405         spin_unlock(&delayed_refs->lock);
 406
 407         mutex_lock(&head->mutex);
 408         spin_lock(&delayed_refs->lock);
 409         if (RB_EMPTY_NODE(&head->href_node)) {
 410                 mutex_unlock(&head->mutex);
 411                 btrfs_put_delayed_ref_head(head);
 412                 return -EAGAIN;
 413         }
 414         btrfs_put_delayed_ref_head(head);
 415         return 0;
 416 }
 417
 418 static inline void drop_delayed_ref(struct btrfs_trans_handle *trans,
 419                                     struct btrfs_delayed_ref_root *delayed_refs,
 420                                     struct btrfs_delayed_ref_head *head,
 421                                     struct btrfs_delayed_ref_node *ref)
 422 {
 423         lockdep_assert_held(&head->lock);
 424         rb_erase_cached(&ref->ref_node, &head->ref_tree);
 425         RB_CLEAR_NODE(&ref->ref_node);
 426         if (!list_empty(&ref->add_list))
 427                 list_del(&ref->add_list);
 428         ref->in_tree = 0;
 429         btrfs_put_delayed_ref(ref);
 430         atomic_dec(&delayed_refs->num_entries);
 431 }
 432
 433 static bool merge_ref(struct btrfs_trans_handle *trans,
 434                       struct btrfs_delayed_ref_root *delayed_refs,
 435                       struct btrfs_delayed_ref_head *head,
 436                       struct btrfs_delayed_ref_node *ref,
 437                       u64 seq)
 438 {
 439         struct btrfs_delayed_ref_node *next;
 440         struct rb_node *node = rb_next(&ref->ref_node);
 441         bool done = false;
 442
 443         while (!done && node) {
 444                 int mod;
 445
 446                 next = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
 447                 node = rb_next(node);
 448                 if (seq && next->seq >= seq)
 449                         break;
 450                 if (comp_refs(ref, next, false))
 451                         break;
 452
 453                 if (ref->action == next->action) {
 454                         mod = next->ref_mod;
 455                 } else {
 456                         if (ref->ref_mod < next->ref_mod) {
 457                                 swap(ref, next);
 458                                 done = true;
 459                         }
 460                         mod = -next->ref_mod;
 461                 }
 462
 463                 drop_delayed_ref(trans, delayed_refs, head, next);
 464                 ref->ref_mod += mod;
 465                 if (ref->ref_mod == 0) {
 466                         drop_delayed_ref(trans, delayed_refs, head, ref);
 467                         done = true;
 468                 } else {
 469                         /*
 470                          * Can't have multiples of the same ref on a tree block.
 471                          */
 472                         WARN_ON(ref->type == BTRFS_TREE_BLOCK_REF_KEY ||
 473                                 ref->type == BTRFS_SHARED_BLOCK_REF_KEY);
 474                 }
 475         }
 476
 477         return done;
 478 }
 479
 480 void btrfs_merge_delayed_refs(struct btrfs_trans_handle *trans,
 481                               struct btrfs_delayed_ref_root *delayed_refs,
 482                               struct btrfs_delayed_ref_head *head)
 483 {
 484         struct btrfs_fs_info *fs_info = trans->fs_info;
 485         struct btrfs_delayed_ref_node *ref;
 486         struct rb_node *node;
 487         u64 seq = 0;
 488
 489         lockdep_assert_held(&head->lock);
 490
 491         if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
 492                 return;
 493
 494         /* We don't have too many refs to merge for data. */
 495         if (head->is_data)
 496                 return;
 497
 498         seq = btrfs_tree_mod_log_lowest_seq(fs_info);
 499 again:
 500         for (node = rb_first_cached(&head->ref_tree); node;
 501              node = rb_next(node)) {
 502                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
 503                 if (seq && ref->seq >= seq)
 504                         continue;
 505                 if (merge_ref(trans, delayed_refs, head, ref, seq))
 506                         goto again;
 507         }
 508 }
 509
 510 int btrfs_check_delayed_seq(struct btrfs_fs_info *fs_info, u64 seq)
 511 {
 512         int ret = 0;
 513         u64 min_seq = btrfs_tree_mod_log_lowest_seq(fs_info);
 514
 515         if (min_seq != 0 && seq >= min_seq) {
