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                 /*
1162                  * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
1163                  * flag
1164                  */
1165                 if (bch2_err_matches(ret, ENOSPC) &&
1166                     (flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
1167                     watermark != BCH_WATERMARK_reclaim) {
1168                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
1169                         goto err;
1170                 }
1171
1172                 closure_init_stack(&cl);
1173
1174                 do {
1175                         ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, &cl);
1176
1177                         bch2_trans_unlock(trans);
1178                         closure_sync(&cl);
1179                 } while (bch2_err_matches(ret, BCH_ERR_operation_blocked));
1180         }
1181
1182         if (ret) {
1183                 trace_and_count(c, btree_reserve_get_fail, trans->fn,
1184                                 _RET_IP_, nr_nodes[0] + nr_nodes[1], ret);
1185                 goto err;
1186         }
1187
1188         ret = bch2_trans_relock(trans);
1189         if (ret)
1190                 goto err;
1191
1192         bch2_trans_verify_not_restarted(trans, restart_count);
1193         return as;
1194 err:
1195         bch2_btree_update_free(as, trans);
1196         return ERR_PTR(ret);
1197 }
1198
1199 /* Btree root updates: */
1200
1201 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1202 {
1203         /* Root nodes cannot be reaped */
1204         mutex_lock(&c->btree_cache.lock);
1205         list_del_init(&b->list);
1206         mutex_unlock(&c->btree_cache.lock);
1207
1208         mutex_lock(&c->btree_root_lock);
1209         BUG_ON(btree_node_root(c, b) &&
1210                (b->c.level < btree_node_root(c, b)->c.level ||
1211                 !btree_node_dying(btree_node_root(c, b))));
1212
1213         bch2_btree_id_root(c, b->c.btree_id)->b = b;
1214         mutex_unlock(&c->btree_root_lock);
1215
1216         bch2_recalc_btree_reserve(c);
1217 }
1218
1219 /**
1220  * bch_btree_set_root - update the root in memory and on disk
1221  *
1222  * To ensure forward progress, the current task must not be holding any
1223  * btree node write locks. However, you must hold an intent lock on the
1224  * old root.
1225  *
1226  * Note: This allocates a journal entry but doesn't add any keys to
1227  * it.  All the btree roots are part of every journal write, so there
1228  * is nothing new to be done.  This just guarantees that there is a
1229  * journal write.
1230  */
1231 static void bch2_btree_set_root(struct btree_update *as,
1232                                 struct btree_trans *trans,
1233                                 struct btree_path *path,
1234                                 struct btree *b)
1235 {
1236         struct bch_fs *c = as->c;
1237         struct btree *old;
1238
1239         trace_and_count(c, btree_node_set_root, c, b);
1240
1241         old = btree_node_root(c, b);
1242
1243         /*
1244          * Ensure no one is using the old root while we switch to the
1245          * new root:
1246          */
1247         bch2_btree_node_lock_write_nofail(trans, path, &old->c);
1248
1249         bch2_btree_set_root_inmem(c, b);
1250
1251         btree_update_updated_root(as, b);
1252
1253         /*
1254          * Unlock old root after new root is visible:
1255          *
1256          * The new root isn't persistent, but that's ok: we still have
1257          * an intent lock on the new root, and any updates that would
1258          * depend on the new root would have to update the new root.
1259          */
1260         bch2_btree_node_unlock_write(trans, path, old);
1261 }
1262
1263 /* Interior node updates: */
1264
1265 static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
1266                                         struct btree_trans *trans,
1267                                         struct btree_path *path,
1268                                         struct btree *b,
1269                                         struct btree_node_iter *node_iter,
1270                                         struct bkey_i *insert)
1271 {
1272         struct bch_fs *c = as->c;
1273         struct bkey_packed *k;
1274         struct printbuf buf = PRINTBUF;
1275         unsigned long old, new, v;
1276
1277         BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 &&
1278                !btree_ptr_sectors_written(insert));
1279
1280         if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
1281                 bch2_journal_key_overwritten(c, b->c.btree_id, b->c.level, insert->k.p);
1282
1283         if (bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1284                               btree_node_type(b), WRITE, &buf) ?:
1285             bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert), &buf)) {
1286                 printbuf_reset(&buf);
1287                 prt_printf(&buf, "inserting invalid bkey\n  ");
1288                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
1289                 prt_printf(&buf, "\n  ");
1290                 bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1291                                   btree_node_type(b), WRITE, &buf);
1292                 bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert), &buf);
1293
1294                 bch2_fs_inconsistent(c, "%s", buf.buf);
1295                 dump_stack();
1296         }
1297
1298         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1299                ARRAY_SIZE(as->journal_entries));
1300
1301         as->journal_u64s +=
1302                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1303                                   BCH_JSET_ENTRY_btree_keys,
1304                                   b->c.btree_id, b->c.level,
1305                                   insert, insert->k.u64s);
1306
1307         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1308                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1309                 bch2_btree_node_iter_advance(node_iter, b);
1310
1311         bch2_btree_bset_insert_key(trans, path, b, node_iter, insert);
1312         set_btree_node_dirty_acct(c, b);
1313
1314         v = READ_ONCE(b->flags);
1315         do {
1316                 old = new = v;
1317
1318                 new &= ~BTREE_WRITE_TYPE_MASK;
1319                 new |= BTREE_WRITE_interior;
1320                 new |= 1 << BTREE_NODE_need_write;
1321         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1322
1323         printbuf_exit(&buf);
1324 }
1325
1326 static void
1327 __bch2_btree_insert_keys_interior(struct btree_update *as,
1328                                   struct btree_trans *trans,
1329                                   struct btree_path *path,
1330                                   struct btree *b,
1331                                   struct btree_node_iter node_iter,
1332                                   struct keylist *keys)
