fs/btrfs/delayed-ref.c

   1 /*
   2  * Copyright (C) 2009 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/sched.h>
  20 #include <linux/slab.h>
  21 #include <linux/sort.h>
  22 #include "ctree.h"
  23 #include "delayed-ref.h"
  24 #include "transaction.h"
  25
  26 struct kmem_cache *btrfs_delayed_ref_head_cachep;
  27 struct kmem_cache *btrfs_delayed_tree_ref_cachep;
  28 struct kmem_cache *btrfs_delayed_data_ref_cachep;
  29 struct kmem_cache *btrfs_delayed_extent_op_cachep;
  30 /*
  31  * delayed back reference update tracking.  For subvolume trees
  32  * we queue up extent allocations and backref maintenance for
  33  * delayed processing.   This avoids deep call chains where we
  34  * add extents in the middle of btrfs_search_slot, and it allows
  35  * us to buffer up frequently modified backrefs in an rb tree instead
  36  * of hammering updates on the extent allocation tree.
  37  */
  38
  39 /*
  40  * compare two delayed tree backrefs with same bytenr and type
  41  */
  42 static int comp_tree_refs(struct btrfs_delayed_tree_ref *ref2,
  43                           struct btrfs_delayed_tree_ref *ref1)
  44 {
  45         if (ref1->root < ref2->root)
  46                 return -1;
  47         if (ref1->root > ref2->root)
  48                 return 1;
  49         if (ref1->parent < ref2->parent)
  50                 return -1;
  51         if (ref1->parent > ref2->parent)
  52                 return 1;
  53         return 0;
  54 }
  55
  56 /*
  57  * compare two delayed data backrefs with same bytenr and type
  58  */
  59 static int comp_data_refs(struct btrfs_delayed_data_ref *ref2,
  60                           struct btrfs_delayed_data_ref *ref1)
  61 {
  62         if (ref1->node.type == BTRFS_EXTENT_DATA_REF_KEY) {
  63                 if (ref1->root < ref2->root)
  64                         return -1;
  65                 if (ref1->root > ref2->root)
  66                         return 1;
  67                 if (ref1->objectid < ref2->objectid)
  68                         return -1;
  69                 if (ref1->objectid > ref2->objectid)
  70                         return 1;
  71                 if (ref1->offset < ref2->offset)
  72                         return -1;
  73                 if (ref1->offset > ref2->offset)
  74                         return 1;
  75         } else {
  76                 if (ref1->parent < ref2->parent)
  77                         return -1;
  78                 if (ref1->parent > ref2->parent)
  79                         return 1;
  80         }
  81         return 0;
  82 }
  83
  84 /*
  85  * entries in the rb tree are ordered by the byte number of the extent,
  86  * type of the delayed backrefs and content of delayed backrefs.
  87  */
  88 static int comp_entry(struct btrfs_delayed_ref_node *ref2,
  89                       struct btrfs_delayed_ref_node *ref1,
  90                       bool compare_seq)
  91 {
  92         if (ref1->bytenr < ref2->bytenr)
  93                 return -1;
  94         if (ref1->bytenr > ref2->bytenr)
  95                 return 1;
  96         if (ref1->is_head && ref2->is_head)
  97                 return 0;
  98         if (ref2->is_head)
  99                 return -1;
 100         if (ref1->is_head)
 101                 return 1;
 102         if (ref1->type < ref2->type)
 103                 return -1;
 104         if (ref1->type > ref2->type)
 105                 return 1;
 106         /* merging of sequenced refs is not allowed */
 107         if (compare_seq) {
 108                 if (ref1->seq < ref2->seq)
 109                         return -1;
 110                 if (ref1->seq > ref2->seq)
 111                         return 1;
 112         }
 113         if (ref1->type == BTRFS_TREE_BLOCK_REF_KEY ||
 114             ref1->type == BTRFS_SHARED_BLOCK_REF_KEY) {
 115                 return comp_tree_refs(btrfs_delayed_node_to_tree_ref(ref2),
 116                                       btrfs_delayed_node_to_tree_ref(ref1));
 117         } else if (ref1->type == BTRFS_EXTENT_DATA_REF_KEY ||
 118                    ref1->type == BTRFS_SHARED_DATA_REF_KEY) {
 119                 return comp_data_refs(btrfs_delayed_node_to_data_ref(ref2),
 120                                       btrfs_delayed_node_to_data_ref(ref1));
 121         }
 122         BUG();
 123         return 0;
 124 }
 125
 126 /*
 127  * insert a new ref into the rbtree.  This returns any existing refs
 128  * for the same (bytenr,parent) tuple, or NULL if the new node was properly
 129  * inserted.
