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         char buf1[100], buf2[100];
  39
  40         BUG_ON(!b->c.level);
  41
  42         if (!test_bit(BCH_FS_BTREE_INTERIOR_REPLAY_DONE, &c->flags))
  43                 return;
  44
  45         bch2_btree_node_iter_init_from_start(&iter, b);
  46
  47         while (1) {
  48                 k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked);
  49                 if (k.k->type != KEY_TYPE_btree_ptr_v2)
  50                         break;
  51                 bp = bkey_s_c_to_btree_ptr_v2(k);
  52
  53                 if (bkey_cmp(next_node, bp.v->min_key)) {
  54                         bch2_dump_btree_node(c, b);
  55                         panic("expected next min_key %s got %s\n",
  56                               (bch2_bpos_to_text(&PBUF(buf1), next_node), buf1),
  57                               (bch2_bpos_to_text(&PBUF(buf2), bp.v->min_key), buf2));
  58                 }
  59
  60                 bch2_btree_node_iter_advance(&iter, b);
  61
  62                 if (bch2_btree_node_iter_end(&iter)) {
  63                         if (bkey_cmp(k.k->p, b->key.k.p)) {
  64                                 bch2_dump_btree_node(c, b);
  65                                 panic("expected end %s got %s\n",
  66                                       (bch2_bpos_to_text(&PBUF(buf1), b->key.k.p), buf1),
  67                                       (bch2_bpos_to_text(&PBUF(buf2), k.k->p), buf2));
  68                         }
  69                         break;
  70                 }
  71
  72                 next_node = bkey_successor(k.k->p);
  73         }
  74 #endif
  75 }
  76
  77 /* Calculate ideal packed bkey format for new btree nodes: */
  78
  79 void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b)
  80 {
  81         struct bkey_packed *k;
  82         struct bset_tree *t;
  83         struct bkey uk;
  84
  85         bch2_bkey_format_add_pos(s, b->data->min_key);
  86
  87         for_each_bset(b, t)
  88                 bset_tree_for_each_key(b, t, k)
  89                         if (!bkey_deleted(k)) {
  90                                 uk = bkey_unpack_key(b, k);
  91                                 bch2_bkey_format_add_key(s, &uk);
  92                         }
  93 }
  94
  95 static struct bkey_format bch2_btree_calc_format(struct btree *b)
  96 {
  97         struct bkey_format_state s;
  98
  99         bch2_bkey_format_init(&s);
 100         __bch2_btree_calc_format(&s, b);
 101
 102         return bch2_bkey_format_done(&s);
 103 }
 104
 105 static size_t btree_node_u64s_with_format(struct btree *b,
 106                                           struct bkey_format *new_f)
 107 {
 108         struct bkey_format *old_f = &b->format;
 109
 110         /* stupid integer promotion rules */
 111         ssize_t delta =
 112             (((int) new_f->key_u64s - old_f->key_u64s) *
 113              (int) b->nr.packed_keys) +
 114             (((int) new_f->key_u64s - BKEY_U64s) *
 115              (int) b->nr.unpacked_keys);
 116
 117         BUG_ON(delta + b->nr.live_u64s < 0);
 118
 119         return b->nr.live_u64s + delta;
 120 }
 121
 122 /**
 123  * btree_node_format_fits - check if we could rewrite node with a new format
 124  *
 125  * This assumes all keys can pack with the new format -- it just checks if
 126  * the re-packed keys would fit inside the node itself.
 127  */
 128 bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
 129                                  struct bkey_format *new_f)
 130 {
 131         size_t u64s = btree_node_u64s_with_format(b, new_f);
 132
 133         return __vstruct_bytes(struct btree_node, u64s) < btree_bytes(c);
 134 }
 135
 136 /* Btree node freeing/allocation: */
 137
 138 static void __btree_node_free(struct bch_fs *c, struct btree *b)
 139 {
 140         trace_btree_node_free(c, b);
 141
 142         BUG_ON(btree_node_dirty(b));
 143         BUG_ON(btree_node_need_write(b));
 144         BUG_ON(b == btree_node_root(c, b));
 145         BUG_ON(b->ob.nr);
 146         BUG_ON(!list_empty(&b->write_blocked));
 147         BUG_ON(b->will_make_reachable);
 148
 149         clear_btree_node_noevict(b);
 150
 151         bch2_btree_node_hash_remove(&c->btree_cache, b);
 152
 153         mutex_lock(&c->btree_cache.lock);
 154         list_move(&b->list, &c->btree_cache.freeable);
 155         mutex_unlock(&c->btree_cache.lock);
 156 }
 157
 158 void bch2_btree_node_free_never_inserted(struct bch_fs *c, struct btree *b)
 159 {
 160         struct open_buckets ob = b->ob;
 161
 162         b->ob.nr = 0;
 163
 164         clear_btree_node_dirty(c, b);
 165
 166         btree_node_lock_type(c, b, SIX_LOCK_write);
 167         __btree_node_free(c, b);
 168         six_unlock_write(&b->c.lock);
 169
 170         bch2_open_buckets_put(c, &ob);
 171 }
 172
 173 void bch2_btree_node_free_inmem(struct bch_fs *c, struct btree *b,
 174                                 struct btree_iter *iter)
 175 {
 176         struct btree_iter *linked;
 177
 178         trans_for_each_iter(iter->trans, linked)
 179                 BUG_ON(linked->l[b->c.level].b == b);
 180
 181         six_lock_write(&b->c.lock, NULL, NULL);
 182         __btree_node_free(c, b);
 183         six_unlock_write(&b->c.lock);
 184         six_unlock_intent(&b->c.lock);
 185 }
 186
 187 static struct btree *__bch2_btree_node_alloc(struct bch_fs *c,
 188                                              struct disk_reservation *res,
 189                                              struct closure *cl,
 190                                              unsigned flags)
 191 {
 192         struct write_point *wp;
 193         struct btree *b;
 194         __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
 195         struct open_buckets ob = { .nr = 0 };
 196         struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
 197         unsigned nr_reserve;
 198         enum alloc_reserve alloc_reserve;
 199
 200         if (flags & BTREE_INSERT_USE_RESERVE) {
 201                 nr_reserve      = 0;
 202                 alloc_reserve   = RESERVE_BTREE_MOVINGGC;
 203         } else {
 204                 nr_reserve      = BTREE_NODE_RESERVE;
 205                 alloc_reserve   = RESERVE_BTREE;
 206         }
 207
 208         mutex_lock(&c->btree_reserve_cache_lock);
 209         if (c->btree_reserve_cache_nr > nr_reserve) {
 210                 struct btree_alloc *a =
 211                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
 212
 213                 ob = a->ob;
 214                 bkey_copy(&tmp.k, &a->k);
 215                 mutex_unlock(&c->btree_reserve_cache_lock);
 216                 goto mem_alloc;
 217         }
 218         mutex_unlock(&c->btree_reserve_cache_lock);
 219
 220 retry:
 221         wp = bch2_alloc_sectors_start(c,
 222                                       c->opts.metadata_target ?:
 223                                       c->opts.foreground_target,
 224                                       0,
 225                                       writepoint_ptr(&c->btree_write_point),
 226                                       &devs_have,
 227                                       res->nr_replicas,
 228                                       c->opts.metadata_replicas_required,
 229                                       alloc_reserve, 0, cl);
 230         if (IS_ERR(wp))
 231                 return ERR_CAST(wp);
 232
 233         if (wp->sectors_free < c->opts.btree_node_size) {
 234                 struct open_bucket *ob;
 235                 unsigned i;
 236
 237                 open_bucket_for_each(c, &wp->ptrs, ob, i)
 238                         if (ob->sectors_free < c->opts.btree_node_size)
 239                                 ob->sectors_free = 0;
 240
 241                 bch2_alloc_sectors_done(c, wp);
 242                 goto retry;
 243         }
 244
 245         if (c->sb.features & (1ULL << BCH_FEATURE_btree_ptr_v2))
 246                 bkey_btree_ptr_v2_init(&tmp.k);
 247         else
 248                 bkey_btree_ptr_init(&tmp.k);
 249
 250         bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, c->opts.btree_node_size);
 251
 252         bch2_open_bucket_get(c, wp, &ob);
 253         bch2_alloc_sectors_done(c, wp);
 254 mem_alloc:
 255         b = bch2_btree_node_mem_alloc(c);
 256
 257         /* we hold cannibalize_lock: */
 258         BUG_ON(IS_ERR(b));
 259         BUG_ON(b->ob.nr);
 260
 261         bkey_copy(&b->key, &tmp.k);
 262         b->ob = ob;
 263
 264         return b;
 265 }
 266
 267 static struct btree *bch2_btree_node_alloc(struct btree_update *as, unsigned level)
 268 {
 269         struct bch_fs *c = as->c;
 270         struct btree *b;
 271         int ret;
 272
 273         BUG_ON(level >= BTREE_MAX_DEPTH);
 274         BUG_ON(!as->nr_prealloc_nodes);
 275
 276         b = as->prealloc_nodes[--as->nr_prealloc_nodes];
 277
 278         set_btree_node_accessed(b);
 279         set_btree_node_dirty(c, b);
 280         set_btree_node_need_write(b);
 281
 282         bch2_bset_init_first(b, &b->data->keys);
 283         b->c.level      = level;
 284         b->c.btree_id   = as->btree_id;
 285         b->version_ondisk = c->sb.version;
 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         }
 301
