fs/btrfs/delayed-ref.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2009 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/slab.h>
   8 #include <linux/sort.h>
   9 #include "messages.h"
  10 #include "ctree.h"
  11 #include "delayed-ref.h"
  12 #include "transaction.h"
  13 #include "qgroup.h"
  14 #include "space-info.h"
  15 #include "tree-mod-log.h"
  16 #include "fs.h"
  17
  18 struct kmem_cache *btrfs_delayed_ref_head_cachep;
  19 struct kmem_cache *btrfs_delayed_tree_ref_cachep;
  20 struct kmem_cache *btrfs_delayed_data_ref_cachep;
  21 struct kmem_cache *btrfs_delayed_extent_op_cachep;
  22 /*
  23  * delayed back reference update tracking.  For subvolume trees
  24  * we queue up extent allocations and backref maintenance for
  25  * delayed processing.   This avoids deep call chains where we
  26  * add extents in the middle of btrfs_search_slot, and it allows
  27  * us to buffer up frequently modified backrefs in an rb tree instead
  28  * of hammering updates on the extent allocation tree.
  29  */
  30
  31 bool btrfs_check_space_for_delayed_refs(struct btrfs_fs_info *fs_info)
  32 {
  33         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
  34         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
  35         bool ret = false;
  36         u64 reserved;
  37
  38         spin_lock(&global_rsv->lock);
  39         reserved = global_rsv->reserved;
  40         spin_unlock(&global_rsv->lock);
  41
  42         /*
  43          * Since the global reserve is just kind of magic we don't really want
  44          * to rely on it to save our bacon, so if our size is more than the
  45          * delayed_refs_rsv and the global rsv then it's time to think about
  46          * bailing.
  47          */
  48         spin_lock(&delayed_refs_rsv->lock);
  49         reserved += delayed_refs_rsv->reserved;
  50         if (delayed_refs_rsv->size >= reserved)
  51                 ret = true;
  52         spin_unlock(&delayed_refs_rsv->lock);
  53         return ret;
  54 }
  55
  56 /*
  57  * Release a ref head's reservation.
  58  *
  59  * @fs_info:  the filesystem
  60  * @nr:       number of items to drop
  61  *
  62  * Drops the delayed ref head's count from the delayed refs rsv and free any
  63  * excess reservation we had.
  64  */
  65 void btrfs_delayed_refs_rsv_release(struct btrfs_fs_info *fs_info, int nr)
  66 {
  67         struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
  68         const u64 num_bytes = btrfs_calc_delayed_ref_bytes(fs_info, nr);
  69         u64 released = 0;
  70
  71         released = btrfs_block_rsv_release(fs_info, block_rsv, num_bytes, NULL);
  72         if (released)
  73                 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
  74                                               0, released, 0);
  75 }
  76
  77 /*
  78  * Adjust the size of the delayed refs rsv.
  79  *
  80  * This is to be called anytime we may have adjusted trans->delayed_ref_updates,
  81  * it'll calculate the additional size and add it to the delayed_refs_rsv.
  82  */
  83 void btrfs_update_delayed_refs_rsv(struct btrfs_trans_handle *trans)
  84 {
  85         struct btrfs_fs_info *fs_info = trans->fs_info;
  86         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_refs_rsv;
  87         u64 num_bytes;
  88
  89         if (!trans->delayed_ref_updates)
  90                 return;
  91
  92         num_bytes = btrfs_calc_delayed_ref_bytes(fs_info,
  93                                                  trans->delayed_ref_updates);
  94
  95         spin_lock(&delayed_rsv->lock);
  96         delayed_rsv->size += num_bytes;
  97         delayed_rsv->full = false;
  98         spin_unlock(&delayed_rsv->lock);
  99         trans->delayed_ref_updates = 0;
 100 }
 101
 102 /*
 103  * Transfer bytes to our delayed refs rsv.
 104  *
 105  * @fs_info:   the filesystem
 106  * @src:       source block rsv to transfer from
 107  * @num_bytes: number of bytes to transfer
 108  *
 109  * This transfers up to the num_bytes amount from the src rsv to the
 110  * delayed_refs_rsv.  Any extra bytes are returned to the space info.
 111  */
 112 void btrfs_migrate_to_delayed_refs_rsv(struct btrfs_fs_info *fs_info,
 113                                        struct btrfs_block_rsv *src,
 114                                        u64 num_bytes)
 115 {
 116         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
 117         u64 to_free = 0;
 118
 119         spin_lock(&src->lock);
 120         src->reserved -= num_bytes;
 121         src->size -= num_bytes;
 122         spin_unlock(&src->lock);
 123
 124         spin_lock(&delayed_refs_rsv->lock);
 125         if (delayed_refs_rsv->size > delayed_refs_rsv->reserved) {
 126                 u64 delta = delayed_refs_rsv->size -
 127                         delayed_refs_rsv->reserved;
 128                 if (num_bytes > delta) {
 129                         to_free = num_bytes - delta;
 130                         num_bytes = delta;
 131                 }
 132         } else {
 133                 to_free = num_bytes;
 134                 num_bytes = 0;
 135         }
 136
 137         if (num_bytes)
 138                 delayed_refs_rsv->reserved += num_bytes;
 139         if (delayed_refs_rsv->reserved >= delayed_refs_rsv->size)
 140                 delayed_refs_rsv->full = true;
 141         spin_unlock(&delayed_refs_rsv->lock);
 142
 143         if (num_bytes)
 144                 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
 145                                               0, num_bytes, 1);
 146         if (to_free)
 147                 btrfs_space_info_free_bytes_may_use(fs_info,
 148                                 delayed_refs_rsv->space_info, to_free);
 149 }
 150
 151 /*
 152  * Refill based on our delayed refs usage.