 516                 btrfs_debug(fs_info,
 517                             "holding back delayed_ref %#x.%x, lowest is %#x.%x",
 518                             (u32)(seq >> 32), (u32)seq,
 519                             (u32)(min_seq >> 32), (u32)min_seq);
 520                 ret = 1;
 521         }
 522
 523         return ret;
 524 }
 525
 526 struct btrfs_delayed_ref_head *btrfs_select_ref_head(
 527                 struct btrfs_delayed_ref_root *delayed_refs)
 528 {
 529         struct btrfs_delayed_ref_head *head;
 530
 531 again:
 532         head = find_ref_head(delayed_refs, delayed_refs->run_delayed_start,
 533                              true);
 534         if (!head && delayed_refs->run_delayed_start != 0) {
 535                 delayed_refs->run_delayed_start = 0;
 536                 head = find_first_ref_head(delayed_refs);
 537         }
 538         if (!head)
 539                 return NULL;
 540
 541         while (head->processing) {
 542                 struct rb_node *node;
 543
 544                 node = rb_next(&head->href_node);
 545                 if (!node) {
 546                         if (delayed_refs->run_delayed_start == 0)
 547                                 return NULL;
 548                         delayed_refs->run_delayed_start = 0;
 549                         goto again;
 550                 }
 551                 head = rb_entry(node, struct btrfs_delayed_ref_head,
 552                                 href_node);
 553         }
 554
 555         head->processing = 1;
 556         WARN_ON(delayed_refs->num_heads_ready == 0);
 557         delayed_refs->num_heads_ready--;
 558         delayed_refs->run_delayed_start = head->bytenr +
 559                 head->num_bytes;
 560         return head;
 561 }
 562
 563 void btrfs_delete_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
 564                            struct btrfs_delayed_ref_head *head)
 565 {
 566         lockdep_assert_held(&delayed_refs->lock);
 567         lockdep_assert_held(&head->lock);
 568
 569         rb_erase_cached(&head->href_node, &delayed_refs->href_root);
 570         RB_CLEAR_NODE(&head->href_node);
 571         atomic_dec(&delayed_refs->num_entries);
 572         delayed_refs->num_heads--;
 573         if (head->processing == 0)
 574                 delayed_refs->num_heads_ready--;
 575 }
 576
 577 /*
 578  * Helper to insert the ref_node to the tail or merge with tail.
 579  *
 580  * Return 0 for insert.
 581  * Return >0 for merge.
 582  */
 583 static int insert_delayed_ref(struct btrfs_trans_handle *trans,
 584                               struct btrfs_delayed_ref_root *root,
 585                               struct btrfs_delayed_ref_head *href,
 586                               struct btrfs_delayed_ref_node *ref)
 587 {
 588         struct btrfs_delayed_ref_node *exist;
 589         int mod;
 590         int ret = 0;
 591
 592         spin_lock(&href->lock);
 593         exist = tree_insert(&href->ref_tree, ref);
 594         if (!exist)
 595                 goto inserted;
 596
 597         /* Now we are sure we can merge */
 598         ret = 1;
 599         if (exist->action == ref->action) {
 600                 mod = ref->ref_mod;
 601         } else {
 602                 /* Need to change action */
 603                 if (exist->ref_mod < ref->ref_mod) {
 604                         exist->action = ref->action;
 605                         mod = -exist->ref_mod;
 606                         exist->ref_mod = ref->ref_mod;
 607                         if (ref->action == BTRFS_ADD_DELAYED_REF)
 608                                 list_add_tail(&exist->add_list,
 609                                               &href->ref_add_list);
 610                         else if (ref->action == BTRFS_DROP_DELAYED_REF) {
 611                                 ASSERT(!list_empty(&exist->add_list));
 612                                 list_del(&exist->add_list);
 613                         } else {
 614                                 ASSERT(0);
 615                         }
 616                 } else
 617                         mod = -ref->ref_mod;
 618         }
 619         exist->ref_mod += mod;
 620
 621         /* remove existing tail if its ref_mod is zero */
 622         if (exist->ref_mod == 0)
 623                 drop_delayed_ref(trans, root, href, exist);
 624         spin_unlock(&href->lock);
 625         return ret;
