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)
 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->seq != ref->seq)
 409                 goto add_tail;
 410
 411         if ((exist->type == BTRFS_TREE_BLOCK_REF_KEY ||
 412              exist->type == BTRFS_SHARED_BLOCK_REF_KEY) &&
 413             comp_tree_refs(btrfs_delayed_node_to_tree_ref(exist),
 414                            btrfs_delayed_node_to_tree_ref(ref),
 415                            ref->type))
 416                 goto add_tail;
 417         if ((exist->type == BTRFS_EXTENT_DATA_REF_KEY ||
 418              exist->type == BTRFS_SHARED_DATA_REF_KEY) &&
 419             comp_data_refs(btrfs_delayed_node_to_data_ref(exist),
 420                            btrfs_delayed_node_to_data_ref(ref)))
 421                 goto add_tail;
 422
 423         /* Now we are sure we can merge */
 424         ret = 1;
 425         if (exist->action == ref->action) {
 426                 mod = ref->ref_mod;
 427         } else {
 428                 /* Need to change action */
 429                 if (exist->ref_mod < ref->ref_mod) {
 430                         exist->action = ref->action;
 431                         mod = -exist->ref_mod;
 432                         exist->ref_mod = ref->ref_mod;
 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
 444 add_tail:
 445         list_add_tail(&ref->list, &href->ref_list);
 446         atomic_inc(&root->num_entries);
 447         trans->delayed_ref_updates++;
 448         spin_unlock(&href->lock);
 449         return ret;
 450 }
 451
 452 /*
 453  * helper function to update the accounting in the head ref
 454  * existing and update must have the same bytenr
 455  */
 456 static noinline void
 457 update_existing_head_ref(struct btrfs_delayed_ref_root *delayed_refs,
 458                          struct btrfs_delayed_ref_node *existing,
 459                          struct btrfs_delayed_ref_node *update)
 460 {
 461         struct btrfs_delayed_ref_head *existing_ref;
 462         struct btrfs_delayed_ref_head *ref;
 463         int old_ref_mod;
 464
 465         existing_ref = btrfs_delayed_node_to_head(existing);
 466         ref = btrfs_delayed_node_to_head(update);
 467         BUG_ON(existing_ref->is_data != ref->is_data);
 468
 469         spin_lock(&existing_ref->lock);
 470         if (ref->must_insert_reserved) {
 471                 /* if the extent was freed and then
 472                  * reallocated before the delayed ref
 473                  * entries were processed, we can end up
 474                  * with an existing head ref without
 475                  * the must_insert_reserved flag set.
 476                  * Set it again here
 477                  */
 478                 existing_ref->must_insert_reserved = ref->must_insert_reserved;
 479
 480                 /*
 481                  * update the num_bytes so we make sure the accounting
 482                  * is done correctly
 483                  */
 484                 existing->num_bytes = update->num_bytes;
 485
 486         }
 487
 488         if (ref->extent_op) {
 489                 if (!existing_ref->extent_op) {
 490                         existing_ref->extent_op = ref->extent_op;
 491                 } else {
 492                         if (ref->extent_op->update_key) {
 493                                 memcpy(&existing_ref->extent_op->key,
 494                                        &ref->extent_op->key,
 495                                        sizeof(ref->extent_op->key));
 496                                 existing_ref->extent_op->update_key = true;
 497                         }
 498                         if (ref->extent_op->update_flags) {
 499                                 existing_ref->extent_op->flags_to_set |=
 500                                         ref->extent_op->flags_to_set;
 501                                 existing_ref->extent_op->update_flags = true;
 502                         }
 503                         btrfs_free_delayed_extent_op(ref->extent_op);
 504                 }
 505         }
 506         /*
 507          * update the reference mod on the head to reflect this new operation,
 508          * only need the lock for this case cause we could be processing it
 509          * currently, for refs we just added we know we're a-ok.
 510          */
 511         old_ref_mod = existing_ref->total_ref_mod;
 512         existing->ref_mod += update->ref_mod;
 513         existing_ref->total_ref_mod += update->ref_mod;
 514
 515         /*
 516          * If we are going to from a positive ref mod to a negative or vice
 517          * versa we need to make sure to adjust pending_csums accordingly.
