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         enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
 250         unsigned nr_reserve = watermark > BCH_WATERMARK_reclaim
 251                 ? BTREE_NODE_RESERVE
 252                 : 0;
 253         int ret;
 254
 255         mutex_lock(&c->btree_reserve_cache_lock);
 256         if (c->btree_reserve_cache_nr > nr_reserve) {
 257                 struct btree_alloc *a =
 258                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
 259
 260                 ob = a->ob;
 261                 bkey_copy(&tmp.k, &a->k);
 262                 mutex_unlock(&c->btree_reserve_cache_lock);
 263                 goto mem_alloc;
 264         }
 265         mutex_unlock(&c->btree_reserve_cache_lock);
 266
 267 retry:
 268         ret = bch2_alloc_sectors_start_trans(trans,
 269                                       c->opts.metadata_target ?:
 270                                       c->opts.foreground_target,
 271                                       0,
 272                                       writepoint_ptr(&c->btree_write_point),
 273                                       &devs_have,
 274                                       res->nr_replicas,
 275                                       c->opts.metadata_replicas_required,
 276                                       watermark, 0, cl, &wp);
 277         if (unlikely(ret))
 278                 return ERR_PTR(ret);
 279
 280         if (wp->sectors_free < btree_sectors(c)) {
 281                 struct open_bucket *ob;
 282                 unsigned i;
 283
 284                 open_bucket_for_each(c, &wp->ptrs, ob, i)
 285                         if (ob->sectors_free < btree_sectors(c))
 286                                 ob->sectors_free = 0;
 287
 288                 bch2_alloc_sectors_done(c, wp);
 289                 goto retry;
 290         }
 291
 292         bkey_btree_ptr_v2_init(&tmp.k);
 293         bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, btree_sectors(c), false);
 294
 295         bch2_open_bucket_get(c, wp, &ob);
 296         bch2_alloc_sectors_done(c, wp);
 297 mem_alloc:
 298         b = bch2_btree_node_mem_alloc(trans, interior_node);
 299         six_unlock_write(&b->c.lock);
 300         six_unlock_intent(&b->c.lock);
 301
 302         /* we hold cannibalize_lock: */
 303         BUG_ON(IS_ERR(b));
 304         BUG_ON(b->ob.nr);
 305
 306         bkey_copy(&b->key, &tmp.k);
 307         b->ob = ob;
 308
 309         return b;
 310 }
 311
 312 static struct btree *bch2_btree_node_alloc(struct btree_update *as,
 313                                            struct btree_trans *trans,
 314                                            unsigned level)
 315 {
 316         struct bch_fs *c = as->c;
 317         struct btree *b;
 318         struct prealloc_nodes *p = &as->prealloc_nodes[!!level];
 319         int ret;
 320
 321         BUG_ON(level >= BTREE_MAX_DEPTH);
 322         BUG_ON(!p->nr);
 323
 324         b = p->b[--p->nr];
 325
 326         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
 327         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
 328
 329         set_btree_node_accessed(b);
 330         set_btree_node_dirty_acct(c, b);
 331         set_btree_node_need_write(b);
 332
 333         bch2_bset_init_first(b, &b->data->keys);
 334         b->c.level      = level;
 335         b->c.btree_id   = as->btree_id;
 336         b->version_ondisk = c->sb.version;
 337
 338         memset(&b->nr, 0, sizeof(b->nr));
 339         b->data->magic = cpu_to_le64(bset_magic(c));
 340         memset(&b->data->_ptr, 0, sizeof(b->data->_ptr));
 341         b->data->flags = 0;
 342         SET_BTREE_NODE_ID(b->data, as->btree_id);
 343         SET_BTREE_NODE_LEVEL(b->data, level);
 344
 345         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 346                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
 347
 348                 bp->v.mem_ptr           = 0;
 349                 bp->v.seq               = b->data->keys.seq;
 350                 bp->v.sectors_written   = 0;
 351         }
 352
 353         SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
 354
 355         bch2_btree_build_aux_trees(b);
 356
 357         ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
 358         BUG_ON(ret);
 359
 360         trace_and_count(c, btree_node_alloc, c, b);
 361         bch2_increment_clock(c, btree_sectors(c), WRITE);
 362         return b;
 363 }
 364
 365 static void btree_set_min(struct btree *b, struct bpos pos)
 366 {
 367         if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
 368                 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
 369         b->data->min_key = pos;
 370 }
 371
 372 static void btree_set_max(struct btree *b, struct bpos pos)
 373 {
 374         b->key.k.p = pos;
 375         b->data->max_key = pos;
 376 }
 377
 378 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
 379                                                        struct btree_trans *trans,
 380                                                        struct btree *b)
 381 {
 382         struct btree *n = bch2_btree_node_alloc(as, trans, b->c.level);
 383         struct bkey_format format = bch2_btree_calc_format(b);
 384
 385         /*
 386          * The keys might expand with the new format - if they wouldn't fit in
 387          * the btree node anymore, use the old format for now:
 388          */
 389         if (!bch2_btree_node_format_fits(as->c, b, &format))
 390                 format = b->format;
 391
 392         SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
 393
 394         btree_set_min(n, b->data->min_key);
 395         btree_set_max(n, b->data->max_key);
 396
 397         n->data->format         = format;
 398         btree_node_set_format(n, format);
 399
 400         bch2_btree_sort_into(as->c, n, b);
 401
 402         btree_node_reset_sib_u64s(n);
 403         return n;
 404 }
 405
 406 static struct btree *__btree_root_alloc(struct btree_update *as,
 407                                 struct btree_trans *trans, unsigned level)
 408 {
 409         struct btree *b = bch2_btree_node_alloc(as, trans, level);
 410
 411         btree_set_min(b, POS_MIN);
 412         btree_set_max(b, SPOS_MAX);
 413         b->data->format = bch2_btree_calc_format(b);
 414
 415         btree_node_set_format(b, b->data->format);
 416         bch2_btree_build_aux_trees(b);
 417
 418         return b;
 419 }
 420
 421 static void bch2_btree_reserve_put(struct btree_update *as, struct btree_trans *trans)
 422 {
 423         struct bch_fs *c = as->c;
 424         struct prealloc_nodes *p;
 425
 426         for (p = as->prealloc_nodes;
 427              p < as->prealloc_nodes + ARRAY_SIZE(as->prealloc_nodes);
 428              p++) {
 429                 while (p->nr) {
 430                         struct btree *b = p->b[--p->nr];
 431
 432                         mutex_lock(&c->btree_reserve_cache_lock);
 433
 434                         if (c->btree_reserve_cache_nr <
 435                             ARRAY_SIZE(c->btree_reserve_cache)) {
 436                                 struct btree_alloc *a =
 437                                         &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
 438
 439                                 a->ob = b->ob;
 440                                 b->ob.nr = 0;
 441                                 bkey_copy(&a->k, &b->key);
 442                         } else {
 443                                 bch2_open_buckets_put(c, &b->ob);
 444                         }
 445
 446                         mutex_unlock(&c->btree_reserve_cache_lock);
 447
 448                         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
 449                         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
 450                         __btree_node_free(c, b);
 451                         six_unlock_write(&b->c.lock);
 452                         six_unlock_intent(&b->c.lock);
 453                 }
 454         }
 455 }
 456
 457 static int bch2_btree_reserve_get(struct btree_trans *trans,
 458                                   struct btree_update *as,
 459                                   unsigned nr_nodes[2],
 460                                   unsigned flags,
 461                                   struct closure *cl)
 462 {
 463         struct bch_fs *c = as->c;
 464         struct btree *b;
 465         unsigned interior;
 466         int ret = 0;
 467
 468         BUG_ON(nr_nodes[0] + nr_nodes[1] > BTREE_RESERVE_MAX);
 469
 470         /*
 471          * Protects reaping from the btree node cache and using the btree node
 472          * open bucket reserve:
 473          *
 474          * BTREE_INSERT_NOWAIT only applies to btree node allocation, not
 475          * blocking on this lock:
 476          */
 477         ret = bch2_btree_cache_cannibalize_lock(c, cl);
 478         if (ret)
 479                 return ret;
 480
 481         for (interior = 0; interior < 2; interior++) {
 482                 struct prealloc_nodes *p = as->prealloc_nodes + interior;
 483
 484                 while (p->nr < nr_nodes[interior]) {
 485                         b = __bch2_btree_node_alloc(trans, &as->disk_res,
 486                                         flags & BTREE_INSERT_NOWAIT ? NULL : cl,
 487                                         interior, flags);
 488                         if (IS_ERR(b)) {
 489                                 ret = PTR_ERR(b);
 490                                 goto err;
 491                         }
 492
 493                         p->b[p->nr++] = b;
 494                 }
 495         }
 496 err:
 497         bch2_btree_cache_cannibalize_unlock(c);
 498         return ret;
 499 }
 500
 501 /* Asynchronous interior node update machinery */
 502
 503 static void bch2_btree_update_free(struct btree_update *as, struct btree_trans *trans)