1333 {
1334         struct bkey_i *insert = bch2_keylist_front(keys);
1335         struct bkey_packed *k;
1336
1337         BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1338
1339         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1340                (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1341                 ;
1342
1343         while (!bch2_keylist_empty(keys)) {
1344                 struct bkey_i *k = bch2_keylist_front(keys);
1345
1346                 if (bpos_gt(k->k.p, b->key.k.p))
1347                         break;
1348
1349                 bch2_insert_fixup_btree_ptr(as, trans, path, b, &node_iter, k);
1350                 bch2_keylist_pop_front(keys);
1351         }
1352 }
1353
1354 /*
1355  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1356  * node)
1357  */
1358 static void __btree_split_node(struct btree_update *as,
1359                                struct btree_trans *trans,
1360                                struct btree *b,
1361                                struct btree *n[2])
1362 {
1363         struct bkey_packed *k;
1364         struct bpos n1_pos = POS_MIN;
1365         struct btree_node_iter iter;
1366         struct bset *bsets[2];
1367         struct bkey_format_state format[2];
1368         struct bkey_packed *out[2];
1369         struct bkey uk;
1370         unsigned u64s, n1_u64s = (b->nr.live_u64s * 3) / 5;
1371         int i;
1372
1373         for (i = 0; i < 2; i++) {
1374                 BUG_ON(n[i]->nsets != 1);
1375
1376                 bsets[i] = btree_bset_first(n[i]);
1377                 out[i] = bsets[i]->start;
1378
1379                 SET_BTREE_NODE_SEQ(n[i]->data, BTREE_NODE_SEQ(b->data) + 1);
1380                 bch2_bkey_format_init(&format[i]);
1381         }
1382
1383         u64s = 0;
1384         for_each_btree_node_key(b, k, &iter) {
1385                 if (bkey_deleted(k))
1386                         continue;
1387
1388                 i = u64s >= n1_u64s;
1389                 u64s += k->u64s;
1390                 uk = bkey_unpack_key(b, k);
1391                 if (!i)
1392                         n1_pos = uk.p;
1393                 bch2_bkey_format_add_key(&format[i], &uk);
1394         }
1395
1396         btree_set_min(n[0], b->data->min_key);
1397         btree_set_max(n[0], n1_pos);
1398         btree_set_min(n[1], bpos_successor(n1_pos));
1399         btree_set_max(n[1], b->data->max_key);
1400
1401         for (i = 0; i < 2; i++) {
1402                 bch2_bkey_format_add_pos(&format[i], n[i]->data->min_key);
1403                 bch2_bkey_format_add_pos(&format[i], n[i]->data->max_key);
1404
1405                 n[i]->data->format = bch2_bkey_format_done(&format[i]);
1406                 btree_node_set_format(n[i], n[i]->data->format);
1407         }
1408
1409         u64s = 0;
1410         for_each_btree_node_key(b, k, &iter) {
1411                 if (bkey_deleted(k))
1412                         continue;
1413
1414                 i = u64s >= n1_u64s;
1415                 u64s += k->u64s;
1416
1417                 if (bch2_bkey_transform(&n[i]->format, out[i], bkey_packed(k)
1418                                         ? &b->format: &bch2_bkey_format_current, k))
1419                         out[i]->format = KEY_FORMAT_LOCAL_BTREE;
1420                 else
1421                         bch2_bkey_unpack(b, (void *) out[i], k);
1422
1423                 out[i]->needs_whiteout = false;
1424
1425                 btree_keys_account_key_add(&n[i]->nr, 0, out[i]);
1426                 out[i] = bkey_p_next(out[i]);
1427         }
1428
1429         for (i = 0; i < 2; i++) {
1430                 bsets[i]->u64s = cpu_to_le16((u64 *) out[i] - bsets[i]->_data);
1431
1432                 BUG_ON(!bsets[i]->u64s);
1433
1434                 set_btree_bset_end(n[i], n[i]->set);
1435
1436                 btree_node_reset_sib_u64s(n[i]);
1437
1438                 bch2_verify_btree_nr_keys(n[i]);
1439
1440                 if (b->c.level)
1441                         btree_node_interior_verify(as->c, n[i]);
1442         }
1443 }
1444
1445 /*
1446  * For updates to interior nodes, we've got to do the insert before we split
1447  * because the stuff we're inserting has to be inserted atomically. Post split,
1448  * the keys might have to go in different nodes and the split would no longer be
1449  * atomic.
1450  *
1451  * Worse, if the insert is from btree node coalescing, if we do the insert after
1452  * we do the split (and pick the pivot) - the pivot we pick might be between
1453  * nodes that were coalesced, and thus in the middle of a child node post
1454  * coalescing:
1455  */
1456 static void btree_split_insert_keys(struct btree_update *as,
1457                                     struct btree_trans *trans,
1458                                     struct btree_path *path,
1459                                     struct btree *b,
1460                                     struct keylist *keys)
1461 {
1462         if (!bch2_keylist_empty(keys) &&
1463             bpos_le(bch2_keylist_front(keys)->k.p, b->data->max_key)) {
1464                 struct btree_node_iter node_iter;
1465
1466                 bch2_btree_node_iter_init(&node_iter, b, &bch2_keylist_front(keys)->k.p);
1467
1468                 __bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
1469
1470                 btree_node_interior_verify(as->c, b);
1471         }
1472 }
1473
1474 static int btree_split(struct btree_update *as, struct btree_trans *trans,
1475                        struct btree_path *path, struct btree *b,
1476                        struct keylist *keys, unsigned flags)
1477 {
1478         struct bch_fs *c = as->c;
1479         struct btree *parent = btree_node_parent(path, b);
1480         struct btree *n1, *n2 = NULL, *n3 = NULL;
1481         struct btree_path *path1 = NULL, *path2 = NULL;
1482         u64 start_time = local_clock();
1483         int ret = 0;
1484
1485         BUG_ON(!parent && (b != btree_node_root(c, b)));
1486         BUG_ON(parent && !btree_node_intent_locked(path, b->c.level + 1));
1487
1488         bch2_btree_interior_update_will_free_node(as, b);
1489
1490         if (b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c)) {
1491                 struct btree *n[2];
1492
1493                 trace_and_count(c, btree_node_split, c, b);
1494
1495                 n[0] = n1 = bch2_btree_node_alloc(as, trans, b->c.level);
1496                 n[1] = n2 = bch2_btree_node_alloc(as, trans, b->c.level);
1497
1498                 __btree_split_node(as, trans, b, n);
1499
1500                 if (keys) {
1501                         btree_split_insert_keys(as, trans, path, n1, keys);
1502                         btree_split_insert_keys(as, trans, path, n2, keys);
1503                         BUG_ON(!bch2_keylist_empty(keys));
1504                 }
1505
1506                 bch2_btree_build_aux_trees(n2);
1507                 bch2_btree_build_aux_trees(n1);
1508
1509                 bch2_btree_update_add_new_node(as, n1);
1510                 bch2_btree_update_add_new_node(as, n2);
