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                 if (bkey_cmp(next_node, bp.v->min_key)) {
  53                         bch2_dump_btree_node(c, b);
  54                         panic("expected next min_key %llu:%llu got %llu:%llu\n",
  55                               next_node.inode,
  56                               next_node.offset,
  57                               bp.v->min_key.inode,
  58                               bp.v->min_key.offset);
  59                 }
  60
  61                 bch2_btree_node_iter_advance(&iter, b);
  62
  63                 if (bch2_btree_node_iter_end(&iter)) {
  64
  65                         if (bkey_cmp(k.k->p, b->key.k.p)) {
  66                                 bch2_dump_btree_node(c, b);
  67                                 panic("expected end %llu:%llu got %llu:%llu\n",
  68                                       b->key.k.p.inode,
  69                                       b->key.k.p.offset,
  70                                       k.k->p.inode,
  71                                       k.k->p.offset);
  72                         }
  73                         break;
  74                 }
  75
  76                 next_node = bkey_successor(k.k->p);
  77         }
  78 #endif
  79 }
  80
  81 /* Calculate ideal packed bkey format for new btree nodes: */
  82
  83 void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b)
  84 {
  85         struct bkey_packed *k;
  86         struct bset_tree *t;
  87         struct bkey uk;
  88
  89         bch2_bkey_format_add_pos(s, b->data->min_key);
  90
  91         for_each_bset(b, t)
  92                 bset_tree_for_each_key(b, t, k)
  93                         if (!bkey_whiteout(k)) {
  94                                 uk = bkey_unpack_key(b, k);
  95                                 bch2_bkey_format_add_key(s, &uk);
  96                         }
  97 }
  98
  99 static struct bkey_format bch2_btree_calc_format(struct btree *b)
 100 {
 101         struct bkey_format_state s;
 102
 103         bch2_bkey_format_init(&s);
 104         __bch2_btree_calc_format(&s, b);
 105
 106         return bch2_bkey_format_done(&s);
 107 }
 108
 109 static size_t btree_node_u64s_with_format(struct btree *b,
 110                                           struct bkey_format *new_f)
 111 {
 112         struct bkey_format *old_f = &b->format;
 113
 114         /* stupid integer promotion rules */
 115         ssize_t delta =
 116             (((int) new_f->key_u64s - old_f->key_u64s) *
 117              (int) b->nr.packed_keys) +
 118             (((int) new_f->key_u64s - BKEY_U64s) *
 119              (int) b->nr.unpacked_keys);
 120
 121         BUG_ON(delta + b->nr.live_u64s < 0);
 122
 123         return b->nr.live_u64s + delta;
 124 }
 125
 126 /**
 127  * btree_node_format_fits - check if we could rewrite node with a new format
 128  *
 129  * This assumes all keys can pack with the new format -- it just checks if
 130  * the re-packed keys would fit inside the node itself.
 131  */
 132 bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
 133                                  struct bkey_format *new_f)
 134 {
 135         size_t u64s = btree_node_u64s_with_format(b, new_f);
 136
 137         return __vstruct_bytes(struct btree_node, u64s) < btree_bytes(c);
 138 }
 139
 140 /* Btree node freeing/allocation: */
 141
 142 static void __btree_node_free(struct bch_fs *c, struct btree *b)
 143 {
 144         trace_btree_node_free(c, b);
 145
 146         BUG_ON(btree_node_dirty(b));
 147         BUG_ON(btree_node_need_write(b));
 148         BUG_ON(b == btree_node_root(c, b));
 149         BUG_ON(b->ob.nr);
 150         BUG_ON(!list_empty(&b->write_blocked));
 151         BUG_ON(b->will_make_reachable);
 152
 153         clear_btree_node_noevict(b);
 154
 155         bch2_btree_node_hash_remove(&c->btree_cache, b);
 156
 157         mutex_lock(&c->btree_cache.lock);
 158         list_move(&b->list, &c->btree_cache.freeable);
 159         mutex_unlock(&c->btree_cache.lock);
 160 }
 161
 162 void bch2_btree_node_free_never_inserted(struct bch_fs *c, struct btree *b)
 163 {
 164         struct open_buckets ob = b->ob;
 165
 166         b->ob.nr = 0;
 167
 168         clear_btree_node_dirty(c, b);
 169
 170         btree_node_lock_type(c, b, SIX_LOCK_write);
 171         __btree_node_free(c, b);
 172         six_unlock_write(&b->c.lock);
 173
 174         bch2_open_buckets_put(c, &ob);
 175 }
 176
 177 void bch2_btree_node_free_inmem(struct bch_fs *c, struct btree *b,
 178                                 struct btree_iter *iter)
 179 {
 180         struct btree_iter *linked;
 181
 182         trans_for_each_iter(iter->trans, linked)
 183                 BUG_ON(linked->l[b->c.level].b == b);
 184
 185         six_lock_write(&b->c.lock, NULL, NULL);
 186         __btree_node_free(c, b);
 187         six_unlock_write(&b->c.lock);
 188         six_unlock_intent(&b->c.lock);
 189 }
 190
 191 static struct btree *__bch2_btree_node_alloc(struct bch_fs *c,
 192                                              struct disk_reservation *res,
 193                                              struct closure *cl,
 194                                              unsigned flags)
 195 {
 196         struct write_point *wp;
 197         struct btree *b;
 198         __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
 199         struct open_buckets ob = { .nr = 0 };
 200         struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
 201         unsigned nr_reserve;
 202         enum alloc_reserve alloc_reserve;
 203
 204         if (flags & BTREE_INSERT_USE_RESERVE) {
 205                 nr_reserve      = 0;
 206                 alloc_reserve   = RESERVE_BTREE_MOVINGGC;
 207         } else {
 208                 nr_reserve      = BTREE_NODE_RESERVE;
 209                 alloc_reserve   = RESERVE_BTREE;
 210         }
 211
 212         mutex_lock(&c->btree_reserve_cache_lock);
 213         if (c->btree_reserve_cache_nr > nr_reserve) {
 214                 struct btree_alloc *a =
 215                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
 216
 217                 ob = a->ob;
 218                 bkey_copy(&tmp.k, &a->k);
 219                 mutex_unlock(&c->btree_reserve_cache_lock);
 220                 goto mem_alloc;
 221         }
 222         mutex_unlock(&c->btree_reserve_cache_lock);
 223
 224 retry:
 225         wp = bch2_alloc_sectors_start(c, c->opts.foreground_target, 0,
 226                                       writepoint_ptr(&c->btree_write_point),
 227                                       &devs_have,
 228                                       res->nr_replicas,
 229                                       c->opts.metadata_replicas_required,
 230                                       alloc_reserve, 0, cl);
 231         if (IS_ERR(wp))
 232                 return ERR_CAST(wp);
 233
 234         if (wp->sectors_free < c->opts.btree_node_size) {
 235                 struct open_bucket *ob;
 236                 unsigned i;
 237
 238                 open_bucket_for_each(c, &wp->ptrs, ob, i)
 239                         if (ob->sectors_free < c->opts.btree_node_size)
 240                                 ob->sectors_free = 0;
 241
 242                 bch2_alloc_sectors_done(c, wp);
 243                 goto retry;
 244         }
 245
 246         if (c->sb.features & (1ULL << BCH_FEATURE_btree_ptr_v2))
 247                 bkey_btree_ptr_v2_init(&tmp.k);
 248         else
 249                 bkey_btree_ptr_init(&tmp.k);
 250
 251         bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, c->opts.btree_node_size);
 252
 253         bch2_open_bucket_get(c, wp, &ob);
 254         bch2_alloc_sectors_done(c, wp);
 255 mem_alloc:
 256         b = bch2_btree_node_mem_alloc(c);
 257
 258         /* we hold cannibalize_lock: */
 259         BUG_ON(IS_ERR(b));
 260         BUG_ON(b->ob.nr);
 261
 262         bkey_copy(&b->key, &tmp.k);
 263         b->ob = ob;
 264
 265         return b;
 266 }
 267
 268 static struct btree *bch2_btree_node_alloc(struct btree_update *as, unsigned level)
 269 {
 270         struct bch_fs *c = as->c;
 271         struct btree *b;
 272         int ret;
 273
 274         BUG_ON(level >= BTREE_MAX_DEPTH);
 275         BUG_ON(!as->nr_prealloc_nodes);
 276
 277         b = as->prealloc_nodes[--as->nr_prealloc_nodes];
 278
 279         set_btree_node_accessed(b);
 280         set_btree_node_dirty(c, b);
 281         set_btree_node_need_write(b);
 282
 283         bch2_bset_init_first(b, &b->data->keys);
 284         b->c.level      = level;
 285         b->c.btree_id   = as->btree_id;
 286
 287         memset(&b->nr, 0, sizeof(b->nr));
 288         b->data->magic = cpu_to_le64(bset_magic(c));
 289         b->data->flags = 0;
 290         SET_BTREE_NODE_ID(b->data, as->btree_id);
 291         SET_BTREE_NODE_LEVEL(b->data, level);
 292         b->data->ptr = bch2_bkey_ptrs_c(bkey_i_to_s_c(&b->key)).start->ptr;
 293
 294         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 295                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
 296
 297                 bp->v.mem_ptr           = 0;
 298                 bp->v.seq               = b->data->keys.seq;
 299                 bp->v.sectors_written   = 0;
 300                 bp->v.sectors           = cpu_to_le16(c->opts.btree_node_size);
 301         }
 302
 303         if (c->sb.features & (1ULL << BCH_FEATURE_new_extent_overwrite))
 304                 SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