 504 {
 505         struct bch_fs *c = as->c;
 506
 507         if (as->took_gc_lock)
 508                 up_read(&c->gc_lock);
 509         as->took_gc_lock = false;
 510
 511         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 512
 513         bch2_journal_pin_drop(&c->journal, &as->journal);
 514         bch2_journal_pin_flush(&c->journal, &as->journal);
 515         bch2_disk_reservation_put(c, &as->disk_res);
 516         bch2_btree_reserve_put(as, trans);
 517
 518         bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_total],
 519                                as->start_time);
 520
 521         mutex_lock(&c->btree_interior_update_lock);
 522         list_del(&as->unwritten_list);
 523         list_del(&as->list);
 524
 525         closure_debug_destroy(&as->cl);
 526         mempool_free(as, &c->btree_interior_update_pool);
 527
 528         /*
 529          * Have to do the wakeup with btree_interior_update_lock still held,
 530          * since being on btree_interior_update_list is our ref on @c:
 531          */
 532         closure_wake_up(&c->btree_interior_update_wait);
 533
 534         mutex_unlock(&c->btree_interior_update_lock);
 535 }
 536
 537 static void btree_update_add_key(struct btree_update *as,
 538                                  struct keylist *keys, struct btree *b)
 539 {
 540         struct bkey_i *k = &b->key;
 541
 542         BUG_ON(bch2_keylist_u64s(keys) + k->k.u64s >
 543                ARRAY_SIZE(as->_old_keys));
 544
 545         bkey_copy(keys->top, k);
 546         bkey_i_to_btree_ptr_v2(keys->top)->v.mem_ptr = b->c.level + 1;
 547
 548         bch2_keylist_push(keys);
 549 }
 550
 551 /*
 552  * The transactional part of an interior btree node update, where we journal the
 553  * update we did to the interior node and update alloc info:
 554  */
 555 static int btree_update_nodes_written_trans(struct btree_trans *trans,
 556                                             struct btree_update *as)
 557 {
 558         struct bkey_i *k;
 559         int ret;
 560
 561         ret = darray_make_room(&trans->extra_journal_entries, as->journal_u64s);
 562         if (ret)
 563                 return ret;
 564
 565         memcpy(&darray_top(trans->extra_journal_entries),
 566                as->journal_entries,
 567                as->journal_u64s * sizeof(u64));
 568         trans->extra_journal_entries.nr += as->journal_u64s;
 569
 570         trans->journal_pin = &as->journal;
 571
 572         for_each_keylist_key(&as->old_keys, k) {
 573                 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
 574
 575                 ret = bch2_trans_mark_old(trans, as->btree_id, level, bkey_i_to_s_c(k), 0);
 576                 if (ret)
 577                         return ret;
 578         }
 579
 580         for_each_keylist_key(&as->new_keys, k) {
 581                 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
 582
 583                 ret = bch2_trans_mark_new(trans, as->btree_id, level, k, 0);
 584                 if (ret)
 585                         return ret;
 586         }
 587
 588         return 0;
 589 }
 590
 591 static void btree_update_nodes_written(struct btree_update *as)
 592 {
 593         struct bch_fs *c = as->c;
 594         struct btree *b;
 595         struct btree_trans trans;
 596         u64 journal_seq = 0;
 597         unsigned i;
 598         int ret;
 599
 600         bch2_trans_init(&trans, c, 0, 512);
 601         /*
 602          * If we're already in an error state, it might be because a btree node
 603          * was never written, and we might be trying to free that same btree
 604          * node here, but it won't have been marked as allocated and we'll see
 605          * spurious disk usage inconsistencies in the transactional part below
 606          * if we don't skip it:
 607          */
 608         ret = bch2_journal_error(&c->journal);
 609         if (ret)
 610                 goto err;
 611
 612         /*
 613          * Wait for any in flight writes to finish before we free the old nodes
 614          * on disk:
 615          */
 616         for (i = 0; i < as->nr_old_nodes; i++) {
 617                 __le64 seq;
 618
 619                 b = as->old_nodes[i];
 620
 621                 btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
 622                 seq = b->data ? b->data->keys.seq : 0;
 623                 six_unlock_read(&b->c.lock);
 624
 625                 if (seq == as->old_nodes_seq[i])
 626                         wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight_inner,
 627                                        TASK_UNINTERRUPTIBLE);
 628         }
 629
 630         /*
 631          * We did an update to a parent node where the pointers we added pointed
 632          * to child nodes that weren't written yet: now, the child nodes have
 633          * been written so we can write out the update to the interior node.
 634          */
 635
 636         /*
 637          * We can't call into journal reclaim here: we'd block on the journal
 638          * reclaim lock, but we may need to release the open buckets we have
 639          * pinned in order for other btree updates to make forward progress, and
 640          * journal reclaim does btree updates when flushing bkey_cached entries,
 641          * which may require allocations as well.
 642          */
 643         ret = commit_do(&trans, &as->disk_res, &journal_seq,
 644                         BCH_WATERMARK_reclaim|
 645                         BTREE_INSERT_NOFAIL|
 646                         BTREE_INSERT_NOCHECK_RW|
 647                         BTREE_INSERT_JOURNAL_RECLAIM,
 648                         btree_update_nodes_written_trans(&trans, as));
 649         bch2_trans_unlock(&trans);
 650
 651         bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
 652                              "%s(): error %s", __func__, bch2_err_str(ret));
 653 err:
 654         if (as->b) {
 655                 struct btree_path *path;
 656
 657                 b = as->b;
 658                 path = get_unlocked_mut_path(&trans, as->btree_id, b->c.level, b->key.k.p);
 659                 /*
 660                  * @b is the node we did the final insert into:
 661                  *
 662                  * On failure to get a journal reservation, we still have to
 663                  * unblock the write and allow most of the write path to happen
 664                  * so that shutdown works, but the i->journal_seq mechanism
 665                  * won't work to prevent the btree write from being visible (we
 666                  * didn't get a journal sequence number) - instead
 667                  * __bch2_btree_node_write() doesn't do the actual write if
 668                  * we're in journal error state:
 669                  */
 670
 671                 /*
 672                  * Ensure transaction is unlocked before using
 673                  * btree_node_lock_nopath() (the use of which is always suspect,
 674                  * we need to work on removing this in the future)
 675                  *
 676                  * It should be, but get_unlocked_mut_path() -> bch2_path_get()
 677                  * calls bch2_path_upgrade(), before we call path_make_mut(), so
 678                  * we may rarely end up with a locked path besides the one we
 679                  * have here:
 680                  */
 681                 bch2_trans_unlock(&trans);
 682                 btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_intent);
 683                 mark_btree_node_locked(&trans, path, b->c.level, SIX_LOCK_intent);
 684                 path->l[b->c.level].lock_seq = six_lock_seq(&b->c.lock);
 685                 path->l[b->c.level].b = b;
 686
 687                 bch2_btree_node_lock_write_nofail(&trans, path, &b->c);
 688
 689                 mutex_lock(&c->btree_interior_update_lock);
 690
 691                 list_del(&as->write_blocked_list);
 692                 if (list_empty(&b->write_blocked))
 693                         clear_btree_node_write_blocked(b);
 694
 695                 /*
 696                  * Node might have been freed, recheck under
 697                  * btree_interior_update_lock:
 698                  */
 699                 if (as->b == b) {
 700                         struct bset *i = btree_bset_last(b);
 701
 702                         BUG_ON(!b->c.level);
 703                         BUG_ON(!btree_node_dirty(b));
 704
 705                         if (!ret) {
 706                                 i->journal_seq = cpu_to_le64(
 707                                                              max(journal_seq,
 708                                                                  le64_to_cpu(i->journal_seq)));
 709
 710                                 bch2_btree_add_journal_pin(c, b, journal_seq);
 711                         } else {
 712                                 /*
 713                                  * If we didn't get a journal sequence number we
 714                                  * can't write this btree node, because recovery
 715                                  * won't know to ignore this write:
 716                                  */
 717                                 set_btree_node_never_write(b);
 718                         }
 719                 }
 720
 721                 mutex_unlock(&c->btree_interior_update_lock);
 722
 723                 mark_btree_node_locked_noreset(path, b->c.level, SIX_LOCK_intent);
 724                 six_unlock_write(&b->c.lock);
 725
 726                 btree_node_write_if_need(c, b, SIX_LOCK_intent);
 727                 btree_node_unlock(&trans, path, b->c.level);
 728                 bch2_path_put(&trans, path, true);
 729         }
 730
 731         bch2_journal_pin_drop(&c->journal, &as->journal);
 732
 733         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 734
