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
  14 struct kmem_cache *btrfs_delayed_ref_head_cachep;
  15 struct kmem_cache *btrfs_delayed_tree_ref_cachep;
  16 struct kmem_cache *btrfs_delayed_data_ref_cachep;
  17 struct kmem_cache *btrfs_delayed_extent_op_cachep;
  18 /*
  19  * delayed back reference update tracking.  For subvolume trees
  20  * we queue up extent allocations and backref maintenance for
  21  * delayed processing.   This avoids deep call chains where we
  22  * add extents in the middle of btrfs_search_slot, and it allows
  23  * us to buffer up frequently modified backrefs in an rb tree instead
  24  * of hammering updates on the extent allocation tree.
  25  */
  26
  27 /*
  28  * compare two delayed tree backrefs with same bytenr and type
  29  */
  30 static int comp_tree_refs(struct btrfs_delayed_tree_ref *ref1,
  31                           struct btrfs_delayed_tree_ref *ref2)
  32 {
  33         if (ref1->node.type == BTRFS_TREE_BLOCK_REF_KEY) {
  34                 if (ref1->root < ref2->root)
  35                         return -1;
  36                 if (ref1->root > ref2->root)
  37                         return 1;
  38         } else {
  39                 if (ref1->parent < ref2->parent)
  40                         return -1;
  41                 if (ref1->parent > ref2->parent)
  42                         return 1;
  43         }
  44         return 0;
  45 }
  46
  47 /*
  48  * compare two delayed data backrefs with same bytenr and type
  49  */
  50 static int comp_data_refs(struct btrfs_delayed_data_ref *ref1,
  51                           struct btrfs_delayed_data_ref *ref2)
  52 {
  53         if (ref1->node.type == BTRFS_EXTENT_DATA_REF_KEY) {
  54                 if (ref1->root < ref2->root)
  55                         return -1;
  56                 if (ref1->root > ref2->root)
  57                         return 1;
  58                 if (ref1->objectid < ref2->objectid)
  59                         return -1;
  60                 if (ref1->objectid > ref2->objectid)
  61                         return 1;
  62                 if (ref1->offset < ref2->offset)
  63                         return -1;
  64                 if (ref1->offset > ref2->offset)
  65                         return 1;
  66         } else {
  67                 if (ref1->parent < ref2->parent)
  68                         return -1;
  69                 if (ref1->parent > ref2->parent)
  70                         return 1;
  71         }
  72         return 0;
  73 }
  74
  75 static int comp_refs(struct btrfs_delayed_ref_node *ref1,
  76                      struct btrfs_delayed_ref_node *ref2,
  77                      bool check_seq)
  78 {
  79         int ret = 0;
  80
  81         if (ref1->type < ref2->type)
  82                 return -1;
  83         if (ref1->type > ref2->type)
  84                 return 1;
  85         if (ref1->type == BTRFS_TREE_BLOCK_REF_KEY ||
  86             ref1->type == BTRFS_SHARED_BLOCK_REF_KEY)
  87                 ret = comp_tree_refs(btrfs_delayed_node_to_tree_ref(ref1),
  88                                      btrfs_delayed_node_to_tree_ref(ref2));
  89         else
  90                 ret = comp_data_refs(btrfs_delayed_node_to_data_ref(ref1),
  91                                      btrfs_delayed_node_to_data_ref(ref2));
  92         if (ret)
  93                 return ret;
  94         if (check_seq) {
  95                 if (ref1->seq < ref2->seq)
  96                         return -1;
  97                 if (ref1->seq > ref2->seq)
  98                         return 1;
  99         }
 100         return 0;
 101 }
 102
 103 /* insert a new ref to head ref rbtree */
 104 static struct btrfs_delayed_ref_head *htree_insert(struct rb_root *root,
 105                                                    struct rb_node *node)
 106 {
 107         struct rb_node **p = &root->rb_node;
 108         struct rb_node *parent_node = NULL;
 109         struct btrfs_delayed_ref_head *entry;
 110         struct btrfs_delayed_ref_head *ins;
 111         u64 bytenr;
 112
 113         ins = rb_entry(node, struct btrfs_delayed_ref_head, href_node);
 114         bytenr = ins->bytenr;
 115         while (*p) {
 116                 parent_node = *p;
 117                 entry = rb_entry(parent_node, struct btrfs_delayed_ref_head,
 118                                  href_node);
 119
 120                 if (bytenr < entry->bytenr)
 121                         p = &(*p)->rb_left;
 122                 else if (bytenr > entry->bytenr)
 123                         p = &(*p)->rb_right;
 124                 else
 125                         return entry;
 126         }
 127
 128         rb_link_node(node, parent_node, p);
 129         rb_insert_color(node, root);
 130         return NULL;
 131 }
 132
 133 static struct btrfs_delayed_ref_node* tree_insert(struct rb_root *root,
 134                 struct btrfs_delayed_ref_node *ins)
 135 {
 136         struct rb_node **p = &root->rb_node;
 137         struct rb_node *node = &ins->ref_node;
 138         struct rb_node *parent_node = NULL;
 139         struct btrfs_delayed_ref_node *entry;
 140
 141         while (*p) {
 142                 int comp;
 143
 144                 parent_node = *p;
 145                 entry = rb_entry(parent_node, struct btrfs_delayed_ref_node,
 146                                  ref_node);
 147                 comp = comp_refs(ins, entry, true);
 148                 if (comp < 0)
 149                         p = &(*p)->rb_left;
 150                 else if (comp > 0)
 151                         p = &(*p)->rb_right;
 152                 else
 153                         return entry;
 154         }
 155
 156         rb_link_node(node, parent_node, p);
 157         rb_insert_color(node, root);
 158         return NULL;
 159 }
 160
 161 /*
 162  * find an head entry based on bytenr. This returns the delayed ref
 163  * head if it was able to find one, or NULL if nothing was in that spot.