 153  *
 154  * @fs_info: the filesystem
 155  * @flush:   control how we can flush for this reservation.
 156  *
 157  * This will refill the delayed block_rsv up to 1 items size worth of space and
 158  * will return -ENOSPC if we can't make the reservation.
 159  */
 160 int btrfs_delayed_refs_rsv_refill(struct btrfs_fs_info *fs_info,
 161                                   enum btrfs_reserve_flush_enum flush)
 162 {
 163         struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
 164         u64 limit = btrfs_calc_delayed_ref_bytes(fs_info, 1);
 165         u64 num_bytes = 0;
 166         int ret = -ENOSPC;
 167
 168         spin_lock(&block_rsv->lock);
 169         if (block_rsv->reserved < block_rsv->size) {
 170                 num_bytes = block_rsv->size - block_rsv->reserved;
 171                 num_bytes = min(num_bytes, limit);
 172         }
 173         spin_unlock(&block_rsv->lock);
 174
 175         if (!num_bytes)
 176                 return 0;
 177
 178         ret = btrfs_reserve_metadata_bytes(fs_info, block_rsv, num_bytes, flush);
 179         if (ret)
 180                 return ret;
 181         btrfs_block_rsv_add_bytes(block_rsv, num_bytes, false);
 182         trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
 183                                       0, num_bytes, 1);
 184         return 0;
 185 }
 186
 187 /*
 188  * compare two delayed tree backrefs with same bytenr and type
 189  */
 190 static int comp_tree_refs(struct btrfs_delayed_tree_ref *ref1,
 191                           struct btrfs_delayed_tree_ref *ref2)
 192 {
 193         if (ref1->node.type == BTRFS_TREE_BLOCK_REF_KEY) {
 194                 if (ref1->root < ref2->root)
 195                         return -1;
 196                 if (ref1->root > ref2->root)
 197                         return 1;
 198         } else {
 199                 if (ref1->parent < ref2->parent)
 200                         return -1;
 201                 if (ref1->parent > ref2->parent)
 202                         return 1;
 203         }
 204         return 0;
 205 }
 206
 207 /*
 208  * compare two delayed data backrefs with same bytenr and type
 209  */
 210 static int comp_data_refs(struct btrfs_delayed_data_ref *ref1,
 211                           struct btrfs_delayed_data_ref *ref2)
 212 {
 213         if (ref1->node.type == BTRFS_EXTENT_DATA_REF_KEY) {
 214                 if (ref1->root < ref2->root)
 215                         return -1;
 216                 if (ref1->root > ref2->root)
 217                         return 1;
 218                 if (ref1->objectid < ref2->objectid)
 219                         return -1;
 220                 if (ref1->objectid > ref2->objectid)
 221                         return 1;
 222                 if (ref1->offset < ref2->offset)
 223                         return -1;
 224                 if (ref1->offset > ref2->offset)
 225                         return 1;
 226         } else {
 227                 if (ref1->parent < ref2->parent)
 228                         return -1;
 229                 if (ref1->parent > ref2->parent)
 230                         return 1;
 231         }
 232         return 0;
 233 }
 234
 235 static int comp_refs(struct btrfs_delayed_ref_node *ref1,
 236                      struct btrfs_delayed_ref_node *ref2,
 237                      bool check_seq)
 238 {
 239         int ret = 0;
 240
 241         if (ref1->type < ref2->type)
 242                 return -1;
 243         if (ref1->type > ref2->type)
 244                 return 1;
 245         if (ref1->type == BTRFS_TREE_BLOCK_REF_KEY ||
 246             ref1->type == BTRFS_SHARED_BLOCK_REF_KEY)
 247                 ret = comp_tree_refs(btrfs_delayed_node_to_tree_ref(ref1),
 248                                      btrfs_delayed_node_to_tree_ref(ref2));
 249         else
 250                 ret = comp_data_refs(btrfs_delayed_node_to_data_ref(ref1),
 251                                      btrfs_delayed_node_to_data_ref(ref2));
 252         if (ret)
 253                 return ret;
 254         if (check_seq) {
 255                 if (ref1->seq < ref2->seq)
 256                         return -1;
 257                 if (ref1->seq > ref2->seq)
 258                         return 1;
 259         }
 260         return 0;
 261 }
 262
 263 /* insert a new ref to head ref rbtree */
 264 static struct btrfs_delayed_ref_head *htree_insert(struct rb_root_cached *root,
 265                                                    struct rb_node *node)
 266 {
 267         struct rb_node **p = &root->rb_root.rb_node;
 268         struct rb_node *parent_node = NULL;
 269         struct btrfs_delayed_ref_head *entry;
 270         struct btrfs_delayed_ref_head *ins;
 271         u64 bytenr;
 272         bool leftmost = true;
 273
 274         ins = rb_entry(node, struct btrfs_delayed_ref_head, href_node);
 275         bytenr = ins->bytenr;
 276         while (*p) {
 277                 parent_node = *p;
 278                 entry = rb_entry(parent_node, struct btrfs_delayed_ref_head,
 279                                  href_node);
 280
 281                 if (bytenr < entry->bytenr) {
 282                         p = &(*p)->rb_left;
 283                 } else if (bytenr > entry->bytenr) {
 284                         p = &(*p)->rb_right;
 285                         leftmost = false;
 286                 } else {
 287                         return entry;
 288                 }
 289         }
 290
 291         rb_link_node(node, parent_node, p);
 292         rb_insert_color_cached(node, root, leftmost);
 293         return NULL;
 294 }
 295
 296 static struct btrfs_delayed_ref_node* tree_insert(struct rb_root_cached *root,
 297                 struct btrfs_delayed_ref_node *ins)
 298 {
 299         struct rb_node **p = &root->rb_root.rb_node;
 300         struct rb_node *node = &ins->ref_node;