 305
 306         if (btree_node_is_extents(b) &&
 307             !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data)) {
 308                 set_btree_node_old_extent_overwrite(b);
 309                 set_btree_node_need_rewrite(b);
 310         }
 311
 312         bch2_btree_build_aux_trees(b);
 313
 314         ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
 315         BUG_ON(ret);
 316
 317         trace_btree_node_alloc(c, b);
 318         return b;
 319 }
 320
 321 static void btree_set_min(struct btree *b, struct bpos pos)
 322 {
 323         if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
 324                 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
 325         b->data->min_key = pos;
 326 }
 327
 328 static void btree_set_max(struct btree *b, struct bpos pos)
 329 {
 330         b->key.k.p = pos;
 331         b->data->max_key = pos;
 332 }
 333
 334 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *as,
 335                                                   struct btree *b,
 336                                                   struct bkey_format format)
 337 {
 338         struct btree *n;
 339
 340         n = bch2_btree_node_alloc(as, b->c.level);
 341
 342         SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
 343
 344         btree_set_min(n, b->data->min_key);
 345         btree_set_max(n, b->data->max_key);
 346
 347         n->data->format         = format;
 348         btree_node_set_format(n, format);
 349
 350         bch2_btree_sort_into(as->c, n, b);
 351
 352         btree_node_reset_sib_u64s(n);
 353
 354         n->key.k.p = b->key.k.p;
 355         return n;
 356 }
 357
 358 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
 359                                                        struct btree *b)
 360 {
 361         struct bkey_format new_f = bch2_btree_calc_format(b);
 362
 363         /*
 364          * The keys might expand with the new format - if they wouldn't fit in
 365          * the btree node anymore, use the old format for now:
 366          */
 367         if (!bch2_btree_node_format_fits(as->c, b, &new_f))
 368                 new_f = b->format;
 369
 370         return __bch2_btree_node_alloc_replacement(as, b, new_f);
 371 }
 372
 373 static struct btree *__btree_root_alloc(struct btree_update *as, unsigned level)
 374 {
 375         struct btree *b = bch2_btree_node_alloc(as, level);
 376
 377         btree_set_min(b, POS_MIN);
 378         btree_set_max(b, POS_MAX);
 379         b->data->format = bch2_btree_calc_format(b);
 380
 381         btree_node_set_format(b, b->data->format);
 382         bch2_btree_build_aux_trees(b);
 383
 384         bch2_btree_update_add_new_node(as, b);
 385         six_unlock_write(&b->c.lock);
 386
 387         return b;
 388 }
 389
 390 static void bch2_btree_reserve_put(struct btree_update *as)
 391 {
 392         struct bch_fs *c = as->c;
 393
 394         mutex_lock(&c->btree_reserve_cache_lock);
 395
 396         while (as->nr_prealloc_nodes) {
 397                 struct btree *b = as->prealloc_nodes[--as->nr_prealloc_nodes];
 398
 399                 six_unlock_write(&b->c.lock);
 400
 401                 if (c->btree_reserve_cache_nr <
 402                     ARRAY_SIZE(c->btree_reserve_cache)) {
 403                         struct btree_alloc *a =
 404                                 &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
 405
 406                         a->ob = b->ob;
 407                         b->ob.nr = 0;
 408                         bkey_copy(&a->k, &b->key);
 409                 } else {
 410                         bch2_open_buckets_put(c, &b->ob);
 411                 }
 412
 413                 btree_node_lock_type(c, b, SIX_LOCK_write);
 414                 __btree_node_free(c, b);
 415                 six_unlock_write(&b->c.lock);
 416
 417                 six_unlock_intent(&b->c.lock);
 418         }
 419
 420         mutex_unlock(&c->btree_reserve_cache_lock);
 421 }
 422
 423 static int bch2_btree_reserve_get(struct btree_update *as, unsigned nr_nodes,
 424                                   unsigned flags, struct closure *cl)
 425 {
 426         struct bch_fs *c = as->c;
 427         struct btree *b;
 428         int ret;
 429
 430         BUG_ON(nr_nodes > BTREE_RESERVE_MAX);
 431
 432         /*
 433          * Protects reaping from the btree node cache and using the btree node
 434          * open bucket reserve:
 435          */
 436         ret = bch2_btree_cache_cannibalize_lock(c, cl);
 437         if (ret)
 438                 return ret;
 439
 440         while (as->nr_prealloc_nodes < nr_nodes) {
 441                 b = __bch2_btree_node_alloc(c, &as->disk_res,
 442                                             flags & BTREE_INSERT_NOWAIT
 443                                             ? NULL : cl, flags);
 444                 if (IS_ERR(b)) {
 445                         ret = PTR_ERR(b);
 446                         goto err_free;
 447                 }
 448
 449                 ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(&b->key));
 450                 if (ret)
 451                         goto err_free;
 452
 453                 as->prealloc_nodes[as->nr_prealloc_nodes++] = b;
 454         }
 455
 456         bch2_btree_cache_cannibalize_unlock(c);
 457         return 0;
 458 err_free:
 459         bch2_btree_cache_cannibalize_unlock(c);
 460         trace_btree_reserve_get_fail(c, nr_nodes, cl);
 461         return ret;
 462 }
 463
 464 /* Asynchronous interior node update machinery */
 465
 466 static void bch2_btree_update_free(struct btree_update *as)
 467 {
 468         struct bch_fs *c = as->c;
 469
 470         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 471
 472         bch2_journal_pin_drop(&c->journal, &as->journal);
 473         bch2_journal_pin_flush(&c->journal, &as->journal);
 474         bch2_disk_reservation_put(c, &as->disk_res);
 475         bch2_btree_reserve_put(as);
 476
 477         mutex_lock(&c->btree_interior_update_lock);
 478         list_del(&as->unwritten_list);
 479         list_del(&as->list);
 480         mutex_unlock(&c->btree_interior_update_lock);
 481
 482         closure_debug_destroy(&as->cl);
 483         mempool_free(as, &c->btree_interior_update_pool);
 484
 485         closure_wake_up(&c->btree_interior_update_wait);
 486 }
 487
 488 static void btree_update_will_delete_key(struct btree_update *as,
 489                                          struct bkey_i *k)
 490 {
 491         BUG_ON(bch2_keylist_u64s(&as->old_keys) + k->k.u64s >
 492                ARRAY_SIZE(as->_old_keys));
 493         bch2_keylist_add(&as->old_keys, k);
 494 }
 495
 496 static void btree_update_will_add_key(struct btree_update *as,
 497                                       struct bkey_i *k)
 498 {
 499         BUG_ON(bch2_keylist_u64s(&as->new_keys) + k->k.u64s >
 500                ARRAY_SIZE(as->_new_keys));
 501         bch2_keylist_add(&as->new_keys, k);
 502 }
 503
 504 /*
 505  * The transactional part of an interior btree node update, where we journal the
 506  * update we did to the interior node and update alloc info:
 507  */
 508 static int btree_update_nodes_written_trans(struct btree_trans *trans,
 509                                             struct btree_update *as)
 510 {
 511         struct bkey_i *k;
 512         int ret;
 513
 514         trans->extra_journal_entries = (void *) &as->journal_entries[0];
 515         trans->extra_journal_entry_u64s = as->journal_u64s;
 516         trans->journal_pin = &as->journal;
 517
 518         for_each_keylist_key(&as->new_keys, k) {
 519                 ret = bch2_trans_mark_key(trans,
 520                                           bkey_s_c_null,
 521                                           bkey_i_to_s_c(k),
 522                                           0, 0, BTREE_TRIGGER_INSERT);
 523                 if (ret)
 524                         return ret;
 525         }
 526
 527         for_each_keylist_key(&as->old_keys, k) {
 528                 ret = bch2_trans_mark_key(trans,
 529                                           bkey_i_to_s_c(k),
 530                                           bkey_s_c_null,
 531                                           0, 0, BTREE_TRIGGER_OVERWRITE);
 532                 if (ret)
 533                         return ret;
 534         }
 535
 536         return 0;
 537 }
 538
 539 static void btree_update_nodes_written(struct btree_update *as)
 540 {
 541         struct bch_fs *c = as->c;
 542         struct btree *b = as->b;
 543         struct btree_trans trans;
 544         u64 journal_seq = 0;
 545         unsigned i;
 546         int ret;
 547
 548         /*
 549          * If we're already in an error state, it might be because a btree node
 550          * was never written, and we might be trying to free that same btree
 551          * node here, but it won't have been marked as allocated and we'll see
 552          * spurious disk usage inconsistencies in the transactional part below
 553          * if we don't skip it:
 554          */
 555         ret = bch2_journal_error(&c->journal);
 556         if (ret)
 557                 goto err;
 558
 559         BUG_ON(!journal_pin_active(&as->journal));
 560
 561         /*
 562          * We did an update to a parent node where the pointers we added pointed
 563          * to child nodes that weren't written yet: now, the child nodes have
 564          * been written so we can write out the update to the interior node.
