btrfs/delayed-ref.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2009 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/slab.h>
   8 #include <linux/sort.h>
   9 #include "ctree.h"
  10 #include "delayed-ref.h"
  11 #include "transaction.h"
  12 #include "qgroup.h"
  13 #include "space-info.h"
  14 #include "tree-mod-log.h"
  15
  16 struct kmem_cache *btrfs_delayed_ref_head_cachep;
  17 struct kmem_cache *btrfs_delayed_tree_ref_cachep;
  18 struct kmem_cache *btrfs_delayed_data_ref_cachep;
  19 struct kmem_cache *btrfs_delayed_extent_op_cachep;
  20 /*
  21  * delayed back reference update tracking.  For subvolume trees
  22  * we queue up extent allocations and backref maintenance for
  23  * delayed processing.   This avoids deep call chains where we
  24  * add extents in the middle of btrfs_search_slot, and it allows
  25  * us to buffer up frequently modified backrefs in an rb tree instead
  26  * of hammering updates on the extent allocation tree.
  27  */
  28
  29 bool btrfs_check_space_for_delayed_refs(struct btrfs_fs_info *fs_info)
  30 {
  31         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
  32         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
  33         bool ret = false;
  34         u64 reserved;
  35
  36         spin_lock(&global_rsv->lock);
  37         reserved = global_rsv->reserved;
  38         spin_unlock(&global_rsv->lock);
  39
  40         /*
  41          * Since the global reserve is just kind of magic we don't really want
  42          * to rely on it to save our bacon, so if our size is more than the
  43          * delayed_refs_rsv and the global rsv then it's time to think about
  44          * bailing.
  45          */
  46         spin_lock(&delayed_refs_rsv->lock);
  47         reserved += delayed_refs_rsv->reserved;
  48         if (delayed_refs_rsv->size >= reserved)
  49                 ret = true;
  50         spin_unlock(&delayed_refs_rsv->lock);
  51         return ret;
  52 }
  53
  54 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans)
  55 {
  56         u64 num_entries =
  57                 atomic_read(&trans->transaction->delayed_refs.num_entries);
  58         u64 avg_runtime;
  59         u64 val;
  60
  61         smp_mb();
  62         avg_runtime = trans->fs_info->avg_delayed_ref_runtime;
  63         val = num_entries * avg_runtime;
  64         if (val >= NSEC_PER_SEC)
  65                 return 1;
  66         if (val >= NSEC_PER_SEC / 2)
  67                 return 2;
  68
  69         return btrfs_check_space_for_delayed_refs(trans->fs_info);
  70 }
  71
  72 /**
  73  * Release a ref head's reservation
  74  *
  75  * @fs_info:  the filesystem
  76  * @nr:       number of items to drop
  77  *
  78  * This drops the delayed ref head's count from the delayed refs rsv and frees
  79  * any excess reservation we had.
  80  */
  81 void btrfs_delayed_refs_rsv_release(struct btrfs_fs_info *fs_info, int nr)
  82 {
  83         struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
  84         u64 num_bytes = btrfs_calc_insert_metadata_size(fs_info, nr);
  85         u64 released = 0;
  86
  87         /*
  88          * We have to check the mount option here because we could be enabling
  89          * the free space tree for the first time and don't have the compat_ro
  90          * option set yet.
  91          *
  92          * We need extra reservations if we have the free space tree because
  93          * we'll have to modify that tree as well.
  94          */
  95         if (btrfs_test_opt(fs_info, FREE_SPACE_TREE))
  96                 num_bytes *= 2;
  97
  98         released = btrfs_block_rsv_release(fs_info, block_rsv, num_bytes, NULL);
  99         if (released)
 100                 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
 101                                               0, released, 0);
 102 }
 103
 104 /*
 105  * btrfs_update_delayed_refs_rsv - adjust the size of the delayed refs rsv
 106  * @trans - the trans that may have generated delayed refs
 107  *
 108  * This is to be called anytime we may have adjusted trans->delayed_ref_updates,
 109  * it'll calculate the additional size and add it to the delayed_refs_rsv.
 110  */
 111 void btrfs_update_delayed_refs_rsv(struct btrfs_trans_handle *trans)
 112 {
 113         struct btrfs_fs_info *fs_info = trans->fs_info;
 114         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_refs_rsv;
 115         u64 num_bytes;
 116
 117         if (!trans->delayed_ref_updates)
 118                 return;
 119
 120         num_bytes = btrfs_calc_insert_metadata_size(fs_info,
 121                                                     trans->delayed_ref_updates);
 122         /*
 123          * We have to check the mount option here because we could be enabling
 124          * the free space tree for the first time and don't have the compat_ro
 125          * option set yet.
 126          *
 127          * We need extra reservations if we have the free space tree because
 128          * we'll have to modify that tree as well.
 129          */
 130         if (btrfs_test_opt(fs_info, FREE_SPACE_TREE))
 131                 num_bytes *= 2;
 132
 133         spin_lock(&delayed_rsv->lock);
 134         delayed_rsv->size += num_bytes;
 135         delayed_rsv->full = false;
 136         spin_unlock(&delayed_rsv->lock);
 137         trans->delayed_ref_updates = 0;
 138 }
 139
 140 /**
 141  * Transfer bytes to our delayed refs rsv
 142  *
 143  * @fs_info:   the filesystem
 144  * @num_bytes: number of bytes to transfer
 145  *
 146  * This transfers up to the num_bytes amount, previously reserved, to the
 147  * delayed_refs_rsv.  Any extra bytes are returned to the space info.
