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 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans)
  57 {
  58         u64 num_entries =
  59                 atomic_read(&trans->transaction->delayed_refs.num_entries);
  60         u64 avg_runtime;
  61         u64 val;
  62
  63         smp_mb();
  64         avg_runtime = trans->fs_info->avg_delayed_ref_runtime;
  65         val = num_entries * avg_runtime;
  66         if (val >= NSEC_PER_SEC)
  67                 return 1;
  68         if (val >= NSEC_PER_SEC / 2)
  69                 return 2;
  70
  71         return btrfs_check_space_for_delayed_refs(trans->fs_info);
  72 }
  73
  74 /*
  75  * Release a ref head's reservation.
  76  *
  77  * @fs_info:  the filesystem
  78  * @nr:       number of items to drop
  79  *
  80  * Drops the delayed ref head's count from the delayed refs rsv and free any
  81  * excess reservation we had.
  82  */
  83 void btrfs_delayed_refs_rsv_release(struct btrfs_fs_info *fs_info, int nr)
  84 {
  85         struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
  86         u64 num_bytes = btrfs_calc_insert_metadata_size(fs_info, nr);
  87         u64 released = 0;
  88
  89         /*
  90          * We have to check the mount option here because we could be enabling
  91          * the free space tree for the first time and don't have the compat_ro
  92          * option set yet.
  93          *
  94          * We need extra reservations if we have the free space tree because
  95          * we'll have to modify that tree as well.
  96          */
  97         if (btrfs_test_opt(fs_info, FREE_SPACE_TREE))
  98                 num_bytes *= 2;
  99
 100         released = btrfs_block_rsv_release(fs_info, block_rsv, num_bytes, NULL);
 101         if (released)
 102                 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
 103                                               0, released, 0);
 104 }
 105
 106 /*
 107  * Adjust the size of the delayed refs rsv.
 108  *
 109  * This is to be called anytime we may have adjusted trans->delayed_ref_updates,
 110  * it'll calculate the additional size and add it to the delayed_refs_rsv.
 111  */
 112 void btrfs_update_delayed_refs_rsv(struct btrfs_trans_handle *trans)
 113 {
 114         struct btrfs_fs_info *fs_info = trans->fs_info;
 115         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_refs_rsv;
 116         u64 num_bytes;
 117
 118         if (!trans->delayed_ref_updates)
 119                 return;
 120
 121         num_bytes = btrfs_calc_insert_metadata_size(fs_info,
 122                                                     trans->delayed_ref_updates);
 123         /*
 124          * We have to check the mount option here because we could be enabling
 125          * the free space tree for the first time and don't have the compat_ro
 126          * option set yet.
 127          *
 128          * We need extra reservations if we have the free space tree because
 129          * we'll have to modify that tree as well.
 130          */
 131         if (btrfs_test_opt(fs_info, FREE_SPACE_TREE))
 132                 num_bytes *= 2;
 133
 134         spin_lock(&delayed_rsv->lock);
 135         delayed_rsv->size += num_bytes;
 136         delayed_rsv->full = false;
 137         spin_unlock(&delayed_rsv->lock);
 138         trans->delayed_ref_updates = 0;
 139 }
 140
 141 /*
 142  * Transfer bytes to our delayed refs rsv.
 143  *
 144  * @fs_info:   the filesystem
 145  * @src:       source block rsv to transfer from
 146  * @num_bytes: number of bytes to transfer
 147  *
 148  * This transfers up to the num_bytes amount from the src rsv to the
 149  * delayed_refs_rsv.  Any extra bytes are returned to the space info.
 150  */
 151 void btrfs_migrate_to_delayed_refs_rsv(struct btrfs_fs_info *fs_info,
 152                                        struct btrfs_block_rsv *src,
 153                                        u64 num_bytes)
 154 {
 155         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
 156         u64 to_free = 0;
 157
 158         spin_lock(&src->lock);
 159         src->reserved -= num_bytes;
 160         src->size -= num_bytes;
 161         spin_unlock(&src->lock);
 162
 163         spin_lock(&delayed_refs_rsv->lock);
 164         if (delayed_refs_rsv->size > delayed_refs_rsv->reserved) {
 165                 u64 delta = delayed_refs_rsv->size -
 166                         delayed_refs_rsv->reserved;
 167                 if (num_bytes > delta) {
 168                         to_free = num_bytes - delta;
 169                         num_bytes = delta;
 170                 }
 171         } else {
 172                 to_free = num_bytes;
 173                 num_bytes = 0;
 174         }
 175
 176         if (num_bytes)
 177                 delayed_refs_rsv->reserved += num_bytes;
 178         if (delayed_refs_rsv->reserved >= delayed_refs_rsv->size)
 179                 delayed_refs_rsv->full = true;
 180         spin_unlock(&delayed_refs_rsv->lock);
 181
 182         if (num_bytes)
 183                 trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
 184                                               0, num_bytes, 1);
 185         if (to_free)
 186                 btrfs_space_info_free_bytes_may_use(fs_info,
 187                                 delayed_refs_rsv->space_info, to_free);
 188 }
 189
 190 /*
 191  * Refill based on our delayed refs usage.
 192  *
 193  * @fs_info: the filesystem
 194  * @flush:   control how we can flush for this reservation.
 195  *
 196  * This will refill the delayed block_rsv up to 1 items size worth of space and
 197  * will return -ENOSPC if we can't make the reservation.
