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_trans_handle *trans,
 441                                     struct btrfs_delayed_ref_root *delayed_refs,
 442                                     struct btrfs_delayed_ref_head *head,
 443                                     struct btrfs_delayed_ref_node *ref)
 444 {
 445         lockdep_assert_held(&head->lock);
 446         rb_erase_cached(&ref->ref_node, &head->ref_tree);
 447         RB_CLEAR_NODE(&ref->ref_node);
 448         if (!list_empty(&ref->add_list))
 449                 list_del(&ref->add_list);
 450         ref->in_tree = 0;
 451         btrfs_put_delayed_ref(ref);
 452         atomic_dec(&delayed_refs->num_entries);
 453 }
 454
 455 static bool merge_ref(struct btrfs_trans_handle *trans,
 456                       struct btrfs_delayed_ref_root *delayed_refs,
 457                       struct btrfs_delayed_ref_head *head,
 458                       struct btrfs_delayed_ref_node *ref,
 459                       u64 seq)
 460 {
 461         struct btrfs_delayed_ref_node *next;
 462         struct rb_node *node = rb_next(&ref->ref_node);
 463         bool done = false;
 464
 465         while (!done && node) {
 466                 int mod;
 467
 468                 next = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
 469                 node = rb_next(node);
 470                 if (seq && next->seq >= seq)
 471                         break;
 472                 if (comp_refs(ref, next, false))
 473                         break;
 474
 475                 if (ref->action == next->action) {
 476                         mod = next->ref_mod;
 477                 } else {
 478                         if (ref->ref_mod < next->ref_mod) {
 479                                 swap(ref, next);
 480                                 done = true;
 481                         }
 482                         mod = -next->ref_mod;
 483                 }
 484
 485                 drop_delayed_ref(trans, delayed_refs, head, next);
 486                 ref->ref_mod += mod;
 487                 if (ref->ref_mod == 0) {
 488                         drop_delayed_ref(trans, delayed_refs, head, ref);
 489                         done = true;
 490                 } else {
 491                         /*
 492                          * Can't have multiples of the same ref on a tree block.
 493                          */
 494                         WARN_ON(ref->type == BTRFS_TREE_BLOCK_REF_KEY ||
 495                                 ref->type == BTRFS_SHARED_BLOCK_REF_KEY);
 496                 }
 497         }
 498
 499         return done;
 500 }
 501
 502 void btrfs_merge_delayed_refs(struct btrfs_trans_handle *trans,
 503                               struct btrfs_delayed_ref_root *delayed_refs,
 504                               struct btrfs_delayed_ref_head *head)
 505 {
 506         struct btrfs_fs_info *fs_info = trans->fs_info;
 507         struct btrfs_delayed_ref_node *ref;
 508         struct rb_node *node;
 509         u64 seq = 0;
 510
 511         lockdep_assert_held(&head->lock);
 512
 513         if (RB_EMPTY_ROOT(&head->ref_tree.rb_root))
 514                 return;
 515
 516         /* We don't have too many refs to merge for data. */
 517         if (head->is_data)
 518                 return;
 519
 520         seq = btrfs_tree_mod_log_lowest_seq(fs_info);
 521 again:
 522         for (node = rb_first_cached(&head->ref_tree); node;
 523              node = rb_next(node)) {
 524                 ref = rb_entry(node, struct btrfs_delayed_ref_node, ref_node);
 525                 if (seq && ref->seq >= seq)
 526                         continue;
 527                 if (merge_ref(trans, delayed_refs, head, ref, seq))
 528                         goto again;
 529         }
 530 }
 531
 532 int btrfs_check_delayed_seq(struct btrfs_fs_info *fs_info, u64 seq)
 533 {
 534         int ret = 0;
 535         u64 min_seq = btrfs_tree_mod_log_lowest_seq(fs_info);
 536
 537         if (min_seq != 0 && seq >= min_seq) {
 538                 btrfs_debug(fs_info,
 539                             "holding back delayed_ref %llu, lowest is %llu",
 540                             seq, min_seq);
 541                 ret = 1;
 542         }
 543
 544         return ret;
 545 }
 546
 547 struct btrfs_delayed_ref_head *btrfs_select_ref_head(
 548                 struct btrfs_delayed_ref_root *delayed_refs)
 549 {
 550         struct btrfs_delayed_ref_head *head;
 551
 552 again:
 553         head = find_ref_head(delayed_refs, delayed_refs->run_delayed_start,
 554                              true);
 555         if (!head && delayed_refs->run_delayed_start != 0) {
 556                 delayed_refs->run_delayed_start = 0;
 557                 head = find_first_ref_head(delayed_refs);
 558         }
 559         if (!head)