 130  */
 131 static struct btrfs_delayed_ref_node *tree_insert(struct rb_root *root,
 132                                                   struct rb_node *node)
 133 {
 134         struct rb_node **p = &root->rb_node;
 135         struct rb_node *parent_node = NULL;
 136         struct btrfs_delayed_ref_node *entry;
 137         struct btrfs_delayed_ref_node *ins;
 138         int cmp;
 139
 140         ins = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
 141         while (*p) {
 142                 parent_node = *p;
 143                 entry = rb_entry(parent_node, struct btrfs_delayed_ref_node,
 144                                  rb_node);
 145
 146                 cmp = comp_entry(entry, ins, 1);
 147                 if (cmp < 0)
 148                         p = &(*p)->rb_left;
 149                 else if (cmp > 0)
 150                         p = &(*p)->rb_right;
 151                 else
 152                         return entry;
 153         }
 154
 155         rb_link_node(node, parent_node, p);
 156         rb_insert_color(node, root);
 157         return NULL;
 158 }
 159
 160 /*
 161  * find an head entry based on bytenr. This returns the delayed ref
 162  * head if it was able to find one, or NULL if nothing was in that spot.
 163  * If return_bigger is given, the next bigger entry is returned if no exact
 164  * match is found.
 165  */
 166 static struct btrfs_delayed_ref_node *find_ref_head(struct rb_root *root,
 167                                   u64 bytenr,
 168                                   struct btrfs_delayed_ref_node **last,
 169                                   int return_bigger)
 170 {
 171         struct rb_node *n;
 172         struct btrfs_delayed_ref_node *entry;
 173         int cmp = 0;
 174
 175 again:
 176         n = root->rb_node;
 177         entry = NULL;
 178         while (n) {
 179                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
 180                 WARN_ON(!entry->in_tree);
 181                 if (last)
 182                         *last = entry;
 183
 184                 if (bytenr < entry->bytenr)
 185                         cmp = -1;
 186                 else if (bytenr > entry->bytenr)
 187                         cmp = 1;
 188                 else if (!btrfs_delayed_ref_is_head(entry))
 189                         cmp = 1;
 190                 else
 191                         cmp = 0;
 192
 193                 if (cmp < 0)
 194                         n = n->rb_left;
 195                 else if (cmp > 0)
 196                         n = n->rb_right;
 197                 else
 198                         return entry;
 199         }
 200         if (entry && return_bigger) {
 201                 if (cmp > 0) {
 202                         n = rb_next(&entry->rb_node);
 203                         if (!n)
 204                                 n = rb_first(root);
 205                         entry = rb_entry(n, struct btrfs_delayed_ref_node,
 206                                          rb_node);
 207                         bytenr = entry->bytenr;
 208                         return_bigger = 0;
 209                         goto again;
 210                 }
 211                 return entry;
 212         }
 213         return NULL;
 214 }
 215
 216 int btrfs_delayed_ref_lock(struct btrfs_trans_handle *trans,
 217                            struct btrfs_delayed_ref_head *head)
 218 {
 219         struct btrfs_delayed_ref_root *delayed_refs;
 220
 221         delayed_refs = &trans->transaction->delayed_refs;
 222         assert_spin_locked(&delayed_refs->lock);
 223         if (mutex_trylock(&head->mutex))
 224                 return 0;
 225
 226         atomic_inc(&head->node.refs);
 227         spin_unlock(&delayed_refs->lock);
 228
 229         mutex_lock(&head->mutex);
 230         spin_lock(&delayed_refs->lock);
 231         if (!head->node.in_tree) {
 232                 mutex_unlock(&head->mutex);
 233                 btrfs_put_delayed_ref(&head->node);
 234                 return -EAGAIN;
 235         }
 236         btrfs_put_delayed_ref(&head->node);
 237         return 0;
 238 }
 239
 240 static void inline drop_delayed_ref(struct btrfs_trans_handle *trans,
 241                                     struct btrfs_delayed_ref_root *delayed_refs,
 242                                     struct btrfs_delayed_ref_node *ref)
 243 {
 244         rb_erase(&ref->rb_node, &delayed_refs->root);
 245         ref->in_tree = 0;
 246         btrfs_put_delayed_ref(ref);
 247         delayed_refs->num_entries--;
 248         if (trans->delayed_ref_updates)
 249                 trans->delayed_ref_updates--;
 250 }
 251
 252 static int merge_ref(struct btrfs_trans_handle *trans,
 253                      struct btrfs_delayed_ref_root *delayed_refs,
 254                      struct btrfs_delayed_ref_node *ref, u64 seq)
 255 {
 256         struct rb_node *node;
 257         int merged = 0;
 258         int mod = 0;
 259         int done = 0;