 565          */
 566
 567         /*
 568          * We can't call into journal reclaim here: we'd block on the journal
 569          * reclaim lock, but we may need to release the open buckets we have
 570          * pinned in order for other btree updates to make forward progress, and
 571          * journal reclaim does btree updates when flushing bkey_cached entries,
 572          * which may require allocations as well.
 573          */
 574         bch2_trans_init(&trans, c, 0, 512);
 575         ret = __bch2_trans_do(&trans, &as->disk_res, &journal_seq,
 576                               BTREE_INSERT_NOFAIL|
 577                               BTREE_INSERT_NOCHECK_RW|
 578                               BTREE_INSERT_JOURNAL_RECLAIM|
 579                               BTREE_INSERT_JOURNAL_RESERVED,
 580                               btree_update_nodes_written_trans(&trans, as));
 581         bch2_trans_exit(&trans);
 582
 583         bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
 584                              "error %i in btree_update_nodes_written()", ret);
 585 err:
 586         if (b) {
 587                 /*
 588                  * @b is the node we did the final insert into:
 589                  *
 590                  * On failure to get a journal reservation, we still have to
 591                  * unblock the write and allow most of the write path to happen
 592                  * so that shutdown works, but the i->journal_seq mechanism
 593                  * won't work to prevent the btree write from being visible (we
 594                  * didn't get a journal sequence number) - instead
 595                  * __bch2_btree_node_write() doesn't do the actual write if
 596                  * we're in journal error state:
 597                  */
 598
 599                 btree_node_lock_type(c, b, SIX_LOCK_intent);
 600                 btree_node_lock_type(c, b, SIX_LOCK_write);
 601                 mutex_lock(&c->btree_interior_update_lock);
 602
 603                 list_del(&as->write_blocked_list);
 604
 605                 /*
 606                  * Node might have been freed, recheck under
 607                  * btree_interior_update_lock:
 608                  */
 609                 if (as->b == b) {
 610                         struct bset *i = btree_bset_last(b);
 611
 612                         BUG_ON(!b->c.level);
 613                         BUG_ON(!btree_node_dirty(b));
 614
 615                         if (!ret) {
 616                                 i->journal_seq = cpu_to_le64(
 617                                         max(journal_seq,
 618                                             le64_to_cpu(i->journal_seq)));
 619
 620                                 bch2_btree_add_journal_pin(c, b, journal_seq);
 621                         } else {
 622                                 /*
 623                                  * If we didn't get a journal sequence number we
 624                                  * can't write this btree node, because recovery
 625                                  * won't know to ignore this write:
 626                                  */
 627                                 set_btree_node_never_write(b);
 628                         }
 629                 }
 630
 631                 mutex_unlock(&c->btree_interior_update_lock);
 632                 six_unlock_write(&b->c.lock);
 633
 634                 btree_node_write_if_need(c, b, SIX_LOCK_intent);
 635                 six_unlock_intent(&b->c.lock);
 636         }
 637
 638         bch2_journal_pin_drop(&c->journal, &as->journal);
 639
 640         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 641
 642         mutex_lock(&c->btree_interior_update_lock);
 643         for (i = 0; i < as->nr_new_nodes; i++) {
 644                 b = as->new_nodes[i];
 645
 646                 BUG_ON(b->will_make_reachable != (unsigned long) as);
 647                 b->will_make_reachable = 0;
 648         }
 649         mutex_unlock(&c->btree_interior_update_lock);
 650
 651         for (i = 0; i < as->nr_new_nodes; i++) {
 652                 b = as->new_nodes[i];
 653
 654                 btree_node_lock_type(c, b, SIX_LOCK_read);
 655                 btree_node_write_if_need(c, b, SIX_LOCK_read);
 656                 six_unlock_read(&b->c.lock);
 657         }
 658
 659         for (i = 0; i < as->nr_open_buckets; i++)
 660                 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
 661
 662         bch2_btree_update_free(as);
 663 }
 664
 665 static void btree_interior_update_work(struct work_struct *work)
 666 {
 667         struct bch_fs *c =
 668                 container_of(work, struct bch_fs, btree_interior_update_work);
 669         struct btree_update *as;
 670
 671         while (1) {
 672                 mutex_lock(&c->btree_interior_update_lock);
 673                 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
 674                                               struct btree_update, unwritten_list);
 675                 if (as && !as->nodes_written)
 676                         as = NULL;
 677                 mutex_unlock(&c->btree_interior_update_lock);
 678
 679                 if (!as)
 680                         break;
 681
 682                 btree_update_nodes_written(as);
 683         }
 684 }
 685
 686 static void btree_update_set_nodes_written(struct closure *cl)
 687 {
 688         struct btree_update *as = container_of(cl, struct btree_update, cl);
 689         struct bch_fs *c = as->c;
 690
 691         mutex_lock(&c->btree_interior_update_lock);
 692         as->nodes_written = true;
 693         mutex_unlock(&c->btree_interior_update_lock);
 694
 695         queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
 696 }
 697
 698 /*
 699  * We're updating @b with pointers to nodes that haven't finished writing yet:
 700  * block @b from being written until @as completes
 701  */
 702 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
 703 {
 704         struct bch_fs *c = as->c;
 705
 706         mutex_lock(&c->btree_interior_update_lock);
 707         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 708
 709         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 710         BUG_ON(!btree_node_dirty(b));
 711
 712         as->mode        = BTREE_INTERIOR_UPDATING_NODE;
 713         as->b           = b;
 714         list_add(&as->write_blocked_list, &b->write_blocked);
 715
 716         mutex_unlock(&c->btree_interior_update_lock);
 717 }
 718
 719 static void btree_update_reparent(struct btree_update *as,
 720                                   struct btree_update *child)
 721 {
 722         struct bch_fs *c = as->c;
 723
 724         lockdep_assert_held(&c->btree_interior_update_lock);
 725
 726         child->b = NULL;
 727         child->mode = BTREE_INTERIOR_UPDATING_AS;
 728
 729         bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL);
 730 }
 731
 732 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
 733 {
 734         struct bkey_i *insert = &b->key;
 735         struct bch_fs *c = as->c;
 736
 737         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 738
 739         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
 740                ARRAY_SIZE(as->journal_entries));
 741
 742         as->journal_u64s +=
 743                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
 744                                   BCH_JSET_ENTRY_btree_root,
 745                                   b->c.btree_id, b->c.level,
 746                                   insert, insert->k.u64s);
 747
 748         mutex_lock(&c->btree_interior_update_lock);
 749         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 750
 751         as->mode        = BTREE_INTERIOR_UPDATING_ROOT;
 752         mutex_unlock(&c->btree_interior_update_lock);
 753 }
 754
 755 /*
 756  * bch2_btree_update_add_new_node:
 757  *
 758  * This causes @as to wait on @b to be written, before it gets to
 759  * bch2_btree_update_nodes_written
 760  *
 761  * Additionally, it sets b->will_make_reachable to prevent any additional writes
 762  * to @b from happening besides the first until @b is reachable on disk
 763  *
 764  * And it adds @b to the list of @as's new nodes, so that we can update sector
 765  * counts in bch2_btree_update_nodes_written:
 766  */
 767 void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
 768 {
 769         struct bch_fs *c = as->c;
 770
 771         closure_get(&as->cl);
 772
 773         mutex_lock(&c->btree_interior_update_lock);
 774         BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
 775         BUG_ON(b->will_make_reachable);
 776
 777         as->new_nodes[as->nr_new_nodes++] = b;
 778         b->will_make_reachable = 1UL|(unsigned long) as;
 779
 780         mutex_unlock(&c->btree_interior_update_lock);
 781
 782         btree_update_will_add_key(as, &b->key);
 783 }
 784
 785 /*
 786  * returns true if @b was a new node
 787  */
 788 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
 789 {
 790         struct btree_update *as;
 791         unsigned long v;
 792         unsigned i;
 793
 794         mutex_lock(&c->btree_interior_update_lock);
 795         /*
 796          * When b->will_make_reachable != 0, it owns a ref on as->cl that's
 797          * dropped when it gets written by bch2_btree_complete_write - the
 798          * xchg() is for synchronization with bch2_btree_complete_write:
 799          */
 800         v = xchg(&b->will_make_reachable, 0);
 801         as = (struct btree_update *) (v & ~1UL);
 802
 803         if (!as) {
 804                 mutex_unlock(&c->btree_interior_update_lock);
 805                 return;
 806         }
 807
 808         for (i = 0; i < as->nr_new_nodes; i++)
 809                 if (as->new_nodes[i] == b)
 810                         goto found;
 811
 812         BUG();
 813 found:
 814         array_remove_item(as->new_nodes, as->nr_new_nodes, i);
 815         mutex_unlock(&c->btree_interior_update_lock);
 816
 817         if (v & 1)
 818                 closure_put(&as->cl);
 819 }
 820
 821 void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
 822 {
 823         while (b->ob.nr)
 824                 as->open_buckets[as->nr_open_buckets++] =
 825                         b->ob.v[--b->ob.nr];
 826 }
 827
 828 /*
 829  * @b is being split/rewritten: it may have pointers to not-yet-written btree
 830  * nodes and thus outstanding btree_updates - redirect @b's
 831  * btree_updates to point to this btree_update:
 832  */
 833 void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 834                                                struct btree *b)
 835 {
 836         struct bch_fs *c = as->c;
 837         struct btree_update *p, *n;
 838         struct btree_write *w;
 839
 840         set_btree_node_dying(b);
 841
 842         if (btree_node_fake(b))
 843                 return;
 844
 845         mutex_lock(&c->btree_interior_update_lock);
 846
 847         /*
 848          * Does this node have any btree_update operations preventing
 849          * it from being written?
