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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "alloc_foreground.h"
   5 #include "bkey_methods.h"
   6 #include "btree_cache.h"
   7 #include "btree_gc.h"
   8 #include "btree_update.h"
   9 #include "btree_update_interior.h"
  10 #include "btree_io.h"
  11 #include "btree_iter.h"
  12 #include "btree_locking.h"
  13 #include "buckets.h"
  14 #include "error.h"
  15 #include "extents.h"
  16 #include "journal.h"
  17 #include "journal_reclaim.h"
  18 #include "keylist.h"
  19 #include "replicas.h"
  20 #include "super-io.h"
  21
  22 #include <linux/random.h>
  23 #include <trace/events/bcachefs.h>
  24
  25 /* Debug code: */
  26
  27 /*
  28  * Verify that child nodes correctly span parent node's range:
  29  */
  30 static void btree_node_interior_verify(struct bch_fs *c, struct btree *b)
  31 {
  32 #ifdef CONFIG_BCACHEFS_DEBUG
  33         struct bpos next_node = b->data->min_key;
  34         struct btree_node_iter iter;
  35         struct bkey_s_c k;
  36         struct bkey_s_c_btree_ptr_v2 bp;
  37         struct bkey unpacked;
  38
  39         BUG_ON(!b->c.level);
  40
  41         if (!test_bit(BCH_FS_BTREE_INTERIOR_REPLAY_DONE, &c->flags))
  42                 return;
  43
  44         bch2_btree_node_iter_init_from_start(&iter, b);
  45
  46         while (1) {
  47                 k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked);
  48                 if (k.k->type != KEY_TYPE_btree_ptr_v2)
  49                         break;
  50                 bp = bkey_s_c_to_btree_ptr_v2(k);
  51
  52                 BUG_ON(bkey_cmp(next_node, bp.v->min_key));
  53
  54                 bch2_btree_node_iter_advance(&iter, b);
  55
  56                 if (bch2_btree_node_iter_end(&iter)) {
  57                         BUG_ON(bkey_cmp(k.k->p, b->key.k.p));
  58                         break;
  59                 }
  60
  61                 next_node = bkey_successor(k.k->p);
  62         }
  63 #endif
  64 }
  65
  66 /* Calculate ideal packed bkey format for new btree nodes: */
  67
  68 void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b)
  69 {
  70         struct bkey_packed *k;
  71         struct bset_tree *t;
  72         struct bkey uk;
  73
  74         bch2_bkey_format_add_pos(s, b->data->min_key);
  75
  76         for_each_bset(b, t)
  77                 bset_tree_for_each_key(b, t, k)
  78                         if (!bkey_whiteout(k)) {
  79                                 uk = bkey_unpack_key(b, k);
  80                                 bch2_bkey_format_add_key(s, &uk);
  81                         }
  82 }
  83
  84 static struct bkey_format bch2_btree_calc_format(struct btree *b)
  85 {
  86         struct bkey_format_state s;
  87
  88         bch2_bkey_format_init(&s);
  89         __bch2_btree_calc_format(&s, b);
  90
  91         return bch2_bkey_format_done(&s);
  92 }
  93
  94 static size_t btree_node_u64s_with_format(struct btree *b,
  95                                           struct bkey_format *new_f)
  96 {
  97         struct bkey_format *old_f = &b->format;
  98
  99         /* stupid integer promotion rules */
 100         ssize_t delta =
 101             (((int) new_f->key_u64s - old_f->key_u64s) *
 102              (int) b->nr.packed_keys) +
 103             (((int) new_f->key_u64s - BKEY_U64s) *
 104              (int) b->nr.unpacked_keys);
 105
 106         BUG_ON(delta + b->nr.live_u64s < 0);
 107
 108         return b->nr.live_u64s + delta;
 109 }
 110
 111 /**
 112  * btree_node_format_fits - check if we could rewrite node with a new format
 113  *
 114  * This assumes all keys can pack with the new format -- it just checks if
 115  * the re-packed keys would fit inside the node itself.
 116  */
 117 bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
 118                                  struct bkey_format *new_f)
 119 {
 120         size_t u64s = btree_node_u64s_with_format(b, new_f);
 121
 122         return __vstruct_bytes(struct btree_node, u64s) < btree_bytes(c);
 123 }
 124
 125 /* Btree node freeing/allocation: */
 126
 127 static void __btree_node_free(struct bch_fs *c, struct btree *b)
 128 {
 129         trace_btree_node_free(c, b);
 130
 131         BUG_ON(btree_node_dirty(b));
 132         BUG_ON(btree_node_need_write(b));
 133         BUG_ON(b == btree_node_root(c, b));
 134         BUG_ON(b->ob.nr);
 135         BUG_ON(!list_empty(&b->write_blocked));
 136         BUG_ON(b->will_make_reachable);
 137
 138         clear_btree_node_noevict(b);
 139
 140         bch2_btree_node_hash_remove(&c->btree_cache, b);
 141
 142         mutex_lock(&c->btree_cache.lock);
 143         list_move(&b->list, &c->btree_cache.freeable);
 144         mutex_unlock(&c->btree_cache.lock);
 145 }
 146
 147 void bch2_btree_node_free_never_inserted(struct bch_fs *c, struct btree *b)
 148 {
 149         struct open_buckets ob = b->ob;
 150
 151         b->ob.nr = 0;
 152
 153         clear_btree_node_dirty(c, b);
 154
 155         btree_node_lock_type(c, b, SIX_LOCK_write);
 156         __btree_node_free(c, b);
 157         six_unlock_write(&b->c.lock);
 158
 159         bch2_open_buckets_put(c, &ob);
 160 }
 161
 162 void bch2_btree_node_free_inmem(struct bch_fs *c, struct btree *b,
 163                                 struct btree_iter *iter)
 164 {
 165         struct btree_iter *linked;
 166
 167         trans_for_each_iter(iter->trans, linked)
 168                 BUG_ON(linked->l[b->c.level].b == b);
 169
 170         six_lock_write(&b->c.lock, NULL, NULL);
 171         __btree_node_free(c, b);
 172         six_unlock_write(&b->c.lock);
 173         six_unlock_intent(&b->c.lock);
 174 }
 175
 176 static struct btree *__bch2_btree_node_alloc(struct bch_fs *c,
 177                                              struct disk_reservation *res,
 178                                              struct closure *cl,
 179                                              unsigned flags)
 180 {
 181         struct write_point *wp;
 182         struct btree *b;
 183         BKEY_PADDED(k) tmp;
 184         struct open_buckets ob = { .nr = 0 };
 185         struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
 186         unsigned nr_reserve;
 187         enum alloc_reserve alloc_reserve;
 188
 189         if (flags & BTREE_INSERT_USE_ALLOC_RESERVE) {
 190                 nr_reserve      = 0;
 191                 alloc_reserve   = RESERVE_ALLOC;
 192         } else if (flags & BTREE_INSERT_USE_RESERVE) {
 193                 nr_reserve      = BTREE_NODE_RESERVE / 2;
 194                 alloc_reserve   = RESERVE_BTREE;
 195         } else {
 196                 nr_reserve      = BTREE_NODE_RESERVE;
 197                 alloc_reserve   = RESERVE_NONE;
 198         }
 199
 200         mutex_lock(&c->btree_reserve_cache_lock);
 201         if (c->btree_reserve_cache_nr > nr_reserve) {
 202                 struct btree_alloc *a =
 203                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
 204
 205                 ob = a->ob;
 206                 bkey_copy(&tmp.k, &a->k);
 207                 mutex_unlock(&c->btree_reserve_cache_lock);
 208                 goto mem_alloc;
 209         }
 210         mutex_unlock(&c->btree_reserve_cache_lock);
 211
 212 retry:
 213         wp = bch2_alloc_sectors_start(c, c->opts.foreground_target, 0,
 214                                       writepoint_ptr(&c->btree_write_point),
 215                                       &devs_have,
 216                                       res->nr_replicas,
 217                                       c->opts.metadata_replicas_required,
 218                                       alloc_reserve, 0, cl);
 219         if (IS_ERR(wp))
 220                 return ERR_CAST(wp);
 221
 222         if (wp->sectors_free < c->opts.btree_node_size) {
 223                 struct open_bucket *ob;
 224                 unsigned i;
 225
 226                 open_bucket_for_each(c, &wp->ptrs, ob, i)
 227                         if (ob->sectors_free < c->opts.btree_node_size)
 228                                 ob->sectors_free = 0;
 229
 230                 bch2_alloc_sectors_done(c, wp);
 231                 goto retry;
 232         }
 233
 234         if (c->sb.features & (1ULL << BCH_FEATURE_btree_ptr_v2))
 235                 bkey_btree_ptr_v2_init(&tmp.k);
 236         else
 237                 bkey_btree_ptr_init(&tmp.k);
 238
 239         bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, c->opts.btree_node_size);
 240
 241         bch2_open_bucket_get(c, wp, &ob);
 242         bch2_alloc_sectors_done(c, wp);
 243 mem_alloc:
 244         b = bch2_btree_node_mem_alloc(c);
 245
 246         /* we hold cannibalize_lock: */
 247         BUG_ON(IS_ERR(b));
 248         BUG_ON(b->ob.nr);
 249
 250         bkey_copy(&b->key, &tmp.k);
 251         b->ob = ob;
 252
 253         return b;
 254 }
 255
 256 static struct btree *bch2_btree_node_alloc(struct btree_update *as, unsigned level)
 257 {
 258         struct bch_fs *c = as->c;
 259         struct btree *b;
 260         int ret;
 261
 262         BUG_ON(level >= BTREE_MAX_DEPTH);
 263         BUG_ON(!as->nr_prealloc_nodes);
 264
 265         b = as->prealloc_nodes[--as->nr_prealloc_nodes];
 266
 267         set_btree_node_accessed(b);
 268         set_btree_node_dirty(c, b);
 269         set_btree_node_need_write(b);
 270
 271         bch2_bset_init_first(b, &b->data->keys);
 272         b->c.level      = level;
 273         b->c.btree_id   = as->btree_id;
 274
 275         memset(&b->nr, 0, sizeof(b->nr));
