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 #include "qgroup.h"
  26
  27 struct kmem_cache *btrfs_delayed_ref_head_cachep;
  28 struct kmem_cache *btrfs_delayed_tree_ref_cachep;
  29 struct kmem_cache *btrfs_delayed_data_ref_cachep;
  30 struct kmem_cache *btrfs_delayed_extent_op_cachep;
  31 /*
  32  * delayed back reference update tracking.  For subvolume trees
  33  * we queue up extent allocations and backref maintenance for
  34  * delayed processing.   This avoids deep call chains where we
  35  * add extents in the middle of btrfs_search_slot, and it allows
  36  * us to buffer up frequently modified backrefs in an rb tree instead
  37  * of hammering updates on the extent allocation tree.
  38  */
  39
  40 /*
  41  * compare two delayed tree backrefs with same bytenr and type
  42  */
  43 static int comp_tree_refs(struct btrfs_delayed_tree_ref *ref2,
  44                           struct btrfs_delayed_tree_ref *ref1, int type)
  45 {
  46         if (type == BTRFS_TREE_BLOCK_REF_KEY) {
  47                 if (ref1->root < ref2->root)
  48                         return -1;
  49                 if (ref1->root > ref2->root)
  50                         return 1;
  51         } else {
  52                 if (ref1->parent < ref2->parent)
  53                         return -1;
  54                 if (ref1->parent > ref2->parent)
  55                         return 1;
  56         }
  57         return 0;
  58 }
  59
  60 /*
  61  * compare two delayed data backrefs with same bytenr and type
  62  */
  63 static int comp_data_refs(struct btrfs_delayed_data_ref *ref2,
  64                           struct btrfs_delayed_data_ref *ref1)
  65 {
  66         if (ref1->node.type == BTRFS_EXTENT_DATA_REF_KEY) {
  67                 if (ref1->root < ref2->root)
  68                         return -1;
  69                 if (ref1->root > ref2->root)
  70                         return 1;
  71                 if (ref1->objectid < ref2->objectid)
  72                         return -1;
  73                 if (ref1->objectid > ref2->objectid)
  74                         return 1;
  75                 if (ref1->offset < ref2->offset)
  76                         return -1;
  77                 if (ref1->offset > ref2->offset)
  78                         return 1;
  79         } else {
  80                 if (ref1->parent < ref2->parent)
  81                         return -1;
  82                 if (ref1->parent > ref2->parent)
  83                         return 1;
  84         }
  85         return 0;
  86 }
  87
  88 /* insert a new ref to head ref rbtree */
  89 static struct btrfs_delayed_ref_head *htree_insert(struct rb_root *root,
  90                                                    struct rb_node *node)
  91 {
  92         struct rb_node **p = &root->rb_node;
  93         struct rb_node *parent_node = NULL;
  94         struct btrfs_delayed_ref_head *entry;
  95         struct btrfs_delayed_ref_head *ins;
  96         u64 bytenr;
  97
  98         ins = rb_entry(node, struct btrfs_delayed_ref_head, href_node);
  99         bytenr = ins->node.bytenr;
 100         while (*p) {
 101                 parent_node = *p;
 102                 entry = rb_entry(parent_node, struct btrfs_delayed_ref_head,
 103                                  href_node);
 104
 105                 if (bytenr < entry->node.bytenr)
 106                         p = &(*p)->rb_left;
 107                 else if (bytenr > entry->node.bytenr)
 108                         p = &(*p)->rb_right;
 109                 else
 110                         return entry;
 111         }
 112
 113         rb_link_node(node, parent_node, p);
 114         rb_insert_color(node, root);
 115         return NULL;
 116 }
 117
 118 /*
 119  * find an head entry based on bytenr. This returns the delayed ref
 120  * head if it was able to find one, or NULL if nothing was in that spot.
 121  * If return_bigger is given, the next bigger entry is returned if no exact
 122  * match is found.