 560                 return NULL;
 561
 562         while (head->processing) {
 563                 struct rb_node *node;
 564
 565                 node = rb_next(&head->href_node);
 566                 if (!node) {
 567                         if (delayed_refs->run_delayed_start == 0)
 568                                 return NULL;
 569                         delayed_refs->run_delayed_start = 0;
 570                         goto again;
 571                 }
 572                 head = rb_entry(node, struct btrfs_delayed_ref_head,
 573                                 href_node);
 574         }
 575
 576         head->processing = 1;
 577         WARN_ON(delayed_refs->num_heads_ready == 0);
 578         delayed_refs->num_heads_ready--;
 579         delayed_refs->run_delayed_start = head->bytenr +
 580                 head->num_bytes;
 581         return head;
 582 }
 583
 584 void btrfs_delete_ref_head(struct btrfs_delayed_ref_root *delayed_refs,
 585                            struct btrfs_delayed_ref_head *head)
 586 {
 587         lockdep_assert_held(&delayed_refs->lock);
 588         lockdep_assert_held(&head->lock);
 589
 590         rb_erase_cached(&head->href_node, &delayed_refs->href_root);
 591         RB_CLEAR_NODE(&head->href_node);
 592         atomic_dec(&delayed_refs->num_entries);
 593         delayed_refs->num_heads--;
 594         if (head->processing == 0)
 595                 delayed_refs->num_heads_ready--;
 596 }
 597
 598 /*
 599  * Helper to insert the ref_node to the tail or merge with tail.
 600  *
 601  * Return 0 for insert.
 602  * Return >0 for merge.
 603  */
 604 static int insert_delayed_ref(struct btrfs_trans_handle *trans,
 605                               struct btrfs_delayed_ref_root *root,
 606                               struct btrfs_delayed_ref_head *href,
 607                               struct btrfs_delayed_ref_node *ref)
 608 {
 609         struct btrfs_delayed_ref_node *exist;
 610         int mod;
 611         int ret = 0;
 612
 613         spin_lock(&href->lock);
 614         exist = tree_insert(&href->ref_tree, ref);
 615         if (!exist)
 616                 goto inserted;
 617
 618         /* Now we are sure we can merge */
 619         ret = 1;
 620         if (exist->action == ref->action) {
 621                 mod = ref->ref_mod;
 622         } else {
 623                 /* Need to change action */
 624                 if (exist->ref_mod < ref->ref_mod) {
 625                         exist->action = ref->action;
 626                         mod = -exist->ref_mod;
 627                         exist->ref_mod = ref->ref_mod;
 628                         if (ref->action == BTRFS_ADD_DELAYED_REF)
 629                                 list_add_tail(&exist->add_list,
 630                                               &href->ref_add_list);
 631                         else if (ref->action == BTRFS_DROP_DELAYED_REF) {
 632                                 ASSERT(!list_empty(&exist->add_list));
 633                                 list_del(&exist->add_list);
 634                         } else {
 635                                 ASSERT(0);
 636                         }
 637                 } else
 638                         mod = -ref->ref_mod;
 639         }
 640         exist->ref_mod += mod;
 641
 642         /* remove existing tail if its ref_mod is zero */
 643         if (exist->ref_mod == 0)
 644                 drop_delayed_ref(trans, root, href, exist);
 645         spin_unlock(&href->lock);
 646         return ret;
 647 inserted:
 648         if (ref->action == BTRFS_ADD_DELAYED_REF)
 649                 list_add_tail(&ref->add_list, &href->ref_add_list);
 650         atomic_inc(&root->num_entries);
 651         spin_unlock(&href->lock);
 652         return ret;
 653 }
 654
 655 /*
 656  * helper function to update the accounting in the head ref
 657  * existing and update must have the same bytenr
 658  */
 659 static noinline void update_existing_head_ref(struct btrfs_trans_handle *trans,
 660                          struct btrfs_delayed_ref_head *existing,
 661                          struct btrfs_delayed_ref_head *update)
 662 {
 663         struct btrfs_delayed_ref_root *delayed_refs =
 664                 &trans->transaction->delayed_refs;
 665         struct btrfs_fs_info *fs_info = trans->fs_info;
 666         int old_ref_mod;
 667
 668         BUG_ON(existing->is_data != update->is_data);
 669
 670         spin_lock(&existing->lock);
 671         if (update->must_insert_reserved) {
 672                 /* if the extent was freed and then
 673                  * reallocated before the delayed ref
 674                  * entries were processed, we can end up
 675                  * with an existing head ref without
 676                  * the must_insert_reserved flag set.