 276         b->data->magic = cpu_to_le64(bset_magic(c));
 277         b->data->flags = 0;
 278         SET_BTREE_NODE_ID(b->data, as->btree_id);
 279         SET_BTREE_NODE_LEVEL(b->data, level);
 280         b->data->ptr = bch2_bkey_ptrs_c(bkey_i_to_s_c(&b->key)).start->ptr;
 281
 282         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 283                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
 284
 285                 bp->v.mem_ptr           = 0;
 286                 bp->v.seq               = b->data->keys.seq;
 287                 bp->v.sectors_written   = 0;
 288                 bp->v.sectors           = cpu_to_le16(c->opts.btree_node_size);
 289         }
 290
 291         if (c->sb.features & (1ULL << BCH_FEATURE_new_extent_overwrite))
 292                 SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
 293
 294         if (btree_node_is_extents(b) &&
 295             !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data)) {
 296                 set_btree_node_old_extent_overwrite(b);
 297                 set_btree_node_need_rewrite(b);
 298         }
 299
 300         bch2_btree_build_aux_trees(b);
 301
 302         ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
 303         BUG_ON(ret);
 304
 305         trace_btree_node_alloc(c, b);
 306         return b;
 307 }
 308
 309 static void btree_set_min(struct btree *b, struct bpos pos)
 310 {
 311         if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
 312                 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
 313         b->data->min_key = pos;
 314 }
 315
 316 static void btree_set_max(struct btree *b, struct bpos pos)
 317 {
 318         b->key.k.p = pos;
 319         b->data->max_key = pos;
 320 }
 321
 322 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *as,
 323                                                   struct btree *b,
 324                                                   struct bkey_format format)
 325 {
 326         struct btree *n;
 327
 328         n = bch2_btree_node_alloc(as, b->c.level);
 329
 330         SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
 331
 332         btree_set_min(n, b->data->min_key);
 333         btree_set_max(n, b->data->max_key);
 334
 335         n->data->format         = format;
 336         btree_node_set_format(n, format);
 337
 338         bch2_btree_sort_into(as->c, n, b);
 339
 340         btree_node_reset_sib_u64s(n);
 341
 342         n->key.k.p = b->key.k.p;
 343         return n;
 344 }
 345
 346 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
 347                                                        struct btree *b)
 348 {
 349         struct bkey_format new_f = bch2_btree_calc_format(b);
 350
 351         /*
 352          * The keys might expand with the new format - if they wouldn't fit in
 353          * the btree node anymore, use the old format for now:
 354          */
 355         if (!bch2_btree_node_format_fits(as->c, b, &new_f))
 356                 new_f = b->format;
 357
 358         return __bch2_btree_node_alloc_replacement(as, b, new_f);
 359 }
 360
 361 static struct btree *__btree_root_alloc(struct btree_update *as, unsigned level)
 362 {
 363         struct btree *b = bch2_btree_node_alloc(as, level);
 364
 365         btree_set_min(b, POS_MIN);
 366         btree_set_max(b, POS_MAX);
 367         b->data->format = bch2_btree_calc_format(b);
 368
 369         btree_node_set_format(b, b->data->format);
 370         bch2_btree_build_aux_trees(b);
 371
 372         bch2_btree_update_add_new_node(as, b);
 373         six_unlock_write(&b->c.lock);
 374
 375         return b;
 376 }
 377
 378 static void bch2_btree_reserve_put(struct btree_update *as)
 379 {
 380         struct bch_fs *c = as->c;
 381
 382         mutex_lock(&c->btree_reserve_cache_lock);
 383
 384         while (as->nr_prealloc_nodes) {
 385                 struct btree *b = as->prealloc_nodes[--as->nr_prealloc_nodes];
 386
 387                 six_unlock_write(&b->c.lock);
 388
 389                 if (c->btree_reserve_cache_nr <
 390                     ARRAY_SIZE(c->btree_reserve_cache)) {
 391                         struct btree_alloc *a =
 392                                 &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
 393
 394                         a->ob = b->ob;
 395                         b->ob.nr = 0;
 396                         bkey_copy(&a->k, &b->key);
 397                 } else {
 398                         bch2_open_buckets_put(c, &b->ob);
 399                 }
 400
 401                 btree_node_lock_type(c, b, SIX_LOCK_write);
 402                 __btree_node_free(c, b);
 403                 six_unlock_write(&b->c.lock);
 404
 405                 six_unlock_intent(&b->c.lock);
 406         }
 407
 408         mutex_unlock(&c->btree_reserve_cache_lock);
 409 }
 410
 411 static int bch2_btree_reserve_get(struct btree_update *as, unsigned nr_nodes,
 412                                   unsigned flags, struct closure *cl)
 413 {
 414         struct bch_fs *c = as->c;
 415         struct btree *b;
 416         int ret;
 417
 418         BUG_ON(nr_nodes > BTREE_RESERVE_MAX);
 419
 420         /*
 421          * Protects reaping from the btree node cache and using the btree node
 422          * open bucket reserve:
 423          */
 424         ret = bch2_btree_cache_cannibalize_lock(c, cl);
 425         if (ret)
 426                 return ret;
 427
 428         while (as->nr_prealloc_nodes < nr_nodes) {
 429                 b = __bch2_btree_node_alloc(c, &as->disk_res,
 430                                             flags & BTREE_INSERT_NOWAIT
 431                                             ? NULL : cl, flags);
 432                 if (IS_ERR(b)) {
 433                         ret = PTR_ERR(b);
 434                         goto err_free;
 435                 }
 436
 437                 ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(&b->key));
 438                 if (ret)
 439                         goto err_free;
 440
 441                 as->prealloc_nodes[as->nr_prealloc_nodes++] = b;
 442         }
 443
 444         bch2_btree_cache_cannibalize_unlock(c);
 445         return 0;
 446 err_free:
 447         bch2_btree_cache_cannibalize_unlock(c);
 448         trace_btree_reserve_get_fail(c, nr_nodes, cl);
 449         return ret;
 450 }
 451
 452 /* Asynchronous interior node update machinery */
 453
 454 static void bch2_btree_update_free(struct btree_update *as)
 455 {
 456         struct bch_fs *c = as->c;
 457
 458         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 459
 460         bch2_journal_pin_drop(&c->journal, &as->journal);
 461         bch2_journal_pin_flush(&c->journal, &as->journal);
 462         bch2_disk_reservation_put(c, &as->disk_res);
 463         bch2_btree_reserve_put(as);
 464
 465         mutex_lock(&c->btree_interior_update_lock);
 466         list_del(&as->unwritten_list);
 467         list_del(&as->list);
 468         mutex_unlock(&c->btree_interior_update_lock);
 469
 470         closure_debug_destroy(&as->cl);
 471         mempool_free(as, &c->btree_interior_update_pool);
 472
 473         closure_wake_up(&c->btree_interior_update_wait);
 474 }
 475
 476 static void btree_update_will_delete_key(struct btree_update *as,
 477                                          struct bkey_i *k)
 478 {
 479         BUG_ON(bch2_keylist_u64s(&as->old_keys) + k->k.u64s >
 480                ARRAY_SIZE(as->_old_keys));
 481         bch2_keylist_add(&as->old_keys, k);
 482 }
 483
 484 static void btree_update_will_add_key(struct btree_update *as,
 485                                       struct bkey_i *k)
 486 {
 487         BUG_ON(bch2_keylist_u64s(&as->new_keys) + k->k.u64s >
 488                ARRAY_SIZE(as->_new_keys));
 489         bch2_keylist_add(&as->new_keys, k);
 490 }
 491
 492 /*
 493  * The transactional part of an interior btree node update, where we journal the
 494  * update we did to the interior node and update alloc info:
 495  */
 496 static int btree_update_nodes_written_trans(struct btree_trans *trans,
 497                                             struct btree_update *as)
 498 {
 499         struct bkey_i *k;
 500         int ret;
 501
 502         trans->extra_journal_entries = (void *) &as->journal_entries[0];
 503         trans->extra_journal_entry_u64s = as->journal_u64s;
 504         trans->journal_pin = &as->journal;
 505
 506         for_each_keylist_key(&as->new_keys, k) {
 507                 ret = bch2_trans_mark_key(trans, bkey_i_to_s_c(k),
 508                                           0, 0, BTREE_TRIGGER_INSERT);
 509                 if (ret)
 510                         return ret;
 511         }
 512
 513         for_each_keylist_key(&as->old_keys, k) {
 514                 ret = bch2_trans_mark_key(trans, bkey_i_to_s_c(k),
 515                                           0, 0, BTREE_TRIGGER_OVERWRITE);
 516                 if (ret)
 517                         return ret;
 518         }
 519
 520         return 0;
 521 }
 522
 523 static void btree_update_nodes_written(struct btree_update *as)
 524 {
 525         struct bch_fs *c = as->c;
 526         struct btree *b = as->b;
 527         struct btree_trans trans;
 528         u64 journal_seq = 0;
 529         unsigned i;
 530         int ret;
 531
 532         /*
 533          * We did an update to a parent node where the pointers we added pointed
 534          * to child nodes that weren't written yet: now, the child nodes have
 535          * been written so we can write out the update to the interior node.
