git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/btree_update_interior.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "alloc_foreground.h"
   5 #include "bkey_methods.h"
   6 #include "btree_cache.h"
   7 #include "btree_gc.h"
   8 #include "btree_update.h"
   9 #include "btree_update_interior.h"
  10 #include "btree_io.h"
  11 #include "btree_iter.h"
  12 #include "btree_locking.h"
  13 #include "buckets.h"
  14 #include "clock.h"
  15 #include "error.h"
  16 #include "extents.h"
  17 #include "journal.h"
  18 #include "journal_reclaim.h"
  19 #include "keylist.h"
  20 #include "recovery.h"
  21 #include "replicas.h"
  22 #include "super-io.h"
  23 #include "trace.h"
  24
  25 #include <linux/random.h>
  26
  27 static int bch2_btree_insert_node(struct btree_update *, struct btree_trans *,
  28                                   struct btree_path *, struct btree *,
  29                                   struct keylist *, unsigned);
  30 static void bch2_btree_update_add_new_node(struct btree_update *, struct btree *);
  31
  32 static struct btree_path *get_unlocked_mut_path(struct btree_trans *trans,
  33                                                 enum btree_id btree_id,
  34                                                 unsigned level,
  35                                                 struct bpos pos)
  36 {
  37         struct btree_path *path;
  38
  39         path = bch2_path_get(trans, btree_id, pos, level + 1, level,
  40                              BTREE_ITER_NOPRESERVE|
  41                              BTREE_ITER_INTENT, _RET_IP_);
  42         path = bch2_btree_path_make_mut(trans, path, true, _RET_IP_);
  43         bch2_btree_path_downgrade(trans, path);
  44         __bch2_btree_path_unlock(trans, path);
  45         return path;
  46 }
  47
  48 /* Debug code: */
  49
  50 /*
  51  * Verify that child nodes correctly span parent node's range:
  52  */
  53 static void btree_node_interior_verify(struct bch_fs *c, struct btree *b)
  54 {
  55 #ifdef CONFIG_BCACHEFS_DEBUG
  56         struct bpos next_node = b->data->min_key;
  57         struct btree_node_iter iter;
  58         struct bkey_s_c k;
  59         struct bkey_s_c_btree_ptr_v2 bp;
  60         struct bkey unpacked;
  61         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
  62
  63         BUG_ON(!b->c.level);
  64
  65         if (!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))
  66                 return;
  67
  68         bch2_btree_node_iter_init_from_start(&iter, b);
  69
  70         while (1) {
  71                 k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked);
  72                 if (k.k->type != KEY_TYPE_btree_ptr_v2)
  73                         break;
  74                 bp = bkey_s_c_to_btree_ptr_v2(k);
  75
  76                 if (!bpos_eq(next_node, bp.v->min_key)) {
  77                         bch2_dump_btree_node(c, b);
  78                         bch2_bpos_to_text(&buf1, next_node);
  79                         bch2_bpos_to_text(&buf2, bp.v->min_key);
  80                         panic("expected next min_key %s got %s\n", buf1.buf, buf2.buf);
  81                 }
  82
  83                 bch2_btree_node_iter_advance(&iter, b);
  84
  85                 if (bch2_btree_node_iter_end(&iter)) {
  86                         if (!bpos_eq(k.k->p, b->key.k.p)) {
  87                                 bch2_dump_btree_node(c, b);
  88                                 bch2_bpos_to_text(&buf1, b->key.k.p);
  89                                 bch2_bpos_to_text(&buf2, k.k->p);
  90                                 panic("expected end %s got %s\n", buf1.buf, buf2.buf);
  91                         }
  92                         break;
  93                 }
  94
  95                 next_node = bpos_successor(k.k->p);
  96         }
  97 #endif
  98 }
  99
 100 /* Calculate ideal packed bkey format for new btree nodes: */
 101
 102 void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b)
 103 {
 104         struct bkey_packed *k;
 105         struct bset_tree *t;
 106         struct bkey uk;
 107
 108         for_each_bset(b, t)
 109                 bset_tree_for_each_key(b, t, k)
 110                         if (!bkey_deleted(k)) {
 111                                 uk = bkey_unpack_key(b, k);
 112                                 bch2_bkey_format_add_key(s, &uk);
 113                         }
 114 }
 115
 116 static struct bkey_format bch2_btree_calc_format(struct btree *b)
 117 {
 118         struct bkey_format_state s;
 119
 120         bch2_bkey_format_init(&s);
 121         bch2_bkey_format_add_pos(&s, b->data->min_key);
 122         bch2_bkey_format_add_pos(&s, b->data->max_key);
 123         __bch2_btree_calc_format(&s, b);
 124
 125         return bch2_bkey_format_done(&s);
 126 }
 127
 128 static size_t btree_node_u64s_with_format(struct btree *b,
 129                                           struct bkey_format *new_f)
 130 {
 131         struct bkey_format *old_f = &b->format;
 132
 133         /* stupid integer promotion rules */
 134         ssize_t delta =
 135             (((int) new_f->key_u64s - old_f->key_u64s) *
 136              (int) b->nr.packed_keys) +
 137             (((int) new_f->key_u64s - BKEY_U64s) *
 138              (int) b->nr.unpacked_keys);
 139
 140         BUG_ON(delta + b->nr.live_u64s < 0);
 141
 142         return b->nr.live_u64s + delta;
 143 }
 144
 145 /**
 146  * btree_node_format_fits - check if we could rewrite node with a new format
 147  *
 148  * This assumes all keys can pack with the new format -- it just checks if
 149  * the re-packed keys would fit inside the node itself.
 150  */
 151 bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
 152                                  struct bkey_format *new_f)
 153 {
 154         size_t u64s = btree_node_u64s_with_format(b, new_f);
 155
 156         return __vstruct_bytes(struct btree_node, u64s) < btree_bytes(c);
 157 }
 158
 159 /* Btree node freeing/allocation: */
 160
 161 static void __btree_node_free(struct bch_fs *c, struct btree *b)
 162 {
 163         trace_and_count(c, btree_node_free, c, b);
 164
 165         BUG_ON(btree_node_write_blocked(b));
 166         BUG_ON(btree_node_dirty(b));
 167         BUG_ON(btree_node_need_write(b));
 168         BUG_ON(b == btree_node_root(c, b));
 169         BUG_ON(b->ob.nr);
 170         BUG_ON(!list_empty(&b->write_blocked));
 171         BUG_ON(b->will_make_reachable);
 172
 173         clear_btree_node_noevict(b);
 174
 175         mutex_lock(&c->btree_cache.lock);
 176         list_move(&b->list, &c->btree_cache.freeable);
 177         mutex_unlock(&c->btree_cache.lock);
 178 }
 179
 180 static void bch2_btree_node_free_inmem(struct btree_trans *trans,
 181                                        struct btree_path *path,
 182                                        struct btree *b)
 183 {
 184         struct bch_fs *c = trans->c;
 185         unsigned level = b->c.level;
 186
 187         bch2_btree_node_lock_write_nofail(trans, path, &b->c);
 188         bch2_btree_node_hash_remove(&c->btree_cache, b);
 189         __btree_node_free(c, b);
 190         six_unlock_write(&b->c.lock);
 191         mark_btree_node_locked_noreset(path, level, SIX_LOCK_intent);
 192
 193         trans_for_each_path(trans, path)
 194                 if (path->l[level].b == b) {
 195                         btree_node_unlock(trans, path, level);
 196                         path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
 197                 }
 198 }
 199
 200 static void bch2_btree_node_free_never_used(struct btree_update *as,
 201                                             struct btree_trans *trans,
 202                                             struct btree *b)
 203 {
 204         struct bch_fs *c = as->c;
 205         struct prealloc_nodes *p = &as->prealloc_nodes[b->c.lock.readers != NULL];
 206         struct btree_path *path;
 207         unsigned level = b->c.level;
 208
 209         BUG_ON(!list_empty(&b->write_blocked));
 210         BUG_ON(b->will_make_reachable != (1UL|(unsigned long) as));
 211
 212         b->will_make_reachable = 0;
 213         closure_put(&as->cl);
 214
 215         clear_btree_node_will_make_reachable(b);
 216         clear_btree_node_accessed(b);
 217         clear_btree_node_dirty_acct(c, b);
 218         clear_btree_node_need_write(b);
 219
 220         mutex_lock(&c->btree_cache.lock);
 221         list_del_init(&b->list);
 222         bch2_btree_node_hash_remove(&c->btree_cache, b);
 223         mutex_unlock(&c->btree_cache.lock);
 224
 225         BUG_ON(p->nr >= ARRAY_SIZE(p->b));
 226         p->b[p->nr++] = b;
 227
 228         six_unlock_intent(&b->c.lock);
 229
 230         trans_for_each_path(trans, path)
 231                 if (path->l[level].b == b) {
 232                         btree_node_unlock(trans, path, level);
 233                         path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
 234                 }
 235 }
 236
 237 static struct btree *__bch2_btree_node_alloc(struct btree_trans *trans,
 238                                              struct disk_reservation *res,
 239                                              struct closure *cl,
 240                                              bool interior_node,
 241                                              unsigned flags)
 242 {
 243         struct bch_fs *c = trans->c;
 244         struct write_point *wp;
 245         struct btree *b;
 246         BKEY_PADDED_ONSTACK(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
 247         struct open_buckets ob = { .nr = 0 };
 248         struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
 249         unsigned nr_reserve;
 250         enum alloc_reserve alloc_reserve;
 251         int ret;
 252
 253         if (flags & BTREE_INSERT_USE_RESERVE) {
 254                 nr_reserve      = 0;
 255                 alloc_reserve   = RESERVE_btree_movinggc;
 256         } else {
 257                 nr_reserve      = BTREE_NODE_RESERVE;
 258                 alloc_reserve   = RESERVE_btree;
 259         }
 260
 261         mutex_lock(&c->btree_reserve_cache_lock);
 262         if (c->btree_reserve_cache_nr > nr_reserve) {
 263                 struct btree_alloc *a =
 264                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
 265
 266                 ob = a->ob;
 267                 bkey_copy(&tmp.k, &a->k);
 268                 mutex_unlock(&c->btree_reserve_cache_lock);
 269                 goto mem_alloc;
 270         }
 271         mutex_unlock(&c->btree_reserve_cache_lock);
 272
 273 retry:
 274         ret = bch2_alloc_sectors_start_trans(trans,
 275                                       c->opts.metadata_target ?:
 276                                       c->opts.foreground_target,
 277                                       0,
 278                                       writepoint_ptr(&c->btree_write_point),
 279                                       &devs_have,
 280                                       res->nr_replicas,
 281                                       c->opts.metadata_replicas_required,
 282                                       alloc_reserve, 0, cl, &wp);
 283         if (unlikely(ret))
 284                 return ERR_PTR(ret);
 285
 286         if (wp->sectors_free < btree_sectors(c)) {
 287                 struct open_bucket *ob;
 288                 unsigned i;
 289
 290                 open_bucket_for_each(c, &wp->ptrs, ob, i)
 291                         if (ob->sectors_free < btree_sectors(c))
 292                                 ob->sectors_free = 0;
 293
 294                 bch2_alloc_sectors_done(c, wp);
 295                 goto retry;
 296         }
 297
 298         bkey_btree_ptr_v2_init(&tmp.k);
 299         bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, btree_sectors(c), false);
 300
 301         bch2_open_bucket_get(c, wp, &ob);
 302         bch2_alloc_sectors_done(c, wp);
 303 mem_alloc:
 304         b = bch2_btree_node_mem_alloc(trans, interior_node);
 305         six_unlock_write(&b->c.lock);
 306         six_unlock_intent(&b->c.lock);
 307
 308         /* we hold cannibalize_lock: */
 309         BUG_ON(IS_ERR(b));
 310         BUG_ON(b->ob.nr);
 311
 312         bkey_copy(&b->key, &tmp.k);
 313         b->ob = ob;
 314
 315         return b;
 316 }
 317