 677                  * Set it again here
 678                  */
 679                 existing->must_insert_reserved = update->must_insert_reserved;
 680
 681                 /*
 682                  * update the num_bytes so we make sure the accounting
 683                  * is done correctly
 684                  */
 685                 existing->num_bytes = update->num_bytes;
 686
 687         }
 688
 689         if (update->extent_op) {
 690                 if (!existing->extent_op) {
 691                         existing->extent_op = update->extent_op;
 692                 } else {
 693                         if (update->extent_op->update_key) {
 694                                 memcpy(&existing->extent_op->key,
 695                                        &update->extent_op->key,
 696                                        sizeof(update->extent_op->key));
 697                                 existing->extent_op->update_key = true;
 698                         }
 699                         if (update->extent_op->update_flags) {
 700                                 existing->extent_op->flags_to_set |=
 701                                         update->extent_op->flags_to_set;
 702                                 existing->extent_op->update_flags = true;
 703                         }
 704                         btrfs_free_delayed_extent_op(update->extent_op);
 705                 }
 706         }
 707         /*
 708          * update the reference mod on the head to reflect this new operation,
 709          * only need the lock for this case cause we could be processing it
 710          * currently, for refs we just added we know we're a-ok.
 711          */
 712         old_ref_mod = existing->total_ref_mod;
 713         existing->ref_mod += update->ref_mod;
 714         existing->total_ref_mod += update->ref_mod;
 715
 716         /*
 717          * If we are going to from a positive ref mod to a negative or vice
 718          * versa we need to make sure to adjust pending_csums accordingly.
 719          */
 720         if (existing->is_data) {
 721                 u64 csum_leaves =
 722                         btrfs_csum_bytes_to_leaves(fs_info,
 723                                                    existing->num_bytes);
 724
 725                 if (existing->total_ref_mod >= 0 && old_ref_mod < 0) {
 726                         delayed_refs->pending_csums -= existing->num_bytes;
 727                         btrfs_delayed_refs_rsv_release(fs_info, csum_leaves);
 728                 }
 729                 if (existing->total_ref_mod < 0 && old_ref_mod >= 0) {
 730                         delayed_refs->pending_csums += existing->num_bytes;
 731                         trans->delayed_ref_updates += csum_leaves;
 732                 }
 733         }
 734
 735         spin_unlock(&existing->lock);
 736 }
 737
 738 static void init_delayed_ref_head(struct btrfs_delayed_ref_head *head_ref,
 739                                   struct btrfs_qgroup_extent_record *qrecord,
 740                                   u64 bytenr, u64 num_bytes, u64 ref_root,
 741                                   u64 reserved, int action, bool is_data,
 742                                   bool is_system)
 743 {
 744         int count_mod = 1;
 745         int must_insert_reserved = 0;
 746
 747         /* If reserved is provided, it must be a data extent. */
 748         BUG_ON(!is_data && reserved);
 749
 750         /*
 751          * The head node stores the sum of all the mods, so dropping a ref
 752          * should drop the sum in the head node by one.
 753          */
 754         if (action == BTRFS_UPDATE_DELAYED_HEAD)
 755                 count_mod = 0;
 756         else if (action == BTRFS_DROP_DELAYED_REF)
 757                 count_mod = -1;
 758
 759         /*
 760          * BTRFS_ADD_DELAYED_EXTENT means that we need to update the reserved
 761          * accounting when the extent is finally added, or if a later
 762          * modification deletes the delayed ref without ever inserting the
 763          * extent into the extent allocation tree.  ref->must_insert_reserved
 764          * is the flag used to record that accounting mods are required.
 765          *
 766          * Once we record must_insert_reserved, switch the action to
 767          * BTRFS_ADD_DELAYED_REF because other special casing is not required.