 536          */
 537
 538         /*
 539          * We can't call into journal reclaim here: we'd block on the journal
 540          * reclaim lock, but we may need to release the open buckets we have
 541          * pinned in order for other btree updates to make forward progress, and
 542          * journal reclaim does btree updates when flushing bkey_cached entries,
 543          * which may require allocations as well.
 544          */
 545         bch2_trans_init(&trans, c, 0, 512);
 546         ret = __bch2_trans_do(&trans, &as->disk_res, &journal_seq,
 547                               BTREE_INSERT_NOFAIL|
 548                               BTREE_INSERT_USE_RESERVE|
 549                               BTREE_INSERT_USE_ALLOC_RESERVE|
 550                               BTREE_INSERT_NOCHECK_RW|
 551                               BTREE_INSERT_JOURNAL_RECLAIM|
 552                               BTREE_INSERT_JOURNAL_RESERVED,
 553                               btree_update_nodes_written_trans(&trans, as));
 554         bch2_trans_exit(&trans);
 555         BUG_ON(ret && !bch2_journal_error(&c->journal));
 556
 557         if (b) {
 558                 /*
 559                  * @b is the node we did the final insert into:
 560                  *
 561                  * On failure to get a journal reservation, we still have to
 562                  * unblock the write and allow most of the write path to happen
 563                  * so that shutdown works, but the i->journal_seq mechanism
 564                  * won't work to prevent the btree write from being visible (we
 565                  * didn't get a journal sequence number) - instead
 566                  * __bch2_btree_node_write() doesn't do the actual write if
 567                  * we're in journal error state:
 568                  */
 569
 570                 btree_node_lock_type(c, b, SIX_LOCK_intent);
 571                 btree_node_lock_type(c, b, SIX_LOCK_write);
 572                 mutex_lock(&c->btree_interior_update_lock);
 573
 574                 list_del(&as->write_blocked_list);
 575
 576                 if (!ret && as->b == b) {
 577                         struct bset *i = btree_bset_last(b);
 578
 579                         BUG_ON(!b->c.level);
 580                         BUG_ON(!btree_node_dirty(b));
 581
 582                         i->journal_seq = cpu_to_le64(
 583                                 max(journal_seq,
 584                                     le64_to_cpu(i->journal_seq)));
 585
 586                         bch2_btree_add_journal_pin(c, b, journal_seq);
 587                 }
 588
 589                 mutex_unlock(&c->btree_interior_update_lock);
 590                 six_unlock_write(&b->c.lock);
 591
 592                 btree_node_write_if_need(c, b, SIX_LOCK_intent);
 593                 six_unlock_intent(&b->c.lock);
 594         }
 595
 596         bch2_journal_pin_drop(&c->journal, &as->journal);
 597
 598         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 599
 600         mutex_lock(&c->btree_interior_update_lock);
 601         for (i = 0; i < as->nr_new_nodes; i++) {
 602                 b = as->new_nodes[i];
 603
 604                 BUG_ON(b->will_make_reachable != (unsigned long) as);
 605                 b->will_make_reachable = 0;
 606         }
 607         mutex_unlock(&c->btree_interior_update_lock);
 608
 609         for (i = 0; i < as->nr_new_nodes; i++) {
 610                 b = as->new_nodes[i];
 611
 612                 btree_node_lock_type(c, b, SIX_LOCK_read);
 613                 btree_node_write_if_need(c, b, SIX_LOCK_read);
 614                 six_unlock_read(&b->c.lock);
 615         }
 616
 617         for (i = 0; i < as->nr_open_buckets; i++)
 618                 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
 619
 620         bch2_btree_update_free(as);
 621 }
 622
 623 static void btree_interior_update_work(struct work_struct *work)
 624 {
 625         struct bch_fs *c =
 626                 container_of(work, struct bch_fs, btree_interior_update_work);
 627         struct btree_update *as;
 628
 629         while (1) {
 630                 mutex_lock(&c->btree_interior_update_lock);
 631                 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
 632                                               struct btree_update, unwritten_list);
 633                 if (as && !as->nodes_written)
 634                         as = NULL;
 635                 mutex_unlock(&c->btree_interior_update_lock);
 636
 637                 if (!as)
 638                         break;
 639
 640                 btree_update_nodes_written(as);
 641         }
 642 }
 643
 644 static void btree_update_set_nodes_written(struct closure *cl)
 645 {
 646         struct btree_update *as = container_of(cl, struct btree_update, cl);
 647         struct bch_fs *c = as->c;
 648
 649         mutex_lock(&c->btree_interior_update_lock);
 650         as->nodes_written = true;
 651         mutex_unlock(&c->btree_interior_update_lock);
 652
 653         queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
 654 }
 655
 656 /*
 657  * We're updating @b with pointers to nodes that haven't finished writing yet:
 658  * block @b from being written until @as completes
 659  */
 660 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
 661 {
 662         struct bch_fs *c = as->c;
 663
 664         mutex_lock(&c->btree_interior_update_lock);
 665         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 666
 667         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 668         BUG_ON(!btree_node_dirty(b));
 669
 670         as->mode        = BTREE_INTERIOR_UPDATING_NODE;
 671         as->b           = b;
 672         list_add(&as->write_blocked_list, &b->write_blocked);
 673
 674         mutex_unlock(&c->btree_interior_update_lock);
 675 }
 676
 677 static void btree_update_reparent(struct btree_update *as,
 678                                   struct btree_update *child)
 679 {
 680         struct bch_fs *c = as->c;
 681
 682         lockdep_assert_held(&c->btree_interior_update_lock);
 683
 684         child->b = NULL;
 685         child->mode = BTREE_INTERIOR_UPDATING_AS;
 686
 687         /*
 688          * When we write a new btree root, we have to drop our journal pin
 689          * _before_ the new nodes are technically reachable; see
 690          * btree_update_nodes_written().
 691          *
 692          * This goes for journal pins that are recursively blocked on us - so,
 693          * just transfer the journal pin to the new interior update so
 694          * btree_update_nodes_written() can drop it.
 695          */
 696         bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL);
 697         bch2_journal_pin_drop(&c->journal, &child->journal);
 698 }
 699
 700 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
 701 {
 702         struct bkey_i *insert = &b->key;
 703         struct bch_fs *c = as->c;
 704
 705         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 706
 707         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
 708                ARRAY_SIZE(as->journal_entries));
 709
 710         as->journal_u64s +=
 711                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
 712                                   BCH_JSET_ENTRY_btree_root,
 713                                   b->c.btree_id, b->c.level,
 714                                   insert, insert->k.u64s);
 715
 716         mutex_lock(&c->btree_interior_update_lock);
 717         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 718
 719         as->mode        = BTREE_INTERIOR_UPDATING_ROOT;
 720         mutex_unlock(&c->btree_interior_update_lock);
 721 }
 722
 723 /*
 724  * bch2_btree_update_add_new_node:
 725  *
 726  * This causes @as to wait on @b to be written, before it gets to
 727  * bch2_btree_update_nodes_written
 728  *
 729  * Additionally, it sets b->will_make_reachable to prevent any additional writes
 730  * to @b from happening besides the first until @b is reachable on disk
 731  *
 732  * And it adds @b to the list of @as's new nodes, so that we can update sector
 733  * counts in bch2_btree_update_nodes_written:
 734  */
 735 void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
 736 {
 737         struct bch_fs *c = as->c;
 738
 739         closure_get(&as->cl);
 740
 741         mutex_lock(&c->btree_interior_update_lock);
 742         BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
 743         BUG_ON(b->will_make_reachable);
 744
 745         as->new_nodes[as->nr_new_nodes++] = b;
 746         b->will_make_reachable = 1UL|(unsigned long) as;
 747
 748         mutex_unlock(&c->btree_interior_update_lock);
 749
 750         btree_update_will_add_key(as, &b->key);
 751 }
 752
 753 /*
 754  * returns true if @b was a new node
 755  */
 756 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
 757 {
 758         struct btree_update *as;
 759         unsigned long v;
 760         unsigned i;
 761
 762         mutex_lock(&c->btree_interior_update_lock);
 763         /*
 764          * When b->will_make_reachable != 0, it owns a ref on as->cl that's
 765          * dropped when it gets written by bch2_btree_complete_write - the
 766          * xchg() is for synchronization with bch2_btree_complete_write:
 767          */
 768         v = xchg(&b->will_make_reachable, 0);
 769         as = (struct btree_update *) (v & ~1UL);
 770
 771         if (!as) {
 772                 mutex_unlock(&c->btree_interior_update_lock);
 773                 return;
 774         }
 775
 776         for (i = 0; i < as->nr_new_nodes; i++)
 777                 if (as->new_nodes[i] == b)
 778                         goto found;
 779
 780         BUG();
 781 found:
 782         array_remove_item(as->new_nodes, as->nr_new_nodes, i);
 783         mutex_unlock(&c->btree_interior_update_lock);
 784
 785         if (v & 1)
 786                 closure_put(&as->cl);
 787 }
 788
 789 void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
 790 {
 791         while (b->ob.nr)
 792                 as->open_buckets[as->nr_open_buckets++] =
 793                         b->ob.v[--b->ob.nr];
 794 }
 795
 796 /*
 797  * @b is being split/rewritten: it may have pointers to not-yet-written btree
 798  * nodes and thus outstanding btree_updates - redirect @b's
 799  * btree_updates to point to this btree_update:
 800  */
 801 void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 802                                                struct btree *b)
 803 {
 804         struct bch_fs *c = as->c;
 805         struct btree_update *p, *n;
 806         struct btree_write *w;
 807
 808         set_btree_node_dying(b);
 809
 810         if (btree_node_fake(b))
 811                 return;
 812
 813         mutex_lock(&c->btree_interior_update_lock);
 814
 815         /*
 816          * Does this node have any btree_update operations preventing
 817          * it from being written?