 626 inserted:
 627         if (ref->action == BTRFS_ADD_DELAYED_REF)
 628                 list_add_tail(&ref->add_list, &href->ref_add_list);
 629         atomic_inc(&root->num_entries);
 630         spin_unlock(&href->lock);
 631         return ret;
 632 }
 633
 634 /*
 635  * helper function to update the accounting in the head ref
 636  * existing and update must have the same bytenr
 637  */
 638 static noinline void update_existing_head_ref(struct btrfs_trans_handle *trans,
 639                          struct btrfs_delayed_ref_head *existing,
 640                          struct btrfs_delayed_ref_head *update)
 641 {
 642         struct btrfs_delayed_ref_root *delayed_refs =
 643                 &trans->transaction->delayed_refs;
 644         struct btrfs_fs_info *fs_info = trans->fs_info;
 645         u64 flags = btrfs_ref_head_to_space_flags(existing);
 646         int old_ref_mod;
 647
 648         BUG_ON(existing->is_data != update->is_data);
 649
 650         spin_lock(&existing->lock);
 651         if (update->must_insert_reserved) {
 652                 /* if the extent was freed and then
 653                  * reallocated before the delayed ref
 654                  * entries were processed, we can end up
 655                  * with an existing head ref without
 656                  * the must_insert_reserved flag set.
 657                  * Set it again here
 658                  */
 659                 existing->must_insert_reserved = update->must_insert_reserved;
 660
 661                 /*
 662                  * update the num_bytes so we make sure the accounting
 663                  * is done correctly
 664                  */
 665                 existing->num_bytes = update->num_bytes;
 666
 667         }
 668
 669         if (update->extent_op) {
 670                 if (!existing->extent_op) {
 671                         existing->extent_op = update->extent_op;
 672                 } else {
 673                         if (update->extent_op->update_key) {
 674                                 memcpy(&existing->extent_op->key,
 675                                        &update->extent_op->key,
 676                                        sizeof(update->extent_op->key));
 677                                 existing->extent_op->update_key = true;
 678                         }
 679                         if (update->extent_op->update_flags) {
 680                                 existing->extent_op->flags_to_set |=
 681                                         update->extent_op->flags_to_set;
 682                                 existing->extent_op->update_flags = true;
 683                         }
 684                         btrfs_free_delayed_extent_op(update->extent_op);
 685                 }
 686         }
 687         /*
 688          * update the reference mod on the head to reflect this new operation,
 689          * only need the lock for this case cause we could be processing it
 690          * currently, for refs we just added we know we're a-ok.
 691          */
 692         old_ref_mod = existing->total_ref_mod;
 693         existing->ref_mod += update->ref_mod;
 694         existing->total_ref_mod += update->ref_mod;
 695
 696         /*
 697          * If we are going to from a positive ref mod to a negative or vice
 698          * versa we need to make sure to adjust pending_csums accordingly.
 699          */
 700         if (existing->is_data) {
 701                 u64 csum_leaves =
 702                         btrfs_csum_bytes_to_leaves(fs_info,
 703                                                    existing->num_bytes);
 704
 705                 if (existing->total_ref_mod >= 0 && old_ref_mod < 0) {
 706                         delayed_refs->pending_csums -= existing->num_bytes;
 707                         btrfs_delayed_refs_rsv_release(fs_info, csum_leaves);
 708                 }
 709                 if (existing->total_ref_mod < 0 && old_ref_mod >= 0) {
 710                         delayed_refs->pending_csums += existing->num_bytes;
 711                         trans->delayed_ref_updates += csum_leaves;
 712                 }
 713         }
 714
 715         /*
 716          * This handles the following conditions:
 717          *
 718          * 1. We had a ref mod of 0 or more and went negative, indicating that
 719          *    we may be freeing space, so add our space to the
 720          *    total_bytes_pinned counter.