 164  * If return_bigger is given, the next bigger entry is returned if no exact
 165  * match is found.
 166  */
 167 static struct btrfs_delayed_ref_head *
 168 find_ref_head(struct rb_root *root, u64 bytenr,
 169               int return_bigger)
 170 {
 171         struct rb_node *n;
 172         struct btrfs_delayed_ref_head *entry;
 173
 174         n = root->rb_node;
 175         entry = NULL;
 176         while (n) {
 177                 entry = rb_entry(n, struct btrfs_delayed_ref_head, href_node);
 178
 179                 if (bytenr < entry->bytenr)
 180                         n = n->rb_left;
 181                 else if (bytenr > entry->bytenr)
 182                         n = n->rb_right;
 183                 else
 184                         return entry;
 185         }
 186         if (entry && return_bigger) {
 187                 if (bytenr > entry->bytenr) {
 188                         n = rb_next(&entry->href_node);
 189                         if (!n)
 190                                 n = rb_first(root);
 191                         entry = rb_entry(n, struct btrfs_delayed_ref_head,
 192                                          href_node);
 193                         return entry;
 194                 }
 195                 return entry;
 196         }
 197         return NULL;
 198 }
 199
 200 int btrfs_delayed_ref_lock(struct btrfs_trans_handle *trans,
 201                            struct btrfs_delayed_ref_head *head)
 202 {
 203         struct btrfs_delayed_ref_root *delayed_refs;
 204
 205         delayed_refs = &trans->transaction->delayed_refs;
 206         lockdep_assert_held(&delayed_refs->lock);
 207         if (mutex_trylock(&head->mutex))
 208                 return 0;
 209
 210         refcount_inc(&head->refs);
 211         spin_unlock(&delayed_refs->lock);
 212
 213         mutex_lock(&head->mutex);
 214         spin_lock(&delayed_refs->lock);
 215         if (RB_EMPTY_NODE(&head->href_node)) {
 216                 mutex_unlock(&head->mutex);
 217                 btrfs_put_delayed_ref_head(head);
 218                 return -EAGAIN;
 219         }
 220         btrfs_put_delayed_ref_head(head);
 221         return 0;
 222 }
 223
 224 static inline void drop_delayed_ref(struct btrfs_trans_handle *trans,
 225                                     struct btrfs_delayed_ref_root *delayed_refs,
 226                                     struct btrfs_delayed_ref_head *head,
 227                                     struct btrfs_delayed_ref_node *ref)
 228 {
 229         lockdep_assert_held(&head->lock);
 230         rb_erase(&ref->ref_node, &head->ref_tree);
 231         RB_CLEAR_NODE(&ref->ref_node);
 232         if (!list_empty(&ref->add_list))
 233                 list_del(&ref->add_list);
 234         ref->in_tree = 0;
 235         btrfs_put_delayed_ref(ref);
 236         atomic_dec(&delayed_refs->num_entries);
 237 }
 238
 239 static bool merge_ref(struct btrfs_trans_handle *trans,
 240                       struct btrfs_delayed_ref_root *delayed_refs,
 241                       struct btrfs_delayed_ref_head *head,
 242                       struct btrfs_delayed_ref_node *ref,
 243                       u64 seq)
 244 {
 245         struct btrfs_delayed_ref_node *next;
 246         struct rb_node *node = rb_next(&ref->ref_node);
 247         bool done = false;
 248
 249         while (!done && node) {
 250                 int mod;
 251
 252                 next = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
 253                 node = rb_next(node);
 254                 if (seq && next->seq >= seq)
 255                         break;
 256                 if (comp_refs(ref, next, false))
 257                         break;
 258
 259                 if (ref->action == next->action) {
 260                         mod = next->ref_mod;
 261                 } else {
 262                         if (ref->ref_mod < next->ref_mod) {
 263                                 swap(ref, next);
 264                                 done = true;
 265                         }
 266                         mod = -next->ref_mod;
 267                 }
 268
 269                 drop_delayed_ref(trans, delayed_refs, head, next);
 270                 ref->ref_mod += mod;
 271                 if (ref->ref_mod == 0) {
 272                         drop_delayed_ref(trans, delayed_refs, head, ref);
 273                         done = true;
 274                 } else {
 275                         /*
 276                          * Can't have multiples of the same ref on a tree block.