 123  */
 124 static struct btrfs_delayed_ref_head *
 125 find_ref_head(struct rb_root *root, u64 bytenr,
 126               int return_bigger)
 127 {
 128         struct rb_node *n;
 129         struct btrfs_delayed_ref_head *entry;
 130
 131         n = root->rb_node;
 132         entry = NULL;
 133         while (n) {
 134                 entry = rb_entry(n, struct btrfs_delayed_ref_head, href_node);
 135
 136                 if (bytenr < entry->node.bytenr)
 137                         n = n->rb_left;
 138                 else if (bytenr > entry->node.bytenr)
 139                         n = n->rb_right;
 140                 else
 141                         return entry;
 142         }
 143         if (entry && return_bigger) {
 144                 if (bytenr > entry->node.bytenr) {
 145                         n = rb_next(&entry->href_node);
 146                         if (!n)
 147                                 n = rb_first(root);
 148                         entry = rb_entry(n, struct btrfs_delayed_ref_head,
 149                                          href_node);
 150                         return entry;
 151                 }
 152                 return entry;
 153         }
 154         return NULL;
 155 }
 156
 157 int btrfs_delayed_ref_lock(struct btrfs_trans_handle *trans,
 158                            struct btrfs_delayed_ref_head *head)
 159 {
 160         struct btrfs_delayed_ref_root *delayed_refs;
 161
 162         delayed_refs = &trans->transaction->delayed_refs;
 163         assert_spin_locked(&delayed_refs->lock);
 164         if (mutex_trylock(&head->mutex))
 165                 return 0;
 166
 167         atomic_inc(&head->node.refs);
 168         spin_unlock(&delayed_refs->lock);
 169
 170         mutex_lock(&head->mutex);
 171         spin_lock(&delayed_refs->lock);
 172         if (!head->node.in_tree) {
 173                 mutex_unlock(&head->mutex);
 174                 btrfs_put_delayed_ref(&head->node);
 175                 return -EAGAIN;
 176         }
 177         btrfs_put_delayed_ref(&head->node);
 178         return 0;
 179 }
 180
 181 static inline void drop_delayed_ref(struct btrfs_trans_handle *trans,
 182                                     struct btrfs_delayed_ref_root *delayed_refs,
 183                                     struct btrfs_delayed_ref_head *head,
 184                                     struct btrfs_delayed_ref_node *ref)
 185 {
 186         if (btrfs_delayed_ref_is_head(ref)) {
 187                 head = btrfs_delayed_node_to_head(ref);
 188                 rb_erase(&head->href_node, &delayed_refs->href_root);
 189         } else {
 190                 assert_spin_locked(&head->lock);
 191                 list_del(&ref->list);
 192         }
 193         ref->in_tree = 0;
 194         btrfs_put_delayed_ref(ref);
 195         atomic_dec(&delayed_refs->num_entries);
 196         if (trans->delayed_ref_updates)
 197                 trans->delayed_ref_updates--;
 198 }
 199
 200 static bool merge_ref(struct btrfs_trans_handle *trans,
 201                       struct btrfs_delayed_ref_root *delayed_refs,
 202                       struct btrfs_delayed_ref_head *head,
 203                       struct btrfs_delayed_ref_node *ref,
 204                       u64 seq)
 205 {
 206         struct btrfs_delayed_ref_node *next;
 207         bool done = false;
 208
 209         next = list_first_entry(&head->ref_list, struct btrfs_delayed_ref_node,
 210                                 list);
 211         while (!done && &next->list != &head->ref_list) {
 212                 int mod;
 213                 struct btrfs_delayed_ref_node *next2;
 214
 215                 next2 = list_next_entry(next, list);
 216
 217                 if (next == ref)
 218                         goto next;
 219
 220                 if (seq && next->seq >= seq)
 221                         goto next;
 222
 223                 if (next->type != ref->type || next->no_quota != ref->no_quota)
 224                         goto next;
 225
 226                 if ((ref->type == BTRFS_TREE_BLOCK_REF_KEY ||
 227                      ref->type == BTRFS_SHARED_BLOCK_REF_KEY) &&
 228                     comp_tree_refs(btrfs_delayed_node_to_tree_ref(ref),
 229                                    btrfs_delayed_node_to_tree_ref(next),
 230                                    ref->type))
 231                         goto next;
 232                 if ((ref->type == BTRFS_EXTENT_DATA_REF_KEY ||
 233                      ref->type == BTRFS_SHARED_DATA_REF_KEY) &&
 234                     comp_data_refs(btrfs_delayed_node_to_data_ref(ref),
 235                                    btrfs_delayed_node_to_data_ref(next)))
 236                         goto next;
 237
 238                 if (ref->action == next->action) {
 239                         mod = next->ref_mod;
 240                 } else {
 241                         if (ref->ref_mod < next->ref_mod) {
 242                                 swap(ref, next);
 243                                 done = true;
 244                         }
 245                         mod = -next->ref_mod;
 246                 }
 247
 248                 drop_delayed_ref(trans, delayed_refs, head, next);
 249                 ref->ref_mod += mod;
 250                 if (ref->ref_mod == 0) {
 251                         drop_delayed_ref(trans, delayed_refs, head, ref);
 252                         done = true;
 253                 } else {
 254                         /*
 255                          * Can't have multiples of the same ref on a tree block.