 768          */
 769         if (action == BTRFS_ADD_DELAYED_EXTENT)
 770                 must_insert_reserved = 1;
 771         else
 772                 must_insert_reserved = 0;
 773
 774         refcount_set(&head_ref->refs, 1);
 775         head_ref->bytenr = bytenr;
 776         head_ref->num_bytes = num_bytes;
 777         head_ref->ref_mod = count_mod;
 778         head_ref->must_insert_reserved = must_insert_reserved;
 779         head_ref->is_data = is_data;
 780         head_ref->is_system = is_system;
 781         head_ref->ref_tree = RB_ROOT_CACHED;
 782         INIT_LIST_HEAD(&head_ref->ref_add_list);
 783         RB_CLEAR_NODE(&head_ref->href_node);
 784         head_ref->processing = 0;
 785         head_ref->total_ref_mod = count_mod;
 786         spin_lock_init(&head_ref->lock);
 787         mutex_init(&head_ref->mutex);
 788
 789         if (qrecord) {
 790                 if (ref_root && reserved) {
 791                         qrecord->data_rsv = reserved;
 792                         qrecord->data_rsv_refroot = ref_root;
 793                 }
 794                 qrecord->bytenr = bytenr;
 795                 qrecord->num_bytes = num_bytes;
 796                 qrecord->old_roots = NULL;
 797         }
 798 }
 799
 800 /*
 801  * helper function to actually insert a head node into the rbtree.
 802  * this does all the dirty work in terms of maintaining the correct
 803  * overall modification count.
 804  */
 805 static noinline struct btrfs_delayed_ref_head *
 806 add_delayed_ref_head(struct btrfs_trans_handle *trans,
 807                      struct btrfs_delayed_ref_head *head_ref,
 808                      struct btrfs_qgroup_extent_record *qrecord,
 809                      int action, int *qrecord_inserted_ret)
 810 {
 811         struct btrfs_delayed_ref_head *existing;
 812         struct btrfs_delayed_ref_root *delayed_refs;
 813         int qrecord_inserted = 0;
 814
 815         delayed_refs = &trans->transaction->delayed_refs;
 816
 817         /* Record qgroup extent info if provided */
 818         if (qrecord) {
 819                 if (btrfs_qgroup_trace_extent_nolock(trans->fs_info,
 820                                         delayed_refs, qrecord))
 821                         kfree(qrecord);
 822                 else
 823                         qrecord_inserted = 1;
 824         }
 825
 826         trace_add_delayed_ref_head(trans->fs_info, head_ref, action);
 827
 828         existing = htree_insert(&delayed_refs->href_root,
 829                                 &head_ref->href_node);
 830         if (existing) {
 831                 update_existing_head_ref(trans, existing, head_ref);
 832                 /*
 833                  * we've updated the existing ref, free the newly
 834                  * allocated ref
 835                  */
 836                 kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
 837                 head_ref = existing;
 838         } else {
 839                 if (head_ref->is_data && head_ref->ref_mod < 0) {
 840                         delayed_refs->pending_csums += head_ref->num_bytes;
 841                         trans->delayed_ref_updates +=
 842                                 btrfs_csum_bytes_to_leaves(trans->fs_info,
 843                                                            head_ref->num_bytes);
 844                 }
 845                 delayed_refs->num_heads++;
 846                 delayed_refs->num_heads_ready++;
 847                 atomic_inc(&delayed_refs->num_entries);
 848                 trans->delayed_ref_updates++;
 849         }
 850         if (qrecord_inserted_ret)
 851                 *qrecord_inserted_ret = qrecord_inserted;
 852
 853         return head_ref;
 854 }
 855
 856 /*
 857  * init_delayed_ref_common - Initialize the structure which represents a
 858  *                           modification to a an extent.
 859  *
 860  * @fs_info:    Internal to the mounted filesystem mount structure.
 861  *
 862  * @ref:        The structure which is going to be initialized.
 863  *
 864  * @bytenr:     The logical address of the extent for which a modification is
 865  *              going to be recorded.
 866  *
 867  * @num_bytes:  Size of the extent whose modification is being recorded.
 868  *
 869  * @ref_root:   The id of the root where this modification has originated, this
 870  *              can be either one of the well-known metadata trees or the
 871  *              subvolume id which references this extent.