 735         mutex_lock(&c->btree_interior_update_lock);
 736         for (i = 0; i < as->nr_new_nodes; i++) {
 737                 b = as->new_nodes[i];
 738
 739                 BUG_ON(b->will_make_reachable != (unsigned long) as);
 740                 b->will_make_reachable = 0;
 741                 clear_btree_node_will_make_reachable(b);
 742         }
 743         mutex_unlock(&c->btree_interior_update_lock);
 744
 745         for (i = 0; i < as->nr_new_nodes; i++) {
 746                 b = as->new_nodes[i];
 747
 748                 btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
 749                 btree_node_write_if_need(c, b, SIX_LOCK_read);
 750                 six_unlock_read(&b->c.lock);
 751         }
 752
 753         for (i = 0; i < as->nr_open_buckets; i++)
 754                 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
 755
 756         bch2_btree_update_free(as, &trans);
 757         bch2_trans_exit(&trans);
 758 }
 759
 760 static void btree_interior_update_work(struct work_struct *work)
 761 {
 762         struct bch_fs *c =
 763                 container_of(work, struct bch_fs, btree_interior_update_work);
 764         struct btree_update *as;
 765
 766         while (1) {
 767                 mutex_lock(&c->btree_interior_update_lock);
 768                 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
 769                                               struct btree_update, unwritten_list);
 770                 if (as && !as->nodes_written)
 771                         as = NULL;
 772                 mutex_unlock(&c->btree_interior_update_lock);
 773
 774                 if (!as)
 775                         break;
 776
 777                 btree_update_nodes_written(as);
 778         }
 779 }
 780
 781 static void btree_update_set_nodes_written(struct closure *cl)
 782 {
 783         struct btree_update *as = container_of(cl, struct btree_update, cl);
 784         struct bch_fs *c = as->c;
 785
 786         mutex_lock(&c->btree_interior_update_lock);
 787         as->nodes_written = true;
 788         mutex_unlock(&c->btree_interior_update_lock);
 789
 790         queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
 791 }
 792
 793 /*
 794  * We're updating @b with pointers to nodes that haven't finished writing yet:
 795  * block @b from being written until @as completes
 796  */
 797 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
 798 {
 799         struct bch_fs *c = as->c;
 800
 801         mutex_lock(&c->btree_interior_update_lock);
 802         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 803
 804         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 805         BUG_ON(!btree_node_dirty(b));
 806         BUG_ON(!b->c.level);
 807
 808         as->mode        = BTREE_INTERIOR_UPDATING_NODE;
 809         as->b           = b;
 810
 811         set_btree_node_write_blocked(b);
 812         list_add(&as->write_blocked_list, &b->write_blocked);
 813
 814         mutex_unlock(&c->btree_interior_update_lock);
 815 }
 816
 817 static void btree_update_reparent(struct btree_update *as,
 818                                   struct btree_update *child)
 819 {
 820         struct bch_fs *c = as->c;
 821
 822         lockdep_assert_held(&c->btree_interior_update_lock);
 823
 824         child->b = NULL;
 825         child->mode = BTREE_INTERIOR_UPDATING_AS;
 826
 827         bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL);
 828 }
 829
 830 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
 831 {
 832         struct bkey_i *insert = &b->key;
 833         struct bch_fs *c = as->c;
 834
 835         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 836
 837         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
 838                ARRAY_SIZE(as->journal_entries));
 839
 840         as->journal_u64s +=
 841                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
 842                                   BCH_JSET_ENTRY_btree_root,
 843                                   b->c.btree_id, b->c.level,
 844                                   insert, insert->k.u64s);
 845
 846         mutex_lock(&c->btree_interior_update_lock);
 847         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 848
 849         as->mode        = BTREE_INTERIOR_UPDATING_ROOT;
 850         mutex_unlock(&c->btree_interior_update_lock);
 851 }
 852
 853 /*
 854  * bch2_btree_update_add_new_node:
 855  *
 856  * This causes @as to wait on @b to be written, before it gets to
 857  * bch2_btree_update_nodes_written
 858  *
 859  * Additionally, it sets b->will_make_reachable to prevent any additional writes
 860  * to @b from happening besides the first until @b is reachable on disk
 861  *
 862  * And it adds @b to the list of @as's new nodes, so that we can update sector
 863  * counts in bch2_btree_update_nodes_written:
 864  */
 865 static void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
 866 {
 867         struct bch_fs *c = as->c;
 868
 869         closure_get(&as->cl);
 870
 871         mutex_lock(&c->btree_interior_update_lock);
 872         BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
 873         BUG_ON(b->will_make_reachable);
 874
 875         as->new_nodes[as->nr_new_nodes++] = b;
 876         b->will_make_reachable = 1UL|(unsigned long) as;
 877         set_btree_node_will_make_reachable(b);
 878
 879         mutex_unlock(&c->btree_interior_update_lock);
 880
 881         btree_update_add_key(as, &as->new_keys, b);
 882
 883         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 884                 unsigned bytes = vstruct_end(&b->data->keys) - (void *) b->data;
 885                 unsigned sectors = round_up(bytes, block_bytes(c)) >> 9;
 886
 887                 bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written =
 888                         cpu_to_le16(sectors);
 889         }
 890 }
 891
 892 /*
 893  * returns true if @b was a new node
 894  */
 895 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
 896 {
 897         struct btree_update *as;
 898         unsigned long v;
 899         unsigned i;
 900
 901         mutex_lock(&c->btree_interior_update_lock);
 902         /*
 903          * When b->will_make_reachable != 0, it owns a ref on as->cl that's
 904          * dropped when it gets written by bch2_btree_complete_write - the
 905          * xchg() is for synchronization with bch2_btree_complete_write:
 906          */
 907         v = xchg(&b->will_make_reachable, 0);
 908         clear_btree_node_will_make_reachable(b);
 909         as = (struct btree_update *) (v & ~1UL);
 910
 911         if (!as) {
 912                 mutex_unlock(&c->btree_interior_update_lock);
 913                 return;
 914         }
 915
 916         for (i = 0; i < as->nr_new_nodes; i++)
 917                 if (as->new_nodes[i] == b)
 918                         goto found;
 919
 920         BUG();
 921 found:
 922         array_remove_item(as->new_nodes, as->nr_new_nodes, i);
 923         mutex_unlock(&c->btree_interior_update_lock);
 924
 925         if (v & 1)
 926                 closure_put(&as->cl);
 927 }
 928
 929 static void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
 930 {
 931         while (b->ob.nr)
 932                 as->open_buckets[as->nr_open_buckets++] =
 933                         b->ob.v[--b->ob.nr];
 934 }
 935
 936 /*
 937  * @b is being split/rewritten: it may have pointers to not-yet-written btree
 938  * nodes and thus outstanding btree_updates - redirect @b's
 939  * btree_updates to point to this btree_update:
 940  */
 941 static void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 942                                                       struct btree *b)
 943 {
 944         struct bch_fs *c = as->c;
 945         struct btree_update *p, *n;
 946         struct btree_write *w;
 947
 948         set_btree_node_dying(b);
 949
 950         if (btree_node_fake(b))
 951                 return;
 952
 953         mutex_lock(&c->btree_interior_update_lock);
 954
 955         /*
 956          * Does this node have any btree_update operations preventing
 957          * it from being written?
 958          *
 959          * If so, redirect them to point to this btree_update: we can
 960          * write out our new nodes, but we won't make them visible until those
 961          * operations complete
 962          */
 963         list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
 964                 list_del_init(&p->write_blocked_list);
 965                 btree_update_reparent(as, p);
 966
 967                 /*
 968                  * for flush_held_btree_writes() waiting on updates to flush or
 969                  * nodes to be writeable:
 970                  */
 971                 closure_wake_up(&c->btree_interior_update_wait);
 972         }
 973
 974         clear_btree_node_dirty_acct(c, b);
 975         clear_btree_node_need_write(b);
 976         clear_btree_node_write_blocked(b);
 977
 978         /*
 979          * Does this node have unwritten data that has a pin on the journal?
 980          *
 981          * If so, transfer that pin to the btree_update operation -
 982          * note that if we're freeing multiple nodes, we only need to keep the
 983          * oldest pin of any of the nodes we're freeing. We'll release the pin
 984          * when the new nodes are persistent and reachable on disk:
 985          */
 986         w = btree_current_write(b);
 987         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 988         bch2_journal_pin_drop(&c->journal, &w->journal);
 989
 990         w = btree_prev_write(b);
 991         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 992         bch2_journal_pin_drop(&c->journal, &w->journal);
 993
 994         mutex_unlock(&c->btree_interior_update_lock);
 995
 996         /*
 997          * Is this a node that isn't reachable on disk yet?