 256                          */
 257                         WARN_ON(ref->type == BTRFS_TREE_BLOCK_REF_KEY ||
 258                                 ref->type == BTRFS_SHARED_BLOCK_REF_KEY);
 259                 }
 260 next:
 261                 next = next2;
 262         }
 263
 264         return done;
 265 }
 266
 267 void btrfs_merge_delayed_refs(struct btrfs_trans_handle *trans,
 268                               struct btrfs_fs_info *fs_info,
 269                               struct btrfs_delayed_ref_root *delayed_refs,
 270                               struct btrfs_delayed_ref_head *head)
 271 {
 272         struct btrfs_delayed_ref_node *ref;
 273         u64 seq = 0;
 274
 275         assert_spin_locked(&head->lock);
 276
 277         if (list_empty(&head->ref_list))
 278                 return;
 279
 280         /* We don't have too many refs to merge for data. */
 281         if (head->is_data)
 282                 return;
 283
 284         spin_lock(&fs_info->tree_mod_seq_lock);
 285         if (!list_empty(&fs_info->tree_mod_seq_list)) {
 286                 struct seq_list *elem;
 287
 288                 elem = list_first_entry(&fs_info->tree_mod_seq_list,
 289                                         struct seq_list, list);
 290                 seq = elem->seq;
 291         }
 292         spin_unlock(&fs_info->tree_mod_seq_lock);
 293
 294         ref = list_first_entry(&head->ref_list, struct btrfs_delayed_ref_node,
 295                                list);
 296         while (&ref->list != &head->ref_list) {
 297                 if (seq && ref->seq >= seq)
 298                         goto next;
 299
 300                 if (merge_ref(trans, delayed_refs, head, ref, seq)) {
 301                         if (list_empty(&head->ref_list))
 302                                 break;
 303                         ref = list_first_entry(&head->ref_list,
 304                                                struct btrfs_delayed_ref_node,
 305                                                list);
 306                         continue;
 307                 }
 308 next:
 309                 ref = list_next_entry(ref, list);
 310         }
 311 }
 312
 313 int btrfs_check_delayed_seq(struct btrfs_fs_info *fs_info,
 314                             struct btrfs_delayed_ref_root *delayed_refs,
 315                             u64 seq)
 316 {
 317         struct seq_list *elem;
 318         int ret = 0;
 319
 320         spin_lock(&fs_info->tree_mod_seq_lock);
 321         if (!list_empty(&fs_info->tree_mod_seq_list)) {
 322                 elem = list_first_entry(&fs_info->tree_mod_seq_list,
 323                                         struct seq_list, list);
 324                 if (seq >= elem->seq) {
 325                         pr_debug("holding back delayed_ref %#x.%x, lowest is %#x.%x (%p)\n",
 326                                  (u32)(seq >> 32), (u32)seq,
 327                                  (u32)(elem->seq >> 32), (u32)elem->seq,
 328                                  delayed_refs);
 329                         ret = 1;
 330                 }
 331         }
 332
 333         spin_unlock(&fs_info->tree_mod_seq_lock);
 334         return ret;
 335 }
 336
 337 struct btrfs_delayed_ref_head *
 338 btrfs_select_ref_head(struct btrfs_trans_handle *trans)
 339 {
 340         struct btrfs_delayed_ref_root *delayed_refs;
 341         struct btrfs_delayed_ref_head *head;
 342         u64 start;
 343         bool loop = false;
 344
 345         delayed_refs = &trans->transaction->delayed_refs;
 346
 347 again:
 348         start = delayed_refs->run_delayed_start;
 349         head = find_ref_head(&delayed_refs->href_root, start, 1);
 350         if (!head && !loop) {
 351                 delayed_refs->run_delayed_start = 0;
 352                 start = 0;
 353                 loop = true;
 354                 head = find_ref_head(&delayed_refs->href_root, start, 1);
 355                 if (!head)
 356                         return NULL;
 357         } else if (!head && loop) {
 358                 return NULL;
 359         }
 360
 361         while (head->processing) {
 362                 struct rb_node *node;
 363
 364                 node = rb_next(&head->href_node);
 365                 if (!node) {
 366                         if (loop)
 367                                 return NULL;
 368                         delayed_refs->run_delayed_start = 0;
 369                         start = 0;
 370                         loop = true;
 371                         goto again;
 372                 }
 373                 head = rb_entry(node, struct btrfs_delayed_ref_head,
 374                                 href_node);
 375         }
 376
 377         head->processing = 1;
 378         WARN_ON(delayed_refs->num_heads_ready == 0);
 379         delayed_refs->num_heads_ready--;
 380         delayed_refs->run_delayed_start = head->node.bytenr +
 381                 head->node.num_bytes;
 382         return head;
 383 }
 384
 385 /*
 386  * Helper to insert the ref_node to the tail or merge with tail.
 387  *
 388  * Return 0 for insert.
 389  * Return >0 for merge.