 318 static struct btree *bch2_btree_node_alloc(struct btree_update *as,
 319                                            struct btree_trans *trans,
 320                                            unsigned level)
 321 {
 322         struct bch_fs *c = as->c;
 323         struct btree *b;
 324         struct prealloc_nodes *p = &as->prealloc_nodes[!!level];
 325         int ret;
 326
 327         BUG_ON(level >= BTREE_MAX_DEPTH);
 328         BUG_ON(!p->nr);
 329
 330         b = p->b[--p->nr];
 331
 332         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
 333         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
 334
 335         set_btree_node_accessed(b);
 336         set_btree_node_dirty_acct(c, b);
 337         set_btree_node_need_write(b);
 338
 339         bch2_bset_init_first(b, &b->data->keys);
 340         b->c.level      = level;
 341         b->c.btree_id   = as->btree_id;
 342         b->version_ondisk = c->sb.version;
 343
 344         memset(&b->nr, 0, sizeof(b->nr));
 345         b->data->magic = cpu_to_le64(bset_magic(c));
 346         memset(&b->data->_ptr, 0, sizeof(b->data->_ptr));
 347         b->data->flags = 0;
 348         SET_BTREE_NODE_ID(b->data, as->btree_id);
 349         SET_BTREE_NODE_LEVEL(b->data, level);
 350
 351         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 352                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
 353
 354                 bp->v.mem_ptr           = 0;
 355                 bp->v.seq               = b->data->keys.seq;
 356                 bp->v.sectors_written   = 0;
 357         }
 358
 359         SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
 360
 361         bch2_btree_build_aux_trees(b);
 362
 363         ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
 364         BUG_ON(ret);
 365
 366         trace_and_count(c, btree_node_alloc, c, b);
 367         bch2_increment_clock(c, btree_sectors(c), WRITE);
 368         return b;
 369 }
 370
 371 static void btree_set_min(struct btree *b, struct bpos pos)
 372 {
 373         if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
 374                 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
 375         b->data->min_key = pos;
 376 }
 377
 378 static void btree_set_max(struct btree *b, struct bpos pos)
 379 {
 380         b->key.k.p = pos;
 381         b->data->max_key = pos;
 382 }
 383
 384 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
 385                                                        struct btree_trans *trans,
 386                                                        struct btree *b)
 387 {
 388         struct btree *n = bch2_btree_node_alloc(as, trans, b->c.level);
 389         struct bkey_format format = bch2_btree_calc_format(b);
 390
 391         /*
 392          * The keys might expand with the new format - if they wouldn't fit in
 393          * the btree node anymore, use the old format for now:
 394          */
 395         if (!bch2_btree_node_format_fits(as->c, b, &format))
 396                 format = b->format;
 397
 398         SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
 399
 400         btree_set_min(n, b->data->min_key);
 401         btree_set_max(n, b->data->max_key);
 402
 403         n->data->format         = format;
 404         btree_node_set_format(n, format);
 405
 406         bch2_btree_sort_into(as->c, n, b);
 407
 408         btree_node_reset_sib_u64s(n);
 409         return n;
 410 }
 411
 412 static struct btree *__btree_root_alloc(struct btree_update *as,
 413                                 struct btree_trans *trans, unsigned level)
 414 {
 415         struct btree *b = bch2_btree_node_alloc(as, trans, level);
 416
 417         btree_set_min(b, POS_MIN);
 418         btree_set_max(b, SPOS_MAX);
 419         b->data->format = bch2_btree_calc_format(b);
 420
 421         btree_node_set_format(b, b->data->format);
 422         bch2_btree_build_aux_trees(b);
 423
 424         return b;
 425 }
 426
 427 static void bch2_btree_reserve_put(struct btree_update *as, struct btree_trans *trans)
 428 {
 429         struct bch_fs *c = as->c;
 430         struct prealloc_nodes *p;
 431
 432         for (p = as->prealloc_nodes;
 433              p < as->prealloc_nodes + ARRAY_SIZE(as->prealloc_nodes);
 434              p++) {
 435                 while (p->nr) {
 436                         struct btree *b = p->b[--p->nr];
 437
 438                         mutex_lock(&c->btree_reserve_cache_lock);
 439
 440                         if (c->btree_reserve_cache_nr <
 441                             ARRAY_SIZE(c->btree_reserve_cache)) {
 442                                 struct btree_alloc *a =
 443                                         &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
 444
 445                                 a->ob = b->ob;
 446                                 b->ob.nr = 0;
 447                                 bkey_copy(&a->k, &b->key);
 448                         } else {
 449                                 bch2_open_buckets_put(c, &b->ob);
 450                         }
 451
 452                         mutex_unlock(&c->btree_reserve_cache_lock);
 453
 454                         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
 455                         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
 456                         __btree_node_free(c, b);
 457                         six_unlock_write(&b->c.lock);
 458                         six_unlock_intent(&b->c.lock);
 459                 }
 460         }
 461 }
 462
 463 static int bch2_btree_reserve_get(struct btree_trans *trans,
 464                                   struct btree_update *as,
 465                                   unsigned nr_nodes[2],
 466                                   unsigned flags,
 467                                   struct closure *cl)
 468 {
 469         struct bch_fs *c = as->c;
 470         struct btree *b;
 471         unsigned interior;
 472         int ret = 0;
 473
 474         BUG_ON(nr_nodes[0] + nr_nodes[1] > BTREE_RESERVE_MAX);
 475
 476         /*
 477          * Protects reaping from the btree node cache and using the btree node
 478          * open bucket reserve:
 479          *
 480          * BTREE_INSERT_NOWAIT only applies to btree node allocation, not
 481          * blocking on this lock:
 482          */
 483         ret = bch2_btree_cache_cannibalize_lock(c, cl);
 484         if (ret)
 485                 return ret;
 486
 487         for (interior = 0; interior < 2; interior++) {
 488                 struct prealloc_nodes *p = as->prealloc_nodes + interior;
 489
 490                 while (p->nr < nr_nodes[interior]) {
 491                         b = __bch2_btree_node_alloc(trans, &as->disk_res,
 492                                         flags & BTREE_INSERT_NOWAIT ? NULL : cl,
 493                                         interior, flags);
 494                         if (IS_ERR(b)) {
 495                                 ret = PTR_ERR(b);
 496                                 goto err;
 497                         }
 498
 499                         p->b[p->nr++] = b;
 500                 }
 501         }
 502 err:
 503         bch2_btree_cache_cannibalize_unlock(c);
 504         return ret;
 505 }
 506
 507 /* Asynchronous interior node update machinery */
 508
 509 static void bch2_btree_update_free(struct btree_update *as, struct btree_trans *trans)
 510 {
 511         struct bch_fs *c = as->c;
 512
 513         if (as->took_gc_lock)
 514                 up_read(&c->gc_lock);
 515         as->took_gc_lock = false;
 516
 517         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 518
 519         bch2_journal_pin_drop(&c->journal, &as->journal);
 520         bch2_journal_pin_flush(&c->journal, &as->journal);
 521         bch2_disk_reservation_put(c, &as->disk_res);
 522         bch2_btree_reserve_put(as, trans);
 523
 524         bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_total],
 525                                as->start_time);
 526
 527         mutex_lock(&c->btree_interior_update_lock);
 528         list_del(&as->unwritten_list);
 529         list_del(&as->list);
 530
 531         closure_debug_destroy(&as->cl);
 532         mempool_free(as, &c->btree_interior_update_pool);
 533
 534         /*
 535          * Have to do the wakeup with btree_interior_update_lock still held,
 536          * since being on btree_interior_update_list is our ref on @c:
 537          */
 538         closure_wake_up(&c->btree_interior_update_wait);
 539
 540         mutex_unlock(&c->btree_interior_update_lock);
 541 }
 542
 543 static void btree_update_add_key(struct btree_update *as,
 544                                  struct keylist *keys, struct btree *b)
 545 {
 546         struct bkey_i *k = &b->key;
 547
 548         BUG_ON(bch2_keylist_u64s(keys) + k->k.u64s >
 549                ARRAY_SIZE(as->_old_keys));
 550
 551         bkey_copy(keys->top, k);
 552         bkey_i_to_btree_ptr_v2(keys->top)->v.mem_ptr = b->c.level + 1;
 553
 554         bch2_keylist_push(keys);
 555 }
 556
 557 /*
 558  * The transactional part of an interior btree node update, where we journal the
 559  * update we did to the interior node and update alloc info:
 560  */
 561 static int btree_update_nodes_written_trans(struct btree_trans *trans,
 562                                             struct btree_update *as)
 563 {
 564         struct bkey_i *k;
 565         int ret;
 566
 567         ret = darray_make_room(&trans->extra_journal_entries, as->journal_u64s);
 568         if (ret)
 569                 return ret;
 570
 571         memcpy(&darray_top(trans->extra_journal_entries),
 572                as->journal_entries,
 573                as->journal_u64s * sizeof(u64));
 574         trans->extra_journal_entries.nr += as->journal_u64s;
 575
 576         trans->journal_pin = &as->journal;
 577
 578         for_each_keylist_key(&as->old_keys, k) {
 579                 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
 580
 581                 ret = bch2_trans_mark_old(trans, as->btree_id, level, bkey_i_to_s_c(k), 0);
 582                 if (ret)
 583                         return ret;
 584         }
 585
 586         for_each_keylist_key(&as->new_keys, k) {
 587                 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
 588
 589                 ret = bch2_trans_mark_new(trans, as->btree_id, level, k, 0);
 590                 if (ret)
 591                         return ret;
 592         }
 593
 594         return 0;
 595 }
 596
 597 static void btree_update_nodes_written(struct btree_update *as)
 598 {
 599         struct bch_fs *c = as->c;
 600         struct btree *b;
 601         struct btree_trans trans;
 602         u64 journal_seq = 0;
 603         unsigned i;
 604         int ret;
 605
 606         bch2_trans_init(&trans, c, 0, 512);
 607         /*
 608          * If we're already in an error state, it might be because a btree node
 609          * was never written, and we might be trying to free that same btree
 610          * node here, but it won't have been marked as allocated and we'll see
 611          * spurious disk usage inconsistencies in the transactional part below
 612          * if we don't skip it:
 613          */
 614         ret = bch2_journal_error(&c->journal);
 615         if (ret)
 616                 goto err;
 617
 618         /*
 619          * Wait for any in flight writes to finish before we free the old nodes
 620          * on disk:
 621          */
 622         for (i = 0; i < as->nr_old_nodes; i++) {
 623                 __le64 seq;
 624
 625                 b = as->old_nodes[i];
 626
 627                 btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
 628                 seq = b->data ? b->data->keys.seq : 0;
 629                 six_unlock_read(&b->c.lock);
 630
 631                 if (seq == as->old_nodes_seq[i])
 632                         wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight_inner,
 633                                        TASK_UNINTERRUPTIBLE);
 634         }
 635
 636         /*
 637          * We did an update to a parent node where the pointers we added pointed
 638          * to child nodes that weren't written yet: now, the child nodes have
 639          * been written so we can write out the update to the interior node.
