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_journal_iter.h"
   9 #include "btree_update.h"
  10 #include "btree_update_interior.h"
  11 #include "btree_io.h"
  12 #include "btree_iter.h"
  13 #include "btree_locking.h"
  14 #include "buckets.h"
  15 #include "clock.h"
  16 #include "error.h"
  17 #include "extents.h"
  18 #include "journal.h"
  19 #include "journal_reclaim.h"
  20 #include "keylist.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                                   btree_path_idx_t, struct btree *,
  29                                   struct keylist *, unsigned);
  30 static void bch2_btree_update_add_new_node(struct btree_update *, struct btree *);
  31
  32 static btree_path_idx_t get_unlocked_mut_path(struct btree_trans *trans,
  33                                               enum btree_id btree_id,
  34                                               unsigned level,
  35                                               struct bpos pos)
  36 {
  37         btree_path_idx_t path_idx = bch2_path_get(trans, btree_id, pos, level + 1, level,
  38                              BTREE_ITER_NOPRESERVE|
  39                              BTREE_ITER_INTENT, _RET_IP_);
  40         path_idx = bch2_btree_path_make_mut(trans, path_idx, true, _RET_IP_);
  41
  42         struct btree_path *path = trans->paths + path_idx;
  43         bch2_btree_path_downgrade(trans, path);
  44         __bch2_btree_path_unlock(trans, path);
  45         return path_idx;
  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  * bch2_btree_node_format_fits - check if we could rewrite node with a new format
 147  *
 148  * @c:          filesystem handle
 149  * @b:          btree node to rewrite
 150  * @new_f:      bkey format to translate keys to
 151  *
 152  * Returns: true if all re-packed keys will be able to fit in a new node.
 153  *
 154  * Assumes all keys will successfully pack with the new format.
 155  */
 156 bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
 157                                  struct bkey_format *new_f)
 158 {
 159         size_t u64s = btree_node_u64s_with_format(b, new_f);
 160
 161         return __vstruct_bytes(struct btree_node, u64s) < btree_bytes(c);
 162 }
 163
 164 /* Btree node freeing/allocation: */
 165
 166 static void __btree_node_free(struct btree_trans *trans, struct btree *b)
 167 {
 168         struct bch_fs *c = trans->c;
 169
 170         trace_and_count(c, btree_node_free, trans, b);
 171
 172         BUG_ON(btree_node_write_blocked(b));
 173         BUG_ON(btree_node_dirty(b));
 174         BUG_ON(btree_node_need_write(b));
 175         BUG_ON(b == btree_node_root(c, b));
 176         BUG_ON(b->ob.nr);
 177         BUG_ON(!list_empty(&b->write_blocked));
 178         BUG_ON(b->will_make_reachable);
 179
 180         clear_btree_node_noevict(b);
 181
 182         mutex_lock(&c->btree_cache.lock);
 183         list_move(&b->list, &c->btree_cache.freeable);
 184         mutex_unlock(&c->btree_cache.lock);
 185 }
 186
 187 static void bch2_btree_node_free_inmem(struct btree_trans *trans,
 188                                        struct btree_path *path,
 189                                        struct btree *b)
 190 {
 191         struct bch_fs *c = trans->c;
 192         unsigned i, level = b->c.level;
 193
 194         bch2_btree_node_lock_write_nofail(trans, path, &b->c);
 195         bch2_btree_node_hash_remove(&c->btree_cache, b);
 196         __btree_node_free(trans, b);
 197         six_unlock_write(&b->c.lock);
 198         mark_btree_node_locked_noreset(path, level, BTREE_NODE_INTENT_LOCKED);
 199
 200         trans_for_each_path(trans, path, i)
 201                 if (path->l[level].b == b) {
 202                         btree_node_unlock(trans, path, level);
 203                         path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
 204                 }
 205 }
 206
 207 static void bch2_btree_node_free_never_used(struct btree_update *as,
 208                                             struct btree_trans *trans,
 209                                             struct btree *b)
 210 {
 211         struct bch_fs *c = as->c;
 212         struct prealloc_nodes *p = &as->prealloc_nodes[b->c.lock.readers != NULL];
 213         struct btree_path *path;
 214         unsigned i, level = b->c.level;
 215
 216         BUG_ON(!list_empty(&b->write_blocked));
 217         BUG_ON(b->will_make_reachable != (1UL|(unsigned long) as));
 218
 219         b->will_make_reachable = 0;
 220         closure_put(&as->cl);
 221
 222         clear_btree_node_will_make_reachable(b);
 223         clear_btree_node_accessed(b);
 224         clear_btree_node_dirty_acct(c, b);
 225         clear_btree_node_need_write(b);
 226
 227         mutex_lock(&c->btree_cache.lock);
 228         list_del_init(&b->list);
 229         bch2_btree_node_hash_remove(&c->btree_cache, b);
 230         mutex_unlock(&c->btree_cache.lock);
 231
 232         BUG_ON(p->nr >= ARRAY_SIZE(p->b));
 233         p->b[p->nr++] = b;
 234
 235         six_unlock_intent(&b->c.lock);
 236
 237         trans_for_each_path(trans, path, i)
 238                 if (path->l[level].b == b) {
 239                         btree_node_unlock(trans, path, level);
 240                         path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
 241                 }
 242 }
 243
 244 static struct btree *__bch2_btree_node_alloc(struct btree_trans *trans,
 245                                              struct disk_reservation *res,
 246                                              struct closure *cl,
 247                                              bool interior_node,
 248                                              unsigned flags)
 249 {
 250         struct bch_fs *c = trans->c;
 251         struct write_point *wp;
 252         struct btree *b;
 253         BKEY_PADDED_ONSTACK(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
 254         struct open_buckets obs = { .nr = 0 };
 255         struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
 256         enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
 257         unsigned nr_reserve = watermark > BCH_WATERMARK_reclaim
 258                 ? BTREE_NODE_RESERVE
 259                 : 0;
 260         int ret;
 261
 262         mutex_lock(&c->btree_reserve_cache_lock);
 263         if (c->btree_reserve_cache_nr > nr_reserve) {
 264                 struct btree_alloc *a =
 265                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
 266
 267                 obs = a->ob;
 268                 bkey_copy(&tmp.k, &a->k);
 269                 mutex_unlock(&c->btree_reserve_cache_lock);
 270                 goto mem_alloc;
 271         }
 272         mutex_unlock(&c->btree_reserve_cache_lock);
 273
 274 retry:
 275         ret = bch2_alloc_sectors_start_trans(trans,
 276                                       c->opts.metadata_target ?:
 277                                       c->opts.foreground_target,
 278                                       0,
 279                                       writepoint_ptr(&c->btree_write_point),
 280                                       &devs_have,
 281                                       res->nr_replicas,
 282                                       c->opts.metadata_replicas_required,
 283                                       watermark, 0, cl, &wp);
 284         if (unlikely(ret))
 285                 return ERR_PTR(ret);
 286
 287         if (wp->sectors_free < btree_sectors(c)) {
 288                 struct open_bucket *ob;
 289                 unsigned i;
 290
 291                 open_bucket_for_each(c, &wp->ptrs, ob, i)
 292                         if (ob->sectors_free < btree_sectors(c))
 293                                 ob->sectors_free = 0;
 294
 295                 bch2_alloc_sectors_done(c, wp);
 296                 goto retry;
 297         }
 298
 299         bkey_btree_ptr_v2_init(&tmp.k);
 300         bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, btree_sectors(c), false);
 301
 302         bch2_open_bucket_get(c, wp, &obs);
 303         bch2_alloc_sectors_done(c, wp);
 304 mem_alloc:
 305         b = bch2_btree_node_mem_alloc(trans, interior_node);
 306         six_unlock_write(&b->c.lock);
 307         six_unlock_intent(&b->c.lock);
 308
 309         /* we hold cannibalize_lock: */
 310         BUG_ON(IS_ERR(b));
 311         BUG_ON(b->ob.nr);
 312
 313         bkey_copy(&b->key, &tmp.k);
 314         b->ob = obs;
 315
 316         return b;
 317 }
 318
 319 static struct btree *bch2_btree_node_alloc(struct btree_update *as,
 320                                            struct btree_trans *trans,
 321                                            unsigned level)
 322 {
 323         struct bch_fs *c = as->c;
 324         struct btree *b;
 325         struct prealloc_nodes *p = &as->prealloc_nodes[!!level];
 326         int ret;
 327
 328         BUG_ON(level >= BTREE_MAX_DEPTH);
 329         BUG_ON(!p->nr);
 330
 331         b = p->b[--p->nr];
 332
 333         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
 334         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
 335
 336         set_btree_node_accessed(b);
 337         set_btree_node_dirty_acct(c, b);
 338         set_btree_node_need_write(b);
 339
 340         bch2_bset_init_first(b, &b->data->keys);
 341         b->c.level      = level;
 342         b->c.btree_id   = as->btree_id;
 343         b->version_ondisk = c->sb.version;
 344
 345         memset(&b->nr, 0, sizeof(b->nr));
 346         b->data->magic = cpu_to_le64(bset_magic(c));