 198  */
 199 int btrfs_delayed_refs_rsv_refill(struct btrfs_fs_info *fs_info,
 200                                   enum btrfs_reserve_flush_enum flush)
 201 {
 202         struct btrfs_block_rsv *block_rsv = &fs_info->delayed_refs_rsv;
 203         u64 limit = btrfs_calc_insert_metadata_size(fs_info, 1);
 204         u64 num_bytes = 0;
 205         int ret = -ENOSPC;
 206
 207         spin_lock(&block_rsv->lock);
 208         if (block_rsv->reserved < block_rsv->size) {
 209                 num_bytes = block_rsv->size - block_rsv->reserved;
 210                 num_bytes = min(num_bytes, limit);
 211         }
 212         spin_unlock(&block_rsv->lock);
 213
 214         if (!num_bytes)
 215                 return 0;
 216
 217         ret = btrfs_reserve_metadata_bytes(fs_info, block_rsv, num_bytes, flush);
 218         if (ret)
 219                 return ret;
 220         btrfs_block_rsv_add_bytes(block_rsv, num_bytes, 0);
 221         trace_btrfs_space_reservation(fs_info, "delayed_refs_rsv",
 222                                       0, num_bytes, 1);
 223         return 0;
 224 }
 225
 226 /*
 227  * compare two delayed tree backrefs with same bytenr and type
 228  */
 229 static int comp_tree_refs(struct btrfs_delayed_tree_ref *ref1,
 230                           struct btrfs_delayed_tree_ref *ref2)
 231 {
 232         if (ref1->node.type == BTRFS_TREE_BLOCK_REF_KEY) {
 233                 if (ref1->root < ref2->root)
 234                         return -1;
 235                 if (ref1->root > ref2->root)
 236                         return 1;
 237         } else {
 238                 if (ref1->parent < ref2->parent)
 239                         return -1;
 240                 if (ref1->parent > ref2->parent)
 241                         return 1;
 242         }
 243         return 0;
 244 }
 245
 246 /*
 247  * compare two delayed data backrefs with same bytenr and type
 248  */
 249 static int comp_data_refs(struct btrfs_delayed_data_ref *ref1,
 250                           struct btrfs_delayed_data_ref *ref2)
 251 {
 252         if (ref1->node.type == BTRFS_EXTENT_DATA_REF_KEY) {
 253                 if (ref1->root < ref2->root)
 254                         return -1;
 255                 if (ref1->root > ref2->root)
 256                         return 1;
 257                 if (ref1->objectid < ref2->objectid)
 258                         return -1;
 259                 if (ref1->objectid > ref2->objectid)
 260                         return 1;
 261                 if (ref1->offset < ref2->offset)
 262                         return -1;
 263                 if (ref1->offset > ref2->offset)
 264                         return 1;
 265         } else {
 266                 if (ref1->parent < ref2->parent)
 267                         return -1;
 268                 if (ref1->parent > ref2->parent)
 269                         return 1;
 270         }
 271         return 0;
 272 }
 273
 274 static int comp_refs(struct btrfs_delayed_ref_node *ref1,
 275                      struct btrfs_delayed_ref_node *ref2,
 276                      bool check_seq)
 277 {
 278         int ret = 0;
 279
 280         if (ref1->type < ref2->type)
 281                 return -1;
 282         if (ref1->type > ref2->type)
 283                 return 1;
 284         if (ref1->type == BTRFS_TREE_BLOCK_REF_KEY ||
 285             ref1->type == BTRFS_SHARED_BLOCK_REF_KEY)
 286                 ret = comp_tree_refs(btrfs_delayed_node_to_tree_ref(ref1),
 287                                      btrfs_delayed_node_to_tree_ref(ref2));
 288         else
 289                 ret = comp_data_refs(btrfs_delayed_node_to_data_ref(ref1),
 290                                      btrfs_delayed_node_to_data_ref(ref2));
 291         if (ret)
 292                 return ret;
 293         if (check_seq) {
 294                 if (ref1->seq < ref2->seq)
 295                         return -1;
 296                 if (ref1->seq > ref2->seq)
 297                         return 1;
 298         }
 299         return 0;
 300 }
 301
 302 /* insert a new ref to head ref rbtree */
 303 static struct btrfs_delayed_ref_head *htree_insert(struct rb_root_cached *root,
 304                                                    struct rb_node *node)
 305 {
 306         struct rb_node **p = &root->rb_root.rb_node;
 307         struct rb_node *parent_node = NULL;
 308         struct btrfs_delayed_ref_head *entry;
 309         struct btrfs_delayed_ref_head *ins;
 310         u64 bytenr;
 311         bool leftmost = true;
 312
 313         ins = rb_entry(node, struct btrfs_delayed_ref_head, href_node);
 314         bytenr = ins->bytenr;
 315         while (*p) {
 316                 parent_node = *p;
 317                 entry = rb_entry(parent_node, struct btrfs_delayed_ref_head,
 318                                  href_node);
 319
 320                 if (bytenr < entry->bytenr) {
 321                         p = &(*p)->rb_left;
 322                 } else if (bytenr > entry->bytenr) {
 323                         p = &(*p)->rb_right;
 324                         leftmost = false;
 325                 } else {
 326                         return entry;
 327                 }
 328         }
 329
 330         rb_link_node(node, parent_node, p);
 331         rb_insert_color_cached(node, root, leftmost);
 332         return NULL;
 333 }
 334
 335 static struct btrfs_delayed_ref_node* tree_insert(struct rb_root_cached *root,
 336                 struct btrfs_delayed_ref_node *ins)
 337 {
 338         struct rb_node **p = &root->rb_root.rb_node;
 339         struct rb_node *node = &ins->ref_node;
 340         struct rb_node *parent_node = NULL;
 341         struct btrfs_delayed_ref_node *entry;
 342         bool leftmost = true;
 343
 344         while (*p) {
 345                 int comp;
 346
 347                 parent_node = *p;
 348                 entry = rb_entry(parent_node, struct btrfs_delayed_ref_node,
 349                                  ref_node);
 350                 comp = comp_refs(ins, entry, true);
 351                 if (comp < 0) {
 352                         p = &(*p)->rb_left;
 353                 } else if (comp > 0) {
 354                         p = &(*p)->rb_right;
 355                         leftmost = false;
 356                 } else {
 357                         return entry;
 358                 }
 359         }
 360
 361         rb_link_node(node, parent_node, p);
 362         rb_insert_color_cached(node, root, leftmost);
 363         return NULL;
 364 }
 365
 366 static struct btrfs_delayed_ref_head *find_first_ref_head(
 367                 struct btrfs_delayed_ref_root *dr)
 368 {
 369         struct rb_node *n;
 370         struct btrfs_delayed_ref_head *entry;
 371
 372         n = rb_first_cached(&dr->href_root);
 373         if (!n)
 374                 return NULL;
 375
 376         entry = rb_entry(n, struct btrfs_delayed_ref_head, href_node);
 377
 378         return entry;
 379 }
 380
 381 /*
 382  * Find a head entry based on bytenr. This returns the delayed ref head if it
 383  * was able to find one, or NULL if nothing was in that spot.  If return_bigger
 384  * is given, the next bigger entry is returned if no exact match is found.