 998          *
 999          * Nodes that aren't reachable yet have writes blocked until they're
1000          * reachable - now that we've cancelled any pending writes and moved
1001          * things waiting on that write to wait on this update, we can drop this
1002          * node from the list of nodes that the other update is making
1003          * reachable, prior to freeing it:
1004          */
1005         btree_update_drop_new_node(c, b);
1006
1007         btree_update_add_key(as, &as->old_keys, b);
1008
1009         as->old_nodes[as->nr_old_nodes] = b;
1010         as->old_nodes_seq[as->nr_old_nodes] = b->data->keys.seq;
1011         as->nr_old_nodes++;
1012 }
1013
1014 static void bch2_btree_update_done(struct btree_update *as, struct btree_trans *trans)
1015 {
1016         struct bch_fs *c = as->c;
1017         u64 start_time = as->start_time;
1018
1019         BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
1020
1021         if (as->took_gc_lock)
1022                 up_read(&as->c->gc_lock);
1023         as->took_gc_lock = false;
1024
1025         bch2_btree_reserve_put(as, trans);
1026
1027         continue_at(&as->cl, btree_update_set_nodes_written,
1028                     as->c->btree_interior_update_worker);
1029
1030         bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_foreground],
1031                                start_time);
1032 }
1033
1034 static struct btree_update *
1035 bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
1036                         unsigned level, bool split, unsigned flags)
1037 {
1038         struct bch_fs *c = trans->c;
1039         struct btree_update *as;
1040         u64 start_time = local_clock();
1041         int disk_res_flags = (flags & BTREE_INSERT_NOFAIL)
1042                 ? BCH_DISK_RESERVATION_NOFAIL : 0;
1043         unsigned nr_nodes[2] = { 0, 0 };
1044         unsigned update_level = level;
1045         enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
1046         unsigned journal_flags = 0;
1047         int ret = 0;
1048         u32 restart_count = trans->restart_count;
1049
1050         BUG_ON(!path->should_be_locked);
1051
1052         if (watermark == BCH_WATERMARK_copygc)
1053                 watermark = BCH_WATERMARK_btree_copygc;
1054         if (watermark < BCH_WATERMARK_btree)
1055                 watermark = BCH_WATERMARK_btree;
1056
1057         flags &= ~BCH_WATERMARK_MASK;
1058         flags |= watermark;
1059
1060         if (flags & BTREE_INSERT_JOURNAL_RECLAIM)
1061                 journal_flags |= JOURNAL_RES_GET_NONBLOCK;
1062         journal_flags |= watermark;
1063
1064         while (1) {
1065                 nr_nodes[!!update_level] += 1 + split;
1066                 update_level++;
1067
1068                 ret = bch2_btree_path_upgrade(trans, path, update_level + 1);
1069                 if (ret)
1070                         return ERR_PTR(ret);
1071
1072                 if (!btree_path_node(path, update_level)) {
1073                         /* Allocating new root? */
1074                         nr_nodes[1] += split;
1075                         update_level = BTREE_MAX_DEPTH;
1076                         break;
1077                 }
1078
1079                 if (bch2_btree_node_insert_fits(c, path->l[update_level].b,
1080                                         BKEY_BTREE_PTR_U64s_MAX * (1 + split)))
1081                         break;
1082
1083                 split = path->l[update_level].b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c);
1084         }
1085
1086         if (flags & BTREE_INSERT_GC_LOCK_HELD)
1087                 lockdep_assert_held(&c->gc_lock);
1088         else if (!down_read_trylock(&c->gc_lock)) {
1089                 ret = drop_locks_do(trans, (down_read(&c->gc_lock), 0));
1090                 if (ret) {
1091                         up_read(&c->gc_lock);
1092                         return ERR_PTR(ret);
1093                 }
1094         }
1095
1096         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOFS);
1097         memset(as, 0, sizeof(*as));
1098         closure_init(&as->cl, NULL);
1099         as->c           = c;
1100         as->start_time  = start_time;
1101         as->mode        = BTREE_INTERIOR_NO_UPDATE;
1102         as->took_gc_lock = !(flags & BTREE_INSERT_GC_LOCK_HELD);
1103         as->btree_id    = path->btree_id;
1104         as->update_level = update_level;
1105         INIT_LIST_HEAD(&as->list);
1106         INIT_LIST_HEAD(&as->unwritten_list);
1107         INIT_LIST_HEAD(&as->write_blocked_list);
1108         bch2_keylist_init(&as->old_keys, as->_old_keys);
1109         bch2_keylist_init(&as->new_keys, as->_new_keys);
1110         bch2_keylist_init(&as->parent_keys, as->inline_keys);
1111
1112         mutex_lock(&c->btree_interior_update_lock);
1113         list_add_tail(&as->list, &c->btree_interior_update_list);
1114         mutex_unlock(&c->btree_interior_update_lock);
1115
1116         /*
1117          * We don't want to allocate if we're in an error state, that can cause
1118          * deadlock on emergency shutdown due to open buckets getting stuck in
1119          * the btree_reserve_cache after allocator shutdown has cleared it out.
1120          * This check needs to come after adding us to the btree_interior_update
1121          * list but before calling bch2_btree_reserve_get, to synchronize with
1122          * __bch2_fs_read_only().
1123          */
1124         ret = bch2_journal_error(&c->journal);
1125         if (ret)
1126                 goto err;
1127
1128         ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
1129                                       BTREE_UPDATE_JOURNAL_RES,
1130                                       journal_flags|JOURNAL_RES_GET_NONBLOCK);
1131         if (ret) {
1132                 if (flags & BTREE_INSERT_JOURNAL_RECLAIM) {
1133                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
1134                         goto err;
1135                 }
1136
1137                 ret = drop_locks_do(trans,
1138                         bch2_journal_preres_get(&c->journal, &as->journal_preres,
1139                                               BTREE_UPDATE_JOURNAL_RES,
1140                                               journal_flags));
1141                 if (ret == -BCH_ERR_journal_preres_get_blocked) {
1142                         trace_and_count(c, trans_restart_journal_preres_get, trans, _RET_IP_, journal_flags);
1143                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_journal_preres_get);
1144                 }
1145                 if (ret)
1146                         goto err;
1147         }
1148
1149         ret = bch2_disk_reservation_get(c, &as->disk_res,
1150                         (nr_nodes[0] + nr_nodes[1]) * btree_sectors(c),
1151                         c->opts.metadata_replicas,
1152                         disk_res_flags);
1153         if (ret)
1154                 goto err;
1155
1156         ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, NULL);
1157         if (bch2_err_matches(ret, ENOSPC) ||
1158             bch2_err_matches(ret, ENOMEM)) {
1159                 struct closure cl;
1160
1161                 closure_init_stack(&cl);
1162
1163                 do {
1164                         ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, &cl);
1165
1166                         bch2_trans_unlock(trans);
1167                         closure_sync(&cl);
1168                 } while (bch2_err_matches(ret, BCH_ERR_operation_blocked));
1169         }
1170
1171         if (ret) {
1172                 trace_and_count(c, btree_reserve_get_fail, trans->fn,
1173                                 _RET_IP_, nr_nodes[0] + nr_nodes[1], ret);
1174                 goto err;
1175         }
1176
1177         ret = bch2_trans_relock(trans);
1178         if (ret)
1179                 goto err;
1180
1181         bch2_trans_verify_not_restarted(trans, restart_count);
1182         return as;
1183 err:
1184         bch2_btree_update_free(as, trans);
1185         return ERR_PTR(ret);
1186 }
1187
1188 /* Btree root updates: */
1189
1190 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1191 {
1192         /* Root nodes cannot be reaped */
1193         mutex_lock(&c->btree_cache.lock);
1194         list_del_init(&b->list);
1195         mutex_unlock(&c->btree_cache.lock);
1196
1197         mutex_lock(&c->btree_root_lock);
1198         BUG_ON(btree_node_root(c, b) &&
1199                (b->c.level < btree_node_root(c, b)->c.level ||
1200                 !btree_node_dying(btree_node_root(c, b))));
1201
1202         bch2_btree_id_root(c, b->c.btree_id)->b = b;
1203         mutex_unlock(&c->btree_root_lock);
1204
1205         bch2_recalc_btree_reserve(c);
1206 }
1207
1208 /**
1209  * bch_btree_set_root - update the root in memory and on disk
1210  *
1211  * To ensure forward progress, the current task must not be holding any
1212  * btree node write locks. However, you must hold an intent lock on the
1213  * old root.
1214  *
1215  * Note: This allocates a journal entry but doesn't add any keys to
1216  * it.  All the btree roots are part of every journal write, so there
1217  * is nothing new to be done.  This just guarantees that there is a
1218  * journal write.
1219  */
1220 static void bch2_btree_set_root(struct btree_update *as,
1221                                 struct btree_trans *trans,
1222                                 struct btree_path *path,
1223                                 struct btree *b)
1224 {
1225         struct bch_fs *c = as->c;
1226         struct btree *old;
1227
1228         trace_and_count(c, btree_node_set_root, c, b);
1229
1230         old = btree_node_root(c, b);
1231
1232         /*
1233          * Ensure no one is using the old root while we switch to the
1234          * new root:
1235          */
1236         bch2_btree_node_lock_write_nofail(trans, path, &old->c);
1237
1238         bch2_btree_set_root_inmem(c, b);
1239
1240         btree_update_updated_root(as, b);
1241
1242         /*
1243          * Unlock old root after new root is visible:
1244          *
1245          * The new root isn't persistent, but that's ok: we still have
1246          * an intent lock on the new root, and any updates that would
1247          * depend on the new root would have to update the new root.