 277                          */
 278                         WARN_ON(ref->type == BTRFS_TREE_BLOCK_REF_KEY ||
 279                                 ref->type == BTRFS_SHARED_BLOCK_REF_KEY);
 280                 }
 281         }
 282
 283         return done;
 284 }
 285
 286 void btrfs_merge_delayed_refs(struct btrfs_trans_handle *trans,
 287                               struct btrfs_delayed_ref_root *delayed_refs,
 288                               struct btrfs_delayed_ref_head *head)
 289 {
 290         struct btrfs_fs_info *fs_info = trans->fs_info;
 291         struct btrfs_delayed_ref_node *ref;
 292         struct rb_node *node;
 293         u64 seq = 0;
 294
 295         lockdep_assert_held(&head->lock);
 296
 297         if (RB_EMPTY_ROOT(&head->ref_tree))
 298                 return;
 299
 300         /* We don't have too many refs to merge for data. */
 301         if (head->is_data)
 302                 return;
 303
 304         read_lock(&fs_info->tree_mod_log_lock);
 305         if (!list_empty(&fs_info->tree_mod_seq_list)) {
 306                 struct seq_list *elem;
 307
 308                 elem = list_first_entry(&fs_info->tree_mod_seq_list,
 309                                         struct seq_list, list);
 310                 seq = elem->seq;
 311         }
 312         read_unlock(&fs_info->tree_mod_log_lock);
 313
 314 again:
 315         for (node = rb_first(&head->ref_tree); node; node = rb_next(node)) {
 316                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
 317                 if (seq && ref->seq >= seq)
 318                         continue;
 319                 if (merge_ref(trans, delayed_refs, head, ref, seq))
 320                         goto again;
 321         }
 322 }
 323
 324 int btrfs_check_delayed_seq(struct btrfs_fs_info *fs_info, u64 seq)
 325 {
 326         struct seq_list *elem;
 327         int ret = 0;
 328
 329         read_lock(&fs_info->tree_mod_log_lock);
 330         if (!list_empty(&fs_info->tree_mod_seq_list)) {
 331                 elem = list_first_entry(&fs_info->tree_mod_seq_list,
 332                                         struct seq_list, list);
 333                 if (seq >= elem->seq) {
 334                         btrfs_debug(fs_info,
 335                                 "holding back delayed_ref %#x.%x, lowest is %#x.%x",
 336                                 (u32)(seq >> 32), (u32)seq,
 337                                 (u32)(elem->seq >> 32), (u32)elem->seq);
 338                         ret = 1;
 339                 }
 340         }
 341
 342         read_unlock(&fs_info->tree_mod_log_lock);
 343         return ret;
 344 }
 345
 346 struct btrfs_delayed_ref_head *
 347 btrfs_select_ref_head(struct btrfs_trans_handle *trans)
 348 {
 349         struct btrfs_delayed_ref_root *delayed_refs;
 350         struct btrfs_delayed_ref_head *head;
 351         u64 start;
 352         bool loop = false;
 353
 354         delayed_refs = &trans->transaction->delayed_refs;
 355
 356 again:
 357         start = delayed_refs->run_delayed_start;
 358         head = find_ref_head(&delayed_refs->href_root, start, 1);
 359         if (!head && !loop) {
 360                 delayed_refs->run_delayed_start = 0;
 361                 start = 0;
 362                 loop = true;
 363                 head = find_ref_head(&delayed_refs->href_root, start, 1);
 364                 if (!head)
 365                         return NULL;
 366         } else if (!head && loop) {
 367                 return NULL;
 368         }
 369
 370         while (head->processing) {
 371                 struct rb_node *node;
 372
 373                 node = rb_next(&head->href_node);
 374                 if (!node) {
 375                         if (loop)
 376                                 return NULL;
 377                         delayed_refs->run_delayed_start = 0;
 378                         start = 0;
 379                         loop = true;
 380                         goto again;
 381                 }
 382                 head = rb_entry(node, struct btrfs_delayed_ref_head,
 383                                 href_node);
 384         }
 385
 386         head->processing = 1;
 387         WARN_ON(delayed_refs->num_heads_ready == 0);
 388         delayed_refs->num_heads_ready--;
 389         delayed_refs->run_delayed_start = head->bytenr +
 390                 head->num_bytes;
 391         return head;
 392 }
 393
 394 /*
 395  * Helper to insert the ref_node to the tail or merge with tail.