 850          *
 851          * If so, redirect them to point to this btree_update: we can
 852          * write out our new nodes, but we won't make them visible until those
 853          * operations complete
 854          */
 855         list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
 856                 list_del_init(&p->write_blocked_list);
 857                 btree_update_reparent(as, p);
 858
 859                 /*
 860                  * for flush_held_btree_writes() waiting on updates to flush or
 861                  * nodes to be writeable:
 862                  */
 863                 closure_wake_up(&c->btree_interior_update_wait);
 864         }
 865
 866         clear_btree_node_dirty(c, b);
 867         clear_btree_node_need_write(b);
 868
 869         /*
 870          * Does this node have unwritten data that has a pin on the journal?
 871          *
 872          * If so, transfer that pin to the btree_update operation -
 873          * note that if we're freeing multiple nodes, we only need to keep the
 874          * oldest pin of any of the nodes we're freeing. We'll release the pin
 875          * when the new nodes are persistent and reachable on disk:
 876          */
 877         w = btree_current_write(b);
 878         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 879         bch2_journal_pin_drop(&c->journal, &w->journal);
 880
 881         w = btree_prev_write(b);
 882         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 883         bch2_journal_pin_drop(&c->journal, &w->journal);
 884
 885         mutex_unlock(&c->btree_interior_update_lock);
 886
 887         /*
 888          * Is this a node that isn't reachable on disk yet?
 889          *
 890          * Nodes that aren't reachable yet have writes blocked until they're
 891          * reachable - now that we've cancelled any pending writes and moved
 892          * things waiting on that write to wait on this update, we can drop this
 893          * node from the list of nodes that the other update is making
 894          * reachable, prior to freeing it:
 895          */
 896         btree_update_drop_new_node(c, b);
 897
 898         btree_update_will_delete_key(as, &b->key);
 899 }
 900
 901 void bch2_btree_update_done(struct btree_update *as)
 902 {
 903         BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
 904
 905         bch2_btree_reserve_put(as);
 906
 907         continue_at(&as->cl, btree_update_set_nodes_written, system_freezable_wq);
 908 }
 909
 910 struct btree_update *
 911 bch2_btree_update_start(struct btree_trans *trans, enum btree_id id,
 912                         unsigned nr_nodes, unsigned flags,
 913                         struct closure *cl)
 914 {
 915         struct bch_fs *c = trans->c;
 916         struct btree_update *as;
 917         int disk_res_flags = (flags & BTREE_INSERT_NOFAIL)
 918                 ? BCH_DISK_RESERVATION_NOFAIL : 0;
 919         int journal_flags = (flags & BTREE_INSERT_JOURNAL_RESERVED)
 920                 ? JOURNAL_RES_GET_RECLAIM : 0;
 921         int ret = 0;
 922
 923         /*
 924          * This check isn't necessary for correctness - it's just to potentially
 925          * prevent us from doing a lot of work that'll end up being wasted:
 926          */
 927         ret = bch2_journal_error(&c->journal);
 928         if (ret)
 929                 return ERR_PTR(ret);
 930
 931         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
 932         memset(as, 0, sizeof(*as));
 933         closure_init(&as->cl, NULL);
 934         as->c           = c;
 935         as->mode        = BTREE_INTERIOR_NO_UPDATE;
 936         as->btree_id    = id;
 937         INIT_LIST_HEAD(&as->list);
 938         INIT_LIST_HEAD(&as->unwritten_list);
 939         INIT_LIST_HEAD(&as->write_blocked_list);
 940         bch2_keylist_init(&as->old_keys, as->_old_keys);
 941         bch2_keylist_init(&as->new_keys, as->_new_keys);
 942         bch2_keylist_init(&as->parent_keys, as->inline_keys);
 943
 944         ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
 945                                       BTREE_UPDATE_JOURNAL_RES,
 946                                       journal_flags|JOURNAL_RES_GET_NONBLOCK);
 947         if (ret == -EAGAIN) {
 948                 if (flags & BTREE_INSERT_NOUNLOCK)
 949                         return ERR_PTR(-EINTR);
 950
 951                 bch2_trans_unlock(trans);
 952
 953                 ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
 954                                 BTREE_UPDATE_JOURNAL_RES,
 955                                 journal_flags);
 956                 if (ret)
 957                         return ERR_PTR(ret);
 958
 959                 if (!bch2_trans_relock(trans)) {
 960                         ret = -EINTR;
 961                         goto err;
 962                 }
 963         }
 964
 965         ret = bch2_disk_reservation_get(c, &as->disk_res,
 966                         nr_nodes * c->opts.btree_node_size,
 967                         c->opts.metadata_replicas,
 968                         disk_res_flags);
 969         if (ret)
 970                 goto err;
 971
 972         ret = bch2_btree_reserve_get(as, nr_nodes, flags, cl);
 973         if (ret)
 974                 goto err;
 975
 976         bch2_journal_pin_add(&c->journal,
 977                              atomic64_read(&c->journal.seq),
 978                              &as->journal, NULL);
 979
 980         mutex_lock(&c->btree_interior_update_lock);
 981         list_add_tail(&as->list, &c->btree_interior_update_list);
 982         mutex_unlock(&c->btree_interior_update_lock);
 983
 984         return as;
 985 err:
 986         bch2_btree_update_free(as);
 987         return ERR_PTR(ret);
 988 }
 989
 990 /* Btree root updates: */
 991
 992 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
 993 {
 994         /* Root nodes cannot be reaped */
 995         mutex_lock(&c->btree_cache.lock);
 996         list_del_init(&b->list);
 997         mutex_unlock(&c->btree_cache.lock);
 998
 999         mutex_lock(&c->btree_root_lock);
1000         BUG_ON(btree_node_root(c, b) &&
1001                (b->c.level < btree_node_root(c, b)->c.level ||
1002                 !btree_node_dying(btree_node_root(c, b))));
1003
1004         btree_node_root(c, b) = b;
1005         mutex_unlock(&c->btree_root_lock);
1006
1007         bch2_recalc_btree_reserve(c);
1008 }
1009
1010 /**
1011  * bch_btree_set_root - update the root in memory and on disk
1012  *
1013  * To ensure forward progress, the current task must not be holding any
1014  * btree node write locks. However, you must hold an intent lock on the
1015  * old root.
1016  *
1017  * Note: This allocates a journal entry but doesn't add any keys to
1018  * it.  All the btree roots are part of every journal write, so there
1019  * is nothing new to be done.  This just guarantees that there is a
1020  * journal write.
1021  */
1022 static void bch2_btree_set_root(struct btree_update *as, struct btree *b,
1023                                 struct btree_iter *iter)
1024 {
1025         struct bch_fs *c = as->c;
1026         struct btree *old;
1027
1028         trace_btree_set_root(c, b);
1029         BUG_ON(!b->written &&
1030                !test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags));
1031
1032         old = btree_node_root(c, b);
1033
1034         /*
1035          * Ensure no one is using the old root while we switch to the
1036          * new root:
1037          */
1038         bch2_btree_node_lock_write(old, iter);
1039
1040         bch2_btree_set_root_inmem(c, b);
1041
1042         btree_update_updated_root(as, b);
1043
1044         /*
1045          * Unlock old root after new root is visible:
1046          *
1047          * The new root isn't persistent, but that's ok: we still have
1048          * an intent lock on the new root, and any updates that would
1049          * depend on the new root would have to update the new root.