 148  */
 149 void btrfs_migrate_to_delayed_refs_rsv(struct btrfs_fs_info *fs_info,
 150                                        u64 num_bytes)
 151 {
 152         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
 153         u64 to_free = 0;
 154
 155         spin_lock(&delayed_refs_rsv->lock);
 156         if (delayed_refs_rsv->size > delayed_refs_rsv->reserved) {
 157                 u64 delta = delayed_refs_rsv->size -
 158                         delayed_refs_rsv->reserved;
 159                 if (num_bytes > delta) {
 160                         to_free = num_bytes - delta;
 161                         num_bytes = delta;
 162                 }
 163         } else {
 164                 to_free = num_bytes;
 165                 num_bytes = 0;
 166         }
 167
 168         if (num_bytes)
 169                 delayed_refs_rsv->reserved += num_bytes;
 170         if (delayed_refs_rsv->reserved >= delayed_refs_rsv->size)
 171                 delayed_refs_rsv->full = true;
 172         spin_unlock(&delayed_refs_rsv->lock);
 173
 174         if (num_bytes)
 175                 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
 176                                               0, num_bytes, 1);
 177         if (to_free)
 178                 btrfs_space_info_free_bytes_may_use(fs_info,
 179                                 delayed_refs_rsv->space_info, to_free);
 180 }
 181
 182 /**
 183  * Refill based on our delayed refs usage
 184  *
 185  * @fs_info: the filesystem
 186  * @flush:   control how we can flush for this reservation.
 187  *
 188  * This will refill the delayed block_rsv up to 1 items size worth of space and
 189  * will return -ENOSPC if we can't make the reservation.
 190  */
 191 int btrfs_delayed_refs_rsv_refill(struct btrfs_fs_info *fs_info,
 192                                   enum btrfs_reserve_flush_enum flush)
 193 {
 194         struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
 195         u64 limit = btrfs_calc_insert_metadata_size(fs_info, 1);
 196         u64 num_bytes = 0;
 197         int ret = -ENOSPC;
 198
 199         spin_lock(&block_rsv->lock);
 200         if (block_rsv->reserved < block_rsv->size) {
 201                 num_bytes = block_rsv->size - block_rsv->reserved;
 202                 num_bytes = min(num_bytes, limit);
 203         }
 204         spin_unlock(&block_rsv->lock);
 205
 206         if (!num_bytes)
 207                 return 0;
 208
 209         ret = btrfs_reserve_metadata_bytes(fs_info, block_rsv, num_bytes, flush);
 210         if (ret)
 211                 return ret;
 212         btrfs_block_rsv_add_bytes(block_rsv, num_bytes, 0);
 213         trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
 214                                       0, num_bytes, 1);
 215         return 0;
 216 }
 217
 218 /*
 219  * compare two delayed tree backrefs with same bytenr and type
 220  */
 221 static int comp_tree_refs(struct btrfs_delayed_tree_ref *ref1,
 222                           struct btrfs_delayed_tree_ref *ref2)
 223 {
 224         if (ref1->node.type == BTRFS_TREE_BLOCK_REF_KEY) {
 225                 if (ref1->root < ref2->root)
 226                         return -1;
 227                 if (ref1->root > ref2->root)
 228                         return 1;
 229         } else {
 230                 if (ref1->parent < ref2->parent)
 231                         return -1;
 232                 if (ref1->parent > ref2->parent)
 233                         return 1;
 234         }
 235         return 0;
 236 }
 237
 238 /*
 239  * compare two delayed data backrefs with same bytenr and type
 240  */
 241 static int comp_data_refs(struct btrfs_delayed_data_ref *ref1,
 242                           struct btrfs_delayed_data_ref *ref2)
 243 {
 244         if (ref1->node.type == BTRFS_EXTENT_DATA_REF_KEY) {
 245                 if (ref1->root < ref2->root)
 246                         return -1;
 247                 if (ref1->root > ref2->root)
 248                         return 1;
 249                 if (ref1->objectid < ref2->objectid)
 250                         return -1;
 251                 if (ref1->objectid > ref2->objectid)
 252                         return 1;
 253                 if (ref1->offset < ref2->offset)
 254                         return -1;
 255                 if (ref1->offset > ref2->offset)
 256                         return 1;
 257         } else {
 258                 if (ref1->parent < ref2->parent)
 259                         return -1;
 260                 if (ref1->parent > ref2->parent)
 261                         return 1;
 262         }
 263         return 0;
 264 }
 265
 266 static int comp_refs(struct btrfs_delayed_ref_node *ref1,
 267                      struct btrfs_delayed_ref_node *ref2,
 268                      bool check_seq)
 269 {
 270         int ret = 0;
 271
 272         if (ref1->type < ref2->type)
 273                 return -1;
 274         if (ref1->type > ref2->type)
 275                 return 1;
 276         if (ref1->type == BTRFS_TREE_BLOCK_REF_KEY ||
 277             ref1->type == BTRFS_SHARED_BLOCK_REF_KEY)
 278                 ret = comp_tree_refs(btrfs_delayed_node_to_tree_ref(ref1),
 279                                      btrfs_delayed_node_to_tree_ref(ref2));
 280         else
 281                 ret = comp_data_refs(btrfs_delayed_node_to_data_ref(ref1),
 282                                      btrfs_delayed_node_to_data_ref(ref2));
 283         if (ret)
 284                 return ret;
 285         if (check_seq) {
 286                 if (ref1->seq < ref2->seq)
 287                         return -1;
 288                 if (ref1->seq > ref2->seq)
 289                         return 1;
 290         }
 291         return 0;
 292 }
 293
 294 /* insert a new ref to head ref rbtree */
 295 static struct btrfs_delayed_ref_head *htree_insert(struct rb_root_cached *root,
 296                                                    struct rb_node *node)
 297 {
 298         struct rb_node **p = &root->rb_root.rb_node;
 299         struct rb_node *parent_node = NULL;
 300         struct btrfs_delayed_ref_head *entry;
 301         struct btrfs_delayed_ref_head *ins;
 302         u64 bytenr;
 303         bool leftmost = true;
 304
 305         ins = rb_entry(node, struct btrfs_delayed_ref_head, href_node);
 306         bytenr = ins->bytenr;
 307         while (*p) {
 308                 parent_node = *p;
 309                 entry = rb_entry(parent_node, struct btrfs_delayed_ref_head,