 872  *
 873  * @action:     Can be one of BTRFS_ADD_DELAYED_REF/BTRFS_DROP_DELAYED_REF or
 874  *              BTRFS_ADD_DELAYED_EXTENT
 875  *
 876  * @ref_type:   Holds the type of the extent which is being recorded, can be
 877  *              one of BTRFS_SHARED_BLOCK_REF_KEY/BTRFS_TREE_BLOCK_REF_KEY
 878  *              when recording a metadata extent or BTRFS_SHARED_DATA_REF_KEY/
 879  *              BTRFS_EXTENT_DATA_REF_KEY when recording data extent
 880  */
 881 static void init_delayed_ref_common(struct btrfs_fs_info *fs_info,
 882                                     struct btrfs_delayed_ref_node *ref,
 883                                     u64 bytenr, u64 num_bytes, u64 ref_root,
 884                                     int action, u8 ref_type)
 885 {
 886         u64 seq = 0;
 887
 888         if (action == BTRFS_ADD_DELAYED_EXTENT)
 889                 action = BTRFS_ADD_DELAYED_REF;
 890
 891         if (is_fstree(ref_root))
 892                 seq = atomic64_read(&fs_info->tree_mod_seq);
 893
 894         refcount_set(&ref->refs, 1);
 895         ref->bytenr = bytenr;
 896         ref->num_bytes = num_bytes;
 897         ref->ref_mod = 1;
 898         ref->action = action;
 899         ref->is_head = 0;
 900         ref->in_tree = 1;
 901         ref->seq = seq;
 902         ref->type = ref_type;
 903         RB_CLEAR_NODE(&ref->ref_node);
 904         INIT_LIST_HEAD(&ref->add_list);
 905 }
 906
 907 /*
 908  * add a delayed tree ref.  This does all of the accounting required
 909  * to make sure the delayed ref is eventually processed before this
 910  * transaction commits.
 911  */
 912 int btrfs_add_delayed_tree_ref(struct btrfs_trans_handle *trans,
 913                                struct btrfs_ref *generic_ref,
 914                                struct btrfs_delayed_extent_op *extent_op)
 915 {
 916         struct btrfs_fs_info *fs_info = trans->fs_info;
 917         struct btrfs_delayed_tree_ref *ref;
 918         struct btrfs_delayed_ref_head *head_ref;
 919         struct btrfs_delayed_ref_root *delayed_refs;
 920         struct btrfs_qgroup_extent_record *record = NULL;
 921         int qrecord_inserted;
 922         bool is_system;
 923         int action = generic_ref->action;
 924         int level = generic_ref->tree_ref.level;
 925         int ret;
 926         u64 bytenr = generic_ref->bytenr;
 927         u64 num_bytes = generic_ref->len;
 928         u64 parent = generic_ref->parent;
 929         u8 ref_type;
 930
 931         is_system = (generic_ref->tree_ref.owning_root == BTRFS_CHUNK_TREE_OBJECTID);
 932
 933         ASSERT(generic_ref->type == BTRFS_REF_METADATA && generic_ref->action);
 934         ref = kmem_cache_alloc(btrfs_delayed_tree_ref_cachep, GFP_NOFS);
 935         if (!ref)
 936                 return -ENOMEM;
 937
 938         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
 939         if (!head_ref) {
 940                 kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
 941                 return -ENOMEM;
 942         }
 943
 944         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) &&
 945             !generic_ref->skip_qgroup) {
 946                 record = kzalloc(sizeof(*record), GFP_NOFS);
 947                 if (!record) {
 948                         kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
 949                         kmem_cache_free(btrfs_delayed_ref_head_cachep, head_ref);
 950                         return -ENOMEM;
 951                 }
 952         }
 953
 954         if (parent)
 955                 ref_type = BTRFS_SHARED_BLOCK_REF_KEY;
 956         else
 957                 ref_type = BTRFS_TREE_BLOCK_REF_KEY;
 958
 959         init_delayed_ref_common(fs_info, &ref->node, bytenr, num_bytes,
 960                                 generic_ref->tree_ref.owning_root, action,
 961                                 ref_type);
 962         ref->root = generic_ref->tree_ref.owning_root;
 963         ref->parent = parent;
 964         ref->level = level;
 965
 966         init_delayed_ref_head(head_ref, record, bytenr, num_bytes,
 967                               generic_ref->tree_ref.owning_root, 0, action,
 968                               false, is_system);
 969         head_ref->extent_op = extent_op;
 970
 971         delayed_refs = &trans->transaction->delayed_refs;
 972         spin_lock(&delayed_refs->lock);
 973
 974         /*
 975          * insert both the head node and the new ref without dropping
 976          * the spin lock
 977          */
 978         head_ref = add_delayed_ref_head(trans, head_ref, record,
 979                                         action, &qrecord_inserted);
 980
 981         ret = insert_delayed_ref(trans, delayed_refs, head_ref, &ref->node);
 982         spin_unlock(&delayed_refs->lock);
 983
 984         /*
 985          * Need to update the delayed_refs_rsv with any changes we may have
 986          * made.