 260
 261         node = rb_prev(&ref->rb_node);
 262         while (node) {
 263                 struct btrfs_delayed_ref_node *next;
 264
 265                 next = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
 266                 node = rb_prev(node);
 267                 if (next->bytenr != ref->bytenr)
 268                         break;
 269                 if (seq && next->seq >= seq)
 270                         break;
 271                 if (comp_entry(ref, next, 0))
 272                         continue;
 273
 274                 if (ref->action == next->action) {
 275                         mod = next->ref_mod;
 276                 } else {
 277                         if (ref->ref_mod < next->ref_mod) {
 278                                 struct btrfs_delayed_ref_node *tmp;
 279
 280                                 tmp = ref;
 281                                 ref = next;
 282                                 next = tmp;
 283                                 done = 1;
 284                         }
 285                         mod = -next->ref_mod;
 286                 }
 287
 288                 merged++;
 289                 drop_delayed_ref(trans, delayed_refs, next);
 290                 ref->ref_mod += mod;
 291                 if (ref->ref_mod == 0) {
 292                         drop_delayed_ref(trans, delayed_refs, ref);
 293                         break;
 294                 } else {
 295                         /*
 296                          * You can't have multiples of the same ref on a tree
 297                          * block.
 298                          */
 299                         WARN_ON(ref->type == BTRFS_TREE_BLOCK_REF_KEY ||
 300                                 ref->type == BTRFS_SHARED_BLOCK_REF_KEY);
 301                 }
 302
 303                 if (done)
 304                         break;
 305                 node = rb_prev(&ref->rb_node);
 306         }
 307
 308         return merged;
 309 }
 310
 311 void btrfs_merge_delayed_refs(struct btrfs_trans_handle *trans,
 312                               struct btrfs_fs_info *fs_info,
 313                               struct btrfs_delayed_ref_root *delayed_refs,
 314                               struct btrfs_delayed_ref_head *head)
 315 {
 316         struct rb_node *node;
 317         u64 seq = 0;
 318
 319         spin_lock(&fs_info->tree_mod_seq_lock);
 320         if (!list_empty(&fs_info->tree_mod_seq_list)) {
 321                 struct seq_list *elem;
 322
 323                 elem = list_first_entry(&fs_info->tree_mod_seq_list,
 324                                         struct seq_list, list);
 325                 seq = elem->seq;
 326         }
 327         spin_unlock(&fs_info->tree_mod_seq_lock);
 328
 329         node = rb_prev(&head->node.rb_node);
 330         while (node) {
 331                 struct btrfs_delayed_ref_node *ref;
 332
 333                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
 334                                rb_node);
 335                 if (ref->bytenr != head->node.bytenr)
 336                         break;
 337
 338                 /* We can't merge refs that are outside of our seq count */
 339                 if (seq && ref->seq >= seq)
 340                         break;
 341                 if (merge_ref(trans, delayed_refs, ref, seq))
 342                         node = rb_prev(&head->node.rb_node);
 343                 else
 344                         node = rb_prev(node);
 345         }
 346 }
 347
 348 int btrfs_check_delayed_seq(struct btrfs_fs_info *fs_info,
 349                             struct btrfs_delayed_ref_root *delayed_refs,
 350                             u64 seq)
 351 {
 352         struct seq_list *elem;
 353         int ret = 0;
 354
 355         spin_lock(&fs_info->tree_mod_seq_lock);
 356         if (!list_empty(&fs_info->tree_mod_seq_list)) {
 357                 elem = list_first_entry(&fs_info->tree_mod_seq_list,
 358                                         struct seq_list, list);
 359                 if (seq >= elem->seq) {
 360                         pr_debug("holding back delayed_ref %llu, lowest is "
 361                                  "%llu (%p)\n", seq, elem->seq, delayed_refs);
 362                         ret = 1;
 363                 }
 364         }
 365
 366         spin_unlock(&fs_info->tree_mod_seq_lock);
 367         return ret;
 368 }
 369
 370 int btrfs_find_ref_cluster(struct btrfs_trans_handle *trans,
 371                            struct list_head *cluster, u64 start)
 372 {
 373         int count = 0;
 374         struct btrfs_delayed_ref_root *delayed_refs;
 375         struct rb_node *node;
 376         struct btrfs_delayed_ref_node *ref;
 377         struct btrfs_delayed_ref_head *head;
 378
 379         delayed_refs = &trans->transaction->delayed_refs;
 380         if (start == 0) {
 381                 node = rb_first(&delayed_refs->root);
 382         } else {
 383                 ref = NULL;