1511                 six_unlock_write(&n2->c.lock);
1512                 six_unlock_write(&n1->c.lock);
1513
1514                 path1 = get_unlocked_mut_path(trans, path->btree_id, n1->c.level, n1->key.k.p);
1515                 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1516                 mark_btree_node_locked(trans, path1, n1->c.level, SIX_LOCK_intent);
1517                 bch2_btree_path_level_init(trans, path1, n1);
1518
1519                 path2 = get_unlocked_mut_path(trans, path->btree_id, n2->c.level, n2->key.k.p);
1520                 six_lock_increment(&n2->c.lock, SIX_LOCK_intent);
1521                 mark_btree_node_locked(trans, path2, n2->c.level, SIX_LOCK_intent);
1522                 bch2_btree_path_level_init(trans, path2, n2);
1523
1524                 /*
1525                  * Note that on recursive parent_keys == keys, so we
1526                  * can't start adding new keys to parent_keys before emptying it
1527                  * out (which we did with btree_split_insert_keys() above)
1528                  */
1529                 bch2_keylist_add(&as->parent_keys, &n1->key);
1530                 bch2_keylist_add(&as->parent_keys, &n2->key);
1531
1532                 if (!parent) {
1533                         /* Depth increases, make a new root */
1534                         n3 = __btree_root_alloc(as, trans, b->c.level + 1);
1535
1536                         bch2_btree_update_add_new_node(as, n3);
1537                         six_unlock_write(&n3->c.lock);
1538
1539                         path2->locks_want++;
1540                         BUG_ON(btree_node_locked(path2, n3->c.level));
1541                         six_lock_increment(&n3->c.lock, SIX_LOCK_intent);
1542                         mark_btree_node_locked(trans, path2, n3->c.level, SIX_LOCK_intent);
1543                         bch2_btree_path_level_init(trans, path2, n3);
1544
1545                         n3->sib_u64s[0] = U16_MAX;
1546                         n3->sib_u64s[1] = U16_MAX;
1547
1548                         btree_split_insert_keys(as, trans, path, n3, &as->parent_keys);
1549                 }
1550         } else {
1551                 trace_and_count(c, btree_node_compact, c, b);
1552
1553                 n1 = bch2_btree_node_alloc_replacement(as, trans, b);
1554
1555                 if (keys) {
1556                         btree_split_insert_keys(as, trans, path, n1, keys);
1557                         BUG_ON(!bch2_keylist_empty(keys));
1558                 }
1559
1560                 bch2_btree_build_aux_trees(n1);
1561                 bch2_btree_update_add_new_node(as, n1);
1562                 six_unlock_write(&n1->c.lock);
1563
1564                 path1 = get_unlocked_mut_path(trans, path->btree_id, n1->c.level, n1->key.k.p);
1565                 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1566                 mark_btree_node_locked(trans, path1, n1->c.level, SIX_LOCK_intent);
1567                 bch2_btree_path_level_init(trans, path1, n1);
1568
1569                 if (parent)
1570                         bch2_keylist_add(&as->parent_keys, &n1->key);
1571         }
1572
1573         /* New nodes all written, now make them visible: */
1574
1575         if (parent) {
1576                 /* Split a non root node */
1577                 ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1578                 if (ret)
1579                         goto err;
1580         } else if (n3) {
1581                 bch2_btree_set_root(as, trans, path, n3);
1582         } else {
1583                 /* Root filled up but didn't need to be split */
1584                 bch2_btree_set_root(as, trans, path, n1);
1585         }
1586
1587         if (n3) {
1588                 bch2_btree_update_get_open_buckets(as, n3);
1589                 bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0);
1590         }
1591         if (n2) {
1592                 bch2_btree_update_get_open_buckets(as, n2);
1593                 bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0);
1594         }
1595         bch2_btree_update_get_open_buckets(as, n1);
1596         bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1597
1598         /*
1599          * The old node must be freed (in memory) _before_ unlocking the new
1600          * nodes - else another thread could re-acquire a read lock on the old
1601          * node after another thread has locked and updated the new node, thus
1602          * seeing stale data:
1603          */
1604         bch2_btree_node_free_inmem(trans, path, b);
1605
1606         if (n3)
1607                 bch2_trans_node_add(trans, n3);
1608         if (n2)
1609                 bch2_trans_node_add(trans, n2);
1610         bch2_trans_node_add(trans, n1);
1611
1612         if (n3)
1613                 six_unlock_intent(&n3->c.lock);
1614         if (n2)
1615                 six_unlock_intent(&n2->c.lock);
1616         six_unlock_intent(&n1->c.lock);
1617 out:
1618         if (path2) {
1619                 __bch2_btree_path_unlock(trans, path2);
1620                 bch2_path_put(trans, path2, true);
1621         }
1622         if (path1) {
1623                 __bch2_btree_path_unlock(trans, path1);
1624                 bch2_path_put(trans, path1, true);
1625         }
1626
1627         bch2_trans_verify_locks(trans);
1628
1629         bch2_time_stats_update(&c->times[n2
1630                                ? BCH_TIME_btree_node_split
1631                                : BCH_TIME_btree_node_compact],
1632                                start_time);
1633         return ret;
1634 err:
1635         if (n3)
1636                 bch2_btree_node_free_never_used(as, trans, n3);
1637         if (n2)
1638                 bch2_btree_node_free_never_used(as, trans, n2);
1639         bch2_btree_node_free_never_used(as, trans, n1);
1640         goto out;
1641 }
1642
1643 static void
1644 bch2_btree_insert_keys_interior(struct btree_update *as,
1645                                 struct btree_trans *trans,
1646                                 struct btree_path *path,
1647                                 struct btree *b,
1648                                 struct keylist *keys)
1649 {
1650         struct btree_path *linked;
1651
1652         __bch2_btree_insert_keys_interior(as, trans, path, b,
1653                                           path->l[b->c.level].iter, keys);
1654
1655         btree_update_updated_node(as, b);
1656
1657         trans_for_each_path_with_node(trans, b, linked)
1658                 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1659
1660         bch2_trans_verify_paths(trans);
1661 }
1662
1663 /**
1664  * bch_btree_insert_node - insert bkeys into a given btree node
1665  *
1666  * @iter:               btree iterator
1667  * @keys:               list of keys to insert
1668  * @hook:               insert callback
1669  * @persistent:         if not null, @persistent will wait on journal write
1670  *
1671  * Inserts as many keys as it can into a given btree node, splitting it if full.