 818          *
 819          * If so, redirect them to point to this btree_update: we can
 820          * write out our new nodes, but we won't make them visible until those
 821          * operations complete
 822          */
 823         list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
 824                 list_del_init(&p->write_blocked_list);
 825                 btree_update_reparent(as, p);
 826
 827                 /*
 828                  * for flush_held_btree_writes() waiting on updates to flush or
 829                  * nodes to be writeable:
 830                  */
 831                 closure_wake_up(&c->btree_interior_update_wait);
 832         }
 833
 834         clear_btree_node_dirty(c, b);
 835         clear_btree_node_need_write(b);
 836
 837         /*
 838          * Does this node have unwritten data that has a pin on the journal?
 839          *
 840          * If so, transfer that pin to the btree_update operation -
 841          * note that if we're freeing multiple nodes, we only need to keep the
 842          * oldest pin of any of the nodes we're freeing. We'll release the pin
 843          * when the new nodes are persistent and reachable on disk:
 844          */
 845         w = btree_current_write(b);
 846         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 847         bch2_journal_pin_drop(&c->journal, &w->journal);
 848
 849         w = btree_prev_write(b);
 850         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 851         bch2_journal_pin_drop(&c->journal, &w->journal);
 852
 853         mutex_unlock(&c->btree_interior_update_lock);
 854
 855         /*
 856          * Is this a node that isn't reachable on disk yet?
 857          *
 858          * Nodes that aren't reachable yet have writes blocked until they're
 859          * reachable - now that we've cancelled any pending writes and moved
 860          * things waiting on that write to wait on this update, we can drop this
 861          * node from the list of nodes that the other update is making
 862          * reachable, prior to freeing it:
 863          */
 864         btree_update_drop_new_node(c, b);
 865
 866         btree_update_will_delete_key(as, &b->key);
 867 }
 868
 869 void bch2_btree_update_done(struct btree_update *as)
 870 {
 871         BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
 872
 873         bch2_btree_reserve_put(as);
 874
 875         continue_at(&as->cl, btree_update_set_nodes_written, system_freezable_wq);
 876 }
 877
 878 struct btree_update *
 879 bch2_btree_update_start(struct btree_trans *trans, enum btree_id id,
 880                         unsigned nr_nodes, unsigned flags,
 881                         struct closure *cl)
 882 {
 883         struct bch_fs *c = trans->c;
 884         struct btree_update *as;
 885         int disk_res_flags = (flags & BTREE_INSERT_NOFAIL)
 886                 ? BCH_DISK_RESERVATION_NOFAIL : 0;
 887         int journal_flags = (flags & BTREE_INSERT_JOURNAL_RESERVED)
 888                 ? JOURNAL_RES_GET_RECLAIM : 0;
 889         int ret = 0;
 890
 891         /*
 892          * This check isn't necessary for correctness - it's just to potentially
 893          * prevent us from doing a lot of work that'll end up being wasted:
 894          */
 895         ret = bch2_journal_error(&c->journal);
 896         if (ret)
 897                 return ERR_PTR(ret);
 898
 899         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
 900         memset(as, 0, sizeof(*as));
 901         closure_init(&as->cl, NULL);
 902         as->c           = c;
 903         as->mode        = BTREE_INTERIOR_NO_UPDATE;
 904         as->btree_id    = id;
 905         INIT_LIST_HEAD(&as->list);
 906         INIT_LIST_HEAD(&as->unwritten_list);
 907         INIT_LIST_HEAD(&as->write_blocked_list);
 908         bch2_keylist_init(&as->old_keys, as->_old_keys);
 909         bch2_keylist_init(&as->new_keys, as->_new_keys);
 910         bch2_keylist_init(&as->parent_keys, as->inline_keys);
 911
 912         ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
 913                                       BTREE_UPDATE_JOURNAL_RES,
 914                                       journal_flags|JOURNAL_RES_GET_NONBLOCK);
 915         if (ret == -EAGAIN) {
 916                 if (flags & BTREE_INSERT_NOUNLOCK)
 917                         return ERR_PTR(-EINTR);
 918
 919                 bch2_trans_unlock(trans);
 920
 921                 ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
 922                                 BTREE_UPDATE_JOURNAL_RES,
 923                                 journal_flags);
 924                 if (ret)
 925                         return ERR_PTR(ret);
 926
 927                 if (!bch2_trans_relock(trans)) {
 928                         ret = -EINTR;
 929                         goto err;
 930                 }
 931         }
 932
 933         ret = bch2_disk_reservation_get(c, &as->disk_res,
 934                         nr_nodes * c->opts.btree_node_size,
 935                         c->opts.metadata_replicas,
 936                         disk_res_flags);
 937         if (ret)
 938                 goto err;
 939
 940         ret = bch2_btree_reserve_get(as, nr_nodes, flags, cl);
 941         if (ret)
 942                 goto err;
 943
 944         mutex_lock(&c->btree_interior_update_lock);
 945         list_add_tail(&as->list, &c->btree_interior_update_list);
 946         mutex_unlock(&c->btree_interior_update_lock);
 947
 948         return as;
 949 err:
 950         bch2_btree_update_free(as);
 951         return ERR_PTR(ret);
 952 }
 953
 954 /* Btree root updates: */
 955
 956 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
 957 {
 958         /* Root nodes cannot be reaped */
 959         mutex_lock(&c->btree_cache.lock);
 960         list_del_init(&b->list);
 961         mutex_unlock(&c->btree_cache.lock);
 962
 963         mutex_lock(&c->btree_root_lock);
 964         BUG_ON(btree_node_root(c, b) &&
 965                (b->c.level < btree_node_root(c, b)->c.level ||
 966                 !btree_node_dying(btree_node_root(c, b))));
 967
 968         btree_node_root(c, b) = b;
 969         mutex_unlock(&c->btree_root_lock);
 970
 971         bch2_recalc_btree_reserve(c);
 972 }
 973
 974 /**
 975  * bch_btree_set_root - update the root in memory and on disk
 976  *
 977  * To ensure forward progress, the current task must not be holding any
 978  * btree node write locks. However, you must hold an intent lock on the
 979  * old root.
 980  *
 981  * Note: This allocates a journal entry but doesn't add any keys to
 982  * it.  All the btree roots are part of every journal write, so there
 983  * is nothing new to be done.  This just guarantees that there is a
 984  * journal write.
 985  */
 986 static void bch2_btree_set_root(struct btree_update *as, struct btree *b,
 987                                 struct btree_iter *iter)
 988 {
 989         struct bch_fs *c = as->c;
 990         struct btree *old;
 991
 992         trace_btree_set_root(c, b);
 993         BUG_ON(!b->written &&
 994                !test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags));
 995
 996         old = btree_node_root(c, b);
 997
 998         /*
 999          * Ensure no one is using the old root while we switch to the
1000          * new root:
1001          */
1002         bch2_btree_node_lock_write(old, iter);
1003
1004         bch2_btree_set_root_inmem(c, b);
1005
1006         btree_update_updated_root(as, b);
1007
1008         /*
1009          * Unlock old root after new root is visible:
1010          *
1011          * The new root isn't persistent, but that's ok: we still have
1012          * an intent lock on the new root, and any updates that would
1013          * depend on the new root would have to update the new root.