 301         struct rb_node *parent_node = NULL;
 302         struct btrfs_delayed_ref_node *entry;
 303         bool leftmost = true;
 304
 305         while (*p) {
 306                 int comp;
 307
 308                 parent_node = *p;
 309                 entry = rb_entry(parent_node, struct btrfs_delayed_ref_node,
 310                                  ref_node);
 311                 comp = comp_refs(ins, entry, true);
 312                 if (comp < 0) {
 313                         p = &(*p)->rb_left;
 314                 } else if (comp > 0) {
 315                         p = &(*p)->rb_right;
 316                         leftmost = false;
 317                 } else {
 318                         return entry;
 319                 }
 320         }
 321
 322         rb_link_node(node, parent_node, p);
 323         rb_insert_color_cached(node, root, leftmost);
 324         return NULL;
 325 }
 326
 327 static struct btrfs_delayed_ref_head *find_first_ref_head(
 328                 struct btrfs_delayed_ref_root *dr)
 329 {
 330         struct rb_node *n;
 331         struct btrfs_delayed_ref_head *entry;
 332
 333         n = rb_first_cached(&dr->href_root);
 334         if (!n)
 335                 return NULL;
 336
 337         entry = rb_entry(n, struct btrfs_delayed_ref_head, href_node);
 338
 339         return entry;
 340 }
 341
 342 /*
 343  * Find a head entry based on bytenr. This returns the delayed ref head if it
 344  * was able to find one, or NULL if nothing was in that spot.  If return_bigger
 345  * is given, the next bigger entry is returned if no exact match is found.
 346  */
 347 static struct btrfs_delayed_ref_head *find_ref_head(
 348                 struct btrfs_delayed_ref_root *dr, u64 bytenr,
 349                 bool return_bigger)
 350 {
 351         struct rb_root *root = &dr->href_root.rb_root;
 352         struct rb_node *n;
 353         struct btrfs_delayed_ref_head *entry;
 354
 355         n = root->rb_node;
 356         entry = NULL;
 357         while (n) {
 358                 entry = rb_entry(n, struct btrfs_delayed_ref_head, href_node);
 359
 360                 if (bytenr < entry->bytenr)
 361                         n = n->rb_left;
 362                 else if (bytenr > entry->bytenr)
 363                         n = n->rb_right;
 364                 else
 365                         return entry;
 366         }
 367         if (entry && return_bigger) {
 368                 if (bytenr > entry->bytenr) {
 369                         n = rb_next(&entry->href_node);
 370                         if (!n)
 371                                 return NULL;
 372                         entry = rb_entry(n, struct btrfs_delayed_ref_head,
 373                                          href_node);
 374                 }
 375                 return entry;
 376         }
 377         return NULL;
 378 }
 379
 380 int btrfs_delayed_ref_lock(struct btrfs_delayed_ref_root *delayed_refs,
 381                            struct btrfs_delayed_ref_head *head)
 382 {
 383         lockdep_assert_held(&delayed_refs->lock);
 384         if (mutex_trylock(&head->mutex))
 385                 return 0;
 386
 387         refcount_inc(&head->refs);
 388         spin_unlock(&delayed_refs->lock);
 389
 390         mutex_lock(&head->mutex);
 391         spin_lock(&delayed_refs->lock);
 392         if (RB_EMPTY_NODE(&head->href_node)) {
 393                 mutex_unlock(&head->mutex);
 394                 btrfs_put_delayed_ref_head(head);
 395                 return -EAGAIN;
 396         }
 397         btrfs_put_delayed_ref_head(head);
 398         return 0;
 399 }
 400
 401 static inline void drop_delayed_ref(struct btrfs_delayed_ref_root *delayed_refs,
 402                                     struct btrfs_delayed_ref_head *head,
 403                                     struct btrfs_delayed_ref_node *ref)
 404 {
 405         lockdep_assert_held(&head->lock);
 406         rb_erase_cached(&ref->ref_node, &head->ref_tree);
 407         RB_CLEAR_NODE(&ref->ref_node);
 408         if (!list_empty(&ref->add_list))
 409                 list_del(&ref->add_list);
 410         ref->in_tree = 0;
 411         btrfs_put_delayed_ref(ref);
 412         atomic_dec(&delayed_refs->num_entries);
 413 }
 414
 415 static bool merge_ref(struct btrfs_delayed_ref_root *delayed_refs,
 416                       struct btrfs_delayed_ref_head *head,
 417                       struct btrfs_delayed_ref_node *ref,
 418                       u64 seq)
 419 {
 420         struct btrfs_delayed_ref_node *next;
 421         struct rb_node *node = rb_next(&ref->ref_node);
 422         bool done = false;
 423
 424         while (!done && node) {
 425                 int mod;
 426
 427                 next = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
 428                 node = rb_next(node);
 429                 if (seq && next->seq >= seq)
 430                         break;
 431                 if (comp_refs(ref, next, false))
 432                         break;
 433
 434                 if (ref->action == next->action) {
 435                         mod = next->ref_mod;
 436                 } else {
 437                         if (ref->ref_mod < next->ref_mod) {
 438                                 swap(ref, next);
 439                                 done = true;
 440                         }
 441                         mod = -next->ref_mod;
 442                 }
 443
 444                 drop_delayed_ref(delayed_refs, head, next);
 445                 ref->ref_mod += mod;
 446                 if (ref->ref_mod == 0) {
 447                         drop_delayed_ref(delayed_refs, head, ref);
 448                         done = true;
 449                 } else {
 450                         /*
 451                          * Can't have multiples of the same ref on a tree block.