1672  * If a split occurred, this function will return early. This can only happen
1673  * for leaf nodes -- inserts into interior nodes have to be atomic.
1674  */
1675 static int bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans,
1676                                   struct btree_path *path, struct btree *b,
1677                                   struct keylist *keys, unsigned flags)
1678 {
1679         struct bch_fs *c = as->c;
1680         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1681         int old_live_u64s = b->nr.live_u64s;
1682         int live_u64s_added, u64s_added;
1683         int ret;
1684
1685         lockdep_assert_held(&c->gc_lock);
1686         BUG_ON(!btree_node_intent_locked(path, b->c.level));
1687         BUG_ON(!b->c.level);
1688         BUG_ON(!as || as->b);
1689         bch2_verify_keylist_sorted(keys);
1690
1691         ret = bch2_btree_node_lock_write(trans, path, &b->c);
1692         if (ret)
1693                 return ret;
1694
1695         bch2_btree_node_prep_for_write(trans, path, b);
1696
1697         if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1698                 bch2_btree_node_unlock_write(trans, path, b);
1699                 goto split;
1700         }
1701
1702         btree_node_interior_verify(c, b);
1703
1704         bch2_btree_insert_keys_interior(as, trans, path, b, keys);
1705
1706         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1707         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1708
1709         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1710                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1711         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1712                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1713
1714         if (u64s_added > live_u64s_added &&
1715             bch2_maybe_compact_whiteouts(c, b))
1716                 bch2_trans_node_reinit_iter(trans, b);
1717
1718         bch2_btree_node_unlock_write(trans, path, b);
1719
1720         btree_node_interior_verify(c, b);
1721         return 0;
1722 split:
1723         /*
1724          * We could attempt to avoid the transaction restart, by calling
1725          * bch2_btree_path_upgrade() and allocating more nodes:
1726          */
1727         if (b->c.level >= as->update_level) {
1728                 trace_and_count(c, trans_restart_split_race, trans, _THIS_IP_, b);
1729                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_split_race);
1730         }
1731
1732         return btree_split(as, trans, path, b, keys, flags);
1733 }
1734
1735 int bch2_btree_split_leaf(struct btree_trans *trans,
1736                           struct btree_path *path,
1737                           unsigned flags)
1738 {
1739         struct btree *b = path_l(path)->b;
1740         struct btree_update *as;
1741         unsigned l;
1742         int ret = 0;
1743
1744         as = bch2_btree_update_start(trans, path, path->level,
1745                                      true, flags);
1746         if (IS_ERR(as))
1747                 return PTR_ERR(as);
1748
1749         ret = btree_split(as, trans, path, b, NULL, flags);
1750         if (ret) {
1751                 bch2_btree_update_free(as, trans);
1752                 return ret;
1753         }
1754
1755         bch2_btree_update_done(as, trans);
1756
1757         for (l = path->level + 1; btree_node_intent_locked(path, l) && !ret; l++)
1758                 ret = bch2_foreground_maybe_merge(trans, path, l, flags);
1759
1760         return ret;
1761 }
1762
1763 int __bch2_foreground_maybe_merge(struct btree_trans *trans,
1764                                   struct btree_path *path,
1765                                   unsigned level,
1766                                   unsigned flags,
1767                                   enum btree_node_sibling sib)
1768 {
1769         struct bch_fs *c = trans->c;
1770         struct btree_path *sib_path = NULL, *new_path = NULL;
1771         struct btree_update *as;
1772         struct bkey_format_state new_s;
1773         struct bkey_format new_f;
1774         struct bkey_i delete;
1775         struct btree *b, *m, *n, *prev, *next, *parent;
1776         struct bpos sib_pos;
1777         size_t sib_u64s;
1778         u64 start_time = local_clock();
1779         int ret = 0;
1780
1781         BUG_ON(!path->should_be_locked);
1782         BUG_ON(!btree_node_locked(path, level));
1783
1784         b = path->l[level].b;
1785
1786         if ((sib == btree_prev_sib && bpos_eq(b->data->min_key, POS_MIN)) ||
1787             (sib == btree_next_sib && bpos_eq(b->data->max_key, SPOS_MAX))) {
1788                 b->sib_u64s[sib] = U16_MAX;
1789                 return 0;
1790         }
1791
1792         sib_pos = sib == btree_prev_sib
1793                 ? bpos_predecessor(b->data->min_key)
1794                 : bpos_successor(b->data->max_key);
1795
1796         sib_path = bch2_path_get(trans, path->btree_id, sib_pos,
1797                                  U8_MAX, level, BTREE_ITER_INTENT, _THIS_IP_);
1798         ret = bch2_btree_path_traverse(trans, sib_path, false);
1799         if (ret)
1800                 goto err;
1801
1802         btree_path_set_should_be_locked(sib_path);
1803
1804         m = sib_path->l[level].b;
1805
1806         if (btree_node_parent(path, b) !=
1807             btree_node_parent(sib_path, m)) {
1808                 b->sib_u64s[sib] = U16_MAX;
1809                 goto out;
1810         }
1811
1812         if (sib == btree_prev_sib) {
1813                 prev = m;
1814                 next = b;
1815         } else {
1816                 prev = b;
1817                 next = m;
1818         }
1819
1820         if (!bpos_eq(bpos_successor(prev->data->max_key), next->data->min_key)) {
1821                 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1822
1823                 bch2_bpos_to_text(&buf1, prev->data->max_key);
1824                 bch2_bpos_to_text(&buf2, next->data->min_key);
1825                 bch_err(c,
1826                         "%s(): btree topology error:\n"
1827                         "  prev ends at   %s\n"
1828                         "  next starts at %s",