 385  */
 386 static struct btrfs_delayed_ref_head *find_ref_head(
 387                 struct btrfs_delayed_ref_root *dr, u64 bytenr,
 388                 bool return_bigger)
 389 {
 390         struct rb_root *root = &dr->href_root.rb_root;
 391         struct rb_node *n;
 392         struct btrfs_delayed_ref_head *entry;
 393
 394         n = root->rb_node;
 395         entry = NULL;
 396         while (n) {
 397                 entry = rb_entry(n, struct btrfs_delayed_ref_head, href_node);
 398
 399                 if (bytenr < entry->bytenr)
 400                         n = n->rb_left;
 401                 else if (bytenr > entry->bytenr)
 402                         n = n->rb_right;
 403                 else
 404                         return entry;
 405         }
 406         if (entry && return_bigger) {
 407                 if (bytenr > entry->bytenr) {
 408                         n = rb_next(&entry->href_node);
 409                         if (!n)
 410                                 return NULL;
 411                         entry = rb_entry(n, struct btrfs_delayed_ref_head,
 412                                          href_node);
 413                 }
 414                 return entry;
 415         }
 416         return NULL;
 417 }
 418
 419 int btrfs_delayed_ref_lock(struct btrfs_delayed_ref_root *delayed_refs,
 420                            struct btrfs_delayed_ref_head *head)
 421 {
 422         lockdep_assert_held(&delayed_refs->lock);
 423         if (mutex_trylock(&head->mutex))
 424                 return 0;
 425
 426         refcount_inc(&head->refs);
 427         spin_unlock(&delayed_refs->lock);
 428
 429         mutex_lock(&head->mutex);
 430         spin_lock(&delayed_refs->lock);
 431         if (RB_EMPTY_NODE(&head->href_node)) {
 432                 mutex_unlock(&head->mutex);
 433                 btrfs_put_delayed_ref_head(head);
 434                 return -EAGAIN;
 435         }
 436         btrfs_put_delayed_ref_head(head);
 437         return 0;
 438 }
 439
 440 static inline void drop_delayed_ref(struct btrfs_delayed_ref_root *delayed_refs,
 441                                     struct btrfs_delayed_ref_head *head,
 442                                     struct btrfs_delayed_ref_node *ref)
 443 {
 444         lockdep_assert_held(&head->lock);
 445         rb_erase_cached(&ref->ref_node, &head->ref_tree);
 446         RB_CLEAR_NODE(&ref->ref_node);
 447         if (!list_empty(&ref->add_list))
 448                 list_del(&ref->add_list);
 449         ref->in_tree = 0;
 450         btrfs_put_delayed_ref(ref);
 451         atomic_dec(&delayed_refs->num_entries);
 452 }
 453
 454 static bool merge_ref(struct btrfs_delayed_ref_root *delayed_refs,
 455                       struct btrfs_delayed_ref_head *head,
 456                       struct btrfs_delayed_ref_node *ref,
 457                       u64 seq)
 458 {
 459         struct btrfs_delayed_ref_node *next;
 460         struct rb_node *node = rb_next(&ref->ref_node);
 461         bool done = false;
 462
 463         while (!done && node) {
 464                 int mod;
 465
 466                 next = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
 467                 node = rb_next(node);
 468                 if (seq && next->seq >= seq)
 469                         break;
 470                 if (comp_refs(ref, next, false))
 471                         break;
 472
 473                 if (ref->action == next->action) {
 474                         mod = next->ref_mod;
 475                 } else {
 476                         if (ref->ref_mod < next->ref_mod) {
 477                                 swap(ref, next);
 478                                 done = true;
 479                         }
 480                         mod = -next->ref_mod;
 481                 }
 482
 483                 drop_delayed_ref(delayed_refs, head, next);
 484                 ref->ref_mod += mod;
 485                 if (ref->ref_mod == 0) {
 486                         drop_delayed_ref(delayed_refs, head, ref);
 487                         done = true;
 488                 } else {
 489                         /*
 490                          * Can't have multiples of the same ref on a tree block.