1248          */
1249         bch2_btree_node_unlock_write(trans, path, old);
1250 }
1251
1252 /* Interior node updates: */
1253
1254 static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
1255                                         struct btree_trans *trans,
1256                                         struct btree_path *path,
1257                                         struct btree *b,
1258                                         struct btree_node_iter *node_iter,
1259                                         struct bkey_i *insert)
1260 {
1261         struct bch_fs *c = as->c;
1262         struct bkey_packed *k;
1263         struct printbuf buf = PRINTBUF;
1264         unsigned long old, new, v;
1265
1266         BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 &&
1267                !btree_ptr_sectors_written(insert));
1268
1269         if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
1270                 bch2_journal_key_overwritten(c, b->c.btree_id, b->c.level, insert->k.p);
1271
1272         if (bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1273                               btree_node_type(b), WRITE, &buf) ?:
1274             bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert), &buf)) {
1275                 printbuf_reset(&buf);
1276                 prt_printf(&buf, "inserting invalid bkey\n  ");
1277                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
1278                 prt_printf(&buf, "\n  ");
1279                 bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1280                                   btree_node_type(b), WRITE, &buf);
1281                 bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert), &buf);
1282
1283                 bch2_fs_inconsistent(c, "%s", buf.buf);
1284                 dump_stack();
1285         }
1286
1287         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1288                ARRAY_SIZE(as->journal_entries));
1289
1290         as->journal_u64s +=
1291                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1292                                   BCH_JSET_ENTRY_btree_keys,
1293                                   b->c.btree_id, b->c.level,
1294                                   insert, insert->k.u64s);
1295
1296         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1297                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1298                 bch2_btree_node_iter_advance(node_iter, b);
1299
1300         bch2_btree_bset_insert_key(trans, path, b, node_iter, insert);
1301         set_btree_node_dirty_acct(c, b);
1302
1303         v = READ_ONCE(b->flags);
1304         do {
1305                 old = new = v;
1306
1307                 new &= ~BTREE_WRITE_TYPE_MASK;
1308                 new |= BTREE_WRITE_interior;
1309                 new |= 1 << BTREE_NODE_need_write;
1310         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1311
1312         printbuf_exit(&buf);
1313 }
1314
1315 static void
1316 __bch2_btree_insert_keys_interior(struct btree_update *as,
1317                                   struct btree_trans *trans,
1318                                   struct btree_path *path,
1319                                   struct btree *b,
1320                                   struct btree_node_iter node_iter,
1321                                   struct keylist *keys)
1322 {
1323         struct bkey_i *insert = bch2_keylist_front(keys);
1324         struct bkey_packed *k;
1325
1326         BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1327
1328         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1329                (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1330                 ;
1331
1332         while (!bch2_keylist_empty(keys)) {
1333                 struct bkey_i *k = bch2_keylist_front(keys);
1334
1335                 if (bpos_gt(k->k.p, b->key.k.p))
1336                         break;
1337
1338                 bch2_insert_fixup_btree_ptr(as, trans, path, b, &node_iter, k);
1339                 bch2_keylist_pop_front(keys);
1340         }
1341 }
1342
1343 /*
1344  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1345  * node)
1346  */
1347 static void __btree_split_node(struct btree_update *as,
1348                                struct btree_trans *trans,
1349                                struct btree *b,
1350                                struct btree *n[2])
1351 {
1352         struct bkey_packed *k;
1353         struct bpos n1_pos = POS_MIN;
1354         struct btree_node_iter iter;
1355         struct bset *bsets[2];
1356         struct bkey_format_state format[2];
1357         struct bkey_packed *out[2];
1358         struct bkey uk;
1359         unsigned u64s, n1_u64s = (b->nr.live_u64s * 3) / 5;
1360         int i;
1361
1362         for (i = 0; i < 2; i++) {
1363                 BUG_ON(n[i]->nsets != 1);
1364
1365                 bsets[i] = btree_bset_first(n[i]);
1366                 out[i] = bsets[i]->start;
1367
1368                 SET_BTREE_NODE_SEQ(n[i]->data, BTREE_NODE_SEQ(b->data) + 1);
1369                 bch2_bkey_format_init(&format[i]);
1370         }
1371
1372         u64s = 0;
1373         for_each_btree_node_key(b, k, &iter) {
1374                 if (bkey_deleted(k))
1375                         continue;
1376
1377                 i = u64s >= n1_u64s;
1378                 u64s += k->u64s;
1379                 uk = bkey_unpack_key(b, k);
1380                 if (!i)
1381                         n1_pos = uk.p;
1382                 bch2_bkey_format_add_key(&format[i], &uk);
1383         }
1384
1385         btree_set_min(n[0], b->data->min_key);
1386         btree_set_max(n[0], n1_pos);
1387         btree_set_min(n[1], bpos_successor(n1_pos));
1388         btree_set_max(n[1], b->data->max_key);
1389
1390         for (i = 0; i < 2; i++) {
1391                 bch2_bkey_format_add_pos(&format[i], n[i]->data->min_key);
1392                 bch2_bkey_format_add_pos(&format[i], n[i]->data->max_key);
1393
1394                 n[i]->data->format = bch2_bkey_format_done(&format[i]);
1395                 btree_node_set_format(n[i], n[i]->data->format);
1396         }
1397
1398         u64s = 0;
1399         for_each_btree_node_key(b, k, &iter) {
1400                 if (bkey_deleted(k))
1401                         continue;
1402
1403                 i = u64s >= n1_u64s;
1404                 u64s += k->u64s;
1405
1406                 if (bch2_bkey_transform(&n[i]->format, out[i], bkey_packed(k)
1407                                         ? &b->format: &bch2_bkey_format_current, k))
1408                         out[i]->format = KEY_FORMAT_LOCAL_BTREE;
1409                 else
1410                         bch2_bkey_unpack(b, (void *) out[i], k);
1411
1412                 out[i]->needs_whiteout = false;
1413
1414                 btree_keys_account_key_add(&n[i]->nr, 0, out[i]);
1415                 out[i] = bkey_p_next(out[i]);
1416         }
1417
1418         for (i = 0; i < 2; i++) {
1419                 bsets[i]->u64s = cpu_to_le16((u64 *) out[i] - bsets[i]->_data);
1420
1421                 BUG_ON(!bsets[i]->u64s);
1422
1423                 set_btree_bset_end(n[i], n[i]->set);
1424
1425                 btree_node_reset_sib_u64s(n[i]);
1426
1427                 bch2_verify_btree_nr_keys(n[i]);
1428
1429                 if (b->c.level)
1430                         btree_node_interior_verify(as->c, n[i]);
1431         }
1432 }
1433
1434 /*
1435  * For updates to interior nodes, we've got to do the insert before we split
1436  * because the stuff we're inserting has to be inserted atomically. Post split,
1437  * the keys might have to go in different nodes and the split would no longer be
1438  * atomic.
1439  *
1440  * Worse, if the insert is from btree node coalescing, if we do the insert after
1441  * we do the split (and pick the pivot) - the pivot we pick might be between
1442  * nodes that were coalesced, and thus in the middle of a child node post
1443  * coalescing:
1444  */
1445 static void btree_split_insert_keys(struct btree_update *as,
1446                                     struct btree_trans *trans,
1447                                     struct btree_path *path,
1448                                     struct btree *b,
1449                                     struct keylist *keys)
1450 {
1451         if (!bch2_keylist_empty(keys) &&
1452             bpos_le(bch2_keylist_front(keys)->k.p, b->data->max_key)) {
1453                 struct btree_node_iter node_iter;
1454
1455                 bch2_btree_node_iter_init(&node_iter, b, &bch2_keylist_front(keys)->k.p);
1456
1457                 __bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
1458
1459                 btree_node_interior_verify(as->c, b);
1460         }
1461 }
1462
1463 static int btree_split(struct btree_update *as, struct btree_trans *trans,
1464                        struct btree_path *path, struct btree *b,
1465                        struct keylist *keys, unsigned flags)
1466 {
1467         struct bch_fs *c = as->c;
1468         struct btree *parent = btree_node_parent(path, b);
1469         struct btree *n1, *n2 = NULL, *n3 = NULL;
1470         struct btree_path *path1 = NULL, *path2 = NULL;
1471         u64 start_time = local_clock();
1472         int ret = 0;
1473
1474         BUG_ON(!parent && (b != btree_node_root(c, b)));
1475         BUG_ON(parent && !btree_node_intent_locked(path, b->c.level + 1));
1476
1477         bch2_btree_interior_update_will_free_node(as, b);
1478
1479         if (b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c)) {
1480                 struct btree *n[2];
1481
1482                 trace_and_count(c, btree_node_split, c, b);
1483
1484                 n[0] = n1 = bch2_btree_node_alloc(as, trans, b->c.level);
1485                 n[1] = n2 = bch2_btree_node_alloc(as, trans, b->c.level);
1486
1487                 __btree_split_node(as, trans, b, n);
1488
1489                 if (keys) {
1490                         btree_split_insert_keys(as, trans, path, n1, keys);
1491                         btree_split_insert_keys(as, trans, path, n2, keys);
1492                         BUG_ON(!bch2_keylist_empty(keys));
1493                 }
1494
1495                 bch2_btree_build_aux_trees(n2);
1496                 bch2_btree_build_aux_trees(n1);
1497
1498                 bch2_btree_update_add_new_node(as, n1);
1499                 bch2_btree_update_add_new_node(as, n2);