1050          */
1051         bch2_btree_node_unlock_write(old, iter);
1052 }
1053
1054 /* Interior node updates: */
1055
1056 static void bch2_insert_fixup_btree_ptr(struct btree_update *as, struct btree *b,
1057                                         struct btree_iter *iter,
1058                                         struct bkey_i *insert,
1059                                         struct btree_node_iter *node_iter)
1060 {
1061         struct bch_fs *c = as->c;
1062         struct bkey_packed *k;
1063         const char *invalid;
1064
1065         invalid = bch2_bkey_invalid(c, bkey_i_to_s_c(insert), btree_node_type(b)) ?:
1066                 bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert));
1067         if (invalid) {
1068                 char buf[160];
1069
1070                 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(insert));
1071                 bch2_fs_inconsistent(c, "inserting invalid bkey %s: %s", buf, invalid);
1072                 dump_stack();
1073         }
1074
1075         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1076                ARRAY_SIZE(as->journal_entries));
1077
1078         as->journal_u64s +=
1079                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1080                                   BCH_JSET_ENTRY_btree_keys,
1081                                   b->c.btree_id, b->c.level,
1082                                   insert, insert->k.u64s);
1083
1084         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1085                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1086                 bch2_btree_node_iter_advance(node_iter, b);
1087
1088         bch2_btree_bset_insert_key(iter, b, node_iter, insert);
1089         set_btree_node_dirty(c, b);
1090         set_btree_node_need_write(b);
1091 }
1092
1093 /*
1094  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1095  * node)
1096  */
1097 static struct btree *__btree_split_node(struct btree_update *as,
1098                                         struct btree *n1,
1099                                         struct btree_iter *iter)
1100 {
1101         size_t nr_packed = 0, nr_unpacked = 0;
1102         struct btree *n2;
1103         struct bset *set1, *set2;
1104         struct bkey_packed *k, *prev = NULL;
1105
1106         n2 = bch2_btree_node_alloc(as, n1->c.level);
1107         bch2_btree_update_add_new_node(as, n2);
1108
1109         n2->data->max_key       = n1->data->max_key;
1110         n2->data->format        = n1->format;
1111         SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data));
1112         n2->key.k.p = n1->key.k.p;
1113
1114         btree_node_set_format(n2, n2->data->format);
1115
1116         set1 = btree_bset_first(n1);
1117         set2 = btree_bset_first(n2);
1118
1119         /*
1120          * Has to be a linear search because we don't have an auxiliary
1121          * search tree yet
1122          */
1123         k = set1->start;
1124         while (1) {
1125                 struct bkey_packed *n = bkey_next_skip_noops(k, vstruct_last(set1));
1126
1127                 if (n == vstruct_last(set1))
1128                         break;
1129                 if (k->_data - set1->_data >= (le16_to_cpu(set1->u64s) * 3) / 5)
1130                         break;
1131
1132                 if (bkey_packed(k))
1133                         nr_packed++;
1134                 else
1135                         nr_unpacked++;
1136
1137                 prev = k;
1138                 k = n;
1139         }
1140
1141         BUG_ON(!prev);
1142
1143         btree_set_max(n1, bkey_unpack_pos(n1, prev));
1144         btree_set_min(n2, bkey_successor(n1->key.k.p));
1145
1146         set2->u64s = cpu_to_le16((u64 *) vstruct_end(set1) - (u64 *) k);
1147         set1->u64s = cpu_to_le16(le16_to_cpu(set1->u64s) - le16_to_cpu(set2->u64s));
1148
1149         set_btree_bset_end(n1, n1->set);
1150         set_btree_bset_end(n2, n2->set);
1151
1152         n2->nr.live_u64s        = le16_to_cpu(set2->u64s);
1153         n2->nr.bset_u64s[0]     = le16_to_cpu(set2->u64s);
1154         n2->nr.packed_keys      = n1->nr.packed_keys - nr_packed;
1155         n2->nr.unpacked_keys    = n1->nr.unpacked_keys - nr_unpacked;
1156
1157         n1->nr.live_u64s        = le16_to_cpu(set1->u64s);
1158         n1->nr.bset_u64s[0]     = le16_to_cpu(set1->u64s);
1159         n1->nr.packed_keys      = nr_packed;
1160         n1->nr.unpacked_keys    = nr_unpacked;
1161
1162         BUG_ON(!set1->u64s);
1163         BUG_ON(!set2->u64s);
1164
1165         memcpy_u64s(set2->start,
1166                     vstruct_end(set1),
1167                     le16_to_cpu(set2->u64s));
1168
1169         btree_node_reset_sib_u64s(n1);
1170         btree_node_reset_sib_u64s(n2);
1171
1172         bch2_verify_btree_nr_keys(n1);
1173         bch2_verify_btree_nr_keys(n2);
1174
1175         if (n1->c.level) {
1176                 btree_node_interior_verify(as->c, n1);
1177                 btree_node_interior_verify(as->c, n2);
1178         }
1179
1180         return n2;
1181 }
1182
1183 /*
1184  * For updates to interior nodes, we've got to do the insert before we split
1185  * because the stuff we're inserting has to be inserted atomically. Post split,
1186  * the keys might have to go in different nodes and the split would no longer be
1187  * atomic.
1188  *
1189  * Worse, if the insert is from btree node coalescing, if we do the insert after
1190  * we do the split (and pick the pivot) - the pivot we pick might be between
1191  * nodes that were coalesced, and thus in the middle of a child node post
1192  * coalescing:
1193  */
1194 static void btree_split_insert_keys(struct btree_update *as, struct btree *b,
1195                                     struct btree_iter *iter,
1196                                     struct keylist *keys)
1197 {
1198         struct btree_node_iter node_iter;
1199         struct bkey_i *k = bch2_keylist_front(keys);
1200         struct bkey_packed *src, *dst, *n;
1201         struct bset *i;
1202
1203         BUG_ON(btree_node_type(b) != BKEY_TYPE_BTREE);
1204
1205         bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
1206
1207         while (!bch2_keylist_empty(keys)) {
1208                 k = bch2_keylist_front(keys);
1209
1210                 bch2_insert_fixup_btree_ptr(as, b, iter, k, &node_iter);
1211                 bch2_keylist_pop_front(keys);
1212         }
1213
1214         /*
1215          * We can't tolerate whiteouts here - with whiteouts there can be
1216          * duplicate keys, and it would be rather bad if we picked a duplicate
1217          * for the pivot:
1218          */
1219         i = btree_bset_first(b);
1220         src = dst = i->start;
1221         while (src != vstruct_last(i)) {
1222                 n = bkey_next_skip_noops(src, vstruct_last(i));
1223                 if (!bkey_deleted(src)) {
1224                         memmove_u64s_down(dst, src, src->u64s);
1225                         dst = bkey_next(dst);
1226                 }
1227                 src = n;
1228         }
1229
1230         /* Also clear out the unwritten whiteouts area: */
1231         b->whiteout_u64s = 0;
1232
1233         i->u64s = cpu_to_le16((u64 *) dst - i->_data);
1234         set_btree_bset_end(b, b->set);
1235
1236         BUG_ON(b->nsets != 1 ||
1237                b->nr.live_u64s != le16_to_cpu(btree_bset_first(b)->u64s));
1238
1239         btree_node_interior_verify(as->c, b);
1240 }
1241
1242 static void btree_split(struct btree_update *as, struct btree *b,
1243                         struct btree_iter *iter, struct keylist *keys,
1244                         unsigned flags)
1245 {
1246         struct bch_fs *c = as->c;
1247         struct btree *parent = btree_node_parent(iter, b);
1248         struct btree *n1, *n2 = NULL, *n3 = NULL;
1249         u64 start_time = local_clock();
1250
1251         BUG_ON(!parent && (b != btree_node_root(c, b)));
1252         BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->c.level));
1253
1254         bch2_btree_interior_update_will_free_node(as, b);
1255
1256         n1 = bch2_btree_node_alloc_replacement(as, b);
1257         bch2_btree_update_add_new_node(as, n1);
1258
1259         if (keys)
1260                 btree_split_insert_keys(as, n1, iter, keys);
1261
1262         if (bset_u64s(&n1->set[0]) > BTREE_SPLIT_THRESHOLD(c)) {
1263                 trace_btree_split(c, b);
1264
1265                 n2 = __btree_split_node(as, n1, iter);
1266
1267                 bch2_btree_build_aux_trees(n2);
1268                 bch2_btree_build_aux_trees(n1);
1269                 six_unlock_write(&n2->c.lock);
1270                 six_unlock_write(&n1->c.lock);
1271
1272                 bch2_btree_node_write(c, n2, SIX_LOCK_intent);
1273
1274                 /*
1275                  * Note that on recursive parent_keys == keys, so we
1276                  * can't start adding new keys to parent_keys before emptying it
1277                  * out (which we did with btree_split_insert_keys() above)