 491                          */
 492                         WARN_ON(ref->type == BTRFS_TREE_BLOCK_REF_KEY ||
 493                                 ref->type == BTRFS_SHARED_BLOCK_REF_KEY);
 494                 }
 495         }
 496
 497         return done;
 498 }
 499
 500 void btrfs_merge_delayed_refs(struct btrfs_fs_info *fs_info,
 501                               struct btrfs_delayed_ref_root *delayed_refs,
 502                               struct btrfs_delayed_ref_head *head)
 503 {
 504         struct btrfs_delayed_ref_node *ref;
 505         struct rb_node *node;
 506         u64 seq = 0;
 507
 508         lockdep_assert_held(&head->lock);
 509
 510         if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
 511                 return;
 512
 513         /* We don't have too many refs to merge for data. */
 514         if (head->is_data)
 515                 return;
 516
 517         seq = btrfs_tree_mod_log_lowest_seq(fs_info);
 518 again:
 519         for (node = rb_first_cached(&head->ref_tree); node;
 520              node = rb_next(node)) {
 521                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
 522                 if (seq && ref->seq >= seq)
 523                         continue;
 524                 if (merge_ref(delayed_refs, head, ref, seq))
 525                         goto again;
 526         }
 527 }
 528
 529 int btrfs_check_delayed_seq(struct btrfs_fs_info *fs_info, u64 seq)
 530 {
 531         int ret = 0;
 532         u64 min_seq = btrfs_tree_mod_log_lowest_seq(fs_info);
 533
 534         if (min_seq != 0 && seq >= min_seq) {
 535                 btrfs_debug(fs_info,
 536                             "holding back delayed_ref %llu, lowest is %llu",
 537                             seq, min_seq);
 538                 ret = 1;
 539         }
 540
 541         return ret;
 542 }
 543
 544 struct btrfs_delayed_ref_head *btrfs_select_ref_head(
 545                 struct btrfs_delayed_ref_root *delayed_refs)
 546 {
 547         struct btrfs_delayed_ref_head *head;
 548
 549 again:
 550         head = find_ref_head(delayed_refs, delayed_refs->run_delayed_start,
 551                              true);
 552         if (!head && delayed_refs->run_delayed_start != 0) {
 553                 delayed_refs->run_delayed_start = 0;
 554                 head = find_first_ref_head(delayed_refs);
 555         }
 556         if (!head)
 557                 return NULL;
 558
 559         while (head->processing) {
 560                 struct rb_node *node;
 561
 562                 node = rb_next(&head->href_node);
 563                 if (!node) {
 564                         if (delayed_refs->run_delayed_start == 0)
 565                                 return NULL;
 566                         delayed_refs->run_delayed_start = 0;
 567                         goto again;
 568                 }
 569                 head = rb_entry(node, struct btrfs_delayed_ref_head,
 570                                 href_node);
 571         }
 572
 573         head->processing = 1;
 574         WARN_ON(delayed_refs->num_heads_ready == 0);
 575         delayed_refs->num_heads_ready--;
 576         delayed_refs->run_delayed_start = head->bytenr +
 577                 head->num_bytes;
 578         return head;
 579 }
 580
 581 void btrfs_delete_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
 582                            struct btrfs_delayed_ref_head *head)
 583 {
 584         lockdep_assert_held(&delayed_refs->lock);
 585         lockdep_assert_held(&head->lock);
 586
 587         rb_erase_cached(&head->href_node, &delayed_refs->href_root);
 588         RB_CLEAR_NODE(&head->href_node);
 589         atomic_dec(&delayed_refs->num_entries);
 590         delayed_refs->num_heads--;
 591         if (head->processing == 0)
 592                 delayed_refs->num_heads_ready--;
 593 }
 594
 595 /*
 596  * Helper to insert the ref_node to the tail or merge with tail.
 597  *
 598  * Return 0 for insert.
 599  * Return >0 for merge.
 600  */
 601 static int insert_delayed_ref(struct btrfs_delayed_ref_root *root,
 602                               struct btrfs_delayed_ref_head *href,
 603                               struct btrfs_delayed_ref_node *ref)
 604 {
 605         struct btrfs_delayed_ref_node *exist;
 606         int mod;
 607         int ret = 0;
 608
 609         spin_lock(&href->lock);
 610         exist = tree_insert(&href->ref_tree, ref);
 611         if (!exist)
 612                 goto inserted;
 613
 614         /* Now we are sure we can merge */
 615         ret = 1;
 616         if (exist->action == ref->action) {
 617                 mod = ref->ref_mod;
 618         } else {
 619                 /* Need to change action */
 620                 if (exist->ref_mod < ref->ref_mod) {
 621                         exist->action = ref->action;
 622                         mod = -exist->ref_mod;
 623                         exist->ref_mod = ref->ref_mod;
 624                         if (ref->action == BTRFS_ADD_DELAYED_REF)
 625                                 list_add_tail(&exist->add_list,
 626                                               &href->ref_add_list);
 627                         else if (ref->action == BTRFS_DROP_DELAYED_REF) {
 628                                 ASSERT(!list_empty(&exist->add_list));
 629                                 list_del(&exist->add_list);
 630                         } else {
 631                                 ASSERT(0);
 632                         }
 633                 } else
 634                         mod = -ref->ref_mod;
 635         }
 636         exist->ref_mod += mod;
 637
 638         /* remove existing tail if its ref_mod is zero */
 639         if (exist->ref_mod == 0)
 640                 drop_delayed_ref(root, href, exist);
 641         spin_unlock(&href->lock);
 642         return ret;
 643 inserted:
 644         if (ref->action == BTRFS_ADD_DELAYED_REF)
 645                 list_add_tail(&ref->add_list, &href->ref_add_list);
 646         atomic_inc(&root->num_entries);
 647         spin_unlock(&href->lock);
 648         return ret;
 649 }
 650
 651 /*
 652  * helper function to update the accounting in the head ref
 653  * existing and update must have the same bytenr
 654  */
 655 static noinline void update_existing_head_ref(struct btrfs_trans_handle *trans,
 656                          struct btrfs_delayed_ref_head *existing,
 657                          struct btrfs_delayed_ref_head *update)
 658 {
 659         struct btrfs_delayed_ref_root *delayed_refs =
 660                 &trans->transaction->delayed_refs;
 661         struct btrfs_fs_info *fs_info = trans->fs_info;
 662         int old_ref_mod;
 663
 664         BUG_ON(existing->is_data != update->is_data);
 665
 666         spin_lock(&existing->lock);
 667         if (update->must_insert_reserved) {
 668                 /* if the extent was freed and then
 669                  * reallocated before the delayed ref
 670                  * entries were processed, we can end up
 671                  * with an existing head ref without
 672                  * the must_insert_reserved flag set.