 310                                  href_node);
 311
 312                 if (bytenr < entry->bytenr) {
 313                         p = &(*p)->rb_left;
 314                 } else if (bytenr > entry->bytenr) {
 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_node* tree_insert(struct rb_root_cached *root,
 328                 struct btrfs_delayed_ref_node *ins)
 329 {
 330         struct rb_node **p = &root->rb_root.rb_node;
 331         struct rb_node *node = &ins->ref_node;
 332         struct rb_node *parent_node = NULL;
 333         struct btrfs_delayed_ref_node *entry;
 334         bool leftmost = true;
 335
 336         while (*p) {
 337                 int comp;
 338
 339                 parent_node = *p;
 340                 entry = rb_entry(parent_node, struct btrfs_delayed_ref_node,
 341                                  ref_node);
 342                 comp = comp_refs(ins, entry, true);
 343                 if (comp < 0) {
 344                         p = &(*p)->rb_left;
 345                 } else if (comp > 0) {
 346                         p = &(*p)->rb_right;
 347                         leftmost = false;
 348                 } else {
 349                         return entry;
 350                 }
 351         }
 352
 353         rb_link_node(node, parent_node, p);
 354         rb_insert_color_cached(node, root, leftmost);
 355         return NULL;
 356 }
 357
 358 static struct btrfs_delayed_ref_head *find_first_ref_head(
 359                 struct btrfs_delayed_ref_root *dr)
 360 {
 361         struct rb_node *n;
 362         struct btrfs_delayed_ref_head *entry;
 363
 364         n = rb_first_cached(&dr->href_root);
 365         if (!n)
 366                 return NULL;
 367
 368         entry = rb_entry(n, struct btrfs_delayed_ref_head, href_node);
 369
 370         return entry;
 371 }
 372
 373 /*
 374  * Find a head entry based on bytenr. This returns the delayed ref head if it
 375  * was able to find one, or NULL if nothing was in that spot.  If return_bigger
 376  * is given, the next bigger entry is returned if no exact match is found.
 377  */
 378 static struct btrfs_delayed_ref_head *find_ref_head(
 379                 struct btrfs_delayed_ref_root *dr, u64 bytenr,
 380                 bool return_bigger)
 381 {
 382         struct rb_root *root = &dr->href_root.rb_root;
 383         struct rb_node *n;
 384         struct btrfs_delayed_ref_head *entry;
 385
 386         n = root->rb_node;
 387         entry = NULL;
 388         while (n) {
 389                 entry = rb_entry(n, struct btrfs_delayed_ref_head, href_node);
 390
 391                 if (bytenr < entry->bytenr)
 392                         n = n->rb_left;
 393                 else if (bytenr > entry->bytenr)
 394                         n = n->rb_right;
 395                 else
 396                         return entry;
 397         }
 398         if (entry && return_bigger) {
 399                 if (bytenr > entry->bytenr) {
 400                         n = rb_next(&entry->href_node);
 401                         if (!n)
 402                                 return NULL;
 403                         entry = rb_entry(n, struct btrfs_delayed_ref_head,
 404                                          href_node);
 405                 }
 406                 return entry;
 407         }
 408         return NULL;
 409 }
 410
 411 int btrfs_delayed_ref_lock(struct btrfs_delayed_ref_root *delayed_refs,
 412                            struct btrfs_delayed_ref_head *head)
 413 {
 414         lockdep_assert_held(&delayed_refs->lock);
 415         if (mutex_trylock(&head->mutex))
 416                 return 0;
 417
 418         refcount_inc(&head->refs);
 419         spin_unlock(&delayed_refs->lock);
 420
 421         mutex_lock(&head->mutex);
 422         spin_lock(&delayed_refs->lock);
 423         if (RB_EMPTY_NODE(&head->href_node)) {
 424                 mutex_unlock(&head->mutex);
 425                 btrfs_put_delayed_ref_head(head);
 426                 return -EAGAIN;
 427         }
 428         btrfs_put_delayed_ref_head(head);
 429         return 0;
 430 }
 431
 432 static inline void drop_delayed_ref(struct btrfs_trans_handle *trans,
 433                                     struct btrfs_delayed_ref_root *delayed_refs,
 434                                     struct btrfs_delayed_ref_head *head,
 435                                     struct btrfs_delayed_ref_node *ref)
 436 {
 437         lockdep_assert_held(&head->lock);
 438         rb_erase_cached(&ref->ref_node, &head->ref_tree);
 439         RB_CLEAR_NODE(&ref->ref_node);
 440         if (!list_empty(&ref->add_list))
 441                 list_del(&ref->add_list);
 442         ref->in_tree = 0;
 443         btrfs_put_delayed_ref(ref);
 444         atomic_dec(&delayed_refs->num_entries);
 445 }
 446
 447 static bool merge_ref(struct btrfs_trans_handle *trans,
 448                       struct btrfs_delayed_ref_root *delayed_refs,
 449                       struct btrfs_delayed_ref_head *head,
 450                       struct btrfs_delayed_ref_node *ref,
 451                       u64 seq)
 452 {
 453         struct btrfs_delayed_ref_node *next;
 454         struct rb_node *node = rb_next(&ref->ref_node);
 455         bool done = false;
 456
 457         while (!done && node) {
 458                 int mod;
 459
 460                 next = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
 461                 node = rb_next(node);
 462                 if (seq && next->seq >= seq)
 463                         break;
 464                 if (comp_refs(ref, next, false))
 465                         break;
 466
 467                 if (ref->action == next->action) {
 468                         mod = next->ref_mod;
 469                 } else {
 470                         if (ref->ref_mod < next->ref_mod) {
 471                                 swap(ref, next);
 472                                 done = true;
 473                         }
 474                         mod = -next->ref_mod;
 475                 }
 476
 477                 drop_delayed_ref(trans, delayed_refs, head, next);
 478                 ref->ref_mod += mod;
 479                 if (ref->ref_mod == 0) {
 480                         drop_delayed_ref(trans, delayed_refs, head, ref);
 481                         done = true;
 482                 } else {
 483                         /*
 484                          * Can't have multiples of the same ref on a tree block.
