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 static void bch2_btree_insert_node(struct btree_update *, struct btree_trans *,
  26                                    struct btree_path *, struct btree *,
  27                                    struct keylist *, unsigned);
  28 static void bch2_btree_update_add_new_node(struct btree_update *, struct btree *);
  29
  30 /* Debug code: */
  31
  32 /*
  33  * Verify that child nodes correctly span parent node's range:
  34  */
  35 static void btree_node_interior_verify(struct bch_fs *c, struct btree *b)
  36 {
  37 #ifdef CONFIG_BCACHEFS_DEBUG
  38         struct bpos next_node = b->data->min_key;
  39         struct btree_node_iter iter;
  40         struct bkey_s_c k;
  41         struct bkey_s_c_btree_ptr_v2 bp;
  42         struct bkey unpacked;
  43         char buf1[100], buf2[100];
  44
  45         BUG_ON(!b->c.level);
  46
  47         if (!test_bit(BCH_FS_BTREE_INTERIOR_REPLAY_DONE, &c->flags))
  48                 return;
  49
  50         bch2_btree_node_iter_init_from_start(&iter, b);
  51
  52         while (1) {
  53                 k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked);
  54                 if (k.k->type != KEY_TYPE_btree_ptr_v2)
  55                         break;
  56                 bp = bkey_s_c_to_btree_ptr_v2(k);
  57
  58                 if (bpos_cmp(next_node, bp.v->min_key)) {
  59                         bch2_dump_btree_node(c, b);
  60                         panic("expected next min_key %s got %s\n",
  61                               (bch2_bpos_to_text(&PBUF(buf1), next_node), buf1),
  62                               (bch2_bpos_to_text(&PBUF(buf2), bp.v->min_key), buf2));
  63                 }
  64
  65                 bch2_btree_node_iter_advance(&iter, b);
  66
  67                 if (bch2_btree_node_iter_end(&iter)) {
  68                         if (bpos_cmp(k.k->p, b->key.k.p)) {
  69                                 bch2_dump_btree_node(c, b);
  70                                 panic("expected end %s got %s\n",
  71                                       (bch2_bpos_to_text(&PBUF(buf1), b->key.k.p), buf1),
  72                                       (bch2_bpos_to_text(&PBUF(buf2), k.k->p), buf2));
  73                         }
  74                         break;
  75                 }
  76
  77                 next_node = bpos_successor(k.k->p);
  78         }
  79 #endif
  80 }
  81
  82 /* Calculate ideal packed bkey format for new btree nodes: */
  83
  84 void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b)
  85 {
  86         struct bkey_packed *k;
  87         struct bset_tree *t;
  88         struct bkey uk;
  89
  90         for_each_bset(b, t)
  91                 bset_tree_for_each_key(b, t, k)
  92                         if (!bkey_deleted(k)) {
  93                                 uk = bkey_unpack_key(b, k);
  94                                 bch2_bkey_format_add_key(s, &uk);
  95                         }
  96 }
  97
  98 static struct bkey_format bch2_btree_calc_format(struct btree *b)
  99 {
 100         struct bkey_format_state s;
 101
 102         bch2_bkey_format_init(&s);
 103         bch2_bkey_format_add_pos(&s, b->data->min_key);
 104         bch2_bkey_format_add_pos(&s, b->data->max_key);
 105         __bch2_btree_calc_format(&s, b);
 106
 107         return bch2_bkey_format_done(&s);
 108 }
 109
 110 static size_t btree_node_u64s_with_format(struct btree *b,
 111                                           struct bkey_format *new_f)
 112 {
 113         struct bkey_format *old_f = &b->format;
 114
 115         /* stupid integer promotion rules */
 116         ssize_t delta =
 117             (((int) new_f->key_u64s - old_f->key_u64s) *
 118              (int) b->nr.packed_keys) +
 119             (((int) new_f->key_u64s - BKEY_U64s) *
 120              (int) b->nr.unpacked_keys);
 121
 122         BUG_ON(delta + b->nr.live_u64s < 0);
 123
 124         return b->nr.live_u64s + delta;
 125 }
 126
 127 /**
 128  * btree_node_format_fits - check if we could rewrite node with a new format
 129  *
 130  * This assumes all keys can pack with the new format -- it just checks if
 131  * the re-packed keys would fit inside the node itself.
 132  */
 133 bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
 134                                  struct bkey_format *new_f)
 135 {
 136         size_t u64s = btree_node_u64s_with_format(b, new_f);
 137
 138         return __vstruct_bytes(struct btree_node, u64s) < btree_bytes(c);
 139 }
 140
 141 /* Btree node freeing/allocation: */
 142
 143 static void __btree_node_free(struct bch_fs *c, struct btree *b)
 144 {
 145         trace_btree_node_free(c, b);
 146
 147         BUG_ON(btree_node_dirty(b));
 148         BUG_ON(btree_node_need_write(b));
 149         BUG_ON(b == btree_node_root(c, b));
 150         BUG_ON(b->ob.nr);
 151         BUG_ON(!list_empty(&b->write_blocked));
 152         BUG_ON(b->will_make_reachable);
 153
 154         clear_btree_node_noevict(b);
 155
 156         mutex_lock(&c->btree_cache.lock);
 157         list_move(&b->list, &c->btree_cache.freeable);
 158         mutex_unlock(&c->btree_cache.lock);
 159 }
 160
 161 static void bch2_btree_node_free_inmem(struct btree_trans *trans,
 162                                        struct btree *b)
 163 {
 164         struct bch_fs *c = trans->c;
 165         struct btree_path *path;
 166
 167         trans_for_each_path(trans, path)
 168                 BUG_ON(path->l[b->c.level].b == b &&
 169                        path->l[b->c.level].lock_seq == b->c.lock.state.seq);
 170
 171         six_lock_write(&b->c.lock, NULL, NULL);
 172
 173         bch2_btree_node_hash_remove(&c->btree_cache, b);
 174         __btree_node_free(c, b);
 175
 176         six_unlock_write(&b->c.lock);
 177         six_unlock_intent(&b->c.lock);
 178 }
 179
 180 static struct btree *__bch2_btree_node_alloc(struct bch_fs *c,
 181                                              struct disk_reservation *res,
 182                                              struct closure *cl,
 183                                              unsigned flags)
 184 {
 185         struct write_point *wp;
 186         struct btree *b;
 187         __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
 188         struct open_buckets ob = { .nr = 0 };
 189         struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
 190         unsigned nr_reserve;
 191         enum alloc_reserve alloc_reserve;
 192
 193         if (flags & BTREE_INSERT_USE_RESERVE) {
 194                 nr_reserve      = 0;
 195                 alloc_reserve   = RESERVE_BTREE_MOVINGGC;
 196         } else {
 197                 nr_reserve      = BTREE_NODE_RESERVE;
 198                 alloc_reserve   = RESERVE_BTREE;
 199         }
 200
 201         mutex_lock(&c->btree_reserve_cache_lock);
 202         if (c->btree_reserve_cache_nr > nr_reserve) {
 203                 struct btree_alloc *a =
 204                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
 205
 206                 ob = a->ob;
 207                 bkey_copy(&tmp.k, &a->k);
 208                 mutex_unlock(&c->btree_reserve_cache_lock);
 209                 goto mem_alloc;
 210         }
 211         mutex_unlock(&c->btree_reserve_cache_lock);
 212
 213 retry:
 214         wp = bch2_alloc_sectors_start(c,
 215                                       c->opts.metadata_target ?:
 216                                       c->opts.foreground_target,
 217                                       0,
 218                                       writepoint_ptr(&c->btree_write_point),
 219                                       &devs_have,
 220                                       res->nr_replicas,
 221                                       c->opts.metadata_replicas_required,
 222                                       alloc_reserve, 0, cl);
 223         if (IS_ERR(wp))
 224                 return ERR_CAST(wp);
 225
 226         if (wp->sectors_free < c->opts.btree_node_size) {
 227                 struct open_bucket *ob;
 228                 unsigned i;
 229
 230                 open_bucket_for_each(c, &wp->ptrs, ob, i)
 231                         if (ob->sectors_free < c->opts.btree_node_size)
 232                                 ob->sectors_free = 0;
 233
 234                 bch2_alloc_sectors_done(c, wp);
 235                 goto retry;
 236         }
 237
 238         bkey_btree_ptr_v2_init(&tmp.k);
 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         b->version_ondisk = c->sb.version;
 275
 276         memset(&b->nr, 0, sizeof(b->nr));
 277         b->data->magic = cpu_to_le64(bset_magic(c));
 278         memset(&b->data->_ptr, 0, sizeof(b->data->_ptr));
 279         b->data->flags = 0;
 280         SET_BTREE_NODE_ID(b->data, as->btree_id);
 281         SET_BTREE_NODE_LEVEL(b->data, level);
 282
 283         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 284                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
 285
 286                 bp->v.mem_ptr           = 0;
 287                 bp->v.seq               = b->data->keys.seq;
 288                 bp->v.sectors_written   = 0;
 289         }
 290
 291         SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
 292
 293         bch2_btree_build_aux_trees(b);
 294
 295         ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
 296         BUG_ON(ret);
 297
 298         trace_btree_node_alloc(c, b);
 299         return b;
 300 }
 301
 302 static void btree_set_min(struct btree *b, struct bpos pos)
 303 {
 304         if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
 305                 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
 306         b->data->min_key = pos;
 307 }
 308
 309 static void btree_set_max(struct btree *b, struct bpos pos)