 302         SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
 303
 304         bch2_btree_build_aux_trees(b);
 305
 306         ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
 307         BUG_ON(ret);
 308
 309         trace_btree_node_alloc(c, b);
 310         return b;
 311 }
 312
 313 static void btree_set_min(struct btree *b, struct bpos pos)
 314 {
 315         if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
 316                 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
 317         b->data->min_key = pos;
 318 }
 319
 320 static void btree_set_max(struct btree *b, struct bpos pos)
 321 {
 322         b->key.k.p = pos;
 323         b->data->max_key = pos;
 324 }
 325
 326 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *as,
 327                                                   struct btree *b,
 328                                                   struct bkey_format format)
 329 {
 330         struct btree *n;
 331
 332         n = bch2_btree_node_alloc(as, b->c.level);
 333
 334         SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
 335
 336         btree_set_min(n, b->data->min_key);
 337         btree_set_max(n, b->data->max_key);
 338
 339         n->data->format         = format;
 340         btree_node_set_format(n, format);
 341
 342         bch2_btree_sort_into(as->c, n, b);
 343
 344         btree_node_reset_sib_u64s(n);
 345
 346         n->key.k.p = b->key.k.p;
 347         return n;
 348 }
 349
 350 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
 351                                                        struct btree *b)
 352 {
 353         struct bkey_format new_f = bch2_btree_calc_format(b);
 354
 355         /*
 356          * The keys might expand with the new format - if they wouldn't fit in
 357          * the btree node anymore, use the old format for now:
 358          */
 359         if (!bch2_btree_node_format_fits(as->c, b, &new_f))
 360                 new_f = b->format;
 361
 362         return __bch2_btree_node_alloc_replacement(as, b, new_f);
 363 }
 364
 365 static struct btree *__btree_root_alloc(struct btree_update *as, unsigned level)
 366 {
 367         struct btree *b = bch2_btree_node_alloc(as, level);
 368
 369         btree_set_min(b, POS_MIN);
 370         btree_set_max(b, POS_MAX);
 371         b->data->format = bch2_btree_calc_format(b);
 372
 373         btree_node_set_format(b, b->data->format);
 374         bch2_btree_build_aux_trees(b);
 375
 376         bch2_btree_update_add_new_node(as, b);
 377         six_unlock_write(&b->c.lock);
 378
 379         return b;
 380 }
 381
 382 static void bch2_btree_reserve_put(struct btree_update *as)
 383 {
 384         struct bch_fs *c = as->c;
 385
 386         mutex_lock(&c->btree_reserve_cache_lock);
 387
 388         while (as->nr_prealloc_nodes) {
 389                 struct btree *b = as->prealloc_nodes[--as->nr_prealloc_nodes];
 390
 391                 six_unlock_write(&b->c.lock);
 392
 393                 if (c->btree_reserve_cache_nr <
 394                     ARRAY_SIZE(c->btree_reserve_cache)) {
 395                         struct btree_alloc *a =
 396                                 &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
 397
 398                         a->ob = b->ob;
 399                         b->ob.nr = 0;
 400                         bkey_copy(&a->k, &b->key);
 401                 } else {
 402                         bch2_open_buckets_put(c, &b->ob);
 403                 }
 404
 405                 btree_node_lock_type(c, b, SIX_LOCK_write);
 406                 __btree_node_free(c, b);
 407                 six_unlock_write(&b->c.lock);
 408
 409                 six_unlock_intent(&b->c.lock);
 410         }
 411
 412         mutex_unlock(&c->btree_reserve_cache_lock);
 413 }
 414
 415 static int bch2_btree_reserve_get(struct btree_update *as, unsigned nr_nodes,
 416                                   unsigned flags, struct closure *cl)
 417 {
 418         struct bch_fs *c = as->c;
 419         struct btree *b;
 420         int ret;
 421
 422         BUG_ON(nr_nodes > BTREE_RESERVE_MAX);
 423
 424         /*
 425          * Protects reaping from the btree node cache and using the btree node
 426          * open bucket reserve:
 427          */
 428         ret = bch2_btree_cache_cannibalize_lock(c, cl);
 429         if (ret)
 430                 return ret;
 431
 432         while (as->nr_prealloc_nodes < nr_nodes) {
 433                 b = __bch2_btree_node_alloc(c, &as->disk_res,
 434                                             flags & BTREE_INSERT_NOWAIT
 435                                             ? NULL : cl, flags);
 436                 if (IS_ERR(b)) {
 437                         ret = PTR_ERR(b);
 438                         goto err_free;
 439                 }
 440
 441                 ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(&b->key));
 442                 if (ret)
 443                         goto err_free;
 444
 445                 as->prealloc_nodes[as->nr_prealloc_nodes++] = b;
 446         }
 447
 448         bch2_btree_cache_cannibalize_unlock(c);
 449         return 0;
 450 err_free:
 451         bch2_btree_cache_cannibalize_unlock(c);
 452         trace_btree_reserve_get_fail(c, nr_nodes, cl);
 453         return ret;
 454 }
 455
 456 /* Asynchronous interior node update machinery */
 457
 458 static void bch2_btree_update_free(struct btree_update *as)
 459 {
 460         struct bch_fs *c = as->c;
 461
 462         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 463
 464         bch2_journal_pin_drop(&c->journal, &as->journal);
 465         bch2_journal_pin_flush(&c->journal, &as->journal);
 466         bch2_disk_reservation_put(c, &as->disk_res);
 467         bch2_btree_reserve_put(as);
 468
 469         mutex_lock(&c->btree_interior_update_lock);
 470         list_del(&as->unwritten_list);
 471         list_del(&as->list);
 472         mutex_unlock(&c->btree_interior_update_lock);
 473
 474         closure_debug_destroy(&as->cl);
 475         mempool_free(as, &c->btree_interior_update_pool);
 476
 477         closure_wake_up(&c->btree_interior_update_wait);
 478 }
 479
 480 static void btree_update_will_delete_key(struct btree_update *as,
 481                                          struct bkey_i *k)
 482 {
 483         BUG_ON(bch2_keylist_u64s(&as->old_keys) + k->k.u64s >
 484                ARRAY_SIZE(as->_old_keys));
 485         bch2_keylist_add(&as->old_keys, k);
 486 }
 487
 488 static void btree_update_will_add_key(struct btree_update *as,
 489                                       struct bkey_i *k)
 490 {
 491         BUG_ON(bch2_keylist_u64s(&as->new_keys) + k->k.u64s >
 492                ARRAY_SIZE(as->_new_keys));
 493         bch2_keylist_add(&as->new_keys, k);
 494 }
 495
 496 /*
 497  * The transactional part of an interior btree node update, where we journal the
 498  * update we did to the interior node and update alloc info:
 499  */
 500 static int btree_update_nodes_written_trans(struct btree_trans *trans,
 501                                             struct btree_update *as)
 502 {
 503         struct bkey_i *k;
 504         int ret;
 505
 506         trans->extra_journal_entries = (void *) &as->journal_entries[0];
 507         trans->extra_journal_entry_u64s = as->journal_u64s;
 508         trans->journal_pin = &as->journal;
 509
 510         for_each_keylist_key(&as->new_keys, k) {
 511                 ret = bch2_trans_mark_key(trans,
 512                                           bkey_s_c_null,
 513                                           bkey_i_to_s_c(k),
 514                                           0, 0, BTREE_TRIGGER_INSERT);
 515                 if (ret)
 516                         return ret;
 517         }
 518
 519         for_each_keylist_key(&as->old_keys, k) {
 520                 ret = bch2_trans_mark_key(trans,
 521                                           bkey_i_to_s_c(k),
 522                                           bkey_s_c_null,
 523                                           0, 0, BTREE_TRIGGER_OVERWRITE);
 524                 if (ret)
 525                         return ret;
 526         }
 527
 528         return 0;
 529 }
 530
 531 static void btree_update_nodes_written(struct btree_update *as)
 532 {
 533         struct bch_fs *c = as->c;
 534         struct btree *b = as->b;
 535         struct btree_trans trans;
 536         u64 journal_seq = 0;
 537         unsigned i;
 538         int ret;
 539
 540         /*
 541          * If we're already in an error state, it might be because a btree node
 542          * was never written, and we might be trying to free that same btree
 543          * node here, but it won't have been marked as allocated and we'll see
 544          * spurious disk usage inconsistencies in the transactional part below
 545          * if we don't skip it:
 546          */
 547         ret = bch2_journal_error(&c->journal);
 548         if (ret)
 549                 goto err;
 550
 551         BUG_ON(!journal_pin_active(&as->journal));
 552
 553         /*
 554          * We did an update to a parent node where the pointers we added pointed
 555          * to child nodes that weren't written yet: now, the child nodes have
 556          * been written so we can write out the update to the interior node.