 384                 find_ref_head(&delayed_refs->root, start + 1, &ref, 1);
 385                 if (ref) {
 386                         node = &ref->rb_node;
 387                 } else
 388                         node = rb_first(&delayed_refs->root);
 389         }
 390 again:
 391         while (node && count < 32) {
 392                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
 393                 if (btrfs_delayed_ref_is_head(ref)) {
 394                         head = btrfs_delayed_node_to_head(ref);
 395                         if (list_empty(&head->cluster)) {
 396                                 list_add_tail(&head->cluster, cluster);
 397                                 delayed_refs->run_delayed_start =
 398                                         head->node.bytenr;
 399                                 count++;
 400
 401                                 WARN_ON(delayed_refs->num_heads_ready == 0);
 402                                 delayed_refs->num_heads_ready--;
 403                         } else if (count) {
 404                                 /* the goal of the clustering is to find extents
 405                                  * that are likely to end up in the same extent
 406                                  * leaf on disk.  So, we don't want them spread
 407                                  * all over the tree.  Stop now if we've hit
 408                                  * a head that was already in use
 409                                  */
 410                                 break;
 411                         }
 412                 }
 413                 node = rb_next(node);
 414         }
 415         if (count) {
 416                 return 0;
 417         } else if (start) {
 418                 /*
 419                  * we've gone to the end of the rbtree without finding any
 420                  * clusters.  start from the beginning and try again
 421                  */
 422                 start = 0;
 423                 node = rb_first(&delayed_refs->root);
 424                 goto again;
 425         }
 426         return 1;
 427 }
 428
 429 /*
 430  * helper function to update an extent delayed ref in the
 431  * rbtree.  existing and update must both have the same
 432  * bytenr and parent
 433  *
 434  * This may free existing if the update cancels out whatever
 435  * operation it was doing.
 436  */
 437 static noinline void
 438 update_existing_ref(struct btrfs_trans_handle *trans,
 439                     struct btrfs_delayed_ref_root *delayed_refs,
 440                     struct btrfs_delayed_ref_node *existing,
 441                     struct btrfs_delayed_ref_node *update)
 442 {
 443         if (update->action != existing->action) {
 444                 /*
 445                  * this is effectively undoing either an add or a
 446                  * drop.  We decrement the ref_mod, and if it goes
 447                  * down to zero we just delete the entry without
 448                  * every changing the extent allocation tree.
 449                  */
 450                 existing->ref_mod--;
 451                 if (existing->ref_mod == 0)
 452                         drop_delayed_ref(trans, delayed_refs, existing);
 453                 else
 454                         WARN_ON(existing->type == BTRFS_TREE_BLOCK_REF_KEY ||
 455                                 existing->type == BTRFS_SHARED_BLOCK_REF_KEY);
 456         } else {
 457                 WARN_ON(existing->type == BTRFS_TREE_BLOCK_REF_KEY ||
 458                         existing->type == BTRFS_SHARED_BLOCK_REF_KEY);
 459                 /*
 460                  * the action on the existing ref matches
 461                  * the action on the ref we're trying to add.
 462                  * Bump the ref_mod by one so the backref that
 463                  * is eventually added/removed has the correct
 464                  * reference count
 465                  */
 466                 existing->ref_mod += update->ref_mod;
 467         }
 468 }
 469
 470 /*
 471  * helper function to update the accounting in the head ref
 472  * existing and update must have the same bytenr
 473  */
 474 static noinline void
 475 update_existing_head_ref(struct btrfs_delayed_ref_node *existing,
 476                          struct btrfs_delayed_ref_node *update)
 477 {
 478         struct btrfs_delayed_ref_head *existing_ref;
 479         struct btrfs_delayed_ref_head *ref;
 480
 481         existing_ref = btrfs_delayed_node_to_head(existing);
 482         ref = btrfs_delayed_node_to_head(update);
 483         BUG_ON(existing_ref->is_data != ref->is_data);
 484
 485         if (ref->must_insert_reserved) {
 486                 /* if the extent was freed and then
 487                  * reallocated before the delayed ref
 488                  * entries were processed, we can end up
 489                  * with an existing head ref without
 490                  * the must_insert_reserved flag set.