 987          */
 988         btrfs_update_delayed_refs_rsv(trans);
 989
 990         trace_add_delayed_tree_ref(fs_info, &ref->node, ref,
 991                                    action == BTRFS_ADD_DELAYED_EXTENT ?
 992                                    BTRFS_ADD_DELAYED_REF : action);
 993         if (ret > 0)
 994                 kmem_cache_free(btrfs_delayed_tree_ref_cachep, ref);
 995
 996         if (qrecord_inserted)
 997                 btrfs_qgroup_trace_extent_post(trans, record);
 998
 999         return 0;
1000 }
1001
1002 /*
1003  * add a delayed data ref. it's similar to btrfs_add_delayed_tree_ref.
1004  */
1005 int btrfs_add_delayed_data_ref(struct btrfs_trans_handle *trans,
1006                                struct btrfs_ref *generic_ref,
1007                                u64 reserved)
1008 {
1009         struct btrfs_fs_info *fs_info = trans->fs_info;
1010         struct btrfs_delayed_data_ref *ref;
1011         struct btrfs_delayed_ref_head *head_ref;
1012         struct btrfs_delayed_ref_root *delayed_refs;
1013         struct btrfs_qgroup_extent_record *record = NULL;
1014         int qrecord_inserted;
1015         int action = generic_ref->action;
1016         int ret;
1017         u64 bytenr = generic_ref->bytenr;
1018         u64 num_bytes = generic_ref->len;
1019         u64 parent = generic_ref->parent;
1020         u64 ref_root = generic_ref->data_ref.owning_root;
1021         u64 owner = generic_ref->data_ref.ino;
1022         u64 offset = generic_ref->data_ref.offset;
1023         u8 ref_type;
1024
1025         ASSERT(generic_ref->type == BTRFS_REF_DATA && action);
1026         ref = kmem_cache_alloc(btrfs_delayed_data_ref_cachep, GFP_NOFS);
1027         if (!ref)
1028                 return -ENOMEM;
1029
1030         if (parent)
1031                 ref_type = BTRFS_SHARED_DATA_REF_KEY;
1032         else
1033                 ref_type = BTRFS_EXTENT_DATA_REF_KEY;
1034         init_delayed_ref_common(fs_info, &ref->node, bytenr, num_bytes,
1035                                 ref_root, action, ref_type);
1036         ref->root = ref_root;
1037         ref->parent = parent;
1038         ref->objectid = owner;
1039         ref->offset = offset;
1040
1041
1042         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
1043         if (!head_ref) {
1044                 kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
1045                 return -ENOMEM;
1046         }
1047
1048         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) &&
1049             !generic_ref->skip_qgroup) {
1050                 record = kzalloc(sizeof(*record), GFP_NOFS);
1051                 if (!record) {
1052                         kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
1053                         kmem_cache_free(btrfs_delayed_ref_head_cachep,
1054                                         head_ref);
1055                         return -ENOMEM;
1056                 }
1057         }
1058
1059         init_delayed_ref_head(head_ref, record, bytenr, num_bytes, ref_root,
1060                               reserved, action, true, false);
1061         head_ref->extent_op = NULL;
1062
1063         delayed_refs = &trans->transaction->delayed_refs;
1064         spin_lock(&delayed_refs->lock);
1065
1066         /*
1067          * insert both the head node and the new ref without dropping
1068          * the spin lock
1069          */
1070         head_ref = add_delayed_ref_head(trans, head_ref, record,
1071                                         action, &qrecord_inserted);
1072
1073         ret = insert_delayed_ref(trans, delayed_refs, head_ref, &ref->node);
1074         spin_unlock(&delayed_refs->lock);
1075
1076         /*
1077          * Need to update the delayed_refs_rsv with any changes we may have
1078          * made.
1079          */
1080         btrfs_update_delayed_refs_rsv(trans);
1081
1082         trace_add_delayed_data_ref(trans->fs_info, &ref->node, ref,
1083                                    action == BTRFS_ADD_DELAYED_EXTENT ?