1014          */
1015         bch2_btree_node_unlock_write(old, iter);
1016 }
1017
1018 /* Interior node updates: */
1019
1020 static void bch2_insert_fixup_btree_ptr(struct btree_update *as, struct btree *b,
1021                                         struct btree_iter *iter,
1022                                         struct bkey_i *insert,
1023                                         struct btree_node_iter *node_iter)
1024 {
1025         struct bch_fs *c = as->c;
1026         struct bkey_packed *k;
1027         const char *invalid;
1028
1029         invalid = bch2_bkey_invalid(c, bkey_i_to_s_c(insert), btree_node_type(b));
1030         if (invalid) {
1031                 char buf[160];
1032
1033                 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(insert));
1034                 bch2_fs_inconsistent(c, "inserting invalid bkey %s: %s", buf, invalid);
1035                 dump_stack();
1036         }
1037
1038         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1039                ARRAY_SIZE(as->journal_entries));
1040
1041         as->journal_u64s +=
1042                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1043                                   BCH_JSET_ENTRY_btree_keys,
1044                                   b->c.btree_id, b->c.level,
1045                                   insert, insert->k.u64s);
1046
1047         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1048                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1049                 bch2_btree_node_iter_advance(node_iter, b);
1050
1051         bch2_btree_bset_insert_key(iter, b, node_iter, insert);
1052         set_btree_node_dirty(c, b);
1053         set_btree_node_need_write(b);
1054 }
1055
1056 /*
1057  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1058  * node)
1059  */
1060 static struct btree *__btree_split_node(struct btree_update *as,
1061                                         struct btree *n1,
1062                                         struct btree_iter *iter)
1063 {
1064         size_t nr_packed = 0, nr_unpacked = 0;
1065         struct btree *n2;
1066         struct bset *set1, *set2;
1067         struct bkey_packed *k, *prev = NULL;
1068
1069         n2 = bch2_btree_node_alloc(as, n1->c.level);
1070         bch2_btree_update_add_new_node(as, n2);
1071
1072         n2->data->max_key       = n1->data->max_key;
1073         n2->data->format        = n1->format;
1074         SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data));
1075         n2->key.k.p = n1->key.k.p;
1076
1077         btree_node_set_format(n2, n2->data->format);
1078
1079         set1 = btree_bset_first(n1);
1080         set2 = btree_bset_first(n2);
1081
1082         /*
1083          * Has to be a linear search because we don't have an auxiliary
1084          * search tree yet
1085          */
1086         k = set1->start;
1087         while (1) {
1088                 struct bkey_packed *n = bkey_next_skip_noops(k, vstruct_last(set1));
1089
1090                 if (n == vstruct_last(set1))
1091                         break;
1092                 if (k->_data - set1->_data >= (le16_to_cpu(set1->u64s) * 3) / 5)
1093                         break;
1094
1095                 if (bkey_packed(k))
1096                         nr_packed++;
1097                 else
1098                         nr_unpacked++;
1099
1100                 prev = k;
1101                 k = n;
1102         }
1103
1104         BUG_ON(!prev);
1105
1106         btree_set_max(n1, bkey_unpack_pos(n1, prev));
1107         btree_set_min(n2, bkey_successor(n1->key.k.p));
1108
1109         set2->u64s = cpu_to_le16((u64 *) vstruct_end(set1) - (u64 *) k);
1110         set1->u64s = cpu_to_le16(le16_to_cpu(set1->u64s) - le16_to_cpu(set2->u64s));
1111
1112         set_btree_bset_end(n1, n1->set);
1113         set_btree_bset_end(n2, n2->set);
1114
1115         n2->nr.live_u64s        = le16_to_cpu(set2->u64s);
1116         n2->nr.bset_u64s[0]     = le16_to_cpu(set2->u64s);
1117         n2->nr.packed_keys      = n1->nr.packed_keys - nr_packed;
1118         n2->nr.unpacked_keys    = n1->nr.unpacked_keys - nr_unpacked;
1119
1120         n1->nr.live_u64s        = le16_to_cpu(set1->u64s);
1121         n1->nr.bset_u64s[0]     = le16_to_cpu(set1->u64s);
1122         n1->nr.packed_keys      = nr_packed;
1123         n1->nr.unpacked_keys    = nr_unpacked;
1124
1125         BUG_ON(!set1->u64s);
1126         BUG_ON(!set2->u64s);
1127
1128         memcpy_u64s(set2->start,
1129                     vstruct_end(set1),
1130                     le16_to_cpu(set2->u64s));
1131
1132         btree_node_reset_sib_u64s(n1);
1133         btree_node_reset_sib_u64s(n2);
1134
1135         bch2_verify_btree_nr_keys(n1);
1136         bch2_verify_btree_nr_keys(n2);
1137
1138         if (n1->c.level) {
1139                 btree_node_interior_verify(as->c, n1);
1140                 btree_node_interior_verify(as->c, n2);
1141         }
1142
1143         return n2;
1144 }
1145
1146 /*
1147  * For updates to interior nodes, we've got to do the insert before we split
1148  * because the stuff we're inserting has to be inserted atomically. Post split,
1149  * the keys might have to go in different nodes and the split would no longer be
1150  * atomic.
1151  *
1152  * Worse, if the insert is from btree node coalescing, if we do the insert after
1153  * we do the split (and pick the pivot) - the pivot we pick might be between
1154  * nodes that were coalesced, and thus in the middle of a child node post
1155  * coalescing:
1156  */
1157 static void btree_split_insert_keys(struct btree_update *as, struct btree *b,
1158                                     struct btree_iter *iter,
1159                                     struct keylist *keys)
1160 {
1161         struct btree_node_iter node_iter;
1162         struct bkey_i *k = bch2_keylist_front(keys);
1163         struct bkey_packed *src, *dst, *n;
1164         struct bset *i;
1165
1166         BUG_ON(btree_node_type(b) != BKEY_TYPE_BTREE);
1167
1168         bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
1169
1170         while (!bch2_keylist_empty(keys)) {
1171                 k = bch2_keylist_front(keys);
1172
1173                 bch2_insert_fixup_btree_ptr(as, b, iter, k, &node_iter);
1174                 bch2_keylist_pop_front(keys);
1175         }
1176
1177         /*
1178          * We can't tolerate whiteouts here - with whiteouts there can be
1179          * duplicate keys, and it would be rather bad if we picked a duplicate
1180          * for the pivot:
1181          */
1182         i = btree_bset_first(b);
1183         src = dst = i->start;
1184         while (src != vstruct_last(i)) {
1185                 n = bkey_next_skip_noops(src, vstruct_last(i));
1186                 if (!bkey_deleted(src)) {
1187                         memmove_u64s_down(dst, src, src->u64s);
1188                         dst = bkey_next(dst);
1189                 }
1190                 src = n;
1191         }
1192
1193         i->u64s = cpu_to_le16((u64 *) dst - i->_data);
1194         set_btree_bset_end(b, b->set);
1195
1196         BUG_ON(b->nsets != 1 ||
1197                b->nr.live_u64s != le16_to_cpu(btree_bset_first(b)->u64s));
1198
1199         btree_node_interior_verify(as->c, b);
1200 }
1201
1202 static void btree_split(struct btree_update *as, struct btree *b,
1203                         struct btree_iter *iter, struct keylist *keys,
1204                         unsigned flags)
1205 {
1206         struct bch_fs *c = as->c;
1207         struct btree *parent = btree_node_parent(iter, b);
1208         struct btree *n1, *n2 = NULL, *n3 = NULL;
1209         u64 start_time = local_clock();
1210
1211         BUG_ON(!parent && (b != btree_node_root(c, b)));
1212         BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->c.level));
1213
1214         bch2_btree_interior_update_will_free_node(as, b);
1215
1216         n1 = bch2_btree_node_alloc_replacement(as, b);
1217         bch2_btree_update_add_new_node(as, n1);
1218
1219         if (keys)
1220                 btree_split_insert_keys(as, n1, iter, keys);
1221
1222         if (bset_u64s(&n1->set[0]) > BTREE_SPLIT_THRESHOLD(c)) {
1223                 trace_btree_split(c, b);
1224
1225                 n2 = __btree_split_node(as, n1, iter);
1226
1227                 bch2_btree_build_aux_trees(n2);
1228                 bch2_btree_build_aux_trees(n1);
1229                 six_unlock_write(&n2->c.lock);
1230                 six_unlock_write(&n1->c.lock);
1231
1232                 bch2_btree_node_write(c, n2, SIX_LOCK_intent);
1233
1234                 /*
1235                  * Note that on recursive parent_keys == keys, so we
1236                  * can't start adding new keys to parent_keys before emptying it
1237                  * out (which we did with btree_split_insert_keys() above)
1238                  */
1239                 bch2_keylist_add(&as->parent_keys, &n1->key);
1240                 bch2_keylist_add(&as->parent_keys, &n2->key);
1241
1242                 if (!parent) {
1243                         /* Depth increases, make a new root */