1829                         __func__, buf1.buf, buf2.buf);
1830                 printbuf_exit(&buf1);
1831                 printbuf_exit(&buf2);
1832                 bch2_topology_error(c);
1833                 ret = -EIO;
1834                 goto err;
1835         }
1836
1837         bch2_bkey_format_init(&new_s);
1838         bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
1839         __bch2_btree_calc_format(&new_s, prev);
1840         __bch2_btree_calc_format(&new_s, next);
1841         bch2_bkey_format_add_pos(&new_s, next->data->max_key);
1842         new_f = bch2_bkey_format_done(&new_s);
1843
1844         sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1845                 btree_node_u64s_with_format(m, &new_f);
1846
1847         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1848                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1849                 sib_u64s /= 2;
1850                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1851         }
1852
1853         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1854         sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
1855         b->sib_u64s[sib] = sib_u64s;
1856
1857         if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
1858                 goto out;
1859
1860         parent = btree_node_parent(path, b);
1861         as = bch2_btree_update_start(trans, path, level, false,
1862                                      BTREE_INSERT_NOFAIL|flags);
1863         ret = PTR_ERR_OR_ZERO(as);
1864         if (ret)
1865                 goto err;
1866
1867         trace_and_count(c, btree_node_merge, c, b);
1868
1869         bch2_btree_interior_update_will_free_node(as, b);
1870         bch2_btree_interior_update_will_free_node(as, m);
1871
1872         n = bch2_btree_node_alloc(as, trans, b->c.level);
1873
1874         SET_BTREE_NODE_SEQ(n->data,
1875                            max(BTREE_NODE_SEQ(b->data),
1876                                BTREE_NODE_SEQ(m->data)) + 1);
1877
1878         btree_set_min(n, prev->data->min_key);
1879         btree_set_max(n, next->data->max_key);
1880
1881         n->data->format  = new_f;
1882         btree_node_set_format(n, new_f);
1883
1884         bch2_btree_sort_into(c, n, prev);
1885         bch2_btree_sort_into(c, n, next);
1886
1887         bch2_btree_build_aux_trees(n);
1888         bch2_btree_update_add_new_node(as, n);
1889         six_unlock_write(&n->c.lock);
1890
1891         new_path = get_unlocked_mut_path(trans, path->btree_id, n->c.level, n->key.k.p);
1892         six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1893         mark_btree_node_locked(trans, new_path, n->c.level, SIX_LOCK_intent);
1894         bch2_btree_path_level_init(trans, new_path, n);
1895
1896         bkey_init(&delete.k);
1897         delete.k.p = prev->key.k.p;
1898         bch2_keylist_add(&as->parent_keys, &delete);
1899         bch2_keylist_add(&as->parent_keys, &n->key);
1900
1901         bch2_trans_verify_paths(trans);
1902
1903         ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1904         if (ret)
1905                 goto err_free_update;
1906
1907         bch2_trans_verify_paths(trans);
1908
1909         bch2_btree_update_get_open_buckets(as, n);
1910         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1911
1912         bch2_btree_node_free_inmem(trans, path, b);
1913         bch2_btree_node_free_inmem(trans, sib_path, m);
1914
1915         bch2_trans_node_add(trans, n);
1916
1917         bch2_trans_verify_paths(trans);
1918
1919         six_unlock_intent(&n->c.lock);
1920
1921         bch2_btree_update_done(as, trans);
1922
1923         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time);
1924 out:
1925 err:
1926         if (new_path)
1927                 bch2_path_put(trans, new_path, true);
1928         bch2_path_put(trans, sib_path, true);
1929         bch2_trans_verify_locks(trans);
1930         return ret;
1931 err_free_update:
1932         bch2_btree_node_free_never_used(as, trans, n);
1933         bch2_btree_update_free(as, trans);
1934         goto out;
1935 }
1936
1937 /**
1938  * bch_btree_node_rewrite - Rewrite/move a btree node
1939  */
1940 int bch2_btree_node_rewrite(struct btree_trans *trans,
1941                             struct btree_iter *iter,
1942                             struct btree *b,
1943                             unsigned flags)
1944 {
1945         struct bch_fs *c = trans->c;
1946         struct btree_path *new_path = NULL;
1947         struct btree *n, *parent;
1948         struct btree_update *as;
1949         int ret;
1950
1951         flags |= BTREE_INSERT_NOFAIL;
1952
1953         parent = btree_node_parent(iter->path, b);
1954         as = bch2_btree_update_start(trans, iter->path, b->c.level,
1955                                      false, flags);
1956         ret = PTR_ERR_OR_ZERO(as);
1957         if (ret)
1958                 goto out;
1959
1960         bch2_btree_interior_update_will_free_node(as, b);
1961
1962         n = bch2_btree_node_alloc_replacement(as, trans, b);
1963
1964         bch2_btree_build_aux_trees(n);
1965         bch2_btree_update_add_new_node(as, n);
1966         six_unlock_write(&n->c.lock);
1967
1968         new_path = get_unlocked_mut_path(trans, iter->btree_id, n->c.level, n->key.k.p);
1969         six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1970         mark_btree_node_locked(trans, new_path, n->c.level, SIX_LOCK_intent);
1971         bch2_btree_path_level_init(trans, new_path, n);
1972
1973         trace_and_count(c, btree_node_rewrite, c, b);
1974
1975         if (parent) {
1976                 bch2_keylist_add(&as->parent_keys, &n->key);
1977                 ret = bch2_btree_insert_node(as, trans, iter->path, parent,
1978                                              &as->parent_keys, flags);
1979                 if (ret)
1980                         goto err;
1981         } else {
1982                 bch2_btree_set_root(as, trans, iter->path, n);
1983         }
1984
1985         bch2_btree_update_get_open_buckets(as, n);
1986         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1987
1988         bch2_btree_node_free_inmem(trans, iter->path, b);
1989
1990         bch2_trans_node_add(trans, n);