 396  *
 397  * Return 0 for insert.
 398  * Return >0 for merge.
 399  */
 400 static int insert_delayed_ref(struct btrfs_trans_handle *trans,
 401                               struct btrfs_delayed_ref_root *root,
 402                               struct btrfs_delayed_ref_head *href,
 403                               struct btrfs_delayed_ref_node *ref)
 404 {
 405         struct btrfs_delayed_ref_node *exist;
 406         int mod;
 407         int ret = 0;
 408
 409         spin_lock(&href->lock);
 410         exist = tree_insert(&href->ref_tree, ref);
 411         if (!exist)
 412                 goto inserted;
 413
 414         /* Now we are sure we can merge */
 415         ret = 1;
 416         if (exist->action == ref->action) {
 417                 mod = ref->ref_mod;
 418         } else {
 419                 /* Need to change action */
 420                 if (exist->ref_mod < ref->ref_mod) {
 421                         exist->action = ref->action;
 422                         mod = -exist->ref_mod;
 423                         exist->ref_mod = ref->ref_mod;
 424                         if (ref->action == BTRFS_ADD_DELAYED_REF)
 425                                 list_add_tail(&exist->add_list,
 426                                               &href->ref_add_list);
 427                         else if (ref->action == BTRFS_DROP_DELAYED_REF) {
 428                                 ASSERT(!list_empty(&exist->add_list));
 429                                 list_del(&exist->add_list);
 430                         } else {
 431                                 ASSERT(0);
 432                         }
 433                 } else
 434                         mod = -ref->ref_mod;
 435         }
 436         exist->ref_mod += mod;
 437
 438         /* remove existing tail if its ref_mod is zero */
 439         if (exist->ref_mod == 0)
 440                 drop_delayed_ref(trans, root, href, exist);
 441         spin_unlock(&href->lock);
 442         return ret;
 443 inserted:
 444         if (ref->action == BTRFS_ADD_DELAYED_REF)
 445                 list_add_tail(&ref->add_list, &href->ref_add_list);
 446         atomic_inc(&root->num_entries);
 447         spin_unlock(&href->lock);
 448         return ret;
 449 }
 450
 451 /*
 452  * helper function to update the accounting in the head ref
 453  * existing and update must have the same bytenr
 454  */
 455 static noinline void
 456 update_existing_head_ref(struct btrfs_delayed_ref_root *delayed_refs,
 457                          struct btrfs_delayed_ref_head *existing,
 458                          struct btrfs_delayed_ref_head *update,
 459                          int *old_ref_mod_ret)
 460 {
 461         int old_ref_mod;
 462
 463         BUG_ON(existing->is_data != update->is_data);
 464
 465         spin_lock(&existing->lock);
 466         if (update->must_insert_reserved) {
 467                 /* if the extent was freed and then
 468                  * reallocated before the delayed ref
 469                  * entries were processed, we can end up
 470                  * with an existing head ref without
 471                  * the must_insert_reserved flag set.
 472                  * Set it again here
 473                  */
 474                 existing->must_insert_reserved = update->must_insert_reserved;
 475
 476                 /*
 477                  * update the num_bytes so we make sure the accounting
 478                  * is done correctly
 479                  */
 480                 existing->num_bytes = update->num_bytes;
 481
 482         }
 483
 484         if (update->extent_op) {
 485                 if (!existing->extent_op) {
 486                         existing->extent_op = update->extent_op;
 487                 } else {
 488                         if (update->extent_op->update_key) {
 489                                 memcpy(&existing->extent_op->key,
 490                                        &update->extent_op->key,
 491                                        sizeof(update->extent_op->key));
 492                                 existing->extent_op->update_key = true;
 493                         }
 494                         if (update->extent_op->update_flags) {
 495                                 existing->extent_op->flags_to_set |=
 496                                         update->extent_op->flags_to_set;
 497                                 existing->extent_op->update_flags = true;
 498                         }
 499                         btrfs_free_delayed_extent_op(update->extent_op);
 500                 }
 501         }
 502         /*
 503          * update the reference mod on the head to reflect this new operation,
 504          * only need the lock for this case cause we could be processing it
 505          * currently, for refs we just added we know we're a-ok.