 310 {
 311         b->key.k.p = pos;
 312         b->data->max_key = pos;
 313 }
 314
 315 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *as,
 316                                                   struct btree *b,
 317                                                   struct bkey_format format)
 318 {
 319         struct btree *n;
 320
 321         n = bch2_btree_node_alloc(as, b->c.level);
 322
 323         SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
 324
 325         btree_set_min(n, b->data->min_key);
 326         btree_set_max(n, b->data->max_key);
 327
 328         n->data->format         = format;
 329         btree_node_set_format(n, format);
 330
 331         bch2_btree_sort_into(as->c, n, b);
 332
 333         btree_node_reset_sib_u64s(n);
 334
 335         n->key.k.p = b->key.k.p;
 336         return n;
 337 }
 338
 339 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
 340                                                        struct btree *b)
 341 {
 342         struct bkey_format new_f = bch2_btree_calc_format(b);
 343
 344         /*
 345          * The keys might expand with the new format - if they wouldn't fit in
 346          * the btree node anymore, use the old format for now:
 347          */
 348         if (!bch2_btree_node_format_fits(as->c, b, &new_f))
 349                 new_f = b->format;
 350
 351         return __bch2_btree_node_alloc_replacement(as, b, new_f);
 352 }
 353
 354 static struct btree *__btree_root_alloc(struct btree_update *as, unsigned level)
 355 {
 356         struct btree *b = bch2_btree_node_alloc(as, level);
 357
 358         btree_set_min(b, POS_MIN);
 359         btree_set_max(b, SPOS_MAX);
 360         b->data->format = bch2_btree_calc_format(b);
 361
 362         btree_node_set_format(b, b->data->format);
 363         bch2_btree_build_aux_trees(b);
 364
 365         bch2_btree_update_add_new_node(as, b);
 366         six_unlock_write(&b->c.lock);
 367
 368         return b;
 369 }
 370
 371 static void bch2_btree_reserve_put(struct btree_update *as)
 372 {
 373         struct bch_fs *c = as->c;
 374
 375         mutex_lock(&c->btree_reserve_cache_lock);
 376
 377         while (as->nr_prealloc_nodes) {
 378                 struct btree *b = as->prealloc_nodes[--as->nr_prealloc_nodes];
 379
 380                 six_unlock_write(&b->c.lock);
 381
 382                 if (c->btree_reserve_cache_nr <
 383                     ARRAY_SIZE(c->btree_reserve_cache)) {
 384                         struct btree_alloc *a =
 385                                 &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
 386
 387                         a->ob = b->ob;
 388                         b->ob.nr = 0;
 389                         bkey_copy(&a->k, &b->key);
 390                 } else {
 391                         bch2_open_buckets_put(c, &b->ob);
 392                 }
 393
 394                 btree_node_lock_type(c, b, SIX_LOCK_write);
 395                 __btree_node_free(c, b);
 396                 six_unlock_write(&b->c.lock);
 397
 398                 six_unlock_intent(&b->c.lock);
 399         }
 400
 401         mutex_unlock(&c->btree_reserve_cache_lock);
 402 }
 403
 404 static int bch2_btree_reserve_get(struct btree_update *as, unsigned nr_nodes,
 405                                   unsigned flags, struct closure *cl)
 406 {
 407         struct bch_fs *c = as->c;
 408         struct btree *b;
 409         int ret;
 410
 411         BUG_ON(nr_nodes > BTREE_RESERVE_MAX);
 412
 413         /*
 414          * Protects reaping from the btree node cache and using the btree node
 415          * open bucket reserve:
 416          */
 417         ret = bch2_btree_cache_cannibalize_lock(c, cl);
 418         if (ret)
 419                 return ret;
 420
 421         while (as->nr_prealloc_nodes < nr_nodes) {
 422                 b = __bch2_btree_node_alloc(c, &as->disk_res,
 423                                             flags & BTREE_INSERT_NOWAIT
 424                                             ? NULL : cl, flags);
 425                 if (IS_ERR(b)) {
 426                         ret = PTR_ERR(b);
 427                         goto err_free;
 428                 }
 429
 430                 as->prealloc_nodes[as->nr_prealloc_nodes++] = b;
 431         }
 432
 433         bch2_btree_cache_cannibalize_unlock(c);
 434         return 0;
 435 err_free:
 436         bch2_btree_cache_cannibalize_unlock(c);
 437         trace_btree_reserve_get_fail(c, nr_nodes, cl);
 438         return ret;
 439 }
 440
 441 /* Asynchronous interior node update machinery */
 442
 443 static void bch2_btree_update_free(struct btree_update *as)
 444 {
 445         struct bch_fs *c = as->c;
 446
 447         if (as->took_gc_lock)
 448                 up_read(&c->gc_lock);
 449         as->took_gc_lock = false;
 450
 451         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 452
 453         bch2_journal_pin_drop(&c->journal, &as->journal);
 454         bch2_journal_pin_flush(&c->journal, &as->journal);
 455         bch2_disk_reservation_put(c, &as->disk_res);
 456         bch2_btree_reserve_put(as);
 457
 458         mutex_lock(&c->btree_interior_update_lock);
 459         list_del(&as->unwritten_list);
 460         list_del(&as->list);
 461         mutex_unlock(&c->btree_interior_update_lock);
 462
 463         closure_debug_destroy(&as->cl);
 464         mempool_free(as, &c->btree_interior_update_pool);
 465
 466         closure_wake_up(&c->btree_interior_update_wait);
 467 }
 468
 469 static void btree_update_will_delete_key(struct btree_update *as,
 470                                          struct bkey_i *k)
 471 {
 472         BUG_ON(bch2_keylist_u64s(&as->old_keys) + k->k.u64s >
 473                ARRAY_SIZE(as->_old_keys));
 474         bch2_keylist_add(&as->old_keys, k);
 475 }
 476
 477 static void btree_update_will_add_key(struct btree_update *as,
 478                                       struct bkey_i *k)
 479 {
 480         BUG_ON(bch2_keylist_u64s(&as->new_keys) + k->k.u64s >
 481                ARRAY_SIZE(as->_new_keys));
 482         bch2_keylist_add(&as->new_keys, k);
 483 }
 484
 485 /*
 486  * The transactional part of an interior btree node update, where we journal the
 487  * update we did to the interior node and update alloc info:
 488  */
 489 static int btree_update_nodes_written_trans(struct btree_trans *trans,
 490                                             struct btree_update *as)
 491 {
 492         struct bkey_i *k;
 493         int ret;
 494
 495         trans->extra_journal_entries = (void *) &as->journal_entries[0];
 496         trans->extra_journal_entry_u64s = as->journal_u64s;
 497         trans->journal_pin = &as->journal;
 498
 499         for_each_keylist_key(&as->new_keys, k) {
 500                 ret = bch2_trans_mark_key(trans,
 501                                           bkey_s_c_null,
 502                                           bkey_i_to_s_c(k),
 503                                           BTREE_TRIGGER_INSERT);
 504                 if (ret)
 505                         return ret;
 506         }
 507
 508         for_each_keylist_key(&as->old_keys, k) {
 509                 ret = bch2_trans_mark_key(trans,
 510                                           bkey_i_to_s_c(k),
 511                                           bkey_s_c_null,
 512                                           BTREE_TRIGGER_OVERWRITE);
 513                 if (ret)
 514                         return ret;
 515         }
 516
 517         return 0;
 518 }
 519
 520 static void btree_update_nodes_written(struct btree_update *as)
 521 {
 522         struct bch_fs *c = as->c;
 523         struct btree *b = as->b;
 524         struct btree_trans trans;
 525         u64 journal_seq = 0;
 526         unsigned i;
 527         int ret;
 528
 529         /*
 530          * If we're already in an error state, it might be because a btree node
 531          * was never written, and we might be trying to free that same btree
 532          * node here, but it won't have been marked as allocated and we'll see
 533          * spurious disk usage inconsistencies in the transactional part below
 534          * if we don't skip it:
 535          */
 536         ret = bch2_journal_error(&c->journal);
 537         if (ret)
 538                 goto err;
 539
 540         BUG_ON(!journal_pin_active(&as->journal));
 541
 542         /*
 543          * Wait for any in flight writes to finish before we free the old nodes
 544          * on disk:
 545          */
 546         for (i = 0; i < as->nr_old_nodes; i++) {
 547                 struct btree *old = as->old_nodes[i];
 548                 __le64 seq;
 549
 550                 six_lock_read(&old->c.lock, NULL, NULL);
 551                 seq = old->data ? old->data->keys.seq : 0;
 552                 six_unlock_read(&old->c.lock);
 553
 554                 if (seq == as->old_nodes_seq[i])
 555                         wait_on_bit_io(&old->flags, BTREE_NODE_write_in_flight_inner,
 556                                        TASK_UNINTERRUPTIBLE);
 557         }
 558
 559         /*
 560          * We did an update to a parent node where the pointers we added pointed
 561          * to child nodes that weren't written yet: now, the child nodes have
 562          * been written so we can write out the update to the interior node.
 563          */
 564
 565         /*
 566          * We can't call into journal reclaim here: we'd block on the journal
 567          * reclaim lock, but we may need to release the open buckets we have
 568          * pinned in order for other btree updates to make forward progress, and
 569          * journal reclaim does btree updates when flushing bkey_cached entries,
 570          * which may require allocations as well.