1084                                    BTRFS_ADD_DELAYED_REF : action);
1085         if (ret > 0)
1086                 kmem_cache_free(btrfs_delayed_data_ref_cachep, ref);
1087
1088
1089         if (qrecord_inserted)
1090                 return btrfs_qgroup_trace_extent_post(trans, record);
1091         return 0;
1092 }
1093
1094 int btrfs_add_delayed_extent_op(struct btrfs_trans_handle *trans,
1095                                 u64 bytenr, u64 num_bytes,
1096                                 struct btrfs_delayed_extent_op *extent_op)
1097 {
1098         struct btrfs_delayed_ref_head *head_ref;
1099         struct btrfs_delayed_ref_root *delayed_refs;
1100
1101         head_ref = kmem_cache_alloc(btrfs_delayed_ref_head_cachep, GFP_NOFS);
1102         if (!head_ref)
1103                 return -ENOMEM;
1104
1105         init_delayed_ref_head(head_ref, NULL, bytenr, num_bytes, 0, 0,
1106                               BTRFS_UPDATE_DELAYED_HEAD, false, false);
1107         head_ref->extent_op = extent_op;
1108
1109         delayed_refs = &trans->transaction->delayed_refs;
1110         spin_lock(&delayed_refs->lock);
1111
1112         add_delayed_ref_head(trans, head_ref, NULL, BTRFS_UPDATE_DELAYED_HEAD,
1113                              NULL);
1114
1115         spin_unlock(&delayed_refs->lock);
1116
1117         /*
1118          * Need to update the delayed_refs_rsv with any changes we may have
1119          * made.
1120          */
1121         btrfs_update_delayed_refs_rsv(trans);
1122         return 0;
1123 }
1124
1125 /*
1126  * This does a simple search for the head node for a given extent.  Returns the
1127  * head node if found, or NULL if not.
1128  */
1129 struct btrfs_delayed_ref_head *
1130 btrfs_find_delayed_ref_head(struct btrfs_delayed_ref_root *delayed_refs, u64 bytenr)
1131 {
1132         lockdep_assert_held(&delayed_refs->lock);
1133
1134         return find_ref_head(delayed_refs, bytenr, false);
1135 }
1136
1137 void __cold btrfs_delayed_ref_exit(void)
1138 {
1139         kmem_cache_destroy(btrfs_delayed_ref_head_cachep);
1140         kmem_cache_destroy(btrfs_delayed_tree_ref_cachep);
1141         kmem_cache_destroy(btrfs_delayed_data_ref_cachep);
1142         kmem_cache_destroy(btrfs_delayed_extent_op_cachep);
1143 }
1144
1145 int __init btrfs_delayed_ref_init(void)
1146 {
1147         btrfs_delayed_ref_head_cachep = kmem_cache_create(
1148                                 "btrfs_delayed_ref_head",
1149                                 sizeof(struct btrfs_delayed_ref_head), 0,
1150                                 SLAB_MEM_SPREAD, NULL);
1151         if (!btrfs_delayed_ref_head_cachep)
1152                 goto fail;
1153
1154         btrfs_delayed_tree_ref_cachep = kmem_cache_create(
1155                                 "btrfs_delayed_tree_ref",
1156                                 sizeof(struct btrfs_delayed_tree_ref), 0,
1157                                 SLAB_MEM_SPREAD, NULL);
1158         if (!btrfs_delayed_tree_ref_cachep)
1159                 goto fail;
1160
1161         btrfs_delayed_data_ref_cachep = kmem_cache_create(
1162                                 "btrfs_delayed_data_ref",
1163                                 sizeof(struct btrfs_delayed_data_ref), 0,
1164                                 SLAB_MEM_SPREAD, NULL);
1165         if (!btrfs_delayed_data_ref_cachep)
1166                 goto fail;
1167
1168         btrfs_delayed_extent_op_cachep = kmem_cache_create(
1169                                 "btrfs_delayed_extent_op",
1170                                 sizeof(struct btrfs_delayed_extent_op), 0,
1171                                 SLAB_MEM_SPREAD, NULL);
1172         if (!btrfs_delayed_extent_op_cachep)
1173                 goto fail;
1174
1175         return 0;
1176 fail:
1177         btrfs_delayed_ref_exit();
1178         return -ENOMEM;
1179 }