 518          */
 519         if (existing_ref->is_data) {
 520                 if (existing_ref->total_ref_mod >= 0 && old_ref_mod < 0)
 521                         delayed_refs->pending_csums -= existing->num_bytes;
 522                 if (existing_ref->total_ref_mod < 0 && old_ref_mod >= 0)
 523                         delayed_refs->pending_csums += existing->num_bytes;
 524         }
 525         spin_unlock(&existing_ref->lock);
 526 }
 527
 528 /*
 529  * helper function to actually insert a head node into the rbtree.
 530  * this does all the dirty work in terms of maintaining the correct
 531  * overall modification count.
 532  */
 533 static noinline struct btrfs_delayed_ref_head *
 534 add_delayed_ref_head(struct btrfs_fs_info *fs_info,
 535                      struct btrfs_trans_handle *trans,
 536                      struct btrfs_delayed_ref_node *ref,
 537                      struct btrfs_qgroup_extent_record *qrecord,
 538                      u64 bytenr, u64 num_bytes, u64 ref_root, u64 reserved,
 539                      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         /* If reserved is provided, it must be a data extent. */
 549         BUG_ON(!is_data && reserved);
 550
 551         /*
 552          * the head node stores the sum of all the mods, so dropping a ref
 553          * should drop the sum in the head node by one.
 554          */
 555         if (action == BTRFS_UPDATE_DELAYED_HEAD)
 556                 count_mod = 0;
 557         else if (action == BTRFS_DROP_DELAYED_REF)
 558                 count_mod = -1;
 559
 560         /*
 561          * BTRFS_ADD_DELAYED_EXTENT means that we need to update
 562          * the reserved accounting when the extent is finally added, or
 563          * if a later modification deletes the delayed ref without ever
 564          * inserting the extent into the extent allocation tree.
 565          * ref->must_insert_reserved is the flag used to record
 566          * that accounting mods are required.
 567          *
 568          * Once we record must_insert_reserved, switch the action to
 569          * BTRFS_ADD_DELAYED_REF because other special casing is not required.
 570          */
 571         if (action == BTRFS_ADD_DELAYED_EXTENT)
 572                 must_insert_reserved = 1;
 573         else
 574                 must_insert_reserved = 0;
 575
 576         delayed_refs = &trans->transaction->delayed_refs;
 577
 578         /* first set the basic ref node struct up */
 579         atomic_set(&ref->refs, 1);
 580         ref->bytenr = bytenr;
 581         ref->num_bytes = num_bytes;
 582         ref->ref_mod = count_mod;
 583         ref->type  = 0;
 584         ref->action  = 0;
 585         ref->is_head = 1;
 586         ref->in_tree = 1;
 587         ref->seq = 0;
 588
 589         head_ref = btrfs_delayed_node_to_head(ref);
 590         head_ref->must_insert_reserved = must_insert_reserved;
 591         head_ref->is_data = is_data;
 592         INIT_LIST_HEAD(&head_ref->ref_list);
 593         head_ref->processing = 0;
 594         head_ref->total_ref_mod = count_mod;
 595         head_ref->qgroup_reserved = 0;
 596         head_ref->qgroup_ref_root = 0;
 597
 598         /* Record qgroup extent info if provided */
 599         if (qrecord) {
 600                 if (ref_root && reserved) {
 601                         head_ref->qgroup_ref_root = ref_root;
 602                         head_ref->qgroup_reserved = reserved;
 603                 }
 604
 605                 qrecord->bytenr = bytenr;
 606                 qrecord->num_bytes = num_bytes;
 607                 qrecord->old_roots = NULL;
 608
 609                 qexisting = btrfs_qgroup_insert_dirty_extent(fs_info,
 610                                                              delayed_refs,
 611                                                              qrecord);
 612                 if (qexisting)
 613                         kfree(qrecord);
 614         }
 615
 616         spin_lock_init(&head_ref->lock);
 617         mutex_init(&head_ref->mutex);
 618
 619         trace_add_delayed_ref_head(fs_info, ref, head_ref, action);
 620
 621         existing = htree_insert(&delayed_refs->href_root,
 622                                 &head_ref->href_node);
 623         if (existing) {
 624                 WARN_ON(ref_root && reserved && existing->qgroup_ref_root
 625                         && existing->qgroup_reserved);
 626                 update_existing_head_ref(delayed_refs, &existing->node, ref);
 627                 /*
 628                  * we've updated the existing ref, free the newly
 629                  * allocated ref
 630                  */
 631                 kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
 632                 head_ref = existing;
 633         } else {
 634                 if (is_data && count_mod < 0)
 635                         delayed_refs->pending_csums += num_bytes;
 636                 delayed_refs->num_heads++;
 637                 delayed_refs->num_heads_ready++;
 638                 atomic_inc(&delayed_refs->num_entries);
 639                 trans->delayed_ref_updates++;
 640         }
 641         return head_ref;
 642 }
 643
 644 /*
 645  * helper to insert a delayed tree ref into the rbtree.