 721          * 2. We were negative and went to 0 or positive, so no longer can say
 722          *    that the space would be pinned, decrement our counter from the
 723          *    total_bytes_pinned counter.
 724          * 3. We are now at 0 and have ->must_insert_reserved set, which means
 725          *    this was a new allocation and then we dropped it, and thus must
 726          *    add our space to the total_bytes_pinned counter.
 727          */
 728         if (existing->total_ref_mod < 0 && old_ref_mod >= 0)
 729                 btrfs_mod_total_bytes_pinned(fs_info, flags, existing->num_bytes);
 730         else if (existing->total_ref_mod >= 0 && old_ref_mod < 0)
 731                 btrfs_mod_total_bytes_pinned(fs_info, flags, -existing->num_bytes);
 732         else if (existing->total_ref_mod == 0 && existing->must_insert_reserved)
 733                 btrfs_mod_total_bytes_pinned(fs_info, flags, existing->num_bytes);
 734
 735         spin_unlock(&existing->lock);
 736 }
 737
 738 static void init_delayed_ref_head(struct btrfs_delayed_ref_head *head_ref,
 739                                   struct btrfs_qgroup_extent_record *qrecord,
 740                                   u64 bytenr, u64 num_bytes, u64 ref_root,
 741                                   u64 reserved, int action, bool is_data,
 742                                   bool is_system)
 743 {
 744         int count_mod = 1;
 745         int must_insert_reserved = 0;
 746
 747         /* If reserved is provided, it must be a data extent. */
 748         BUG_ON(!is_data && reserved);
 749
 750         /*
 751          * The head node stores the sum of all the mods, so dropping a ref
 752          * should drop the sum in the head node by one.
 753          */
 754         if (action == BTRFS_UPDATE_DELAYED_HEAD)
 755                 count_mod = 0;
 756         else if (action == BTRFS_DROP_DELAYED_REF)
 757                 count_mod = -1;
 758
 759         /*
 760          * BTRFS_ADD_DELAYED_EXTENT means that we need to update the reserved
 761          * accounting when the extent is finally added, or if a later
 762          * modification deletes the delayed ref without ever inserting the
 763          * extent into the extent allocation tree.  ref->must_insert_reserved
 764          * is the flag used to record that accounting mods are required.
 765          *
 766          * Once we record must_insert_reserved, switch the action to
 767          * BTRFS_ADD_DELAYED_REF because other special casing is not required.
 768          */
 769         if (action == BTRFS_ADD_DELAYED_EXTENT)
 770                 must_insert_reserved = 1;
 771         else
 772                 must_insert_reserved = 0;
 773
 774         refcount_set(&head_ref->refs, 1);
 775         head_ref->bytenr = bytenr;
 776         head_ref->num_bytes = num_bytes;
 777         head_ref->ref_mod = count_mod;
 778         head_ref->must_insert_reserved = must_insert_reserved;
 779         head_ref->is_data = is_data;
 780         head_ref->is_system = is_system;
 781         head_ref->ref_tree = RB_ROOT_CACHED;
 782         INIT_LIST_HEAD(&head_ref->ref_add_list);
 783         RB_CLEAR_NODE(&head_ref->href_node);
 784         head_ref->processing = 0;
 785         head_ref->total_ref_mod = count_mod;
 786         spin_lock_init(&head_ref->lock);
 787         mutex_init(&head_ref->mutex);
 788
 789         if (qrecord) {
 790                 if (ref_root && reserved) {
 791                         qrecord->data_rsv = reserved;
 792                         qrecord->data_rsv_refroot = ref_root;
 793                 }
 794                 qrecord->bytenr = bytenr;
 795                 qrecord->num_bytes = num_bytes;
 796                 qrecord->old_roots = NULL;
 797         }
 798 }
 799
 800 /*
 801  * helper function to actually insert a head node into the rbtree.
 802  * this does all the dirty work in terms of maintaining the correct
 803  * overall modification count.