 673                  * Set it again here
 674                  */
 675                 existing->must_insert_reserved = update->must_insert_reserved;
 676
 677                 /*
 678                  * update the num_bytes so we make sure the accounting
 679                  * is done correctly
 680                  */
 681                 existing->num_bytes = update->num_bytes;
 682
 683         }
 684
 685         if (update->extent_op) {
 686                 if (!existing->extent_op) {
 687                         existing->extent_op = update->extent_op;
 688                 } else {
 689                         if (update->extent_op->update_key) {
 690                                 memcpy(&existing->extent_op->key,
 691                                        &update->extent_op->key,
 692                                        sizeof(update->extent_op->key));
 693                                 existing->extent_op->update_key = true;
 694                         }
 695                         if (update->extent_op->update_flags) {
 696                                 existing->extent_op->flags_to_set |=
 697                                         update->extent_op->flags_to_set;
 698                                 existing->extent_op->update_flags = true;
 699                         }
 700                         btrfs_free_delayed_extent_op(update->extent_op);
 701                 }
 702         }
 703         /*
 704          * update the reference mod on the head to reflect this new operation,
 705          * only need the lock for this case cause we could be processing it
 706          * currently, for refs we just added we know we're a-ok.
 707          */
 708         old_ref_mod = existing->total_ref_mod;
 709         existing->ref_mod += update->ref_mod;
 710         existing->total_ref_mod += update->ref_mod;
 711
 712         /*
 713          * If we are going to from a positive ref mod to a negative or vice
 714          * versa we need to make sure to adjust pending_csums accordingly.
 715          */
 716         if (existing->is_data) {
 717                 u64 csum_leaves =
 718                         btrfs_csum_bytes_to_leaves(fs_info,
 719                                                    existing->num_bytes);
 720
 721                 if (existing->total_ref_mod >= 0 && old_ref_mod < 0) {
 722                         delayed_refs->pending_csums -= existing->num_bytes;
 723                         btrfs_delayed_refs_rsv_release(fs_info, csum_leaves);
 724                 }
 725                 if (existing->total_ref_mod < 0 && old_ref_mod >= 0) {
 726                         delayed_refs->pending_csums += existing->num_bytes;
 727                         trans->delayed_ref_updates += csum_leaves;
 728                 }
 729         }
 730
 731         spin_unlock(&existing->lock);
 732 }
 733
 734 static void init_delayed_ref_head(struct btrfs_delayed_ref_head *head_ref,
 735                                   struct btrfs_qgroup_extent_record *qrecord,
 736                                   u64 bytenr, u64 num_bytes, u64 ref_root,
 737                                   u64 reserved, int action, bool is_data,
 738                                   bool is_system)
 739 {
 740         int count_mod = 1;
 741         int must_insert_reserved = 0;
 742
 743         /* If reserved is provided, it must be a data extent. */
 744         BUG_ON(!is_data && reserved);
 745
 746         /*
 747          * The head node stores the sum of all the mods, so dropping a ref
 748          * should drop the sum in the head node by one.
 749          */
 750         if (action == BTRFS_UPDATE_DELAYED_HEAD)
 751                 count_mod = 0;
 752         else if (action == BTRFS_DROP_DELAYED_REF)
 753                 count_mod = -1;
 754
 755         /*
 756          * BTRFS_ADD_DELAYED_EXTENT means that we need to update the reserved
 757          * accounting when the extent is finally added, or if a later
 758          * modification deletes the delayed ref without ever inserting the
 759          * extent into the extent allocation tree.  ref->must_insert_reserved
 760          * is the flag used to record that accounting mods are required.
 761          *
 762          * Once we record must_insert_reserved, switch the action to
 763          * BTRFS_ADD_DELAYED_REF because other special casing is not required.