 640          */
 641
 642         /*
 643          * We can't call into journal reclaim here: we'd block on the journal
 644          * reclaim lock, but we may need to release the open buckets we have
 645          * pinned in order for other btree updates to make forward progress, and
 646          * journal reclaim does btree updates when flushing bkey_cached entries,
 647          * which may require allocations as well.
 648          */
 649         ret = commit_do(&trans, &as->disk_res, &journal_seq,
 650                         BTREE_INSERT_NOFAIL|
 651                         BTREE_INSERT_NOCHECK_RW|
 652                         BTREE_INSERT_USE_RESERVE|
 653                         BTREE_INSERT_JOURNAL_RECLAIM|
 654                         JOURNAL_WATERMARK_reserved,
 655                         btree_update_nodes_written_trans(&trans, as));
 656         bch2_trans_unlock(&trans);
 657
 658         bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
 659                              "%s(): error %s", __func__, bch2_err_str(ret));
 660 err:
 661         if (as->b) {
 662                 struct btree_path *path;
 663
 664                 b = as->b;
 665                 path = get_unlocked_mut_path(&trans, as->btree_id, b->c.level, b->key.k.p);
 666                 /*
 667                  * @b is the node we did the final insert into:
 668                  *
 669                  * On failure to get a journal reservation, we still have to
 670                  * unblock the write and allow most of the write path to happen
 671                  * so that shutdown works, but the i->journal_seq mechanism
 672                  * won't work to prevent the btree write from being visible (we
 673                  * didn't get a journal sequence number) - instead
 674                  * __bch2_btree_node_write() doesn't do the actual write if
 675                  * we're in journal error state:
 676                  */
 677
 678                 /*
 679                  * Ensure transaction is unlocked before using
 680                  * btree_node_lock_nopath() (the use of which is always suspect,
 681                  * we need to work on removing this in the future)
 682                  *
 683                  * It should be, but get_unlocked_mut_path() -> bch2_path_get()
 684                  * calls bch2_path_upgrade(), before we call path_make_mut(), so
 685                  * we may rarely end up with a locked path besides the one we
 686                  * have here:
 687                  */
 688                 bch2_trans_unlock(&trans);
 689                 btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_intent);
 690                 mark_btree_node_locked(&trans, path, b->c.level, SIX_LOCK_intent);
 691                 path->l[b->c.level].lock_seq = six_lock_seq(&b->c.lock);
 692                 path->l[b->c.level].b = b;
 693
 694                 bch2_btree_node_lock_write_nofail(&trans, path, &b->c);
 695
 696                 mutex_lock(&c->btree_interior_update_lock);
 697
 698                 list_del(&as->write_blocked_list);
 699                 if (list_empty(&b->write_blocked))
 700                         clear_btree_node_write_blocked(b);
 701
 702                 /*
 703                  * Node might have been freed, recheck under
 704                  * btree_interior_update_lock:
 705                  */
 706                 if (as->b == b) {
 707                         struct bset *i = btree_bset_last(b);
 708
 709                         BUG_ON(!b->c.level);
 710                         BUG_ON(!btree_node_dirty(b));
 711
 712                         if (!ret) {
 713                                 i->journal_seq = cpu_to_le64(
 714                                                              max(journal_seq,
 715                                                                  le64_to_cpu(i->journal_seq)));
 716
 717                                 bch2_btree_add_journal_pin(c, b, journal_seq);
 718                         } else {
 719                                 /*
 720                                  * If we didn't get a journal sequence number we
 721                                  * can't write this btree node, because recovery
 722                                  * won't know to ignore this write:
 723                                  */
 724                                 set_btree_node_never_write(b);
 725                         }
 726                 }
 727
 728                 mutex_unlock(&c->btree_interior_update_lock);
 729
 730                 mark_btree_node_locked_noreset(path, b->c.level, SIX_LOCK_intent);
 731                 six_unlock_write(&b->c.lock);
 732
 733                 btree_node_write_if_need(c, b, SIX_LOCK_intent);
 734                 btree_node_unlock(&trans, path, b->c.level);
 735                 bch2_path_put(&trans, path, true);
 736         }
 737
 738         bch2_journal_pin_drop(&c->journal, &as->journal);
 739
 740         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 741
 742         mutex_lock(&c->btree_interior_update_lock);
 743         for (i = 0; i < as->nr_new_nodes; i++) {
 744                 b = as->new_nodes[i];
 745
 746                 BUG_ON(b->will_make_reachable != (unsigned long) as);
 747                 b->will_make_reachable = 0;
 748                 clear_btree_node_will_make_reachable(b);
 749         }
 750         mutex_unlock(&c->btree_interior_update_lock);
 751
 752         for (i = 0; i < as->nr_new_nodes; i++) {
 753                 b = as->new_nodes[i];
 754
 755                 btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
 756                 btree_node_write_if_need(c, b, SIX_LOCK_read);
 757                 six_unlock_read(&b->c.lock);
 758         }
 759
 760         for (i = 0; i < as->nr_open_buckets; i++)
 761                 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
 762
 763         bch2_btree_update_free(as, &trans);
 764         bch2_trans_exit(&trans);
 765 }
 766
 767 static void btree_interior_update_work(struct work_struct *work)
 768 {
 769         struct bch_fs *c =
 770                 container_of(work, struct bch_fs, btree_interior_update_work);
 771         struct btree_update *as;
 772
 773         while (1) {
 774                 mutex_lock(&c->btree_interior_update_lock);
 775                 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
 776                                               struct btree_update, unwritten_list);
 777                 if (as && !as->nodes_written)
 778                         as = NULL;
 779                 mutex_unlock(&c->btree_interior_update_lock);
 780
 781                 if (!as)
 782                         break;
 783
 784                 btree_update_nodes_written(as);
 785         }
 786 }
 787
 788 static void btree_update_set_nodes_written(struct closure *cl)
 789 {
 790         struct btree_update *as = container_of(cl, struct btree_update, cl);
 791         struct bch_fs *c = as->c;
 792
 793         mutex_lock(&c->btree_interior_update_lock);
 794         as->nodes_written = true;
 795         mutex_unlock(&c->btree_interior_update_lock);
 796
 797         queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
 798 }
 799
 800 /*
 801  * We're updating @b with pointers to nodes that haven't finished writing yet:
 802  * block @b from being written until @as completes
 803  */
 804 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
 805 {
 806         struct bch_fs *c = as->c;
 807
 808         mutex_lock(&c->btree_interior_update_lock);
 809         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 810
 811         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 812         BUG_ON(!btree_node_dirty(b));
 813         BUG_ON(!b->c.level);
 814
 815         as->mode        = BTREE_INTERIOR_UPDATING_NODE;
 816         as->b           = b;
 817
 818         set_btree_node_write_blocked(b);
 819         list_add(&as->write_blocked_list, &b->write_blocked);
 820
 821         mutex_unlock(&c->btree_interior_update_lock);
 822 }
 823
 824 static void btree_update_reparent(struct btree_update *as,
 825                                   struct btree_update *child)
 826 {
 827         struct bch_fs *c = as->c;
 828
 829         lockdep_assert_held(&c->btree_interior_update_lock);
 830
 831         child->b = NULL;
 832         child->mode = BTREE_INTERIOR_UPDATING_AS;
 833
 834         bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL);
 835 }
 836
 837 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
 838 {
 839         struct bkey_i *insert = &b->key;
 840         struct bch_fs *c = as->c;
 841
 842         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 843
 844         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
 845                ARRAY_SIZE(as->journal_entries));
 846
 847         as->journal_u64s +=
 848                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
 849                                   BCH_JSET_ENTRY_btree_root,
 850                                   b->c.btree_id, b->c.level,
 851                                   insert, insert->k.u64s);
 852
 853         mutex_lock(&c->btree_interior_update_lock);
 854         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 855
 856         as->mode        = BTREE_INTERIOR_UPDATING_ROOT;
 857         mutex_unlock(&c->btree_interior_update_lock);
 858 }
 859
 860 /*
 861  * bch2_btree_update_add_new_node:
 862  *
 863  * This causes @as to wait on @b to be written, before it gets to
 864  * bch2_btree_update_nodes_written
 865  *
 866  * Additionally, it sets b->will_make_reachable to prevent any additional writes
 867  * to @b from happening besides the first until @b is reachable on disk
 868  *
 869  * And it adds @b to the list of @as's new nodes, so that we can update sector
 870  * counts in bch2_btree_update_nodes_written:
 871  */
 872 static void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
 873 {
 874         struct bch_fs *c = as->c;
 875
 876         closure_get(&as->cl);
 877
 878         mutex_lock(&c->btree_interior_update_lock);
 879         BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
 880         BUG_ON(b->will_make_reachable);
 881
 882         as->new_nodes[as->nr_new_nodes++] = b;
 883         b->will_make_reachable = 1UL|(unsigned long) as;
 884         set_btree_node_will_make_reachable(b);
 885
 886         mutex_unlock(&c->btree_interior_update_lock);
 887
 888         btree_update_add_key(as, &as->new_keys, b);
 889
 890         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 891                 unsigned bytes = vstruct_end(&b->data->keys) - (void *) b->data;
 892                 unsigned sectors = round_up(bytes, block_bytes(c)) >> 9;
 893
 894                 bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written =
 895                         cpu_to_le16(sectors);
 896         }
 897 }
 898
 899 /*
 900  * returns true if @b was a new node
 901  */
 902 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
 903 {
 904         struct btree_update *as;
 905         unsigned long v;
 906         unsigned i;
 907
 908         mutex_lock(&c->btree_interior_update_lock);
 909         /*
 910          * When b->will_make_reachable != 0, it owns a ref on as->cl that's
 911          * dropped when it gets written by bch2_btree_complete_write - the
 912          * xchg() is for synchronization with bch2_btree_complete_write:
 913          */
 914         v = xchg(&b->will_make_reachable, 0);
 915         clear_btree_node_will_make_reachable(b);
 916         as = (struct btree_update *) (v & ~1UL);
 917
 918         if (!as) {
 919                 mutex_unlock(&c->btree_interior_update_lock);
 920                 return;
 921         }
 922
 923         for (i = 0; i < as->nr_new_nodes; i++)
 924                 if (as->new_nodes[i] == b)
 925                         goto found;
 926
 927         BUG();
 928 found:
 929         array_remove_item(as->new_nodes, as->nr_new_nodes, i);
 930         mutex_unlock(&c->btree_interior_update_lock);
 931
 932         if (v & 1)
 933                 closure_put(&as->cl);
 934 }
 935
 936 static void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
 937 {
 938         while (b->ob.nr)
 939                 as->open_buckets[as->nr_open_buckets++] =
 940                         b->ob.v[--b->ob.nr];
 941 }
 942
 943 /*
 944  * @b is being split/rewritten: it may have pointers to not-yet-written btree
 945  * nodes and thus outstanding btree_updates - redirect @b's
 946  * btree_updates to point to this btree_update:
 947  */
 948 static void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 949                                                       struct btree *b)
 950 {
 951         struct bch_fs *c = as->c;
 952         struct btree_update *p, *n;
 953         struct btree_write *w;
 954
 955         set_btree_node_dying(b);
 956
 957         if (btree_node_fake(b))
 958                 return;
 959
 960         mutex_lock(&c->btree_interior_update_lock);
 961
 962         /*
 963          * Does this node have any btree_update operations preventing
 964          * it from being written?