 390  */
 391 static int
 392 add_delayed_ref_tail_merge(struct btrfs_trans_handle *trans,
 393                            struct btrfs_delayed_ref_root *root,
 394                            struct btrfs_delayed_ref_head *href,
 395                            struct btrfs_delayed_ref_node *ref)
 396 {
 397         struct btrfs_delayed_ref_node *exist;
 398         int mod;
 399         int ret = 0;
 400
 401         spin_lock(&href->lock);
 402         /* Check whether we can merge the tail node with ref */
 403         if (list_empty(&href->ref_list))
 404                 goto add_tail;
 405         exist = list_entry(href->ref_list.prev, struct btrfs_delayed_ref_node,
 406                            list);
 407         /* No need to compare bytenr nor is_head */
 408         if (exist->type != ref->type || exist->no_quota != ref->no_quota ||
 409             exist->seq != ref->seq)
 410                 goto add_tail;
 411
 412         if ((exist->type == BTRFS_TREE_BLOCK_REF_KEY ||
 413              exist->type == BTRFS_SHARED_BLOCK_REF_KEY) &&
 414             comp_tree_refs(btrfs_delayed_node_to_tree_ref(exist),
 415                            btrfs_delayed_node_to_tree_ref(ref),
 416                            ref->type))
 417                 goto add_tail;
 418         if ((exist->type == BTRFS_EXTENT_DATA_REF_KEY ||
 419              exist->type == BTRFS_SHARED_DATA_REF_KEY) &&
 420             comp_data_refs(btrfs_delayed_node_to_data_ref(exist),
 421                            btrfs_delayed_node_to_data_ref(ref)))
 422                 goto add_tail;
 423
 424         /* Now we are sure we can merge */
 425         ret = 1;
 426         if (exist->action == ref->action) {
 427                 mod = ref->ref_mod;
 428         } else {
 429                 /* Need to change action */
 430                 if (exist->ref_mod < ref->ref_mod) {
 431                         exist->action = ref->action;
 432                         mod = -exist->ref_mod;
 433                         exist->ref_mod = ref->ref_mod;
 434                 } else
 435                         mod = -ref->ref_mod;
 436         }
 437         exist->ref_mod += mod;
 438
 439         /* remove existing tail if its ref_mod is zero */
 440         if (exist->ref_mod == 0)
 441                 drop_delayed_ref(trans, root, href, exist);
 442         spin_unlock(&href->lock);
 443         return ret;
 444
 445 add_tail:
 446         list_add_tail(&ref->list, &href->ref_list);
 447         atomic_inc(&root->num_entries);
 448         trans->delayed_ref_updates++;
 449         spin_unlock(&href->lock);
 450         return ret;
 451 }
 452
 453 /*
 454  * helper function to update the accounting in the head ref
 455  * existing and update must have the same bytenr
 456  */
 457 static noinline void
 458 update_existing_head_ref(struct btrfs_delayed_ref_root *delayed_refs,
 459                          struct btrfs_delayed_ref_node *existing,
 460                          struct btrfs_delayed_ref_node *update)
 461 {
 462         struct btrfs_delayed_ref_head *existing_ref;
 463         struct btrfs_delayed_ref_head *ref;
 464         int old_ref_mod;
 465
 466         existing_ref = btrfs_delayed_node_to_head(existing);
 467         ref = btrfs_delayed_node_to_head(update);
 468         BUG_ON(existing_ref->is_data != ref->is_data);
 469
 470         spin_lock(&existing_ref->lock);
 471         if (ref->must_insert_reserved) {
 472                 /* if the extent was freed and then
 473                  * reallocated before the delayed ref
 474                  * entries were processed, we can end up
 475                  * with an existing head ref without
 476                  * the must_insert_reserved flag set.
 477                  * Set it again here
 478                  */
 479                 existing_ref->must_insert_reserved = ref->must_insert_reserved;
 480
 481                 /*
 482                  * update the num_bytes so we make sure the accounting
 483                  * is done correctly
 484                  */
 485                 existing->num_bytes = update->num_bytes;
 486
 487         }
 488
 489         if (ref->extent_op) {
 490                 if (!existing_ref->extent_op) {
 491                         existing_ref->extent_op = ref->extent_op;
 492                 } else {
 493                         if (ref->extent_op->update_key) {
 494                                 memcpy(&existing_ref->extent_op->key,
 495                                        &ref->extent_op->key,
 496                                        sizeof(ref->extent_op->key));
 497                                 existing_ref->extent_op->update_key = 1;
 498                         }
 499                         if (ref->extent_op->update_flags) {
 500                                 existing_ref->extent_op->flags_to_set |=
 501                                         ref->extent_op->flags_to_set;
 502                                 existing_ref->extent_op->update_flags = 1;
 503                         }
 504                         btrfs_free_delayed_extent_op(ref->extent_op);
 505                 }
 506         }
 507         /*
 508          * update the reference mod on the head to reflect this new operation,
 509          * only need the lock for this case cause we could be processing it
 510          * currently, for refs we just added we know we're a-ok.