 485                          */
 486                         WARN_ON(ref->type == BTRFS_TREE_BLOCK_REF_KEY ||
 487                                 ref->type == BTRFS_SHARED_BLOCK_REF_KEY);
 488                 }
 489         }
 490
 491         return done;
 492 }
 493
 494 void btrfs_merge_delayed_refs(struct btrfs_trans_handle *trans,
 495                               struct btrfs_delayed_ref_root *delayed_refs,
 496                               struct btrfs_delayed_ref_head *head)
 497 {
 498         struct btrfs_fs_info *fs_info = trans->fs_info;
 499         struct btrfs_delayed_ref_node *ref;
 500         struct rb_node *node;
 501         u64 seq = 0;
 502
 503         lockdep_assert_held(&head->lock);
 504
 505         if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
 506                 return;
 507
 508         /* We don't have too many refs to merge for data. */
 509         if (head->is_data)
 510                 return;
 511
 512         seq = btrfs_tree_mod_log_lowest_seq(fs_info);
 513 again:
 514         for (node = rb_first_cached(&head->ref_tree); node;
 515              node = rb_next(node)) {
 516                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
 517                 if (seq && ref->seq >= seq)
 518                         continue;
 519                 if (merge_ref(trans, delayed_refs, head, ref, seq))
 520                         goto again;
 521         }
 522 }
 523
 524 int btrfs_check_delayed_seq(struct btrfs_fs_info *fs_info, u64 seq)
 525 {
 526         int ret = 0;
 527         u64 min_seq = btrfs_tree_mod_log_lowest_seq(fs_info);
 528
 529         if (min_seq != 0 && seq >= min_seq) {
 530                 btrfs_debug(fs_info,
 531                             "holding back delayed_ref %llu, lowest is %llu",
 532                             seq, min_seq);
 533                 ret = 1;
 534         }
 535
 536         return ret;
 537 }
 538
 539 struct btrfs_delayed_ref_head *btrfs_select_ref_head(
 540                 struct btrfs_delayed_ref_root *delayed_refs)
 541 {
 542         struct btrfs_delayed_ref_head *head;
 543
 544 again:
 545         head = find_ref_head(delayed_refs, delayed_refs->run_delayed_start,
 546                              true);
 547         if (!head && delayed_refs->run_delayed_start != 0) {
 548                 delayed_refs->run_delayed_start = 0;
 549                 head = find_first_ref_head(delayed_refs);
 550         }
 551         if (!head)
 552                 return NULL;
 553
 554         while (head->processing) {
 555                 struct rb_node *node;
 556
 557                 node = rb_next(&head->href_node);
 558                 if (!node) {
 559                         if (delayed_refs->run_delayed_start == 0)
 560                                 return NULL;
 561                         delayed_refs->run_delayed_start = 0;
 562                         goto again;
 563                 }
 564                 head = rb_entry(node, struct btrfs_delayed_ref_head,
 565                                 href_node);
 566         }
 567
 568         head->processing = 1;
 569         WARN_ON(delayed_refs->num_heads_ready == 0);
 570         delayed_refs->num_heads_ready--;
 571         delayed_refs->run_delayed_start = head->bytenr +
 572                 head->num_bytes;
 573         return head;
 574 }
 575
 576 void btrfs_delete_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
 577                            struct btrfs_delayed_ref_head *head)
 578 {
 579         lockdep_assert_held(&delayed_refs->lock);
 580         lockdep_assert_held(&head->lock);
 581
 582         rb_erase_cached(&head->href_node, &delayed_refs->href_root);
 583         RB_CLEAR_NODE(&head->href_node);
 584         atomic_dec(&delayed_refs->num_entries);
 585         delayed_refs->num_heads--;
 586         if (head->processing == 0)
 587                 delayed_refs->num_heads_ready--;
 588 }
 589
 590 /*
 591  * Helper to insert the ref_node to the tail or merge with tail.
 592  *
 593  * Return 0 for insert.
 594  * Return >0 for merge.