 804  */
 805 static noinline struct btrfs_delayed_ref_head *
 806 add_delayed_ref_head(struct btrfs_trans_handle *trans,
 807                      struct btrfs_delayed_ref_head *head_ref,
 808                      struct btrfs_qgroup_extent_record *qrecord,
 809                      int action, int *qrecord_inserted_ret)
 810 {
 811         struct btrfs_delayed_ref_head *existing;
 812         struct btrfs_delayed_ref_root *delayed_refs;
 813         int qrecord_inserted = 0;
 814
 815         delayed_refs = &trans->transaction->delayed_refs;
 816
 817         /* Record qgroup extent info if provided */
 818         if (qrecord) {
 819                 if (btrfs_qgroup_trace_extent_nolock(trans->fs_info,
 820                                         delayed_refs, qrecord))
 821                         kfree(qrecord);
 822                 else
 823                         qrecord_inserted = 1;
 824         }
 825
 826         trace_add_delayed_ref_head(trans->fs_info, head_ref, action);
 827
 828         existing = htree_insert(&delayed_refs->href_root,
 829                                 &head_ref->href_node);
 830         if (existing) {
 831                 update_existing_head_ref(trans, existing, head_ref);
 832                 /*
 833                  * we've updated the existing ref, free the newly
 834                  * allocated ref
 835                  */
 836                 kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
 837                 head_ref = existing;
 838         } else {
 839                 u64 flags = btrfs_ref_head_to_space_flags(head_ref);
 840
 841                 if (head_ref->is_data && head_ref->ref_mod < 0) {
 842                         delayed_refs->pending_csums += head_ref->num_bytes;
 843                         trans->delayed_ref_updates +=
 844                                 btrfs_csum_bytes_to_leaves(trans->fs_info,
 845                                                            head_ref->num_bytes);
 846                 }
 847                 if (head_ref->ref_mod < 0)
 848                         btrfs_mod_total_bytes_pinned(trans->fs_info, flags,
 849                                                      head_ref->num_bytes);
 850                 delayed_refs->num_heads++;
 851                 delayed_refs->num_heads_ready++;
 852                 atomic_inc(&delayed_refs->num_entries);
 853                 trans->delayed_ref_updates++;
 854         }
 855         if (qrecord_inserted_ret)
 856                 *qrecord_inserted_ret = qrecord_inserted;
 857
 858         return head_ref;
 859 }
 860
 861 /*
 862  * init_delayed_ref_common - Initialize the structure which represents a
 863  *                           modification to a an extent.
 864  *
 865  * @fs_info:    Internal to the mounted filesystem mount structure.
 866  *
 867  * @ref:        The structure which is going to be initialized.
 868  *
 869  * @bytenr:     The logical address of the extent for which a modification is
 870  *              going to be recorded.
 871  *
 872  * @num_bytes:  Size of the extent whose modification is being recorded.
 873  *
 874  * @ref_root:   The id of the root where this modification has originated, this
 875  *              can be either one of the well-known metadata trees or the
 876  *              subvolume id which references this extent.
 877  *
 878  * @action:     Can be one of BTRFS_ADD_DELAYED_REF/BTRFS_DROP_DELAYED_REF or
 879  *              BTRFS_ADD_DELAYED_EXTENT
 880  *
 881  * @ref_type:   Holds the type of the extent which is being recorded, can be
 882  *              one of BTRFS_SHARED_BLOCK_REF_KEY/BTRFS_TREE_BLOCK_REF_KEY
 883  *              when recording a metadata extent or BTRFS_SHARED_DATA_REF_KEY/
 884  *              BTRFS_EXTENT_DATA_REF_KEY when recording data extent
 885  */
 886 static void init_delayed_ref_common(struct btrfs_fs_info *fs_info,
 887                                     struct btrfs_delayed_ref_node *ref,
 888                                     u64 bytenr, u64 num_bytes, u64 ref_root,
 889                                     int action, u8 ref_type)
 890 {
 891         u64 seq = 0;
 892
 893         if (action == BTRFS_ADD_DELAYED_EXTENT)
 894                 action = BTRFS_ADD_DELAYED_REF;
 895
 896         if (is_fstree(ref_root))
 897                 seq = atomic64_read(&fs_info->tree_mod_seq);
 898
 899         refcount_set(&ref->refs, 1);
 900         ref->bytenr = bytenr;
 901         ref->num_bytes = num_bytes;
 902         ref->ref_mod = 1;
 903         ref->action = action;
 904         ref->is_head = 0;
 905         ref->in_tree = 1;
 906         ref->seq = seq;
 907         ref->type = ref_type;
 908         RB_CLEAR_NODE(&ref->ref_node);
 909         INIT_LIST_HEAD(&ref->add_list);
 910 }
 911
 912 /*
 913  * add a delayed tree ref.  This does all of the accounting required
 914  * to make sure the delayed ref is eventually processed before this
 915  * transaction commits.