1244                         n3 = __btree_root_alloc(as, b->c.level + 1);
1245
1246                         n3->sib_u64s[0] = U16_MAX;
1247                         n3->sib_u64s[1] = U16_MAX;
1248
1249                         btree_split_insert_keys(as, n3, iter, &as->parent_keys);
1250
1251                         bch2_btree_node_write(c, n3, SIX_LOCK_intent);
1252                 }
1253         } else {
1254                 trace_btree_compact(c, b);
1255
1256                 bch2_btree_build_aux_trees(n1);
1257                 six_unlock_write(&n1->c.lock);
1258
1259                 if (parent)
1260                         bch2_keylist_add(&as->parent_keys, &n1->key);
1261         }
1262
1263         bch2_btree_node_write(c, n1, SIX_LOCK_intent);
1264
1265         /* New nodes all written, now make them visible: */
1266
1267         if (parent) {
1268                 /* Split a non root node */
1269                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1270         } else if (n3) {
1271                 bch2_btree_set_root(as, n3, iter);
1272         } else {
1273                 /* Root filled up but didn't need to be split */
1274                 bch2_btree_set_root(as, n1, iter);
1275         }
1276
1277         bch2_btree_update_get_open_buckets(as, n1);
1278         if (n2)
1279                 bch2_btree_update_get_open_buckets(as, n2);
1280         if (n3)
1281                 bch2_btree_update_get_open_buckets(as, n3);
1282
1283         /* Successful split, update the iterator to point to the new nodes: */
1284
1285         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1286         bch2_btree_iter_node_drop(iter, b);
1287         if (n3)
1288                 bch2_btree_iter_node_replace(iter, n3);
1289         if (n2)
1290                 bch2_btree_iter_node_replace(iter, n2);
1291         bch2_btree_iter_node_replace(iter, n1);
1292
1293         /*
1294          * The old node must be freed (in memory) _before_ unlocking the new
1295          * nodes - else another thread could re-acquire a read lock on the old
1296          * node after another thread has locked and updated the new node, thus
1297          * seeing stale data:
1298          */
1299         bch2_btree_node_free_inmem(c, b, iter);
1300
1301         if (n3)
1302                 six_unlock_intent(&n3->c.lock);
1303         if (n2)
1304                 six_unlock_intent(&n2->c.lock);
1305         six_unlock_intent(&n1->c.lock);
1306
1307         bch2_btree_trans_verify_locks(iter->trans);
1308
1309         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_split],
1310                                start_time);
1311 }
1312
1313 static void
1314 bch2_btree_insert_keys_interior(struct btree_update *as, struct btree *b,
1315                                 struct btree_iter *iter, struct keylist *keys)
1316 {
1317         struct btree_iter *linked;
1318         struct btree_node_iter node_iter;
1319         struct bkey_i *insert = bch2_keylist_front(keys);
1320         struct bkey_packed *k;
1321
1322         /* Don't screw up @iter's position: */
1323         node_iter = iter->l[b->c.level].iter;
1324
1325         /*
1326          * btree_split(), btree_gc_coalesce() will insert keys before
1327          * the iterator's current position - they know the keys go in
1328          * the node the iterator points to:
1329          */
1330         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1331                (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1332                 ;
1333
1334         for_each_keylist_key(keys, insert)
1335                 bch2_insert_fixup_btree_ptr(as, b, iter, insert, &node_iter);
1336
1337         btree_update_updated_node(as, b);
1338
1339         trans_for_each_iter_with_node(iter->trans, b, linked)
1340                 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1341
1342         bch2_btree_trans_verify_iters(iter->trans, b);
1343 }
1344
1345 /**
1346  * bch_btree_insert_node - insert bkeys into a given btree node
1347  *
1348  * @iter:               btree iterator
1349  * @keys:               list of keys to insert
1350  * @hook:               insert callback
1351  * @persistent:         if not null, @persistent will wait on journal write
1352  *
1353  * Inserts as many keys as it can into a given btree node, splitting it if full.
1354  * If a split occurred, this function will return early. This can only happen
1355  * for leaf nodes -- inserts into interior nodes have to be atomic.
1356  */
1357 void bch2_btree_insert_node(struct btree_update *as, struct btree *b,
1358                             struct btree_iter *iter, struct keylist *keys,
1359                             unsigned flags)
1360 {
1361         struct bch_fs *c = as->c;
1362         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1363         int old_live_u64s = b->nr.live_u64s;
1364         int live_u64s_added, u64s_added;
1365
1366         BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->c.level));
1367         BUG_ON(!b->c.level);
1368         BUG_ON(!as || as->b);
1369         bch2_verify_keylist_sorted(keys);
1370
1371         if (as->must_rewrite)
1372                 goto split;
1373
1374         bch2_btree_node_lock_for_insert(c, b, iter);
1375
1376         if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1377                 bch2_btree_node_unlock_write(b, iter);
1378                 goto split;
1379         }
1380
1381         bch2_btree_insert_keys_interior(as, b, iter, keys);
1382
1383         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1384         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1385
1386         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1387                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1388         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1389                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1390
1391         if (u64s_added > live_u64s_added &&
1392             bch2_maybe_compact_whiteouts(c, b))
1393                 bch2_btree_iter_reinit_node(iter, b);
1394
1395         bch2_btree_node_unlock_write(b, iter);
1396
1397         btree_node_interior_verify(c, b);
1398
1399         /*
1400          * when called from the btree_split path the new nodes aren't added to
1401          * the btree iterator yet, so the merge path's unlock/wait/relock dance
1402          * won't work:
1403          */
1404         bch2_foreground_maybe_merge(c, iter, b->c.level,
1405                                     flags|BTREE_INSERT_NOUNLOCK);
1406         return;
1407 split:
1408         btree_split(as, b, iter, keys, flags);
1409 }
1410
1411 int bch2_btree_split_leaf(struct bch_fs *c, struct btree_iter *iter,
1412                           unsigned flags)
1413 {
1414         struct btree_trans *trans = iter->trans;
1415         struct btree *b = iter_l(iter)->b;
1416         struct btree_update *as;
1417         struct closure cl;
1418         int ret = 0;
1419         struct btree_insert_entry *i;
1420
1421         /*
1422          * We already have a disk reservation and open buckets pinned; this
1423          * allocation must not block:
1424          */
1425         trans_for_each_update(trans, i)
1426                 if (btree_node_type_needs_gc(i->iter->btree_id))
1427                         flags |= BTREE_INSERT_USE_RESERVE;
1428
1429         closure_init_stack(&cl);
1430
1431         /* Hack, because gc and splitting nodes doesn't mix yet: */
1432         if (!(flags & BTREE_INSERT_GC_LOCK_HELD) &&
1433             !down_read_trylock(&c->gc_lock)) {
1434                 if (flags & BTREE_INSERT_NOUNLOCK) {
1435                         trace_transaction_restart_ip(trans->ip, _THIS_IP_);
1436                         return -EINTR;
1437                 }
1438
1439                 bch2_trans_unlock(trans);
1440                 down_read(&c->gc_lock);
1441
1442                 if (!bch2_trans_relock(trans))
1443                         ret = -EINTR;
1444         }
1445
1446         /*
1447          * XXX: figure out how far we might need to split,
1448          * instead of locking/reserving all the way to the root:
1449          */
1450         if (!bch2_btree_iter_upgrade(iter, U8_MAX)) {
1451                 trace_trans_restart_iter_upgrade(trans->ip);
1452                 ret = -EINTR;
1453                 goto out;
1454         }
1455
1456         as = bch2_btree_update_start(trans, iter->btree_id,
1457                 btree_update_reserve_required(c, b), flags,
1458                 !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1459         if (IS_ERR(as)) {
1460                 ret = PTR_ERR(as);
1461                 if (ret == -EAGAIN) {
1462                         BUG_ON(flags & BTREE_INSERT_NOUNLOCK);
1463                         bch2_trans_unlock(trans);
1464                         ret = -EINTR;
1465
1466                         trace_transaction_restart_ip(trans->ip, _THIS_IP_);
1467                 }
1468                 goto out;
1469         }
1470
1471         btree_split(as, b, iter, NULL, flags);