 965          *
 966          * If so, redirect them to point to this btree_update: we can
 967          * write out our new nodes, but we won't make them visible until those
 968          * operations complete
 969          */
 970         list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
 971                 list_del_init(&p->write_blocked_list);
 972                 btree_update_reparent(as, p);
 973
 974                 /*
 975                  * for flush_held_btree_writes() waiting on updates to flush or
 976                  * nodes to be writeable:
 977                  */
 978                 closure_wake_up(&c->btree_interior_update_wait);
 979         }
 980
 981         clear_btree_node_dirty_acct(c, b);
 982         clear_btree_node_need_write(b);
 983         clear_btree_node_write_blocked(b);
 984
 985         /*
 986          * Does this node have unwritten data that has a pin on the journal?
 987          *
 988          * If so, transfer that pin to the btree_update operation -
 989          * note that if we're freeing multiple nodes, we only need to keep the
 990          * oldest pin of any of the nodes we're freeing. We'll release the pin
 991          * when the new nodes are persistent and reachable on disk:
 992          */
 993         w = btree_current_write(b);
 994         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 995         bch2_journal_pin_drop(&c->journal, &w->journal);
 996
 997         w = btree_prev_write(b);
 998         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 999         bch2_journal_pin_drop(&c->journal, &w->journal);
1000
1001         mutex_unlock(&c->btree_interior_update_lock);
1002
1003         /*
1004          * Is this a node that isn't reachable on disk yet?
1005          *
1006          * Nodes that aren't reachable yet have writes blocked until they're
1007          * reachable - now that we've cancelled any pending writes and moved
1008          * things waiting on that write to wait on this update, we can drop this
1009          * node from the list of nodes that the other update is making
1010          * reachable, prior to freeing it:
1011          */
1012         btree_update_drop_new_node(c, b);
1013
1014         btree_update_add_key(as, &as->old_keys, b);
1015
1016         as->old_nodes[as->nr_old_nodes] = b;
1017         as->old_nodes_seq[as->nr_old_nodes] = b->data->keys.seq;
1018         as->nr_old_nodes++;
1019 }
1020
1021 static void bch2_btree_update_done(struct btree_update *as, struct btree_trans *trans)
1022 {
1023         struct bch_fs *c = as->c;
1024         u64 start_time = as->start_time;
1025
1026         BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
1027
1028         if (as->took_gc_lock)
1029                 up_read(&as->c->gc_lock);
1030         as->took_gc_lock = false;
1031
1032         bch2_btree_reserve_put(as, trans);
1033
1034         continue_at(&as->cl, btree_update_set_nodes_written,
1035                     as->c->btree_interior_update_worker);
1036
1037         bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_foreground],
1038                                start_time);
1039 }
1040
1041 static struct btree_update *
1042 bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
1043                         unsigned level, bool split, unsigned flags)
1044 {
1045         struct bch_fs *c = trans->c;
1046         struct btree_update *as;
1047         u64 start_time = local_clock();
1048         int disk_res_flags = (flags & BTREE_INSERT_NOFAIL)
1049                 ? BCH_DISK_RESERVATION_NOFAIL : 0;
1050         unsigned nr_nodes[2] = { 0, 0 };
1051         unsigned update_level = level;
1052         int journal_flags = flags & JOURNAL_WATERMARK_MASK;
1053         int ret = 0;
1054         u32 restart_count = trans->restart_count;
1055
1056         BUG_ON(!path->should_be_locked);
1057
1058         if (flags & BTREE_INSERT_JOURNAL_RECLAIM)
1059                 journal_flags |= JOURNAL_RES_GET_NONBLOCK;
1060
1061         while (1) {
1062                 nr_nodes[!!update_level] += 1 + split;
1063                 update_level++;
1064
1065                 ret = bch2_btree_path_upgrade(trans, path, update_level + 1);
1066                 if (ret)
1067                         return ERR_PTR(ret);
1068
1069                 if (!btree_path_node(path, update_level)) {
1070                         /* Allocating new root? */
1071                         nr_nodes[1] += split;
1072                         update_level = BTREE_MAX_DEPTH;
1073                         break;
1074                 }
1075
1076                 if (bch2_btree_node_insert_fits(c, path->l[update_level].b,
1077                                         BKEY_BTREE_PTR_U64s_MAX * (1 + split)))
1078                         break;
1079
1080                 split = true;
1081         }
1082
1083         if (flags & BTREE_INSERT_GC_LOCK_HELD)
1084                 lockdep_assert_held(&c->gc_lock);
1085         else if (!down_read_trylock(&c->gc_lock)) {
1086                 ret = drop_locks_do(trans, (down_read(&c->gc_lock), 0));
1087                 if (ret) {
1088                         up_read(&c->gc_lock);
1089                         return ERR_PTR(ret);
1090                 }
1091         }
1092
1093         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOFS);
1094         memset(as, 0, sizeof(*as));
1095         closure_init(&as->cl, NULL);
1096         as->c           = c;
1097         as->start_time  = start_time;
1098         as->mode        = BTREE_INTERIOR_NO_UPDATE;
1099         as->took_gc_lock = !(flags & BTREE_INSERT_GC_LOCK_HELD);
1100         as->btree_id    = path->btree_id;
1101         as->update_level = update_level;
1102         INIT_LIST_HEAD(&as->list);
1103         INIT_LIST_HEAD(&as->unwritten_list);
1104         INIT_LIST_HEAD(&as->write_blocked_list);
1105         bch2_keylist_init(&as->old_keys, as->_old_keys);
1106         bch2_keylist_init(&as->new_keys, as->_new_keys);
1107         bch2_keylist_init(&as->parent_keys, as->inline_keys);
1108
1109         mutex_lock(&c->btree_interior_update_lock);
1110         list_add_tail(&as->list, &c->btree_interior_update_list);
1111         mutex_unlock(&c->btree_interior_update_lock);
1112
1113         /*
1114          * We don't want to allocate if we're in an error state, that can cause
1115          * deadlock on emergency shutdown due to open buckets getting stuck in
1116          * the btree_reserve_cache after allocator shutdown has cleared it out.
1117          * This check needs to come after adding us to the btree_interior_update
1118          * list but before calling bch2_btree_reserve_get, to synchronize with
1119          * __bch2_fs_read_only().
1120          */
1121         ret = bch2_journal_error(&c->journal);
1122         if (ret)
1123                 goto err;
1124
1125         ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
1126                                       BTREE_UPDATE_JOURNAL_RES,
1127                                       journal_flags|JOURNAL_RES_GET_NONBLOCK);
1128         if (ret) {
1129                 if (flags & BTREE_INSERT_JOURNAL_RECLAIM) {
1130                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
1131                         goto err;
1132                 }
1133
1134                 ret = drop_locks_do(trans,
1135                         bch2_journal_preres_get(&c->journal, &as->journal_preres,
1136                                               BTREE_UPDATE_JOURNAL_RES,
1137                                               journal_flags));
1138                 if (ret == -BCH_ERR_journal_preres_get_blocked) {
1139                         trace_and_count(c, trans_restart_journal_preres_get, trans, _RET_IP_, journal_flags);
1140                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_journal_preres_get);
1141                 }
1142                 if (ret)
1143                         goto err;
1144         }
1145
1146         ret = bch2_disk_reservation_get(c, &as->disk_res,
1147                         (nr_nodes[0] + nr_nodes[1]) * btree_sectors(c),
1148                         c->opts.metadata_replicas,
1149                         disk_res_flags);
1150         if (ret)
1151                 goto err;
1152
1153         ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, NULL);
1154         if (bch2_err_matches(ret, ENOSPC) ||
1155             bch2_err_matches(ret, ENOMEM)) {
1156                 struct closure cl;
1157
1158                 closure_init_stack(&cl);
1159
1160                 do {
1161                         ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, &cl);
1162
1163                         bch2_trans_unlock(trans);
1164                         closure_sync(&cl);
1165                 } while (bch2_err_matches(ret, BCH_ERR_operation_blocked));
1166         }
1167
1168         if (ret) {
1169                 trace_and_count(c, btree_reserve_get_fail, trans->fn,
1170                                 _RET_IP_, nr_nodes[0] + nr_nodes[1], ret);
1171                 goto err;
1172         }
1173
1174         ret = bch2_trans_relock(trans);
1175         if (ret)
1176                 goto err;
1177
1178         bch2_trans_verify_not_restarted(trans, restart_count);
1179         return as;
1180 err:
1181         bch2_btree_update_free(as, trans);
1182         return ERR_PTR(ret);
1183 }
1184
1185 /* Btree root updates: */
1186
1187 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1188 {
1189         /* Root nodes cannot be reaped */
1190         mutex_lock(&c->btree_cache.lock);
1191         list_del_init(&b->list);
1192         mutex_unlock(&c->btree_cache.lock);
1193
1194         mutex_lock(&c->btree_root_lock);
1195         BUG_ON(btree_node_root(c, b) &&
1196                (b->c.level < btree_node_root(c, b)->c.level ||
1197                 !btree_node_dying(btree_node_root(c, b))));
1198
1199         btree_node_root(c, b) = b;
1200         mutex_unlock(&c->btree_root_lock);
1201
1202         bch2_recalc_btree_reserve(c);
1203 }
1204
1205 /**
1206  * bch_btree_set_root - update the root in memory and on disk
1207  *
1208  * To ensure forward progress, the current task must not be holding any
1209  * btree node write locks. However, you must hold an intent lock on the
1210  * old root.
1211  *
1212  * Note: This allocates a journal entry but doesn't add any keys to
1213  * it.  All the btree roots are part of every journal write, so there
1214  * is nothing new to be done.  This just guarantees that there is a
1215  * journal write.
1216  */
1217 static void bch2_btree_set_root(struct btree_update *as,
1218                                 struct btree_trans *trans,
1219                                 struct btree_path *path,
1220                                 struct btree *b)
1221 {
1222         struct bch_fs *c = as->c;
1223         struct btree *old;
1224
1225         trace_and_count(c, btree_node_set_root, c, b);
1226
1227         old = btree_node_root(c, b);
1228
1229         /*
1230          * Ensure no one is using the old root while we switch to the
1231          * new root:
1232          */
1233         bch2_btree_node_lock_write_nofail(trans, path, &old->c);
1234
1235         bch2_btree_set_root_inmem(c, b);
1236
1237         btree_update_updated_root(as, b);
1238
1239         /*
1240          * Unlock old root after new root is visible:
1241          *
1242          * The new root isn't persistent, but that's ok: we still have
1243          * an intent lock on the new root, and any updates that would
1244          * depend on the new root would have to update the new root.