 452                          */
 453                         WARN_ON(ref->type == BTRFS_TREE_BLOCK_REF_KEY ||
 454                                 ref->type == BTRFS_SHARED_BLOCK_REF_KEY);
 455                 }
 456         }
 457
 458         return done;
 459 }
 460
 461 void btrfs_merge_delayed_refs(struct btrfs_fs_info *fs_info,
 462                               struct btrfs_delayed_ref_root *delayed_refs,
 463                               struct btrfs_delayed_ref_head *head)
 464 {
 465         struct btrfs_delayed_ref_node *ref;
 466         struct rb_node *node;
 467         u64 seq = 0;
 468
 469         lockdep_assert_held(&head->lock);
 470
 471         if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
 472                 return;
 473
 474         /* We don't have too many refs to merge for data. */
 475         if (head->is_data)
 476                 return;
 477
 478         seq = btrfs_tree_mod_log_lowest_seq(fs_info);
 479 again:
 480         for (node = rb_first_cached(&head->ref_tree); node;
 481              node = rb_next(node)) {
 482                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
 483                 if (seq && ref->seq >= seq)
 484                         continue;
 485                 if (merge_ref(delayed_refs, head, ref, seq))
 486                         goto again;
 487         }
 488 }
 489
 490 int btrfs_check_delayed_seq(struct btrfs_fs_info *fs_info, u64 seq)
 491 {
 492         int ret = 0;
 493         u64 min_seq = btrfs_tree_mod_log_lowest_seq(fs_info);
 494
 495         if (min_seq != 0 && seq >= min_seq) {
 496                 btrfs_debug(fs_info,
 497                             "holding back delayed_ref %llu, lowest is %llu",
 498                             seq, min_seq);
 499                 ret = 1;
 500         }
 501
 502         return ret;
 503 }
 504
 505 struct btrfs_delayed_ref_head *btrfs_select_ref_head(
 506                 struct btrfs_delayed_ref_root *delayed_refs)
 507 {
 508         struct btrfs_delayed_ref_head *head;
 509
 510 again:
 511         head = find_ref_head(delayed_refs, delayed_refs->run_delayed_start,
 512                              true);
 513         if (!head && delayed_refs->run_delayed_start != 0) {
 514                 delayed_refs->run_delayed_start = 0;
 515                 head = find_first_ref_head(delayed_refs);
 516         }
 517         if (!head)
 518                 return NULL;
 519
 520         while (head->processing) {
 521                 struct rb_node *node;
 522
 523                 node = rb_next(&head->href_node);
 524                 if (!node) {
 525                         if (delayed_refs->run_delayed_start == 0)
 526                                 return NULL;
 527                         delayed_refs->run_delayed_start = 0;
 528                         goto again;
 529                 }
 530                 head = rb_entry(node, struct btrfs_delayed_ref_head,
 531                                 href_node);
 532         }
 533
 534         head->processing = 1;
 535         WARN_ON(delayed_refs->num_heads_ready == 0);
 536         delayed_refs->num_heads_ready--;
 537         delayed_refs->run_delayed_start = head->bytenr +
 538                 head->num_bytes;
 539         return head;
 540 }
 541
 542 void btrfs_delete_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
 543                            struct btrfs_delayed_ref_head *head)
 544 {
 545         lockdep_assert_held(&delayed_refs->lock);
 546         lockdep_assert_held(&head->lock);
 547
 548         rb_erase_cached(&head->href_node, &delayed_refs->href_root);
 549         RB_CLEAR_NODE(&head->href_node);
 550         atomic_dec(&delayed_refs->num_entries);
 551         delayed_refs->num_heads--;
 552         if (head->processing == 0)
 553                 delayed_refs->num_heads_ready--;
 554 }
 555
 556 /*
 557  * Helper to insert the ref_node to the tail or merge with tail.
 558  *
 559  * Return 0 for insert.
 560  * Return >0 for merge.