 571          */
 572         bch2_trans_init(&trans, c, 0, 512);
 573         ret = __bch2_trans_do(&trans, &as->disk_res, &journal_seq,
 574                               BTREE_INSERT_NOFAIL|
 575                               BTREE_INSERT_NOCHECK_RW|
 576                               BTREE_INSERT_JOURNAL_RECLAIM|
 577                               BTREE_INSERT_JOURNAL_RESERVED,
 578                               btree_update_nodes_written_trans(&trans, as));
 579         bch2_trans_exit(&trans);
 580
 581         bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
 582                              "error %i in btree_update_nodes_written()", ret);
 583 err:
 584         if (b) {
 585                 /*
 586                  * @b is the node we did the final insert into:
 587                  *
 588                  * On failure to get a journal reservation, we still have to
 589                  * unblock the write and allow most of the write path to happen
 590                  * so that shutdown works, but the i->journal_seq mechanism
 591                  * won't work to prevent the btree write from being visible (we
 592                  * didn't get a journal sequence number) - instead
 593                  * __bch2_btree_node_write() doesn't do the actual write if
 594                  * we're in journal error state:
 595                  */
 596
 597                 btree_node_lock_type(c, b, SIX_LOCK_intent);
 598                 btree_node_lock_type(c, b, SIX_LOCK_write);
 599                 mutex_lock(&c->btree_interior_update_lock);
 600
 601                 list_del(&as->write_blocked_list);
 602
 603                 /*
 604                  * Node might have been freed, recheck under
 605                  * btree_interior_update_lock:
 606                  */
 607                 if (as->b == b) {
 608                         struct bset *i = btree_bset_last(b);
 609
 610                         BUG_ON(!b->c.level);
 611                         BUG_ON(!btree_node_dirty(b));
 612
 613                         if (!ret) {
 614                                 i->journal_seq = cpu_to_le64(
 615                                         max(journal_seq,
 616                                             le64_to_cpu(i->journal_seq)));
 617
 618                                 bch2_btree_add_journal_pin(c, b, journal_seq);
 619                         } else {
 620                                 /*
 621                                  * If we didn't get a journal sequence number we
 622                                  * can't write this btree node, because recovery
 623                                  * won't know to ignore this write:
 624                                  */
 625                                 set_btree_node_never_write(b);
 626                         }
 627                 }
 628
 629                 mutex_unlock(&c->btree_interior_update_lock);
 630                 six_unlock_write(&b->c.lock);
 631
 632                 btree_node_write_if_need(c, b, SIX_LOCK_intent);
 633                 six_unlock_intent(&b->c.lock);
 634         }
 635
 636         bch2_journal_pin_drop(&c->journal, &as->journal);
 637
 638         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 639
 640         mutex_lock(&c->btree_interior_update_lock);
 641         for (i = 0; i < as->nr_new_nodes; i++) {
 642                 b = as->new_nodes[i];
 643
 644                 BUG_ON(b->will_make_reachable != (unsigned long) as);
 645                 b->will_make_reachable = 0;
 646         }
 647         mutex_unlock(&c->btree_interior_update_lock);
 648
 649         for (i = 0; i < as->nr_new_nodes; i++) {
 650                 b = as->new_nodes[i];
 651
 652                 btree_node_lock_type(c, b, SIX_LOCK_read);
 653                 btree_node_write_if_need(c, b, SIX_LOCK_read);
 654                 six_unlock_read(&b->c.lock);
 655         }
 656
 657         for (i = 0; i < as->nr_open_buckets; i++)
 658                 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
 659
 660         bch2_btree_update_free(as);
 661 }
 662
 663 static void btree_interior_update_work(struct work_struct *work)
 664 {
 665         struct bch_fs *c =
 666                 container_of(work, struct bch_fs, btree_interior_update_work);
 667         struct btree_update *as;
 668
 669         while (1) {
 670                 mutex_lock(&c->btree_interior_update_lock);
 671                 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
 672                                               struct btree_update, unwritten_list);
 673                 if (as && !as->nodes_written)
 674                         as = NULL;
 675                 mutex_unlock(&c->btree_interior_update_lock);
 676
 677                 if (!as)
 678                         break;
 679
 680                 btree_update_nodes_written(as);
 681         }
 682 }
 683
 684 static void btree_update_set_nodes_written(struct closure *cl)
 685 {
 686         struct btree_update *as = container_of(cl, struct btree_update, cl);
 687         struct bch_fs *c = as->c;
 688
 689         mutex_lock(&c->btree_interior_update_lock);
 690         as->nodes_written = true;
 691         mutex_unlock(&c->btree_interior_update_lock);
 692
 693         queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
 694 }
 695
 696 /*
 697  * We're updating @b with pointers to nodes that haven't finished writing yet:
 698  * block @b from being written until @as completes
 699  */
 700 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
 701 {
 702         struct bch_fs *c = as->c;
 703
 704         mutex_lock(&c->btree_interior_update_lock);
 705         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 706
 707         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 708         BUG_ON(!btree_node_dirty(b));
 709
 710         as->mode        = BTREE_INTERIOR_UPDATING_NODE;
 711         as->b           = b;
 712         list_add(&as->write_blocked_list, &b->write_blocked);
 713
 714         mutex_unlock(&c->btree_interior_update_lock);
 715 }
 716
 717 static void btree_update_reparent(struct btree_update *as,
 718                                   struct btree_update *child)
 719 {
 720         struct bch_fs *c = as->c;
 721
 722         lockdep_assert_held(&c->btree_interior_update_lock);
 723
 724         child->b = NULL;
 725         child->mode = BTREE_INTERIOR_UPDATING_AS;
 726
 727         bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL);
 728 }
 729
 730 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
 731 {
 732         struct bkey_i *insert = &b->key;
 733         struct bch_fs *c = as->c;
 734
 735         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 736
 737         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
 738                ARRAY_SIZE(as->journal_entries));
 739
 740         as->journal_u64s +=
 741                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
 742                                   BCH_JSET_ENTRY_btree_root,
 743                                   b->c.btree_id, b->c.level,
 744                                   insert, insert->k.u64s);
 745
 746         mutex_lock(&c->btree_interior_update_lock);
 747         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 748
 749         as->mode        = BTREE_INTERIOR_UPDATING_ROOT;
 750         mutex_unlock(&c->btree_interior_update_lock);
 751 }
 752
 753 /*
 754  * bch2_btree_update_add_new_node:
 755  *
 756  * This causes @as to wait on @b to be written, before it gets to
 757  * bch2_btree_update_nodes_written
 758  *
 759  * Additionally, it sets b->will_make_reachable to prevent any additional writes
 760  * to @b from happening besides the first until @b is reachable on disk
 761  *
 762  * And it adds @b to the list of @as's new nodes, so that we can update sector
 763  * counts in bch2_btree_update_nodes_written:
 764  */
 765 static void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
 766 {
 767         struct bch_fs *c = as->c;
 768
 769         closure_get(&as->cl);
 770
 771         mutex_lock(&c->btree_interior_update_lock);
 772         BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
 773         BUG_ON(b->will_make_reachable);
 774
 775         as->new_nodes[as->nr_new_nodes++] = b;
 776         b->will_make_reachable = 1UL|(unsigned long) as;
 777
 778         mutex_unlock(&c->btree_interior_update_lock);
 779
 780         btree_update_will_add_key(as, &b->key);
 781 }
 782
 783 /*
 784  * returns true if @b was a new node
 785  */
 786 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
 787 {
 788         struct btree_update *as;
 789         unsigned long v;
 790         unsigned i;
 791
 792         mutex_lock(&c->btree_interior_update_lock);
 793         /*
 794          * When b->will_make_reachable != 0, it owns a ref on as->cl that's
 795          * dropped when it gets written by bch2_btree_complete_write - the
 796          * xchg() is for synchronization with bch2_btree_complete_write:
 797          */
 798         v = xchg(&b->will_make_reachable, 0);
 799         as = (struct btree_update *) (v & ~1UL);
 800
 801         if (!as) {
 802                 mutex_unlock(&c->btree_interior_update_lock);
 803                 return;
 804         }
 805
 806         for (i = 0; i < as->nr_new_nodes; i++)
 807                 if (as->new_nodes[i] == b)
 808                         goto found;
 809
 810         BUG();
 811 found:
 812         array_remove_item(as->new_nodes, as->nr_new_nodes, i);
 813         mutex_unlock(&c->btree_interior_update_lock);
 814
 815         if (v & 1)
 816                 closure_put(&as->cl);
 817 }
 818
 819 static void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
 820 {
 821         while (b->ob.nr)
 822                 as->open_buckets[as->nr_open_buckets++] =
 823                         b->ob.v[--b->ob.nr];
 824 }
 825
 826 /*
 827  * @b is being split/rewritten: it may have pointers to not-yet-written btree
 828  * nodes and thus outstanding btree_updates - redirect @b's
 829  * btree_updates to point to this btree_update:
 830  */
 831 static void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 832                                                struct btree *b)
 833 {
 834         struct bch_fs *c = as->c;
 835         struct btree_update *p, *n;
 836         struct btree_write *w;
 837
 838         set_btree_node_dying(b);
 839
 840         if (btree_node_fake(b))
 841                 return;
 842
 843         mutex_lock(&c->btree_interior_update_lock);
 844
 845         /*
 846          * Does this node have any btree_update operations preventing
 847          * it from being written?