 764          */
 765         if (action == BTRFS_ADD_DELAYED_EXTENT)
 766                 must_insert_reserved = 1;
 767         else
 768                 must_insert_reserved = 0;
 769
 770         refcount_set(&head_ref->refs, 1);
 771         head_ref->bytenr = bytenr;
 772         head_ref->num_bytes = num_bytes;
 773         head_ref->ref_mod = count_mod;
 774         head_ref->must_insert_reserved = must_insert_reserved;
 775         head_ref->is_data = is_data;
 776         head_ref->is_system = is_system;
 777         head_ref->ref_tree = RB_ROOT_CACHED;
 778         INIT_LIST_HEAD(&head_ref->ref_add_list);
 779         RB_CLEAR_NODE(&head_ref->href_node);
 780         head_ref->processing = 0;
 781         head_ref->total_ref_mod = count_mod;
 782         spin_lock_init(&head_ref->lock);
 783         mutex_init(&head_ref->mutex);
 784
 785         if (qrecord) {
 786                 if (ref_root && reserved) {
 787                         qrecord->data_rsv = reserved;
 788                         qrecord->data_rsv_refroot = ref_root;
 789                 }
 790                 qrecord->bytenr = bytenr;
 791                 qrecord->num_bytes = num_bytes;
 792                 qrecord->old_roots = NULL;
 793         }
 794 }
 795
 796 /*
 797  * helper function to actually insert a head node into the rbtree.
 798  * this does all the dirty work in terms of maintaining the correct
 799  * overall modification count.
 800  */
 801 static noinline struct btrfs_delayed_ref_head *
 802 add_delayed_ref_head(struct btrfs_trans_handle *trans,
 803                      struct btrfs_delayed_ref_head *head_ref,
 804                      struct btrfs_qgroup_extent_record *qrecord,
 805                      int action, int *qrecord_inserted_ret)
 806 {
 807         struct btrfs_delayed_ref_head *existing;
 808         struct btrfs_delayed_ref_root *delayed_refs;
 809         int qrecord_inserted = 0;
 810
 811         delayed_refs = &trans->transaction->delayed_refs;
 812
 813         /* Record qgroup extent info if provided */
 814         if (qrecord) {
 815                 if (btrfs_qgroup_trace_extent_nolock(trans->fs_info,
 816                                         delayed_refs, qrecord))
 817                         kfree(qrecord);
 818                 else
 819                         qrecord_inserted = 1;
 820         }
 821
 822         trace_add_delayed_ref_head(trans->fs_info, head_ref, action);
 823
 824         existing = htree_insert(&delayed_refs->href_root,
 825                                 &head_ref->href_node);
 826         if (existing) {
 827                 update_existing_head_ref(trans, existing, head_ref);
 828                 /*
 829                  * we've updated the existing ref, free the newly
 830                  * allocated ref
 831                  */
 832                 kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
 833                 head_ref = existing;
 834         } else {
 835                 if (head_ref->is_data && head_ref->ref_mod < 0) {
 836                         delayed_refs->pending_csums += head_ref->num_bytes;
 837                         trans->delayed_ref_updates +=
 838                                 btrfs_csum_bytes_to_leaves(trans->fs_info,
 839                                                            head_ref->num_bytes);
 840                 }
 841                 delayed_refs->num_heads++;
 842                 delayed_refs->num_heads_ready++;
 843                 atomic_inc(&delayed_refs->num_entries);
 844                 trans->delayed_ref_updates++;
 845         }
 846         if (qrecord_inserted_ret)
 847                 *qrecord_inserted_ret = qrecord_inserted;
 848
 849         return head_ref;
 850 }
 851
 852 /*
 853  * init_delayed_ref_common - Initialize the structure which represents a
 854  *                           modification to a an extent.
 855  *
 856  * @fs_info:    Internal to the mounted filesystem mount structure.
 857  *
 858  * @ref:        The structure which is going to be initialized.
 859  *
 860  * @bytenr:     The logical address of the extent for which a modification is
 861  *              going to be recorded.
 862  *
 863  * @num_bytes:  Size of the extent whose modification is being recorded.
 864  *
 865  * @ref_root:   The id of the root where this modification has originated, this
 866  *              can be either one of the well-known metadata trees or the
 867  *              subvolume id which references this extent.
 868  *
 869  * @action:     Can be one of BTRFS_ADD_DELAYED_REF/BTRFS_DROP_DELAYED_REF or
 870  *              BTRFS_ADD_DELAYED_EXTENT
 871  *
 872  * @ref_type:   Holds the type of the extent which is being recorded, can be
 873  *              one of BTRFS_SHARED_BLOCK_REF_KEY/BTRFS_TREE_BLOCK_REF_KEY
 874  *              when recording a metadata extent or BTRFS_SHARED_DATA_REF_KEY/
 875  *              BTRFS_EXTENT_DATA_REF_KEY when recording data extent
 876  */
 877 static void init_delayed_ref_common(struct btrfs_fs_info *fs_info,
 878                                     struct btrfs_delayed_ref_node *ref,
 879                                     u64 bytenr, u64 num_bytes, u64 ref_root,
 880                                     int action, u8 ref_type)
 881 {
 882         u64 seq = 0;
 883
 884         if (action == BTRFS_ADD_DELAYED_EXTENT)
 885                 action = BTRFS_ADD_DELAYED_REF;
 886
 887         if (is_fstree(ref_root))
 888                 seq = atomic64_read(&fs_info->tree_mod_seq);
 889
 890         refcount_set(&ref->refs, 1);
 891         ref->bytenr = bytenr;
 892         ref->num_bytes = num_bytes;
 893         ref->ref_mod = 1;
 894         ref->action = action;
 895         ref->is_head = 0;
 896         ref->in_tree = 1;
 897         ref->seq = seq;
 898         ref->type = ref_type;
 899         RB_CLEAR_NODE(&ref->ref_node);
 900         INIT_LIST_HEAD(&ref->add_list);
 901 }
 902
 903 /*
 904  * add a delayed tree ref.  This does all of the accounting required
 905  * to make sure the delayed ref is eventually processed before this
 906  * transaction commits.