 347         memset(&b->data->_ptr, 0, sizeof(b->data->_ptr));
 348         b->data->flags = 0;
 349         SET_BTREE_NODE_ID(b->data, as->btree_id);
 350         SET_BTREE_NODE_LEVEL(b->data, level);
 351
 352         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 353                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
 354
 355                 bp->v.mem_ptr           = 0;
 356                 bp->v.seq               = b->data->keys.seq;
 357                 bp->v.sectors_written   = 0;
 358         }
 359
 360         SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
 361
 362         bch2_btree_build_aux_trees(b);
 363
 364         ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
 365         BUG_ON(ret);
 366
 367         trace_and_count(c, btree_node_alloc, trans, b);
 368         bch2_increment_clock(c, btree_sectors(c), WRITE);
 369         return b;
 370 }
 371
 372 static void btree_set_min(struct btree *b, struct bpos pos)
 373 {
 374         if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
 375                 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
 376         b->data->min_key = pos;
 377 }
 378
 379 static void btree_set_max(struct btree *b, struct bpos pos)
 380 {
 381         b->key.k.p = pos;
 382         b->data->max_key = pos;
 383 }
 384
 385 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
 386                                                        struct btree_trans *trans,
 387                                                        struct btree *b)
 388 {
 389         struct btree *n = bch2_btree_node_alloc(as, trans, b->c.level);
 390         struct bkey_format format = bch2_btree_calc_format(b);
 391
 392         /*
 393          * The keys might expand with the new format - if they wouldn't fit in
 394          * the btree node anymore, use the old format for now:
 395          */
 396         if (!bch2_btree_node_format_fits(as->c, b, &format))
 397                 format = b->format;
 398
 399         SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
 400
 401         btree_set_min(n, b->data->min_key);
 402         btree_set_max(n, b->data->max_key);
 403
 404         n->data->format         = format;
 405         btree_node_set_format(n, format);
 406
 407         bch2_btree_sort_into(as->c, n, b);
 408
 409         btree_node_reset_sib_u64s(n);
 410         return n;
 411 }
 412
 413 static struct btree *__btree_root_alloc(struct btree_update *as,
 414                                 struct btree_trans *trans, unsigned level)
 415 {
 416         struct btree *b = bch2_btree_node_alloc(as, trans, level);
 417
 418         btree_set_min(b, POS_MIN);
 419         btree_set_max(b, SPOS_MAX);
 420         b->data->format = bch2_btree_calc_format(b);
 421
 422         btree_node_set_format(b, b->data->format);
 423         bch2_btree_build_aux_trees(b);
 424
 425         return b;
 426 }
 427
 428 static void bch2_btree_reserve_put(struct btree_update *as, struct btree_trans *trans)
 429 {
 430         struct bch_fs *c = as->c;
 431         struct prealloc_nodes *p;
 432
 433         for (p = as->prealloc_nodes;
 434              p < as->prealloc_nodes + ARRAY_SIZE(as->prealloc_nodes);
 435              p++) {
 436                 while (p->nr) {
 437                         struct btree *b = p->b[--p->nr];
 438
 439                         mutex_lock(&c->btree_reserve_cache_lock);
 440
 441                         if (c->btree_reserve_cache_nr <
 442                             ARRAY_SIZE(c->btree_reserve_cache)) {
 443                                 struct btree_alloc *a =
 444                                         &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
 445
 446                                 a->ob = b->ob;
 447                                 b->ob.nr = 0;
 448                                 bkey_copy(&a->k, &b->key);
 449                         } else {
 450                                 bch2_open_buckets_put(c, &b->ob);
 451                         }
 452
 453                         mutex_unlock(&c->btree_reserve_cache_lock);
 454
 455                         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
 456                         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
 457                         __btree_node_free(trans, b);
 458                         six_unlock_write(&b->c.lock);
 459                         six_unlock_intent(&b->c.lock);
 460                 }
 461         }
 462 }
 463
 464 static int bch2_btree_reserve_get(struct btree_trans *trans,
 465                                   struct btree_update *as,
 466                                   unsigned nr_nodes[2],
 467                                   unsigned flags,
 468                                   struct closure *cl)
 469 {
 470         struct btree *b;
 471         unsigned interior;
 472         int ret = 0;
 473
 474         BUG_ON(nr_nodes[0] + nr_nodes[1] > BTREE_RESERVE_MAX);
 475
 476         /*
 477          * Protects reaping from the btree node cache and using the btree node
 478          * open bucket reserve:
 479          */
 480         ret = bch2_btree_cache_cannibalize_lock(trans, cl);
 481         if (ret)
 482                 return ret;
 483
 484         for (interior = 0; interior < 2; interior++) {
 485                 struct prealloc_nodes *p = as->prealloc_nodes + interior;
 486
 487                 while (p->nr < nr_nodes[interior]) {
 488                         b = __bch2_btree_node_alloc(trans, &as->disk_res, cl,
 489                                                     interior, flags);
 490                         if (IS_ERR(b)) {
 491                                 ret = PTR_ERR(b);
 492                                 goto err;
 493                         }
 494
 495                         p->b[p->nr++] = b;
 496                 }
 497         }
 498 err:
 499         bch2_btree_cache_cannibalize_unlock(trans);
 500         return ret;
 501 }
 502
 503 /* Asynchronous interior node update machinery */
 504
 505 static void bch2_btree_update_free(struct btree_update *as, struct btree_trans *trans)
 506 {
 507         struct bch_fs *c = as->c;
 508
 509         if (as->took_gc_lock)
 510                 up_read(&c->gc_lock);
 511         as->took_gc_lock = false;
 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
 560         struct jset_entry *e = bch2_trans_jset_entry_alloc(trans, as->journal_u64s);
 561         int ret = PTR_ERR_OR_ZERO(e);
 562         if (ret)
 563                 return ret;
 564
 565         memcpy(e, as->journal_entries, as->journal_u64s * sizeof(u64));
 566
 567         trans->journal_pin = &as->journal;
 568
 569         for_each_keylist_key(&as->old_keys, k) {
 570                 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
 571
 572                 ret = bch2_trans_mark_old(trans, as->btree_id, level, bkey_i_to_s_c(k), 0);
 573                 if (ret)
 574                         return ret;
 575         }
 576
 577         for_each_keylist_key(&as->new_keys, k) {
 578                 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
 579
 580                 ret = bch2_trans_mark_new(trans, as->btree_id, level, k, 0);
 581                 if (ret)
 582                         return ret;
 583         }
 584
 585         return 0;
 586 }
 587
 588 static void btree_update_nodes_written(struct btree_update *as)
 589 {
 590         struct bch_fs *c = as->c;
 591         struct btree *b;
 592         struct btree_trans *trans = bch2_trans_get(c);
 593         u64 journal_seq = 0;
 594         unsigned i;
 595         int ret;
 596
 597         /*
 598          * If we're already in an error state, it might be because a btree node
 599          * was never written, and we might be trying to free that same btree
 600          * node here, but it won't have been marked as allocated and we'll see
 601          * spurious disk usage inconsistencies in the transactional part below
 602          * if we don't skip it:
 603          */
 604         ret = bch2_journal_error(&c->journal);
 605         if (ret)
 606                 goto err;
 607
 608         /*
 609          * Wait for any in flight writes to finish before we free the old nodes
 610          * on disk:
 611          */
 612         for (i = 0; i < as->nr_old_nodes; i++) {
 613                 __le64 seq;
 614
 615                 b = as->old_nodes[i];
 616
 617                 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
 618                 seq = b->data ? b->data->keys.seq : 0;
 619                 six_unlock_read(&b->c.lock);
 620
 621                 if (seq == as->old_nodes_seq[i])
 622                         wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight_inner,
 623                                        TASK_UNINTERRUPTIBLE);
 624         }
 625
 626         /*
 627          * We did an update to a parent node where the pointers we added pointed
 628          * to child nodes that weren't written yet: now, the child nodes have
 629          * been written so we can write out the update to the interior node.
 630          */
 631
 632         /*
 633          * We can't call into journal reclaim here: we'd block on the journal
 634          * reclaim lock, but we may need to release the open buckets we have
 635          * pinned in order for other btree updates to make forward progress, and
 636          * journal reclaim does btree updates when flushing bkey_cached entries,
 637          * which may require allocations as well.