 561  */
 562 static int insert_delayed_ref(struct btrfs_delayed_ref_root *root,
 563                               struct btrfs_delayed_ref_head *href,
 564                               struct btrfs_delayed_ref_node *ref)
 565 {
 566         struct btrfs_delayed_ref_node *exist;
 567         int mod;
 568         int ret = 0;
 569
 570         spin_lock(&href->lock);
 571         exist = tree_insert(&href->ref_tree, ref);
 572         if (!exist)
 573                 goto inserted;
 574
 575         /* Now we are sure we can merge */
 576         ret = 1;
 577         if (exist->action == ref->action) {
 578                 mod = ref->ref_mod;
 579         } else {
 580                 /* Need to change action */
 581                 if (exist->ref_mod < ref->ref_mod) {
 582                         exist->action = ref->action;
 583                         mod = -exist->ref_mod;
 584                         exist->ref_mod = ref->ref_mod;
 585                         if (ref->action == BTRFS_ADD_DELAYED_REF)
 586                                 list_add_tail(&exist->add_list,
 587                                               &href->ref_add_list);
 588                         else if (ref->action == BTRFS_DROP_DELAYED_REF) {
 589                                 ASSERT(!list_empty(&exist->add_list));
 590                                 list_del(&exist->add_list);
 591                         } else {
 592                                 ASSERT(0);
 593                         }
 594                 } else
 595                         mod = -ref->ref_mod;
 596         }
 597         exist->ref_mod += mod;
 598
 599         /* remove existing tail if its ref_mod is zero */
 600         if (exist->ref_mod == 0)
 601                 drop_delayed_ref(root, href, exist);
 602         spin_unlock(&href->lock);
 603         return ret;
 604 inserted:
 605         if (ref->action == BTRFS_ADD_DELAYED_REF)
 606                 list_add_tail(&ref->add_list, &href->ref_add_list);
 607         atomic_inc(&root->num_entries);
 608         spin_unlock(&href->lock);
 609         return ret;
 610 }
 611
 612 /*
 613  * helper function to update the accounting in the head ref
 614  * existing and update must have the same bytenr
 615  */
 616 static noinline void update_existing_head_ref(struct btrfs_trans_handle *trans,
 617                          struct btrfs_delayed_ref_head *existing,
 618                          struct btrfs_delayed_ref_head *update)
 619 {
 620         struct btrfs_delayed_ref_root *delayed_refs =
 621                 &trans->transaction->delayed_refs;
 622         struct btrfs_fs_info *fs_info = trans->fs_info;
 623         int old_ref_mod;
 624
 625         BUG_ON(existing->is_data != update->is_data);
 626
 627         spin_lock(&existing->lock);
 628         if (update->must_insert_reserved) {
 629                 /* if the extent was freed and then
 630                  * reallocated before the delayed ref
 631                  * entries were processed, we can end up
 632                  * with an existing head ref without
 633                  * the must_insert_reserved flag set.
 634                  * Set it again here
 635                  */
 636                 existing->must_insert_reserved = update->must_insert_reserved;
 637
 638                 /*
 639                  * update the num_bytes so we make sure the accounting
 640                  * is done correctly
 641                  */
 642                 existing->num_bytes = update->num_bytes;
 643
 644         }
 645
 646         if (update->extent_op) {
 647                 if (!existing->extent_op) {
 648                         existing->extent_op = update->extent_op;
 649                 } else {
 650                         if (update->extent_op->update_key) {
 651                                 memcpy(&existing->extent_op->key,
 652                                        &update->extent_op->key,
 653                                        sizeof(update->extent_op->key));
 654                                 existing->extent_op->update_key = true;
 655                         }
 656                         if (update->extent_op->update_flags) {
 657                                 existing->extent_op->flags_to_set |=
 658                                         update->extent_op->flags_to_set;
 659                                 existing->extent_op->update_flags = true;
 660                         }
 661                         btrfs_free_delayed_extent_op(update->extent_op);
 662                 }
 663         }
 664         /*
 665          * update the reference mod on the head to reflect this new operation,
 666          * only need the lock for this case cause we could be processing it
 667          * currently, for refs we just added we know we're a-ok.
 668          */
 669         old_ref_mod = existing->total_ref_mod;
 670         existing->ref_mod += update->ref_mod;
 671         existing->total_ref_mod += update->ref_mod;
 672
 673         /*
 674          * If we are going to from a positive ref mod to a negative or vice
 675          * versa we need to make sure to adjust pending_csums accordingly.
 676          */
 677         if (existing->is_data) {
 678                 u64 csum_leaves =
 679                         btrfs_csum_bytes_to_leaves(fs_info,
 680                                                    existing->num_bytes);
 681
 682                 if (existing->total_ref_mod >= 0 && old_ref_mod < 0) {
 683                         delayed_refs->pending_csums -= existing->num_bytes;
 684                         btrfs_delayed_refs_rsv_release(fs_info, csum_leaves);
 685                 }
 686                 if (existing->total_ref_mod < 0 && old_ref_mod >= 0) {
 687                         delayed_refs->pending_csums += existing->num_bytes;
 688                         trans->delayed_ref_updates += csum_leaves;
 689                 }
 690         }
 691
 692         spin_unlock(&existing->lock);
 693 }
 694
 695 static void init_delayed_ref_head(struct btrfs_delayed_ref_head *head_ref,
 696                                   struct btrfs_qgroup_extent_record *qrecord,
 697                                   u64 bytenr, u64 num_bytes, u64 ref_root,
 698                                   u64 reserved, int action, bool is_data,
 699                                   bool is_system)
 700 {
 701         int count_mod = 1;
 702         int must_insert_reserved = 0;
 703
 704         /* If reserved is provided, it must be a data extent. */
 705         BUG_ON(!is_data && reserved);
 706
 707         /*
 708          * The head node stores the sum of all the mods, so dropping a ref
 709          * should drop the sum in the head node by one.
 710          */
 711         if (action == BTRFS_UPDATE_DELAYED_HEAD)
 712                 count_mod = 0;
 713         else if (action == BTRFS_DROP_DELAYED_REF)
 714                 count_mod = -1;
 715
 716         /*
 717          * BTRFS_ADD_DELAYED_EXTENT means that we need to update the reserved
 718          * accounting when the extent is finally added, or if a later
 719          * modification deletes the delayed ref without ever inserting the
 720          * extent into the extent allocation tree.  ref->must_insert_reserved
 721          * is the flag used to record that accounting mods are required.
 722          *
 723          * Once we record must_insert_reserved, switch the action to
 724          * BTRFS_ADD_DELAYED_REF because other special casing is not required.