 595  */
 596 static int insert_delayed_ref(struct btrfs_trans_handle *trans,
 597                               struct btrfs_delayed_ref_root *root,
 598                               struct btrfs_delayed_ref_head *href,
 599                               struct btrfs_delayed_ref_node *ref)
 600 {
 601         struct btrfs_delayed_ref_node *exist;
 602         int mod;
 603         int ret = 0;
 604
 605         spin_lock(&href->lock);
 606         exist = tree_insert(&href->ref_tree, ref);
 607         if (!exist)
 608                 goto inserted;
 609
 610         /* Now we are sure we can merge */
 611         ret = 1;
 612         if (exist->action == ref->action) {
 613                 mod = ref->ref_mod;
 614         } else {
 615                 /* Need to change action */
 616                 if (exist->ref_mod < ref->ref_mod) {
 617                         exist->action = ref->action;
 618                         mod = -exist->ref_mod;
 619                         exist->ref_mod = ref->ref_mod;
 620                         if (ref->action == BTRFS_ADD_DELAYED_REF)
 621                                 list_add_tail(&exist->add_list,
 622                                               &href->ref_add_list);
 623                         else if (ref->action == BTRFS_DROP_DELAYED_REF) {
 624                                 ASSERT(!list_empty(&exist->add_list));
 625                                 list_del(&exist->add_list);
 626                         } else {
 627                                 ASSERT(0);
 628                         }
 629                 } else
 630                         mod = -ref->ref_mod;
 631         }
 632         exist->ref_mod += mod;
 633
 634         /* remove existing tail if its ref_mod is zero */
 635         if (exist->ref_mod == 0)
 636                 drop_delayed_ref(trans, root, href, exist);
 637         spin_unlock(&href->lock);
 638         return ret;
 639 inserted:
 640         if (ref->action == BTRFS_ADD_DELAYED_REF)
 641                 list_add_tail(&ref->add_list, &href->ref_add_list);
 642         atomic_inc(&root->num_entries);
 643         spin_unlock(&href->lock);
 644         return ret;
 645 }
 646
 647 /*
 648  * helper function to update the accounting in the head ref
 649  * existing and update must have the same bytenr
 650  */
 651 static noinline void update_existing_head_ref(struct btrfs_trans_handle *trans,
 652                          struct btrfs_delayed_ref_head *existing,
 653                          struct btrfs_delayed_ref_head *update)
 654 {
 655         struct btrfs_delayed_ref_root *delayed_refs =
 656                 &trans->transaction->delayed_refs;
 657         struct btrfs_fs_info *fs_info = trans->fs_info;
 658         int old_ref_mod;
 659
 660         BUG_ON(existing->is_data != update->is_data);
 661
 662         spin_lock(&existing->lock);
 663         if (update->must_insert_reserved) {
 664                 /* if the extent was freed and then
 665                  * reallocated before the delayed ref
 666                  * entries were processed, we can end up
 667                  * with an existing head ref without
 668                  * the must_insert_reserved flag set.
 669                  * Set it again here
 670                  */
 671                 existing->must_insert_reserved = update->must_insert_reserved;
 672
 673                 /*
 674                  * update the num_bytes so we make sure the accounting
 675                  * is done correctly
 676                  */
 677                 existing->num_bytes = update->num_bytes;
 678
 679         }
 680
 681         if (update->extent_op) {
 682                 if (!existing->extent_op) {
 683                         existing->extent_op = update->extent_op;
 684                 } else {
 685                         if (update->extent_op->update_key) {
 686                                 memcpy(&existing->extent_op->key,
 687                                        &update->extent_op->key,
 688                                        sizeof(update->extent_op->key));
 689                                 existing->extent_op->update_key = true;
 690                         }
 691                         if (update->extent_op->update_flags) {
 692                                 existing->extent_op->flags_to_set |=
 693                                         update->extent_op->flags_to_set;
 694                                 existing->extent_op->update_flags = true;
 695                         }
 696                         btrfs_free_delayed_extent_op(update->extent_op);
 697                 }
 698         }
 699         /*
 700          * update the reference mod on the head to reflect this new operation,
 701          * only need the lock for this case cause we could be processing it
 702          * currently, for refs we just added we know we're a-ok.
 703          */
 704         old_ref_mod = existing->total_ref_mod;
 705         existing->ref_mod += update->ref_mod;
 706         existing->total_ref_mod += update->ref_mod;
 707
 708         /*
 709          * If we are going to from a positive ref mod to a negative or vice
 710          * versa we need to make sure to adjust pending_csums accordingly.
 711          */
 712         if (existing->is_data) {
 713                 u64 csum_leaves =
 714                         btrfs_csum_bytes_to_leaves(fs_info,
 715                                                    existing->num_bytes);
 716
 717                 if (existing->total_ref_mod >= 0 && old_ref_mod < 0) {
 718                         delayed_refs->pending_csums -= existing->num_bytes;
 719                         btrfs_delayed_refs_rsv_release(fs_info, csum_leaves);
 720                 }
 721                 if (existing->total_ref_mod < 0 && old_ref_mod >= 0) {
 722                         delayed_refs->pending_csums += existing->num_bytes;
 723                         trans->delayed_ref_updates += csum_leaves;
 724                 }
 725         }
 726
 727         spin_unlock(&existing->lock);
 728 }
 729
 730 static void init_delayed_ref_head(struct btrfs_delayed_ref_head *head_ref,
 731                                   struct btrfs_qgroup_extent_record *qrecord,
 732                                   u64 bytenr, u64 num_bytes, u64 ref_root,
 733                                   u64 reserved, int action, bool is_data,
 734                                   bool is_system)
 735 {
 736         int count_mod = 1;
 737         int must_insert_reserved = 0;
 738
 739         /* If reserved is provided, it must be a data extent. */
 740         BUG_ON(!is_data && reserved);
 741
 742         /*
 743          * The head node stores the sum of all the mods, so dropping a ref
 744          * should drop the sum in the head node by one.
 745          */
 746         if (action == BTRFS_UPDATE_DELAYED_HEAD)
 747                 count_mod = 0;
 748         else if (action == BTRFS_DROP_DELAYED_REF)
 749                 count_mod = -1;
 750
 751         /*
 752          * BTRFS_ADD_DELAYED_EXTENT means that we need to update the reserved
 753          * accounting when the extent is finally added, or if a later
 754          * modification deletes the delayed ref without ever inserting the
 755          * extent into the extent allocation tree.  ref->must_insert_reserved
 756          * is the flag used to record that accounting mods are required.
 757          *
 758          * Once we record must_insert_reserved, switch the action to
 759          * BTRFS_ADD_DELAYED_REF because other special casing is not required.