1245          */
1246         bch2_btree_node_unlock_write(trans, path, old);
1247 }
1248
1249 /* Interior node updates: */
1250
1251 static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
1252                                         struct btree_trans *trans,
1253                                         struct btree_path *path,
1254                                         struct btree *b,
1255                                         struct btree_node_iter *node_iter,
1256                                         struct bkey_i *insert)
1257 {
1258         struct bch_fs *c = as->c;
1259         struct bkey_packed *k;
1260         struct printbuf buf = PRINTBUF;
1261         unsigned long old, new, v;
1262
1263         BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 &&
1264                !btree_ptr_sectors_written(insert));
1265
1266         if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
1267                 bch2_journal_key_overwritten(c, b->c.btree_id, b->c.level, insert->k.p);
1268
1269         if (bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1270                               btree_node_type(b), WRITE, &buf) ?:
1271             bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert), &buf)) {
1272                 printbuf_reset(&buf);
1273                 prt_printf(&buf, "inserting invalid bkey\n  ");
1274                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
1275                 prt_printf(&buf, "\n  ");
1276                 bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1277                                   btree_node_type(b), WRITE, &buf);
1278                 bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert), &buf);
1279
1280                 bch2_fs_inconsistent(c, "%s", buf.buf);
1281                 dump_stack();
1282         }
1283
1284         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1285                ARRAY_SIZE(as->journal_entries));
1286
1287         as->journal_u64s +=
1288                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1289                                   BCH_JSET_ENTRY_btree_keys,
1290                                   b->c.btree_id, b->c.level,
1291                                   insert, insert->k.u64s);
1292
1293         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1294                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1295                 bch2_btree_node_iter_advance(node_iter, b);
1296
1297         bch2_btree_bset_insert_key(trans, path, b, node_iter, insert);
1298         set_btree_node_dirty_acct(c, b);
1299
1300         v = READ_ONCE(b->flags);
1301         do {
1302                 old = new = v;
1303
1304                 new &= ~BTREE_WRITE_TYPE_MASK;
1305                 new |= BTREE_WRITE_interior;
1306                 new |= 1 << BTREE_NODE_need_write;
1307         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1308
1309         printbuf_exit(&buf);
1310 }
1311
1312 static void
1313 __bch2_btree_insert_keys_interior(struct btree_update *as,
1314                                   struct btree_trans *trans,
1315                                   struct btree_path *path,
1316                                   struct btree *b,
1317                                   struct btree_node_iter node_iter,
1318                                   struct keylist *keys)
1319 {
1320         struct bkey_i *insert = bch2_keylist_front(keys);
1321         struct bkey_packed *k;
1322
1323         BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1324
1325         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1326                (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1327                 ;
1328
1329         while (!bch2_keylist_empty(keys)) {
1330                 struct bkey_i *k = bch2_keylist_front(keys);
1331
1332                 if (bpos_gt(k->k.p, b->key.k.p))
1333                         break;
1334
1335                 bch2_insert_fixup_btree_ptr(as, trans, path, b, &node_iter, k);
1336                 bch2_keylist_pop_front(keys);
1337         }
1338 }
1339
1340 /*
1341  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1342  * node)
1343  */
1344 static void __btree_split_node(struct btree_update *as,
1345                                struct btree_trans *trans,
1346                                struct btree *b,
1347                                struct btree *n[2])
1348 {
1349         struct bkey_packed *k;
1350         struct bpos n1_pos = POS_MIN;
1351         struct btree_node_iter iter;
1352         struct bset *bsets[2];
1353         struct bkey_format_state format[2];
1354         struct bkey_packed *out[2];
1355         struct bkey uk;
1356         unsigned u64s, n1_u64s = (b->nr.live_u64s * 3) / 5;
1357         int i;
1358
1359         for (i = 0; i < 2; i++) {
1360                 BUG_ON(n[i]->nsets != 1);
1361
1362                 bsets[i] = btree_bset_first(n[i]);
1363                 out[i] = bsets[i]->start;
1364
1365                 SET_BTREE_NODE_SEQ(n[i]->data, BTREE_NODE_SEQ(b->data) + 1);
1366                 bch2_bkey_format_init(&format[i]);
1367         }
1368
1369         u64s = 0;
1370         for_each_btree_node_key(b, k, &iter) {
1371                 if (bkey_deleted(k))
1372                         continue;
1373
1374                 i = u64s >= n1_u64s;
1375                 u64s += k->u64s;
1376                 uk = bkey_unpack_key(b, k);
1377                 if (!i)
1378                         n1_pos = uk.p;
1379                 bch2_bkey_format_add_key(&format[i], &uk);
1380         }
1381
1382         btree_set_min(n[0], b->data->min_key);
1383         btree_set_max(n[0], n1_pos);
1384         btree_set_min(n[1], bpos_successor(n1_pos));
1385         btree_set_max(n[1], b->data->max_key);
1386
1387         for (i = 0; i < 2; i++) {
1388                 bch2_bkey_format_add_pos(&format[i], n[i]->data->min_key);
1389                 bch2_bkey_format_add_pos(&format[i], n[i]->data->max_key);
1390
1391                 n[i]->data->format = bch2_bkey_format_done(&format[i]);
1392                 btree_node_set_format(n[i], n[i]->data->format);
1393         }
1394
1395         u64s = 0;
1396         for_each_btree_node_key(b, k, &iter) {
1397                 if (bkey_deleted(k))
1398                         continue;
1399
1400                 i = u64s >= n1_u64s;
1401                 u64s += k->u64s;
1402
1403                 if (bch2_bkey_transform(&n[i]->format, out[i], bkey_packed(k)
1404                                         ? &b->format: &bch2_bkey_format_current, k))
1405                         out[i]->format = KEY_FORMAT_LOCAL_BTREE;
1406                 else
1407                         bch2_bkey_unpack(b, (void *) out[i], k);
1408
1409                 out[i]->needs_whiteout = false;
1410
1411                 btree_keys_account_key_add(&n[i]->nr, 0, out[i]);
1412                 out[i] = bkey_p_next(out[i]);
1413         }
1414
1415         for (i = 0; i < 2; i++) {
1416                 bsets[i]->u64s = cpu_to_le16((u64 *) out[i] - bsets[i]->_data);
1417
1418                 BUG_ON(!bsets[i]->u64s);
1419
1420                 set_btree_bset_end(n[i], n[i]->set);
1421
1422                 btree_node_reset_sib_u64s(n[i]);
1423
1424                 bch2_verify_btree_nr_keys(n[i]);
1425
1426                 if (b->c.level)
1427                         btree_node_interior_verify(as->c, n[i]);
1428         }
1429 }
1430
1431 /*
1432  * For updates to interior nodes, we've got to do the insert before we split
1433  * because the stuff we're inserting has to be inserted atomically. Post split,
1434  * the keys might have to go in different nodes and the split would no longer be
1435  * atomic.
1436  *
1437  * Worse, if the insert is from btree node coalescing, if we do the insert after
1438  * we do the split (and pick the pivot) - the pivot we pick might be between
1439  * nodes that were coalesced, and thus in the middle of a child node post
1440  * coalescing:
1441  */
1442 static void btree_split_insert_keys(struct btree_update *as,
1443                                     struct btree_trans *trans,
1444                                     struct btree_path *path,
1445                                     struct btree *b,
1446                                     struct keylist *keys)
1447 {
1448         if (!bch2_keylist_empty(keys) &&
1449             bpos_le(bch2_keylist_front(keys)->k.p, b->data->max_key)) {
1450                 struct btree_node_iter node_iter;
1451
1452                 bch2_btree_node_iter_init(&node_iter, b, &bch2_keylist_front(keys)->k.p);
1453
1454                 __bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
1455
1456                 btree_node_interior_verify(as->c, b);
1457         }
1458 }
1459
1460 static int btree_split(struct btree_update *as, struct btree_trans *trans,
1461                        struct btree_path *path, struct btree *b,
1462                        struct keylist *keys, unsigned flags)
1463 {
1464         struct bch_fs *c = as->c;
1465         struct btree *parent = btree_node_parent(path, b);
1466         struct btree *n1, *n2 = NULL, *n3 = NULL;
1467         struct btree_path *path1 = NULL, *path2 = NULL;
1468         u64 start_time = local_clock();
1469         int ret = 0;
1470
1471         BUG_ON(!parent && (b != btree_node_root(c, b)));
1472         BUG_ON(parent && !btree_node_intent_locked(path, b->c.level + 1));
1473
1474         bch2_btree_interior_update_will_free_node(as, b);
1475
1476         if (b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c)) {
1477                 struct btree *n[2];
1478
1479                 trace_and_count(c, btree_node_split, c, b);
1480
1481                 n[0] = n1 = bch2_btree_node_alloc(as, trans, b->c.level);
1482                 n[1] = n2 = bch2_btree_node_alloc(as, trans, b->c.level);
1483
1484                 __btree_split_node(as, trans, b, n);
1485
1486                 if (keys) {
1487                         btree_split_insert_keys(as, trans, path, n1, keys);
1488                         btree_split_insert_keys(as, trans, path, n2, keys);
1489                         BUG_ON(!bch2_keylist_empty(keys));
1490                 }
1491
1492                 bch2_btree_build_aux_trees(n2);
1493                 bch2_btree_build_aux_trees(n1);
1494
1495                 bch2_btree_update_add_new_node(as, n1);
1496                 bch2_btree_update_add_new_node(as, n2);
1497                 six_unlock_write(&n2->c.lock);
1498                 six_unlock_write(&n1->c.lock);
1499
1500                 path1 = get_unlocked_mut_path(trans, path->btree_id, n1->c.level, n1->key.k.p);
1501                 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1502                 mark_btree_node_locked(trans, path1, n1->c.level, SIX_LOCK_intent);
1503                 bch2_btree_path_level_init(trans, path1, n1);
1504
1505                 path2 = get_unlocked_mut_path(trans, path->btree_id, n2->c.level, n2->key.k.p);
1506                 six_lock_increment(&n2->c.lock, SIX_LOCK_intent);
1507                 mark_btree_node_locked(trans, path2, n2->c.level, SIX_LOCK_intent);
1508                 bch2_btree_path_level_init(trans, path2, n2);
1509
1510                 /*
1511                  * Note that on recursive parent_keys == keys, so we
1512                  * can't start adding new keys to parent_keys before emptying it
1513                  * out (which we did with btree_split_insert_keys() above)
1514                  */
1515                 bch2_keylist_add(&as->parent_keys, &n1->key);
1516                 bch2_keylist_add(&as->parent_keys, &n2->key);
1517
1518                 if (!parent) {
1519                         /* Depth increases, make a new root */
1520                         n3 = __btree_root_alloc(as, trans, b->c.level + 1);
1521
1522                         bch2_btree_update_add_new_node(as, n3);
1523                         six_unlock_write(&n3->c.lock);
1524
1525                         path2->locks_want++;
1526                         BUG_ON(btree_node_locked(path2, n3->c.level));
1527                         six_lock_increment(&n3->c.lock, SIX_LOCK_intent);
1528                         mark_btree_node_locked(trans, path2, n3->c.level, SIX_LOCK_intent);