 511          */
 512         old_ref_mod = existing_ref->total_ref_mod;
 513         existing->ref_mod += update->ref_mod;
 514         existing_ref->total_ref_mod += update->ref_mod;
 515
 516         /*
 517          * If we are going to from a positive ref mod to a negative or vice
 518          * versa we need to make sure to adjust pending_csums accordingly.
 519          */
 520         if (existing_ref->is_data) {
 521                 if (existing_ref->total_ref_mod >= 0 && old_ref_mod < 0)
 522                         delayed_refs->pending_csums -= existing->num_bytes;
 523                 if (existing_ref->total_ref_mod < 0 && old_ref_mod >= 0)
 524                         delayed_refs->pending_csums += existing->num_bytes;
 525         }
 526         spin_unlock(&existing_ref->lock);
 527 }
 528
 529 /*
 530  * helper function to actually insert a head node into the rbtree.
 531  * this does all the dirty work in terms of maintaining the correct
 532  * overall modification count.
 533  */
 534 static noinline struct btrfs_delayed_ref_head *
 535 add_delayed_ref_head(struct btrfs_fs_info *fs_info,
 536                      struct btrfs_trans_handle *trans,
 537                      struct btrfs_delayed_ref_node *ref,
 538                      struct btrfs_qgroup_extent_record *qrecord,
 539                      u64 bytenr, u64 num_bytes, int action, int is_data)
 540 {
 541         struct btrfs_delayed_ref_head *existing;
 542         struct btrfs_delayed_ref_head *head_ref = NULL;
 543         struct btrfs_delayed_ref_root *delayed_refs;
 544         struct btrfs_qgroup_extent_record *qexisting;
 545         int count_mod = 1;
 546         int must_insert_reserved = 0;
 547
 548         /*
 549          * the head node stores the sum of all the mods, so dropping a ref
 550          * should drop the sum in the head node by one.
 551          */
 552         if (action == BTRFS_UPDATE_DELAYED_HEAD)
 553                 count_mod = 0;
 554         else if (action == BTRFS_DROP_DELAYED_REF)
 555                 count_mod = -1;
 556
 557         /*
 558          * BTRFS_ADD_DELAYED_EXTENT means that we need to update
 559          * the reserved accounting when the extent is finally added, or
 560          * if a later modification deletes the delayed ref without ever
 561          * inserting the extent into the extent allocation tree.
 562          * ref->must_insert_reserved is the flag used to record
 563          * that accounting mods are required.
 564          *
 565          * Once we record must_insert_reserved, switch the action to
 566          * BTRFS_ADD_DELAYED_REF because other special casing is not required.
 567          */
 568         if (action == BTRFS_ADD_DELAYED_EXTENT)
 569                 must_insert_reserved = 1;
 570         else
 571                 must_insert_reserved = 0;
 572
 573         delayed_refs = &trans->transaction->delayed_refs;
 574
 575         /* first set the basic ref node struct up */
 576         atomic_set(&ref->refs, 1);
 577         ref->bytenr = bytenr;
 578         ref->num_bytes = num_bytes;
 579         ref->ref_mod = count_mod;
 580         ref->type  = 0;
 581         ref->action  = 0;
 582         ref->is_head = 1;
 583         ref->in_tree = 1;
 584         ref->seq = 0;
 585
 586         head_ref = btrfs_delayed_node_to_head(ref);
 587         head_ref->must_insert_reserved = must_insert_reserved;
 588         head_ref->is_data = is_data;
 589         INIT_LIST_HEAD(&head_ref->ref_list);
 590         head_ref->processing = 0;
 591         head_ref->total_ref_mod = count_mod;
 592         head_ref->qgroup_reserved = 0;
 593         head_ref->qgroup_ref_root = 0;
 594
 595         /* Record qgroup extent info if provided */
 596         if (qrecord) {
 597                 qrecord->bytenr = bytenr;
 598                 qrecord->num_bytes = num_bytes;
 599                 qrecord->old_roots = NULL;
 600
 601                 qexisting = btrfs_qgroup_insert_dirty_extent(delayed_refs,
 602                                                              qrecord);
 603                 if (qexisting)
 604                         kfree(qrecord);
 605         }
 606
 607         spin_lock_init(&head_ref->lock);
 608         mutex_init(&head_ref->mutex);
 609
 610         trace_add_delayed_ref_head(ref, head_ref, action);
 611
 612         existing = htree_insert(&delayed_refs->href_root,
 613                                 &head_ref->href_node);
 614         if (existing) {
 615                 update_existing_head_ref(delayed_refs, &existing->node, ref);
 616                 /*
 617                  * we've updated the existing ref, free the newly
 618                  * allocated ref
 619                  */
 620                 kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
 621                 head_ref = existing;
 622         } else {
 623                 if (is_data && count_mod < 0)
 624                         delayed_refs->pending_csums += num_bytes;
 625                 delayed_refs->num_heads++;
 626                 delayed_refs->num_heads_ready++;
 627                 atomic_inc(&delayed_refs->num_entries);
 628                 trans->delayed_ref_updates++;
 629         }
 630         return head_ref;
 631 }
 632
 633 /*
 634  * helper to insert a delayed tree ref into the rbtree.