1278                  */
1279                 bch2_keylist_add(&as->parent_keys, &n1->key);
1280                 bch2_keylist_add(&as->parent_keys, &n2->key);
1281
1282                 if (!parent) {
1283                         /* Depth increases, make a new root */
1284                         n3 = __btree_root_alloc(as, b->c.level + 1);
1285
1286                         n3->sib_u64s[0] = U16_MAX;
1287                         n3->sib_u64s[1] = U16_MAX;
1288
1289                         btree_split_insert_keys(as, n3, iter, &as->parent_keys);
1290
1291                         bch2_btree_node_write(c, n3, SIX_LOCK_intent);
1292                 }
1293         } else {
1294                 trace_btree_compact(c, b);
1295
1296                 bch2_btree_build_aux_trees(n1);
1297                 six_unlock_write(&n1->c.lock);
1298
1299                 if (parent)
1300                         bch2_keylist_add(&as->parent_keys, &n1->key);
1301         }
1302
1303         bch2_btree_node_write(c, n1, SIX_LOCK_intent);
1304
1305         /* New nodes all written, now make them visible: */
1306
1307         if (parent) {
1308                 /* Split a non root node */
1309                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1310         } else if (n3) {
1311                 bch2_btree_set_root(as, n3, iter);
1312         } else {
1313                 /* Root filled up but didn't need to be split */
1314                 bch2_btree_set_root(as, n1, iter);
1315         }
1316
1317         bch2_btree_update_get_open_buckets(as, n1);
1318         if (n2)
1319                 bch2_btree_update_get_open_buckets(as, n2);
1320         if (n3)
1321                 bch2_btree_update_get_open_buckets(as, n3);
1322
1323         /* Successful split, update the iterator to point to the new nodes: */
1324
1325         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1326         bch2_btree_iter_node_drop(iter, b);
1327         if (n3)
1328                 bch2_btree_iter_node_replace(iter, n3);
1329         if (n2)
1330                 bch2_btree_iter_node_replace(iter, n2);
1331         bch2_btree_iter_node_replace(iter, n1);
1332
1333         /*
1334          * The old node must be freed (in memory) _before_ unlocking the new
1335          * nodes - else another thread could re-acquire a read lock on the old
1336          * node after another thread has locked and updated the new node, thus
1337          * seeing stale data:
1338          */
1339         bch2_btree_node_free_inmem(c, b, iter);
1340
1341         if (n3)
1342                 six_unlock_intent(&n3->c.lock);
1343         if (n2)
1344                 six_unlock_intent(&n2->c.lock);
1345         six_unlock_intent(&n1->c.lock);
1346
1347         bch2_btree_trans_verify_locks(iter->trans);
1348
1349         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_split],
1350                                start_time);
1351 }
1352
1353 static void
1354 bch2_btree_insert_keys_interior(struct btree_update *as, struct btree *b,
1355                                 struct btree_iter *iter, struct keylist *keys)
1356 {
1357         struct btree_iter *linked;
1358         struct btree_node_iter node_iter;
1359         struct bkey_i *insert = bch2_keylist_front(keys);
1360         struct bkey_packed *k;
1361
1362         /* Don't screw up @iter's position: */
1363         node_iter = iter->l[b->c.level].iter;
1364
1365         /*
1366          * btree_split(), btree_gc_coalesce() will insert keys before
1367          * the iterator's current position - they know the keys go in
1368          * the node the iterator points to:
1369          */
1370         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1371                (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1372                 ;
1373
1374         for_each_keylist_key(keys, insert)
1375                 bch2_insert_fixup_btree_ptr(as, b, iter, insert, &node_iter);
1376
1377         btree_update_updated_node(as, b);
1378
1379         trans_for_each_iter_with_node(iter->trans, b, linked)
1380                 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1381
1382         bch2_btree_trans_verify_iters(iter->trans, b);
1383 }
1384
1385 /**
1386  * bch_btree_insert_node - insert bkeys into a given btree node
1387  *
1388  * @iter:               btree iterator
1389  * @keys:               list of keys to insert
1390  * @hook:               insert callback
1391  * @persistent:         if not null, @persistent will wait on journal write
1392  *
1393  * Inserts as many keys as it can into a given btree node, splitting it if full.
1394  * If a split occurred, this function will return early. This can only happen
1395  * for leaf nodes -- inserts into interior nodes have to be atomic.
1396  */
1397 void bch2_btree_insert_node(struct btree_update *as, struct btree *b,
1398                             struct btree_iter *iter, struct keylist *keys,
1399                             unsigned flags)
1400 {
1401         struct bch_fs *c = as->c;
1402         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1403         int old_live_u64s = b->nr.live_u64s;
1404         int live_u64s_added, u64s_added;
1405
1406         BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->c.level));
1407         BUG_ON(!b->c.level);
1408         BUG_ON(!as || as->b);
1409         bch2_verify_keylist_sorted(keys);
1410
1411         bch2_btree_node_lock_for_insert(c, b, iter);
1412
1413         if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1414                 bch2_btree_node_unlock_write(b, iter);
1415                 goto split;
1416         }
1417
1418         btree_node_interior_verify(c, b);
1419
1420         bch2_btree_insert_keys_interior(as, b, iter, keys);
1421
1422         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1423         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1424
1425         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1426                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1427         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1428                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1429
1430         if (u64s_added > live_u64s_added &&
1431             bch2_maybe_compact_whiteouts(c, b))
1432                 bch2_btree_iter_reinit_node(iter, b);
1433
1434         bch2_btree_node_unlock_write(b, iter);
1435
1436         btree_node_interior_verify(c, b);
1437
1438         /*
1439          * when called from the btree_split path the new nodes aren't added to
1440          * the btree iterator yet, so the merge path's unlock/wait/relock dance
1441          * won't work:
1442          */
1443         bch2_foreground_maybe_merge(c, iter, b->c.level,
1444                                     flags|BTREE_INSERT_NOUNLOCK);
1445         return;
1446 split:
1447         btree_split(as, b, iter, keys, flags);
1448 }
1449
1450 int bch2_btree_split_leaf(struct bch_fs *c, struct btree_iter *iter,
1451                           unsigned flags)
1452 {
1453         struct btree_trans *trans = iter->trans;
1454         struct btree *b = iter_l(iter)->b;
1455         struct btree_update *as;
1456         struct closure cl;
1457         int ret = 0;
1458
1459         closure_init_stack(&cl);
1460
1461         /* Hack, because gc and splitting nodes doesn't mix yet: */
1462         if (!(flags & BTREE_INSERT_GC_LOCK_HELD) &&
1463             !down_read_trylock(&c->gc_lock)) {
1464                 if (flags & BTREE_INSERT_NOUNLOCK) {
1465                         trace_transaction_restart_ip(trans->ip, _THIS_IP_);
1466                         return -EINTR;
1467                 }
1468
1469                 bch2_trans_unlock(trans);
1470                 down_read(&c->gc_lock);
1471
1472                 if (!bch2_trans_relock(trans))
1473                         ret = -EINTR;
1474         }
1475
1476         /*
1477          * XXX: figure out how far we might need to split,
1478          * instead of locking/reserving all the way to the root:
1479          */
1480         if (!bch2_btree_iter_upgrade(iter, U8_MAX)) {
1481                 trace_trans_restart_iter_upgrade(trans->ip);
1482                 ret = -EINTR;
1483                 goto out;
1484         }
1485
1486         as = bch2_btree_update_start(trans, iter->btree_id,
1487                 btree_update_reserve_required(c, b), flags,
1488                 !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1489         if (IS_ERR(as)) {
1490                 ret = PTR_ERR(as);
1491                 if (ret == -EAGAIN) {
1492                         BUG_ON(flags & BTREE_INSERT_NOUNLOCK);
1493                         bch2_trans_unlock(trans);
1494                         ret = -EINTR;
1495
1496                         trace_transaction_restart_ip(trans->ip, _THIS_IP_);
1497                 }
1498                 goto out;
1499         }
1500
1501         btree_split(as, b, iter, NULL, flags);
1502         bch2_btree_update_done(as);
1503