 760          */
 761         if (action == BTRFS_ADD_DELAYED_EXTENT)
 762                 must_insert_reserved = 1;
 763         else
 764                 must_insert_reserved = 0;
 765
 766         refcount_set(&head_ref->refs, 1);
 767         head_ref->bytenr = bytenr;
 768         head_ref->num_bytes = num_bytes;
 769         head_ref->ref_mod = count_mod;
 770         head_ref->must_insert_reserved = must_insert_reserved;
 771         head_ref->is_data = is_data;
 772         head_ref->is_system = is_system;
 773         head_ref->ref_tree = RB_ROOT_CACHED;
 774         INIT_LIST_HEAD(&head_ref->ref_add_list);
 775         RB_CLEAR_NODE(&head_ref->href_node);
 776         head_ref->processing = 0;
 777         head_ref->total_ref_mod = count_mod;
 778         spin_lock_init(&head_ref->lock);
 779         mutex_init(&head_ref->mutex);
 780
 781         if (qrecord) {
 782                 if (ref_root && reserved) {
 783                         qrecord->data_rsv = reserved;
 784                         qrecord->data_rsv_refroot = ref_root;
 785                 }
 786                 qrecord->bytenr = bytenr;
 787                 qrecord->num_bytes = num_bytes;
 788                 qrecord->old_roots = NULL;
 789         }
 790 }
 791
 792 /*
 793  * helper function to actually insert a head node into the rbtree.
 794  * this does all the dirty work in terms of maintaining the correct
 795  * overall modification count.
 796  */
 797 static noinline struct btrfs_delayed_ref_head *
 798 add_delayed_ref_head(struct btrfs_trans_handle *trans,
 799                      struct btrfs_delayed_ref_head *head_ref,
 800                      struct btrfs_qgroup_extent_record *qrecord,
 801                      int action, int *qrecord_inserted_ret)
 802 {
 803         struct btrfs_delayed_ref_head *existing;
 804         struct btrfs_delayed_ref_root *delayed_refs;
 805         int qrecord_inserted = 0;
 806
 807         delayed_refs = &trans->transaction->delayed_refs;
 808
 809         /* Record qgroup extent info if provided */
 810         if (qrecord) {
 811                 if (btrfs_qgroup_trace_extent_nolock(trans->fs_info,
 812                                         delayed_refs, qrecord))
 813                         kfree(qrecord);
 814                 else
 815                         qrecord_inserted = 1;
 816         }
 817
 818         trace_add_delayed_ref_head(trans->fs_info, head_ref, action);
 819
 820         existing = htree_insert(&delayed_refs->href_root,
 821                                 &head_ref->href_node);
 822         if (existing) {
 823                 update_existing_head_ref(trans, existing, head_ref);
 824                 /*
 825                  * we've updated the existing ref, free the newly
 826                  * allocated ref
 827                  */
 828                 kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
 829                 head_ref = existing;
 830         } else {
 831                 if (head_ref->is_data && head_ref->ref_mod < 0) {
 832                         delayed_refs->pending_csums += head_ref->num_bytes;
 833                         trans->delayed_ref_updates +=
 834                                 btrfs_csum_bytes_to_leaves(trans->fs_info,
 835                                                            head_ref->num_bytes);
 836                 }
 837                 delayed_refs->num_heads++;
 838                 delayed_refs->num_heads_ready++;
 839                 atomic_inc(&delayed_refs->num_entries);
 840                 trans->delayed_ref_updates++;
 841         }
 842         if (qrecord_inserted_ret)
 843                 *qrecord_inserted_ret = qrecord_inserted;
 844
 845         return head_ref;
 846 }
 847
 848 /*
 849  * init_delayed_ref_common - Initialize the structure which represents a
 850  *                           modification to a an extent.
 851  *
 852  * @fs_info:    Internal to the mounted filesystem mount structure.
 853  *
 854  * @ref:        The structure which is going to be initialized.
 855  *
 856  * @bytenr:     The logical address of the extent for which a modification is
 857  *              going to be recorded.
 858  *
 859  * @num_bytes:  Size of the extent whose modification is being recorded.
 860  *
 861  * @ref_root:   The id of the root where this modification has originated, this
 862  *              can be either one of the well-known metadata trees or the
 863  *              subvolume id which references this extent.
 864  *
 865  * @action:     Can be one of BTRFS_ADD_DELAYED_REF/BTRFS_DROP_DELAYED_REF or
 866  *              BTRFS_ADD_DELAYED_EXTENT
 867  *
 868  * @ref_type:   Holds the type of the extent which is being recorded, can be
 869  *              one of BTRFS_SHARED_BLOCK_REF_KEY/BTRFS_TREE_BLOCK_REF_KEY
 870  *              when recording a metadata extent or BTRFS_SHARED_DATA_REF_KEY/
 871  *              BTRFS_EXTENT_DATA_REF_KEY when recording data extent
 872  */
 873 static void init_delayed_ref_common(struct btrfs_fs_info *fs_info,
 874                                     struct btrfs_delayed_ref_node *ref,
 875                                     u64 bytenr, u64 num_bytes, u64 ref_root,
 876                                     int action, u8 ref_type)
 877 {
 878         u64 seq = 0;
 879
 880         if (action == BTRFS_ADD_DELAYED_EXTENT)
 881                 action = BTRFS_ADD_DELAYED_REF;
 882
 883         if (is_fstree(ref_root))
 884                 seq = atomic64_read(&fs_info->tree_mod_seq);
 885
 886         refcount_set(&ref->refs, 1);
 887         ref->bytenr = bytenr;
 888         ref->num_bytes = num_bytes;
 889         ref->ref_mod = 1;
 890         ref->action = action;
 891         ref->is_head = 0;
 892         ref->in_tree = 1;
 893         ref->seq = seq;
 894         ref->type = ref_type;
 895         RB_CLEAR_NODE(&ref->ref_node);
 896         INIT_LIST_HEAD(&ref->add_list);
 897 }
 898
 899 /*
 900  * add a delayed tree ref.  This does all of the accounting required
 901  * to make sure the delayed ref is eventually processed before this
 902  * transaction commits.