 635  */
 636 static noinline void
 637 add_delayed_tree_ref(struct btrfs_fs_info *fs_info,
 638                      struct btrfs_trans_handle *trans,
 639                      struct btrfs_delayed_ref_head *head_ref,
 640                      struct btrfs_delayed_ref_node *ref, u64 bytenr,
 641                      u64 num_bytes, u64 parent, u64 ref_root, int level,
 642                      int action, int no_quota)
 643 {
 644         struct btrfs_delayed_tree_ref *full_ref;
 645         struct btrfs_delayed_ref_root *delayed_refs;
 646         u64 seq = 0;
 647         int ret;
 648
 649         if (action == BTRFS_ADD_DELAYED_EXTENT)
 650                 action = BTRFS_ADD_DELAYED_REF;
 651
 652         if (is_fstree(ref_root))
 653                 seq = atomic64_read(&fs_info->tree_mod_seq);
 654         delayed_refs = &trans->transaction->delayed_refs;
 655
 656         /* first set the basic ref node struct up */
 657         atomic_set(&ref->refs, 1);
 658         ref->bytenr = bytenr;
 659         ref->num_bytes = num_bytes;
 660         ref->ref_mod = 1;
 661         ref->action = action;
 662         ref->is_head = 0;
 663         ref->in_tree = 1;
 664         ref->no_quota = no_quota;
 665         ref->seq = seq;
 666
 667         full_ref = btrfs_delayed_node_to_tree_ref(ref);
 668         full_ref->parent = parent;
 669         full_ref->root = ref_root;
 670         if (parent)
 671                 ref->type = BTRFS_SHARED_BLOCK_REF_KEY;
 672         else
 673                 ref->type = BTRFS_TREE_BLOCK_REF_KEY;
 674         full_ref->level = level;
 675
 676         trace_add_delayed_tree_ref(ref, full_ref, action);
 677
 678         ret = add_delayed_ref_tail_merge(trans, delayed_refs, head_ref, ref);
 679
 680         /*
 681          * XXX: memory should be freed at the same level allocated.
 682          * But bad practice is anywhere... Follow it now. Need cleanup.
 683          */
 684         if (ret > 0)
 685                 kmem_cache_free(btrfs_delayed_tree_ref_cachep, full_ref);
 686 }
 687
 688 /*
 689  * helper to insert a delayed data ref into the rbtree.
 690  */
 691 static noinline void
 692 add_delayed_data_ref(struct btrfs_fs_info *fs_info,
 693                      struct btrfs_trans_handle *trans,
 694                      struct btrfs_delayed_ref_head *head_ref,
 695                      struct btrfs_delayed_ref_node *ref, u64 bytenr,
 696                      u64 num_bytes, u64 parent, u64 ref_root, u64 owner,
 697                      u64 offset, int action, int no_quota)
 698 {
 699         struct btrfs_delayed_data_ref *full_ref;
 700         struct btrfs_delayed_ref_root *delayed_refs;
 701         u64 seq = 0;
 702         int ret;
 703
 704         if (action == BTRFS_ADD_DELAYED_EXTENT)
 705                 action = BTRFS_ADD_DELAYED_REF;
 706
 707         delayed_refs = &trans->transaction->delayed_refs;
 708
 709         if (is_fstree(ref_root))
 710                 seq = atomic64_read(&fs_info->tree_mod_seq);
 711
 712         /* first set the basic ref node struct up */
 713         atomic_set(&ref->refs, 1);
 714         ref->bytenr = bytenr;
 715         ref->num_bytes = num_bytes;
 716         ref->ref_mod = 1;
 717         ref->action = action;
 718         ref->is_head = 0;
 719         ref->in_tree = 1;
 720         ref->no_quota = no_quota;
 721         ref->seq = seq;
 722
 723         full_ref = btrfs_delayed_node_to_data_ref(ref);
 724         full_ref->parent = parent;
 725         full_ref->root = ref_root;
 726         if (parent)
 727                 ref->type = BTRFS_SHARED_DATA_REF_KEY;
 728         else
 729                 ref->type = BTRFS_EXTENT_DATA_REF_KEY;
 730
 731         full_ref->objectid = owner;
 732         full_ref->offset = offset;
 733
 734         trace_add_delayed_data_ref(ref, full_ref, action);
 735
 736         ret = add_delayed_ref_tail_merge(trans, delayed_refs, head_ref, ref);
 737
 738         if (ret > 0)
 739                 kmem_cache_free(btrfs_delayed_data_ref_cachep, full_ref);
 740 }
 741
 742 /*
 743  * add a delayed tree ref.  This does all of the accounting required
 744  * to make sure the delayed ref is eventually processed before this
 745  * transaction commits.