 557          */
 558
 559         /*
 560          * We can't call into journal reclaim here: we'd block on the journal
 561          * reclaim lock, but we may need to release the open buckets we have
 562          * pinned in order for other btree updates to make forward progress, and
 563          * journal reclaim does btree updates when flushing bkey_cached entries,
 564          * which may require allocations as well.
 565          */
 566         bch2_trans_init(&trans, c, 0, 512);
 567         ret = __bch2_trans_do(&trans, &as->disk_res, &journal_seq,
 568                               BTREE_INSERT_NOFAIL|
 569                               BTREE_INSERT_NOCHECK_RW|
 570                               BTREE_INSERT_JOURNAL_RECLAIM|
 571                               BTREE_INSERT_JOURNAL_RESERVED,
 572                               btree_update_nodes_written_trans(&trans, as));
 573         bch2_trans_exit(&trans);
 574
 575         bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
 576                              "error %i in btree_update_nodes_written()", ret);
 577 err:
 578         if (b) {
 579                 /*
 580                  * @b is the node we did the final insert into:
 581                  *
 582                  * On failure to get a journal reservation, we still have to
 583                  * unblock the write and allow most of the write path to happen
 584                  * so that shutdown works, but the i->journal_seq mechanism
 585                  * won't work to prevent the btree write from being visible (we
 586                  * didn't get a journal sequence number) - instead
 587                  * __bch2_btree_node_write() doesn't do the actual write if
 588                  * we're in journal error state:
 589                  */
 590
 591                 btree_node_lock_type(c, b, SIX_LOCK_intent);
 592                 btree_node_lock_type(c, b, SIX_LOCK_write);
 593                 mutex_lock(&c->btree_interior_update_lock);
 594
 595                 list_del(&as->write_blocked_list);
 596
 597                 /*
 598                  * Node might have been freed, recheck under
 599                  * btree_interior_update_lock:
 600                  */
 601                 if (as->b == b) {
 602                         struct bset *i = btree_bset_last(b);
 603
 604                         BUG_ON(!b->c.level);
 605                         BUG_ON(!btree_node_dirty(b));
 606
 607                         if (!ret) {
 608                                 i->journal_seq = cpu_to_le64(
 609                                         max(journal_seq,
 610                                             le64_to_cpu(i->journal_seq)));
 611
 612                                 bch2_btree_add_journal_pin(c, b, journal_seq);
 613                         } else {
 614                                 /*
 615                                  * If we didn't get a journal sequence number we
 616                                  * can't write this btree node, because recovery
 617                                  * won't know to ignore this write:
 618                                  */
 619                                 set_btree_node_never_write(b);
 620                         }
 621                 }
 622
 623                 mutex_unlock(&c->btree_interior_update_lock);
 624                 six_unlock_write(&b->c.lock);
 625
 626                 btree_node_write_if_need(c, b, SIX_LOCK_intent);
 627                 six_unlock_intent(&b->c.lock);
 628         }
 629
 630         bch2_journal_pin_drop(&c->journal, &as->journal);
 631
 632         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 633
 634         mutex_lock(&c->btree_interior_update_lock);
 635         for (i = 0; i < as->nr_new_nodes; i++) {
 636                 b = as->new_nodes[i];
 637
 638                 BUG_ON(b->will_make_reachable != (unsigned long) as);
 639                 b->will_make_reachable = 0;
 640         }
 641         mutex_unlock(&c->btree_interior_update_lock);
 642
 643         for (i = 0; i < as->nr_new_nodes; i++) {
 644                 b = as->new_nodes[i];
 645
 646                 btree_node_lock_type(c, b, SIX_LOCK_read);
 647                 btree_node_write_if_need(c, b, SIX_LOCK_read);
 648                 six_unlock_read(&b->c.lock);
 649         }
 650
 651         for (i = 0; i < as->nr_open_buckets; i++)
 652                 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
 653
 654         bch2_btree_update_free(as);
 655 }
 656
 657 static void btree_interior_update_work(struct work_struct *work)
 658 {
 659         struct bch_fs *c =
 660                 container_of(work, struct bch_fs, btree_interior_update_work);
 661         struct btree_update *as;
 662
 663         while (1) {
 664                 mutex_lock(&c->btree_interior_update_lock);
 665                 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
 666                                               struct btree_update, unwritten_list);
 667                 if (as && !as->nodes_written)
 668                         as = NULL;
 669                 mutex_unlock(&c->btree_interior_update_lock);
 670
 671                 if (!as)
 672                         break;
 673
 674                 btree_update_nodes_written(as);
 675         }
 676 }
 677
 678 static void btree_update_set_nodes_written(struct closure *cl)
 679 {
 680         struct btree_update *as = container_of(cl, struct btree_update, cl);
 681         struct bch_fs *c = as->c;
 682
 683         mutex_lock(&c->btree_interior_update_lock);
 684         as->nodes_written = true;
 685         mutex_unlock(&c->btree_interior_update_lock);
 686
 687         queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
 688 }
 689
 690 /*
 691  * We're updating @b with pointers to nodes that haven't finished writing yet:
 692  * block @b from being written until @as completes
 693  */
 694 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
 695 {
 696         struct bch_fs *c = as->c;
 697
 698         mutex_lock(&c->btree_interior_update_lock);
 699         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 700
 701         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 702         BUG_ON(!btree_node_dirty(b));
 703
 704         as->mode        = BTREE_INTERIOR_UPDATING_NODE;
 705         as->b           = b;
 706         list_add(&as->write_blocked_list, &b->write_blocked);
 707
 708         mutex_unlock(&c->btree_interior_update_lock);
 709 }
 710
 711 static void btree_update_reparent(struct btree_update *as,
 712                                   struct btree_update *child)
 713 {
 714         struct bch_fs *c = as->c;
 715
 716         lockdep_assert_held(&c->btree_interior_update_lock);
 717
 718         child->b = NULL;
 719         child->mode = BTREE_INTERIOR_UPDATING_AS;
 720
 721         bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL);
 722 }
 723
 724 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
 725 {
 726         struct bkey_i *insert = &b->key;
 727         struct bch_fs *c = as->c;
 728
 729         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 730
 731         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
 732                ARRAY_SIZE(as->journal_entries));
 733
 734         as->journal_u64s +=
 735                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
 736                                   BCH_JSET_ENTRY_btree_root,
 737                                   b->c.btree_id, b->c.level,
 738                                   insert, insert->k.u64s);
 739
 740         mutex_lock(&c->btree_interior_update_lock);
 741         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 742
 743         as->mode        = BTREE_INTERIOR_UPDATING_ROOT;
 744         mutex_unlock(&c->btree_interior_update_lock);
 745 }
 746
 747 /*
 748  * bch2_btree_update_add_new_node:
 749  *
 750  * This causes @as to wait on @b to be written, before it gets to
 751  * bch2_btree_update_nodes_written
 752  *
 753  * Additionally, it sets b->will_make_reachable to prevent any additional writes
 754  * to @b from happening besides the first until @b is reachable on disk
 755  *
 756  * And it adds @b to the list of @as's new nodes, so that we can update sector
 757  * counts in bch2_btree_update_nodes_written:
 758  */
 759 void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
 760 {
 761         struct bch_fs *c = as->c;
 762
 763         closure_get(&as->cl);
 764
 765         mutex_lock(&c->btree_interior_update_lock);
 766         BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
 767         BUG_ON(b->will_make_reachable);
 768
 769         as->new_nodes[as->nr_new_nodes++] = b;
 770         b->will_make_reachable = 1UL|(unsigned long) as;
 771
 772         mutex_unlock(&c->btree_interior_update_lock);
 773
 774         btree_update_will_add_key(as, &b->key);
 775 }
 776
 777 /*
 778  * returns true if @b was a new node
 779  */
 780 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
 781 {
 782         struct btree_update *as;
 783         unsigned long v;
 784         unsigned i;
 785
 786         mutex_lock(&c->btree_interior_update_lock);
 787         /*
 788          * When b->will_make_reachable != 0, it owns a ref on as->cl that's
 789          * dropped when it gets written by bch2_btree_complete_write - the
 790          * xchg() is for synchronization with bch2_btree_complete_write:
 791          */
 792         v = xchg(&b->will_make_reachable, 0);
 793         as = (struct btree_update *) (v & ~1UL);
 794
 795         if (!as) {
 796                 mutex_unlock(&c->btree_interior_update_lock);
 797                 return;
 798         }
 799
 800         for (i = 0; i < as->nr_new_nodes; i++)
 801                 if (as->new_nodes[i] == b)
 802                         goto found;
 803
 804         BUG();
 805 found:
 806         array_remove_item(as->new_nodes, as->nr_new_nodes, i);
 807         mutex_unlock(&c->btree_interior_update_lock);
 808
 809         if (v & 1)
 810                 closure_put(&as->cl);
 811 }
 812
 813 void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
 814 {
 815         while (b->ob.nr)
 816                 as->open_buckets[as->nr_open_buckets++] =
 817                         b->ob.v[--b->ob.nr];
 818 }
 819
 820 /*
 821  * @b is being split/rewritten: it may have pointers to not-yet-written btree
 822  * nodes and thus outstanding btree_updates - redirect @b's
 823  * btree_updates to point to this btree_update:
 824  */
 825 void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 826                                                struct btree *b)
 827 {
 828         struct bch_fs *c = as->c;
 829         struct btree_update *p, *n;
 830         struct btree_write *w;
 831
 832         set_btree_node_dying(b);
 833
 834         if (btree_node_fake(b))
 835                 return;
 836
 837         mutex_lock(&c->btree_interior_update_lock);
 838
 839         /*
 840          * Does this node have any btree_update operations preventing
 841          * it from being written?