 638          */
 639         ret = commit_do(trans, &as->disk_res, &journal_seq,
 640                         BCH_WATERMARK_reclaim|
 641                         BCH_TRANS_COMMIT_no_enospc|
 642                         BCH_TRANS_COMMIT_no_check_rw|
 643                         BCH_TRANS_COMMIT_journal_reclaim,
 644                         btree_update_nodes_written_trans(trans, as));
 645         bch2_trans_unlock(trans);
 646
 647         bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
 648                              "%s(): error %s", __func__, bch2_err_str(ret));
 649 err:
 650         if (as->b) {
 651
 652                 b = as->b;
 653                 btree_path_idx_t path_idx = get_unlocked_mut_path(trans,
 654                                                 as->btree_id, b->c.level, b->key.k.p);
 655                 struct btree_path *path = trans->paths + path_idx;
 656                 /*
 657                  * @b is the node we did the final insert into:
 658                  *
 659                  * On failure to get a journal reservation, we still have to
 660                  * unblock the write and allow most of the write path to happen
 661                  * so that shutdown works, but the i->journal_seq mechanism
 662                  * won't work to prevent the btree write from being visible (we
 663                  * didn't get a journal sequence number) - instead
 664                  * __bch2_btree_node_write() doesn't do the actual write if
 665                  * we're in journal error state:
 666                  */
 667
 668                 /*
 669                  * Ensure transaction is unlocked before using
 670                  * btree_node_lock_nopath() (the use of which is always suspect,
 671                  * we need to work on removing this in the future)
 672                  *
 673                  * It should be, but get_unlocked_mut_path() -> bch2_path_get()
 674                  * calls bch2_path_upgrade(), before we call path_make_mut(), so
 675                  * we may rarely end up with a locked path besides the one we
 676                  * have here:
 677                  */
 678                 bch2_trans_unlock(trans);
 679                 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
 680                 mark_btree_node_locked(trans, path, b->c.level, BTREE_NODE_INTENT_LOCKED);
 681                 path->l[b->c.level].lock_seq = six_lock_seq(&b->c.lock);
 682                 path->l[b->c.level].b = b;
 683
 684                 bch2_btree_node_lock_write_nofail(trans, path, &b->c);
 685
 686                 mutex_lock(&c->btree_interior_update_lock);
 687
 688                 list_del(&as->write_blocked_list);
 689                 if (list_empty(&b->write_blocked))
 690                         clear_btree_node_write_blocked(b);
 691
 692                 /*
 693                  * Node might have been freed, recheck under
 694                  * btree_interior_update_lock:
 695                  */
 696                 if (as->b == b) {
 697                         BUG_ON(!b->c.level);
 698                         BUG_ON(!btree_node_dirty(b));
 699
 700                         if (!ret) {
 701                                 struct bset *last = btree_bset_last(b);
 702
 703                                 last->journal_seq = cpu_to_le64(
 704                                                              max(journal_seq,
 705                                                                  le64_to_cpu(last->journal_seq)));
 706
 707                                 bch2_btree_add_journal_pin(c, b, journal_seq);
 708                         } else {
 709                                 /*
 710                                  * If we didn't get a journal sequence number we
 711                                  * can't write this btree node, because recovery
 712                                  * won't know to ignore this write:
 713                                  */
 714                                 set_btree_node_never_write(b);
 715                         }
 716                 }
 717
 718                 mutex_unlock(&c->btree_interior_update_lock);
 719
 720                 mark_btree_node_locked_noreset(path, b->c.level, BTREE_NODE_INTENT_LOCKED);
 721                 six_unlock_write(&b->c.lock);
 722
 723                 btree_node_write_if_need(c, b, SIX_LOCK_intent);
 724                 btree_node_unlock(trans, path, b->c.level);
 725                 bch2_path_put(trans, path_idx, true);
 726         }
 727
 728         bch2_journal_pin_drop(&c->journal, &as->journal);
 729
 730         mutex_lock(&c->btree_interior_update_lock);
 731         for (i = 0; i < as->nr_new_nodes; i++) {
 732                 b = as->new_nodes[i];
 733
 734                 BUG_ON(b->will_make_reachable != (unsigned long) as);
 735                 b->will_make_reachable = 0;
 736                 clear_btree_node_will_make_reachable(b);
 737         }
 738         mutex_unlock(&c->btree_interior_update_lock);
 739
 740         for (i = 0; i < as->nr_new_nodes; i++) {
 741                 b = as->new_nodes[i];
 742
 743                 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
 744                 btree_node_write_if_need(c, b, SIX_LOCK_read);
 745                 six_unlock_read(&b->c.lock);
 746         }
 747
 748         for (i = 0; i < as->nr_open_buckets; i++)
 749                 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
 750
 751         bch2_btree_update_free(as, trans);
 752         bch2_trans_put(trans);
 753 }
 754
 755 static void btree_interior_update_work(struct work_struct *work)
 756 {
 757         struct bch_fs *c =
 758                 container_of(work, struct bch_fs, btree_interior_update_work);
 759         struct btree_update *as;
 760
 761         while (1) {
 762                 mutex_lock(&c->btree_interior_update_lock);
 763                 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
 764                                               struct btree_update, unwritten_list);
 765                 if (as && !as->nodes_written)
 766                         as = NULL;
 767                 mutex_unlock(&c->btree_interior_update_lock);
 768
 769                 if (!as)
 770                         break;
 771
 772                 btree_update_nodes_written(as);
 773         }
 774 }
 775
 776 static CLOSURE_CALLBACK(btree_update_set_nodes_written)
 777 {
 778         closure_type(as, struct btree_update, cl);
 779         struct bch_fs *c = as->c;
 780
 781         mutex_lock(&c->btree_interior_update_lock);
 782         as->nodes_written = true;
 783         mutex_unlock(&c->btree_interior_update_lock);
 784
 785         queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
 786 }
 787
 788 /*
 789  * We're updating @b with pointers to nodes that haven't finished writing yet:
 790  * block @b from being written until @as completes
 791  */
 792 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
 793 {
 794         struct bch_fs *c = as->c;
 795
 796         mutex_lock(&c->btree_interior_update_lock);
 797         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 798
 799         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 800         BUG_ON(!btree_node_dirty(b));
 801         BUG_ON(!b->c.level);
 802
 803         as->mode        = BTREE_INTERIOR_UPDATING_NODE;
 804         as->b           = b;
 805
 806         set_btree_node_write_blocked(b);
 807         list_add(&as->write_blocked_list, &b->write_blocked);
 808
 809         mutex_unlock(&c->btree_interior_update_lock);
 810 }
 811
 812 static int bch2_update_reparent_journal_pin_flush(struct journal *j,
 813                                 struct journal_entry_pin *_pin, u64 seq)
 814 {
 815         return 0;
 816 }
 817
 818 static void btree_update_reparent(struct btree_update *as,
 819                                   struct btree_update *child)
 820 {
 821         struct bch_fs *c = as->c;
 822
 823         lockdep_assert_held(&c->btree_interior_update_lock);
 824
 825         child->b = NULL;
 826         child->mode = BTREE_INTERIOR_UPDATING_AS;
 827
 828         bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal,
 829                               bch2_update_reparent_journal_pin_flush);
 830 }
 831
 832 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
 833 {
 834         struct bkey_i *insert = &b->key;
 835         struct bch_fs *c = as->c;
 836
 837         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 838
 839         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
 840                ARRAY_SIZE(as->journal_entries));
 841
 842         as->journal_u64s +=
 843                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
 844                                   BCH_JSET_ENTRY_btree_root,
 845                                   b->c.btree_id, b->c.level,
 846                                   insert, insert->k.u64s);
 847
 848         mutex_lock(&c->btree_interior_update_lock);
 849         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 850
 851         as->mode        = BTREE_INTERIOR_UPDATING_ROOT;
 852         mutex_unlock(&c->btree_interior_update_lock);
 853 }
 854
 855 /*
 856  * bch2_btree_update_add_new_node:
 857  *
 858  * This causes @as to wait on @b to be written, before it gets to
 859  * bch2_btree_update_nodes_written
 860  *
 861  * Additionally, it sets b->will_make_reachable to prevent any additional writes
 862  * to @b from happening besides the first until @b is reachable on disk
 863  *
 864  * And it adds @b to the list of @as's new nodes, so that we can update sector
 865  * counts in bch2_btree_update_nodes_written:
 866  */
 867 static void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
 868 {
 869         struct bch_fs *c = as->c;
 870
 871         closure_get(&as->cl);
 872
 873         mutex_lock(&c->btree_interior_update_lock);
 874         BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
 875         BUG_ON(b->will_make_reachable);
 876
 877         as->new_nodes[as->nr_new_nodes++] = b;
 878         b->will_make_reachable = 1UL|(unsigned long) as;
 879         set_btree_node_will_make_reachable(b);
 880
 881         mutex_unlock(&c->btree_interior_update_lock);
 882
 883         btree_update_add_key(as, &as->new_keys, b);
 884
 885         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 886                 unsigned bytes = vstruct_end(&b->data->keys) - (void *) b->data;
 887                 unsigned sectors = round_up(bytes, block_bytes(c)) >> 9;
 888
 889                 bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written =
 890                         cpu_to_le16(sectors);
 891         }
 892 }
 893
 894 /*
 895  * returns true if @b was a new node
 896  */
 897 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
 898 {
 899         struct btree_update *as;
 900         unsigned long v;
 901         unsigned i;
 902
 903         mutex_lock(&c->btree_interior_update_lock);
 904         /*
 905          * When b->will_make_reachable != 0, it owns a ref on as->cl that's
 906          * dropped when it gets written by bch2_btree_complete_write - the
 907          * xchg() is for synchronization with bch2_btree_complete_write:
 908          */
 909         v = xchg(&b->will_make_reachable, 0);
 910         clear_btree_node_will_make_reachable(b);
 911         as = (struct btree_update *) (v & ~1UL);
 912
 913         if (!as) {
 914                 mutex_unlock(&c->btree_interior_update_lock);
 915                 return;
 916         }
 917
 918         for (i = 0; i < as->nr_new_nodes; i++)
 919                 if (as->new_nodes[i] == b)
 920                         goto found;
 921
 922         BUG();
 923 found:
 924         array_remove_item(as->new_nodes, as->nr_new_nodes, i);
 925         mutex_unlock(&c->btree_interior_update_lock);
 926
 927         if (v & 1)
 928                 closure_put(&as->cl);
 929 }
 930
 931 static void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
 932 {
 933         while (b->ob.nr)
 934                 as->open_buckets[as->nr_open_buckets++] =
 935                         b->ob.v[--b->ob.nr];
 936 }
 937
 938 static int bch2_btree_update_will_free_node_journal_pin_flush(struct journal *j,
 939                                 struct journal_entry_pin *_pin, u64 seq)
 940 {
 941         return 0;
 942 }
 943
 944 /*
 945  * @b is being split/rewritten: it may have pointers to not-yet-written btree
 946  * nodes and thus outstanding btree_updates - redirect @b's
 947  * btree_updates to point to this btree_update:
 948  */
 949 static void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 950                                                       struct btree *b)
 951 {
 952         struct bch_fs *c = as->c;
 953         struct btree_update *p, *n;
 954         struct btree_write *w;
 955
 956         set_btree_node_dying(b);
 957
 958         if (btree_node_fake(b))
 959                 return;
 960
 961         mutex_lock(&c->btree_interior_update_lock);
 962
 963         /*
 964          * Does this node have any btree_update operations preventing
 965          * it from being written?
 966          *
 967          * If so, redirect them to point to this btree_update: we can
 968          * write out our new nodes, but we won't make them visible until those
 969          * operations complete
 970          */
 971         list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
 972                 list_del_init(&p->write_blocked_list);
 973                 btree_update_reparent(as, p);
 974
 975                 /*
 976                  * for flush_held_btree_writes() waiting on updates to flush or
 977                  * nodes to be writeable:
 978                  */
 979                 closure_wake_up(&c->btree_interior_update_wait);
 980         }
 981
 982         clear_btree_node_dirty_acct(c, b);
 983         clear_btree_node_need_write(b);
 984         clear_btree_node_write_blocked(b);
 985
 986         /*
 987          * Does this node have unwritten data that has a pin on the journal?
 988          *
 989          * If so, transfer that pin to the btree_update operation -
 990          * note that if we're freeing multiple nodes, we only need to keep the
 991          * oldest pin of any of the nodes we're freeing. We'll release the pin
 992          * when the new nodes are persistent and reachable on disk:
 993          */
 994         w = btree_current_write(b);
 995         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal,
 996                               bch2_btree_update_will_free_node_journal_pin_flush);
 997         bch2_journal_pin_drop(&c->journal, &w->journal);
 998
 999         w = btree_prev_write(b);
1000         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal,
1001                               bch2_btree_update_will_free_node_journal_pin_flush);
1002         bch2_journal_pin_drop(&c->journal, &w->journal);
1003
1004         mutex_unlock(&c->btree_interior_update_lock);
1005
1006         /*
1007          * Is this a node that isn't reachable on disk yet?