 848          *
 849          * If so, redirect them to point to this btree_update: we can
 850          * write out our new nodes, but we won't make them visible until those
 851          * operations complete
 852          */
 853         list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
 854                 list_del_init(&p->write_blocked_list);
 855                 btree_update_reparent(as, p);
 856
 857                 /*
 858                  * for flush_held_btree_writes() waiting on updates to flush or
 859                  * nodes to be writeable:
 860                  */
 861                 closure_wake_up(&c->btree_interior_update_wait);
 862         }
 863
 864         clear_btree_node_dirty(c, b);
 865         clear_btree_node_need_write(b);
 866
 867         /*
 868          * Does this node have unwritten data that has a pin on the journal?
 869          *
 870          * If so, transfer that pin to the btree_update operation -
 871          * note that if we're freeing multiple nodes, we only need to keep the
 872          * oldest pin of any of the nodes we're freeing. We'll release the pin
 873          * when the new nodes are persistent and reachable on disk:
 874          */
 875         w = btree_current_write(b);
 876         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 877         bch2_journal_pin_drop(&c->journal, &w->journal);
 878
 879         w = btree_prev_write(b);
 880         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 881         bch2_journal_pin_drop(&c->journal, &w->journal);
 882
 883         mutex_unlock(&c->btree_interior_update_lock);
 884
 885         /*
 886          * Is this a node that isn't reachable on disk yet?
 887          *
 888          * Nodes that aren't reachable yet have writes blocked until they're
 889          * reachable - now that we've cancelled any pending writes and moved
 890          * things waiting on that write to wait on this update, we can drop this
 891          * node from the list of nodes that the other update is making
 892          * reachable, prior to freeing it:
 893          */
 894         btree_update_drop_new_node(c, b);
 895
 896         btree_update_will_delete_key(as, &b->key);
 897
 898         as->old_nodes[as->nr_old_nodes] = b;
 899         as->old_nodes_seq[as->nr_old_nodes] = b->data->keys.seq;
 900         as->nr_old_nodes++;
 901 }
 902
 903 static void bch2_btree_update_done(struct btree_update *as)
 904 {
 905         BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
 906
 907         if (as->took_gc_lock)
 908                 up_read(&as->c->gc_lock);
 909         as->took_gc_lock = false;
 910
 911         bch2_btree_reserve_put(as);
 912
 913         continue_at(&as->cl, btree_update_set_nodes_written,
 914                     as->c->btree_interior_update_worker);
 915 }
 916
 917 static struct btree_update *
 918 bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
 919                         unsigned level, unsigned nr_nodes, unsigned flags)
 920 {
 921         struct bch_fs *c = trans->c;
 922         struct btree_update *as;
 923         struct closure cl;
 924         int disk_res_flags = (flags & BTREE_INSERT_NOFAIL)
 925                 ? BCH_DISK_RESERVATION_NOFAIL : 0;
 926         int journal_flags = 0;
 927         int ret = 0;
 928
 929         BUG_ON(!path->should_be_locked);
 930
 931         if (flags & BTREE_INSERT_JOURNAL_RESERVED)
 932                 journal_flags |= JOURNAL_RES_GET_RESERVED;
 933
 934         closure_init_stack(&cl);
 935 retry:
 936
 937         /*
 938          * XXX: figure out how far we might need to split,
 939          * instead of locking/reserving all the way to the root:
 940          */
 941         if (!bch2_btree_path_upgrade(trans, path, U8_MAX)) {
 942                 trace_trans_restart_iter_upgrade(trans->ip, _RET_IP_,
 943                                                  path->btree_id, &path->pos);
 944                 ret = btree_trans_restart(trans);
 945                 return ERR_PTR(ret);
 946         }
 947
 948         if (flags & BTREE_INSERT_GC_LOCK_HELD)
 949                 lockdep_assert_held(&c->gc_lock);
 950         else if (!down_read_trylock(&c->gc_lock)) {
 951                 bch2_trans_unlock(trans);
 952                 down_read(&c->gc_lock);
 953                 if (!bch2_trans_relock(trans)) {
 954                         up_read(&c->gc_lock);
 955                         return ERR_PTR(-EINTR);
 956                 }
 957         }
 958
 959         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
 960         memset(as, 0, sizeof(*as));
 961         closure_init(&as->cl, NULL);
 962         as->c           = c;
 963         as->mode        = BTREE_INTERIOR_NO_UPDATE;
 964         as->took_gc_lock = !(flags & BTREE_INSERT_GC_LOCK_HELD);
 965         as->btree_id    = path->btree_id;
 966         INIT_LIST_HEAD(&as->list);
 967         INIT_LIST_HEAD(&as->unwritten_list);
 968         INIT_LIST_HEAD(&as->write_blocked_list);
 969         bch2_keylist_init(&as->old_keys, as->_old_keys);
 970         bch2_keylist_init(&as->new_keys, as->_new_keys);
 971         bch2_keylist_init(&as->parent_keys, as->inline_keys);
 972
 973         mutex_lock(&c->btree_interior_update_lock);
 974         list_add_tail(&as->list, &c->btree_interior_update_list);
 975         mutex_unlock(&c->btree_interior_update_lock);
 976
 977         /*
 978          * We don't want to allocate if we're in an error state, that can cause
 979          * deadlock on emergency shutdown due to open buckets getting stuck in
 980          * the btree_reserve_cache after allocator shutdown has cleared it out.
 981          * This check needs to come after adding us to the btree_interior_update
 982          * list but before calling bch2_btree_reserve_get, to synchronize with
 983          * __bch2_fs_read_only().
 984          */
 985         ret = bch2_journal_error(&c->journal);
 986         if (ret)
 987                 goto err;
 988
 989         ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
 990                                       BTREE_UPDATE_JOURNAL_RES,
 991                                       journal_flags|JOURNAL_RES_GET_NONBLOCK);
 992         if (ret == -EAGAIN) {
 993                 bch2_trans_unlock(trans);
 994
 995                 if (flags & BTREE_INSERT_JOURNAL_RECLAIM) {
 996                         bch2_btree_update_free(as);
 997                         btree_trans_restart(trans);
 998                         return ERR_PTR(ret);
 999                 }
1000
1001                 ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
1002                                 BTREE_UPDATE_JOURNAL_RES,
1003                                 journal_flags);
1004                 if (ret) {
1005                         trace_trans_restart_journal_preres_get(trans->ip, _RET_IP_);
1006                         goto err;
1007                 }
1008
1009                 if (!bch2_trans_relock(trans)) {
1010                         ret = -EINTR;
1011                         goto err;
1012                 }
1013         }
1014
1015         ret = bch2_disk_reservation_get(c, &as->disk_res,
1016                         nr_nodes * c->opts.btree_node_size,
1017                         c->opts.metadata_replicas,
1018                         disk_res_flags);
1019         if (ret)
1020                 goto err;
1021
1022         ret = bch2_btree_reserve_get(as, nr_nodes, flags, &cl);
1023         if (ret)
1024                 goto err;
1025
1026         bch2_journal_pin_add(&c->journal,
1027                              atomic64_read(&c->journal.seq),
1028                              &as->journal, NULL);
1029
1030         return as;
1031 err:
1032         bch2_btree_update_free(as);
1033
1034         if (ret == -EAGAIN) {
1035                 bch2_trans_unlock(trans);
1036                 closure_sync(&cl);
1037                 ret = -EINTR;
1038         }
1039
1040         if (ret == -EINTR && bch2_trans_relock(trans))
1041                 goto retry;
1042
1043         return ERR_PTR(ret);
1044 }
1045
1046 /* Btree root updates: */
1047
1048 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1049 {
1050         /* Root nodes cannot be reaped */
1051         mutex_lock(&c->btree_cache.lock);
1052         list_del_init(&b->list);
1053         mutex_unlock(&c->btree_cache.lock);
1054
1055         if (b->c.level)
1056                 six_lock_pcpu_alloc(&b->c.lock);
1057         else
1058                 six_lock_pcpu_free(&b->c.lock);
1059
1060         mutex_lock(&c->btree_root_lock);
1061         BUG_ON(btree_node_root(c, b) &&
1062                (b->c.level < btree_node_root(c, b)->c.level ||
1063                 !btree_node_dying(btree_node_root(c, b))));
1064
1065         btree_node_root(c, b) = b;
1066         mutex_unlock(&c->btree_root_lock);
1067
1068         bch2_recalc_btree_reserve(c);
1069 }
1070
1071 /**
1072  * bch_btree_set_root - update the root in memory and on disk
1073  *
1074  * To ensure forward progress, the current task must not be holding any
1075  * btree node write locks. However, you must hold an intent lock on the
1076  * old root.
1077  *
1078  * Note: This allocates a journal entry but doesn't add any keys to
1079  * it.  All the btree roots are part of every journal write, so there
1080  * is nothing new to be done.  This just guarantees that there is a
1081  * journal write.
1082  */
1083 static void bch2_btree_set_root(struct btree_update *as,
1084                                 struct btree_trans *trans,
1085                                 struct btree_path *path,
1086                                 struct btree *b)
1087 {
1088         struct bch_fs *c = as->c;
1089         struct btree *old;
1090
1091         trace_btree_set_root(c, b);
1092         BUG_ON(!b->written &&
1093                !test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags));
1094
1095         old = btree_node_root(c, b);
1096
1097         /*
1098          * Ensure no one is using the old root while we switch to the
1099          * new root:
1100          */
1101         bch2_btree_node_lock_write(trans, path, old);
1102
1103         bch2_btree_set_root_inmem(c, b);
1104
1105         btree_update_updated_root(as, b);
1106
1107         /*
1108          * Unlock old root after new root is visible:
1109          *
1110          * The new root isn't persistent, but that's ok: we still have
1111          * an intent lock on the new root, and any updates that would
1112          * depend on the new root would have to update the new root.