 916  */
 917 int btrfs_add_delayed_tree_ref(struct btrfs_trans_handle *trans,
 918                                struct btrfs_ref *generic_ref,
 919                                struct btrfs_delayed_extent_op *extent_op)
 920 {
 921         struct btrfs_fs_info *fs_info = trans->fs_info;
 922         struct btrfs_delayed_tree_ref *ref;
 923         struct btrfs_delayed_ref_head *head_ref;
 924         struct btrfs_delayed_ref_root *delayed_refs;
 925         struct btrfs_qgroup_extent_record *record = NULL;
 926         int qrecord_inserted;
 927         bool is_system;
 928         int action = generic_ref->action;
 929         int level = generic_ref->tree_ref.level;
 930         int ret;
 931         u64 bytenr = generic_ref->bytenr;
 932         u64 num_bytes = generic_ref->len;
 933         u64 parent = generic_ref->parent;
 934         u8 ref_type;
 935
 936         is_system = (generic_ref->real_root == BTRFS_CHUNK_TREE_OBJECTID);
 937
 938         ASSERT(generic_ref->type == BTRFS_REF_METADATA && generic_ref->action);
 939         BUG_ON(extent_op && extent_op->is_data);
 940         ref = kmem_cache_alloc(btrfs_delayed_tree_ref_cachep, GFP_NOFS);
 941         if (!ref)
 942                 return -ENOMEM;
 943
 944         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
 945         if (!head_ref) {
 946                 kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
 947                 return -ENOMEM;
 948         }
 949
 950         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) &&
 951             is_fstree(generic_ref->real_root) &&
 952             is_fstree(generic_ref->tree_ref.root) &&
 953             !generic_ref->skip_qgroup) {
 954                 record = kzalloc(sizeof(*record), GFP_NOFS);
 955                 if (!record) {
 956                         kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
 957                         kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
 958                         return -ENOMEM;
 959                 }
 960         }
 961
 962         if (parent)
 963                 ref_type = BTRFS_SHARED_BLOCK_REF_KEY;
 964         else
 965                 ref_type = BTRFS_TREE_BLOCK_REF_KEY;
 966
 967         init_delayed_ref_common(fs_info, &ref->node, bytenr, num_bytes,
 968                                 generic_ref->tree_ref.root, action, ref_type);
 969         ref->root = generic_ref->tree_ref.root;
 970         ref->parent = parent;
 971         ref->level = level;
 972
 973         init_delayed_ref_head(head_ref, record, bytenr, num_bytes,
 974                               generic_ref->tree_ref.root, 0, action, false,
 975                               is_system);
 976         head_ref->extent_op = extent_op;
 977
 978         delayed_refs = &trans->transaction->delayed_refs;
 979         spin_lock(&delayed_refs->lock);
 980
 981         /*
 982          * insert both the head node and the new ref without dropping
 983          * the spin lock
 984          */
 985         head_ref = add_delayed_ref_head(trans, head_ref, record,
 986                                         action, &qrecord_inserted);
 987
 988         ret = insert_delayed_ref(trans, delayed_refs, head_ref, &ref->node);
 989         spin_unlock(&delayed_refs->lock);
 990
 991         /*
 992          * Need to update the delayed_refs_rsv with any changes we may have
 993          * made.