 842          *
 843          * If so, redirect them to point to this btree_update: we can
 844          * write out our new nodes, but we won't make them visible until those
 845          * operations complete
 846          */
 847         list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
 848                 list_del_init(&p->write_blocked_list);
 849                 btree_update_reparent(as, p);
 850
 851                 /*
 852                  * for flush_held_btree_writes() waiting on updates to flush or
 853                  * nodes to be writeable:
 854                  */
 855                 closure_wake_up(&c->btree_interior_update_wait);
 856         }
 857
 858         clear_btree_node_dirty(c, b);
 859         clear_btree_node_need_write(b);
 860
 861         /*
 862          * Does this node have unwritten data that has a pin on the journal?
 863          *
 864          * If so, transfer that pin to the btree_update operation -
 865          * note that if we're freeing multiple nodes, we only need to keep the
 866          * oldest pin of any of the nodes we're freeing. We'll release the pin
 867          * when the new nodes are persistent and reachable on disk:
 868          */
 869         w = btree_current_write(b);
 870         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 871         bch2_journal_pin_drop(&c->journal, &w->journal);
 872
 873         w = btree_prev_write(b);
 874         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 875         bch2_journal_pin_drop(&c->journal, &w->journal);
 876
 877         mutex_unlock(&c->btree_interior_update_lock);
 878
 879         /*
 880          * Is this a node that isn't reachable on disk yet?
 881          *
 882          * Nodes that aren't reachable yet have writes blocked until they're
 883          * reachable - now that we've cancelled any pending writes and moved
 884          * things waiting on that write to wait on this update, we can drop this
 885          * node from the list of nodes that the other update is making
 886          * reachable, prior to freeing it:
 887          */
 888         btree_update_drop_new_node(c, b);
 889
 890         btree_update_will_delete_key(as, &b->key);
 891 }
 892
 893 void bch2_btree_update_done(struct btree_update *as)
 894 {
 895         BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
 896
 897         bch2_btree_reserve_put(as);
 898
 899         continue_at(&as->cl, btree_update_set_nodes_written, system_freezable_wq);
 900 }
 901
 902 struct btree_update *
 903 bch2_btree_update_start(struct btree_trans *trans, enum btree_id id,
 904                         unsigned nr_nodes, unsigned flags,
 905                         struct closure *cl)
 906 {
 907         struct bch_fs *c = trans->c;
 908         struct btree_update *as;
 909         int disk_res_flags = (flags & BTREE_INSERT_NOFAIL)
 910                 ? BCH_DISK_RESERVATION_NOFAIL : 0;
 911         int journal_flags = (flags & BTREE_INSERT_JOURNAL_RESERVED)
 912                 ? JOURNAL_RES_GET_RECLAIM : 0;
 913         int ret = 0;
 914
 915         /*
 916          * This check isn't necessary for correctness - it's just to potentially
 917          * prevent us from doing a lot of work that'll end up being wasted:
 918          */
 919         ret = bch2_journal_error(&c->journal);
 920         if (ret)
 921                 return ERR_PTR(ret);
 922
 923         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
 924         memset(as, 0, sizeof(*as));
 925         closure_init(&as->cl, NULL);
 926         as->c           = c;
 927         as->mode        = BTREE_INTERIOR_NO_UPDATE;
 928         as->btree_id    = id;
 929         INIT_LIST_HEAD(&as->list);
 930         INIT_LIST_HEAD(&as->unwritten_list);
 931         INIT_LIST_HEAD(&as->write_blocked_list);
 932         bch2_keylist_init(&as->old_keys, as->_old_keys);
 933         bch2_keylist_init(&as->new_keys, as->_new_keys);
 934         bch2_keylist_init(&as->parent_keys, as->inline_keys);
 935
 936         ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
 937                                       BTREE_UPDATE_JOURNAL_RES,
 938                                       journal_flags|JOURNAL_RES_GET_NONBLOCK);
 939         if (ret == -EAGAIN) {
 940                 if (flags & BTREE_INSERT_NOUNLOCK)
 941                         return ERR_PTR(-EINTR);
 942
 943                 bch2_trans_unlock(trans);
 944
 945                 ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
 946                                 BTREE_UPDATE_JOURNAL_RES,
 947                                 journal_flags);
 948                 if (ret)
 949                         return ERR_PTR(ret);
 950
 951                 if (!bch2_trans_relock(trans)) {
 952                         ret = -EINTR;
 953                         goto err;
 954                 }
 955         }
 956
 957         ret = bch2_disk_reservation_get(c, &as->disk_res,
 958                         nr_nodes * c->opts.btree_node_size,
 959                         c->opts.metadata_replicas,
 960                         disk_res_flags);
 961         if (ret)
 962                 goto err;
 963
 964         ret = bch2_btree_reserve_get(as, nr_nodes, flags, cl);
 965         if (ret)
 966                 goto err;
 967
 968         bch2_journal_pin_add(&c->journal,
 969                              atomic64_read(&c->journal.seq),
 970                              &as->journal, NULL);
 971
 972         mutex_lock(&c->btree_interior_update_lock);
 973         list_add_tail(&as->list, &c->btree_interior_update_list);
 974         mutex_unlock(&c->btree_interior_update_lock);
 975
 976         return as;
 977 err:
 978         bch2_btree_update_free(as);
 979         return ERR_PTR(ret);
 980 }
 981
 982 /* Btree root updates: */
 983
 984 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
 985 {
 986         /* Root nodes cannot be reaped */
 987         mutex_lock(&c->btree_cache.lock);
 988         list_del_init(&b->list);
 989         mutex_unlock(&c->btree_cache.lock);
 990
 991         mutex_lock(&c->btree_root_lock);
 992         BUG_ON(btree_node_root(c, b) &&
 993                (b->c.level < btree_node_root(c, b)->c.level ||
 994                 !btree_node_dying(btree_node_root(c, b))));
 995
 996         btree_node_root(c, b) = b;
 997         mutex_unlock(&c->btree_root_lock);
 998
 999         bch2_recalc_btree_reserve(c);
1000 }
1001
1002 /**
1003  * bch_btree_set_root - update the root in memory and on disk
1004  *
1005  * To ensure forward progress, the current task must not be holding any
1006  * btree node write locks. However, you must hold an intent lock on the
1007  * old root.
1008  *
1009  * Note: This allocates a journal entry but doesn't add any keys to
1010  * it.  All the btree roots are part of every journal write, so there
1011  * is nothing new to be done.  This just guarantees that there is a
1012  * journal write.