1529                         bch2_btree_path_level_init(trans, path2, n3);
1530
1531                         n3->sib_u64s[0] = U16_MAX;
1532                         n3->sib_u64s[1] = U16_MAX;
1533
1534                         btree_split_insert_keys(as, trans, path, n3, &as->parent_keys);
1535                 }
1536         } else {
1537                 trace_and_count(c, btree_node_compact, c, b);
1538
1539                 n1 = bch2_btree_node_alloc_replacement(as, trans, b);
1540
1541                 if (keys) {
1542                         btree_split_insert_keys(as, trans, path, n1, keys);
1543                         BUG_ON(!bch2_keylist_empty(keys));
1544                 }
1545
1546                 bch2_btree_build_aux_trees(n1);
1547                 bch2_btree_update_add_new_node(as, n1);
1548                 six_unlock_write(&n1->c.lock);
1549
1550                 path1 = get_unlocked_mut_path(trans, path->btree_id, n1->c.level, n1->key.k.p);
1551                 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1552                 mark_btree_node_locked(trans, path1, n1->c.level, SIX_LOCK_intent);
1553                 bch2_btree_path_level_init(trans, path1, n1);
1554
1555                 if (parent)
1556                         bch2_keylist_add(&as->parent_keys, &n1->key);
1557         }
1558
1559         /* New nodes all written, now make them visible: */
1560
1561         if (parent) {
1562                 /* Split a non root node */
1563                 ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1564                 if (ret)
1565                         goto err;
1566         } else if (n3) {
1567                 bch2_btree_set_root(as, trans, path, n3);
1568         } else {
1569                 /* Root filled up but didn't need to be split */
1570                 bch2_btree_set_root(as, trans, path, n1);
1571         }
1572
1573         if (n3) {
1574                 bch2_btree_update_get_open_buckets(as, n3);
1575                 bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0);
1576         }
1577         if (n2) {
1578                 bch2_btree_update_get_open_buckets(as, n2);
1579                 bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0);
1580         }
1581         bch2_btree_update_get_open_buckets(as, n1);
1582         bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1583
1584         /*
1585          * The old node must be freed (in memory) _before_ unlocking the new
1586          * nodes - else another thread could re-acquire a read lock on the old
1587          * node after another thread has locked and updated the new node, thus
1588          * seeing stale data:
1589          */
1590         bch2_btree_node_free_inmem(trans, path, b);
1591
1592         if (n3)
1593                 bch2_trans_node_add(trans, n3);
1594         if (n2)
1595                 bch2_trans_node_add(trans, n2);
1596         bch2_trans_node_add(trans, n1);
1597
1598         if (n3)
1599                 six_unlock_intent(&n3->c.lock);
1600         if (n2)
1601                 six_unlock_intent(&n2->c.lock);
1602         six_unlock_intent(&n1->c.lock);
1603 out:
1604         if (path2) {
1605                 __bch2_btree_path_unlock(trans, path2);
1606                 bch2_path_put(trans, path2, true);
1607         }
1608         if (path1) {
1609                 __bch2_btree_path_unlock(trans, path1);
1610                 bch2_path_put(trans, path1, true);
1611         }
1612
1613         bch2_trans_verify_locks(trans);
1614
1615         bch2_time_stats_update(&c->times[n2
1616                                ? BCH_TIME_btree_node_split
1617                                : BCH_TIME_btree_node_compact],
1618                                start_time);
1619         return ret;
1620 err:
1621         if (n3)
1622                 bch2_btree_node_free_never_used(as, trans, n3);
1623         if (n2)
1624                 bch2_btree_node_free_never_used(as, trans, n2);
1625         bch2_btree_node_free_never_used(as, trans, n1);
1626         goto out;
1627 }
1628
1629 static void
1630 bch2_btree_insert_keys_interior(struct btree_update *as,
1631                                 struct btree_trans *trans,
1632                                 struct btree_path *path,
1633                                 struct btree *b,
1634                                 struct keylist *keys)
1635 {
1636         struct btree_path *linked;
1637
1638         __bch2_btree_insert_keys_interior(as, trans, path, b,
1639                                           path->l[b->c.level].iter, keys);
1640
1641         btree_update_updated_node(as, b);
1642
1643         trans_for_each_path_with_node(trans, b, linked)
1644                 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1645
1646         bch2_trans_verify_paths(trans);
1647 }
1648
1649 /**
1650  * bch_btree_insert_node - insert bkeys into a given btree node
1651  *
1652  * @iter:               btree iterator
1653  * @keys:               list of keys to insert
1654  * @hook:               insert callback
1655  * @persistent:         if not null, @persistent will wait on journal write
1656  *
1657  * Inserts as many keys as it can into a given btree node, splitting it if full.
1658  * If a split occurred, this function will return early. This can only happen
1659  * for leaf nodes -- inserts into interior nodes have to be atomic.
1660  */
1661 static int bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans,
1662                                   struct btree_path *path, struct btree *b,
1663                                   struct keylist *keys, unsigned flags)
1664 {
1665         struct bch_fs *c = as->c;
1666         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1667         int old_live_u64s = b->nr.live_u64s;
1668         int live_u64s_added, u64s_added;
1669         int ret;
1670
1671         lockdep_assert_held(&c->gc_lock);
1672         BUG_ON(!btree_node_intent_locked(path, b->c.level));
1673         BUG_ON(!b->c.level);
1674         BUG_ON(!as || as->b);
1675         bch2_verify_keylist_sorted(keys);
1676
1677         if ((local_clock() & 63) == 63)
1678                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_split_race);
1679
1680         ret = bch2_btree_node_lock_write(trans, path, &b->c);
1681         if (ret)
1682                 return ret;
1683
1684         bch2_btree_node_prep_for_write(trans, path, b);
1685
1686         if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1687                 bch2_btree_node_unlock_write(trans, path, b);
1688                 goto split;
1689         }
1690
1691         btree_node_interior_verify(c, b);
1692
1693         bch2_btree_insert_keys_interior(as, trans, path, b, keys);
1694
1695         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1696         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1697
1698         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1699                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1700         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1701                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1702
1703         if (u64s_added > live_u64s_added &&
1704             bch2_maybe_compact_whiteouts(c, b))
1705                 bch2_trans_node_reinit_iter(trans, b);
1706
1707         bch2_btree_node_unlock_write(trans, path, b);
1708
1709         btree_node_interior_verify(c, b);
1710         return 0;
1711 split:
1712         /*
1713          * We could attempt to avoid the transaction restart, by calling
1714          * bch2_btree_path_upgrade() and allocating more nodes:
1715          */
1716         if (b->c.level >= as->update_level) {
1717                 trace_and_count(c, trans_restart_split_race, trans, _THIS_IP_, b);
1718                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_split_race);
1719         }
1720
1721         return btree_split(as, trans, path, b, keys, flags);
1722 }
1723
1724 int bch2_btree_split_leaf(struct btree_trans *trans,
1725                           struct btree_path *path,
1726                           unsigned flags)
1727 {
1728         struct btree *b = path_l(path)->b;
1729         struct btree_update *as;
1730         unsigned l;
1731         int ret = 0;
1732
1733         as = bch2_btree_update_start(trans, path, path->level,
1734                                      true, flags);
1735         if (IS_ERR(as))
1736                 return PTR_ERR(as);
1737
1738         ret = btree_split(as, trans, path, b, NULL, flags);
1739         if (ret) {
1740                 bch2_btree_update_free(as, trans);
1741                 return ret;
1742         }
1743
1744         bch2_btree_update_done(as, trans);
1745
1746         for (l = path->level + 1; btree_node_intent_locked(path, l) && !ret; l++)
1747                 ret = bch2_foreground_maybe_merge(trans, path, l, flags);
1748
1749         return ret;
1750 }
1751
1752 int __bch2_foreground_maybe_merge(struct btree_trans *trans,
1753                                   struct btree_path *path,
1754                                   unsigned level,
1755                                   unsigned flags,
1756                                   enum btree_node_sibling sib)
1757 {
1758         struct bch_fs *c = trans->c;
1759         struct btree_path *sib_path = NULL, *new_path = NULL;
1760         struct btree_update *as;
1761         struct bkey_format_state new_s;
1762         struct bkey_format new_f;
1763         struct bkey_i delete;
1764         struct btree *b, *m, *n, *prev, *next, *parent;
1765         struct bpos sib_pos;
1766         size_t sib_u64s;
1767         u64 start_time = local_clock();
1768         int ret = 0;
1769
1770         BUG_ON(!path->should_be_locked);
1771         BUG_ON(!btree_node_locked(path, level));
1772
1773         b = path->l[level].b;
1774
1775         if ((sib == btree_prev_sib && bpos_eq(b->data->min_key, POS_MIN)) ||
1776             (sib == btree_next_sib && bpos_eq(b->data->max_key, SPOS_MAX))) {
1777                 b->sib_u64s[sib] = U16_MAX;
1778                 return 0;
1779         }
1780
1781         sib_pos = sib == btree_prev_sib
1782                 ? bpos_predecessor(b->data->min_key)
1783                 : bpos_successor(b->data->max_key);
1784
1785         sib_path = bch2_path_get(trans, path->btree_id, sib_pos,
1786                                  U8_MAX, level, BTREE_ITER_INTENT, _THIS_IP_);
1787         ret = bch2_btree_path_traverse(trans, sib_path, false);
1788         if (ret)
1789                 goto err;
1790
1791         btree_path_set_should_be_locked(sib_path);
1792
1793         m = sib_path->l[level].b;
1794
1795         if (btree_node_parent(path, b) !=
1796             btree_node_parent(sib_path, m)) {
1797                 b->sib_u64s[sib] = U16_MAX;
1798                 goto out;
1799         }
1800
1801         if (sib == btree_prev_sib) {
1802                 prev = m;
1803                 next = b;
1804         } else {
1805                 prev = b;
1806                 next = m;
1807         }
1808
1809         if (!bpos_eq(bpos_successor(prev->data->max_key), next->data->min_key)) {
1810                 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1811
1812                 bch2_bpos_to_text(&buf1, prev->data->max_key);
1813                 bch2_bpos_to_text(&buf2, next->data->min_key);
1814                 bch_err(c,
1815                         "%s(): btree topology error:\n"
1816                         "  prev ends at   %s\n"
1817                         "  next starts at %s",
1818                         __func__, buf1.buf, buf2.buf);
1819                 printbuf_exit(&buf1);
1820                 printbuf_exit(&buf2);
1821                 bch2_topology_error(c);
1822                 ret = -EIO;
1823                 goto err;
1824         }
1825
1826         bch2_bkey_format_init(&new_s);
1827         bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
1828         __bch2_btree_calc_format(&new_s, prev);
1829         __bch2_btree_calc_format(&new_s, next);
1830         bch2_bkey_format_add_pos(&new_s, next->data->max_key);
1831         new_f = bch2_bkey_format_done(&new_s);
1832