1504         /*
1505          * We haven't successfully inserted yet, so don't downgrade all the way
1506          * back to read locks;
1507          */
1508         __bch2_btree_iter_downgrade(iter, 1);
1509 out:
1510         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1511                 up_read(&c->gc_lock);
1512         closure_sync(&cl);
1513         return ret;
1514 }
1515
1516 void __bch2_foreground_maybe_merge(struct bch_fs *c,
1517                                    struct btree_iter *iter,
1518                                    unsigned level,
1519                                    unsigned flags,
1520                                    enum btree_node_sibling sib)
1521 {
1522         struct btree_trans *trans = iter->trans;
1523         struct btree_update *as;
1524         struct bkey_format_state new_s;
1525         struct bkey_format new_f;
1526         struct bkey_i delete;
1527         struct btree *b, *m, *n, *prev, *next, *parent;
1528         struct closure cl;
1529         size_t sib_u64s;
1530         int ret = 0;
1531
1532         BUG_ON(!btree_node_locked(iter, level));
1533
1534         closure_init_stack(&cl);
1535 retry:
1536         BUG_ON(!btree_node_locked(iter, level));
1537
1538         b = iter->l[level].b;
1539
1540         parent = btree_node_parent(iter, b);
1541         if (!parent)
1542                 goto out;
1543
1544         if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c))
1545                 goto out;
1546
1547         /* XXX: can't be holding read locks */
1548         m = bch2_btree_node_get_sibling(c, iter, b, sib);
1549         if (IS_ERR(m)) {
1550                 ret = PTR_ERR(m);
1551                 goto err;
1552         }
1553
1554         /* NULL means no sibling: */
1555         if (!m) {
1556                 b->sib_u64s[sib] = U16_MAX;
1557                 goto out;
1558         }
1559
1560         if (sib == btree_prev_sib) {
1561                 prev = m;
1562                 next = b;
1563         } else {
1564                 prev = b;
1565                 next = m;
1566         }
1567
1568         bch2_bkey_format_init(&new_s);
1569         __bch2_btree_calc_format(&new_s, b);
1570         __bch2_btree_calc_format(&new_s, m);
1571         new_f = bch2_bkey_format_done(&new_s);
1572
1573         sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1574                 btree_node_u64s_with_format(m, &new_f);
1575
1576         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1577                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1578                 sib_u64s /= 2;
1579                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1580         }
1581
1582         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1583         b->sib_u64s[sib] = sib_u64s;
1584
1585         if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c)) {
1586                 six_unlock_intent(&m->c.lock);
1587                 goto out;
1588         }
1589
1590         /* We're changing btree topology, doesn't mix with gc: */
1591         if (!(flags & BTREE_INSERT_GC_LOCK_HELD) &&
1592             !down_read_trylock(&c->gc_lock))
1593                 goto err_cycle_gc_lock;
1594
1595         if (!bch2_btree_iter_upgrade(iter, U8_MAX)) {
1596                 ret = -EINTR;
1597                 goto err_unlock;
1598         }
1599
1600         as = bch2_btree_update_start(trans, iter->btree_id,
1601                          btree_update_reserve_required(c, parent) + 1,
1602                          flags|
1603                          BTREE_INSERT_NOFAIL|
1604                          BTREE_INSERT_USE_RESERVE,
1605                          !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1606         if (IS_ERR(as)) {
1607                 ret = PTR_ERR(as);
1608                 goto err_unlock;
1609         }
1610
1611         trace_btree_merge(c, b);
1612
1613         bch2_btree_interior_update_will_free_node(as, b);
1614         bch2_btree_interior_update_will_free_node(as, m);
1615
1616         n = bch2_btree_node_alloc(as, b->c.level);
1617         bch2_btree_update_add_new_node(as, n);
1618
1619         btree_set_min(n, prev->data->min_key);
1620         btree_set_max(n, next->data->max_key);
1621         n->data->format         = new_f;
1622
1623         btree_node_set_format(n, new_f);
1624
1625         bch2_btree_sort_into(c, n, prev);
1626         bch2_btree_sort_into(c, n, next);
1627
1628         bch2_btree_build_aux_trees(n);
1629         six_unlock_write(&n->c.lock);
1630
1631         bkey_init(&delete.k);
1632         delete.k.p = prev->key.k.p;
1633         bch2_keylist_add(&as->parent_keys, &delete);
1634         bch2_keylist_add(&as->parent_keys, &n->key);
1635
1636         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1637
1638         bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1639
1640         bch2_btree_update_get_open_buckets(as, n);
1641
1642         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1643         bch2_btree_iter_node_drop(iter, b);
1644         bch2_btree_iter_node_drop(iter, m);
1645
1646         bch2_btree_iter_node_replace(iter, n);
1647
1648         bch2_btree_trans_verify_iters(trans, n);
1649
1650         bch2_btree_node_free_inmem(c, b, iter);
1651         bch2_btree_node_free_inmem(c, m, iter);
1652
1653         six_unlock_intent(&n->c.lock);
1654
1655         bch2_btree_update_done(as);
1656
1657         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1658                 up_read(&c->gc_lock);
1659 out:
1660         bch2_btree_trans_verify_locks(trans);
1661
1662         /*
1663          * Don't downgrade locks here: we're called after successful insert,
1664          * and the caller will downgrade locks after a successful insert
1665          * anyways (in case e.g. a split was required first)
1666          *
1667          * And we're also called when inserting into interior nodes in the
1668          * split path, and downgrading to read locks in there is potentially
1669          * confusing:
1670          */
1671         closure_sync(&cl);
1672         return;
1673
1674 err_cycle_gc_lock:
1675         six_unlock_intent(&m->c.lock);
1676
1677         if (flags & BTREE_INSERT_NOUNLOCK)
1678                 goto out;
1679
1680         bch2_trans_unlock(trans);
1681
1682         down_read(&c->gc_lock);
1683         up_read(&c->gc_lock);
1684         ret = -EINTR;
1685         goto err;
1686
1687 err_unlock:
1688         six_unlock_intent(&m->c.lock);
1689         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1690                 up_read(&c->gc_lock);
1691 err:
1692         BUG_ON(ret == -EAGAIN && (flags & BTREE_INSERT_NOUNLOCK));
1693
1694         if ((ret == -EAGAIN || ret == -EINTR) &&
1695             !(flags & BTREE_INSERT_NOUNLOCK)) {
1696                 bch2_trans_unlock(trans);
1697                 closure_sync(&cl);
1698                 ret = bch2_btree_iter_traverse(iter);
1699                 if (ret)
1700                         goto out;
1701
1702                 goto retry;
1703         }
1704
1705         goto out;
1706 }
1707
1708 static int __btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
1709                                 struct btree *b, unsigned flags,
1710                                 struct closure *cl)
1711 {
1712         struct btree *n, *parent = btree_node_parent(iter, b);
1713         struct btree_update *as;
1714
1715         as = bch2_btree_update_start(iter->trans, iter->btree_id,
1716                 (parent
1717                  ? btree_update_reserve_required(c, parent)
1718                  : 0) + 1,
1719                 flags, cl);
1720         if (IS_ERR(as)) {
1721                 trace_btree_gc_rewrite_node_fail(c, b);
1722                 return PTR_ERR(as);
1723         }
1724
1725         bch2_btree_interior_update_will_free_node(as, b);
1726
1727         n = bch2_btree_node_alloc_replacement(as, b);
1728         bch2_btree_update_add_new_node(as, n);
1729
1730         bch2_btree_build_aux_trees(n);
1731         six_unlock_write(&n->c.lock);
1732
1733         trace_btree_gc_rewrite_node(c, b);
1734
1735         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1736
1737         if (parent) {
1738                 bch2_keylist_add(&as->parent_keys, &n->key);
1739                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1740         } else {
1741                 bch2_btree_set_root(as, n, iter);
1742         }
1743
1744         bch2_btree_update_get_open_buckets(as, n);
1745
1746         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1747         bch2_btree_iter_node_drop(iter, b);
1748         bch2_btree_iter_node_replace(iter, n);
1749         bch2_btree_node_free_inmem(c, b, iter);
1750         six_unlock_intent(&n->c.lock);
1751
1752         bch2_btree_update_done(as);
1753         return 0;
1754 }
1755
1756 /**
1757  * bch_btree_node_rewrite - Rewrite/move a btree node
1758  *
1759  * Returns 0 on success, -EINTR or -EAGAIN on failure (i.e.