 646  */
 647 static noinline void
 648 add_delayed_tree_ref(struct btrfs_fs_info *fs_info,
 649                      struct btrfs_trans_handle *trans,
 650                      struct btrfs_delayed_ref_head *head_ref,
 651                      struct btrfs_delayed_ref_node *ref, u64 bytenr,
 652                      u64 num_bytes, u64 parent, u64 ref_root, int level,
 653                      int action)
 654 {
 655         struct btrfs_delayed_tree_ref *full_ref;
 656         struct btrfs_delayed_ref_root *delayed_refs;
 657         u64 seq = 0;
 658         int ret;
 659
 660         if (action == BTRFS_ADD_DELAYED_EXTENT)
 661                 action = BTRFS_ADD_DELAYED_REF;
 662
 663         if (is_fstree(ref_root))
 664                 seq = atomic64_read(&fs_info->tree_mod_seq);
 665         delayed_refs = &trans->transaction->delayed_refs;
 666
 667         /* first set the basic ref node struct up */
 668         atomic_set(&ref->refs, 1);
 669         ref->bytenr = bytenr;
 670         ref->num_bytes = num_bytes;
 671         ref->ref_mod = 1;
 672         ref->action = action;
 673         ref->is_head = 0;
 674         ref->in_tree = 1;
 675         ref->seq = seq;
 676
 677         full_ref = btrfs_delayed_node_to_tree_ref(ref);
 678         full_ref->parent = parent;
 679         full_ref->root = ref_root;
 680         if (parent)
 681                 ref->type = BTRFS_SHARED_BLOCK_REF_KEY;
 682         else
 683                 ref->type = BTRFS_TREE_BLOCK_REF_KEY;
 684         full_ref->level = level;
 685
 686         trace_add_delayed_tree_ref(fs_info, ref, full_ref, action);
 687
 688         ret = add_delayed_ref_tail_merge(trans, delayed_refs, head_ref, ref);
 689
 690         /*
 691          * XXX: memory should be freed at the same level allocated.
 692          * But bad practice is anywhere... Follow it now. Need cleanup.
 693          */
 694         if (ret > 0)
 695                 kmem_cache_free(btrfs_delayed_tree_ref_cachep, full_ref);
 696 }
 697
 698 /*
 699  * helper to insert a delayed data ref into the rbtree.
 700  */
 701 static noinline void
 702 add_delayed_data_ref(struct btrfs_fs_info *fs_info,
 703                      struct btrfs_trans_handle *trans,
 704                      struct btrfs_delayed_ref_head *head_ref,
 705                      struct btrfs_delayed_ref_node *ref, u64 bytenr,
 706                      u64 num_bytes, u64 parent, u64 ref_root, u64 owner,
 707                      u64 offset, int action)
 708 {
 709         struct btrfs_delayed_data_ref *full_ref;
 710         struct btrfs_delayed_ref_root *delayed_refs;
 711         u64 seq = 0;
 712         int ret;
 713
 714         if (action == BTRFS_ADD_DELAYED_EXTENT)
 715                 action = BTRFS_ADD_DELAYED_REF;
 716
 717         delayed_refs = &trans->transaction->delayed_refs;
 718
 719         if (is_fstree(ref_root))
 720                 seq = atomic64_read(&fs_info->tree_mod_seq);
 721
 722         /* first set the basic ref node struct up */
 723         atomic_set(&ref->refs, 1);
 724         ref->bytenr = bytenr;
 725         ref->num_bytes = num_bytes;
 726         ref->ref_mod = 1;
 727         ref->action = action;
 728         ref->is_head = 0;
 729         ref->in_tree = 1;
 730         ref->seq = seq;
 731
 732         full_ref = btrfs_delayed_node_to_data_ref(ref);
 733         full_ref->parent = parent;
 734         full_ref->root = ref_root;
 735         if (parent)
 736                 ref->type = BTRFS_SHARED_DATA_REF_KEY;
 737         else
 738                 ref->type = BTRFS_EXTENT_DATA_REF_KEY;
 739
 740         full_ref->objectid = owner;
 741         full_ref->offset = offset;
 742
 743         trace_add_delayed_data_ref(fs_info, ref, full_ref, action);
 744
 745         ret = add_delayed_ref_tail_merge(trans, delayed_refs, head_ref, ref);
 746
 747         if (ret > 0)
 748                 kmem_cache_free(btrfs_delayed_data_ref_cachep, full_ref);
 749 }
 750
 751 /*
 752  * add a delayed tree ref.  This does all of the accounting required
 753  * to make sure the delayed ref is eventually processed before this
 754  * transaction commits.