 491                  * Set it again here
 492                  */
 493                 existing_ref->must_insert_reserved = ref->must_insert_reserved;
 494
 495                 /*
 496                  * update the num_bytes so we make sure the accounting
 497                  * is done correctly
 498                  */
 499                 existing->num_bytes = update->num_bytes;
 500
 501         }
 502
 503         if (ref->extent_op) {
 504                 if (!existing_ref->extent_op) {
 505                         existing_ref->extent_op = ref->extent_op;
 506                 } else {
 507                         if (ref->extent_op->update_key) {
 508                                 memcpy(&existing_ref->extent_op->key,
 509                                        &ref->extent_op->key,
 510                                        sizeof(ref->extent_op->key));
 511                                 existing_ref->extent_op->update_key = 1;
 512                         }
 513                         if (ref->extent_op->update_flags) {
 514                                 existing_ref->extent_op->flags_to_set |=
 515                                         ref->extent_op->flags_to_set;
 516                                 existing_ref->extent_op->update_flags = 1;
 517                         }
 518                         btrfs_free_delayed_extent_op(ref->extent_op);
 519                 }
 520         }
 521         /*
 522          * update the reference mod on the head to reflect this new operation
 523          */
 524         existing->ref_mod += update->ref_mod;
 525 }
 526
 527 /*
 528  * helper function to actually insert a head node into the rbtree.
 529  * this does all the dirty work in terms of maintaining the correct
 530  * overall modification count.
 531  */
 532 static noinline void add_delayed_ref_head(struct btrfs_fs_info *fs_info,
 533                                         struct btrfs_trans_handle *trans,
 534                                         struct btrfs_delayed_ref_node *ref,
 535                                         u64 bytenr, u64 num_bytes,
 536                                         int action, int is_data)
 537 {
 538         struct btrfs_delayed_ref_node *existing;
 539         struct btrfs_delayed_ref_head *head_ref = NULL;
 540         struct btrfs_delayed_ref_root *delayed_refs;
 541         int count_mod = 1;
 542         int must_insert_reserved = 0;
 543
 544         /*
 545          * the head node stores the sum of all the mods, so dropping a ref
 546          * should drop the sum in the head node by one.
 547          */
 548         if (action == BTRFS_UPDATE_DELAYED_HEAD)
 549                 count_mod = 0;
 550         else if (action == BTRFS_DROP_DELAYED_REF)
 551                 count_mod = -1;
 552
 553         /*
 554          * BTRFS_ADD_DELAYED_EXTENT means that we need to update
 555          * the reserved accounting when the extent is finally added, or
 556          * if a later modification deletes the delayed ref without ever
 557          * inserting the extent into the extent allocation tree.
 558          * ref->must_insert_reserved is the flag used to record
 559          * that accounting mods are required.
 560          *
 561          * Once we record must_insert_reserved, switch the action to
 562          * BTRFS_ADD_DELAYED_REF because other special casing is not required.
 563          */
 564         if (action == BTRFS_ADD_DELAYED_EXTENT)
 565                 must_insert_reserved = 1;
 566         else
 567                 must_insert_reserved = 0;
 568
 569         delayed_refs = &trans->transaction->delayed_refs;
 570
 571         /* first set the basic ref node struct up */
 572         atomic_set(&ref->refs, 1);
 573         ref->bytenr = bytenr;
 574         ref->num_bytes = num_bytes;
 575         ref->ref_mod = count_mod;
 576         ref->type  = 0;
 577         ref->action  = 0;
 578         ref->is_head = 1;
 579         ref->in_tree = 1;
 580         ref->seq = 0;
 581
 582         head_ref = btrfs_delayed_node_to_head(ref);
 583         head_ref->must_insert_reserved = must_insert_reserved;
 584         head_ref->is_data = is_data;
 585
 586         INIT_LIST_HEAD(&head_ref->cluster);
 587         mutex_init(&head_ref->mutex);
 588
 589         trace_btrfs_delayed_ref_head(ref, head_ref, action);
 590
 591         existing = tree_insert(&delayed_refs->root, &ref->rb_node);
 592
 593         if (existing) {
 594                 update_existing_head_ref(existing, ref);
 595                 /*
 596                  * we've updated the existing ref, free the newly
 597                  * allocated ref
 598                  */
 599                 kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
 600         } else {
 601                 delayed_refs->num_heads++;
 602                 delayed_refs->num_heads_ready++;
 603                 delayed_refs->num_entries++;
 604                 trans->delayed_ref_updates++;
 605         }
 606 }
 607
 608 /*
 609  * helper to insert a delayed tree ref into the rbtree.