1991         six_unlock_intent(&n->c.lock);
1992
1993         bch2_btree_update_done(as, trans);
1994 out:
1995         if (new_path)
1996                 bch2_path_put(trans, new_path, true);
1997         bch2_btree_path_downgrade(trans, iter->path);
1998         return ret;
1999 err:
2000         bch2_btree_node_free_never_used(as, trans, n);
2001         bch2_btree_update_free(as, trans);
2002         goto out;
2003 }
2004
2005 struct async_btree_rewrite {
2006         struct bch_fs           *c;
2007         struct work_struct      work;
2008         struct list_head        list;
2009         enum btree_id           btree_id;
2010         unsigned                level;
2011         struct bpos             pos;
2012         __le64                  seq;
2013 };
2014
2015 static int async_btree_node_rewrite_trans(struct btree_trans *trans,
2016                                           struct async_btree_rewrite *a)
2017 {
2018         struct bch_fs *c = trans->c;
2019         struct btree_iter iter;
2020         struct btree *b;
2021         int ret;
2022
2023         bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos,
2024                                   BTREE_MAX_DEPTH, a->level, 0);
2025         b = bch2_btree_iter_peek_node(&iter);
2026         ret = PTR_ERR_OR_ZERO(b);
2027         if (ret)
2028                 goto out;
2029
2030         if (!b || b->data->keys.seq != a->seq) {
2031                 struct printbuf buf = PRINTBUF;
2032
2033                 if (b)
2034                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
2035                 else
2036                         prt_str(&buf, "(null");
2037                 bch_info(c, "%s: node to rewrite not found:, searching for seq %llu, got\n%s",
2038                          __func__, a->seq, buf.buf);
2039                 printbuf_exit(&buf);
2040                 goto out;
2041         }
2042
2043         ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
2044 out:
2045         bch2_trans_iter_exit(trans, &iter);
2046
2047         return ret;
2048 }
2049
2050 static void async_btree_node_rewrite_work(struct work_struct *work)
2051 {
2052         struct async_btree_rewrite *a =
2053                 container_of(work, struct async_btree_rewrite, work);
2054         struct bch_fs *c = a->c;
2055         int ret;
2056
2057         ret = bch2_trans_do(c, NULL, NULL, 0,
2058                       async_btree_node_rewrite_trans(&trans, a));
2059         if (ret)
2060                 bch_err(c, "%s: error %s", __func__, bch2_err_str(ret));
2061         bch2_write_ref_put(c, BCH_WRITE_REF_node_rewrite);
2062         kfree(a);
2063 }
2064
2065 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
2066 {
2067         struct async_btree_rewrite *a;
2068         int ret;
2069
2070         a = kmalloc(sizeof(*a), GFP_NOFS);
2071         if (!a) {
2072                 bch_err(c, "%s: error allocating memory", __func__);
2073                 return;
2074         }
2075
2076         a->c            = c;
2077         a->btree_id     = b->c.btree_id;
2078         a->level        = b->c.level;
2079         a->pos          = b->key.k.p;
2080         a->seq          = b->data->keys.seq;
2081         INIT_WORK(&a->work, async_btree_node_rewrite_work);
2082
2083         if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
2084                 mutex_lock(&c->pending_node_rewrites_lock);
2085                 list_add(&a->list, &c->pending_node_rewrites);
2086                 mutex_unlock(&c->pending_node_rewrites_lock);
2087                 return;
2088         }
2089
2090         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_node_rewrite)) {
2091                 if (test_bit(BCH_FS_STARTED, &c->flags)) {
2092                         bch_err(c, "%s: error getting c->writes ref", __func__);
2093                         kfree(a);
2094                         return;
2095                 }
2096
2097                 ret = bch2_fs_read_write_early(c);
2098                 if (ret) {
2099                         bch_err(c, "%s: error going read-write: %s",
2100                                 __func__, bch2_err_str(ret));
2101                         kfree(a);
2102                         return;
2103                 }
2104
2105                 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2106         }
2107
2108         queue_work(c->btree_interior_update_worker, &a->work);
2109 }
2110
2111 void bch2_do_pending_node_rewrites(struct bch_fs *c)
2112 {
2113         struct async_btree_rewrite *a, *n;
2114
2115         mutex_lock(&c->pending_node_rewrites_lock);
2116         list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2117                 list_del(&a->list);
2118
2119                 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2120                 queue_work(c->btree_interior_update_worker, &a->work);
2121         }
2122         mutex_unlock(&c->pending_node_rewrites_lock);
2123 }
2124
2125 void bch2_free_pending_node_rewrites(struct bch_fs *c)
2126 {
2127         struct async_btree_rewrite *a, *n;
2128
2129         mutex_lock(&c->pending_node_rewrites_lock);
2130         list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2131                 list_del(&a->list);
2132
2133                 kfree(a);
2134         }
2135         mutex_unlock(&c->pending_node_rewrites_lock);
2136 }
2137
2138 static int __bch2_btree_node_update_key(struct btree_trans *trans,
2139                                         struct btree_iter *iter,
2140                                         struct btree *b, struct btree *new_hash,
2141                                         struct bkey_i *new_key,
2142                                         unsigned commit_flags,
2143                                         bool skip_triggers)
2144 {
2145         struct bch_fs *c = trans->c;
2146         struct btree_iter iter2 = { NULL };
2147         struct btree *parent;
2148         int ret;
2149
2150         if (!skip_triggers) {
2151                 ret = bch2_trans_mark_old(trans, b->c.btree_id, b->c.level + 1,
2152                                           bkey_i_to_s_c(&b->key), 0);
2153                 if (ret)
2154                         return ret;
2155