1013  */
1014 static void bch2_btree_set_root(struct btree_update *as, struct btree *b,
1015                                 struct btree_iter *iter)
1016 {
1017         struct bch_fs *c = as->c;
1018         struct btree *old;
1019
1020         trace_btree_set_root(c, b);
1021         BUG_ON(!b->written &&
1022                !test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags));
1023
1024         old = btree_node_root(c, b);
1025
1026         /*
1027          * Ensure no one is using the old root while we switch to the
1028          * new root:
1029          */
1030         bch2_btree_node_lock_write(old, iter);
1031
1032         bch2_btree_set_root_inmem(c, b);
1033
1034         btree_update_updated_root(as, b);
1035
1036         /*
1037          * Unlock old root after new root is visible:
1038          *
1039          * The new root isn't persistent, but that's ok: we still have
1040          * an intent lock on the new root, and any updates that would
1041          * depend on the new root would have to update the new root.
1042          */
1043         bch2_btree_node_unlock_write(old, iter);
1044 }
1045
1046 /* Interior node updates: */
1047
1048 static void bch2_insert_fixup_btree_ptr(struct btree_update *as, struct btree *b,
1049                                         struct btree_iter *iter,
1050                                         struct bkey_i *insert,
1051                                         struct btree_node_iter *node_iter)
1052 {
1053         struct bch_fs *c = as->c;
1054         struct bkey_packed *k;
1055         const char *invalid;
1056
1057         invalid = bch2_bkey_invalid(c, bkey_i_to_s_c(insert), btree_node_type(b)) ?:
1058                 bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert));
1059         if (invalid) {
1060                 char buf[160];
1061
1062                 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(insert));
1063                 bch2_fs_inconsistent(c, "inserting invalid bkey %s: %s", buf, invalid);
1064                 dump_stack();
1065         }
1066
1067         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1068                ARRAY_SIZE(as->journal_entries));
1069
1070         as->journal_u64s +=
1071                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1072                                   BCH_JSET_ENTRY_btree_keys,
1073                                   b->c.btree_id, b->c.level,
1074                                   insert, insert->k.u64s);
1075
1076         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1077                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1078                 bch2_btree_node_iter_advance(node_iter, b);
1079
1080         bch2_btree_bset_insert_key(iter, b, node_iter, insert);
1081         set_btree_node_dirty(c, b);
1082         set_btree_node_need_write(b);
1083 }
1084
1085 /*
1086  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1087  * node)
1088  */
1089 static struct btree *__btree_split_node(struct btree_update *as,
1090                                         struct btree *n1,
1091                                         struct btree_iter *iter)
1092 {
1093         size_t nr_packed = 0, nr_unpacked = 0;
1094         struct btree *n2;
1095         struct bset *set1, *set2;
1096         struct bkey_packed *k, *prev = NULL;
1097
1098         n2 = bch2_btree_node_alloc(as, n1->c.level);
1099         bch2_btree_update_add_new_node(as, n2);
1100
1101         n2->data->max_key       = n1->data->max_key;
1102         n2->data->format        = n1->format;
1103         SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data));
1104         n2->key.k.p = n1->key.k.p;
1105
1106         btree_node_set_format(n2, n2->data->format);
1107
1108         set1 = btree_bset_first(n1);
1109         set2 = btree_bset_first(n2);
1110
1111         /*
1112          * Has to be a linear search because we don't have an auxiliary
1113          * search tree yet
1114          */
1115         k = set1->start;
1116         while (1) {
1117                 struct bkey_packed *n = bkey_next_skip_noops(k, vstruct_last(set1));
1118
1119                 if (n == vstruct_last(set1))
1120                         break;
1121                 if (k->_data - set1->_data >= (le16_to_cpu(set1->u64s) * 3) / 5)
1122                         break;
1123
1124                 if (bkey_packed(k))
1125                         nr_packed++;
1126                 else
1127                         nr_unpacked++;
1128
1129                 prev = k;
1130                 k = n;
1131         }
1132
1133         BUG_ON(!prev);
1134
1135         btree_set_max(n1, bkey_unpack_pos(n1, prev));
1136         btree_set_min(n2, bkey_successor(n1->key.k.p));
1137
1138         set2->u64s = cpu_to_le16((u64 *) vstruct_end(set1) - (u64 *) k);
1139         set1->u64s = cpu_to_le16(le16_to_cpu(set1->u64s) - le16_to_cpu(set2->u64s));
1140
1141         set_btree_bset_end(n1, n1->set);
1142         set_btree_bset_end(n2, n2->set);
1143
1144         n2->nr.live_u64s        = le16_to_cpu(set2->u64s);
1145         n2->nr.bset_u64s[0]     = le16_to_cpu(set2->u64s);
1146         n2->nr.packed_keys      = n1->nr.packed_keys - nr_packed;
1147         n2->nr.unpacked_keys    = n1->nr.unpacked_keys - nr_unpacked;
1148
1149         n1->nr.live_u64s        = le16_to_cpu(set1->u64s);
1150         n1->nr.bset_u64s[0]     = le16_to_cpu(set1->u64s);
1151         n1->nr.packed_keys      = nr_packed;
1152         n1->nr.unpacked_keys    = nr_unpacked;
1153
1154         BUG_ON(!set1->u64s);
1155         BUG_ON(!set2->u64s);
1156
1157         memcpy_u64s(set2->start,
1158                     vstruct_end(set1),
1159                     le16_to_cpu(set2->u64s));
1160
1161         btree_node_reset_sib_u64s(n1);
1162         btree_node_reset_sib_u64s(n2);
1163
1164         bch2_verify_btree_nr_keys(n1);
1165         bch2_verify_btree_nr_keys(n2);
1166
1167         if (n1->c.level) {
1168                 btree_node_interior_verify(as->c, n1);
1169                 btree_node_interior_verify(as->c, n2);
1170         }
1171
1172         return n2;
1173 }
1174
1175 /*
1176  * For updates to interior nodes, we've got to do the insert before we split
1177  * because the stuff we're inserting has to be inserted atomically. Post split,
1178  * the keys might have to go in different nodes and the split would no longer be
1179  * atomic.
1180  *
1181  * Worse, if the insert is from btree node coalescing, if we do the insert after
1182  * we do the split (and pick the pivot) - the pivot we pick might be between
1183  * nodes that were coalesced, and thus in the middle of a child node post
1184  * coalescing:
1185  */
1186 static void btree_split_insert_keys(struct btree_update *as, struct btree *b,
1187                                     struct btree_iter *iter,
1188                                     struct keylist *keys)
1189 {
1190         struct btree_node_iter node_iter;
1191         struct bkey_i *k = bch2_keylist_front(keys);
1192         struct bkey_packed *src, *dst, *n;
1193         struct bset *i;
1194
1195         BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1196
1197         bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
1198
1199         while (!bch2_keylist_empty(keys)) {
1200                 k = bch2_keylist_front(keys);
1201
1202                 bch2_insert_fixup_btree_ptr(as, b, iter, k, &node_iter);
1203                 bch2_keylist_pop_front(keys);
1204         }
1205
1206         /*
1207          * We can't tolerate whiteouts here - with whiteouts there can be
1208          * duplicate keys, and it would be rather bad if we picked a duplicate
1209          * for the pivot:
1210          */
1211         i = btree_bset_first(b);
1212         src = dst = i->start;
1213         while (src != vstruct_last(i)) {
1214                 n = bkey_next_skip_noops(src, vstruct_last(i));
1215                 if (!bkey_deleted(src)) {
1216                         memmove_u64s_down(dst, src, src->u64s);
1217                         dst = bkey_next(dst);
1218                 }
1219                 src = n;
1220         }
1221
1222         /* Also clear out the unwritten whiteouts area: */
1223         b->whiteout_u64s = 0;
1224
1225         i->u64s = cpu_to_le16((u64 *) dst - i->_data);
1226         set_btree_bset_end(b, b->set);
1227
1228         BUG_ON(b->nsets != 1 ||
1229                b->nr.live_u64s != le16_to_cpu(btree_bset_first(b)->u64s));
1230
1231         btree_node_interior_verify(as->c, b);
1232 }
1233
1234 static void btree_split(struct btree_update *as, struct btree *b,
1235                         struct btree_iter *iter, struct keylist *keys,
1236                         unsigned flags)
1237 {
1238         struct bch_fs *c = as->c;
1239         struct btree *parent = btree_node_parent(iter, b);
1240         struct btree *n1, *n2 = NULL, *n3 = NULL;
1241         u64 start_time = local_clock();
1242
1243         BUG_ON(!parent && (b != btree_node_root(c, b)));
1244         BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->c.level));
1245
1246         bch2_btree_interior_update_will_free_node(as, b);
1247
1248         n1 = bch2_btree_node_alloc_replacement(as, b);
1249         bch2_btree_update_add_new_node(as, n1);
1250
1251         if (keys)
1252                 btree_split_insert_keys(as, n1, iter, keys);
1253