1833         sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1834                 btree_node_u64s_with_format(m, &new_f);
1835
1836         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1837                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1838                 sib_u64s /= 2;
1839                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1840         }
1841
1842         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1843         sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
1844         b->sib_u64s[sib] = sib_u64s;
1845
1846         if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
1847                 goto out;
1848
1849         parent = btree_node_parent(path, b);
1850         as = bch2_btree_update_start(trans, path, level, false,
1851                          BTREE_INSERT_NOFAIL|
1852                          BTREE_INSERT_USE_RESERVE|
1853                          flags);
1854         ret = PTR_ERR_OR_ZERO(as);
1855         if (ret)
1856                 goto err;
1857
1858         trace_and_count(c, btree_node_merge, c, b);
1859
1860         bch2_btree_interior_update_will_free_node(as, b);
1861         bch2_btree_interior_update_will_free_node(as, m);
1862
1863         n = bch2_btree_node_alloc(as, trans, b->c.level);
1864
1865         SET_BTREE_NODE_SEQ(n->data,
1866                            max(BTREE_NODE_SEQ(b->data),
1867                                BTREE_NODE_SEQ(m->data)) + 1);
1868
1869         btree_set_min(n, prev->data->min_key);
1870         btree_set_max(n, next->data->max_key);
1871
1872         n->data->format  = new_f;
1873         btree_node_set_format(n, new_f);
1874
1875         bch2_btree_sort_into(c, n, prev);
1876         bch2_btree_sort_into(c, n, next);
1877
1878         bch2_btree_build_aux_trees(n);
1879         bch2_btree_update_add_new_node(as, n);
1880         six_unlock_write(&n->c.lock);
1881
1882         new_path = get_unlocked_mut_path(trans, path->btree_id, n->c.level, n->key.k.p);
1883         six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1884         mark_btree_node_locked(trans, new_path, n->c.level, SIX_LOCK_intent);
1885         bch2_btree_path_level_init(trans, new_path, n);
1886
1887         bkey_init(&delete.k);
1888         delete.k.p = prev->key.k.p;
1889         bch2_keylist_add(&as->parent_keys, &delete);
1890         bch2_keylist_add(&as->parent_keys, &n->key);
1891
1892         bch2_trans_verify_paths(trans);
1893
1894         ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1895         if (ret)
1896                 goto err_free_update;
1897
1898         bch2_trans_verify_paths(trans);
1899
1900         bch2_btree_update_get_open_buckets(as, n);
1901         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1902
1903         bch2_btree_node_free_inmem(trans, path, b);
1904         bch2_btree_node_free_inmem(trans, sib_path, m);
1905
1906         bch2_trans_node_add(trans, n);
1907
1908         bch2_trans_verify_paths(trans);
1909
1910         six_unlock_intent(&n->c.lock);
1911
1912         bch2_btree_update_done(as, trans);
1913
1914         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time);
1915 out:
1916 err:
1917         if (new_path)
1918                 bch2_path_put(trans, new_path, true);
1919         bch2_path_put(trans, sib_path, true);
1920         bch2_trans_verify_locks(trans);
1921         return ret;
1922 err_free_update:
1923         bch2_btree_node_free_never_used(as, trans, n);
1924         bch2_btree_update_free(as, trans);
1925         goto out;
1926 }
1927
1928 /**
1929  * bch_btree_node_rewrite - Rewrite/move a btree node
1930  */
1931 int bch2_btree_node_rewrite(struct btree_trans *trans,
1932                             struct btree_iter *iter,
1933                             struct btree *b,
1934                             unsigned flags)
1935 {
1936         struct bch_fs *c = trans->c;
1937         struct btree_path *new_path = NULL;
1938         struct btree *n, *parent;
1939         struct btree_update *as;
1940         int ret;
1941
1942         flags |= BTREE_INSERT_NOFAIL;
1943
1944         parent = btree_node_parent(iter->path, b);
1945         as = bch2_btree_update_start(trans, iter->path, b->c.level,
1946                                      false, flags);
1947         ret = PTR_ERR_OR_ZERO(as);
1948         if (ret)
1949                 goto out;
1950
1951         bch2_btree_interior_update_will_free_node(as, b);
1952
1953         n = bch2_btree_node_alloc_replacement(as, trans, b);
1954
1955         bch2_btree_build_aux_trees(n);
1956         bch2_btree_update_add_new_node(as, n);
1957         six_unlock_write(&n->c.lock);
1958
1959         new_path = get_unlocked_mut_path(trans, iter->btree_id, n->c.level, n->key.k.p);
1960         six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1961         mark_btree_node_locked(trans, new_path, n->c.level, SIX_LOCK_intent);
1962         bch2_btree_path_level_init(trans, new_path, n);
1963
1964         trace_and_count(c, btree_node_rewrite, c, b);
1965
1966         if (parent) {
1967                 bch2_keylist_add(&as->parent_keys, &n->key);
1968                 ret = bch2_btree_insert_node(as, trans, iter->path, parent,
1969                                              &as->parent_keys, flags);
1970                 if (ret)
1971                         goto err;
1972         } else {
1973                 bch2_btree_set_root(as, trans, iter->path, n);
1974         }
1975
1976         bch2_btree_update_get_open_buckets(as, n);
1977         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1978
1979         bch2_btree_node_free_inmem(trans, iter->path, b);
1980
1981         bch2_trans_node_add(trans, n);
1982         six_unlock_intent(&n->c.lock);
1983
1984         bch2_btree_update_done(as, trans);
1985 out:
1986         if (new_path)
1987                 bch2_path_put(trans, new_path, true);
1988         bch2_btree_path_downgrade(trans, iter->path);
1989         return ret;
1990 err:
1991         bch2_btree_node_free_never_used(as, trans, n);
1992         bch2_btree_update_free(as, trans);
1993         goto out;
1994 }
1995
1996 struct async_btree_rewrite {
1997         struct bch_fs           *c;
1998         struct work_struct      work;
1999         struct list_head        list;
2000         enum btree_id           btree_id;
2001         unsigned                level;
2002         struct bpos             pos;
2003         __le64                  seq;
2004 };
2005
2006 static int async_btree_node_rewrite_trans(struct btree_trans *trans,
2007                                           struct async_btree_rewrite *a)
2008 {
2009         struct bch_fs *c = trans->c;
2010         struct btree_iter iter;
2011         struct btree *b;
2012         int ret;
2013
2014         bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos,
2015                                   BTREE_MAX_DEPTH, a->level, 0);
2016         b = bch2_btree_iter_peek_node(&iter);
2017         ret = PTR_ERR_OR_ZERO(b);
2018         if (ret)
2019                 goto out;
2020
2021         if (!b || b->data->keys.seq != a->seq) {
2022                 struct printbuf buf = PRINTBUF;
2023
2024                 if (b)
2025                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
2026                 else
2027                         prt_str(&buf, "(null");
2028                 bch_info(c, "%s: node to rewrite not found:, searching for seq %llu, got\n%s",
2029                          __func__, a->seq, buf.buf);
2030                 printbuf_exit(&buf);
2031                 goto out;
2032         }
2033
2034         ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
2035 out:
2036         bch2_trans_iter_exit(trans, &iter);
2037
2038         return ret;
2039 }
2040
2041 void async_btree_node_rewrite_work(struct work_struct *work)
2042 {
2043         struct async_btree_rewrite *a =
2044                 container_of(work, struct async_btree_rewrite, work);
2045         struct bch_fs *c = a->c;
2046         int ret;
2047
2048         ret = bch2_trans_do(c, NULL, NULL, 0,
2049                       async_btree_node_rewrite_trans(&trans, a));
2050         if (ret)
2051                 bch_err(c, "%s: error %s", __func__, bch2_err_str(ret));
2052         bch2_write_ref_put(c, BCH_WRITE_REF_node_rewrite);
2053         kfree(a);
2054 }
2055
2056 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
2057 {
2058         struct async_btree_rewrite *a;
2059         int ret;
2060
2061         a = kmalloc(sizeof(*a), GFP_NOFS);
2062         if (!a) {
2063                 bch_err(c, "%s: error allocating memory", __func__);
2064                 return;
2065         }
2066
2067         a->c            = c;
2068         a->btree_id     = b->c.btree_id;
2069         a->level        = b->c.level;
2070         a->pos          = b->key.k.p;
2071         a->seq          = b->data->keys.seq;
2072         INIT_WORK(&a->work, async_btree_node_rewrite_work);
2073
2074         if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
2075                 mutex_lock(&c->pending_node_rewrites_lock);
2076                 list_add(&a->list, &c->pending_node_rewrites);
2077                 mutex_unlock(&c->pending_node_rewrites_lock);
2078                 return;
2079         }
2080
2081         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_node_rewrite)) {
2082                 if (test_bit(BCH_FS_STARTED, &c->flags)) {
2083                         bch_err(c, "%s: error getting c->writes ref", __func__);
2084                         kfree(a);
2085                         return;
2086                 }
2087
2088                 ret = bch2_fs_read_write_early(c);
2089                 if (ret) {
2090                         bch_err(c, "%s: error going read-write: %s",
2091                                 __func__, bch2_err_str(ret));
2092                         kfree(a);
2093                         return;
2094                 }
2095
2096                 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2097         }
2098
2099         queue_work(c->btree_interior_update_worker, &a->work);
2100 }
2101
2102 void bch2_do_pending_node_rewrites(struct bch_fs *c)
2103 {
2104         struct async_btree_rewrite *a, *n;
2105
2106         mutex_lock(&c->pending_node_rewrites_lock);
2107         list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2108                 list_del(&a->list);
2109
2110                 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2111                 queue_work(c->btree_interior_update_worker, &a->work);
2112         }
2113         mutex_unlock(&c->pending_node_rewrites_lock);
2114 }
2115
2116 void bch2_free_pending_node_rewrites(struct bch_fs *c)
2117 {
2118         struct async_btree_rewrite *a, *n;
2119
2120         mutex_lock(&c->pending_node_rewrites_lock);
2121         list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2122                 list_del(&a->list);
2123
2124                 kfree(a);
2125         }
2126         mutex_unlock(&c->pending_node_rewrites_lock);
2127 }
2128
2129 static int __bch2_btree_node_update_key(struct btree_trans *trans,
2130                                         struct btree_iter *iter,
2131                                         struct btree *b, struct btree *new_hash,
2132                                         struct bkey_i *new_key,
2133                                         bool skip_triggers)
2134 {
2135         struct bch_fs *c = trans->c;
2136         struct btree_iter iter2 = { NULL };
2137         struct btree *parent;
2138         int ret;
2139
2140         if (!skip_triggers) {
2141                 ret = bch2_trans_mark_old(trans, b->c.btree_id, b->c.level + 1,
2142                                           bkey_i_to_s_c(&b->key), 0);
2143                 if (ret)
2144                         return ret;
2145
2146                 ret = bch2_trans_mark_new(trans, b->c.btree_id, b->c.level + 1,
2147                                           new_key, 0);
2148                 if (ret)
2149                         return ret;
2150         }
2151
2152         if (new_hash) {
2153                 bkey_copy(&new_hash->key, new_key);
2154                 ret = bch2_btree_node_hash_insert(&c->btree_cache,