1472         bch2_btree_update_done(as);
1473
1474         /*
1475          * We haven't successfully inserted yet, so don't downgrade all the way
1476          * back to read locks;
1477          */
1478         __bch2_btree_iter_downgrade(iter, 1);
1479 out:
1480         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1481                 up_read(&c->gc_lock);
1482         closure_sync(&cl);
1483         return ret;
1484 }
1485
1486 void __bch2_foreground_maybe_merge(struct bch_fs *c,
1487                                    struct btree_iter *iter,
1488                                    unsigned level,
1489                                    unsigned flags,
1490                                    enum btree_node_sibling sib)
1491 {
1492         struct btree_trans *trans = iter->trans;
1493         struct btree_update *as;
1494         struct bkey_format_state new_s;
1495         struct bkey_format new_f;
1496         struct bkey_i delete;
1497         struct btree *b, *m, *n, *prev, *next, *parent;
1498         struct closure cl;
1499         size_t sib_u64s;
1500         int ret = 0;
1501
1502         BUG_ON(!btree_node_locked(iter, level));
1503
1504         closure_init_stack(&cl);
1505 retry:
1506         BUG_ON(!btree_node_locked(iter, level));
1507
1508         b = iter->l[level].b;
1509
1510         parent = btree_node_parent(iter, b);
1511         if (!parent)
1512                 goto out;
1513
1514         if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c))
1515                 goto out;
1516
1517         /* XXX: can't be holding read locks */
1518         m = bch2_btree_node_get_sibling(c, iter, b, sib);
1519         if (IS_ERR(m)) {
1520                 ret = PTR_ERR(m);
1521                 goto err;
1522         }
1523
1524         /* NULL means no sibling: */
1525         if (!m) {
1526                 b->sib_u64s[sib] = U16_MAX;
1527                 goto out;
1528         }
1529
1530         if (sib == btree_prev_sib) {
1531                 prev = m;
1532                 next = b;
1533         } else {
1534                 prev = b;
1535                 next = m;
1536         }
1537
1538         bch2_bkey_format_init(&new_s);
1539         __bch2_btree_calc_format(&new_s, b);
1540         __bch2_btree_calc_format(&new_s, m);
1541         new_f = bch2_bkey_format_done(&new_s);
1542
1543         sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1544                 btree_node_u64s_with_format(m, &new_f);
1545
1546         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1547                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1548                 sib_u64s /= 2;
1549                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1550         }
1551
1552         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1553         b->sib_u64s[sib] = sib_u64s;
1554
1555         if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c)) {
1556                 six_unlock_intent(&m->c.lock);
1557                 goto out;
1558         }
1559
1560         /* We're changing btree topology, doesn't mix with gc: */
1561         if (!(flags & BTREE_INSERT_GC_LOCK_HELD) &&
1562             !down_read_trylock(&c->gc_lock))
1563                 goto err_cycle_gc_lock;
1564
1565         if (!bch2_btree_iter_upgrade(iter, U8_MAX)) {
1566                 ret = -EINTR;
1567                 goto err_unlock;
1568         }
1569
1570         as = bch2_btree_update_start(trans, iter->btree_id,
1571                          btree_update_reserve_required(c, parent) + 1,
1572                          flags|
1573                          BTREE_INSERT_NOFAIL|
1574                          BTREE_INSERT_USE_RESERVE,
1575                          !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1576         if (IS_ERR(as)) {
1577                 ret = PTR_ERR(as);
1578                 goto err_unlock;
1579         }
1580
1581         trace_btree_merge(c, b);
1582
1583         bch2_btree_interior_update_will_free_node(as, b);
1584         bch2_btree_interior_update_will_free_node(as, m);
1585
1586         n = bch2_btree_node_alloc(as, b->c.level);
1587         bch2_btree_update_add_new_node(as, n);
1588
1589         btree_set_min(n, prev->data->min_key);
1590         btree_set_max(n, next->data->max_key);
1591         n->data->format         = new_f;
1592
1593         btree_node_set_format(n, new_f);
1594
1595         bch2_btree_sort_into(c, n, prev);
1596         bch2_btree_sort_into(c, n, next);
1597
1598         bch2_btree_build_aux_trees(n);
1599         six_unlock_write(&n->c.lock);
1600
1601         bkey_init(&delete.k);
1602         delete.k.p = prev->key.k.p;
1603         bch2_keylist_add(&as->parent_keys, &delete);
1604         bch2_keylist_add(&as->parent_keys, &n->key);
1605
1606         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1607
1608         bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1609
1610         bch2_btree_update_get_open_buckets(as, n);
1611
1612         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1613         bch2_btree_iter_node_drop(iter, b);
1614         bch2_btree_iter_node_drop(iter, m);
1615
1616         bch2_btree_iter_node_replace(iter, n);
1617
1618         bch2_btree_trans_verify_iters(trans, n);
1619
1620         bch2_btree_node_free_inmem(c, b, iter);
1621         bch2_btree_node_free_inmem(c, m, iter);
1622
1623         six_unlock_intent(&n->c.lock);
1624
1625         bch2_btree_update_done(as);
1626
1627         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1628                 up_read(&c->gc_lock);
1629 out:
1630         bch2_btree_trans_verify_locks(trans);
1631
1632         /*
1633          * Don't downgrade locks here: we're called after successful insert,
1634          * and the caller will downgrade locks after a successful insert
1635          * anyways (in case e.g. a split was required first)
1636          *
1637          * And we're also called when inserting into interior nodes in the
1638          * split path, and downgrading to read locks in there is potentially
1639          * confusing:
1640          */
1641         closure_sync(&cl);
1642         return;
1643
1644 err_cycle_gc_lock:
1645         six_unlock_intent(&m->c.lock);
1646
1647         if (flags & BTREE_INSERT_NOUNLOCK)
1648                 goto out;
1649
1650         bch2_trans_unlock(trans);
1651
1652         down_read(&c->gc_lock);
1653         up_read(&c->gc_lock);
1654         ret = -EINTR;
1655         goto err;
1656
1657 err_unlock:
1658         six_unlock_intent(&m->c.lock);
1659         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1660                 up_read(&c->gc_lock);
1661 err:
1662         BUG_ON(ret == -EAGAIN && (flags & BTREE_INSERT_NOUNLOCK));
1663
1664         if ((ret == -EAGAIN || ret == -EINTR) &&
1665             !(flags & BTREE_INSERT_NOUNLOCK)) {
1666                 bch2_trans_unlock(trans);
1667                 closure_sync(&cl);
1668                 ret = bch2_btree_iter_traverse(iter);
1669                 if (ret)
1670                         goto out;
1671
1672                 goto retry;
1673         }
1674
1675         goto out;
1676 }
1677
1678 static int __btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
1679                                 struct btree *b, unsigned flags,
1680                                 struct closure *cl)
1681 {
1682         struct btree *n, *parent = btree_node_parent(iter, b);
1683         struct btree_update *as;
1684
1685         as = bch2_btree_update_start(iter->trans, iter->btree_id,
1686                 (parent
1687                  ? btree_update_reserve_required(c, parent)
1688                  : 0) + 1,
1689                 flags, cl);
1690         if (IS_ERR(as)) {
1691                 trace_btree_gc_rewrite_node_fail(c, b);
1692                 return PTR_ERR(as);
1693         }
1694
1695         bch2_btree_interior_update_will_free_node(as, b);
1696
1697         n = bch2_btree_node_alloc_replacement(as, b);
1698         bch2_btree_update_add_new_node(as, n);
1699
1700         bch2_btree_build_aux_trees(n);
1701         six_unlock_write(&n->c.lock);
1702
1703         trace_btree_gc_rewrite_node(c, b);
1704
1705         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1706
1707         if (parent) {
1708                 bch2_keylist_add(&as->parent_keys, &n->key);
1709                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1710         } else {
1711                 bch2_btree_set_root(as, n, iter);
1712         }
1713
1714         bch2_btree_update_get_open_buckets(as, n);
1715
1716         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1717         bch2_btree_iter_node_drop(iter, b);
1718         bch2_btree_iter_node_replace(iter, n);
1719         bch2_btree_node_free_inmem(c, b, iter);
1720         six_unlock_intent(&n->c.lock);
1721
1722         bch2_btree_update_done(as);
1723         return 0;
1724 }
1725
1726 /**
1727  * bch_btree_node_rewrite - Rewrite/move a btree node
1728  *
1729  * Returns 0 on success, -EINTR or -EAGAIN on failure (i.e.