1008          *
1009          * Nodes that aren't reachable yet have writes blocked until they're
1010          * reachable - now that we've cancelled any pending writes and moved
1011          * things waiting on that write to wait on this update, we can drop this
1012          * node from the list of nodes that the other update is making
1013          * reachable, prior to freeing it:
1014          */
1015         btree_update_drop_new_node(c, b);
1016
1017         btree_update_add_key(as, &as->old_keys, b);
1018
1019         as->old_nodes[as->nr_old_nodes] = b;
1020         as->old_nodes_seq[as->nr_old_nodes] = b->data->keys.seq;
1021         as->nr_old_nodes++;
1022 }
1023
1024 static void bch2_btree_update_done(struct btree_update *as, struct btree_trans *trans)
1025 {
1026         struct bch_fs *c = as->c;
1027         u64 start_time = as->start_time;
1028
1029         BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
1030
1031         if (as->took_gc_lock)
1032                 up_read(&as->c->gc_lock);
1033         as->took_gc_lock = false;
1034
1035         bch2_btree_reserve_put(as, trans);
1036
1037         continue_at(&as->cl, btree_update_set_nodes_written,
1038                     as->c->btree_interior_update_worker);
1039
1040         bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_foreground],
1041                                start_time);
1042 }
1043
1044 static struct btree_update *
1045 bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
1046                         unsigned level, bool split, unsigned flags)
1047 {
1048         struct bch_fs *c = trans->c;
1049         struct btree_update *as;
1050         u64 start_time = local_clock();
1051         int disk_res_flags = (flags & BCH_TRANS_COMMIT_no_enospc)
1052                 ? BCH_DISK_RESERVATION_NOFAIL : 0;
1053         unsigned nr_nodes[2] = { 0, 0 };
1054         unsigned update_level = level;
1055         enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
1056         int ret = 0;
1057         u32 restart_count = trans->restart_count;
1058
1059         BUG_ON(!path->should_be_locked);
1060
1061         if (watermark == BCH_WATERMARK_copygc)
1062                 watermark = BCH_WATERMARK_btree_copygc;
1063         if (watermark < BCH_WATERMARK_btree)
1064                 watermark = BCH_WATERMARK_btree;
1065
1066         flags &= ~BCH_WATERMARK_MASK;
1067         flags |= watermark;
1068
1069         if (!(flags & BCH_TRANS_COMMIT_journal_reclaim) &&
1070             watermark < c->journal.watermark) {
1071                 struct journal_res res = { 0 };
1072
1073                 ret = drop_locks_do(trans,
1074                         bch2_journal_res_get(&c->journal, &res, 1,
1075                                              watermark|JOURNAL_RES_GET_CHECK));
1076                 if (ret)
1077                         return ERR_PTR(ret);
1078         }
1079
1080         while (1) {
1081                 nr_nodes[!!update_level] += 1 + split;
1082                 update_level++;
1083
1084                 ret = bch2_btree_path_upgrade(trans, path, update_level + 1);
1085                 if (ret)
1086                         return ERR_PTR(ret);
1087
1088                 if (!btree_path_node(path, update_level)) {
1089                         /* Allocating new root? */
1090                         nr_nodes[1] += split;
1091                         update_level = BTREE_MAX_DEPTH;
1092                         break;
1093                 }
1094
1095                 /*
1096                  * Always check for space for two keys, even if we won't have to
1097                  * split at prior level - it might have been a merge instead:
1098                  */
1099                 if (bch2_btree_node_insert_fits(c, path->l[update_level].b,
1100                                                 BKEY_BTREE_PTR_U64s_MAX * 2))
1101                         break;
1102
1103                 split = path->l[update_level].b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c);
1104         }
1105
1106         if (!down_read_trylock(&c->gc_lock)) {
1107                 ret = drop_locks_do(trans, (down_read(&c->gc_lock), 0));
1108                 if (ret) {
1109                         up_read(&c->gc_lock);
1110                         return ERR_PTR(ret);
1111                 }
1112         }
1113
1114         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOFS);
1115         memset(as, 0, sizeof(*as));
1116         closure_init(&as->cl, NULL);
1117         as->c           = c;
1118         as->start_time  = start_time;
1119         as->mode        = BTREE_INTERIOR_NO_UPDATE;
1120         as->took_gc_lock = true;
1121         as->btree_id    = path->btree_id;
1122         as->update_level = update_level;
1123         INIT_LIST_HEAD(&as->list);
1124         INIT_LIST_HEAD(&as->unwritten_list);
1125         INIT_LIST_HEAD(&as->write_blocked_list);
1126         bch2_keylist_init(&as->old_keys, as->_old_keys);
1127         bch2_keylist_init(&as->new_keys, as->_new_keys);
1128         bch2_keylist_init(&as->parent_keys, as->inline_keys);
1129
1130         mutex_lock(&c->btree_interior_update_lock);
1131         list_add_tail(&as->list, &c->btree_interior_update_list);
1132         mutex_unlock(&c->btree_interior_update_lock);
1133
1134         /*
1135          * We don't want to allocate if we're in an error state, that can cause
1136          * deadlock on emergency shutdown due to open buckets getting stuck in
1137          * the btree_reserve_cache after allocator shutdown has cleared it out.
1138          * This check needs to come after adding us to the btree_interior_update
1139          * list but before calling bch2_btree_reserve_get, to synchronize with
1140          * __bch2_fs_read_only().
1141          */
1142         ret = bch2_journal_error(&c->journal);
1143         if (ret)
1144                 goto err;
1145
1146         ret = bch2_disk_reservation_get(c, &as->disk_res,
1147                         (nr_nodes[0] + nr_nodes[1]) * btree_sectors(c),
1148                         c->opts.metadata_replicas,
1149                         disk_res_flags);
1150         if (ret)
1151                 goto err;
1152
1153         ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, NULL);
1154         if (bch2_err_matches(ret, ENOSPC) ||
1155             bch2_err_matches(ret, ENOMEM)) {
1156                 struct closure cl;
1157
1158                 /*
1159                  * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
1160                  * flag
1161                  */
1162                 if (bch2_err_matches(ret, ENOSPC) &&
1163                     (flags & BCH_TRANS_COMMIT_journal_reclaim) &&
1164                     watermark != BCH_WATERMARK_reclaim) {
1165                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
1166                         goto err;
1167                 }
1168
1169                 closure_init_stack(&cl);
1170
1171                 do {
1172                         ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, &cl);
1173
1174                         bch2_trans_unlock(trans);
1175                         closure_sync(&cl);
1176                 } while (bch2_err_matches(ret, BCH_ERR_operation_blocked));
1177         }
1178
1179         if (ret) {
1180                 trace_and_count(c, btree_reserve_get_fail, trans->fn,
1181                                 _RET_IP_, nr_nodes[0] + nr_nodes[1], ret);
1182                 goto err;
1183         }
1184
1185         ret = bch2_trans_relock(trans);
1186         if (ret)
1187                 goto err;
1188
1189         bch2_trans_verify_not_restarted(trans, restart_count);
1190         return as;
1191 err:
1192         bch2_btree_update_free(as, trans);
1193         return ERR_PTR(ret);
1194 }
1195
1196 /* Btree root updates: */
1197
1198 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1199 {
1200         /* Root nodes cannot be reaped */
1201         mutex_lock(&c->btree_cache.lock);
1202         list_del_init(&b->list);
1203         mutex_unlock(&c->btree_cache.lock);
1204
1205         mutex_lock(&c->btree_root_lock);
1206         BUG_ON(btree_node_root(c, b) &&
1207                (b->c.level < btree_node_root(c, b)->c.level ||
1208                 !btree_node_dying(btree_node_root(c, b))));
1209
1210         bch2_btree_id_root(c, b->c.btree_id)->b = b;
1211         mutex_unlock(&c->btree_root_lock);
1212
1213         bch2_recalc_btree_reserve(c);
1214 }
1215
1216 static void bch2_btree_set_root(struct btree_update *as,
1217                                 struct btree_trans *trans,
1218                                 struct btree_path *path,
1219                                 struct btree *b)
1220 {
1221         struct bch_fs *c = as->c;
1222         struct btree *old;
1223
1224         trace_and_count(c, btree_node_set_root, trans, b);
1225
1226         old = btree_node_root(c, b);
1227
1228         /*
1229          * Ensure no one is using the old root while we switch to the
1230          * new root:
1231          */
1232         bch2_btree_node_lock_write_nofail(trans, path, &old->c);
1233
1234         bch2_btree_set_root_inmem(c, b);
1235
1236         btree_update_updated_root(as, b);
1237
1238         /*
1239          * Unlock old root after new root is visible:
1240          *
1241          * The new root isn't persistent, but that's ok: we still have
1242          * an intent lock on the new root, and any updates that would
1243          * depend on the new root would have to update the new root.