 907  */
 908 int btrfs_add_delayed_tree_ref(struct btrfs_trans_handle *trans,
 909                                struct btrfs_ref *generic_ref,
 910                                struct btrfs_delayed_extent_op *extent_op)
 911 {
 912         struct btrfs_fs_info *fs_info = trans->fs_info;
 913         struct btrfs_delayed_tree_ref *ref;
 914         struct btrfs_delayed_ref_head *head_ref;
 915         struct btrfs_delayed_ref_root *delayed_refs;
 916         struct btrfs_qgroup_extent_record *record = NULL;
 917         int qrecord_inserted;
 918         bool is_system;
 919         int action = generic_ref->action;
 920         int level = generic_ref->tree_ref.level;
 921         int ret;
 922         u64 bytenr = generic_ref->bytenr;
 923         u64 num_bytes = generic_ref->len;
 924         u64 parent = generic_ref->parent;
 925         u8 ref_type;
 926
 927         is_system = (generic_ref->tree_ref.owning_root == BTRFS_CHUNK_TREE_OBJECTID);
 928
 929         ASSERT(generic_ref->type == BTRFS_REF_METADATA && generic_ref->action);
 930         ref = kmem_cache_alloc(btrfs_delayed_tree_ref_cachep, GFP_NOFS);
 931         if (!ref)
 932                 return -ENOMEM;
 933
 934         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
 935         if (!head_ref) {
 936                 kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
 937                 return -ENOMEM;
 938         }
 939
 940         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) &&
 941             !generic_ref->skip_qgroup) {
 942                 record = kzalloc(sizeof(*record), GFP_NOFS);
 943                 if (!record) {
 944                         kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
 945                         kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
 946                         return -ENOMEM;
 947                 }
 948         }
 949
 950         if (parent)
 951                 ref_type = BTRFS_SHARED_BLOCK_REF_KEY;
 952         else
 953                 ref_type = BTRFS_TREE_BLOCK_REF_KEY;
 954
 955         init_delayed_ref_common(fs_info, &ref->node, bytenr, num_bytes,
 956                                 generic_ref->tree_ref.owning_root, action,
 957                                 ref_type);
 958         ref->root = generic_ref->tree_ref.owning_root;
 959         ref->parent = parent;
 960         ref->level = level;
 961
 962         init_delayed_ref_head(head_ref, record, bytenr, num_bytes,
 963                               generic_ref->tree_ref.owning_root, 0, action,
 964                               false, is_system);
 965         head_ref->extent_op = extent_op;
 966
 967         delayed_refs = &trans->transaction->delayed_refs;
 968         spin_lock(&delayed_refs->lock);
 969
 970         /*
 971          * insert both the head node and the new ref without dropping
 972          * the spin lock
 973          */
 974         head_ref = add_delayed_ref_head(trans, head_ref, record,
 975                                         action, &qrecord_inserted);
 976
 977         ret = insert_delayed_ref(delayed_refs, head_ref, &ref->node);
 978         spin_unlock(&delayed_refs->lock);
 979
 980         /*
 981          * Need to update the delayed_refs_rsv with any changes we may have
 982          * made.
 983          */
 984         btrfs_update_delayed_refs_rsv(trans);
 985
 986         trace_add_delayed_tree_ref(fs_info, &ref->node, ref,
 987                                    action == BTRFS_ADD_DELAYED_EXTENT ?
 988                                    BTRFS_ADD_DELAYED_REF : action);
 989         if (ret > 0)
 990                 kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
 991
 992         if (qrecord_inserted)
 993                 btrfs_qgroup_trace_extent_post(trans, record);
 994
 995         return 0;
 996 }
 997
 998 /*
 999  * add a delayed data ref. it's similar to btrfs_add_delayed_tree_ref.
1000  */
1001 int btrfs_add_delayed_data_ref(struct btrfs_trans_handle *trans,
1002                                struct btrfs_ref *generic_ref,
1003                                u64 reserved)
1004 {
1005         struct btrfs_fs_info *fs_info = trans->fs_info;
1006         struct btrfs_delayed_data_ref *ref;
1007         struct btrfs_delayed_ref_head *head_ref;
1008         struct btrfs_delayed_ref_root *delayed_refs;
1009         struct btrfs_qgroup_extent_record *record = NULL;
1010         int qrecord_inserted;
1011         int action = generic_ref->action;
1012         int ret;
1013         u64 bytenr = generic_ref->bytenr;
1014         u64 num_bytes = generic_ref->len;
1015         u64 parent = generic_ref->parent;
1016         u64 ref_root = generic_ref->data_ref.owning_root;
1017         u64 owner = generic_ref->data_ref.ino;
1018         u64 offset = generic_ref->data_ref.offset;
1019         u8 ref_type;
1020
1021         ASSERT(generic_ref->type == BTRFS_REF_DATA && action);
1022         ref = kmem_cache_alloc(btrfs_delayed_data_ref_cachep, GFP_NOFS);
1023         if (!ref)
1024                 return -ENOMEM;
1025
1026         if (parent)
1027                 ref_type = BTRFS_SHARED_DATA_REF_KEY;
1028         else
1029                 ref_type = BTRFS_EXTENT_DATA_REF_KEY;
1030         init_delayed_ref_common(fs_info, &ref->node, bytenr, num_bytes,
1031                                 ref_root, action, ref_type);
1032         ref->root = ref_root;
1033         ref->parent = parent;
1034         ref->objectid = owner;
1035         ref->offset = offset;
1036
1037
1038         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
1039         if (!head_ref) {
1040                 kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
1041                 return -ENOMEM;
1042         }
1043
1044         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) &&
1045             !generic_ref->skip_qgroup) {
1046                 record = kzalloc(sizeof(*record), GFP_NOFS);
1047                 if (!record) {
1048                         kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
1049                         kmem_cache_free(btrfs_delayed_ref_head_cachep,
1050                                         head_ref);
1051                         return -ENOMEM;
1052                 }
1053         }
1054
1055         init_delayed_ref_head(head_ref, record, bytenr, num_bytes, ref_root,
1056                               reserved, action, true, false);
1057         head_ref->extent_op = NULL;
1058
1059         delayed_refs = &trans->transaction->delayed_refs;
1060         spin_lock(&delayed_refs->lock);
1061
1062         /*
1063          * insert both the head node and the new ref without dropping
1064          * the spin lock
1065          */
1066         head_ref = add_delayed_ref_head(trans, head_ref, record,
1067                                         action, &qrecord_inserted);
1068
1069         ret = insert_delayed_ref(delayed_refs, head_ref, &ref->node);
1070         spin_unlock(&delayed_refs->lock);
1071
1072         /*
1073          * Need to update the delayed_refs_rsv with any changes we may have
1074          * made.