1730  * btree_check_reserve() has to wait)
1731  */
1732 int bch2_btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
1733                             __le64 seq, unsigned flags)
1734 {
1735         struct btree_trans *trans = iter->trans;
1736         struct closure cl;
1737         struct btree *b;
1738         int ret;
1739
1740         flags |= BTREE_INSERT_NOFAIL;
1741
1742         closure_init_stack(&cl);
1743
1744         bch2_btree_iter_upgrade(iter, U8_MAX);
1745
1746         if (!(flags & BTREE_INSERT_GC_LOCK_HELD)) {
1747                 if (!down_read_trylock(&c->gc_lock)) {
1748                         bch2_trans_unlock(trans);
1749                         down_read(&c->gc_lock);
1750                 }
1751         }
1752
1753         while (1) {
1754                 ret = bch2_btree_iter_traverse(iter);
1755                 if (ret)
1756                         break;
1757
1758                 b = bch2_btree_iter_peek_node(iter);
1759                 if (!b || b->data->keys.seq != seq)
1760                         break;
1761
1762                 ret = __btree_node_rewrite(c, iter, b, flags, &cl);
1763                 if (ret != -EAGAIN &&
1764                     ret != -EINTR)
1765                         break;
1766
1767                 bch2_trans_unlock(trans);
1768                 closure_sync(&cl);
1769         }
1770
1771         bch2_btree_iter_downgrade(iter);
1772
1773         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1774                 up_read(&c->gc_lock);
1775
1776         closure_sync(&cl);
1777         return ret;
1778 }
1779
1780 static void __bch2_btree_node_update_key(struct bch_fs *c,
1781                                          struct btree_update *as,
1782                                          struct btree_iter *iter,
1783                                          struct btree *b, struct btree *new_hash,
1784                                          struct bkey_i *new_key)
1785 {
1786         struct btree *parent;
1787         int ret;
1788
1789         btree_update_will_delete_key(as, &b->key);
1790         btree_update_will_add_key(as, new_key);
1791
1792         parent = btree_node_parent(iter, b);
1793         if (parent) {
1794                 if (new_hash) {
1795                         bkey_copy(&new_hash->key, new_key);
1796                         ret = bch2_btree_node_hash_insert(&c->btree_cache,
1797                                         new_hash, b->c.level, b->c.btree_id);
1798                         BUG_ON(ret);
1799                 }
1800
1801                 bch2_keylist_add(&as->parent_keys, new_key);
1802                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, 0);
1803
1804                 if (new_hash) {
1805                         mutex_lock(&c->btree_cache.lock);
1806                         bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
1807
1808                         bch2_btree_node_hash_remove(&c->btree_cache, b);
1809
1810                         bkey_copy(&b->key, new_key);
1811                         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1812                         BUG_ON(ret);
1813                         mutex_unlock(&c->btree_cache.lock);
1814                 } else {
1815                         bkey_copy(&b->key, new_key);
1816                 }
1817         } else {
1818                 BUG_ON(btree_node_root(c, b) != b);
1819
1820                 bch2_btree_node_lock_write(b, iter);
1821                 bkey_copy(&b->key, new_key);
1822
1823                 if (btree_ptr_hash_val(&b->key) != b->hash_val) {
1824                         mutex_lock(&c->btree_cache.lock);
1825                         bch2_btree_node_hash_remove(&c->btree_cache, b);
1826
1827                         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1828                         BUG_ON(ret);
1829                         mutex_unlock(&c->btree_cache.lock);
1830                 }
1831
1832                 btree_update_updated_root(as, b);
1833                 bch2_btree_node_unlock_write(b, iter);
1834         }
1835
1836         bch2_btree_update_done(as);
1837 }
1838
1839 int bch2_btree_node_update_key(struct bch_fs *c, struct btree_iter *iter,
1840                                struct btree *b,
1841                                struct bkey_i *new_key)
1842 {
1843         struct btree *parent = btree_node_parent(iter, b);
1844         struct btree_update *as = NULL;
1845         struct btree *new_hash = NULL;
1846         struct closure cl;
1847         int ret;
1848
1849         closure_init_stack(&cl);
1850
1851         if (!bch2_btree_iter_upgrade(iter, U8_MAX))
1852                 return -EINTR;
1853
1854         if (!down_read_trylock(&c->gc_lock)) {
1855                 bch2_trans_unlock(iter->trans);
1856                 down_read(&c->gc_lock);
1857
1858                 if (!bch2_trans_relock(iter->trans)) {
1859                         ret = -EINTR;
1860                         goto err;
1861                 }
1862         }
1863
1864         /*
1865          * check btree_ptr_hash_val() after @b is locked by
1866          * btree_iter_traverse():
1867          */
1868         if (btree_ptr_hash_val(new_key) != b->hash_val) {
1869                 /* bch2_btree_reserve_get will unlock */
1870                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1871                 if (ret) {
1872                         bch2_trans_unlock(iter->trans);
1873                         up_read(&c->gc_lock);
1874                         closure_sync(&cl);
1875                         down_read(&c->gc_lock);
1876
1877                         if (!bch2_trans_relock(iter->trans)) {
1878                                 ret = -EINTR;
1879                                 goto err;
1880                         }
1881                 }
1882
1883                 new_hash = bch2_btree_node_mem_alloc(c);
1884         }
1885 retry:
1886         as = bch2_btree_update_start(iter->trans, iter->btree_id,
1887                 parent ? btree_update_reserve_required(c, parent) : 0,
1888                 BTREE_INSERT_NOFAIL|
1889                 BTREE_INSERT_USE_RESERVE|
1890                 BTREE_INSERT_USE_ALLOC_RESERVE,
1891                 &cl);
1892
1893         if (IS_ERR(as)) {
1894                 ret = PTR_ERR(as);
1895                 if (ret == -EAGAIN)
1896                         ret = -EINTR;
1897
1898                 if (ret == -EINTR) {
1899                         bch2_trans_unlock(iter->trans);
1900                         up_read(&c->gc_lock);
1901                         closure_sync(&cl);
1902                         down_read(&c->gc_lock);
1903
1904                         if (bch2_trans_relock(iter->trans))
1905                                 goto retry;
1906                 }
1907
1908                 goto err;
1909         }
1910
1911         ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(new_key));
1912         if (ret)
1913                 goto err_free_update;
1914
1915         __bch2_btree_node_update_key(c, as, iter, b, new_hash, new_key);
1916
1917         bch2_btree_iter_downgrade(iter);
1918 err:
1919         if (new_hash) {
1920                 mutex_lock(&c->btree_cache.lock);
1921                 list_move(&new_hash->list, &c->btree_cache.freeable);
1922                 mutex_unlock(&c->btree_cache.lock);
1923
1924                 six_unlock_write(&new_hash->c.lock);
1925                 six_unlock_intent(&new_hash->c.lock);
1926         }
1927         up_read(&c->gc_lock);
1928         closure_sync(&cl);
1929         return ret;
1930 err_free_update:
1931         bch2_btree_update_free(as);
1932         goto err;
1933 }
1934
1935 /* Init code: */
1936
1937 /*
1938  * Only for filesystem bringup, when first reading the btree roots or allocating
1939  * btree roots when initializing a new filesystem:
1940  */
1941 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
1942 {
1943         BUG_ON(btree_node_root(c, b));
1944
1945         bch2_btree_set_root_inmem(c, b);
1946 }
1947
1948 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
1949 {
1950         struct closure cl;
1951         struct btree *b;
1952         int ret;
1953
1954         closure_init_stack(&cl);
1955
1956         do {
1957                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1958                 closure_sync(&cl);
1959         } while (ret);
1960
1961         b = bch2_btree_node_mem_alloc(c);
1962         bch2_btree_cache_cannibalize_unlock(c);
1963
1964         set_btree_node_fake(b);
1965         set_btree_node_need_rewrite(b);
1966         b->c.level      = 0;
1967         b->c.btree_id   = id;
1968
1969         bkey_btree_ptr_init(&b->key);
1970         b->key.k.p = POS_MAX;
1971         *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
1972
1973         bch2_bset_init_first(b, &b->data->keys);
1974         bch2_btree_build_aux_trees(b);
1975
1976         b->data->flags = 0;
1977         btree_set_min(b, POS_MIN);
1978         btree_set_max(b, POS_MAX);
1979         b->data->format = bch2_btree_calc_format(b);
1980         btree_node_set_format(b, b->data->format);
1981
1982         ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
1983                                           b->c.level, b->c.btree_id);
1984         BUG_ON(ret);
1985
1986         bch2_btree_set_root_inmem(c, b);
1987
1988         six_unlock_write(&b->c.lock);
1989         six_unlock_intent(&b->c.lock);
1990 }
1991
1992 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
1993 {
1994         struct btree_update *as;
1995
1996         mutex_lock(&c->btree_interior_update_lock);
1997         list_for_each_entry(as, &c->btree_interior_update_list, list)
1998                 pr_buf(out, "%p m %u w %u r %u j %llu\n",
1999                        as,
2000                        as->mode,
2001                        as->nodes_written,
2002                        atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2003                        as->journal.seq);
2004         mutex_unlock(&c->btree_interior_update_lock);
2005 }
2006
2007 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *c)
2008 {
2009         size_t ret = 0;
2010         struct list_head *i;
2011
2012         mutex_lock(&c->btree_interior_update_lock);
2013         list_for_each(i, &c->btree_interior_update_list)
2014                 ret++;
2015         mutex_unlock(&c->btree_interior_update_lock);
2016
2017         return ret;
2018 }
2019
2020 void bch2_journal_entries_to_btree_roots(struct bch_fs *c, struct jset *jset)
2021 {
2022         struct btree_root *r;
2023         struct jset_entry *entry;
2024
2025         mutex_lock(&c->btree_root_lock);
2026
2027         vstruct_for_each(jset, entry)
2028                 if (entry->type == BCH_JSET_ENTRY_btree_root) {
2029                         r = &c->btree_roots[entry->btree_id];
2030                         r->level = entry->level;
2031                         r->alive = true;
2032                         bkey_copy(&r->key, &entry->start[0]);
2033                 }
2034
2035         mutex_unlock(&c->btree_root_lock);
2036 }
2037
2038 struct jset_entry *
2039 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2040                                     struct jset_entry *start,
2041                                     struct jset_entry *end)
2042 {
2043         struct jset_entry *entry;
2044         unsigned long have = 0;
2045         unsigned i;
2046
2047         for (entry = start; entry < end; entry = vstruct_next(entry))
2048                 if (entry->type == BCH_JSET_ENTRY_btree_root)
2049                         __set_bit(entry->btree_id, &have);
2050
2051         mutex_lock(&c->btree_root_lock);
2052
2053         for (i = 0; i < BTREE_ID_NR; i++)
2054                 if (c->btree_roots[i].alive && !test_bit(i, &have)) {
2055                         journal_entry_set(end,
2056                                           BCH_JSET_ENTRY_btree_root,
2057                                           i, c->btree_roots[i].level,
2058                                           &c->btree_roots[i].key,
2059                                           c->btree_roots[i].key.u64s);
2060                         end = vstruct_next(end);
2061                 }
2062
2063         mutex_unlock(&c->btree_root_lock);
2064
2065         return end;
2066 }
2067
2068 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2069 {
2070         if (c->btree_interior_update_worker)
2071                 destroy_workqueue(c->btree_interior_update_worker);
2072         mempool_exit(&c->btree_interior_update_pool);
2073 }
2074
2075 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2076 {
2077         mutex_init(&c->btree_reserve_cache_lock);
2078         INIT_LIST_HEAD(&c->btree_interior_update_list);
2079         INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2080         mutex_init(&c->btree_interior_update_lock);
2081         INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2082
2083         c->btree_interior_update_worker =
2084                 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2085         if (!c->btree_interior_update_worker)
2086                 return -ENOMEM;
2087
2088         return mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2089                                          sizeof(struct btree_update));
2090 }