 506          */
 507         old_ref_mod = existing->total_ref_mod;
 508         if (old_ref_mod_ret)
 509                 *old_ref_mod_ret = old_ref_mod;
 510         existing->ref_mod += update->ref_mod;
 511         existing->total_ref_mod += update->ref_mod;
 512
 513         /*
 514          * If we are going to from a positive ref mod to a negative or vice
 515          * versa we need to make sure to adjust pending_csums accordingly.
 516          */
 517         if (existing->is_data) {
 518                 if (existing->total_ref_mod >= 0 && old_ref_mod < 0)
 519                         delayed_refs->pending_csums -= existing->num_bytes;
 520                 if (existing->total_ref_mod < 0 && old_ref_mod >= 0)
 521                         delayed_refs->pending_csums += existing->num_bytes;
 522         }
 523         spin_unlock(&existing->lock);
 524 }
 525
 526 static void init_delayed_ref_head(struct btrfs_delayed_ref_head *head_ref,
 527                                   struct btrfs_qgroup_extent_record *qrecord,
 528                                   u64 bytenr, u64 num_bytes, u64 ref_root,
 529                                   u64 reserved, int action, bool is_data,
 530                                   bool is_system)
 531 {
 532         int count_mod = 1;
 533         int must_insert_reserved = 0;
 534
 535         /* If reserved is provided, it must be a data extent. */
 536         BUG_ON(!is_data && reserved);
 537
 538         /*
 539          * The head node stores the sum of all the mods, so dropping a ref
 540          * should drop the sum in the head node by one.
 541          */
 542         if (action == BTRFS_UPDATE_DELAYED_HEAD)
 543                 count_mod = 0;
 544         else if (action == BTRFS_DROP_DELAYED_REF)
 545                 count_mod = -1;
 546
 547         /*
 548          * BTRFS_ADD_DELAYED_EXTENT means that we need to update the reserved
 549          * accounting when the extent is finally added, or if a later
 550          * modification deletes the delayed ref without ever inserting the
 551          * extent into the extent allocation tree.  ref->must_insert_reserved
 552          * is the flag used to record that accounting mods are required.
 553          *
 554          * Once we record must_insert_reserved, switch the action to
 555          * BTRFS_ADD_DELAYED_REF because other special casing is not required.
 556          */
 557         if (action == BTRFS_ADD_DELAYED_EXTENT)
 558                 must_insert_reserved = 1;
 559         else
 560                 must_insert_reserved = 0;
 561
 562         refcount_set(&head_ref->refs, 1);
 563         head_ref->bytenr = bytenr;
 564         head_ref->num_bytes = num_bytes;
 565         head_ref->ref_mod = count_mod;
 566         head_ref->must_insert_reserved = must_insert_reserved;
 567         head_ref->is_data = is_data;
 568         head_ref->is_system = is_system;
 569         head_ref->ref_tree = RB_ROOT;
 570         INIT_LIST_HEAD(&head_ref->ref_add_list);
 571         RB_CLEAR_NODE(&head_ref->href_node);
 572         head_ref->processing = 0;
 573         head_ref->total_ref_mod = count_mod;
 574         head_ref->qgroup_reserved = 0;
 575         head_ref->qgroup_ref_root = 0;
 576         spin_lock_init(&head_ref->lock);
 577         mutex_init(&head_ref->mutex);
 578
 579         if (qrecord) {
 580                 if (ref_root && reserved) {
 581                         head_ref->qgroup_ref_root = ref_root;
 582                         head_ref->qgroup_reserved = reserved;
 583                 }
 584
 585                 qrecord->bytenr = bytenr;
 586                 qrecord->num_bytes = num_bytes;
 587                 qrecord->old_roots = NULL;
 588         }
 589 }
 590
 591 /*
 592  * helper function to actually insert a head node into the rbtree.
 593  * this does all the dirty work in terms of maintaining the correct
 594  * overall modification count.