1075          */
1076         btrfs_update_delayed_refs_rsv(trans);
1077
1078         trace_add_delayed_data_ref(trans->fs_info, &ref->node, ref,
1079                                    action == BTRFS_ADD_DELAYED_EXTENT ?
1080                                    BTRFS_ADD_DELAYED_REF : action);
1081         if (ret > 0)
1082                 kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
1083
1084
1085         if (qrecord_inserted)
1086                 return btrfs_qgroup_trace_extent_post(trans, record);
1087         return 0;
1088 }
1089
1090 int btrfs_add_delayed_extent_op(struct btrfs_trans_handle *trans,
1091                                 u64 bytenr, u64 num_bytes,
1092                                 struct btrfs_delayed_extent_op *extent_op)
1093 {
1094         struct btrfs_delayed_ref_head *head_ref;
1095         struct btrfs_delayed_ref_root *delayed_refs;
1096
1097         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
1098         if (!head_ref)
1099                 return -ENOMEM;
1100
1101         init_delayed_ref_head(head_ref, NULL, bytenr, num_bytes, 0, 0,
1102                               BTRFS_UPDATE_DELAYED_HEAD, false, false);
1103         head_ref->extent_op = extent_op;
1104
1105         delayed_refs = &trans->transaction->delayed_refs;
1106         spin_lock(&delayed_refs->lock);
1107
1108         add_delayed_ref_head(trans, head_ref, NULL, BTRFS_UPDATE_DELAYED_HEAD,
1109                              NULL);
1110
1111         spin_unlock(&delayed_refs->lock);
1112
1113         /*
1114          * Need to update the delayed_refs_rsv with any changes we may have
1115          * made.
1116          */
1117         btrfs_update_delayed_refs_rsv(trans);
1118         return 0;
1119 }
1120
1121 /*
1122  * This does a simple search for the head node for a given extent.  Returns the
1123  * head node if found, or NULL if not.
1124  */
1125 struct btrfs_delayed_ref_head *
1126 btrfs_find_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs, u64 bytenr)
1127 {
1128         lockdep_assert_held(&delayed_refs->lock);
1129
1130         return find_ref_head(delayed_refs, bytenr, false);
1131 }
1132
1133 void __cold btrfs_delayed_ref_exit(void)
1134 {
1135         kmem_cache_destroy(btrfs_delayed_ref_head_cachep);
1136         kmem_cache_destroy(btrfs_delayed_tree_ref_cachep);
1137         kmem_cache_destroy(btrfs_delayed_data_ref_cachep);
1138         kmem_cache_destroy(btrfs_delayed_extent_op_cachep);
1139 }
1140
1141 int __init btrfs_delayed_ref_init(void)
1142 {
1143         btrfs_delayed_ref_head_cachep = kmem_cache_create(
1144                                 "btrfs_delayed_ref_head",
1145                                 sizeof(struct btrfs_delayed_ref_head), 0,
1146                                 SLAB_MEM_SPREAD, NULL);
1147         if (!btrfs_delayed_ref_head_cachep)
1148                 goto fail;
1149
1150         btrfs_delayed_tree_ref_cachep = kmem_cache_create(
1151                                 "btrfs_delayed_tree_ref",
1152                                 sizeof(struct btrfs_delayed_tree_ref), 0,
1153                                 SLAB_MEM_SPREAD, NULL);
1154         if (!btrfs_delayed_tree_ref_cachep)
1155                 goto fail;
1156
1157         btrfs_delayed_data_ref_cachep = kmem_cache_create(
1158                                 "btrfs_delayed_data_ref",
1159                                 sizeof(struct btrfs_delayed_data_ref), 0,
1160                                 SLAB_MEM_SPREAD, NULL);
1161         if (!btrfs_delayed_data_ref_cachep)
1162                 goto fail;
1163
1164         btrfs_delayed_extent_op_cachep = kmem_cache_create(
1165                                 "btrfs_delayed_extent_op",
1166                                 sizeof(struct btrfs_delayed_extent_op), 0,
1167                                 SLAB_MEM_SPREAD, NULL);
1168         if (!btrfs_delayed_extent_op_cachep)
1169                 goto fail;
1170
1171         return 0;
1172 fail:
1173         btrfs_delayed_ref_exit();
1174         return -ENOMEM;
1175 }