 755  */
 756 int btrfs_add_delayed_tree_ref(struct btrfs_fs_info *fs_info,
 757                                struct btrfs_trans_handle *trans,
 758                                u64 bytenr, u64 num_bytes, u64 parent,
 759                                u64 ref_root,  int level, int action,
 760                                struct btrfs_delayed_extent_op *extent_op)
 761 {
 762         struct btrfs_delayed_tree_ref *ref;
 763         struct btrfs_delayed_ref_head *head_ref;
 764         struct btrfs_delayed_ref_root *delayed_refs;
 765         struct btrfs_qgroup_extent_record *record = NULL;
 766
 767         BUG_ON(extent_op && extent_op->is_data);
 768         ref = kmem_cache_alloc(btrfs_delayed_tree_ref_cachep, GFP_NOFS);
 769         if (!ref)
 770                 return -ENOMEM;
 771
 772         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
 773         if (!head_ref)
 774                 goto free_ref;
 775
 776         if (fs_info->quota_enabled && is_fstree(ref_root)) {
 777                 record = kmalloc(sizeof(*record), GFP_NOFS);
 778                 if (!record)
 779                         goto free_head_ref;
 780         }
 781
 782         head_ref->extent_op = extent_op;
 783
 784         delayed_refs = &trans->transaction->delayed_refs;
 785         spin_lock(&delayed_refs->lock);
 786
 787         /*
 788          * insert both the head node and the new ref without dropping
 789          * the spin lock
 790          */
 791         head_ref = add_delayed_ref_head(fs_info, trans, &head_ref->node, record,
 792                                         bytenr, num_bytes, 0, 0, action, 0);
 793
 794         add_delayed_tree_ref(fs_info, trans, head_ref, &ref->node, bytenr,
 795                              num_bytes, parent, ref_root, level, action);
 796         spin_unlock(&delayed_refs->lock);
 797
 798         return 0;
 799
 800 free_head_ref:
 801         kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
 802 free_ref:
 803         kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
 804
 805         return -ENOMEM;
 806 }
 807
 808 /*
 809  * add a delayed data ref. it's similar to btrfs_add_delayed_tree_ref.
 810  */
 811 int btrfs_add_delayed_data_ref(struct btrfs_fs_info *fs_info,
 812                                struct btrfs_trans_handle *trans,
 813                                u64 bytenr, u64 num_bytes,
 814                                u64 parent, u64 ref_root,
 815                                u64 owner, u64 offset, u64 reserved, int action,
 816                                struct btrfs_delayed_extent_op *extent_op)
 817 {
 818         struct btrfs_delayed_data_ref *ref;
 819         struct btrfs_delayed_ref_head *head_ref;
 820         struct btrfs_delayed_ref_root *delayed_refs;
 821         struct btrfs_qgroup_extent_record *record = NULL;
 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, ref_root, reserved,
 855                                         action, 1);
 856
 857         add_delayed_data_ref(fs_info, trans, head_ref, &ref->node, bytenr,
 858                                    num_bytes, parent, ref_root, owner, offset,
 859                                    action);
 860         spin_unlock(&delayed_refs->lock);
 861
 862         return 0;
 863 }
 864
 865 int btrfs_add_delayed_qgroup_reserve(struct btrfs_fs_info *fs_info,
 866                                      struct btrfs_trans_handle *trans,
 867                                      u64 ref_root, u64 bytenr, u64 num_bytes)
 868 {
 869         struct btrfs_delayed_ref_root *delayed_refs;
 870         struct btrfs_delayed_ref_head *ref_head;
 871         int ret = 0;
 872
 873         if (!fs_info->quota_enabled || !is_fstree(ref_root))
 874                 return 0;
 875
 876         delayed_refs = &trans->transaction->delayed_refs;
 877
 878         spin_lock(&delayed_refs->lock);