1500                 six_unlock_write(&n2->c.lock);
1501                 six_unlock_write(&n1->c.lock);
1502
1503                 path1 = get_unlocked_mut_path(trans, path->btree_id, n1->c.level, n1->key.k.p);
1504                 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1505                 mark_btree_node_locked(trans, path1, n1->c.level, SIX_LOCK_intent);
1506                 bch2_btree_path_level_init(trans, path1, n1);
1507
1508                 path2 = get_unlocked_mut_path(trans, path->btree_id, n2->c.level, n2->key.k.p);
1509                 six_lock_increment(&n2->c.lock, SIX_LOCK_intent);
1510                 mark_btree_node_locked(trans, path2, n2->c.level, SIX_LOCK_intent);
1511                 bch2_btree_path_level_init(trans, path2, n2);
1512
1513                 /*
1514                  * Note that on recursive parent_keys == keys, so we
1515                  * can't start adding new keys to parent_keys before emptying it
1516                  * out (which we did with btree_split_insert_keys() above)
1517                  */
1518                 bch2_keylist_add(&as->parent_keys, &n1->key);
1519                 bch2_keylist_add(&as->parent_keys, &n2->key);
1520
1521                 if (!parent) {
1522                         /* Depth increases, make a new root */
1523                         n3 = __btree_root_alloc(as, trans, b->c.level + 1);
1524
1525                         bch2_btree_update_add_new_node(as, n3);
1526                         six_unlock_write(&n3->c.lock);
1527
1528                         path2->locks_want++;
1529                         BUG_ON(btree_node_locked(path2, n3->c.level));
1530                         six_lock_increment(&n3->c.lock, SIX_LOCK_intent);
1531                         mark_btree_node_locked(trans, path2, n3->c.level, SIX_LOCK_intent);
1532                         bch2_btree_path_level_init(trans, path2, n3);
1533
1534                         n3->sib_u64s[0] = U16_MAX;
1535                         n3->sib_u64s[1] = U16_MAX;
1536
1537                         btree_split_insert_keys(as, trans, path, n3, &as->parent_keys);
1538                 }
1539         } else {
1540                 trace_and_count(c, btree_node_compact, c, b);
1541
1542                 n1 = bch2_btree_node_alloc_replacement(as, trans, b);
1543
1544                 if (keys) {
1545                         btree_split_insert_keys(as, trans, path, n1, keys);
1546                         BUG_ON(!bch2_keylist_empty(keys));
1547                 }
1548
1549                 bch2_btree_build_aux_trees(n1);
1550                 bch2_btree_update_add_new_node(as, n1);
1551                 six_unlock_write(&n1->c.lock);
1552
1553                 path1 = get_unlocked_mut_path(trans, path->btree_id, n1->c.level, n1->key.k.p);
1554                 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1555                 mark_btree_node_locked(trans, path1, n1->c.level, SIX_LOCK_intent);
1556                 bch2_btree_path_level_init(trans, path1, n1);
1557
1558                 if (parent)
1559                         bch2_keylist_add(&as->parent_keys, &n1->key);
1560         }
1561
1562         /* New nodes all written, now make them visible: */
1563
1564         if (parent) {
1565                 /* Split a non root node */
1566                 ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1567                 if (ret)
1568                         goto err;
1569         } else if (n3) {
1570                 bch2_btree_set_root(as, trans, path, n3);
1571         } else {
1572                 /* Root filled up but didn't need to be split */
1573                 bch2_btree_set_root(as, trans, path, n1);
1574         }
1575
1576         if (n3) {
1577                 bch2_btree_update_get_open_buckets(as, n3);
1578                 bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0);
1579         }
1580         if (n2) {
1581                 bch2_btree_update_get_open_buckets(as, n2);
1582                 bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0);
1583         }
1584         bch2_btree_update_get_open_buckets(as, n1);
1585         bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1586
1587         /*
1588          * The old node must be freed (in memory) _before_ unlocking the new
1589          * nodes - else another thread could re-acquire a read lock on the old
1590          * node after another thread has locked and updated the new node, thus
1591          * seeing stale data:
1592          */
1593         bch2_btree_node_free_inmem(trans, path, b);
1594
1595         if (n3)
1596                 bch2_trans_node_add(trans, n3);
1597         if (n2)
1598                 bch2_trans_node_add(trans, n2);
1599         bch2_trans_node_add(trans, n1);
1600
1601         if (n3)
1602                 six_unlock_intent(&n3->c.lock);
1603         if (n2)
1604                 six_unlock_intent(&n2->c.lock);
1605         six_unlock_intent(&n1->c.lock);
1606 out:
1607         if (path2) {
1608                 __bch2_btree_path_unlock(trans, path2);
1609                 bch2_path_put(trans, path2, true);
1610         }
1611         if (path1) {
1612                 __bch2_btree_path_unlock(trans, path1);
1613                 bch2_path_put(trans, path1, true);
1614         }
1615
1616         bch2_trans_verify_locks(trans);
1617
1618         bch2_time_stats_update(&c->times[n2
1619                                ? BCH_TIME_btree_node_split
1620                                : BCH_TIME_btree_node_compact],
1621                                start_time);
1622         return ret;
1623 err:
1624         if (n3)
1625                 bch2_btree_node_free_never_used(as, trans, n3);
1626         if (n2)
1627                 bch2_btree_node_free_never_used(as, trans, n2);
1628         bch2_btree_node_free_never_used(as, trans, n1);
1629         goto out;
1630 }
1631
1632 static void
1633 bch2_btree_insert_keys_interior(struct btree_update *as,
1634                                 struct btree_trans *trans,
1635                                 struct btree_path *path,
1636                                 struct btree *b,
1637                                 struct keylist *keys)
1638 {
1639         struct btree_path *linked;
1640
1641         __bch2_btree_insert_keys_interior(as, trans, path, b,
1642                                           path->l[b->c.level].iter, keys);
1643
1644         btree_update_updated_node(as, b);
1645
1646         trans_for_each_path_with_node(trans, b, linked)
1647                 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1648
1649         bch2_trans_verify_paths(trans);
1650 }
1651
1652 /**
1653  * bch_btree_insert_node - insert bkeys into a given btree node
1654  *
1655  * @iter:               btree iterator
1656  * @keys:               list of keys to insert
1657  * @hook:               insert callback
1658  * @persistent:         if not null, @persistent will wait on journal write
1659  *
1660  * Inserts as many keys as it can into a given btree node, splitting it if full.
1661  * If a split occurred, this function will return early. This can only happen
1662  * for leaf nodes -- inserts into interior nodes have to be atomic.
1663  */
1664 static int bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans,
1665                                   struct btree_path *path, struct btree *b,
1666                                   struct keylist *keys, unsigned flags)
1667 {
1668         struct bch_fs *c = as->c;
1669         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1670         int old_live_u64s = b->nr.live_u64s;
1671         int live_u64s_added, u64s_added;
1672         int ret;
1673
1674         lockdep_assert_held(&c->gc_lock);
1675         BUG_ON(!btree_node_intent_locked(path, b->c.level));
1676         BUG_ON(!b->c.level);
1677         BUG_ON(!as || as->b);
1678         bch2_verify_keylist_sorted(keys);
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|flags);
1852         ret = PTR_ERR_OR_ZERO(as);
1853         if (ret)
1854                 goto err;
1855
1856         trace_and_count(c, btree_node_merge, c, b);
1857
1858         bch2_btree_interior_update_will_free_node(as, b);
1859         bch2_btree_interior_update_will_free_node(as, m);
1860
1861         n = bch2_btree_node_alloc(as, trans, b->c.level);
1862
1863         SET_BTREE_NODE_SEQ(n->data,
1864                            max(BTREE_NODE_SEQ(b->data),
1865                                BTREE_NODE_SEQ(m->data)) + 1);
1866
1867         btree_set_min(n, prev->data->min_key);
1868         btree_set_max(n, next->data->max_key);
1869
1870         n->data->format  = new_f;
1871         btree_node_set_format(n, new_f);
1872
1873         bch2_btree_sort_into(c, n, prev);
1874         bch2_btree_sort_into(c, n, next);
1875
1876         bch2_btree_build_aux_trees(n);
1877         bch2_btree_update_add_new_node(as, n);
1878         six_unlock_write(&n->c.lock);
1879
1880         new_path = get_unlocked_mut_path(trans, path->btree_id, n->c.level, n->key.k.p);
1881         six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1882         mark_btree_node_locked(trans, new_path, n->c.level, SIX_LOCK_intent);
1883         bch2_btree_path_level_init(trans, new_path, n);
1884
1885         bkey_init(&delete.k);
1886         delete.k.p = prev->key.k.p;
1887         bch2_keylist_add(&as->parent_keys, &delete);
1888         bch2_keylist_add(&as->parent_keys, &n->key);
1889
1890         bch2_trans_verify_paths(trans);
1891
1892         ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1893         if (ret)
1894                 goto err_free_update;
1895
1896         bch2_trans_verify_paths(trans);
1897
1898         bch2_btree_update_get_open_buckets(as, n);
1899         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1900
1901         bch2_btree_node_free_inmem(trans, path, b);
1902         bch2_btree_node_free_inmem(trans, sib_path, m);
1903
1904         bch2_trans_node_add(trans, n);
1905
1906         bch2_trans_verify_paths(trans);
1907
1908         six_unlock_intent(&n->c.lock);
1909
1910         bch2_btree_update_done(as, trans);
1911
1912         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time);
1913 out:
1914 err:
1915         if (new_path)
1916                 bch2_path_put(trans, new_path, true);
1917         bch2_path_put(trans, sib_path, true);
1918         bch2_trans_verify_locks(trans);
1919         return ret;
1920 err_free_update:
1921         bch2_btree_node_free_never_used(as, trans, n);