1254         if (bset_u64s(&n1->set[0]) > BTREE_SPLIT_THRESHOLD(c)) {
1255                 trace_btree_split(c, b);
1256
1257                 n2 = __btree_split_node(as, n1, iter);
1258
1259                 bch2_btree_build_aux_trees(n2);
1260                 bch2_btree_build_aux_trees(n1);
1261                 six_unlock_write(&n2->c.lock);
1262                 six_unlock_write(&n1->c.lock);
1263
1264                 bch2_btree_node_write(c, n2, SIX_LOCK_intent);
1265
1266                 /*
1267                  * Note that on recursive parent_keys == keys, so we
1268                  * can't start adding new keys to parent_keys before emptying it
1269                  * out (which we did with btree_split_insert_keys() above)
1270                  */
1271                 bch2_keylist_add(&as->parent_keys, &n1->key);
1272                 bch2_keylist_add(&as->parent_keys, &n2->key);
1273
1274                 if (!parent) {
1275                         /* Depth increases, make a new root */
1276                         n3 = __btree_root_alloc(as, b->c.level + 1);
1277
1278                         n3->sib_u64s[0] = U16_MAX;
1279                         n3->sib_u64s[1] = U16_MAX;
1280
1281                         btree_split_insert_keys(as, n3, iter, &as->parent_keys);
1282
1283                         bch2_btree_node_write(c, n3, SIX_LOCK_intent);
1284                 }
1285         } else {
1286                 trace_btree_compact(c, b);
1287
1288                 bch2_btree_build_aux_trees(n1);
1289                 six_unlock_write(&n1->c.lock);
1290
1291                 if (parent)
1292                         bch2_keylist_add(&as->parent_keys, &n1->key);
1293         }
1294
1295         bch2_btree_node_write(c, n1, SIX_LOCK_intent);
1296
1297         /* New nodes all written, now make them visible: */
1298
1299         if (parent) {
1300                 /* Split a non root node */
1301                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1302         } else if (n3) {
1303                 bch2_btree_set_root(as, n3, iter);
1304         } else {
1305                 /* Root filled up but didn't need to be split */
1306                 bch2_btree_set_root(as, n1, iter);
1307         }
1308
1309         bch2_btree_update_get_open_buckets(as, n1);
1310         if (n2)
1311                 bch2_btree_update_get_open_buckets(as, n2);
1312         if (n3)
1313                 bch2_btree_update_get_open_buckets(as, n3);
1314
1315         /* Successful split, update the iterator to point to the new nodes: */
1316
1317         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1318         bch2_btree_iter_node_drop(iter, b);
1319         if (n3)
1320                 bch2_btree_iter_node_replace(iter, n3);
1321         if (n2)
1322                 bch2_btree_iter_node_replace(iter, n2);
1323         bch2_btree_iter_node_replace(iter, n1);
1324
1325         /*
1326          * The old node must be freed (in memory) _before_ unlocking the new
1327          * nodes - else another thread could re-acquire a read lock on the old
1328          * node after another thread has locked and updated the new node, thus
1329          * seeing stale data:
1330          */
1331         bch2_btree_node_free_inmem(c, b, iter);
1332
1333         if (n3)
1334                 six_unlock_intent(&n3->c.lock);
1335         if (n2)
1336                 six_unlock_intent(&n2->c.lock);
1337         six_unlock_intent(&n1->c.lock);
1338
1339         bch2_btree_trans_verify_locks(iter->trans);
1340
1341         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_split],
1342                                start_time);
1343 }
1344
1345 static void
1346 bch2_btree_insert_keys_interior(struct btree_update *as, struct btree *b,
1347                                 struct btree_iter *iter, struct keylist *keys)
1348 {
1349         struct btree_iter *linked;
1350         struct btree_node_iter node_iter;
1351         struct bkey_i *insert = bch2_keylist_front(keys);
1352         struct bkey_packed *k;
1353
1354         /* Don't screw up @iter's position: */
1355         node_iter = iter->l[b->c.level].iter;
1356
1357         /*
1358          * btree_split(), btree_gc_coalesce() will insert keys before
1359          * the iterator's current position - they know the keys go in
1360          * the node the iterator points to:
1361          */
1362         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1363                (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1364                 ;
1365
1366         for_each_keylist_key(keys, insert)
1367                 bch2_insert_fixup_btree_ptr(as, b, iter, insert, &node_iter);
1368
1369         btree_update_updated_node(as, b);
1370
1371         trans_for_each_iter_with_node(iter->trans, b, linked)
1372                 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1373
1374         bch2_btree_trans_verify_iters(iter->trans, b);
1375 }
1376
1377 /**
1378  * bch_btree_insert_node - insert bkeys into a given btree node
1379  *
1380  * @iter:               btree iterator
1381  * @keys:               list of keys to insert
1382  * @hook:               insert callback
1383  * @persistent:         if not null, @persistent will wait on journal write
1384  *
1385  * Inserts as many keys as it can into a given btree node, splitting it if full.
1386  * If a split occurred, this function will return early. This can only happen
1387  * for leaf nodes -- inserts into interior nodes have to be atomic.
1388  */
1389 void bch2_btree_insert_node(struct btree_update *as, struct btree *b,
1390                             struct btree_iter *iter, struct keylist *keys,
1391                             unsigned flags)
1392 {
1393         struct bch_fs *c = as->c;
1394         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1395         int old_live_u64s = b->nr.live_u64s;
1396         int live_u64s_added, u64s_added;
1397
1398         BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->c.level));
1399         BUG_ON(!b->c.level);
1400         BUG_ON(!as || as->b);
1401         bch2_verify_keylist_sorted(keys);
1402
1403         bch2_btree_node_lock_for_insert(c, b, iter);
1404
1405         if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1406                 bch2_btree_node_unlock_write(b, iter);
1407                 goto split;
1408         }
1409
1410         btree_node_interior_verify(c, b);
1411
1412         bch2_btree_insert_keys_interior(as, b, iter, keys);
1413
1414         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1415         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1416
1417         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1418                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1419         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1420                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1421
1422         if (u64s_added > live_u64s_added &&
1423             bch2_maybe_compact_whiteouts(c, b))
1424                 bch2_btree_iter_reinit_node(iter, b);
1425
1426         bch2_btree_node_unlock_write(b, iter);
1427
1428         btree_node_interior_verify(c, b);
1429
1430         /*
1431          * when called from the btree_split path the new nodes aren't added to
1432          * the btree iterator yet, so the merge path's unlock/wait/relock dance
1433          * won't work:
1434          */
1435         bch2_foreground_maybe_merge(c, iter, b->c.level,
1436                                     flags|BTREE_INSERT_NOUNLOCK);
1437         return;
1438 split:
1439         btree_split(as, b, iter, keys, flags);
1440 }
1441
1442 int bch2_btree_split_leaf(struct bch_fs *c, struct btree_iter *iter,
1443                           unsigned flags)
1444 {
1445         struct btree_trans *trans = iter->trans;
1446         struct btree *b = iter_l(iter)->b;
1447         struct btree_update *as;
1448         struct closure cl;
1449         int ret = 0;
1450
1451         closure_init_stack(&cl);
1452
1453         /* Hack, because gc and splitting nodes doesn't mix yet: */
1454         if (!(flags & BTREE_INSERT_GC_LOCK_HELD) &&
1455             !down_read_trylock(&c->gc_lock)) {
1456                 if (flags & BTREE_INSERT_NOUNLOCK) {
1457                         trace_transaction_restart_ip(trans->ip, _THIS_IP_);
1458                         return -EINTR;
1459                 }
1460
1461                 bch2_trans_unlock(trans);
1462                 down_read(&c->gc_lock);
1463
1464                 if (!bch2_trans_relock(trans))
1465                         ret = -EINTR;
1466         }
1467
1468         /*
1469          * XXX: figure out how far we might need to split,
1470          * instead of locking/reserving all the way to the root:
1471          */
1472         if (!bch2_btree_iter_upgrade(iter, U8_MAX)) {
1473                 trace_trans_restart_iter_upgrade(trans->ip);
1474                 ret = -EINTR;
1475                 goto out;
1476         }
1477
1478         as = bch2_btree_update_start(trans, iter->btree_id,
1479                 btree_update_reserve_required(c, b), flags,
1480                 !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1481         if (IS_ERR(as)) {
1482                 ret = PTR_ERR(as);
1483                 if (ret == -EAGAIN) {
1484                         BUG_ON(flags & BTREE_INSERT_NOUNLOCK);