1113          */
1114         bch2_btree_node_unlock_write(trans, path, old);
1115 }
1116
1117 /* Interior node updates: */
1118
1119 static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
1120                                         struct btree_trans *trans,
1121                                         struct btree_path *path,
1122                                         struct btree *b,
1123                                         struct btree_node_iter *node_iter,
1124                                         struct bkey_i *insert)
1125 {
1126         struct bch_fs *c = as->c;
1127         struct bkey_packed *k;
1128         const char *invalid;
1129
1130         BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 &&
1131                !btree_ptr_sectors_written(insert));
1132
1133         invalid = bch2_bkey_invalid(c, bkey_i_to_s_c(insert), btree_node_type(b)) ?:
1134                 bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert));
1135         if (invalid) {
1136                 char buf[160];
1137
1138                 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(insert));
1139                 bch2_fs_inconsistent(c, "inserting invalid bkey %s: %s", buf, invalid);
1140                 dump_stack();
1141         }
1142
1143         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1144                ARRAY_SIZE(as->journal_entries));
1145
1146         as->journal_u64s +=
1147                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1148                                   BCH_JSET_ENTRY_btree_keys,
1149                                   b->c.btree_id, b->c.level,
1150                                   insert, insert->k.u64s);
1151
1152         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1153                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1154                 bch2_btree_node_iter_advance(node_iter, b);
1155
1156         bch2_btree_bset_insert_key(trans, path, b, node_iter, insert);
1157         set_btree_node_dirty(c, b);
1158         set_btree_node_need_write(b);
1159 }
1160
1161 static void
1162 __bch2_btree_insert_keys_interior(struct btree_update *as,
1163                                   struct btree_trans *trans,
1164                                   struct btree_path *path,
1165                                   struct btree *b,
1166                                   struct btree_node_iter node_iter,
1167                                   struct keylist *keys)
1168 {
1169         struct bkey_i *insert = bch2_keylist_front(keys);
1170         struct bkey_packed *k;
1171
1172         BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1173
1174         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1175                (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1176                 ;
1177
1178         while (!bch2_keylist_empty(keys)) {
1179                 bch2_insert_fixup_btree_ptr(as, trans, path, b,
1180                                 &node_iter, bch2_keylist_front(keys));
1181                 bch2_keylist_pop_front(keys);
1182         }
1183 }
1184
1185 /*
1186  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1187  * node)
1188  */
1189 static struct btree *__btree_split_node(struct btree_update *as,
1190                                         struct btree *n1)
1191 {
1192         struct bkey_format_state s;
1193         size_t nr_packed = 0, nr_unpacked = 0;
1194         struct btree *n2;
1195         struct bset *set1, *set2;
1196         struct bkey_packed *k, *set2_start, *set2_end, *out, *prev = NULL;
1197         struct bpos n1_pos;
1198
1199         n2 = bch2_btree_node_alloc(as, n1->c.level);
1200         bch2_btree_update_add_new_node(as, n2);
1201
1202         n2->data->max_key       = n1->data->max_key;
1203         n2->data->format        = n1->format;
1204         SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data));
1205         n2->key.k.p = n1->key.k.p;
1206
1207         set1 = btree_bset_first(n1);
1208         set2 = btree_bset_first(n2);
1209
1210         /*
1211          * Has to be a linear search because we don't have an auxiliary
1212          * search tree yet
1213          */
1214         k = set1->start;
1215         while (1) {
1216                 struct bkey_packed *n = bkey_next(k);
1217
1218                 if (n == vstruct_last(set1))
1219                         break;
1220                 if (k->_data - set1->_data >= (le16_to_cpu(set1->u64s) * 3) / 5)
1221                         break;
1222
1223                 if (bkey_packed(k))
1224                         nr_packed++;
1225                 else
1226                         nr_unpacked++;
1227
1228                 prev = k;
1229                 k = n;
1230         }
1231
1232         BUG_ON(!prev);
1233         set2_start      = k;
1234         set2_end        = vstruct_last(set1);
1235
1236         set1->u64s = cpu_to_le16((u64 *) set2_start - set1->_data);
1237         set_btree_bset_end(n1, n1->set);
1238
1239         n1->nr.live_u64s        = le16_to_cpu(set1->u64s);
1240         n1->nr.bset_u64s[0]     = le16_to_cpu(set1->u64s);
1241         n1->nr.packed_keys      = nr_packed;
1242         n1->nr.unpacked_keys    = nr_unpacked;
1243
1244         n1_pos = bkey_unpack_pos(n1, prev);
1245         if (as->c->sb.version < bcachefs_metadata_version_snapshot)
1246                 n1_pos.snapshot = U32_MAX;
1247
1248         btree_set_max(n1, n1_pos);
1249         btree_set_min(n2, bpos_successor(n1->key.k.p));
1250
1251         bch2_bkey_format_init(&s);
1252         bch2_bkey_format_add_pos(&s, n2->data->min_key);
1253         bch2_bkey_format_add_pos(&s, n2->data->max_key);
1254
1255         for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1256                 struct bkey uk = bkey_unpack_key(n1, k);
1257                 bch2_bkey_format_add_key(&s, &uk);
1258         }
1259
1260         n2->data->format = bch2_bkey_format_done(&s);
1261         btree_node_set_format(n2, n2->data->format);
1262
1263         out = set2->start;
1264         memset(&n2->nr, 0, sizeof(n2->nr));
1265
1266         for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1267                 BUG_ON(!bch2_bkey_transform(&n2->format, out, bkey_packed(k)
1268                                        ? &n1->format : &bch2_bkey_format_current, k));
1269                 out->format = KEY_FORMAT_LOCAL_BTREE;
1270                 btree_keys_account_key_add(&n2->nr, 0, out);
1271                 out = bkey_next(out);
1272         }
1273
1274         set2->u64s = cpu_to_le16((u64 *) out - set2->_data);
1275         set_btree_bset_end(n2, n2->set);
1276
1277         BUG_ON(!set1->u64s);
1278         BUG_ON(!set2->u64s);
1279
1280         btree_node_reset_sib_u64s(n1);
1281         btree_node_reset_sib_u64s(n2);
1282
1283         bch2_verify_btree_nr_keys(n1);
1284         bch2_verify_btree_nr_keys(n2);
1285
1286         if (n1->c.level) {
1287                 btree_node_interior_verify(as->c, n1);
1288                 btree_node_interior_verify(as->c, n2);
1289         }
1290
1291         return n2;
1292 }
1293
1294 /*
1295  * For updates to interior nodes, we've got to do the insert before we split
1296  * because the stuff we're inserting has to be inserted atomically. Post split,
1297  * the keys might have to go in different nodes and the split would no longer be
1298  * atomic.