2156                 ret = bch2_trans_mark_new(trans, b->c.btree_id, b->c.level + 1,
2157                                           new_key, 0);
2158                 if (ret)
2159                         return ret;
2160         }
2161
2162         if (new_hash) {
2163                 bkey_copy(&new_hash->key, new_key);
2164                 ret = bch2_btree_node_hash_insert(&c->btree_cache,
2165                                 new_hash, b->c.level, b->c.btree_id);
2166                 BUG_ON(ret);
2167         }
2168
2169         parent = btree_node_parent(iter->path, b);
2170         if (parent) {
2171                 bch2_trans_copy_iter(&iter2, iter);
2172
2173                 iter2.path = bch2_btree_path_make_mut(trans, iter2.path,
2174                                 iter2.flags & BTREE_ITER_INTENT,
2175                                 _THIS_IP_);
2176
2177                 BUG_ON(iter2.path->level != b->c.level);
2178                 BUG_ON(!bpos_eq(iter2.path->pos, new_key->k.p));
2179
2180                 btree_path_set_level_up(trans, iter2.path);
2181
2182                 trans->paths_sorted = false;
2183
2184                 ret   = bch2_btree_iter_traverse(&iter2) ?:
2185                         bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_NORUN);
2186                 if (ret)
2187                         goto err;
2188         } else {
2189                 BUG_ON(btree_node_root(c, b) != b);
2190
2191                 ret = darray_make_room(&trans->extra_journal_entries,
2192                                        jset_u64s(new_key->k.u64s));
2193                 if (ret)
2194                         return ret;
2195
2196                 journal_entry_set((void *) &darray_top(trans->extra_journal_entries),
2197                                   BCH_JSET_ENTRY_btree_root,
2198                                   b->c.btree_id, b->c.level,
2199                                   new_key, new_key->k.u64s);
2200                 trans->extra_journal_entries.nr += jset_u64s(new_key->k.u64s);
2201         }
2202
2203         ret = bch2_trans_commit(trans, NULL, NULL, commit_flags);
2204         if (ret)
2205                 goto err;
2206
2207         bch2_btree_node_lock_write_nofail(trans, iter->path, &b->c);
2208
2209         if (new_hash) {
2210                 mutex_lock(&c->btree_cache.lock);
2211                 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
2212                 bch2_btree_node_hash_remove(&c->btree_cache, b);
2213
2214                 bkey_copy(&b->key, new_key);
2215                 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
2216                 BUG_ON(ret);
2217                 mutex_unlock(&c->btree_cache.lock);
2218         } else {
2219                 bkey_copy(&b->key, new_key);
2220         }
2221
2222         bch2_btree_node_unlock_write(trans, iter->path, b);
2223 out:
2224         bch2_trans_iter_exit(trans, &iter2);
2225         return ret;
2226 err:
2227         if (new_hash) {
2228                 mutex_lock(&c->btree_cache.lock);
2229                 bch2_btree_node_hash_remove(&c->btree_cache, b);
2230                 mutex_unlock(&c->btree_cache.lock);
2231         }
2232         goto out;
2233 }
2234
2235 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
2236                                struct btree *b, struct bkey_i *new_key,
2237                                unsigned commit_flags, bool skip_triggers)
2238 {
2239         struct bch_fs *c = trans->c;
2240         struct btree *new_hash = NULL;
2241         struct btree_path *path = iter->path;
2242         struct closure cl;
2243         int ret = 0;
2244
2245         ret = bch2_btree_path_upgrade(trans, path, b->c.level + 1);
2246         if (ret)
2247                 return ret;
2248
2249         closure_init_stack(&cl);
2250
2251         /*
2252          * check btree_ptr_hash_val() after @b is locked by
2253          * btree_iter_traverse():
2254          */
2255         if (btree_ptr_hash_val(new_key) != b->hash_val) {
2256                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2257                 if (ret) {
2258                         ret = drop_locks_do(trans, (closure_sync(&cl), 0));
2259                         if (ret)
2260                                 return ret;
2261                 }
2262
2263                 new_hash = bch2_btree_node_mem_alloc(trans, false);
2264         }
2265
2266         path->intent_ref++;
2267         ret = __bch2_btree_node_update_key(trans, iter, b, new_hash, new_key,
2268                                            commit_flags, skip_triggers);
2269         --path->intent_ref;
2270
2271         if (new_hash) {
2272                 mutex_lock(&c->btree_cache.lock);
2273                 list_move(&new_hash->list, &c->btree_cache.freeable);
2274                 mutex_unlock(&c->btree_cache.lock);
2275
2276                 six_unlock_write(&new_hash->c.lock);
2277                 six_unlock_intent(&new_hash->c.lock);
2278         }
2279         closure_sync(&cl);
2280         bch2_btree_cache_cannibalize_unlock(c);
2281         return ret;
2282 }
2283
2284 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
2285                                         struct btree *b, struct bkey_i *new_key,
2286                                         unsigned commit_flags, bool skip_triggers)
2287 {
2288         struct btree_iter iter;
2289         int ret;
2290
2291         bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p,
2292                                   BTREE_MAX_DEPTH, b->c.level,
2293                                   BTREE_ITER_INTENT);
2294         ret = bch2_btree_iter_traverse(&iter);
2295         if (ret)
2296                 goto out;
2297
2298         /* has node been freed? */
2299         if (iter.path->l[b->c.level].b != b) {
2300                 /* node has been freed: */
2301                 BUG_ON(!btree_node_dying(b));
2302                 goto out;
2303         }
2304
2305         BUG_ON(!btree_node_hashed(b));
2306
2307         ret = bch2_btree_node_update_key(trans, &iter, b, new_key,
2308                                          commit_flags, skip_triggers);
2309 out:
2310         bch2_trans_iter_exit(trans, &iter);
2311         return ret;