1485                         bch2_trans_unlock(trans);
1486                         ret = -EINTR;
1487
1488                         trace_transaction_restart_ip(trans->ip, _THIS_IP_);
1489                 }
1490                 goto out;
1491         }
1492
1493         btree_split(as, b, iter, NULL, flags);
1494         bch2_btree_update_done(as);
1495
1496         /*
1497          * We haven't successfully inserted yet, so don't downgrade all the way
1498          * back to read locks;
1499          */
1500         __bch2_btree_iter_downgrade(iter, 1);
1501 out:
1502         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1503                 up_read(&c->gc_lock);
1504         closure_sync(&cl);
1505         return ret;
1506 }
1507
1508 void __bch2_foreground_maybe_merge(struct bch_fs *c,
1509                                    struct btree_iter *iter,
1510                                    unsigned level,
1511                                    unsigned flags,
1512                                    enum btree_node_sibling sib)
1513 {
1514         struct btree_trans *trans = iter->trans;
1515         struct btree_update *as;
1516         struct bkey_format_state new_s;
1517         struct bkey_format new_f;
1518         struct bkey_i delete;
1519         struct btree *b, *m, *n, *prev, *next, *parent;
1520         struct closure cl;
1521         size_t sib_u64s;
1522         int ret = 0;
1523
1524         BUG_ON(!btree_node_locked(iter, level));
1525
1526         closure_init_stack(&cl);
1527 retry:
1528         BUG_ON(!btree_node_locked(iter, level));
1529
1530         b = iter->l[level].b;
1531
1532         parent = btree_node_parent(iter, b);
1533         if (!parent)
1534                 goto out;
1535
1536         if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c))
1537                 goto out;
1538
1539         /* XXX: can't be holding read locks */
1540         m = bch2_btree_node_get_sibling(c, iter, b, sib);
1541         if (IS_ERR(m)) {
1542                 ret = PTR_ERR(m);
1543                 goto err;
1544         }
1545
1546         /* NULL means no sibling: */
1547         if (!m) {
1548                 b->sib_u64s[sib] = U16_MAX;
1549                 goto out;
1550         }
1551
1552         if (sib == btree_prev_sib) {
1553                 prev = m;
1554                 next = b;
1555         } else {
1556                 prev = b;
1557                 next = m;
1558         }
1559
1560         bch2_bkey_format_init(&new_s);
1561         __bch2_btree_calc_format(&new_s, b);
1562         __bch2_btree_calc_format(&new_s, m);
1563         new_f = bch2_bkey_format_done(&new_s);
1564
1565         sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1566                 btree_node_u64s_with_format(m, &new_f);
1567
1568         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1569                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1570                 sib_u64s /= 2;
1571                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1572         }
1573
1574         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1575         b->sib_u64s[sib] = sib_u64s;
1576
1577         if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c)) {
1578                 six_unlock_intent(&m->c.lock);
1579                 goto out;
1580         }
1581
1582         /* We're changing btree topology, doesn't mix with gc: */
1583         if (!(flags & BTREE_INSERT_GC_LOCK_HELD) &&
1584             !down_read_trylock(&c->gc_lock))
1585                 goto err_cycle_gc_lock;
1586
1587         if (!bch2_btree_iter_upgrade(iter, U8_MAX)) {
1588                 ret = -EINTR;
1589                 goto err_unlock;
1590         }
1591
1592         as = bch2_btree_update_start(trans, iter->btree_id,
1593                          btree_update_reserve_required(c, parent) + 1,
1594                          flags|
1595                          BTREE_INSERT_NOFAIL|
1596                          BTREE_INSERT_USE_RESERVE,
1597                          !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1598         if (IS_ERR(as)) {
1599                 ret = PTR_ERR(as);
1600                 goto err_unlock;
1601         }
1602
1603         trace_btree_merge(c, b);
1604
1605         bch2_btree_interior_update_will_free_node(as, b);
1606         bch2_btree_interior_update_will_free_node(as, m);
1607
1608         n = bch2_btree_node_alloc(as, b->c.level);
1609         bch2_btree_update_add_new_node(as, n);
1610
1611         btree_set_min(n, prev->data->min_key);
1612         btree_set_max(n, next->data->max_key);
1613         n->data->format         = new_f;
1614
1615         btree_node_set_format(n, new_f);
1616
1617         bch2_btree_sort_into(c, n, prev);
1618         bch2_btree_sort_into(c, n, next);
1619
1620         bch2_btree_build_aux_trees(n);
1621         six_unlock_write(&n->c.lock);
1622
1623         bkey_init(&delete.k);
1624         delete.k.p = prev->key.k.p;
1625         bch2_keylist_add(&as->parent_keys, &delete);
1626         bch2_keylist_add(&as->parent_keys, &n->key);
1627
1628         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1629
1630         bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1631
1632         bch2_btree_update_get_open_buckets(as, n);
1633
1634         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1635         bch2_btree_iter_node_drop(iter, b);
1636         bch2_btree_iter_node_drop(iter, m);
1637
1638         bch2_btree_iter_node_replace(iter, n);
1639
1640         bch2_btree_trans_verify_iters(trans, n);
1641
1642         bch2_btree_node_free_inmem(c, b, iter);
1643         bch2_btree_node_free_inmem(c, m, iter);
1644
1645         six_unlock_intent(&n->c.lock);
1646
1647         bch2_btree_update_done(as);
1648
1649         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1650                 up_read(&c->gc_lock);
1651 out:
1652         bch2_btree_trans_verify_locks(trans);
1653
1654         /*
1655          * Don't downgrade locks here: we're called after successful insert,
1656          * and the caller will downgrade locks after a successful insert
1657          * anyways (in case e.g. a split was required first)
1658          *
1659          * And we're also called when inserting into interior nodes in the
1660          * split path, and downgrading to read locks in there is potentially
1661          * confusing:
1662          */
1663         closure_sync(&cl);
1664         return;
1665
1666 err_cycle_gc_lock:
1667         six_unlock_intent(&m->c.lock);
1668
1669         if (flags & BTREE_INSERT_NOUNLOCK)
1670                 goto out;
1671
1672         bch2_trans_unlock(trans);
1673
1674         down_read(&c->gc_lock);
1675         up_read(&c->gc_lock);
1676         ret = -EINTR;
1677         goto err;
1678
1679 err_unlock:
1680         six_unlock_intent(&m->c.lock);
1681         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1682                 up_read(&c->gc_lock);
1683 err:
1684         BUG_ON(ret == -EAGAIN && (flags & BTREE_INSERT_NOUNLOCK));
1685
1686         if ((ret == -EAGAIN || ret == -EINTR) &&
1687             !(flags & BTREE_INSERT_NOUNLOCK)) {
1688                 bch2_trans_unlock(trans);
1689                 closure_sync(&cl);
1690                 ret = bch2_btree_iter_traverse(iter);
1691                 if (ret)
1692                         goto out;
1693
1694                 goto retry;
1695         }
1696
1697         goto out;
1698 }
1699
1700 static int __btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
1701                                 struct btree *b, unsigned flags,
1702                                 struct closure *cl)
1703 {
1704         struct btree *n, *parent = btree_node_parent(iter, b);
1705         struct btree_update *as;
1706
1707         as = bch2_btree_update_start(iter->trans, iter->btree_id,
1708                 (parent
1709                  ? btree_update_reserve_required(c, parent)
1710                  : 0) + 1,
1711                 flags, cl);
1712         if (IS_ERR(as)) {
1713                 trace_btree_gc_rewrite_node_fail(c, b);
1714                 return PTR_ERR(as);
1715         }
1716
1717         bch2_btree_interior_update_will_free_node(as, b);
1718
1719         n = bch2_btree_node_alloc_replacement(as, b);
1720         bch2_btree_update_add_new_node(as, n);
1721
1722         bch2_btree_build_aux_trees(n);
1723         six_unlock_write(&n->c.lock);
1724
1725         trace_btree_gc_rewrite_node(c, b);
1726
1727         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1728
1729         if (parent) {
1730                 bch2_keylist_add(&as->parent_keys, &n->key);
1731                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1732         } else {
1733                 bch2_btree_set_root(as, n, iter);
1734         }
1735
1736         bch2_btree_update_get_open_buckets(as, n);
1737
1738         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1739         bch2_btree_iter_node_drop(iter, b);
1740         bch2_btree_iter_node_replace(iter, n);
1741         bch2_btree_node_free_inmem(c, b, iter);
1742         six_unlock_intent(&n->c.lock);
1743
1744         bch2_btree_update_done(as);
1745         return 0;
1746 }
1747
1748 /**
1749  * bch_btree_node_rewrite - Rewrite/move a btree node
1750  *
1751  * Returns 0 on success, -EINTR or -EAGAIN on failure (i.e.