 595  */
 596 static noinline struct btrfs_delayed_ref_head *
 597 add_delayed_ref_head(struct btrfs_trans_handle *trans,
 598                      struct btrfs_delayed_ref_head *head_ref,
 599                      struct btrfs_qgroup_extent_record *qrecord,
 600                      int action, int *qrecord_inserted_ret,
 601                      int *old_ref_mod, int *new_ref_mod)
 602 {
 603         struct btrfs_delayed_ref_head *existing;
 604         struct btrfs_delayed_ref_root *delayed_refs;
 605         int qrecord_inserted = 0;
 606
 607         delayed_refs = &trans->transaction->delayed_refs;
 608
 609         /* Record qgroup extent info if provided */
 610         if (qrecord) {
 611                 if (btrfs_qgroup_trace_extent_nolock(trans->fs_info,
 612                                         delayed_refs, qrecord))
 613                         kfree(qrecord);
 614                 else
 615                         qrecord_inserted = 1;
 616         }
 617
 618         trace_add_delayed_ref_head(trans->fs_info, head_ref, action);
 619
 620         existing = htree_insert(&delayed_refs->href_root,
 621                                 &head_ref->href_node);
 622         if (existing) {
 623                 WARN_ON(qrecord && head_ref->qgroup_ref_root
 624                         && head_ref->qgroup_reserved
 625                         && existing->qgroup_ref_root
 626                         && existing->qgroup_reserved);
 627                 update_existing_head_ref(delayed_refs, existing, head_ref,
 628                                          old_ref_mod);
 629                 /*
 630                  * we've updated the existing ref, free the newly
 631                  * allocated ref
 632                  */
 633                 kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
 634                 head_ref = existing;
 635         } else {
 636                 if (old_ref_mod)
 637                         *old_ref_mod = 0;
 638                 if (head_ref->is_data && head_ref->ref_mod < 0)
 639                         delayed_refs->pending_csums += head_ref->num_bytes;
 640                 delayed_refs->num_heads++;
 641                 delayed_refs->num_heads_ready++;
 642                 atomic_inc(&delayed_refs->num_entries);
 643                 trans->delayed_ref_updates++;
 644         }
 645         if (qrecord_inserted_ret)
 646                 *qrecord_inserted_ret = qrecord_inserted;
 647         if (new_ref_mod)
 648                 *new_ref_mod = head_ref->total_ref_mod;
 649
 650         return head_ref;
 651 }
 652
 653 /*
 654  * init_delayed_ref_common - Initialize the structure which represents a
 655  *                           modification to a an extent.
 656  *
 657  * @fs_info:    Internal to the mounted filesystem mount structure.
 658  *
 659  * @ref:        The structure which is going to be initialized.
 660  *
 661  * @bytenr:     The logical address of the extent for which a modification is
 662  *              going to be recorded.
 663  *
 664  * @num_bytes:  Size of the extent whose modification is being recorded.
 665  *
 666  * @ref_root:   The id of the root where this modification has originated, this
 667  *              can be either one of the well-known metadata trees or the
 668  *              subvolume id which references this extent.
 669  *
 670  * @action:     Can be one of BTRFS_ADD_DELAYED_REF/BTRFS_DROP_DELAYED_REF or
 671  *              BTRFS_ADD_DELAYED_EXTENT
 672  *
 673  * @ref_type:   Holds the type of the extent which is being recorded, can be
 674  *              one of BTRFS_SHARED_BLOCK_REF_KEY/BTRFS_TREE_BLOCK_REF_KEY
 675  *              when recording a metadata extent or BTRFS_SHARED_DATA_REF_KEY/
 676  *              BTRFS_EXTENT_DATA_REF_KEY when recording data extent
 677  */
 678 static void init_delayed_ref_common(struct btrfs_fs_info *fs_info,
 679                                     struct btrfs_delayed_ref_node *ref,
 680                                     u64 bytenr, u64 num_bytes, u64 ref_root,
 681                                     int action, u8 ref_type)
 682 {
 683         u64 seq = 0;
 684
 685         if (action == BTRFS_ADD_DELAYED_EXTENT)
 686                 action = BTRFS_ADD_DELAYED_REF;
 687
 688         if (is_fstree(ref_root))
 689                 seq = atomic64_read(&fs_info->tree_mod_seq);
 690
 691         refcount_set(&ref->refs, 1);
 692         ref->bytenr = bytenr;
 693         ref->num_bytes = num_bytes;
 694         ref->ref_mod = 1;
 695         ref->action = action;
 696         ref->is_head = 0;
 697         ref->in_tree = 1;
 698         ref->seq = seq;
 699         ref->type = ref_type;
 700         RB_CLEAR_NODE(&ref->ref_node);
 701         INIT_LIST_HEAD(&ref->add_list);
 702 }
 703
 704 /*
 705  * add a delayed tree ref.  This does all of the accounting required
 706  * to make sure the delayed ref is eventually processed before this
 707  * transaction commits.