 746  */
 747 int btrfs_add_delayed_tree_ref(struct btrfs_fs_info *fs_info,
 748                                struct btrfs_trans_handle *trans,
 749                                u64 bytenr, u64 num_bytes, u64 parent,
 750                                u64 ref_root,  int level, int action,
 751                                struct btrfs_delayed_extent_op *extent_op,
 752                                int no_quota)
 753 {
 754         struct btrfs_delayed_tree_ref *ref;
 755         struct btrfs_delayed_ref_head *head_ref;
 756         struct btrfs_delayed_ref_root *delayed_refs;
 757         struct btrfs_qgroup_extent_record *record = NULL;
 758
 759         if (!is_fstree(ref_root) || !fs_info->quota_enabled)
 760                 no_quota = 0;
 761
 762         BUG_ON(extent_op && extent_op->is_data);
 763         ref = kmem_cache_alloc(btrfs_delayed_tree_ref_cachep, GFP_NOFS);
 764         if (!ref)
 765                 return -ENOMEM;
 766
 767         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
 768         if (!head_ref)
 769                 goto free_ref;
 770
 771         if (fs_info->quota_enabled && is_fstree(ref_root)) {
 772                 record = kmalloc(sizeof(*record), GFP_NOFS);
 773                 if (!record)
 774                         goto free_head_ref;
 775         }
 776
 777         head_ref->extent_op = extent_op;
 778
 779         delayed_refs = &trans->transaction->delayed_refs;
 780         spin_lock(&delayed_refs->lock);
 781
 782         /*
 783          * insert both the head node and the new ref without dropping
 784          * the spin lock
 785          */
 786         head_ref = add_delayed_ref_head(fs_info, trans, &head_ref->node, record,
 787                                         bytenr, num_bytes, action, 0);
 788
 789         add_delayed_tree_ref(fs_info, trans, head_ref, &ref->node, bytenr,
 790                                    num_bytes, parent, ref_root, level, action,
 791                                    no_quota);
 792         spin_unlock(&delayed_refs->lock);
 793
 794         return 0;
 795
 796 free_head_ref:
 797         kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
 798 free_ref:
 799         kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
 800
 801         return -ENOMEM;
 802 }
 803
 804 /*
 805  * add a delayed data ref. it's similar to btrfs_add_delayed_tree_ref.
 806  */
 807 int btrfs_add_delayed_data_ref(struct btrfs_fs_info *fs_info,
 808                                struct btrfs_trans_handle *trans,
 809                                u64 bytenr, u64 num_bytes,
 810                                u64 parent, u64 ref_root,
 811                                u64 owner, u64 offset, int action,
 812                                struct btrfs_delayed_extent_op *extent_op,
 813                                int no_quota)
 814 {
 815         struct btrfs_delayed_data_ref *ref;
 816         struct btrfs_delayed_ref_head *head_ref;
 817         struct btrfs_delayed_ref_root *delayed_refs;
 818         struct btrfs_qgroup_extent_record *record = NULL;
 819
 820         if (!is_fstree(ref_root) || !fs_info->quota_enabled)
 821                 no_quota = 0;
 822
 823         BUG_ON(extent_op && !extent_op->is_data);
 824         ref = kmem_cache_alloc(btrfs_delayed_data_ref_cachep, GFP_NOFS);
 825         if (!ref)
 826                 return -ENOMEM;
 827
 828         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
 829         if (!head_ref) {
 830                 kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
 831                 return -ENOMEM;
 832         }
 833
 834         if (fs_info->quota_enabled && is_fstree(ref_root)) {
 835                 record = kmalloc(sizeof(*record), GFP_NOFS);
 836                 if (!record) {
 837                         kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
 838                         kmem_cache_free(btrfs_delayed_ref_head_cachep,
 839                                         head_ref);
 840                         return -ENOMEM;
 841                 }
 842         }
 843
 844         head_ref->extent_op = extent_op;
 845
 846         delayed_refs = &trans->transaction->delayed_refs;
 847         spin_lock(&delayed_refs->lock);
 848
 849         /*
 850          * insert both the head node and the new ref without dropping
 851          * the spin lock
 852          */
 853         head_ref = add_delayed_ref_head(fs_info, trans, &head_ref->node, record,
 854                                         bytenr, num_bytes, action, 1);
 855
 856         add_delayed_data_ref(fs_info, trans, head_ref, &ref->node, bytenr,
 857                                    num_bytes, parent, ref_root, owner, offset,
 858                                    action, no_quota);
 859         spin_unlock(&delayed_refs->lock);
 860
 861         return 0;