1244          */
1245         bch2_btree_node_unlock_write(trans, path, old);
1246 }
1247
1248 /* Interior node updates: */
1249
1250 static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
1251                                         struct btree_trans *trans,
1252                                         struct btree_path *path,
1253                                         struct btree *b,
1254                                         struct btree_node_iter *node_iter,
1255                                         struct bkey_i *insert)
1256 {
1257         struct bch_fs *c = as->c;
1258         struct bkey_packed *k;
1259         struct printbuf buf = PRINTBUF;
1260         unsigned long old, new, v;
1261
1262         BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 &&
1263                !btree_ptr_sectors_written(insert));
1264
1265         if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
1266                 bch2_journal_key_overwritten(c, b->c.btree_id, b->c.level, insert->k.p);
1267
1268         if (bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1269                               btree_node_type(b), WRITE, &buf) ?:
1270             bch2_bkey_in_btree_node(c, b, bkey_i_to_s_c(insert), &buf)) {
1271                 printbuf_reset(&buf);
1272                 prt_printf(&buf, "inserting invalid bkey\n  ");
1273                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
1274                 prt_printf(&buf, "\n  ");
1275                 bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1276                                   btree_node_type(b), WRITE, &buf);
1277                 bch2_bkey_in_btree_node(c, b, bkey_i_to_s_c(insert), &buf);
1278
1279                 bch2_fs_inconsistent(c, "%s", buf.buf);
1280                 dump_stack();
1281         }
1282
1283         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1284                ARRAY_SIZE(as->journal_entries));
1285
1286         as->journal_u64s +=
1287                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1288                                   BCH_JSET_ENTRY_btree_keys,
1289                                   b->c.btree_id, b->c.level,
1290                                   insert, insert->k.u64s);
1291
1292         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1293                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1294                 bch2_btree_node_iter_advance(node_iter, b);
1295
1296         bch2_btree_bset_insert_key(trans, path, b, node_iter, insert);
1297         set_btree_node_dirty_acct(c, b);
1298
1299         v = READ_ONCE(b->flags);
1300         do {
1301                 old = new = v;
1302
1303                 new &= ~BTREE_WRITE_TYPE_MASK;
1304                 new |= BTREE_WRITE_interior;
1305                 new |= 1 << BTREE_NODE_need_write;
1306         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1307
1308         printbuf_exit(&buf);
1309 }
1310
1311 static void
1312 __bch2_btree_insert_keys_interior(struct btree_update *as,
1313                                   struct btree_trans *trans,
1314                                   struct btree_path *path,
1315                                   struct btree *b,
1316                                   struct btree_node_iter node_iter,
1317                                   struct keylist *keys)
1318 {
1319         struct bkey_i *insert = bch2_keylist_front(keys);
1320         struct bkey_packed *k;
1321
1322         BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1323
1324         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1325                (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1326                 ;
1327
1328         while (!bch2_keylist_empty(keys)) {
1329                 insert = bch2_keylist_front(keys);
1330
1331                 if (bpos_gt(insert->k.p, b->key.k.p))
1332                         break;
1333
1334                 bch2_insert_fixup_btree_ptr(as, trans, path, b, &node_iter, insert);
1335                 bch2_keylist_pop_front(keys);
1336         }
1337 }
1338
1339 /*
1340  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1341  * node)
1342  */
1343 static void __btree_split_node(struct btree_update *as,
1344                                struct btree_trans *trans,
1345                                struct btree *b,
1346                                struct btree *n[2])
1347 {
1348         struct bkey_packed *k;
1349         struct bpos n1_pos = POS_MIN;
1350         struct btree_node_iter iter;
1351         struct bset *bsets[2];
1352         struct bkey_format_state format[2];
1353         struct bkey_packed *out[2];
1354         struct bkey uk;
1355         unsigned u64s, n1_u64s = (b->nr.live_u64s * 3) / 5;
1356         int i;
1357
1358         for (i = 0; i < 2; i++) {
1359                 BUG_ON(n[i]->nsets != 1);
1360
1361                 bsets[i] = btree_bset_first(n[i]);
1362                 out[i] = bsets[i]->start;
1363
1364                 SET_BTREE_NODE_SEQ(n[i]->data, BTREE_NODE_SEQ(b->data) + 1);
1365                 bch2_bkey_format_init(&format[i]);
1366         }
1367
1368         u64s = 0;
1369         for_each_btree_node_key(b, k, &iter) {
1370                 if (bkey_deleted(k))
1371                         continue;
1372
1373                 i = u64s >= n1_u64s;
1374                 u64s += k->u64s;
1375                 uk = bkey_unpack_key(b, k);
1376                 if (!i)
1377                         n1_pos = uk.p;
1378                 bch2_bkey_format_add_key(&format[i], &uk);
1379         }
1380
1381         btree_set_min(n[0], b->data->min_key);
1382         btree_set_max(n[0], n1_pos);
1383         btree_set_min(n[1], bpos_successor(n1_pos));
1384         btree_set_max(n[1], b->data->max_key);
1385
1386         for (i = 0; i < 2; i++) {
1387                 bch2_bkey_format_add_pos(&format[i], n[i]->data->min_key);
1388                 bch2_bkey_format_add_pos(&format[i], n[i]->data->max_key);
1389
1390                 n[i]->data->format = bch2_bkey_format_done(&format[i]);
1391                 btree_node_set_format(n[i], n[i]->data->format);
1392         }
1393
1394         u64s = 0;
1395         for_each_btree_node_key(b, k, &iter) {
1396                 if (bkey_deleted(k))
1397                         continue;
1398
1399                 i = u64s >= n1_u64s;
1400                 u64s += k->u64s;
1401
1402                 if (bch2_bkey_transform(&n[i]->format, out[i], bkey_packed(k)
1403                                         ? &b->format: &bch2_bkey_format_current, k))
1404                         out[i]->format = KEY_FORMAT_LOCAL_BTREE;
1405                 else
1406                         bch2_bkey_unpack(b, (void *) out[i], k);
1407
1408                 out[i]->needs_whiteout = false;
1409
1410                 btree_keys_account_key_add(&n[i]->nr, 0, out[i]);
1411                 out[i] = bkey_p_next(out[i]);
1412         }
1413
1414         for (i = 0; i < 2; i++) {
1415                 bsets[i]->u64s = cpu_to_le16((u64 *) out[i] - bsets[i]->_data);
1416
1417                 BUG_ON(!bsets[i]->u64s);
1418
1419                 set_btree_bset_end(n[i], n[i]->set);
1420
1421                 btree_node_reset_sib_u64s(n[i]);
1422
1423                 bch2_verify_btree_nr_keys(n[i]);
1424
1425                 if (b->c.level)
1426                         btree_node_interior_verify(as->c, n[i]);
1427         }
1428 }
1429
1430 /*
1431  * For updates to interior nodes, we've got to do the insert before we split
1432  * because the stuff we're inserting has to be inserted atomically. Post split,
1433  * the keys might have to go in different nodes and the split would no longer be
1434  * atomic.
1435  *
1436  * Worse, if the insert is from btree node coalescing, if we do the insert after
1437  * we do the split (and pick the pivot) - the pivot we pick might be between
1438  * nodes that were coalesced, and thus in the middle of a child node post
1439  * coalescing:
1440  */
1441 static void btree_split_insert_keys(struct btree_update *as,
1442                                     struct btree_trans *trans,
1443                                     btree_path_idx_t path_idx,
1444                                     struct btree *b,
1445                                     struct keylist *keys)
1446 {
1447         struct btree_path *path = trans->paths + path_idx;
1448
1449         if (!bch2_keylist_empty(keys) &&
1450             bpos_le(bch2_keylist_front(keys)->k.p, b->data->max_key)) {
1451                 struct btree_node_iter node_iter;
1452
1453                 bch2_btree_node_iter_init(&node_iter, b, &bch2_keylist_front(keys)->k.p);
1454
1455                 __bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
1456
1457                 btree_node_interior_verify(as->c, b);
1458         }
1459 }
1460
1461 static int btree_split(struct btree_update *as, struct btree_trans *trans,
1462                        btree_path_idx_t path, struct btree *b,
1463                        struct keylist *keys, unsigned flags)
1464 {
1465         struct bch_fs *c = as->c;
1466         struct btree *parent = btree_node_parent(trans->paths + path, b);
1467         struct btree *n1, *n2 = NULL, *n3 = NULL;
1468         btree_path_idx_t path1 = 0, path2 = 0;
1469         u64 start_time = local_clock();
1470         int ret = 0;
1471
1472         BUG_ON(!parent && (b != btree_node_root(c, b)));
1473         BUG_ON(parent && !btree_node_intent_locked(trans->paths + path, b->c.level + 1));
1474
1475         bch2_btree_interior_update_will_free_node(as, b);
1476
1477         if (b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c)) {
1478                 struct btree *n[2];
1479
1480                 trace_and_count(c, btree_node_split, trans, b);
1481
1482                 n[0] = n1 = bch2_btree_node_alloc(as, trans, b->c.level);
1483                 n[1] = n2 = bch2_btree_node_alloc(as, trans, b->c.level);
1484
1485                 __btree_split_node(as, trans, b, n);
1486
1487                 if (keys) {
1488                         btree_split_insert_keys(as, trans, path, n1, keys);
1489                         btree_split_insert_keys(as, trans, path, n2, keys);
1490                         BUG_ON(!bch2_keylist_empty(keys));
1491                 }
1492
1493                 bch2_btree_build_aux_trees(n2);
1494                 bch2_btree_build_aux_trees(n1);
1495
1496                 bch2_btree_update_add_new_node(as, n1);
1497                 bch2_btree_update_add_new_node(as, n2);
1498                 six_unlock_write(&n2->c.lock);
1499                 six_unlock_write(&n1->c.lock);
1500
1501                 path1 = get_unlocked_mut_path(trans, as->btree_id, n1->c.level, n1->key.k.p);
1502                 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1503                 mark_btree_node_locked(trans, trans->paths + path1, n1->c.level, BTREE_NODE_INTENT_LOCKED);
1504                 bch2_btree_path_level_init(trans, trans->paths + path1, n1);
1505
1506                 path2 = get_unlocked_mut_path(trans, as->btree_id, n2->c.level, n2->key.k.p);
1507                 six_lock_increment(&n2->c.lock, SIX_LOCK_intent);
1508                 mark_btree_node_locked(trans, trans->paths + path2, n2->c.level, BTREE_NODE_INTENT_LOCKED);
1509                 bch2_btree_path_level_init(trans, trans->paths + path2, n2);
1510
1511                 /*
1512                  * Note that on recursive parent_keys == keys, so we
1513                  * can't start adding new keys to parent_keys before emptying it