 708  */
 709 int btrfs_add_delayed_tree_ref(struct btrfs_trans_handle *trans,
 710                                u64 bytenr, u64 num_bytes, u64 parent,
 711                                u64 ref_root,  int level, int action,
 712                                struct btrfs_delayed_extent_op *extent_op,
 713                                int *old_ref_mod, int *new_ref_mod)
 714 {
 715         struct btrfs_fs_info *fs_info = trans->fs_info;
 716         struct btrfs_delayed_tree_ref *ref;
 717         struct btrfs_delayed_ref_head *head_ref;
 718         struct btrfs_delayed_ref_root *delayed_refs;
 719         struct btrfs_qgroup_extent_record *record = NULL;
 720         int qrecord_inserted;
 721         bool is_system = (ref_root == BTRFS_CHUNK_TREE_OBJECTID);
 722         int ret;
 723         u8 ref_type;
 724
 725         BUG_ON(extent_op && extent_op->is_data);
 726         ref = kmem_cache_alloc(btrfs_delayed_tree_ref_cachep, GFP_NOFS);
 727         if (!ref)
 728                 return -ENOMEM;
 729
 730         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
 731         if (!head_ref) {
 732                 kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
 733                 return -ENOMEM;
 734         }
 735
 736         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) &&
 737             is_fstree(ref_root)) {
 738                 record = kmalloc(sizeof(*record), GFP_NOFS);
 739                 if (!record) {
 740                         kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
 741                         kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
 742                         return -ENOMEM;
 743                 }
 744         }
 745
 746         if (parent)
 747                 ref_type = BTRFS_SHARED_BLOCK_REF_KEY;
 748         else
 749                 ref_type = BTRFS_TREE_BLOCK_REF_KEY;
 750
 751         init_delayed_ref_common(fs_info, &ref->node, bytenr, num_bytes,
 752                                 ref_root, action, ref_type);
 753         ref->root = ref_root;
 754         ref->parent = parent;
 755         ref->level = level;
 756
 757         init_delayed_ref_head(head_ref, record, bytenr, num_bytes,
 758                               ref_root, 0, action, false, is_system);
 759         head_ref->extent_op = extent_op;
 760
 761         delayed_refs = &trans->transaction->delayed_refs;
 762         spin_lock(&delayed_refs->lock);
 763
 764         /*
 765          * insert both the head node and the new ref without dropping
 766          * the spin lock
 767          */
 768         head_ref = add_delayed_ref_head(trans, head_ref, record,
 769                                         action, &qrecord_inserted,
 770                                         old_ref_mod, new_ref_mod);
 771
 772         ret = insert_delayed_ref(trans, delayed_refs, head_ref, &ref->node);
 773         spin_unlock(&delayed_refs->lock);
 774
 775         trace_add_delayed_tree_ref(fs_info, &ref->node, ref,
 776                                    action == BTRFS_ADD_DELAYED_EXTENT ?
 777                                    BTRFS_ADD_DELAYED_REF : action);
 778         if (ret > 0)
 779                 kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
 780
 781         if (qrecord_inserted)
 782                 btrfs_qgroup_trace_extent_post(fs_info, record);
 783
 784         return 0;
 785 }
 786
 787 /*
 788  * add a delayed data ref. it's similar to btrfs_add_delayed_tree_ref.
 789  */
 790 int btrfs_add_delayed_data_ref(struct btrfs_trans_handle *trans,
 791                                u64 bytenr, u64 num_bytes,
 792                                u64 parent, u64 ref_root,
 793                                u64 owner, u64 offset, u64 reserved, int action,
 794                                int *old_ref_mod, int *new_ref_mod)
 795 {
 796         struct btrfs_fs_info *fs_info = trans->fs_info;
 797         struct btrfs_delayed_data_ref *ref;
 798         struct btrfs_delayed_ref_head *head_ref;
 799         struct btrfs_delayed_ref_root *delayed_refs;
 800         struct btrfs_qgroup_extent_record *record = NULL;
 801         int qrecord_inserted;
 802         int ret;
 803         u8 ref_type;
 804
 805         ref = kmem_cache_alloc(btrfs_delayed_data_ref_cachep, GFP_NOFS);
 806         if (!ref)
 807                 return -ENOMEM;
 808
 809         if (parent)
 810                 ref_type = BTRFS_SHARED_DATA_REF_KEY;
 811         else
 812                 ref_type = BTRFS_EXTENT_DATA_REF_KEY;
 813         init_delayed_ref_common(fs_info, &ref->node, bytenr, num_bytes,
 814                                 ref_root, action, ref_type);
 815         ref->root = ref_root;
 816         ref->parent = parent;
 817         ref->objectid = owner;
 818         ref->offset = offset;
 819
 820
 821         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
 822         if (!head_ref) {
 823                 kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
 824                 return -ENOMEM;
 825         }
 826
 827         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) &&
 828             is_fstree(ref_root)) {
 829                 record = kmalloc(sizeof(*record), GFP_NOFS);
 830                 if (!record) {
 831                         kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
 832                         kmem_cache_free(btrfs_delayed_ref_head_cachep,
 833                                         head_ref);
 834                         return -ENOMEM;
 835                 }
 836         }
 837
 838         init_delayed_ref_head(head_ref, record, bytenr, num_bytes, ref_root,
 839                               reserved, action, true, false);
 840         head_ref->extent_op = NULL;
 841
 842         delayed_refs = &trans->transaction->delayed_refs;
 843         spin_lock(&delayed_refs->lock);
 844
 845         /*
 846          * insert both the head node and the new ref without dropping
 847          * the spin lock
 848          */
 849         head_ref = add_delayed_ref_head(trans, head_ref, record,
 850                                         action, &qrecord_inserted,
 851                                         old_ref_mod, new_ref_mod);
 852
 853         ret = insert_delayed_ref(trans, delayed_refs, head_ref, &ref->node);
 854         spin_unlock(&delayed_refs->lock);
 855
 856         trace_add_delayed_data_ref(trans->fs_info, &ref->node, ref,
 857                                    action == BTRFS_ADD_DELAYED_EXTENT ?