1760  * btree_check_reserve() has to wait)
1761  */
1762 int bch2_btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
1763                             __le64 seq, unsigned flags)
1764 {
1765         struct btree_trans *trans = iter->trans;
1766         struct closure cl;
1767         struct btree *b;
1768         int ret;
1769
1770         flags |= BTREE_INSERT_NOFAIL;
1771
1772         closure_init_stack(&cl);
1773
1774         bch2_btree_iter_upgrade(iter, U8_MAX);
1775
1776         if (!(flags & BTREE_INSERT_GC_LOCK_HELD)) {
1777                 if (!down_read_trylock(&c->gc_lock)) {
1778                         bch2_trans_unlock(trans);
1779                         down_read(&c->gc_lock);
1780                 }
1781         }
1782
1783         while (1) {
1784                 ret = bch2_btree_iter_traverse(iter);
1785                 if (ret)
1786                         break;
1787
1788                 b = bch2_btree_iter_peek_node(iter);
1789                 if (!b || b->data->keys.seq != seq)
1790                         break;
1791
1792                 ret = __btree_node_rewrite(c, iter, b, flags, &cl);
1793                 if (ret != -EAGAIN &&
1794                     ret != -EINTR)
1795                         break;
1796
1797                 bch2_trans_unlock(trans);
1798                 closure_sync(&cl);
1799         }
1800
1801         bch2_btree_iter_downgrade(iter);
1802
1803         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1804                 up_read(&c->gc_lock);
1805
1806         closure_sync(&cl);
1807         return ret;
1808 }
1809
1810 static void __bch2_btree_node_update_key(struct bch_fs *c,
1811                                          struct btree_update *as,
1812                                          struct btree_iter *iter,
1813                                          struct btree *b, struct btree *new_hash,
1814                                          struct bkey_i *new_key)
1815 {
1816         struct btree *parent;
1817         int ret;
1818
1819         btree_update_will_delete_key(as, &b->key);
1820         btree_update_will_add_key(as, new_key);
1821
1822         parent = btree_node_parent(iter, b);
1823         if (parent) {
1824                 if (new_hash) {
1825                         bkey_copy(&new_hash->key, new_key);
1826                         ret = bch2_btree_node_hash_insert(&c->btree_cache,
1827                                         new_hash, b->c.level, b->c.btree_id);
1828                         BUG_ON(ret);
1829                 }
1830
1831                 bch2_keylist_add(&as->parent_keys, new_key);
1832                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, 0);
1833
1834                 if (new_hash) {
1835                         mutex_lock(&c->btree_cache.lock);
1836                         bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
1837
1838                         bch2_btree_node_hash_remove(&c->btree_cache, b);
1839
1840                         bkey_copy(&b->key, new_key);
1841                         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1842                         BUG_ON(ret);
1843                         mutex_unlock(&c->btree_cache.lock);
1844                 } else {
1845                         bkey_copy(&b->key, new_key);
1846                 }
1847         } else {
1848                 BUG_ON(btree_node_root(c, b) != b);
1849
1850                 bch2_btree_node_lock_write(b, iter);
1851                 bkey_copy(&b->key, new_key);
1852
1853                 if (btree_ptr_hash_val(&b->key) != b->hash_val) {
1854                         mutex_lock(&c->btree_cache.lock);
1855                         bch2_btree_node_hash_remove(&c->btree_cache, b);
1856
1857                         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1858                         BUG_ON(ret);
1859                         mutex_unlock(&c->btree_cache.lock);
1860                 }
1861
1862                 btree_update_updated_root(as, b);
1863                 bch2_btree_node_unlock_write(b, iter);
1864         }
1865
1866         bch2_btree_update_done(as);
1867 }
1868
1869 int bch2_btree_node_update_key(struct bch_fs *c, struct btree_iter *iter,
1870                                struct btree *b,
1871                                struct bkey_i *new_key)
1872 {
1873         struct btree *parent = btree_node_parent(iter, b);
1874         struct btree_update *as = NULL;
1875         struct btree *new_hash = NULL;
1876         struct closure cl;
1877         int ret;
1878
1879         closure_init_stack(&cl);
1880
1881         if (!bch2_btree_iter_upgrade(iter, U8_MAX))
1882                 return -EINTR;
1883
1884         if (!down_read_trylock(&c->gc_lock)) {
1885                 bch2_trans_unlock(iter->trans);
1886                 down_read(&c->gc_lock);
1887
1888                 if (!bch2_trans_relock(iter->trans)) {
1889                         ret = -EINTR;
1890                         goto err;
1891                 }
1892         }
1893
1894         /*
1895          * check btree_ptr_hash_val() after @b is locked by
1896          * btree_iter_traverse():
1897          */
1898         if (btree_ptr_hash_val(new_key) != b->hash_val) {
1899                 /* bch2_btree_reserve_get will unlock */
1900                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1901                 if (ret) {
1902                         bch2_trans_unlock(iter->trans);
1903                         up_read(&c->gc_lock);
1904                         closure_sync(&cl);
1905                         down_read(&c->gc_lock);
1906
1907                         if (!bch2_trans_relock(iter->trans)) {
1908                                 ret = -EINTR;
1909                                 goto err;
1910                         }
1911                 }
1912
1913                 new_hash = bch2_btree_node_mem_alloc(c);
1914         }
1915 retry:
1916         as = bch2_btree_update_start(iter->trans, iter->btree_id,
1917                 parent ? btree_update_reserve_required(c, parent) : 0,
1918                 BTREE_INSERT_NOFAIL, &cl);
1919
1920         if (IS_ERR(as)) {
1921                 ret = PTR_ERR(as);
1922                 if (ret == -EAGAIN)
1923                         ret = -EINTR;
1924
1925                 if (ret == -EINTR) {
1926                         bch2_trans_unlock(iter->trans);
1927                         up_read(&c->gc_lock);
1928                         closure_sync(&cl);
1929                         down_read(&c->gc_lock);
1930
1931                         if (bch2_trans_relock(iter->trans))
1932                                 goto retry;
1933                 }
1934
1935                 goto err;
1936         }
1937
1938         ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(new_key));
1939         if (ret)
1940                 goto err_free_update;
1941
1942         __bch2_btree_node_update_key(c, as, iter, b, new_hash, new_key);
1943
1944         bch2_btree_iter_downgrade(iter);
1945 err:
1946         if (new_hash) {
1947                 mutex_lock(&c->btree_cache.lock);
1948                 list_move(&new_hash->list, &c->btree_cache.freeable);
1949                 mutex_unlock(&c->btree_cache.lock);
1950
1951                 six_unlock_write(&new_hash->c.lock);
1952                 six_unlock_intent(&new_hash->c.lock);
1953         }
1954         up_read(&c->gc_lock);
1955         closure_sync(&cl);
1956         return ret;
1957 err_free_update:
1958         bch2_btree_update_free(as);
1959         goto err;
1960 }
1961
1962 /* Init code: */
1963
1964 /*
1965  * Only for filesystem bringup, when first reading the btree roots or allocating
1966  * btree roots when initializing a new filesystem:
1967  */
1968 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
1969 {
1970         BUG_ON(btree_node_root(c, b));
1971
1972         bch2_btree_set_root_inmem(c, b);
1973 }
1974
1975 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
1976 {
1977         struct closure cl;
1978         struct btree *b;
1979         int ret;
1980
1981         closure_init_stack(&cl);
1982
1983         do {
1984                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1985                 closure_sync(&cl);
1986         } while (ret);
1987
1988         b = bch2_btree_node_mem_alloc(c);
1989         bch2_btree_cache_cannibalize_unlock(c);
1990
1991         set_btree_node_fake(b);
1992         set_btree_node_need_rewrite(b);
1993         b->c.level      = 0;
1994         b->c.btree_id   = id;
1995
1996         bkey_btree_ptr_init(&b->key);
1997         b->key.k.p = POS_MAX;
1998         *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
1999
2000         bch2_bset_init_first(b, &b->data->keys);
2001         bch2_btree_build_aux_trees(b);
2002
2003         b->data->flags = 0;
2004         btree_set_min(b, POS_MIN);
2005         btree_set_max(b, POS_MAX);
2006         b->data->format = bch2_btree_calc_format(b);
2007         btree_node_set_format(b, b->data->format);
2008
2009         ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2010                                           b->c.level, b->c.btree_id);
2011         BUG_ON(ret);
2012
2013         bch2_btree_set_root_inmem(c, b);
2014
2015         six_unlock_write(&b->c.lock);
2016         six_unlock_intent(&b->c.lock);
2017 }
2018
2019 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2020 {
2021         struct btree_update *as;
2022
2023         mutex_lock(&c->btree_interior_update_lock);
2024         list_for_each_entry(as, &c->btree_interior_update_list, list)
2025                 pr_buf(out, "%p m %u w %u r %u j %llu\n",
2026                        as,
2027                        as->mode,
2028                        as->nodes_written,
2029                        atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2030                        as->journal.seq);
2031         mutex_unlock(&c->btree_interior_update_lock);
2032 }
2033
2034 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *c)
2035 {
2036         size_t ret = 0;
2037         struct list_head *i;
2038
2039         mutex_lock(&c->btree_interior_update_lock);
2040         list_for_each(i, &c->btree_interior_update_list)
2041                 ret++;
2042         mutex_unlock(&c->btree_interior_update_lock);
2043
2044         return ret;
2045 }
2046
2047 void bch2_journal_entries_to_btree_roots(struct bch_fs *c, struct jset *jset)
2048 {
2049         struct btree_root *r;
2050         struct jset_entry *entry;
2051
2052         mutex_lock(&c->btree_root_lock);
2053
2054         vstruct_for_each(jset, entry)
2055                 if (entry->type == BCH_JSET_ENTRY_btree_root) {
2056                         r = &c->btree_roots[entry->btree_id];
2057                         r->level = entry->level;
2058                         r->alive = true;
2059                         bkey_copy(&r->key, &entry->start[0]);
2060                 }
2061
2062         mutex_unlock(&c->btree_root_lock);
2063 }
2064
2065 struct jset_entry *
2066 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2067                                     struct jset_entry *start,
2068                                     struct jset_entry *end)
2069 {
2070         struct jset_entry *entry;
2071         unsigned long have = 0;
2072         unsigned i;
2073
2074         for (entry = start; entry < end; entry = vstruct_next(entry))
2075                 if (entry->type == BCH_JSET_ENTRY_btree_root)
2076                         __set_bit(entry->btree_id, &have);
2077
2078         mutex_lock(&c->btree_root_lock);
2079
2080         for (i = 0; i < BTREE_ID_NR; i++)
2081                 if (c->btree_roots[i].alive && !test_bit(i, &have)) {
2082                         journal_entry_set(end,
2083                                           BCH_JSET_ENTRY_btree_root,
2084                                           i, c->btree_roots[i].level,
2085                                           &c->btree_roots[i].key,
2086                                           c->btree_roots[i].key.u64s);
2087                         end = vstruct_next(end);
2088                 }
2089
2090         mutex_unlock(&c->btree_root_lock);
2091
2092         return end;
2093 }
2094
2095 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2096 {
2097         if (c->btree_interior_update_worker)
2098                 destroy_workqueue(c->btree_interior_update_worker);
2099         mempool_exit(&c->btree_interior_update_pool);
2100 }
2101
2102 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2103 {
2104         mutex_init(&c->btree_reserve_cache_lock);
2105         INIT_LIST_HEAD(&c->btree_interior_update_list);
2106         INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2107         mutex_init(&c->btree_interior_update_lock);
2108         INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2109
2110         c->btree_interior_update_worker =
2111                 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2112         if (!c->btree_interior_update_worker)
2113                 return -ENOMEM;
2114
2115         return mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2116                                          sizeof(struct btree_update));
2117 }