1922         bch2_btree_update_free(as, trans);
1923         goto out;
1924 }
1925
1926 /**
1927  * bch_btree_node_rewrite - Rewrite/move a btree node
1928  */
1929 int bch2_btree_node_rewrite(struct btree_trans *trans,
1930                             struct btree_iter *iter,
1931                             struct btree *b,
1932                             unsigned flags)
1933 {
1934         struct bch_fs *c = trans->c;
1935         struct btree_path *new_path = NULL;
1936         struct btree *n, *parent;
1937         struct btree_update *as;
1938         int ret;
1939
1940         flags |= BTREE_INSERT_NOFAIL;
1941
1942         parent = btree_node_parent(iter->path, b);
1943         as = bch2_btree_update_start(trans, iter->path, b->c.level,
1944                                      false, flags);
1945         ret = PTR_ERR_OR_ZERO(as);
1946         if (ret)
1947                 goto out;
1948
1949         bch2_btree_interior_update_will_free_node(as, b);
1950
1951         n = bch2_btree_node_alloc_replacement(as, trans, b);
1952
1953         bch2_btree_build_aux_trees(n);
1954         bch2_btree_update_add_new_node(as, n);
1955         six_unlock_write(&n->c.lock);
1956
1957         new_path = get_unlocked_mut_path(trans, iter->btree_id, n->c.level, n->key.k.p);
1958         six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1959         mark_btree_node_locked(trans, new_path, n->c.level, SIX_LOCK_intent);
1960         bch2_btree_path_level_init(trans, new_path, n);
1961
1962         trace_and_count(c, btree_node_rewrite, c, b);
1963
1964         if (parent) {
1965                 bch2_keylist_add(&as->parent_keys, &n->key);
1966                 ret = bch2_btree_insert_node(as, trans, iter->path, parent,
1967                                              &as->parent_keys, flags);
1968                 if (ret)
1969                         goto err;
1970         } else {
1971                 bch2_btree_set_root(as, trans, iter->path, n);
1972         }
1973
1974         bch2_btree_update_get_open_buckets(as, n);
1975         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1976
1977         bch2_btree_node_free_inmem(trans, iter->path, b);
1978
1979         bch2_trans_node_add(trans, n);
1980         six_unlock_intent(&n->c.lock);
1981
1982         bch2_btree_update_done(as, trans);
1983 out:
1984         if (new_path)
1985                 bch2_path_put(trans, new_path, true);
1986         bch2_btree_path_downgrade(trans, iter->path);
1987         return ret;
1988 err:
1989         bch2_btree_node_free_never_used(as, trans, n);
1990         bch2_btree_update_free(as, trans);
1991         goto out;
1992 }
1993
1994 struct async_btree_rewrite {
1995         struct bch_fs           *c;
1996         struct work_struct      work;
1997         struct list_head        list;
1998         enum btree_id           btree_id;
1999         unsigned                level;
2000         struct bpos             pos;
2001         __le64                  seq;
2002 };
2003
2004 static int async_btree_node_rewrite_trans(struct btree_trans *trans,
2005                                           struct async_btree_rewrite *a)
2006 {
2007         struct bch_fs *c = trans->c;
2008         struct btree_iter iter;
2009         struct btree *b;
2010         int ret;
2011
2012         bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos,
2013                                   BTREE_MAX_DEPTH, a->level, 0);
2014         b = bch2_btree_iter_peek_node(&iter);
2015         ret = PTR_ERR_OR_ZERO(b);
2016         if (ret)
2017                 goto out;
2018
2019         if (!b || b->data->keys.seq != a->seq) {
2020                 struct printbuf buf = PRINTBUF;
2021
2022                 if (b)
2023                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
2024                 else
2025                         prt_str(&buf, "(null");
2026                 bch_info(c, "%s: node to rewrite not found:, searching for seq %llu, got\n%s",
2027                          __func__, a->seq, buf.buf);
2028                 printbuf_exit(&buf);
2029                 goto out;
2030         }
2031
2032         ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
2033 out:
2034         bch2_trans_iter_exit(trans, &iter);
2035
2036         return ret;
2037 }
2038
2039 static void async_btree_node_rewrite_work(struct work_struct *work)
2040 {
2041         struct async_btree_rewrite *a =
2042                 container_of(work, struct async_btree_rewrite, work);
2043         struct bch_fs *c = a->c;
2044         int ret;
2045
2046         ret = bch2_trans_do(c, NULL, NULL, 0,
2047                       async_btree_node_rewrite_trans(&trans, a));
2048         if (ret)
2049                 bch_err(c, "%s: error %s", __func__, bch2_err_str(ret));
2050         bch2_write_ref_put(c, BCH_WRITE_REF_node_rewrite);
2051         kfree(a);
2052 }
2053
2054 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
2055 {
2056         struct async_btree_rewrite *a;
2057         int ret;
2058
2059         a = kmalloc(sizeof(*a), GFP_NOFS);
2060         if (!a) {
2061                 bch_err(c, "%s: error allocating memory", __func__);
2062                 return;
2063         }
2064
2065         a->c            = c;
2066         a->btree_id     = b->c.btree_id;
2067         a->level        = b->c.level;
2068         a->pos          = b->key.k.p;
2069         a->seq          = b->data->keys.seq;
2070         INIT_WORK(&a->work, async_btree_node_rewrite_work);
2071
2072         if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
2073                 mutex_lock(&c->pending_node_rewrites_lock);
2074                 list_add(&a->list, &c->pending_node_rewrites);
2075                 mutex_unlock(&c->pending_node_rewrites_lock);
2076                 return;
2077         }
2078
2079         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_node_rewrite)) {
2080                 if (test_bit(BCH_FS_STARTED, &c->flags)) {
2081                         bch_err(c, "%s: error getting c->writes ref", __func__);
2082                         kfree(a);
2083                         return;
2084                 }
2085
2086                 ret = bch2_fs_read_write_early(c);
2087                 if (ret) {
2088                         bch_err(c, "%s: error going read-write: %s",
2089                                 __func__, bch2_err_str(ret));
2090                         kfree(a);
2091                         return;
2092                 }
2093
2094                 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2095         }
2096
2097         queue_work(c->btree_interior_update_worker, &a->work);
2098 }
2099
2100 void bch2_do_pending_node_rewrites(struct bch_fs *c)
2101 {
2102         struct async_btree_rewrite *a, *n;
2103
2104         mutex_lock(&c->pending_node_rewrites_lock);
2105         list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2106                 list_del(&a->list);
2107
2108                 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2109                 queue_work(c->btree_interior_update_worker, &a->work);
2110         }
2111         mutex_unlock(&c->pending_node_rewrites_lock);
2112 }
2113
2114 void bch2_free_pending_node_rewrites(struct bch_fs *c)
2115 {
2116         struct async_btree_rewrite *a, *n;
2117
2118         mutex_lock(&c->pending_node_rewrites_lock);
2119         list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2120                 list_del(&a->list);
2121
2122                 kfree(a);
2123         }
2124         mutex_unlock(&c->pending_node_rewrites_lock);
2125 }
2126
2127 static int __bch2_btree_node_update_key(struct btree_trans *trans,
2128                                         struct btree_iter *iter,
2129                                         struct btree *b, struct btree *new_hash,
2130                                         struct bkey_i *new_key,
2131                                         unsigned commit_flags,
2132                                         bool skip_triggers)
2133 {
2134         struct bch_fs *c = trans->c;
2135         struct btree_iter iter2 = { NULL };
2136         struct btree *parent;
2137         int ret;
2138
2139         if (!skip_triggers) {
2140                 ret = bch2_trans_mark_old(trans, b->c.btree_id, b->c.level + 1,
2141                                           bkey_i_to_s_c(&b->key), 0);
2142                 if (ret)
2143                         return ret;
2144
2145                 ret = bch2_trans_mark_new(trans, b->c.btree_id, b->c.level + 1,
2146                                           new_key, 0);
2147                 if (ret)
2148                         return ret;
2149         }
2150
2151         if (new_hash) {
2152                 bkey_copy(&new_hash->key, new_key);
2153                 ret = bch2_btree_node_hash_insert(&c->btree_cache,
2154                                 new_hash, b->c.level, b->c.btree_id);
2155                 BUG_ON(ret);
2156         }
2157
2158         parent = btree_node_parent(iter->path, b);
2159         if (parent) {
2160                 bch2_trans_copy_iter(&iter2, iter);
2161
2162                 iter2.path = bch2_btree_path_make_mut(trans, iter2.path,
2163                                 iter2.flags & BTREE_ITER_INTENT,
2164                                 _THIS_IP_);
2165
2166                 BUG_ON(iter2.path->level != b->c.level);
2167                 BUG_ON(!bpos_eq(iter2.path->pos, new_key->k.p));
2168
2169                 btree_path_set_level_up(trans, iter2.path);
2170
2171                 trans->paths_sorted = false;
2172
2173                 ret   = bch2_btree_iter_traverse(&iter2) ?:
2174                         bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_NORUN);
2175                 if (ret)
2176                         goto err;
2177         } else {
2178                 BUG_ON(btree_node_root(c, b) != b);
2179
2180                 ret = darray_make_room(&trans->extra_journal_entries,
2181                                        jset_u64s(new_key->k.u64s));
2182                 if (ret)
2183                         return ret;
2184
2185                 journal_entry_set((void *) &darray_top(trans->extra_journal_entries),
2186                                   BCH_JSET_ENTRY_btree_root,
2187                                   b->c.btree_id, b->c.level,
2188                                   new_key, new_key->k.u64s);
2189                 trans->extra_journal_entries.nr += jset_u64s(new_key->k.u64s);