1514                  * out (which we did with btree_split_insert_keys() above)
1515                  */
1516                 bch2_keylist_add(&as->parent_keys, &n1->key);
1517                 bch2_keylist_add(&as->parent_keys, &n2->key);
1518
1519                 if (!parent) {
1520                         /* Depth increases, make a new root */
1521                         n3 = __btree_root_alloc(as, trans, b->c.level + 1);
1522
1523                         bch2_btree_update_add_new_node(as, n3);
1524                         six_unlock_write(&n3->c.lock);
1525
1526                         trans->paths[path2].locks_want++;
1527                         BUG_ON(btree_node_locked(trans->paths + path2, n3->c.level));
1528                         six_lock_increment(&n3->c.lock, SIX_LOCK_intent);
1529                         mark_btree_node_locked(trans, trans->paths + path2, n3->c.level, BTREE_NODE_INTENT_LOCKED);
1530                         bch2_btree_path_level_init(trans, trans->paths + path2, n3);
1531
1532                         n3->sib_u64s[0] = U16_MAX;
1533                         n3->sib_u64s[1] = U16_MAX;
1534
1535                         btree_split_insert_keys(as, trans, path, n3, &as->parent_keys);
1536                 }
1537         } else {
1538                 trace_and_count(c, btree_node_compact, trans, b);
1539
1540                 n1 = bch2_btree_node_alloc_replacement(as, trans, b);
1541
1542                 if (keys) {
1543                         btree_split_insert_keys(as, trans, path, n1, keys);
1544                         BUG_ON(!bch2_keylist_empty(keys));
1545                 }
1546
1547                 bch2_btree_build_aux_trees(n1);
1548                 bch2_btree_update_add_new_node(as, n1);
1549                 six_unlock_write(&n1->c.lock);
1550
1551                 path1 = get_unlocked_mut_path(trans, as->btree_id, n1->c.level, n1->key.k.p);
1552                 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1553                 mark_btree_node_locked(trans, trans->paths + path1, n1->c.level, BTREE_NODE_INTENT_LOCKED);
1554                 bch2_btree_path_level_init(trans, trans->paths + path1, n1);
1555
1556                 if (parent)
1557                         bch2_keylist_add(&as->parent_keys, &n1->key);
1558         }
1559
1560         /* New nodes all written, now make them visible: */
1561
1562         if (parent) {
1563                 /* Split a non root node */
1564                 ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1565                 if (ret)
1566                         goto err;
1567         } else if (n3) {
1568                 bch2_btree_set_root(as, trans, trans->paths + path, n3);
1569         } else {
1570                 /* Root filled up but didn't need to be split */
1571                 bch2_btree_set_root(as, trans, trans->paths + path, n1);
1572         }
1573
1574         if (n3) {
1575                 bch2_btree_update_get_open_buckets(as, n3);
1576                 bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0);
1577         }
1578         if (n2) {
1579                 bch2_btree_update_get_open_buckets(as, n2);
1580                 bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0);
1581         }
1582         bch2_btree_update_get_open_buckets(as, n1);
1583         bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1584
1585         /*
1586          * The old node must be freed (in memory) _before_ unlocking the new
1587          * nodes - else another thread could re-acquire a read lock on the old
1588          * node after another thread has locked and updated the new node, thus
1589          * seeing stale data:
1590          */
1591         bch2_btree_node_free_inmem(trans, trans->paths + path, b);
1592
1593         if (n3)
1594                 bch2_trans_node_add(trans, trans->paths + path, n3);
1595         if (n2)
1596                 bch2_trans_node_add(trans, trans->paths + path2, n2);
1597         bch2_trans_node_add(trans, trans->paths + path1, n1);
1598
1599         if (n3)
1600                 six_unlock_intent(&n3->c.lock);
1601         if (n2)
1602                 six_unlock_intent(&n2->c.lock);
1603         six_unlock_intent(&n1->c.lock);
1604 out:
1605         if (path2) {
1606                 __bch2_btree_path_unlock(trans, trans->paths + path2);
1607                 bch2_path_put(trans, path2, true);
1608         }
1609         if (path1) {
1610                 __bch2_btree_path_unlock(trans, trans->paths + path1);
1611                 bch2_path_put(trans, path1, true);
1612         }
1613
1614         bch2_trans_verify_locks(trans);
1615
1616         bch2_time_stats_update(&c->times[n2
1617                                ? BCH_TIME_btree_node_split
1618                                : BCH_TIME_btree_node_compact],
1619                                start_time);
1620         return ret;
1621 err:
1622         if (n3)
1623                 bch2_btree_node_free_never_used(as, trans, n3);
1624         if (n2)
1625                 bch2_btree_node_free_never_used(as, trans, n2);
1626         bch2_btree_node_free_never_used(as, trans, n1);
1627         goto out;
1628 }
1629
1630 static void
1631 bch2_btree_insert_keys_interior(struct btree_update *as,
1632                                 struct btree_trans *trans,
1633                                 struct btree_path *path,
1634                                 struct btree *b,
1635                                 struct keylist *keys)
1636 {
1637         struct btree_path *linked;
1638         unsigned i;
1639
1640         __bch2_btree_insert_keys_interior(as, trans, path, b,
1641                                           path->l[b->c.level].iter, keys);
1642
1643         btree_update_updated_node(as, b);
1644
1645         trans_for_each_path_with_node(trans, b, linked, i)
1646                 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1647
1648         bch2_trans_verify_paths(trans);
1649 }
1650
1651 /**
1652  * bch2_btree_insert_node - insert bkeys into a given btree node
1653  *
1654  * @as:                 btree_update object
1655  * @trans:              btree_trans object
1656  * @path:               path that points to current node
1657  * @b:                  node to insert keys into
1658  * @keys:               list of keys to insert
1659  * @flags:              transaction commit flags
1660  *
1661  * Returns: 0 on success, typically transaction restart error on failure
1662  *
1663  * Inserts as many keys as it can into a given btree node, splitting it if full.
1664  * If a split occurred, this function will return early. This can only happen
1665  * for leaf nodes -- inserts into interior nodes have to be atomic.
1666  */
1667 static int bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans,
1668                                   btree_path_idx_t path_idx, struct btree *b,
1669                                   struct keylist *keys, unsigned flags)
1670 {
1671         struct bch_fs *c = as->c;
1672         struct btree_path *path = trans->paths + path_idx;
1673         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1674         int old_live_u64s = b->nr.live_u64s;
1675         int live_u64s_added, u64s_added;
1676         int ret;
1677
1678         lockdep_assert_held(&c->gc_lock);
1679         BUG_ON(!btree_node_intent_locked(path, b->c.level));
1680         BUG_ON(!b->c.level);
1681         BUG_ON(!as || as->b);
1682         bch2_verify_keylist_sorted(keys);
1683
1684         ret = bch2_btree_node_lock_write(trans, path, &b->c);
1685         if (ret)
1686                 return ret;
1687
1688         bch2_btree_node_prep_for_write(trans, path, b);
1689
1690         if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1691                 bch2_btree_node_unlock_write(trans, path, b);
1692                 goto split;
1693         }
1694
1695         btree_node_interior_verify(c, b);
1696
1697         bch2_btree_insert_keys_interior(as, trans, path, b, keys);
1698
1699         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1700         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1701
1702         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1703                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1704         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1705                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1706
1707         if (u64s_added > live_u64s_added &&
1708             bch2_maybe_compact_whiteouts(c, b))
1709                 bch2_trans_node_reinit_iter(trans, b);
1710
1711         bch2_btree_node_unlock_write(trans, path, b);
1712
1713         btree_node_interior_verify(c, b);
1714         return 0;
1715 split:
1716         /*
1717          * We could attempt to avoid the transaction restart, by calling
1718          * bch2_btree_path_upgrade() and allocating more nodes:
1719          */
1720         if (b->c.level >= as->update_level) {
1721                 trace_and_count(c, trans_restart_split_race, trans, _THIS_IP_, b);
1722                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_split_race);
1723         }
1724
1725         return btree_split(as, trans, path_idx, b, keys, flags);
1726 }
1727
1728 int bch2_btree_split_leaf(struct btree_trans *trans,
1729                           btree_path_idx_t path,
1730                           unsigned flags)
1731 {
1732         /* btree_split & merge may both cause paths array to be reallocated */
1733
1734         struct btree *b = path_l(trans->paths + path)->b;
1735         struct btree_update *as;
1736         unsigned l;
1737         int ret = 0;
1738
1739         as = bch2_btree_update_start(trans, trans->paths + path,
1740                                      trans->paths[path].level,
1741                                      true, flags);
1742         if (IS_ERR(as))
1743                 return PTR_ERR(as);
1744
1745         ret = btree_split(as, trans, path, b, NULL, flags);
1746         if (ret) {
1747                 bch2_btree_update_free(as, trans);
1748                 return ret;
1749         }
1750
1751         bch2_btree_update_done(as, trans);
1752
1753         for (l = trans->paths[path].level + 1;
1754              btree_node_intent_locked(&trans->paths[path], l) && !ret;
1755              l++)
1756                 ret = bch2_foreground_maybe_merge(trans, path, l, flags);
1757
1758         return ret;
1759 }
1760
1761 int __bch2_foreground_maybe_merge(struct btree_trans *trans,
1762                                   btree_path_idx_t path,
1763                                   unsigned level,
1764                                   unsigned flags,
1765                                   enum btree_node_sibling sib)
1766 {
1767         struct bch_fs *c = trans->c;
1768         struct btree_update *as;
1769         struct bkey_format_state new_s;
1770         struct bkey_format new_f;
1771         struct bkey_i delete;
1772         struct btree *b, *m, *n, *prev, *next, *parent;
1773         struct bpos sib_pos;
1774         size_t sib_u64s;
1775         enum btree_id btree = trans->paths[path].btree_id;
1776         btree_path_idx_t sib_path = 0, new_path = 0;
1777         u64 start_time = local_clock();
1778         int ret = 0;
1779
1780         BUG_ON(!trans->paths[path].should_be_locked);
1781         BUG_ON(!btree_node_locked(&trans->paths[path], level));
1782
1783         b = trans->paths[path].l[level].b;
1784
1785         if ((sib == btree_prev_sib && bpos_eq(b->data->min_key, POS_MIN)) ||