 610  */
 611 static noinline void add_delayed_tree_ref(struct btrfs_fs_info *fs_info,
 612                                          struct btrfs_trans_handle *trans,
 613                                          struct btrfs_delayed_ref_node *ref,
 614                                          u64 bytenr, u64 num_bytes, u64 parent,
 615                                          u64 ref_root, int level, int action,
 616                                          int for_cow)
 617 {
 618         struct btrfs_delayed_ref_node *existing;
 619         struct btrfs_delayed_tree_ref *full_ref;
 620         struct btrfs_delayed_ref_root *delayed_refs;
 621         u64 seq = 0;
 622
 623         if (action == BTRFS_ADD_DELAYED_EXTENT)
 624                 action = BTRFS_ADD_DELAYED_REF;
 625
 626         delayed_refs = &trans->transaction->delayed_refs;
 627
 628         /* first set the basic ref node struct up */
 629         atomic_set(&ref->refs, 1);
 630         ref->bytenr = bytenr;
 631         ref->num_bytes = num_bytes;
 632         ref->ref_mod = 1;
 633         ref->action = action;
 634         ref->is_head = 0;
 635         ref->in_tree = 1;
 636
 637         if (need_ref_seq(for_cow, ref_root))
 638                 seq = btrfs_get_tree_mod_seq(fs_info, &trans->delayed_ref_elem);
 639         ref->seq = seq;
 640
 641         full_ref = btrfs_delayed_node_to_tree_ref(ref);
 642         full_ref->parent = parent;
 643         full_ref->root = ref_root;
 644         if (parent)
 645                 ref->type = BTRFS_SHARED_BLOCK_REF_KEY;
 646         else
 647                 ref->type = BTRFS_TREE_BLOCK_REF_KEY;
 648         full_ref->level = level;
 649
 650         trace_btrfs_delayed_tree_ref(ref, full_ref, action);
 651
 652         existing = tree_insert(&delayed_refs->root, &ref->rb_node);
 653
 654         if (existing) {
 655                 update_existing_ref(trans, delayed_refs, existing, ref);
 656                 /*
 657                  * we've updated the existing ref, free the newly
 658                  * allocated ref
 659                  */
 660                 kmem_cache_free(btrfs_delayed_tree_ref_cachep, full_ref);
 661         } else {
 662                 delayed_refs->num_entries++;
 663                 trans->delayed_ref_updates++;
 664         }
 665 }
 666
 667 /*
 668  * helper to insert a delayed data ref into the rbtree.
 669  */
 670 static noinline void add_delayed_data_ref(struct btrfs_fs_info *fs_info,
 671                                          struct btrfs_trans_handle *trans,
 672                                          struct btrfs_delayed_ref_node *ref,
 673                                          u64 bytenr, u64 num_bytes, u64 parent,
 674                                          u64 ref_root, u64 owner, u64 offset,
 675                                          int action, int for_cow)
 676 {
 677         struct btrfs_delayed_ref_node *existing;
 678         struct btrfs_delayed_data_ref *full_ref;
 679         struct btrfs_delayed_ref_root *delayed_refs;
 680         u64 seq = 0;
 681
 682         if (action == BTRFS_ADD_DELAYED_EXTENT)
 683                 action = BTRFS_ADD_DELAYED_REF;
 684
 685         delayed_refs = &trans->transaction->delayed_refs;
 686
 687         /* first set the basic ref node struct up */
 688         atomic_set(&ref->refs, 1);
 689         ref->bytenr = bytenr;
 690         ref->num_bytes = num_bytes;
 691         ref->ref_mod = 1;
 692         ref->action = action;
 693         ref->is_head = 0;
 694         ref->in_tree = 1;
 695
 696         if (need_ref_seq(for_cow, ref_root))
 697                 seq = btrfs_get_tree_mod_seq(fs_info, &trans->delayed_ref_elem);
 698         ref->seq = seq;
 699
 700         full_ref = btrfs_delayed_node_to_data_ref(ref);
 701         full_ref->parent = parent;
 702         full_ref->root = ref_root;
 703         if (parent)
 704                 ref->type = BTRFS_SHARED_DATA_REF_KEY;
 705         else
 706                 ref->type = BTRFS_EXTENT_DATA_REF_KEY;
 707
 708         full_ref->objectid = owner;
 709         full_ref->offset = offset;
 710
 711         trace_btrfs_delayed_data_ref(ref, full_ref, action);
 712
 713         existing = tree_insert(&delayed_refs->root, &ref->rb_node);
 714
 715         if (existing) {
 716                 update_existing_ref(trans, delayed_refs, existing, ref);
 717                 /*
 718                  * we've updated the existing ref, free the newly
 719                  * allocated ref
 720                  */
 721                 kmem_cache_free(btrfs_delayed_data_ref_cachep, full_ref);
 722         } else {
 723                 delayed_refs->num_entries++;
 724                 trans->delayed_ref_updates++;
 725         }
 726 }
 727
 728 /*
 729  * add a delayed tree ref.  This does all of the accounting required
 730  * to make sure the delayed ref is eventually processed before this
 731  * transaction commits.