 725          */
 726         if (action == BTRFS_ADD_DELAYED_EXTENT)
 727                 must_insert_reserved = 1;
 728         else
 729                 must_insert_reserved = 0;
 730
 731         refcount_set(&head_ref->refs, 1);
 732         head_ref->bytenr = bytenr;
 733         head_ref->num_bytes = num_bytes;
 734         head_ref->ref_mod = count_mod;
 735         head_ref->must_insert_reserved = must_insert_reserved;
 736         head_ref->is_data = is_data;
 737         head_ref->is_system = is_system;
 738         head_ref->ref_tree = RB_ROOT_CACHED;
 739         INIT_LIST_HEAD(&head_ref->ref_add_list);
 740         RB_CLEAR_NODE(&head_ref->href_node);
 741         head_ref->processing = 0;
 742         head_ref->total_ref_mod = count_mod;
 743         spin_lock_init(&head_ref->lock);
 744         mutex_init(&head_ref->mutex);
 745
 746         if (qrecord) {
 747                 if (ref_root && reserved) {
 748                         qrecord->data_rsv = reserved;
 749                         qrecord->data_rsv_refroot = ref_root;
 750                 }
 751                 qrecord->bytenr = bytenr;
 752                 qrecord->num_bytes = num_bytes;
 753                 qrecord->old_roots = NULL;
 754         }
 755 }
 756
 757 /*
 758  * helper function to actually insert a head node into the rbtree.
 759  * this does all the dirty work in terms of maintaining the correct
 760  * overall modification count.
 761  */
 762 static noinline struct btrfs_delayed_ref_head *
 763 add_delayed_ref_head(struct btrfs_trans_handle *trans,
 764                      struct btrfs_delayed_ref_head *head_ref,
 765                      struct btrfs_qgroup_extent_record *qrecord,
 766                      int action, int *qrecord_inserted_ret)
 767 {
 768         struct btrfs_delayed_ref_head *existing;
 769         struct btrfs_delayed_ref_root *delayed_refs;
 770         int qrecord_inserted = 0;
 771
 772         delayed_refs = &trans->transaction->delayed_refs;
 773
 774         /* Record qgroup extent info if provided */
 775         if (qrecord) {
 776                 if (btrfs_qgroup_trace_extent_nolock(trans->fs_info,
 777                                         delayed_refs, qrecord))
 778                         kfree(qrecord);
 779                 else
 780                         qrecord_inserted = 1;
 781         }
 782
 783         trace_add_delayed_ref_head(trans->fs_info, head_ref, action);
 784
 785         existing = htree_insert(&delayed_refs->href_root,
 786                                 &head_ref->href_node);
 787         if (existing) {
 788                 update_existing_head_ref(trans, existing, head_ref);
 789                 /*
 790                  * we've updated the existing ref, free the newly
 791                  * allocated ref
 792                  */
 793                 kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
 794                 head_ref = existing;
 795         } else {
 796                 if (head_ref->is_data && head_ref->ref_mod < 0) {
 797                         delayed_refs->pending_csums += head_ref->num_bytes;
 798                         trans->delayed_ref_updates +=
 799                                 btrfs_csum_bytes_to_leaves(trans->fs_info,
 800                                                            head_ref->num_bytes);
 801                 }
 802                 delayed_refs->num_heads++;
 803                 delayed_refs->num_heads_ready++;
 804                 atomic_inc(&delayed_refs->num_entries);
 805                 trans->delayed_ref_updates++;
 806         }
 807         if (qrecord_inserted_ret)
 808                 *qrecord_inserted_ret = qrecord_inserted;
 809
 810         return head_ref;
 811 }
 812
 813 /*
 814  * init_delayed_ref_common - Initialize the structure which represents a
 815  *                           modification to a an extent.
 816  *
 817  * @fs_info:    Internal to the mounted filesystem mount structure.
 818  *
 819  * @ref:        The structure which is going to be initialized.
 820  *
 821  * @bytenr:     The logical address of the extent for which a modification is
 822  *              going to be recorded.
 823  *
 824  * @num_bytes:  Size of the extent whose modification is being recorded.
 825  *
 826  * @ref_root:   The id of the root where this modification has originated, this
 827  *              can be either one of the well-known metadata trees or the
 828  *              subvolume id which references this extent.
 829  *
 830  * @action:     Can be one of BTRFS_ADD_DELAYED_REF/BTRFS_DROP_DELAYED_REF or
 831  *              BTRFS_ADD_DELAYED_EXTENT
 832  *
 833  * @ref_type:   Holds the type of the extent which is being recorded, can be
 834  *              one of BTRFS_SHARED_BLOCK_REF_KEY/BTRFS_TREE_BLOCK_REF_KEY
 835  *              when recording a metadata extent or BTRFS_SHARED_DATA_REF_KEY/
 836  *              BTRFS_EXTENT_DATA_REF_KEY when recording data extent
 837  */
 838 static void init_delayed_ref_common(struct btrfs_fs_info *fs_info,
 839                                     struct btrfs_delayed_ref_node *ref,
 840                                     u64 bytenr, u64 num_bytes, u64 ref_root,
 841                                     int action, u8 ref_type)
 842 {
 843         u64 seq = 0;
 844
 845         if (action == BTRFS_ADD_DELAYED_EXTENT)
 846                 action = BTRFS_ADD_DELAYED_REF;
 847
 848         if (is_fstree(ref_root))
 849                 seq = atomic64_read(&fs_info->tree_mod_seq);
 850
 851         refcount_set(&ref->refs, 1);
 852         ref->bytenr = bytenr;
 853         ref->num_bytes = num_bytes;
 854         ref->ref_mod = 1;
 855         ref->action = action;
 856         ref->is_head = 0;
 857         ref->in_tree = 1;
 858         ref->seq = seq;
 859         ref->type = ref_type;
 860         RB_CLEAR_NODE(&ref->ref_node);
 861         INIT_LIST_HEAD(&ref->add_list);
 862 }
 863
 864 /*
 865  * add a delayed tree ref.  This does all of the accounting required
 866  * to make sure the delayed ref is eventually processed before this
 867  * transaction commits.