 903  */
 904 int btrfs_add_delayed_tree_ref(struct btrfs_trans_handle *trans,
 905                                struct btrfs_ref *generic_ref,
 906                                struct btrfs_delayed_extent_op *extent_op)
 907 {
 908         struct btrfs_fs_info *fs_info = trans->fs_info;
 909         struct btrfs_delayed_tree_ref *ref;
 910         struct btrfs_delayed_ref_head *head_ref;
 911         struct btrfs_delayed_ref_root *delayed_refs;
 912         struct btrfs_qgroup_extent_record *record = NULL;
 913         int qrecord_inserted;
 914         bool is_system;
 915         int action = generic_ref->action;
 916         int level = generic_ref->tree_ref.level;
 917         int ret;
 918         u64 bytenr = generic_ref->bytenr;
 919         u64 num_bytes = generic_ref->len;
 920         u64 parent = generic_ref->parent;
 921         u8 ref_type;
 922
 923         is_system = (generic_ref->tree_ref.owning_root == BTRFS_CHUNK_TREE_OBJECTID);
 924
 925         ASSERT(generic_ref->type == BTRFS_REF_METADATA && generic_ref->action);
 926         ref = kmem_cache_alloc(btrfs_delayed_tree_ref_cachep, GFP_NOFS);
 927         if (!ref)
 928                 return -ENOMEM;
 929
 930         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
 931         if (!head_ref) {
 932                 kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
 933                 return -ENOMEM;
 934         }
 935
 936         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) &&
 937             !generic_ref->skip_qgroup) {
 938                 record = kzalloc(sizeof(*record), GFP_NOFS);
 939                 if (!record) {
 940                         kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
 941                         kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
 942                         return -ENOMEM;
 943                 }
 944         }
 945
 946         if (parent)
 947                 ref_type = BTRFS_SHARED_BLOCK_REF_KEY;
 948         else
 949                 ref_type = BTRFS_TREE_BLOCK_REF_KEY;
 950
 951         init_delayed_ref_common(fs_info, &ref->node, bytenr, num_bytes,
 952                                 generic_ref->tree_ref.owning_root, action,
 953                                 ref_type);
 954         ref->root = generic_ref->tree_ref.owning_root;
 955         ref->parent = parent;
 956         ref->level = level;
 957
 958         init_delayed_ref_head(head_ref, record, bytenr, num_bytes,
 959                               generic_ref->tree_ref.owning_root, 0, action,
 960                               false, is_system);
 961         head_ref->extent_op = extent_op;
 962
 963         delayed_refs = &trans->transaction->delayed_refs;
 964         spin_lock(&delayed_refs->lock);
 965
 966         /*
 967          * insert both the head node and the new ref without dropping
 968          * the spin lock
 969          */
 970         head_ref = add_delayed_ref_head(trans, head_ref, record,
 971                                         action, &qrecord_inserted);
 972
 973         ret = insert_delayed_ref(trans, delayed_refs, head_ref, &ref->node);
 974         spin_unlock(&delayed_refs->lock);
 975
 976         /*
 977          * Need to update the delayed_refs_rsv with any changes we may have
 978          * made.
 979          */
 980         btrfs_update_delayed_refs_rsv(trans);
 981
 982         trace_add_delayed_tree_ref(fs_info, &ref->node, ref,
 983                                    action == BTRFS_ADD_DELAYED_EXTENT ?
 984                                    BTRFS_ADD_DELAYED_REF : action);
 985         if (ret > 0)
 986                 kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
 987
 988         if (qrecord_inserted)
 989                 btrfs_qgroup_trace_extent_post(trans, record);
 990
 991         return 0;
 992 }
 993
 994 /*
 995  * add a delayed data ref. it's similar to btrfs_add_delayed_tree_ref.
 996  */
 997 int btrfs_add_delayed_data_ref(struct btrfs_trans_handle *trans,
 998                                struct btrfs_ref *generic_ref,
 999                                u64 reserved)
1000 {
1001         struct btrfs_fs_info *fs_info = trans->fs_info;
1002         struct btrfs_delayed_data_ref *ref;
1003         struct btrfs_delayed_ref_head *head_ref;
1004         struct btrfs_delayed_ref_root *delayed_refs;
1005         struct btrfs_qgroup_extent_record *record = NULL;
1006         int qrecord_inserted;
1007         int action = generic_ref->action;
1008         int ret;
1009         u64 bytenr = generic_ref->bytenr;
1010         u64 num_bytes = generic_ref->len;
1011         u64 parent = generic_ref->parent;
1012         u64 ref_root = generic_ref->data_ref.owning_root;
1013         u64 owner = generic_ref->data_ref.ino;
1014         u64 offset = generic_ref->data_ref.offset;
1015         u8 ref_type;
1016
1017         ASSERT(generic_ref->type == BTRFS_REF_DATA && action);
1018         ref = kmem_cache_alloc(btrfs_delayed_data_ref_cachep, GFP_NOFS);
1019         if (!ref)
1020                 return -ENOMEM;
1021
1022         if (parent)
1023                 ref_type = BTRFS_SHARED_DATA_REF_KEY;
1024         else
1025                 ref_type = BTRFS_EXTENT_DATA_REF_KEY;
1026         init_delayed_ref_common(fs_info, &ref->node, bytenr, num_bytes,
1027                                 ref_root, action, ref_type);
1028         ref->root = ref_root;
1029         ref->parent = parent;
1030         ref->objectid = owner;
1031         ref->offset = offset;
1032
1033
1034         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
1035         if (!head_ref) {
1036                 kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
1037                 return -ENOMEM;
1038         }
1039
1040         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) &&
1041             !generic_ref->skip_qgroup) {
1042                 record = kzalloc(sizeof(*record), GFP_NOFS);
1043                 if (!record) {
1044                         kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
1045                         kmem_cache_free(btrfs_delayed_ref_head_cachep,
1046                                         head_ref);
1047                         return -ENOMEM;
1048                 }
1049         }
1050
1051         init_delayed_ref_head(head_ref, record, bytenr, num_bytes, ref_root,
1052                               reserved, action, true, false);
1053         head_ref->extent_op = NULL;
1054
1055         delayed_refs = &trans->transaction->delayed_refs;
1056         spin_lock(&delayed_refs->lock);
1057
1058         /*
1059          * insert both the head node and the new ref without dropping
1060          * the spin lock
1061          */
1062         head_ref = add_delayed_ref_head(trans, head_ref, record,
1063                                         action, &qrecord_inserted);
1064
1065         ret = insert_delayed_ref(trans, delayed_refs, head_ref, &ref->node);
1066         spin_unlock(&delayed_refs->lock);
1067
1068         /*
1069          * Need to update the delayed_refs_rsv with any changes we may have
1070          * made.