 732  */
 733 int btrfs_add_delayed_tree_ref(struct btrfs_fs_info *fs_info,
 734                                struct btrfs_trans_handle *trans,
 735                                u64 bytenr, u64 num_bytes, u64 parent,
 736                                u64 ref_root,  int level, int action,
 737                                struct btrfs_delayed_extent_op *extent_op,
 738                                int for_cow)
 739 {
 740         struct btrfs_delayed_tree_ref *ref;
 741         struct btrfs_delayed_ref_head *head_ref;
 742         struct btrfs_delayed_ref_root *delayed_refs;
 743
 744         BUG_ON(extent_op && extent_op->is_data);
 745         ref = kmem_cache_alloc(btrfs_delayed_tree_ref_cachep, GFP_NOFS);
 746         if (!ref)
 747                 return -ENOMEM;
 748
 749         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
 750         if (!head_ref) {
 751                 kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
 752                 return -ENOMEM;
 753         }
 754
 755         head_ref->extent_op = extent_op;
 756
 757         delayed_refs = &trans->transaction->delayed_refs;
 758         spin_lock(&delayed_refs->lock);
 759
 760         /*
 761          * insert both the head node and the new ref without dropping
 762          * the spin lock
 763          */
 764         add_delayed_ref_head(fs_info, trans, &head_ref->node, bytenr,
 765                                    num_bytes, action, 0);
 766
 767         add_delayed_tree_ref(fs_info, trans, &ref->node, bytenr,
 768                                    num_bytes, parent, ref_root, level, action,
 769                                    for_cow);
 770         spin_unlock(&delayed_refs->lock);
 771         if (need_ref_seq(for_cow, ref_root))
 772                 btrfs_qgroup_record_ref(trans, &ref->node, extent_op);
 773
 774         return 0;
 775 }
 776
 777 /*
 778  * add a delayed data ref. it's similar to btrfs_add_delayed_tree_ref.
 779  */
 780 int btrfs_add_delayed_data_ref(struct btrfs_fs_info *fs_info,
 781                                struct btrfs_trans_handle *trans,
 782                                u64 bytenr, u64 num_bytes,
 783                                u64 parent, u64 ref_root,
 784                                u64 owner, u64 offset, int action,
 785                                struct btrfs_delayed_extent_op *extent_op,
 786                                int for_cow)
 787 {
 788         struct btrfs_delayed_data_ref *ref;
 789         struct btrfs_delayed_ref_head *head_ref;
 790         struct btrfs_delayed_ref_root *delayed_refs;
 791
 792         BUG_ON(extent_op && !extent_op->is_data);
 793         ref = kmem_cache_alloc(btrfs_delayed_data_ref_cachep, GFP_NOFS);
 794         if (!ref)
 795                 return -ENOMEM;
 796
 797         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
 798         if (!head_ref) {
 799                 kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
 800                 return -ENOMEM;
 801         }
 802
 803         head_ref->extent_op = extent_op;
 804
 805         delayed_refs = &trans->transaction->delayed_refs;
 806         spin_lock(&delayed_refs->lock);
 807
 808         /*
 809          * insert both the head node and the new ref without dropping
 810          * the spin lock
 811          */
 812         add_delayed_ref_head(fs_info, trans, &head_ref->node, bytenr,
 813                                    num_bytes, action, 1);
 814
 815         add_delayed_data_ref(fs_info, trans, &ref->node, bytenr,
 816                                    num_bytes, parent, ref_root, owner, offset,
 817                                    action, for_cow);
 818         spin_unlock(&delayed_refs->lock);
 819         if (need_ref_seq(for_cow, ref_root))
 820                 btrfs_qgroup_record_ref(trans, &ref->node, extent_op);
 821
 822         return 0;
 823 }
 824