2312 }
2313
2314 /* Init code: */
2315
2316 /*
2317  * Only for filesystem bringup, when first reading the btree roots or allocating
2318  * btree roots when initializing a new filesystem:
2319  */
2320 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2321 {
2322         BUG_ON(btree_node_root(c, b));
2323
2324         bch2_btree_set_root_inmem(c, b);
2325 }
2326
2327 static int __bch2_btree_root_alloc(struct btree_trans *trans, enum btree_id id)
2328 {
2329         struct bch_fs *c = trans->c;
2330         struct closure cl;
2331         struct btree *b;
2332         int ret;
2333
2334         closure_init_stack(&cl);
2335
2336         do {
2337                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2338                 closure_sync(&cl);
2339         } while (ret);
2340
2341         b = bch2_btree_node_mem_alloc(trans, false);
2342         bch2_btree_cache_cannibalize_unlock(c);
2343
2344         set_btree_node_fake(b);
2345         set_btree_node_need_rewrite(b);
2346         b->c.level      = 0;
2347         b->c.btree_id   = id;
2348
2349         bkey_btree_ptr_init(&b->key);
2350         b->key.k.p = SPOS_MAX;
2351         *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2352
2353         bch2_bset_init_first(b, &b->data->keys);
2354         bch2_btree_build_aux_trees(b);
2355
2356         b->data->flags = 0;
2357         btree_set_min(b, POS_MIN);
2358         btree_set_max(b, SPOS_MAX);
2359         b->data->format = bch2_btree_calc_format(b);
2360         btree_node_set_format(b, b->data->format);
2361
2362         ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2363                                           b->c.level, b->c.btree_id);
2364         BUG_ON(ret);
2365
2366         bch2_btree_set_root_inmem(c, b);
2367
2368         six_unlock_write(&b->c.lock);
2369         six_unlock_intent(&b->c.lock);
2370         return 0;
2371 }
2372
2373 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2374 {
2375         bch2_trans_run(c, __bch2_btree_root_alloc(&trans, id));
2376 }
2377
2378 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2379 {
2380         struct btree_update *as;
2381
2382         mutex_lock(&c->btree_interior_update_lock);
2383         list_for_each_entry(as, &c->btree_interior_update_list, list)
2384                 prt_printf(out, "%p m %u w %u r %u j %llu\n",
2385                        as,
2386                        as->mode,
2387                        as->nodes_written,
2388                        atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2389                        as->journal.seq);
2390         mutex_unlock(&c->btree_interior_update_lock);
2391 }
2392
2393 static bool bch2_btree_interior_updates_pending(struct bch_fs *c)
2394 {
2395         bool ret;
2396
2397         mutex_lock(&c->btree_interior_update_lock);
2398         ret = !list_empty(&c->btree_interior_update_list);
2399         mutex_unlock(&c->btree_interior_update_lock);
2400
2401         return ret;
2402 }
2403
2404 bool bch2_btree_interior_updates_flush(struct bch_fs *c)
2405 {
2406         bool ret = bch2_btree_interior_updates_pending(c);
2407
2408         if (ret)
2409                 closure_wait_event(&c->btree_interior_update_wait,
2410                                    !bch2_btree_interior_updates_pending(c));
2411         return ret;
2412 }
2413
2414 void bch2_journal_entry_to_btree_root(struct bch_fs *c, struct jset_entry *entry)
2415 {
2416         struct btree_root *r = bch2_btree_id_root(c, entry->btree_id);
2417
2418         mutex_lock(&c->btree_root_lock);
2419
2420         r->level = entry->level;
2421         r->alive = true;
2422         bkey_copy(&r->key, &entry->start[0]);
2423
2424         mutex_unlock(&c->btree_root_lock);
2425 }
2426
2427 struct jset_entry *
2428 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2429                                     struct jset_entry *start,
2430                                     struct jset_entry *end)
2431 {
2432         struct jset_entry *entry;
2433         unsigned long have = 0;
2434         unsigned i;
2435
2436         for (entry = start; entry < end; entry = vstruct_next(entry))
2437                 if (entry->type == BCH_JSET_ENTRY_btree_root)
2438                         __set_bit(entry->btree_id, &have);
2439
2440         mutex_lock(&c->btree_root_lock);
2441
2442         for (i = 0; i < btree_id_nr_alive(c); i++) {
2443                 struct btree_root *r = bch2_btree_id_root(c, i);
2444
2445                 if (r->alive && !test_bit(i, &have)) {
2446                         journal_entry_set(end, BCH_JSET_ENTRY_btree_root,
2447                                           i, r->level, &r->key, r->key.k.u64s);
2448                         end = vstruct_next(end);
2449                 }
2450         }
2451
2452         mutex_unlock(&c->btree_root_lock);
2453
2454         return end;
2455 }
2456
2457 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2458 {
2459         if (c->btree_interior_update_worker)
2460                 destroy_workqueue(c->btree_interior_update_worker);
2461         mempool_exit(&c->btree_interior_update_pool);
2462 }
2463
2464 void bch2_fs_btree_interior_update_init_early(struct bch_fs *c)
2465 {
2466         mutex_init(&c->btree_reserve_cache_lock);
2467         INIT_LIST_HEAD(&c->btree_interior_update_list);
2468         INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2469         mutex_init(&c->btree_interior_update_lock);
2470         INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2471
2472         INIT_LIST_HEAD(&c->pending_node_rewrites);
2473         mutex_init(&c->pending_node_rewrites_lock);
2474 }
2475
2476 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2477 {
2478         c->btree_interior_update_worker =
2479                 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2480         if (!c->btree_interior_update_worker)
2481                 return -BCH_ERR_ENOMEM_btree_interior_update_worker_init;
2482
2483         if (mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2484                                       sizeof(struct btree_update)))
2485                 return -BCH_ERR_ENOMEM_btree_interior_update_pool_init;
2486
2487         return 0;
2488 }