1299  *
1300  * Worse, if the insert is from btree node coalescing, if we do the insert after
1301  * we do the split (and pick the pivot) - the pivot we pick might be between
1302  * nodes that were coalesced, and thus in the middle of a child node post
1303  * coalescing:
1304  */
1305 static void btree_split_insert_keys(struct btree_update *as,
1306                                     struct btree_trans *trans,
1307                                     struct btree_path *path,
1308                                     struct btree *b,
1309                                     struct keylist *keys)
1310 {
1311         struct btree_node_iter node_iter;
1312         struct bkey_i *k = bch2_keylist_front(keys);
1313         struct bkey_packed *src, *dst, *n;
1314         struct bset *i;
1315
1316         bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
1317
1318         __bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
1319
1320         /*
1321          * We can't tolerate whiteouts here - with whiteouts there can be
1322          * duplicate keys, and it would be rather bad if we picked a duplicate
1323          * for the pivot:
1324          */
1325         i = btree_bset_first(b);
1326         src = dst = i->start;
1327         while (src != vstruct_last(i)) {
1328                 n = bkey_next(src);
1329                 if (!bkey_deleted(src)) {
1330                         memmove_u64s_down(dst, src, src->u64s);
1331                         dst = bkey_next(dst);
1332                 }
1333                 src = n;
1334         }
1335
1336         /* Also clear out the unwritten whiteouts area: */
1337         b->whiteout_u64s = 0;
1338
1339         i->u64s = cpu_to_le16((u64 *) dst - i->_data);
1340         set_btree_bset_end(b, b->set);
1341
1342         BUG_ON(b->nsets != 1 ||
1343                b->nr.live_u64s != le16_to_cpu(btree_bset_first(b)->u64s));
1344
1345         btree_node_interior_verify(as->c, b);
1346 }
1347
1348 static void btree_split(struct btree_update *as, struct btree_trans *trans,
1349                         struct btree_path *path, struct btree *b,
1350                         struct keylist *keys, unsigned flags)
1351 {
1352         struct bch_fs *c = as->c;
1353         struct btree *parent = btree_node_parent(path, b);
1354         struct btree *n1, *n2 = NULL, *n3 = NULL;
1355         u64 start_time = local_clock();
1356
1357         BUG_ON(!parent && (b != btree_node_root(c, b)));
1358         BUG_ON(!btree_node_intent_locked(path, btree_node_root(c, b)->c.level));
1359
1360         bch2_btree_interior_update_will_free_node(as, b);
1361
1362         n1 = bch2_btree_node_alloc_replacement(as, b);
1363         bch2_btree_update_add_new_node(as, n1);
1364
1365         if (keys)
1366                 btree_split_insert_keys(as, trans, path, n1, keys);
1367
1368         if (bset_u64s(&n1->set[0]) > BTREE_SPLIT_THRESHOLD(c)) {
1369                 trace_btree_split(c, b);
1370
1371                 n2 = __btree_split_node(as, n1);
1372
1373                 bch2_btree_build_aux_trees(n2);
1374                 bch2_btree_build_aux_trees(n1);
1375                 six_unlock_write(&n2->c.lock);
1376                 six_unlock_write(&n1->c.lock);
1377
1378                 bch2_btree_node_write(c, n1, SIX_LOCK_intent);
1379                 bch2_btree_node_write(c, n2, SIX_LOCK_intent);
1380
1381                 /*
1382                  * Note that on recursive parent_keys == keys, so we
1383                  * can't start adding new keys to parent_keys before emptying it
1384                  * out (which we did with btree_split_insert_keys() above)
1385                  */
1386                 bch2_keylist_add(&as->parent_keys, &n1->key);
1387                 bch2_keylist_add(&as->parent_keys, &n2->key);
1388
1389                 if (!parent) {
1390                         /* Depth increases, make a new root */
1391                         n3 = __btree_root_alloc(as, b->c.level + 1);
1392
1393                         n3->sib_u64s[0] = U16_MAX;
1394                         n3->sib_u64s[1] = U16_MAX;
1395
1396                         btree_split_insert_keys(as, trans, path, n3, &as->parent_keys);
1397
1398                         bch2_btree_node_write(c, n3, SIX_LOCK_intent);
1399                 }
1400         } else {
1401                 trace_btree_compact(c, b);
1402
1403                 bch2_btree_build_aux_trees(n1);
1404                 six_unlock_write(&n1->c.lock);
1405
1406                 bch2_btree_node_write(c, n1, SIX_LOCK_intent);
1407
1408                 if (parent)
1409                         bch2_keylist_add(&as->parent_keys, &n1->key);
1410         }
1411
1412         /* New nodes all written, now make them visible: */
1413
1414         if (parent) {
1415                 /* Split a non root node */
1416                 bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1417         } else if (n3) {
1418                 bch2_btree_set_root(as, trans, path, n3);
1419         } else {
1420                 /* Root filled up but didn't need to be split */
1421                 bch2_btree_set_root(as, trans, path, n1);
1422         }
1423
1424         bch2_btree_update_get_open_buckets(as, n1);
1425         if (n2)
1426                 bch2_btree_update_get_open_buckets(as, n2);
1427         if (n3)
1428                 bch2_btree_update_get_open_buckets(as, n3);
1429
1430         /* Successful split, update the path to point to the new nodes: */
1431
1432         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1433         if (n3)
1434                 bch2_trans_node_add(trans, n3);
1435         if (n2)
1436                 bch2_trans_node_add(trans, n2);
1437         bch2_trans_node_add(trans, n1);
1438
1439         /*
1440          * The old node must be freed (in memory) _before_ unlocking the new
1441          * nodes - else another thread could re-acquire a read lock on the old
1442          * node after another thread has locked and updated the new node, thus
1443          * seeing stale data:
1444          */
1445         bch2_btree_node_free_inmem(trans, b);
1446
1447         if (n3)
1448                 six_unlock_intent(&n3->c.lock);
1449         if (n2)
1450                 six_unlock_intent(&n2->c.lock);
1451         six_unlock_intent(&n1->c.lock);
1452
1453         bch2_trans_verify_locks(trans);
1454
1455         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_split],
1456                                start_time);
1457 }
1458
1459 static void
1460 bch2_btree_insert_keys_interior(struct btree_update *as,
1461                                 struct btree_trans *trans,
1462                                 struct btree_path *path,
1463                                 struct btree *b,
1464                                 struct keylist *keys)
1465 {
1466         struct btree_path *linked;
1467
1468         __bch2_btree_insert_keys_interior(as, trans, path, b,
1469                                           path->l[b->c.level].iter, keys);
1470
1471         btree_update_updated_node(as, b);
1472
1473         trans_for_each_path_with_node(trans, b, linked)
1474                 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1475
1476         bch2_trans_verify_paths(trans);
1477 }
1478
1479 /**
1480  * bch_btree_insert_node - insert bkeys into a given btree node
1481  *
1482  * @iter:               btree iterator
1483  * @keys:               list of keys to insert
1484  * @hook:               insert callback
1485  * @persistent:         if not null, @persistent will wait on journal write
1486  *
1487  * Inserts as many keys as it can into a given btree node, splitting it if full.
1488  * If a split occurred, this function will return early. This can only happen
1489  * for leaf nodes -- inserts into interior nodes have to be atomic.
1490  */
1491 static void bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans,
1492                                    struct btree_path *path, struct btree *b,
1493                                    struct keylist *keys, unsigned flags)
1494 {
1495         struct bch_fs *c = as->c;
1496         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1497         int old_live_u64s = b->nr.live_u64s;
1498         int live_u64s_added, u64s_added;
1499
1500         lockdep_assert_held(&c->gc_lock);
1501         BUG_ON(!btree_node_intent_locked(path, btree_node_root(c, b)->c.level));
1502         BUG_ON(!b->c.level);
1503         BUG_ON(!as || as->b);
1504         bch2_verify_keylist_sorted(keys);
1505
1506         bch2_btree_node_lock_for_insert(trans, path, b);
1507
1508         if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1509                 bch2_btree_node_unlock_write(trans, path, b);
1510                 goto split;
1511         }
1512
1513         btree_node_interior_verify(c, b);
1514
1515         bch2_btree_insert_keys_interior(as, trans, path, b, keys);
1516
1517         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1518         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1519
1520         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1521                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1522         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1523                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1524
1525         if (u64s_added > live_u64s_added &&
1526             bch2_maybe_compact_whiteouts(c, b))
1527                 bch2_trans_node_reinit_iter(trans, b);
1528
1529         bch2_btree_node_unlock_write(trans, path, b);
1530
1531         btree_node_interior_verify(c, b);
1532         return;
1533 split:
1534         btree_split(as, trans, path, b, keys, flags);
1535 }
1536
1537 int bch2_btree_split_leaf(struct btree_trans *trans,
1538                           struct btree_path *path,
1539                           unsigned flags)
1540 {
1541         struct bch_fs *c = trans->c;
1542         struct btree *b = path_l(path)->b;
1543         struct btree_update *as;
1544         unsigned l;
1545         int ret = 0;
1546
1547         as = bch2_btree_update_start(trans, path, path->level,
1548                 btree_update_reserve_required(c, b), flags);
1549         if (IS_ERR(as))
1550                 return PTR_ERR(as);
1551
1552         btree_split(as, trans, path, b, NULL, flags);
1553         bch2_btree_update_done(as);
1554
1555         for (l = path->level + 1; btree_path_node(path, l) && !ret; l++)
1556                 ret = bch2_foreground_maybe_merge(trans, path, l, flags);
1557
1558         return ret;
1559 }
1560
1561 int __bch2_foreground_maybe_merge(struct btree_trans *trans,
1562                                   struct btree_path *path,
1563                                   unsigned level,
1564                                   unsigned flags,
1565                                   enum btree_node_sibling sib)
1566 {
1567         struct bch_fs *c = trans->c;
1568         struct btree_path *sib_path = NULL;
1569         struct btree_update *as;
1570         struct bkey_format_state new_s;
1571         struct bkey_format new_f;
1572         struct bkey_i delete;
1573         struct btree *b, *m, *n, *prev, *next, *parent;
1574         struct bpos sib_pos;
1575         size_t sib_u64s;
1576         int ret = 0;
1577
1578         BUG_ON(!path->should_be_locked);