2190         }
2191
2192         ret = bch2_trans_commit(trans, NULL, NULL, commit_flags);
2193         if (ret)
2194                 goto err;
2195
2196         bch2_btree_node_lock_write_nofail(trans, iter->path, &b->c);
2197
2198         if (new_hash) {
2199                 mutex_lock(&c->btree_cache.lock);
2200                 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
2201                 bch2_btree_node_hash_remove(&c->btree_cache, b);
2202
2203                 bkey_copy(&b->key, new_key);
2204                 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
2205                 BUG_ON(ret);
2206                 mutex_unlock(&c->btree_cache.lock);
2207         } else {
2208                 bkey_copy(&b->key, new_key);
2209         }
2210
2211         bch2_btree_node_unlock_write(trans, iter->path, b);
2212 out:
2213         bch2_trans_iter_exit(trans, &iter2);
2214         return ret;
2215 err:
2216         if (new_hash) {
2217                 mutex_lock(&c->btree_cache.lock);
2218                 bch2_btree_node_hash_remove(&c->btree_cache, b);
2219                 mutex_unlock(&c->btree_cache.lock);
2220         }
2221         goto out;
2222 }
2223
2224 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
2225                                struct btree *b, struct bkey_i *new_key,
2226                                unsigned commit_flags, bool skip_triggers)
2227 {
2228         struct bch_fs *c = trans->c;
2229         struct btree *new_hash = NULL;
2230         struct btree_path *path = iter->path;
2231         struct closure cl;
2232         int ret = 0;
2233
2234         ret = bch2_btree_path_upgrade(trans, path, b->c.level + 1);
2235         if (ret)
2236                 return ret;
2237
2238         closure_init_stack(&cl);
2239
2240         /*
2241          * check btree_ptr_hash_val() after @b is locked by
2242          * btree_iter_traverse():
2243          */
2244         if (btree_ptr_hash_val(new_key) != b->hash_val) {
2245                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2246                 if (ret) {
2247                         ret = drop_locks_do(trans, (closure_sync(&cl), 0));
2248                         if (ret)
2249                                 return ret;
2250                 }
2251
2252                 new_hash = bch2_btree_node_mem_alloc(trans, false);
2253         }
2254
2255         path->intent_ref++;
2256         ret = __bch2_btree_node_update_key(trans, iter, b, new_hash, new_key,
2257                                            commit_flags, skip_triggers);
2258         --path->intent_ref;
2259
2260         if (new_hash) {
2261                 mutex_lock(&c->btree_cache.lock);
2262                 list_move(&new_hash->list, &c->btree_cache.freeable);
2263                 mutex_unlock(&c->btree_cache.lock);
2264
2265                 six_unlock_write(&new_hash->c.lock);
2266                 six_unlock_intent(&new_hash->c.lock);
2267         }
2268         closure_sync(&cl);
2269         bch2_btree_cache_cannibalize_unlock(c);
2270         return ret;
2271 }
2272
2273 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
2274                                         struct btree *b, struct bkey_i *new_key,
2275                                         unsigned commit_flags, bool skip_triggers)
2276 {
2277         struct btree_iter iter;
2278         int ret;
2279
2280         bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p,
2281                                   BTREE_MAX_DEPTH, b->c.level,
2282                                   BTREE_ITER_INTENT);
2283         ret = bch2_btree_iter_traverse(&iter);
2284         if (ret)
2285                 goto out;
2286
2287         /* has node been freed? */
2288         if (iter.path->l[b->c.level].b != b) {
2289                 /* node has been freed: */
2290                 BUG_ON(!btree_node_dying(b));
2291                 goto out;
2292         }
2293
2294         BUG_ON(!btree_node_hashed(b));
2295
2296         ret = bch2_btree_node_update_key(trans, &iter, b, new_key,
2297                                          commit_flags, skip_triggers);
2298 out:
2299         bch2_trans_iter_exit(trans, &iter);
2300         return ret;
2301 }
2302
2303 /* Init code: */
2304
2305 /*
2306  * Only for filesystem bringup, when first reading the btree roots or allocating
2307  * btree roots when initializing a new filesystem:
2308  */
2309 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2310 {
2311         BUG_ON(btree_node_root(c, b));
2312
2313         bch2_btree_set_root_inmem(c, b);
2314 }
2315
2316 static int __bch2_btree_root_alloc(struct btree_trans *trans, enum btree_id id)
2317 {
2318         struct bch_fs *c = trans->c;
2319         struct closure cl;
2320         struct btree *b;
2321         int ret;
2322
2323         closure_init_stack(&cl);
2324
2325         do {
2326                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2327                 closure_sync(&cl);
2328         } while (ret);
2329
2330         b = bch2_btree_node_mem_alloc(trans, false);
2331         bch2_btree_cache_cannibalize_unlock(c);
2332
2333         set_btree_node_fake(b);
2334         set_btree_node_need_rewrite(b);
2335         b->c.level      = 0;
2336         b->c.btree_id   = id;
2337
2338         bkey_btree_ptr_init(&b->key);
2339         b->key.k.p = SPOS_MAX;
2340         *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2341
2342         bch2_bset_init_first(b, &b->data->keys);
2343         bch2_btree_build_aux_trees(b);
2344
2345         b->data->flags = 0;
2346         btree_set_min(b, POS_MIN);
2347         btree_set_max(b, SPOS_MAX);
2348         b->data->format = bch2_btree_calc_format(b);
2349         btree_node_set_format(b, b->data->format);
2350
2351         ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2352                                           b->c.level, b->c.btree_id);
2353         BUG_ON(ret);
2354
2355         bch2_btree_set_root_inmem(c, b);
2356
2357         six_unlock_write(&b->c.lock);
2358         six_unlock_intent(&b->c.lock);
2359         return 0;
2360 }
2361
2362 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2363 {
2364         bch2_trans_run(c, __bch2_btree_root_alloc(&trans, id));
2365 }
2366
2367 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2368 {
2369         struct btree_update *as;
2370
2371         mutex_lock(&c->btree_interior_update_lock);
2372         list_for_each_entry(as, &c->btree_interior_update_list, list)
2373                 prt_printf(out, "%p m %u w %u r %u j %llu\n",
2374                        as,
2375                        as->mode,
2376                        as->nodes_written,
2377                        atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2378                        as->journal.seq);
2379         mutex_unlock(&c->btree_interior_update_lock);
2380 }
2381
2382 static bool bch2_btree_interior_updates_pending(struct bch_fs *c)
2383 {
2384         bool ret;
2385
2386         mutex_lock(&c->btree_interior_update_lock);
2387         ret = !list_empty(&c->btree_interior_update_list);
2388         mutex_unlock(&c->btree_interior_update_lock);
2389
2390         return ret;
2391 }
2392
2393 bool bch2_btree_interior_updates_flush(struct bch_fs *c)
2394 {
2395         bool ret = bch2_btree_interior_updates_pending(c);
2396
2397         if (ret)
2398                 closure_wait_event(&c->btree_interior_update_wait,
2399                                    !bch2_btree_interior_updates_pending(c));
2400         return ret;
2401 }
2402
2403 void bch2_journal_entry_to_btree_root(struct bch_fs *c, struct jset_entry *entry)
2404 {
2405         struct btree_root *r = bch2_btree_id_root(c, entry->btree_id);
2406
2407         mutex_lock(&c->btree_root_lock);
2408
2409         r->level = entry->level;
2410         r->alive = true;
2411         bkey_copy(&r->key, &entry->start[0]);
2412
2413         mutex_unlock(&c->btree_root_lock);
2414 }
2415
2416 struct jset_entry *
2417 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2418                                     struct jset_entry *start,
2419                                     struct jset_entry *end)
2420 {
2421         struct jset_entry *entry;
2422         unsigned long have = 0;
2423         unsigned i;
2424
2425         for (entry = start; entry < end; entry = vstruct_next(entry))
2426                 if (entry->type == BCH_JSET_ENTRY_btree_root)
2427                         __set_bit(entry->btree_id, &have);
2428
2429         mutex_lock(&c->btree_root_lock);
2430
2431         for (i = 0; i < btree_id_nr_alive(c); i++) {
2432                 struct btree_root *r = bch2_btree_id_root(c, i);
2433
2434                 if (r->alive && !test_bit(i, &have)) {
2435                         journal_entry_set(end, BCH_JSET_ENTRY_btree_root,
2436                                           i, r->level, &r->key, r->key.k.u64s);
2437                         end = vstruct_next(end);
2438                 }
2439         }
2440
2441         mutex_unlock(&c->btree_root_lock);
2442
2443         return end;
2444 }
2445
2446 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2447 {
2448         if (c->btree_interior_update_worker)
2449                 destroy_workqueue(c->btree_interior_update_worker);
2450         mempool_exit(&c->btree_interior_update_pool);
2451 }
2452
2453 void bch2_fs_btree_interior_update_init_early(struct bch_fs *c)
2454 {
2455         mutex_init(&c->btree_reserve_cache_lock);
2456         INIT_LIST_HEAD(&c->btree_interior_update_list);
2457         INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2458         mutex_init(&c->btree_interior_update_lock);
2459         INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2460
2461         INIT_LIST_HEAD(&c->pending_node_rewrites);
2462         mutex_init(&c->pending_node_rewrites_lock);
2463 }
2464
2465 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2466 {
2467         c->btree_interior_update_worker =
2468                 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2469         if (!c->btree_interior_update_worker)
2470                 return -BCH_ERR_ENOMEM_btree_interior_update_worker_init;
2471
2472         if (mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2473                                       sizeof(struct btree_update)))
2474                 return -BCH_ERR_ENOMEM_btree_interior_update_pool_init;
2475
2476         return 0;
2477 }