1786             (sib == btree_next_sib && bpos_eq(b->data->max_key, SPOS_MAX))) {
1787                 b->sib_u64s[sib] = U16_MAX;
1788                 return 0;
1789         }
1790
1791         sib_pos = sib == btree_prev_sib
1792                 ? bpos_predecessor(b->data->min_key)
1793                 : bpos_successor(b->data->max_key);
1794
1795         sib_path = bch2_path_get(trans, btree, sib_pos,
1796                                  U8_MAX, level, BTREE_ITER_INTENT, _THIS_IP_);
1797         ret = bch2_btree_path_traverse(trans, sib_path, false);
1798         if (ret)
1799                 goto err;
1800
1801         btree_path_set_should_be_locked(trans->paths + sib_path);
1802
1803         m = trans->paths[sib_path].l[level].b;
1804
1805         if (btree_node_parent(trans->paths + path, b) !=
1806             btree_node_parent(trans->paths + sib_path, m)) {
1807                 b->sib_u64s[sib] = U16_MAX;
1808                 goto out;
1809         }
1810
1811         if (sib == btree_prev_sib) {
1812                 prev = m;
1813                 next = b;
1814         } else {
1815                 prev = b;
1816                 next = m;
1817         }
1818
1819         if (!bpos_eq(bpos_successor(prev->data->max_key), next->data->min_key)) {
1820                 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1821
1822                 bch2_bpos_to_text(&buf1, prev->data->max_key);
1823                 bch2_bpos_to_text(&buf2, next->data->min_key);
1824                 bch_err(c,
1825                         "%s(): btree topology error:\n"
1826                         "  prev ends at   %s\n"
1827                         "  next starts at %s",
1828                         __func__, buf1.buf, buf2.buf);
1829                 printbuf_exit(&buf1);
1830                 printbuf_exit(&buf2);
1831                 bch2_topology_error(c);
1832                 ret = -EIO;
1833                 goto err;
1834         }
1835
1836         bch2_bkey_format_init(&new_s);
1837         bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
1838         __bch2_btree_calc_format(&new_s, prev);
1839         __bch2_btree_calc_format(&new_s, next);
1840         bch2_bkey_format_add_pos(&new_s, next->data->max_key);
1841         new_f = bch2_bkey_format_done(&new_s);
1842
1843         sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1844                 btree_node_u64s_with_format(m, &new_f);
1845
1846         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1847                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1848                 sib_u64s /= 2;
1849                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1850         }
1851
1852         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1853         sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
1854         b->sib_u64s[sib] = sib_u64s;
1855
1856         if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
1857                 goto out;
1858
1859         parent = btree_node_parent(trans->paths + path, b);
1860         as = bch2_btree_update_start(trans, trans->paths + path, level, false,
1861                                      BCH_TRANS_COMMIT_no_enospc|flags);
1862         ret = PTR_ERR_OR_ZERO(as);
1863         if (ret)
1864                 goto err;
1865
1866         trace_and_count(c, btree_node_merge, trans, b);
1867
1868         bch2_btree_interior_update_will_free_node(as, b);
1869         bch2_btree_interior_update_will_free_node(as, m);
1870
1871         n = bch2_btree_node_alloc(as, trans, b->c.level);
1872
1873         SET_BTREE_NODE_SEQ(n->data,
1874                            max(BTREE_NODE_SEQ(b->data),
1875                                BTREE_NODE_SEQ(m->data)) + 1);
1876
1877         btree_set_min(n, prev->data->min_key);
1878         btree_set_max(n, next->data->max_key);
1879
1880         n->data->format  = new_f;
1881         btree_node_set_format(n, new_f);
1882
1883         bch2_btree_sort_into(c, n, prev);
1884         bch2_btree_sort_into(c, n, next);
1885
1886         bch2_btree_build_aux_trees(n);
1887         bch2_btree_update_add_new_node(as, n);
1888         six_unlock_write(&n->c.lock);
1889
1890         new_path = get_unlocked_mut_path(trans, btree, n->c.level, n->key.k.p);
1891         six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1892         mark_btree_node_locked(trans, trans->paths + new_path, n->c.level, BTREE_NODE_INTENT_LOCKED);
1893         bch2_btree_path_level_init(trans, trans->paths + new_path, n);
1894
1895         bkey_init(&delete.k);
1896         delete.k.p = prev->key.k.p;
1897         bch2_keylist_add(&as->parent_keys, &delete);
1898         bch2_keylist_add(&as->parent_keys, &n->key);
1899
1900         bch2_trans_verify_paths(trans);
1901
1902         ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1903         if (ret)
1904                 goto err_free_update;
1905
1906         bch2_trans_verify_paths(trans);
1907
1908         bch2_btree_update_get_open_buckets(as, n);
1909         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1910
1911         bch2_btree_node_free_inmem(trans, trans->paths + path, b);
1912         bch2_btree_node_free_inmem(trans, trans->paths + sib_path, m);
1913
1914         bch2_trans_node_add(trans, trans->paths + path, n);
1915
1916         bch2_trans_verify_paths(trans);
1917
1918         six_unlock_intent(&n->c.lock);
1919
1920         bch2_btree_update_done(as, trans);
1921
1922         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time);
1923 out:
1924 err:
1925         if (new_path)
1926                 bch2_path_put(trans, new_path, true);
1927         bch2_path_put(trans, sib_path, true);
1928         bch2_trans_verify_locks(trans);
1929         return ret;
1930 err_free_update:
1931         bch2_btree_node_free_never_used(as, trans, n);
1932         bch2_btree_update_free(as, trans);
1933         goto out;
1934 }
1935
1936 int bch2_btree_node_rewrite(struct btree_trans *trans,
1937                             struct btree_iter *iter,
1938                             struct btree *b,
1939                             unsigned flags)
1940 {
1941         struct bch_fs *c = trans->c;
1942         struct btree *n, *parent;
1943         struct btree_update *as;
1944         btree_path_idx_t new_path = 0;
1945         int ret;
1946
1947         flags |= BCH_TRANS_COMMIT_no_enospc;
1948
1949         struct btree_path *path = btree_iter_path(trans, iter);
1950         parent = btree_node_parent(path, b);
1951         as = bch2_btree_update_start(trans, path, b->c.level, false, flags);
1952         ret = PTR_ERR_OR_ZERO(as);
1953         if (ret)
1954                 goto out;
1955
1956         bch2_btree_interior_update_will_free_node(as, b);
1957
1958         n = bch2_btree_node_alloc_replacement(as, trans, b);
1959
1960         bch2_btree_build_aux_trees(n);
1961         bch2_btree_update_add_new_node(as, n);
1962         six_unlock_write(&n->c.lock);
1963
1964         new_path = get_unlocked_mut_path(trans, iter->btree_id, n->c.level, n->key.k.p);
1965         six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1966         mark_btree_node_locked(trans, trans->paths + new_path, n->c.level, BTREE_NODE_INTENT_LOCKED);
1967         bch2_btree_path_level_init(trans, trans->paths + new_path, n);
1968
1969         trace_and_count(c, btree_node_rewrite, trans, b);
1970
1971         if (parent) {
1972                 bch2_keylist_add(&as->parent_keys, &n->key);
1973                 ret = bch2_btree_insert_node(as, trans, iter->path,
1974                                              parent, &as->parent_keys, flags);
1975                 if (ret)
1976                         goto err;
1977         } else {
1978                 bch2_btree_set_root(as, trans, btree_iter_path(trans, iter), n);
1979         }
1980
1981         bch2_btree_update_get_open_buckets(as, n);
1982         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1983
1984         bch2_btree_node_free_inmem(trans, btree_iter_path(trans, iter), b);
1985
1986         bch2_trans_node_add(trans, trans->paths + iter->path, n);
1987         six_unlock_intent(&n->c.lock);
1988
1989         bch2_btree_update_done(as, trans);
1990 out:
1991         if (new_path)
1992                 bch2_path_put(trans, new_path, true);
1993         bch2_trans_downgrade(trans);
1994         return ret;
1995 err:
1996         bch2_btree_node_free_never_used(as, trans, n);
1997         bch2_btree_update_free(as, trans);
1998         goto out;
1999 }
2000
2001 struct async_btree_rewrite {
2002         struct bch_fs           *c;
2003         struct work_struct      work;
2004         struct list_head        list;
2005         enum btree_id           btree_id;
2006         unsigned                level;
2007         struct bpos             pos;
2008         __le64                  seq;
2009 };
2010
2011 static int async_btree_node_rewrite_trans(struct btree_trans *trans,
2012                                           struct async_btree_rewrite *a)
2013 {
2014         struct bch_fs *c = trans->c;
2015         struct btree_iter iter;
2016         struct btree *b;
2017         int ret;
2018
2019         bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos,
2020                                   BTREE_MAX_DEPTH, a->level, 0);
2021         b = bch2_btree_iter_peek_node(&iter);
2022         ret = PTR_ERR_OR_ZERO(b);
2023         if (ret)
2024                 goto out;
2025
2026         if (!b || b->data->keys.seq != a->seq) {
2027                 struct printbuf buf = PRINTBUF;
2028
2029                 if (b)
2030                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
2031                 else
2032                         prt_str(&buf, "(null");
2033                 bch_info(c, "%s: node to rewrite not found:, searching for seq %llu, got\n%s",
2034                          __func__, a->seq, buf.buf);
2035                 printbuf_exit(&buf);
2036                 goto out;
2037         }
2038
2039         ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
2040 out:
2041         bch2_trans_iter_exit(trans, &iter);
2042
2043         return ret;
2044 }
2045
2046 static void async_btree_node_rewrite_work(struct work_struct *work)
2047 {
2048         struct async_btree_rewrite *a =
2049                 container_of(work, struct async_btree_rewrite, work);
2050         struct bch_fs *c = a->c;
2051         int ret;
2052
2053         ret = bch2_trans_do(c, NULL, NULL, 0,
2054                       async_btree_node_rewrite_trans(trans, a));
2055         if (ret)
2056                 bch_err_fn(c, ret);
2057         bch2_write_ref_put(c, BCH_WRITE_REF_node_rewrite);
2058         kfree(a);
2059 }
2060
2061 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
2062 {
2063         struct async_btree_rewrite *a;
2064         int ret;
2065
2066         a = kmalloc(sizeof(*a), GFP_NOFS);
2067         if (!a) {
2068                 bch_err(c, "%s: error allocating memory", __func__);
2069                 return;
2070         }
2071
2072         a->c            = c;
2073         a->btree_id     = b->c.btree_id;
2074         a->level        = b->c.level;
2075         a->pos          = b->key.k.p;
2076         a->seq          = b->data->keys.seq;
2077         INIT_WORK(&a->work, async_btree_node_rewrite_work);
2078
2079         if (unlikely(!test_bit(BCH_FS_may_go_rw, &c->flags))) {
2080                 mutex_lock(&c->pending_node_rewrites_lock);
2081                 list_add(&a->list, &c->pending_node_rewrites);