 862 }
 863
 864 int btrfs_add_delayed_qgroup_reserve(struct btrfs_fs_info *fs_info,
 865                                      struct btrfs_trans_handle *trans,
 866                                      u64 ref_root, u64 bytenr, u64 num_bytes)
 867 {
 868         struct btrfs_delayed_ref_root *delayed_refs;
 869         struct btrfs_delayed_ref_head *ref_head;
 870         int ret = 0;
 871
 872         if (!fs_info->quota_enabled || !is_fstree(ref_root))
 873                 return 0;
 874
 875         delayed_refs = &trans->transaction->delayed_refs;
 876
 877         spin_lock(&delayed_refs->lock);
 878         ref_head = find_ref_head(&delayed_refs->href_root, bytenr, 0);
 879         if (!ref_head) {
 880                 ret = -ENOENT;
 881                 goto out;
 882         }
 883         WARN_ON(ref_head->qgroup_reserved || ref_head->qgroup_ref_root);
 884         ref_head->qgroup_ref_root = ref_root;
 885         ref_head->qgroup_reserved = num_bytes;
 886 out:
 887         spin_unlock(&delayed_refs->lock);
 888         return ret;
 889 }
 890
 891 int btrfs_add_delayed_extent_op(struct btrfs_fs_info *fs_info,
 892                                 struct btrfs_trans_handle *trans,
 893                                 u64 bytenr, u64 num_bytes,
 894                                 struct btrfs_delayed_extent_op *extent_op)
 895 {
 896         struct btrfs_delayed_ref_head *head_ref;
 897         struct btrfs_delayed_ref_root *delayed_refs;
 898
 899         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
 900         if (!head_ref)
 901                 return -ENOMEM;
 902
 903         head_ref->extent_op = extent_op;
 904
 905         delayed_refs = &trans->transaction->delayed_refs;
 906         spin_lock(&delayed_refs->lock);
 907
 908         add_delayed_ref_head(fs_info, trans, &head_ref->node, NULL, bytenr,
 909                              num_bytes, BTRFS_UPDATE_DELAYED_HEAD,
 910                              extent_op->is_data);
 911
 912         spin_unlock(&delayed_refs->lock);
 913         return 0;
 914 }
 915
 916 /*
 917  * this does a simple search for the head node for a given extent.
 918  * It must be called with the delayed ref spinlock held, and it returns
 919  * the head node if any where found, or NULL if not.
 920  */
 921 struct btrfs_delayed_ref_head *
 922 btrfs_find_delayed_ref_head(struct btrfs_trans_handle *trans, u64 bytenr)
 923 {
 924         struct btrfs_delayed_ref_root *delayed_refs;
 925
 926         delayed_refs = &trans->transaction->delayed_refs;
 927         return find_ref_head(&delayed_refs->href_root, bytenr, 0);
 928 }
 929
 930 void btrfs_delayed_ref_exit(void)
 931 {
 932         if (btrfs_delayed_ref_head_cachep)
 933                 kmem_cache_destroy(btrfs_delayed_ref_head_cachep);
 934         if (btrfs_delayed_tree_ref_cachep)
 935                 kmem_cache_destroy(btrfs_delayed_tree_ref_cachep);
 936         if (btrfs_delayed_data_ref_cachep)
 937                 kmem_cache_destroy(btrfs_delayed_data_ref_cachep);
 938         if (btrfs_delayed_extent_op_cachep)
 939                 kmem_cache_destroy(btrfs_delayed_extent_op_cachep);
 940 }
 941
 942 int btrfs_delayed_ref_init(void)
 943 {
 944         btrfs_delayed_ref_head_cachep = kmem_cache_create(
 945                                 "btrfs_delayed_ref_head",
 946                                 sizeof(struct btrfs_delayed_ref_head), 0,
 947                                 SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 948         if (!btrfs_delayed_ref_head_cachep)
 949                 goto fail;
 950
 951         btrfs_delayed_tree_ref_cachep = kmem_cache_create(
 952                                 "btrfs_delayed_tree_ref",
 953                                 sizeof(struct btrfs_delayed_tree_ref), 0,
 954                                 SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 955         if (!btrfs_delayed_tree_ref_cachep)
 956                 goto fail;
 957
 958         btrfs_delayed_data_ref_cachep = kmem_cache_create(
 959                                 "btrfs_delayed_data_ref",
 960                                 sizeof(struct btrfs_delayed_data_ref), 0,
 961                                 SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 962         if (!btrfs_delayed_data_ref_cachep)
 963                 goto fail;
 964
 965         btrfs_delayed_extent_op_cachep = kmem_cache_create(
 966                                 "btrfs_delayed_extent_op",
 967                                 sizeof(struct btrfs_delayed_extent_op), 0,
 968                                 SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
 969         if (!btrfs_delayed_extent_op_cachep)
 970                 goto fail;
 971
 972         return 0;
 973 fail:
 974         btrfs_delayed_ref_exit();
 975         return -ENOMEM;
 976 }