1579         BUG_ON(!btree_node_locked(path, level));
1580
1581         b = path->l[level].b;
1582
1583         if ((sib == btree_prev_sib && !bpos_cmp(b->data->min_key, POS_MIN)) ||
1584             (sib == btree_next_sib && !bpos_cmp(b->data->max_key, SPOS_MAX))) {
1585                 b->sib_u64s[sib] = U16_MAX;
1586                 return 0;
1587         }
1588
1589         sib_pos = sib == btree_prev_sib
1590                 ? bpos_predecessor(b->data->min_key)
1591                 : bpos_successor(b->data->max_key);
1592
1593         sib_path = bch2_path_get(trans, false, path->btree_id,
1594                                  sib_pos, U8_MAX, level, true);
1595         ret = bch2_btree_path_traverse(trans, sib_path, false);
1596         if (ret)
1597                 goto err;
1598
1599         sib_path->should_be_locked = true;
1600
1601         m = sib_path->l[level].b;
1602
1603         if (btree_node_parent(path, b) !=
1604             btree_node_parent(sib_path, m)) {
1605                 b->sib_u64s[sib] = U16_MAX;
1606                 goto out;
1607         }
1608
1609         if (sib == btree_prev_sib) {
1610                 prev = m;
1611                 next = b;
1612         } else {
1613                 prev = b;
1614                 next = m;
1615         }
1616
1617         if (bkey_cmp(bpos_successor(prev->data->max_key), next->data->min_key)) {
1618                 char buf1[100], buf2[100];
1619
1620                 bch2_bpos_to_text(&PBUF(buf1), prev->data->max_key);
1621                 bch2_bpos_to_text(&PBUF(buf2), next->data->min_key);
1622                 bch_err(c,
1623                         "btree topology error in btree merge:\n"
1624                         "  prev ends at   %s\n"
1625                         "  next starts at %s",
1626                         buf1, buf2);
1627                 bch2_topology_error(c);
1628                 ret = -EIO;
1629                 goto err;
1630         }
1631
1632         bch2_bkey_format_init(&new_s);
1633         bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
1634         __bch2_btree_calc_format(&new_s, prev);
1635         __bch2_btree_calc_format(&new_s, next);
1636         bch2_bkey_format_add_pos(&new_s, next->data->max_key);
1637         new_f = bch2_bkey_format_done(&new_s);
1638
1639         sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1640                 btree_node_u64s_with_format(m, &new_f);
1641
1642         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1643                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1644                 sib_u64s /= 2;
1645                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1646         }
1647
1648         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1649         sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
1650         b->sib_u64s[sib] = sib_u64s;
1651
1652         if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
1653                 goto out;
1654
1655         parent = btree_node_parent(path, b);
1656         as = bch2_btree_update_start(trans, path, level,
1657                          btree_update_reserve_required(c, parent) + 1,
1658                          flags|
1659                          BTREE_INSERT_NOFAIL|
1660                          BTREE_INSERT_USE_RESERVE);
1661         ret = PTR_ERR_OR_ZERO(as);
1662         if (ret)
1663                 goto err;
1664
1665         trace_btree_merge(c, b);
1666
1667         bch2_btree_interior_update_will_free_node(as, b);
1668         bch2_btree_interior_update_will_free_node(as, m);
1669
1670         n = bch2_btree_node_alloc(as, b->c.level);
1671         bch2_btree_update_add_new_node(as, n);
1672
1673         btree_set_min(n, prev->data->min_key);
1674         btree_set_max(n, next->data->max_key);
1675         n->data->format         = new_f;
1676
1677         btree_node_set_format(n, new_f);
1678
1679         bch2_btree_sort_into(c, n, prev);
1680         bch2_btree_sort_into(c, n, next);
1681
1682         bch2_btree_build_aux_trees(n);
1683         six_unlock_write(&n->c.lock);
1684
1685         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1686
1687         bkey_init(&delete.k);
1688         delete.k.p = prev->key.k.p;
1689         bch2_keylist_add(&as->parent_keys, &delete);
1690         bch2_keylist_add(&as->parent_keys, &n->key);
1691
1692         bch2_trans_verify_paths(trans);
1693
1694         bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1695
1696         bch2_trans_verify_paths(trans);
1697
1698         bch2_btree_update_get_open_buckets(as, n);
1699
1700         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1701         six_lock_increment(&m->c.lock, SIX_LOCK_intent);
1702
1703         bch2_trans_node_add(trans, n);
1704
1705         bch2_trans_verify_paths(trans);
1706
1707         bch2_btree_node_free_inmem(trans, b);
1708         bch2_btree_node_free_inmem(trans, m);
1709
1710         six_unlock_intent(&n->c.lock);
1711
1712         bch2_btree_update_done(as);
1713 out:
1714 err:
1715         bch2_path_put(trans, sib_path, true);
1716         bch2_trans_verify_locks(trans);
1717         return ret;
1718 }
1719
1720 /**
1721  * bch_btree_node_rewrite - Rewrite/move a btree node
1722  */
1723 int bch2_btree_node_rewrite(struct btree_trans *trans,
1724                             struct btree_iter *iter,
1725                             __le64 seq, unsigned flags)
1726 {
1727         struct bch_fs *c = trans->c;
1728         struct btree *b, *n, *parent;
1729         struct btree_update *as;
1730         int ret;
1731
1732         flags |= BTREE_INSERT_NOFAIL;
1733 retry:
1734         ret = bch2_btree_iter_traverse(iter);
1735         if (ret)
1736                 goto out;
1737
1738         b = bch2_btree_iter_peek_node(iter);
1739         ret = PTR_ERR_OR_ZERO(b);
1740         if (ret)
1741                 goto out;
1742
1743         if (!b || b->data->keys.seq != seq)
1744                 goto out;
1745
1746         parent = btree_node_parent(iter->path, b);
1747         as = bch2_btree_update_start(trans, iter->path, b->c.level,
1748                 (parent
1749                  ? btree_update_reserve_required(c, parent)
1750                  : 0) + 1,
1751                 flags);
1752         ret = PTR_ERR_OR_ZERO(as);
1753         if (ret == -EINTR)
1754                 goto retry;
1755         if (ret) {
1756                 trace_btree_gc_rewrite_node_fail(c, b);
1757                 goto out;
1758         }
1759
1760         bch2_btree_interior_update_will_free_node(as, b);
1761
1762         n = bch2_btree_node_alloc_replacement(as, b);
1763         bch2_btree_update_add_new_node(as, n);
1764
1765         bch2_btree_build_aux_trees(n);
1766         six_unlock_write(&n->c.lock);
1767
1768         trace_btree_gc_rewrite_node(c, b);
1769
1770         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1771
1772         if (parent) {
1773                 bch2_keylist_add(&as->parent_keys, &n->key);
1774                 bch2_btree_insert_node(as, trans, iter->path, parent,
1775                                        &as->parent_keys, flags);
1776         } else {
1777                 bch2_btree_set_root(as, trans, iter->path, n);
1778         }
1779
1780         bch2_btree_update_get_open_buckets(as, n);
1781
1782         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1783         bch2_trans_node_add(trans, n);
1784         bch2_btree_node_free_inmem(trans, b);
1785         six_unlock_intent(&n->c.lock);
1786
1787         bch2_btree_update_done(as);
1788 out:
1789         bch2_btree_path_downgrade(iter->path);
1790         return ret;
1791 }
1792
1793 struct async_btree_rewrite {
1794         struct bch_fs           *c;
1795         struct work_struct      work;
1796         enum btree_id           btree_id;
1797         unsigned                level;
1798         struct bpos             pos;
1799         __le64                  seq;
1800 };
1801
1802 void async_btree_node_rewrite_work(struct work_struct *work)
1803 {
1804         struct async_btree_rewrite *a =
1805                 container_of(work, struct async_btree_rewrite, work);
1806         struct bch_fs *c = a->c;
1807         struct btree_trans trans;
1808         struct btree_iter iter;
1809
1810         bch2_trans_init(&trans, c, 0, 0);
1811         bch2_trans_node_iter_init(&trans, &iter, a->btree_id, a->pos,
1812                                         BTREE_MAX_DEPTH, a->level, 0);
1813         bch2_btree_node_rewrite(&trans, &iter, a->seq, 0);
1814         bch2_trans_iter_exit(&trans, &iter);
1815         bch2_trans_exit(&trans);
1816         percpu_ref_put(&c->writes);
1817         kfree(a);
1818 }
1819
1820 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
1821 {
1822         struct async_btree_rewrite *a;
1823
1824         if (!test_bit(BCH_FS_BTREE_INTERIOR_REPLAY_DONE, &c->flags))
1825                 return;
1826
1827         if (!percpu_ref_tryget(&c->writes))
1828                 return;
1829
1830         a = kmalloc(sizeof(*a), GFP_NOFS);
1831         if (!a) {
1832                 percpu_ref_put(&c->writes);
1833                 return;
1834         }
1835
1836         a->c            = c;
1837         a->btree_id     = b->c.btree_id;
1838         a->level        = b->c.level;
1839         a->pos          = b->key.k.p;
1840         a->seq          = b->data->keys.seq;
1841
1842         INIT_WORK(&a->work, async_btree_node_rewrite_work);
1843         queue_work(c->btree_interior_update_worker, &a->work);
1844 }
1845
1846 static int __bch2_btree_node_update_key(struct btree_trans *trans,
1847                                         struct btree_iter *iter,
1848                                         struct btree *b, struct btree *new_hash,
1849                                         struct bkey_i *new_key,
1850                                         bool skip_triggers)
1851 {
1852         struct bch_fs *c = trans->c;
1853         struct btree_iter iter2 = { NULL };
1854         struct btree *parent;
1855         u64 journal_entries[BKEY_BTREE_PTR_U64s_MAX];
1856         int ret;
1857
1858         if (!skip_triggers) {
1859                 ret = bch2_trans_mark_key(trans,
1860                                           bkey_s_c_null,
1861                                           bkey_i_to_s_c(new_key),
1862                                           BTREE_TRIGGER_INSERT);
1863                 if (ret)
1864                         return ret;
1865
1866                 ret = bch2_trans_mark_key(trans,
1867                                           bkey_i_to_s_c(&b->key),
1868                                           bkey_s_c_null,
1869                                           BTREE_TRIGGER_OVERWRITE);
1870                 if (ret)
1871                         return ret;
1872         }
1873
1874         if (new_hash) {
1875                 bkey_copy(&new_hash->key, new_key);
1876                 ret = bch2_btree_node_hash_insert(&c->btree_cache,
1877                                 new_hash, b->c.level, b->c.btree_id);
1878                 BUG_ON(ret);
1879         }
1880
1881         parent = btree_node_parent(iter->path, b);
1882         if (parent) {
1883                 bch2_trans_copy_iter(&iter2, iter);
1884
1885                 iter2.path = bch2_btree_path_make_mut(trans, iter2.path,