 994          */
 995         btrfs_update_delayed_refs_rsv(trans);
 996
 997         trace_add_delayed_tree_ref(fs_info, &ref->node, ref,
 998                                    action == BTRFS_ADD_DELAYED_EXTENT ?
 999                                    BTRFS_ADD_DELAYED_REF : action);
1000         if (ret > 0)
1001                 kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
1002
1003         if (qrecord_inserted)
1004                 btrfs_qgroup_trace_extent_post(fs_info, record);
1005
1006         return 0;
1007 }
1008
1009 /*
1010  * add a delayed data ref. it's similar to btrfs_add_delayed_tree_ref.
1011  */
1012 int btrfs_add_delayed_data_ref(struct btrfs_trans_handle *trans,
1013                                struct btrfs_ref *generic_ref,
1014                                u64 reserved)
1015 {
1016         struct btrfs_fs_info *fs_info = trans->fs_info;
1017         struct btrfs_delayed_data_ref *ref;
1018         struct btrfs_delayed_ref_head *head_ref;
1019         struct btrfs_delayed_ref_root *delayed_refs;
1020         struct btrfs_qgroup_extent_record *record = NULL;
1021         int qrecord_inserted;
1022         int action = generic_ref->action;
1023         int ret;
1024         u64 bytenr = generic_ref->bytenr;
1025         u64 num_bytes = generic_ref->len;
1026         u64 parent = generic_ref->parent;
1027         u64 ref_root = generic_ref->data_ref.ref_root;
1028         u64 owner = generic_ref->data_ref.ino;
1029         u64 offset = generic_ref->data_ref.offset;
1030         u8 ref_type;
1031
1032         ASSERT(generic_ref->type == BTRFS_REF_DATA && action);
1033         ref = kmem_cache_alloc(btrfs_delayed_data_ref_cachep, GFP_NOFS);
1034         if (!ref)
1035                 return -ENOMEM;
1036
1037         if (parent)
1038                 ref_type = BTRFS_SHARED_DATA_REF_KEY;
1039         else
1040                 ref_type = BTRFS_EXTENT_DATA_REF_KEY;
1041         init_delayed_ref_common(fs_info, &ref->node, bytenr, num_bytes,
1042                                 ref_root, action, ref_type);
1043         ref->root = ref_root;
1044         ref->parent = parent;
1045         ref->objectid = owner;
1046         ref->offset = offset;
1047
1048
1049         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
1050         if (!head_ref) {
1051                 kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
1052                 return -ENOMEM;
1053         }
1054
1055         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) &&
1056             is_fstree(ref_root) &&
1057             is_fstree(generic_ref->real_root) &&
1058             !generic_ref->skip_qgroup) {
1059                 record = kzalloc(sizeof(*record), GFP_NOFS);
1060                 if (!record) {
1061                         kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
1062                         kmem_cache_free(btrfs_delayed_ref_head_cachep,
1063                                         head_ref);
1064                         return -ENOMEM;
1065                 }
1066         }
1067
1068         init_delayed_ref_head(head_ref, record, bytenr, num_bytes, ref_root,
1069                               reserved, action, true, false);
1070         head_ref->extent_op = NULL;
1071
1072         delayed_refs = &trans->transaction->delayed_refs;
1073         spin_lock(&delayed_refs->lock);
1074
1075         /*
1076          * insert both the head node and the new ref without dropping
1077          * the spin lock
1078          */
1079         head_ref = add_delayed_ref_head(trans, head_ref, record,
1080                                         action, &qrecord_inserted);
1081
1082         ret = insert_delayed_ref(trans, delayed_refs, head_ref, &ref->node);
1083         spin_unlock(&delayed_refs->lock);
1084
1085         /*
1086          * Need to update the delayed_refs_rsv with any changes we may have
1087          * made.
1088          */
1089         btrfs_update_delayed_refs_rsv(trans);
1090
1091         trace_add_delayed_data_ref(trans->fs_info, &ref->node, ref,
1092                                    action == BTRFS_ADD_DELAYED_EXTENT ?