1071          */
1072         btrfs_update_delayed_refs_rsv(trans);
1073
1074         trace_add_delayed_data_ref(trans->fs_info, &ref->node, ref,
1075                                    action == BTRFS_ADD_DELAYED_EXTENT ?
1076                                    BTRFS_ADD_DELAYED_REF : action);
1077         if (ret > 0)
1078                 kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
1079
1080
1081         if (qrecord_inserted)
1082                 return btrfs_qgroup_trace_extent_post(trans, record);
1083         return 0;
1084 }
1085
1086 int btrfs_add_delayed_extent_op(struct btrfs_trans_handle *trans,
1087                                 u64 bytenr, u64 num_bytes,
1088                                 struct btrfs_delayed_extent_op *extent_op)
1089 {
1090         struct btrfs_delayed_ref_head *head_ref;
1091         struct btrfs_delayed_ref_root *delayed_refs;
1092
1093         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
1094         if (!head_ref)
1095                 return -ENOMEM;
1096
1097         init_delayed_ref_head(head_ref, NULL, bytenr, num_bytes, 0, 0,
1098                               BTRFS_UPDATE_DELAYED_HEAD, false, false);
1099         head_ref->extent_op = extent_op;
1100
1101         delayed_refs = &trans->transaction->delayed_refs;
1102         spin_lock(&delayed_refs->lock);
1103
1104         add_delayed_ref_head(trans, head_ref, NULL, BTRFS_UPDATE_DELAYED_HEAD,
1105                              NULL);
1106
1107         spin_unlock(&delayed_refs->lock);
1108
1109         /*
1110          * Need to update the delayed_refs_rsv with any changes we may have
1111          * made.
1112          */
1113         btrfs_update_delayed_refs_rsv(trans);
1114         return 0;
1115 }
1116
1117 /*
1118  * This does a simple search for the head node for a given extent.  Returns the
1119  * head node if found, or NULL if not.
1120  */
1121 struct btrfs_delayed_ref_head *
1122 btrfs_find_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs, u64 bytenr)
1123 {
1124         lockdep_assert_held(&delayed_refs->lock);
1125
1126         return find_ref_head(delayed_refs, bytenr, false);
1127 }
1128
1129 void __cold btrfs_delayed_ref_exit(void)
1130 {
1131         kmem_cache_destroy(btrfs_delayed_ref_head_cachep);
1132         kmem_cache_destroy(btrfs_delayed_tree_ref_cachep);
1133         kmem_cache_destroy(btrfs_delayed_data_ref_cachep);
1134         kmem_cache_destroy(btrfs_delayed_extent_op_cachep);
1135 }
1136
1137 int __init btrfs_delayed_ref_init(void)
1138 {
1139         btrfs_delayed_ref_head_cachep = kmem_cache_create(
1140                                 "btrfs_delayed_ref_head",
1141                                 sizeof(struct btrfs_delayed_ref_head), 0,
1142                                 SLAB_MEM_SPREAD, NULL);
1143         if (!btrfs_delayed_ref_head_cachep)
1144                 goto fail;
1145
1146         btrfs_delayed_tree_ref_cachep = kmem_cache_create(
1147                                 "btrfs_delayed_tree_ref",
1148                                 sizeof(struct btrfs_delayed_tree_ref), 0,
1149                                 SLAB_MEM_SPREAD, NULL);
1150         if (!btrfs_delayed_tree_ref_cachep)
1151                 goto fail;
1152
1153         btrfs_delayed_data_ref_cachep = kmem_cache_create(
1154                                 "btrfs_delayed_data_ref",
1155                                 sizeof(struct btrfs_delayed_data_ref), 0,
1156                                 SLAB_MEM_SPREAD, NULL);
1157         if (!btrfs_delayed_data_ref_cachep)
1158                 goto fail;
1159
1160         btrfs_delayed_extent_op_cachep = kmem_cache_create(
1161                                 "btrfs_delayed_extent_op",
1162                                 sizeof(struct btrfs_delayed_extent_op), 0,
1163                                 SLAB_MEM_SPREAD, NULL);
1164         if (!btrfs_delayed_extent_op_cachep)
1165                 goto fail;
1166
1167         return 0;
1168 fail:
1169         btrfs_delayed_ref_exit();
1170         return -ENOMEM;
1171 }