 879         ref_head = find_ref_head(&delayed_refs->href_root, bytenr, 0);
 880         if (!ref_head) {
 881                 ret = -ENOENT;
 882                 goto out;
 883         }
 884         WARN_ON(ref_head->qgroup_reserved || ref_head->qgroup_ref_root);
 885         ref_head->qgroup_ref_root = ref_root;
 886         ref_head->qgroup_reserved = num_bytes;
 887 out:
 888         spin_unlock(&delayed_refs->lock);
 889         return ret;
 890 }
 891
 892 int btrfs_add_delayed_extent_op(struct btrfs_fs_info *fs_info,
 893                                 struct btrfs_trans_handle *trans,
 894                                 u64 bytenr, u64 num_bytes,
 895                                 struct btrfs_delayed_extent_op *extent_op)
 896 {
 897         struct btrfs_delayed_ref_head *head_ref;
 898         struct btrfs_delayed_ref_root *delayed_refs;
 899
 900         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
 901         if (!head_ref)
 902                 return -ENOMEM;
 903
 904         head_ref->extent_op = extent_op;
 905
 906         delayed_refs = &trans->transaction->delayed_refs;
 907         spin_lock(&delayed_refs->lock);
 908
 909         add_delayed_ref_head(fs_info, trans, &head_ref->node, NULL, bytenr,
 910                              num_bytes, 0, 0, BTRFS_UPDATE_DELAYED_HEAD,
 911                              extent_op->is_data);
 912
 913         spin_unlock(&delayed_refs->lock);
 914         return 0;
 915 }
 916
 917 /*
 918  * this does a simple search for the head node for a given extent.
 919  * It must be called with the delayed ref spinlock held, and it returns
 920  * the head node if any where found, or NULL if not.
 921  */
 922 struct btrfs_delayed_ref_head *
 923 btrfs_find_delayed_ref_head(struct btrfs_trans_handle *trans, u64 bytenr)
 924 {
 925         struct btrfs_delayed_ref_root *delayed_refs;
 926
 927         delayed_refs = &trans->transaction->delayed_refs;
 928         return find_ref_head(&delayed_refs->href_root, bytenr, 0);
 929 }
 930
 931 void btrfs_delayed_ref_exit(void)
 932 {
 933         kmem_cache_destroy(btrfs_delayed_ref_head_cachep);
 934         kmem_cache_destroy(btrfs_delayed_tree_ref_cachep);
 935         kmem_cache_destroy(btrfs_delayed_data_ref_cachep);
 936         kmem_cache_destroy(btrfs_delayed_extent_op_cachep);
 937 }
 938
 939 int btrfs_delayed_ref_init(void)
 940 {
 941         btrfs_delayed_ref_head_cachep = kmem_cache_create(
 942                                 "btrfs_delayed_ref_head",
 943                                 sizeof(struct btrfs_delayed_ref_head), 0,
 944                                 SLAB_MEM_SPREAD, NULL);
 945         if (!btrfs_delayed_ref_head_cachep)
 946                 goto fail;
 947
 948         btrfs_delayed_tree_ref_cachep = kmem_cache_create(
 949                                 "btrfs_delayed_tree_ref",
 950                                 sizeof(struct btrfs_delayed_tree_ref), 0,
 951                                 SLAB_MEM_SPREAD, NULL);
 952         if (!btrfs_delayed_tree_ref_cachep)
 953                 goto fail;
 954
 955         btrfs_delayed_data_ref_cachep = kmem_cache_create(
 956                                 "btrfs_delayed_data_ref",
 957                                 sizeof(struct btrfs_delayed_data_ref), 0,
 958                                 SLAB_MEM_SPREAD, NULL);
 959         if (!btrfs_delayed_data_ref_cachep)
 960                 goto fail;
 961
 962         btrfs_delayed_extent_op_cachep = kmem_cache_create(
 963                                 "btrfs_delayed_extent_op",
 964                                 sizeof(struct btrfs_delayed_extent_op), 0,
 965                                 SLAB_MEM_SPREAD, NULL);
 966         if (!btrfs_delayed_extent_op_cachep)
 967                 goto fail;
 968
 969         return 0;
 970 fail:
 971         btrfs_delayed_ref_exit();
 972         return -ENOMEM;
 973 }