2082                 mutex_unlock(&c->pending_node_rewrites_lock);
2083                 return;
2084         }
2085
2086         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_node_rewrite)) {
2087                 if (test_bit(BCH_FS_started, &c->flags)) {
2088                         bch_err(c, "%s: error getting c->writes ref", __func__);
2089                         kfree(a);
2090                         return;
2091                 }
2092
2093                 ret = bch2_fs_read_write_early(c);
2094                 if (ret) {
2095                         bch_err_msg(c, ret, "going read-write");
2096                         kfree(a);
2097                         return;
2098                 }
2099
2100                 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2101         }
2102
2103         queue_work(c->btree_interior_update_worker, &a->work);
2104 }
2105
2106 void bch2_do_pending_node_rewrites(struct bch_fs *c)
2107 {
2108         struct async_btree_rewrite *a, *n;
2109
2110         mutex_lock(&c->pending_node_rewrites_lock);
2111         list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2112                 list_del(&a->list);
2113
2114                 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2115                 queue_work(c->btree_interior_update_worker, &a->work);
2116         }
2117         mutex_unlock(&c->pending_node_rewrites_lock);
2118 }
2119
2120 void bch2_free_pending_node_rewrites(struct bch_fs *c)
2121 {
2122         struct async_btree_rewrite *a, *n;
2123
2124         mutex_lock(&c->pending_node_rewrites_lock);
2125         list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2126                 list_del(&a->list);
2127
2128                 kfree(a);
2129         }
2130         mutex_unlock(&c->pending_node_rewrites_lock);
2131 }
2132
2133 static int __bch2_btree_node_update_key(struct btree_trans *trans,
2134                                         struct btree_iter *iter,
2135                                         struct btree *b, struct btree *new_hash,
2136                                         struct bkey_i *new_key,
2137                                         unsigned commit_flags,
2138                                         bool skip_triggers)
2139 {
2140         struct bch_fs *c = trans->c;
2141         struct btree_iter iter2 = { NULL };
2142         struct btree *parent;
2143         int ret;
2144
2145         if (!skip_triggers) {
2146                 ret = bch2_trans_mark_old(trans, b->c.btree_id, b->c.level + 1,
2147                                           bkey_i_to_s_c(&b->key), 0);
2148                 if (ret)
2149                         return ret;
2150
2151                 ret = bch2_trans_mark_new(trans, b->c.btree_id, b->c.level + 1,
2152                                           new_key, 0);
2153                 if (ret)
2154                         return ret;
2155         }
2156
2157         if (new_hash) {
2158                 bkey_copy(&new_hash->key, new_key);
2159                 ret = bch2_btree_node_hash_insert(&c->btree_cache,
2160                                 new_hash, b->c.level, b->c.btree_id);
2161                 BUG_ON(ret);
2162         }
2163
2164         parent = btree_node_parent(btree_iter_path(trans, iter), b);
2165         if (parent) {
2166                 bch2_trans_copy_iter(&iter2, iter);
2167
2168                 iter2.path = bch2_btree_path_make_mut(trans, iter2.path,
2169                                 iter2.flags & BTREE_ITER_INTENT,
2170                                 _THIS_IP_);
2171
2172                 struct btree_path *path2 = btree_iter_path(trans, &iter2);
2173                 BUG_ON(path2->level != b->c.level);
2174                 BUG_ON(!bpos_eq(path2->pos, new_key->k.p));
2175
2176                 btree_path_set_level_up(trans, path2);
2177
2178                 trans->paths_sorted = false;
2179
2180                 ret   = bch2_btree_iter_traverse(&iter2) ?:
2181                         bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_NORUN);
2182                 if (ret)
2183                         goto err;
2184         } else {
2185                 BUG_ON(btree_node_root(c, b) != b);
2186
2187                 struct jset_entry *e = bch2_trans_jset_entry_alloc(trans,
2188                                        jset_u64s(new_key->k.u64s));
2189                 ret = PTR_ERR_OR_ZERO(e);
2190                 if (ret)
2191                         return ret;
2192
2193                 journal_entry_set(e,
2194                                   BCH_JSET_ENTRY_btree_root,
2195                                   b->c.btree_id, b->c.level,
2196                                   new_key, new_key->k.u64s);
2197         }
2198
2199         ret = bch2_trans_commit(trans, NULL, NULL, commit_flags);
2200         if (ret)
2201                 goto err;
2202
2203         bch2_btree_node_lock_write_nofail(trans, btree_iter_path(trans, iter), &b->c);
2204
2205         if (new_hash) {
2206                 mutex_lock(&c->btree_cache.lock);
2207                 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
2208                 bch2_btree_node_hash_remove(&c->btree_cache, b);
2209
2210                 bkey_copy(&b->key, new_key);
2211                 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
2212                 BUG_ON(ret);
2213                 mutex_unlock(&c->btree_cache.lock);
2214         } else {
2215                 bkey_copy(&b->key, new_key);
2216         }
2217
2218         bch2_btree_node_unlock_write(trans, btree_iter_path(trans, iter), b);
2219 out:
2220         bch2_trans_iter_exit(trans, &iter2);
2221         return ret;
2222 err:
2223         if (new_hash) {
2224                 mutex_lock(&c->btree_cache.lock);
2225                 bch2_btree_node_hash_remove(&c->btree_cache, b);
2226                 mutex_unlock(&c->btree_cache.lock);
2227         }
2228         goto out;
2229 }
2230
2231 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
2232                                struct btree *b, struct bkey_i *new_key,
2233                                unsigned commit_flags, bool skip_triggers)
2234 {
2235         struct bch_fs *c = trans->c;
2236         struct btree *new_hash = NULL;
2237         struct btree_path *path = btree_iter_path(trans, iter);
2238         struct closure cl;
2239         int ret = 0;
2240
2241         ret = bch2_btree_path_upgrade(trans, path, b->c.level + 1);
2242         if (ret)
2243                 return ret;
2244
2245         closure_init_stack(&cl);
2246
2247         /*
2248          * check btree_ptr_hash_val() after @b is locked by
2249          * btree_iter_traverse():
2250          */
2251         if (btree_ptr_hash_val(new_key) != b->hash_val) {
2252                 ret = bch2_btree_cache_cannibalize_lock(trans, &cl);
2253                 if (ret) {
2254                         ret = drop_locks_do(trans, (closure_sync(&cl), 0));
2255                         if (ret)
2256                                 return ret;
2257                 }
2258
2259                 new_hash = bch2_btree_node_mem_alloc(trans, false);
2260         }
2261
2262         path->intent_ref++;
2263         ret = __bch2_btree_node_update_key(trans, iter, b, new_hash, new_key,
2264                                            commit_flags, skip_triggers);
2265         --path->intent_ref;
2266
2267         if (new_hash) {
2268                 mutex_lock(&c->btree_cache.lock);
2269                 list_move(&new_hash->list, &c->btree_cache.freeable);
2270                 mutex_unlock(&c->btree_cache.lock);
2271
2272                 six_unlock_write(&new_hash->c.lock);
2273                 six_unlock_intent(&new_hash->c.lock);
2274         }
2275         closure_sync(&cl);
2276         bch2_btree_cache_cannibalize_unlock(trans);
2277         return ret;
2278 }
2279
2280 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
2281                                         struct btree *b, struct bkey_i *new_key,
2282                                         unsigned commit_flags, bool skip_triggers)
2283 {
2284         struct btree_iter iter;
2285         int ret;
2286
2287         bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p,
2288                                   BTREE_MAX_DEPTH, b->c.level,
2289                                   BTREE_ITER_INTENT);
2290         ret = bch2_btree_iter_traverse(&iter);
2291         if (ret)
2292                 goto out;
2293
2294         /* has node been freed? */
2295         if (btree_iter_path(trans, &iter)->l[b->c.level].b != b) {
2296                 /* node has been freed: */
2297                 BUG_ON(!btree_node_dying(b));
2298                 goto out;
2299         }
2300
2301         BUG_ON(!btree_node_hashed(b));
2302
2303         struct bch_extent_ptr *ptr;
2304         bch2_bkey_drop_ptrs(bkey_i_to_s(new_key), ptr,
2305                             !bch2_bkey_has_device(bkey_i_to_s(&b->key), ptr->dev));
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(trans, &cl);
2338                 closure_sync(&cl);
2339         } while (ret);
2340
2341         b = bch2_btree_node_mem_alloc(trans, false);
2342         bch2_btree_cache_cannibalize_unlock(trans);
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                        closure_nr_remaining(&as->cl),
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, (struct bkey_i *) entry->start);
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 *end,
2430                                     unsigned long skip)
2431 {
2432         unsigned i;
2433
2434         mutex_lock(&c->btree_root_lock);
2435
2436         for (i = 0; i < btree_id_nr_alive(c); i++) {
2437                 struct btree_root *r = bch2_btree_id_root(c, i);
2438
2439                 if (r->alive && !test_bit(i, &skip)) {
2440                         journal_entry_set(end, BCH_JSET_ENTRY_btree_root,
2441                                           i, r->level, &r->key, r->key.k.u64s);
2442                         end = vstruct_next(end);
2443                 }
2444         }
2445
2446         mutex_unlock(&c->btree_root_lock);
2447
2448         return end;
2449 }
2450
2451 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2452 {
2453         if (c->btree_interior_update_worker)
2454                 destroy_workqueue(c->btree_interior_update_worker);
2455         mempool_exit(&c->btree_interior_update_pool);
2456 }
2457
2458 void bch2_fs_btree_interior_update_init_early(struct bch_fs *c)
2459 {
2460         mutex_init(&c->btree_reserve_cache_lock);
2461         INIT_LIST_HEAD(&c->btree_interior_update_list);
2462         INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2463         mutex_init(&c->btree_interior_update_lock);
2464         INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2465
2466         INIT_LIST_HEAD(&c->pending_node_rewrites);
2467         mutex_init(&c->pending_node_rewrites_lock);
2468 }
2469
2470 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2471 {
2472         c->btree_interior_update_worker =
2473                 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2474         if (!c->btree_interior_update_worker)
2475                 return -BCH_ERR_ENOMEM_btree_interior_update_worker_init;
2476
2477         if (mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2478                                       sizeof(struct btree_update)))
2479                 return -BCH_ERR_ENOMEM_btree_interior_update_pool_init;
2480
2481         return 0;
2482 }