2155                                 new_hash, b->c.level, b->c.btree_id);
2156                 BUG_ON(ret);
2157         }
2158
2159         parent = btree_node_parent(iter->path, b);
2160         if (parent) {
2161                 bch2_trans_copy_iter(&iter2, iter);
2162
2163                 iter2.path = bch2_btree_path_make_mut(trans, iter2.path,
2164                                 iter2.flags & BTREE_ITER_INTENT,
2165                                 _THIS_IP_);
2166
2167                 BUG_ON(iter2.path->level != b->c.level);
2168                 BUG_ON(!bpos_eq(iter2.path->pos, new_key->k.p));
2169
2170                 btree_path_set_level_up(trans, iter2.path);
2171
2172                 trans->paths_sorted = false;
2173
2174                 ret   = bch2_btree_iter_traverse(&iter2) ?:
2175                         bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_NORUN);
2176                 if (ret)
2177                         goto err;
2178         } else {
2179                 BUG_ON(btree_node_root(c, b) != b);
2180
2181                 ret = darray_make_room(&trans->extra_journal_entries,
2182                                        jset_u64s(new_key->k.u64s));
2183                 if (ret)
2184                         return ret;
2185
2186                 journal_entry_set((void *) &darray_top(trans->extra_journal_entries),
2187                                   BCH_JSET_ENTRY_btree_root,
2188                                   b->c.btree_id, b->c.level,
2189                                   new_key, new_key->k.u64s);
2190                 trans->extra_journal_entries.nr += jset_u64s(new_key->k.u64s);
2191         }
2192
2193         ret = bch2_trans_commit(trans, NULL, NULL,
2194                                 BTREE_INSERT_NOFAIL|
2195                                 BTREE_INSERT_NOCHECK_RW|
2196                                 BTREE_INSERT_USE_RESERVE|
2197                                 BTREE_INSERT_JOURNAL_RECLAIM|
2198                                 JOURNAL_WATERMARK_reserved);
2199         if (ret)
2200                 goto err;
2201
2202         bch2_btree_node_lock_write_nofail(trans, iter->path, &b->c);
2203
2204         if (new_hash) {
2205                 mutex_lock(&c->btree_cache.lock);
2206                 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
2207                 bch2_btree_node_hash_remove(&c->btree_cache, b);
2208
2209                 bkey_copy(&b->key, new_key);
2210                 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
2211                 BUG_ON(ret);
2212                 mutex_unlock(&c->btree_cache.lock);
2213         } else {
2214                 bkey_copy(&b->key, new_key);
2215         }
2216
2217         bch2_btree_node_unlock_write(trans, iter->path, b);
2218 out:
2219         bch2_trans_iter_exit(trans, &iter2);
2220         return ret;
2221 err:
2222         if (new_hash) {
2223                 mutex_lock(&c->btree_cache.lock);
2224                 bch2_btree_node_hash_remove(&c->btree_cache, b);
2225                 mutex_unlock(&c->btree_cache.lock);
2226         }
2227         goto out;
2228 }
2229
2230 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
2231                                struct btree *b, struct bkey_i *new_key,
2232                                bool skip_triggers)
2233 {
2234         struct bch_fs *c = trans->c;
2235         struct btree *new_hash = NULL;
2236         struct btree_path *path = iter->path;
2237         struct closure cl;
2238         int ret = 0;
2239
2240         ret = bch2_btree_path_upgrade(trans, path, b->c.level + 1);
2241         if (ret)
2242                 return ret;
2243
2244         closure_init_stack(&cl);
2245
2246         /*
2247          * check btree_ptr_hash_val() after @b is locked by
2248          * btree_iter_traverse():
2249          */
2250         if (btree_ptr_hash_val(new_key) != b->hash_val) {
2251                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2252                 if (ret) {
2253                         ret = drop_locks_do(trans, (closure_sync(&cl), 0));
2254                         if (ret)
2255                                 return ret;
2256                 }
2257
2258                 new_hash = bch2_btree_node_mem_alloc(trans, false);
2259         }
2260
2261         path->intent_ref++;
2262         ret = __bch2_btree_node_update_key(trans, iter, b, new_hash,
2263                                            new_key, skip_triggers);
2264         --path->intent_ref;
2265
2266         if (new_hash) {
2267                 mutex_lock(&c->btree_cache.lock);
2268                 list_move(&new_hash->list, &c->btree_cache.freeable);
2269                 mutex_unlock(&c->btree_cache.lock);
2270
2271                 six_unlock_write(&new_hash->c.lock);
2272                 six_unlock_intent(&new_hash->c.lock);
2273         }
2274         closure_sync(&cl);
2275         bch2_btree_cache_cannibalize_unlock(c);
2276         return ret;
2277 }
2278
2279 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
2280                                         struct btree *b, struct bkey_i *new_key,
2281                                         bool skip_triggers)
2282 {
2283         struct btree_iter iter;
2284         int ret;
2285
2286         bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p,
2287                                   BTREE_MAX_DEPTH, b->c.level,
2288                                   BTREE_ITER_INTENT);
2289         ret = bch2_btree_iter_traverse(&iter);
2290         if (ret)
2291                 goto out;
2292
2293         /* has node been freed? */
2294         if (iter.path->l[b->c.level].b != b) {
2295                 /* node has been freed: */
2296                 BUG_ON(!btree_node_dying(b));
2297                 goto out;
2298         }
2299
2300         BUG_ON(!btree_node_hashed(b));
2301
2302         ret = bch2_btree_node_update_key(trans, &iter, b, new_key, skip_triggers);
2303 out:
2304         bch2_trans_iter_exit(trans, &iter);
2305         return ret;
2306 }
2307
2308 /* Init code: */
2309
2310 /*
2311  * Only for filesystem bringup, when first reading the btree roots or allocating
2312  * btree roots when initializing a new filesystem:
2313  */
2314 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2315 {
2316         BUG_ON(btree_node_root(c, b));
2317
2318         bch2_btree_set_root_inmem(c, b);
2319 }
2320
2321 static int __bch2_btree_root_alloc(struct btree_trans *trans, enum btree_id id)
2322 {
2323         struct bch_fs *c = trans->c;
2324         struct closure cl;
2325         struct btree *b;
2326         int ret;
2327
2328         closure_init_stack(&cl);
2329
2330         do {
2331                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2332                 closure_sync(&cl);
2333         } while (ret);
2334
2335         b = bch2_btree_node_mem_alloc(trans, false);
2336         bch2_btree_cache_cannibalize_unlock(c);
2337
2338         set_btree_node_fake(b);
2339         set_btree_node_need_rewrite(b);
2340         b->c.level      = 0;
2341         b->c.btree_id   = id;
2342
2343         bkey_btree_ptr_init(&b->key);
2344         b->key.k.p = SPOS_MAX;
2345         *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2346
2347         bch2_bset_init_first(b, &b->data->keys);
2348         bch2_btree_build_aux_trees(b);
2349
2350         b->data->flags = 0;
2351         btree_set_min(b, POS_MIN);
2352         btree_set_max(b, SPOS_MAX);
2353         b->data->format = bch2_btree_calc_format(b);
2354         btree_node_set_format(b, b->data->format);
2355
2356         ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2357                                           b->c.level, b->c.btree_id);
2358         BUG_ON(ret);
2359
2360         bch2_btree_set_root_inmem(c, b);
2361
2362         six_unlock_write(&b->c.lock);
2363         six_unlock_intent(&b->c.lock);
2364         return 0;
2365 }
2366
2367 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2368 {
2369         bch2_trans_run(c, __bch2_btree_root_alloc(&trans, id));
2370 }
2371
2372 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2373 {
2374         struct btree_update *as;
2375
2376         mutex_lock(&c->btree_interior_update_lock);
2377         list_for_each_entry(as, &c->btree_interior_update_list, list)
2378                 prt_printf(out, "%p m %u w %u r %u j %llu\n",
2379                        as,
2380                        as->mode,
2381                        as->nodes_written,
2382                        atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2383                        as->journal.seq);
2384         mutex_unlock(&c->btree_interior_update_lock);
2385 }
2386
2387 static bool bch2_btree_interior_updates_pending(struct bch_fs *c)
2388 {
2389         bool ret;
2390
2391         mutex_lock(&c->btree_interior_update_lock);
2392         ret = !list_empty(&c->btree_interior_update_list);
2393         mutex_unlock(&c->btree_interior_update_lock);
2394
2395         return ret;
2396 }
2397
2398 bool bch2_btree_interior_updates_flush(struct bch_fs *c)
2399 {
2400         bool ret = bch2_btree_interior_updates_pending(c);
2401
2402         if (ret)
2403                 closure_wait_event(&c->btree_interior_update_wait,
2404                                    !bch2_btree_interior_updates_pending(c));
2405         return ret;
2406 }
2407
2408 void bch2_journal_entry_to_btree_root(struct bch_fs *c, struct jset_entry *entry)
2409 {
2410         struct btree_root *r = &c->btree_roots[entry->btree_id];
2411
2412         mutex_lock(&c->btree_root_lock);
2413
2414         r->level = entry->level;
2415         r->alive = true;
2416         bkey_copy(&r->key, &entry->start[0]);
2417
2418         mutex_unlock(&c->btree_root_lock);
2419 }
2420
2421 struct jset_entry *
2422 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2423                                     struct jset_entry *start,
2424                                     struct jset_entry *end)
2425 {
2426         struct jset_entry *entry;
2427         unsigned long have = 0;
2428         unsigned i;
2429
2430         for (entry = start; entry < end; entry = vstruct_next(entry))
2431                 if (entry->type == BCH_JSET_ENTRY_btree_root)
2432                         __set_bit(entry->btree_id, &have);
2433
2434         mutex_lock(&c->btree_root_lock);
2435
2436         for (i = 0; i < BTREE_ID_NR; i++)
2437                 if (c->btree_roots[i].alive && !test_bit(i, &have)) {
2438                         journal_entry_set(end,
2439                                           BCH_JSET_ENTRY_btree_root,
2440                                           i, c->btree_roots[i].level,
2441                                           &c->btree_roots[i].key,
2442                                           c->btree_roots[i].key.k.u64s);
2443                         end = vstruct_next(end);
2444                 }
2445
2446         mutex_unlock(&c->btree_root_lock);
2447
2448         return end;
2449 }
2450
2451 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2452 {
2453         if (c->btree_interior_update_worker)
2454                 destroy_workqueue(c->btree_interior_update_worker);
2455         mempool_exit(&c->btree_interior_update_pool);
2456 }
2457
2458 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2459 {
2460         mutex_init(&c->btree_reserve_cache_lock);
2461         INIT_LIST_HEAD(&c->btree_interior_update_list);
2462         INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2463         mutex_init(&c->btree_interior_update_lock);
2464         INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2465
2466         INIT_LIST_HEAD(&c->pending_node_rewrites);
2467         mutex_init(&c->pending_node_rewrites_lock);
2468
2469         c->btree_interior_update_worker =
2470                 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2471         if (!c->btree_interior_update_worker)
2472                 return -BCH_ERR_ENOMEM_btree_interior_update_worker_init;
2473
2474         if (mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2475                                       sizeof(struct btree_update)))
2476                 return -BCH_ERR_ENOMEM_btree_interior_update_pool_init;
2477
2478         return 0;
2479 }