 868  */
 869 int btrfs_add_delayed_tree_ref(struct btrfs_trans_handle *trans,
 870                                struct btrfs_ref *generic_ref,
 871                                struct btrfs_delayed_extent_op *extent_op)
 872 {
 873         struct btrfs_fs_info *fs_info = trans->fs_info;
 874         struct btrfs_delayed_tree_ref *ref;
 875         struct btrfs_delayed_ref_head *head_ref;
 876         struct btrfs_delayed_ref_root *delayed_refs;
 877         struct btrfs_qgroup_extent_record *record = NULL;
 878         int qrecord_inserted;
 879         bool is_system;
 880         int action = generic_ref->action;
 881         int level = generic_ref->tree_ref.level;
 882         int ret;
 883         u64 bytenr = generic_ref->bytenr;
 884         u64 num_bytes = generic_ref->len;
 885         u64 parent = generic_ref->parent;
 886         u8 ref_type;
 887
 888         is_system = (generic_ref->tree_ref.owning_root == BTRFS_CHUNK_TREE_OBJECTID);
 889
 890         ASSERT(generic_ref->type == BTRFS_REF_METADATA && generic_ref->action);
 891         ref = kmem_cache_alloc(btrfs_delayed_tree_ref_cachep, GFP_NOFS);
 892         if (!ref)
 893                 return -ENOMEM;
 894
 895         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
 896         if (!head_ref) {
 897                 kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
 898                 return -ENOMEM;
 899         }
 900
 901         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) &&
 902             !generic_ref->skip_qgroup) {
 903                 record = kzalloc(sizeof(*record), GFP_NOFS);
 904                 if (!record) {
 905                         kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
 906                         kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
 907                         return -ENOMEM;
 908                 }
 909         }
 910
 911         if (parent)
 912                 ref_type = BTRFS_SHARED_BLOCK_REF_KEY;
 913         else
 914                 ref_type = BTRFS_TREE_BLOCK_REF_KEY;
 915
 916         init_delayed_ref_common(fs_info, &ref->node, bytenr, num_bytes,
 917                                 generic_ref->tree_ref.owning_root, action,
 918                                 ref_type);
 919         ref->root = generic_ref->tree_ref.owning_root;
 920         ref->parent = parent;
 921         ref->level = level;
 922
 923         init_delayed_ref_head(head_ref, record, bytenr, num_bytes,
 924                               generic_ref->tree_ref.owning_root, 0, action,
 925                               false, is_system);
 926         head_ref->extent_op = extent_op;
 927
 928         delayed_refs = &trans->transaction->delayed_refs;
 929         spin_lock(&delayed_refs->lock);
 930
 931         /*
 932          * insert both the head node and the new ref without dropping
 933          * the spin lock
 934          */
 935         head_ref = add_delayed_ref_head(trans, head_ref, record,
 936                                         action, &qrecord_inserted);
 937
 938         ret = insert_delayed_ref(delayed_refs, head_ref, &ref->node);
 939         spin_unlock(&delayed_refs->lock);
 940
 941         /*
 942          * Need to update the delayed_refs_rsv with any changes we may have
 943          * made.
 944          */
 945         btrfs_update_delayed_refs_rsv(trans);
 946
 947         trace_add_delayed_tree_ref(fs_info, &ref->node, ref,
 948                                    action == BTRFS_ADD_DELAYED_EXTENT ?
 949                                    BTRFS_ADD_DELAYED_REF : action);
 950         if (ret > 0)
 951                 kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
 952
 953         if (qrecord_inserted)
 954                 btrfs_qgroup_trace_extent_post(trans, record);
 955
 956         return 0;
 957 }
 958
 959 /*
 960  * add a delayed data ref. it's similar to btrfs_add_delayed_tree_ref.
 961  */
 962 int btrfs_add_delayed_data_ref(struct btrfs_trans_handle *trans,
 963                                struct btrfs_ref *generic_ref,
 964                                u64 reserved)
 965 {
 966         struct btrfs_fs_info *fs_info = trans->fs_info;
 967         struct btrfs_delayed_data_ref *ref;
 968         struct btrfs_delayed_ref_head *head_ref;
 969         struct btrfs_delayed_ref_root *delayed_refs;
 970         struct btrfs_qgroup_extent_record *record = NULL;
 971         int qrecord_inserted;
 972         int action = generic_ref->action;
 973         int ret;
 974         u64 bytenr = generic_ref->bytenr;
 975         u64 num_bytes = generic_ref->len;
 976         u64 parent = generic_ref->parent;
 977         u64 ref_root = generic_ref->data_ref.owning_root;
 978         u64 owner = generic_ref->data_ref.ino;
 979         u64 offset = generic_ref->data_ref.offset;
 980         u8 ref_type;
 981
 982         ASSERT(generic_ref->type == BTRFS_REF_DATA && action);
 983         ref = kmem_cache_alloc(btrfs_delayed_data_ref_cachep, GFP_NOFS);
 984         if (!ref)
 985                 return -ENOMEM;
 986
 987         if (parent)
 988                 ref_type = BTRFS_SHARED_DATA_REF_KEY;
 989         else
 990                 ref_type = BTRFS_EXTENT_DATA_REF_KEY;
 991         init_delayed_ref_common(fs_info, &ref->node, bytenr, num_bytes,
 992                                 ref_root, action, ref_type);
 993         ref->root = ref_root;
 994         ref->parent = parent;
 995         ref->objectid = owner;
 996         ref->offset = offset;
 997
 998
 999         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
1000         if (!head_ref) {
1001                 kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
1002                 return -ENOMEM;
1003         }
1004
1005         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) &&
1006             !generic_ref->skip_qgroup) {
1007                 record = kzalloc(sizeof(*record), GFP_NOFS);
1008                 if (!record) {
1009                         kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
1010                         kmem_cache_free(btrfs_delayed_ref_head_cachep,
1011                                         head_ref);
1012                         return -ENOMEM;
1013                 }
1014         }
1015
1016         init_delayed_ref_head(head_ref, record, bytenr, num_bytes, ref_root,
1017                               reserved, action, true, false);
1018         head_ref->extent_op = NULL;
1019
1020         delayed_refs = &trans->transaction->delayed_refs;
1021         spin_lock(&delayed_refs->lock);
1022
1023         /*
1024          * insert both the head node and the new ref without dropping
1025          * the spin lock
1026          */
1027         head_ref = add_delayed_ref_head(trans, head_ref, record,
1028                                         action, &qrecord_inserted);
1029
1030         ret = insert_delayed_ref(delayed_refs, head_ref, &ref->node);
1031         spin_unlock(&delayed_refs->lock);
1032
1033         /*
1034          * Need to update the delayed_refs_rsv with any changes we may have
1035          * made.