1886                                 iter2.flags & BTREE_ITER_INTENT);
1887
1888                 BUG_ON(iter2.path->level != b->c.level);
1889                 BUG_ON(bpos_cmp(iter2.path->pos, new_key->k.p));
1890
1891                 btree_node_unlock(iter2.path, iter2.path->level);
1892                 path_l(iter2.path)->b = BTREE_ITER_NO_NODE_UP;
1893                 iter2.path->level++;
1894
1895                 ret   = bch2_btree_iter_traverse(&iter2) ?:
1896                         bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_NORUN);
1897                 if (ret)
1898                         goto err;
1899         } else {
1900                 BUG_ON(btree_node_root(c, b) != b);
1901
1902                 trans->extra_journal_entries = (void *) &journal_entries[0];
1903                 trans->extra_journal_entry_u64s =
1904                         journal_entry_set((void *) &journal_entries[0],
1905                                           BCH_JSET_ENTRY_btree_root,
1906                                           b->c.btree_id, b->c.level,
1907                                           new_key, new_key->k.u64s);
1908         }
1909
1910         ret = bch2_trans_commit(trans, NULL, NULL,
1911                                 BTREE_INSERT_NOFAIL|
1912                                 BTREE_INSERT_NOCHECK_RW|
1913                                 BTREE_INSERT_JOURNAL_RECLAIM|
1914                                 BTREE_INSERT_JOURNAL_RESERVED);
1915         if (ret)
1916                 goto err;
1917
1918         bch2_btree_node_lock_write(trans, iter->path, b);
1919
1920         if (new_hash) {
1921                 mutex_lock(&c->btree_cache.lock);
1922                 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
1923                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1924
1925                 bkey_copy(&b->key, new_key);
1926                 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1927                 BUG_ON(ret);
1928                 mutex_unlock(&c->btree_cache.lock);
1929         } else {
1930                 bkey_copy(&b->key, new_key);
1931         }
1932
1933         bch2_btree_node_unlock_write(trans, iter->path, b);
1934 out:
1935         bch2_trans_iter_exit(trans, &iter2);
1936         return ret;
1937 err:
1938         if (new_hash) {
1939                 mutex_lock(&c->btree_cache.lock);
1940                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1941                 mutex_unlock(&c->btree_cache.lock);
1942         }
1943         goto out;
1944 }
1945
1946 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
1947                                struct btree *b, struct bkey_i *new_key,
1948                                bool skip_triggers)
1949 {
1950         struct bch_fs *c = trans->c;
1951         struct btree *new_hash = NULL;
1952         struct btree_path *path = iter->path;
1953         struct closure cl;
1954         int ret = 0;
1955
1956         if (!btree_node_intent_locked(path, b->c.level) &&
1957             !bch2_btree_path_upgrade(trans, path, b->c.level + 1)) {
1958                 btree_trans_restart(trans);
1959                 return -EINTR;
1960         }
1961
1962         closure_init_stack(&cl);
1963
1964         /*
1965          * check btree_ptr_hash_val() after @b is locked by
1966          * btree_iter_traverse():
1967          */
1968         if (btree_ptr_hash_val(new_key) != b->hash_val) {
1969                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1970                 if (ret) {
1971                         bch2_trans_unlock(trans);
1972                         closure_sync(&cl);
1973                         if (!bch2_trans_relock(trans))
1974                                 return -EINTR;
1975                 }
1976
1977                 new_hash = bch2_btree_node_mem_alloc(c);
1978         }
1979
1980         path->intent_ref++;
1981         ret = __bch2_btree_node_update_key(trans, iter, b, new_hash,
1982                                            new_key, skip_triggers);
1983         --path->intent_ref;
1984
1985         if (new_hash) {
1986                 mutex_lock(&c->btree_cache.lock);
1987                 list_move(&new_hash->list, &c->btree_cache.freeable);
1988                 mutex_unlock(&c->btree_cache.lock);
1989
1990                 six_unlock_write(&new_hash->c.lock);
1991                 six_unlock_intent(&new_hash->c.lock);
1992         }
1993         closure_sync(&cl);
1994         bch2_btree_cache_cannibalize_unlock(c);
1995         return ret;
1996 }
1997
1998 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
1999                                         struct btree *b, struct bkey_i *new_key,
2000                                         bool skip_triggers)
2001 {
2002         struct btree_iter iter;
2003         int ret;
2004
2005         bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p,
2006                                   BTREE_MAX_DEPTH, b->c.level,
2007                                   BTREE_ITER_INTENT);
2008         ret = bch2_btree_iter_traverse(&iter);
2009         if (ret)
2010                 goto out;
2011
2012         /* has node been freed? */
2013         if (iter.path->l[b->c.level].b != b) {
2014                 /* node has been freed: */
2015                 BUG_ON(!btree_node_dying(b));
2016                 goto out;
2017         }
2018
2019         BUG_ON(!btree_node_hashed(b));
2020
2021         ret = bch2_btree_node_update_key(trans, &iter, b, new_key, skip_triggers);
2022 out:
2023         bch2_trans_iter_exit(trans, &iter);
2024         return ret;
2025 }
2026
2027 /* Init code: */
2028
2029 /*
2030  * Only for filesystem bringup, when first reading the btree roots or allocating
2031  * btree roots when initializing a new filesystem:
2032  */
2033 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2034 {
2035         BUG_ON(btree_node_root(c, b));
2036
2037         bch2_btree_set_root_inmem(c, b);
2038 }
2039
2040 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2041 {
2042         struct closure cl;
2043         struct btree *b;
2044         int ret;
2045
2046         closure_init_stack(&cl);
2047
2048         do {
2049                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2050                 closure_sync(&cl);
2051         } while (ret);
2052
2053         b = bch2_btree_node_mem_alloc(c);
2054         bch2_btree_cache_cannibalize_unlock(c);
2055
2056         set_btree_node_fake(b);
2057         set_btree_node_need_rewrite(b);
2058         b->c.level      = 0;
2059         b->c.btree_id   = id;
2060
2061         bkey_btree_ptr_init(&b->key);
2062         b->key.k.p = SPOS_MAX;
2063         *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2064
2065         bch2_bset_init_first(b, &b->data->keys);
2066         bch2_btree_build_aux_trees(b);
2067
2068         b->data->flags = 0;
2069         btree_set_min(b, POS_MIN);
2070         btree_set_max(b, SPOS_MAX);
2071         b->data->format = bch2_btree_calc_format(b);
2072         btree_node_set_format(b, b->data->format);
2073
2074         ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2075                                           b->c.level, b->c.btree_id);
2076         BUG_ON(ret);
2077
2078         bch2_btree_set_root_inmem(c, b);
2079
2080         six_unlock_write(&b->c.lock);
2081         six_unlock_intent(&b->c.lock);
2082 }
2083
2084 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2085 {
2086         struct btree_update *as;
2087
2088         mutex_lock(&c->btree_interior_update_lock);
2089         list_for_each_entry(as, &c->btree_interior_update_list, list)
2090                 pr_buf(out, "%p m %u w %u r %u j %llu\n",
2091                        as,
2092                        as->mode,
2093                        as->nodes_written,
2094                        atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2095                        as->journal.seq);
2096         mutex_unlock(&c->btree_interior_update_lock);
2097 }
2098
2099 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *c)
2100 {
2101         size_t ret = 0;
2102         struct list_head *i;
2103
2104         mutex_lock(&c->btree_interior_update_lock);
2105         list_for_each(i, &c->btree_interior_update_list)
2106                 ret++;
2107         mutex_unlock(&c->btree_interior_update_lock);
2108
2109         return ret;
2110 }
2111
2112 void bch2_journal_entries_to_btree_roots(struct bch_fs *c, struct jset *jset)
2113 {
2114         struct btree_root *r;
2115         struct jset_entry *entry;
2116
2117         mutex_lock(&c->btree_root_lock);
2118
2119         vstruct_for_each(jset, entry)
2120                 if (entry->type == BCH_JSET_ENTRY_btree_root) {
2121                         r = &c->btree_roots[entry->btree_id];
2122                         r->level = entry->level;
2123                         r->alive = true;
2124                         bkey_copy(&r->key, &entry->start[0]);
2125                 }
2126
2127         mutex_unlock(&c->btree_root_lock);
2128 }
2129
2130 struct jset_entry *
2131 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2132                                     struct jset_entry *start,
2133                                     struct jset_entry *end)
2134 {
2135         struct jset_entry *entry;
2136         unsigned long have = 0;
2137         unsigned i;
2138
2139         for (entry = start; entry < end; entry = vstruct_next(entry))
2140                 if (entry->type == BCH_JSET_ENTRY_btree_root)
2141                         __set_bit(entry->btree_id, &have);
2142
2143         mutex_lock(&c->btree_root_lock);
2144
2145         for (i = 0; i < BTREE_ID_NR; i++)
2146                 if (c->btree_roots[i].alive && !test_bit(i, &have)) {
2147                         journal_entry_set(end,
2148                                           BCH_JSET_ENTRY_btree_root,
2149                                           i, c->btree_roots[i].level,
2150                                           &c->btree_roots[i].key,
2151                                           c->btree_roots[i].key.u64s);
2152                         end = vstruct_next(end);
2153                 }
2154
2155         mutex_unlock(&c->btree_root_lock);
2156
2157         return end;
2158 }
2159
2160 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2161 {
2162         if (c->btree_interior_update_worker)
2163                 destroy_workqueue(c->btree_interior_update_worker);
2164         mempool_exit(&c->btree_interior_update_pool);
2165 }
2166
2167 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2168 {
2169         mutex_init(&c->btree_reserve_cache_lock);
2170         INIT_LIST_HEAD(&c->btree_interior_update_list);
2171         INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2172         mutex_init(&c->btree_interior_update_lock);
2173         INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2174
2175         c->btree_interior_update_worker =
2176                 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2177         if (!c->btree_interior_update_worker)
2178                 return -ENOMEM;
2179
2180         return mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2181                                          sizeof(struct btree_update));
2182 }