1093                                    BTRFS_ADD_DELAYED_REF : action);
1094         if (ret > 0)
1095                 kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
1096
1097
1098         if (qrecord_inserted)
1099                 return btrfs_qgroup_trace_extent_post(fs_info, record);
1100         return 0;
1101 }
1102
1103 int btrfs_add_delayed_extent_op(struct btrfs_trans_handle *trans,
1104                                 u64 bytenr, u64 num_bytes,
1105                                 struct btrfs_delayed_extent_op *extent_op)
1106 {
1107         struct btrfs_delayed_ref_head *head_ref;
1108         struct btrfs_delayed_ref_root *delayed_refs;
1109
1110         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
1111         if (!head_ref)
1112                 return -ENOMEM;
1113
1114         init_delayed_ref_head(head_ref, NULL, bytenr, num_bytes, 0, 0,
1115                               BTRFS_UPDATE_DELAYED_HEAD, extent_op->is_data,
1116                               false);
1117         head_ref->extent_op = extent_op;
1118
1119         delayed_refs = &trans->transaction->delayed_refs;
1120         spin_lock(&delayed_refs->lock);
1121
1122         add_delayed_ref_head(trans, head_ref, NULL, BTRFS_UPDATE_DELAYED_HEAD,
1123                              NULL);
1124
1125         spin_unlock(&delayed_refs->lock);
1126
1127         /*
1128          * Need to update the delayed_refs_rsv with any changes we may have
1129          * made.
1130          */
1131         btrfs_update_delayed_refs_rsv(trans);
1132         return 0;
1133 }
1134
1135 /*
1136  * This does a simple search for the head node for a given extent.  Returns the
1137  * head node if found, or NULL if not.
1138  */
1139 struct btrfs_delayed_ref_head *
1140 btrfs_find_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs, u64 bytenr)
1141 {
1142         lockdep_assert_held(&delayed_refs->lock);
1143
1144         return find_ref_head(delayed_refs, bytenr, false);
1145 }
1146
1147 void __cold btrfs_delayed_ref_exit(void)
1148 {
1149         kmem_cache_destroy(btrfs_delayed_ref_head_cachep);
1150         kmem_cache_destroy(btrfs_delayed_tree_ref_cachep);
1151         kmem_cache_destroy(btrfs_delayed_data_ref_cachep);
1152         kmem_cache_destroy(btrfs_delayed_extent_op_cachep);
1153 }
1154
1155 int __init btrfs_delayed_ref_init(void)
1156 {
1157         btrfs_delayed_ref_head_cachep = kmem_cache_create(
1158                                 "btrfs_delayed_ref_head",
1159                                 sizeof(struct btrfs_delayed_ref_head), 0,
1160                                 SLAB_MEM_SPREAD, NULL);
1161         if (!btrfs_delayed_ref_head_cachep)
1162                 goto fail;
1163
1164         btrfs_delayed_tree_ref_cachep = kmem_cache_create(
1165                                 "btrfs_delayed_tree_ref",
1166                                 sizeof(struct btrfs_delayed_tree_ref), 0,
1167                                 SLAB_MEM_SPREAD, NULL);
1168         if (!btrfs_delayed_tree_ref_cachep)
1169                 goto fail;
1170
1171         btrfs_delayed_data_ref_cachep = kmem_cache_create(
1172                                 "btrfs_delayed_data_ref",
1173                                 sizeof(struct btrfs_delayed_data_ref), 0,
1174                                 SLAB_MEM_SPREAD, NULL);
1175         if (!btrfs_delayed_data_ref_cachep)
1176                 goto fail;
1177
1178         btrfs_delayed_extent_op_cachep = kmem_cache_create(
1179                                 "btrfs_delayed_extent_op",
1180                                 sizeof(struct btrfs_delayed_extent_op), 0,
1181                                 SLAB_MEM_SPREAD, NULL);
1182         if (!btrfs_delayed_extent_op_cachep)
1183                 goto fail;
1184
1185         return 0;
1186 fail:
1187         btrfs_delayed_ref_exit();
1188         return -ENOMEM;
1189 }