1036          */
1037         btrfs_update_delayed_refs_rsv(trans);
1038
1039         trace_add_delayed_data_ref(trans->fs_info, &ref->node, ref,
1040                                    action == BTRFS_ADD_DELAYED_EXTENT ?
1041                                    BTRFS_ADD_DELAYED_REF : action);
1042         if (ret > 0)
1043                 kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
1044
1045
1046         if (qrecord_inserted)
1047                 return btrfs_qgroup_trace_extent_post(trans, record);
1048         return 0;
1049 }
1050
1051 int btrfs_add_delayed_extent_op(struct btrfs_trans_handle *trans,
1052                                 u64 bytenr, u64 num_bytes,
1053                                 struct btrfs_delayed_extent_op *extent_op)
1054 {
1055         struct btrfs_delayed_ref_head *head_ref;
1056         struct btrfs_delayed_ref_root *delayed_refs;
1057
1058         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
1059         if (!head_ref)
1060                 return -ENOMEM;
1061
1062         init_delayed_ref_head(head_ref, NULL, bytenr, num_bytes, 0, 0,
1063                               BTRFS_UPDATE_DELAYED_HEAD, false, false);
1064         head_ref->extent_op = extent_op;
1065
1066         delayed_refs = &trans->transaction->delayed_refs;
1067         spin_lock(&delayed_refs->lock);
1068
1069         add_delayed_ref_head(trans, head_ref, NULL, BTRFS_UPDATE_DELAYED_HEAD,
1070                              NULL);
1071
1072         spin_unlock(&delayed_refs->lock);
1073
1074         /*
1075          * Need to update the delayed_refs_rsv with any changes we may have
1076          * made.
1077          */
1078         btrfs_update_delayed_refs_rsv(trans);
1079         return 0;
1080 }
1081
1082 /*
1083  * This does a simple search for the head node for a given extent.  Returns the
1084  * head node if found, or NULL if not.
1085  */
1086 struct btrfs_delayed_ref_head *
1087 btrfs_find_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs, u64 bytenr)
1088 {
1089         lockdep_assert_held(&delayed_refs->lock);
1090
1091         return find_ref_head(delayed_refs, bytenr, false);
1092 }
1093
1094 void __cold btrfs_delayed_ref_exit(void)
1095 {
1096         kmem_cache_destroy(btrfs_delayed_ref_head_cachep);
1097         kmem_cache_destroy(btrfs_delayed_tree_ref_cachep);
1098         kmem_cache_destroy(btrfs_delayed_data_ref_cachep);
1099         kmem_cache_destroy(btrfs_delayed_extent_op_cachep);
1100 }
1101
1102 int __init btrfs_delayed_ref_init(void)
1103 {
1104         btrfs_delayed_ref_head_cachep = kmem_cache_create(
1105                                 "btrfs_delayed_ref_head",
1106                                 sizeof(struct btrfs_delayed_ref_head), 0,
1107                                 SLAB_MEM_SPREAD, NULL);
1108         if (!btrfs_delayed_ref_head_cachep)
1109                 goto fail;
1110
1111         btrfs_delayed_tree_ref_cachep = kmem_cache_create(
1112                                 "btrfs_delayed_tree_ref",
1113                                 sizeof(struct btrfs_delayed_tree_ref), 0,
1114                                 SLAB_MEM_SPREAD, NULL);
1115         if (!btrfs_delayed_tree_ref_cachep)
1116                 goto fail;
1117
1118         btrfs_delayed_data_ref_cachep = kmem_cache_create(
1119                                 "btrfs_delayed_data_ref",
1120                                 sizeof(struct btrfs_delayed_data_ref), 0,
1121                                 SLAB_MEM_SPREAD, NULL);
1122         if (!btrfs_delayed_data_ref_cachep)
1123                 goto fail;
1124
1125         btrfs_delayed_extent_op_cachep = kmem_cache_create(
1126                                 "btrfs_delayed_extent_op",
1127                                 sizeof(struct btrfs_delayed_extent_op), 0,
1128                                 SLAB_MEM_SPREAD, NULL);
1129         if (!btrfs_delayed_extent_op_cachep)
1130                 goto fail;
1131
1132         return 0;
1133 fail:
1134         btrfs_delayed_ref_exit();
1135         return -ENOMEM;
1136 }