 825 int btrfs_add_delayed_extent_op(struct btrfs_fs_info *fs_info,
 826                                 struct btrfs_trans_handle *trans,
 827                                 u64 bytenr, u64 num_bytes,
 828                                 struct btrfs_delayed_extent_op *extent_op)
 829 {
 830         struct btrfs_delayed_ref_head *head_ref;
 831         struct btrfs_delayed_ref_root *delayed_refs;
 832
 833         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
 834         if (!head_ref)
 835                 return -ENOMEM;
 836
 837         head_ref->extent_op = extent_op;
 838
 839         delayed_refs = &trans->transaction->delayed_refs;
 840         spin_lock(&delayed_refs->lock);
 841
 842         add_delayed_ref_head(fs_info, trans, &head_ref->node, bytenr,
 843                                    num_bytes, BTRFS_UPDATE_DELAYED_HEAD,
 844                                    extent_op->is_data);
 845
 846         spin_unlock(&delayed_refs->lock);
 847         return 0;
 848 }
 849
 850 /*
 851  * this does a simple search for the head node for a given extent.
 852  * It must be called with the delayed ref spinlock held, and it returns
 853  * the head node if any where found, or NULL if not.
 854  */
 855 struct btrfs_delayed_ref_head *
 856 btrfs_find_delayed_ref_head(struct btrfs_trans_handle *trans, u64 bytenr)
 857 {
 858         struct btrfs_delayed_ref_node *ref;
 859         struct btrfs_delayed_ref_root *delayed_refs;
 860
 861         delayed_refs = &trans->transaction->delayed_refs;
 862         ref = find_ref_head(&delayed_refs->root, bytenr, NULL, 0);
 863         if (ref)
 864                 return btrfs_delayed_node_to_head(ref);
 865         return NULL;
 866 }
 867
 868 void btrfs_delayed_ref_exit(void)
 869 {
 870         if (btrfs_delayed_ref_head_cachep)
 871                 kmem_cache_destroy(btrfs_delayed_ref_head_cachep);
 872         if (btrfs_delayed_tree_ref_cachep)
 873                 kmem_cache_destroy(btrfs_delayed_tree_ref_cachep);
 874         if (btrfs_delayed_data_ref_cachep)
 875                 kmem_cache_destroy(btrfs_delayed_data_ref_cachep);
 876         if (btrfs_delayed_extent_op_cachep)
 877                 kmem_cache_destroy(btrfs_delayed_extent_op_cachep);
 878 }
 879
 880 int btrfs_delayed_ref_init(void)
 881 {
 882         btrfs_delayed_ref_head_cachep = kmem_cache_create(
 883                                 "btrfs_delayed_ref_head",
 884                                 sizeof(struct btrfs_delayed_ref_head), 0,
 885                                 SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 886         if (!btrfs_delayed_ref_head_cachep)
 887                 goto fail;
 888
 889         btrfs_delayed_tree_ref_cachep = kmem_cache_create(
 890                                 "btrfs_delayed_tree_ref",
 891                                 sizeof(struct btrfs_delayed_tree_ref), 0,
 892                                 SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 893         if (!btrfs_delayed_tree_ref_cachep)
 894                 goto fail;
 895
 896         btrfs_delayed_data_ref_cachep = kmem_cache_create(
 897                                 "btrfs_delayed_data_ref",
 898                                 sizeof(struct btrfs_delayed_data_ref), 0,
 899                                 SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 900         if (!btrfs_delayed_data_ref_cachep)
 901                 goto fail;
 902
 903         btrfs_delayed_extent_op_cachep = kmem_cache_create(
 904                                 "btrfs_delayed_extent_op",
 905                                 sizeof(struct btrfs_delayed_extent_op), 0,
 906                                 SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 907         if (!btrfs_delayed_extent_op_cachep)
 908                 goto fail;
 909
 910         return 0;
 911 fail:
 912         btrfs_delayed_ref_exit();
 913         return -ENOMEM;
 914 }