 858                                    BTRFS_ADD_DELAYED_REF : action);
 859         if (ret > 0)
 860                 kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
 861
 862
 863         if (qrecord_inserted)
 864                 return btrfs_qgroup_trace_extent_post(fs_info, record);
 865         return 0;
 866 }
 867
 868 int btrfs_add_delayed_extent_op(struct btrfs_fs_info *fs_info,
 869                                 struct btrfs_trans_handle *trans,
 870                                 u64 bytenr, u64 num_bytes,
 871                                 struct btrfs_delayed_extent_op *extent_op)
 872 {
 873         struct btrfs_delayed_ref_head *head_ref;
 874         struct btrfs_delayed_ref_root *delayed_refs;
 875
 876         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
 877         if (!head_ref)
 878                 return -ENOMEM;
 879
 880         init_delayed_ref_head(head_ref, NULL, bytenr, num_bytes, 0, 0,
 881                               BTRFS_UPDATE_DELAYED_HEAD, extent_op->is_data,
 882                               false);
 883         head_ref->extent_op = extent_op;
 884
 885         delayed_refs = &trans->transaction->delayed_refs;
 886         spin_lock(&delayed_refs->lock);
 887
 888         add_delayed_ref_head(trans, head_ref, NULL, BTRFS_UPDATE_DELAYED_HEAD,
 889                              NULL, NULL, NULL);
 890
 891         spin_unlock(&delayed_refs->lock);
 892         return 0;
 893 }
 894
 895 /*
 896  * this does a simple search for the head node for a given extent.
 897  * It must be called with the delayed ref spinlock held, and it returns
 898  * the head node if any where found, or NULL if not.
 899  */
 900 struct btrfs_delayed_ref_head *
 901 btrfs_find_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs, u64 bytenr)
 902 {
 903         return find_ref_head(&delayed_refs->href_root, bytenr, 0);
 904 }
 905
 906 void __cold btrfs_delayed_ref_exit(void)
 907 {
 908         kmem_cache_destroy(btrfs_delayed_ref_head_cachep);
 909         kmem_cache_destroy(btrfs_delayed_tree_ref_cachep);
 910         kmem_cache_destroy(btrfs_delayed_data_ref_cachep);
 911         kmem_cache_destroy(btrfs_delayed_extent_op_cachep);
 912 }
 913
 914 int __init btrfs_delayed_ref_init(void)
 915 {
 916         btrfs_delayed_ref_head_cachep = kmem_cache_create(
 917                                 "btrfs_delayed_ref_head",
 918                                 sizeof(struct btrfs_delayed_ref_head), 0,
 919                                 SLAB_MEM_SPREAD, NULL);
 920         if (!btrfs_delayed_ref_head_cachep)
 921                 goto fail;
 922
 923         btrfs_delayed_tree_ref_cachep = kmem_cache_create(
 924                                 "btrfs_delayed_tree_ref",
 925                                 sizeof(struct btrfs_delayed_tree_ref), 0,
 926                                 SLAB_MEM_SPREAD, NULL);
 927         if (!btrfs_delayed_tree_ref_cachep)
 928                 goto fail;
 929
 930         btrfs_delayed_data_ref_cachep = kmem_cache_create(
 931                                 "btrfs_delayed_data_ref",
 932                                 sizeof(struct btrfs_delayed_data_ref), 0,
 933                                 SLAB_MEM_SPREAD, NULL);
 934         if (!btrfs_delayed_data_ref_cachep)
 935                 goto fail;
 936
 937         btrfs_delayed_extent_op_cachep = kmem_cache_create(
 938                                 "btrfs_delayed_extent_op",
 939                                 sizeof(struct btrfs_delayed_extent_op), 0,
 940                                 SLAB_MEM_SPREAD, NULL);
 941         if (!btrfs_delayed_extent_op_cachep)
 942                 goto fail;
 943
 944         return 0;
 945 fail:
 946         btrfs_delayed_ref_exit();
 947         return -ENOMEM;
 948 }