1752  * btree_check_reserve() has to wait)
1753  */
1754 int bch2_btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
1755                             __le64 seq, unsigned flags)
1756 {
1757         struct btree_trans *trans = iter->trans;
1758         struct closure cl;
1759         struct btree *b;
1760         int ret;
1761
1762         flags |= BTREE_INSERT_NOFAIL;
1763
1764         closure_init_stack(&cl);
1765
1766         bch2_btree_iter_upgrade(iter, U8_MAX);
1767
1768         if (!(flags & BTREE_INSERT_GC_LOCK_HELD)) {
1769                 if (!down_read_trylock(&c->gc_lock)) {
1770                         bch2_trans_unlock(trans);
1771                         down_read(&c->gc_lock);
1772                 }
1773         }
1774
1775         while (1) {
1776                 ret = bch2_btree_iter_traverse(iter);
1777                 if (ret)
1778                         break;
1779
1780                 b = bch2_btree_iter_peek_node(iter);
1781                 if (!b || b->data->keys.seq != seq)
1782                         break;
1783
1784                 ret = __btree_node_rewrite(c, iter, b, flags, &cl);
1785                 if (ret != -EAGAIN &&
1786                     ret != -EINTR)
1787                         break;
1788
1789                 bch2_trans_unlock(trans);
1790                 closure_sync(&cl);
1791         }
1792
1793         bch2_btree_iter_downgrade(iter);
1794
1795         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1796                 up_read(&c->gc_lock);
1797
1798         closure_sync(&cl);
1799         return ret;
1800 }
1801
1802 static void __bch2_btree_node_update_key(struct bch_fs *c,
1803                                          struct btree_update *as,
1804                                          struct btree_iter *iter,
1805                                          struct btree *b, struct btree *new_hash,
1806                                          struct bkey_i *new_key)
1807 {
1808         struct btree *parent;
1809         int ret;
1810
1811         btree_update_will_delete_key(as, &b->key);
1812         btree_update_will_add_key(as, new_key);
1813
1814         parent = btree_node_parent(iter, b);
1815         if (parent) {
1816                 if (new_hash) {
1817                         bkey_copy(&new_hash->key, new_key);
1818                         ret = bch2_btree_node_hash_insert(&c->btree_cache,
1819                                         new_hash, b->c.level, b->c.btree_id);
1820                         BUG_ON(ret);
1821                 }
1822
1823                 bch2_keylist_add(&as->parent_keys, new_key);
1824                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, 0);
1825
1826                 if (new_hash) {
1827                         mutex_lock(&c->btree_cache.lock);
1828                         bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
1829
1830                         bch2_btree_node_hash_remove(&c->btree_cache, b);
1831
1832                         bkey_copy(&b->key, new_key);
1833                         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1834                         BUG_ON(ret);
1835                         mutex_unlock(&c->btree_cache.lock);
1836                 } else {
1837                         bkey_copy(&b->key, new_key);
1838                 }
1839         } else {
1840                 BUG_ON(btree_node_root(c, b) != b);
1841
1842                 bch2_btree_node_lock_write(b, iter);
1843                 bkey_copy(&b->key, new_key);
1844
1845                 if (btree_ptr_hash_val(&b->key) != b->hash_val) {
1846                         mutex_lock(&c->btree_cache.lock);
1847                         bch2_btree_node_hash_remove(&c->btree_cache, b);
1848
1849                         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1850                         BUG_ON(ret);
1851                         mutex_unlock(&c->btree_cache.lock);
1852                 }
1853
1854                 btree_update_updated_root(as, b);
1855                 bch2_btree_node_unlock_write(b, iter);
1856         }
1857
1858         bch2_btree_update_done(as);
1859 }
1860
1861 int bch2_btree_node_update_key(struct bch_fs *c, struct btree_iter *iter,
1862                                struct btree *b,
1863                                struct bkey_i *new_key)
1864 {
1865         struct btree *parent = btree_node_parent(iter, b);
1866         struct btree_update *as = NULL;
1867         struct btree *new_hash = NULL;
1868         struct closure cl;
1869         int ret;
1870
1871         closure_init_stack(&cl);
1872
1873         if (!bch2_btree_iter_upgrade(iter, U8_MAX))
1874                 return -EINTR;
1875
1876         if (!down_read_trylock(&c->gc_lock)) {
1877                 bch2_trans_unlock(iter->trans);
1878                 down_read(&c->gc_lock);
1879
1880                 if (!bch2_trans_relock(iter->trans)) {
1881                         ret = -EINTR;
1882                         goto err;
1883                 }
1884         }
1885
1886         /*
1887          * check btree_ptr_hash_val() after @b is locked by
1888          * btree_iter_traverse():
1889          */
1890         if (btree_ptr_hash_val(new_key) != b->hash_val) {
1891                 /* bch2_btree_reserve_get will unlock */
1892                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1893                 if (ret) {
1894                         bch2_trans_unlock(iter->trans);
1895                         up_read(&c->gc_lock);
1896                         closure_sync(&cl);
1897                         down_read(&c->gc_lock);
1898
1899                         if (!bch2_trans_relock(iter->trans)) {
1900                                 ret = -EINTR;
1901                                 goto err;
1902                         }
1903                 }
1904
1905                 new_hash = bch2_btree_node_mem_alloc(c);
1906         }
1907 retry:
1908         as = bch2_btree_update_start(iter->trans, iter->btree_id,
1909                 parent ? btree_update_reserve_required(c, parent) : 0,
1910                 BTREE_INSERT_NOFAIL, &cl);
1911
1912         if (IS_ERR(as)) {
1913                 ret = PTR_ERR(as);
1914                 if (ret == -EAGAIN)
1915                         ret = -EINTR;
1916
1917                 if (ret == -EINTR) {
1918                         bch2_trans_unlock(iter->trans);
1919                         up_read(&c->gc_lock);
1920                         closure_sync(&cl);
1921                         down_read(&c->gc_lock);
1922
1923                         if (bch2_trans_relock(iter->trans))
1924                                 goto retry;
1925                 }
1926
1927                 goto err;
1928         }
1929
1930         ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(new_key));
1931         if (ret)
1932                 goto err_free_update;
1933
1934         __bch2_btree_node_update_key(c, as, iter, b, new_hash, new_key);
1935
1936         bch2_btree_iter_downgrade(iter);
1937 err:
1938         if (new_hash) {
1939                 mutex_lock(&c->btree_cache.lock);
1940                 list_move(&new_hash->list, &c->btree_cache.freeable);
1941                 mutex_unlock(&c->btree_cache.lock);
1942
1943                 six_unlock_write(&new_hash->c.lock);
1944                 six_unlock_intent(&new_hash->c.lock);
1945         }
1946         up_read(&c->gc_lock);
1947         closure_sync(&cl);
1948         return ret;
1949 err_free_update:
1950         bch2_btree_update_free(as);
1951         goto err;
1952 }
1953
1954 /* Init code: */
1955
1956 /*
1957  * Only for filesystem bringup, when first reading the btree roots or allocating
1958  * btree roots when initializing a new filesystem:
1959  */
1960 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
1961 {
1962         BUG_ON(btree_node_root(c, b));
1963
1964         bch2_btree_set_root_inmem(c, b);
1965 }
1966
1967 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
1968 {
1969         struct closure cl;
1970         struct btree *b;
1971         int ret;
1972
1973         closure_init_stack(&cl);
1974
1975         do {
1976                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1977                 closure_sync(&cl);
1978         } while (ret);
1979
1980         b = bch2_btree_node_mem_alloc(c);
1981         bch2_btree_cache_cannibalize_unlock(c);
1982
1983         set_btree_node_fake(b);
1984         set_btree_node_need_rewrite(b);
1985         b->c.level      = 0;
1986         b->c.btree_id   = id;
1987
1988         bkey_btree_ptr_init(&b->key);
1989         b->key.k.p = POS_MAX;
1990         *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
1991
1992         bch2_bset_init_first(b, &b->data->keys);
1993         bch2_btree_build_aux_trees(b);
1994
1995         b->data->flags = 0;
1996         btree_set_min(b, POS_MIN);
1997         btree_set_max(b, POS_MAX);
1998         b->data->format = bch2_btree_calc_format(b);
1999         btree_node_set_format(b, b->data->format);
2000
2001         ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2002                                           b->c.level, b->c.btree_id);
2003         BUG_ON(ret);
2004
2005         bch2_btree_set_root_inmem(c, b);
2006
2007         six_unlock_write(&b->c.lock);
2008         six_unlock_intent(&b->c.lock);
2009 }
2010
2011 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2012 {
2013         struct btree_update *as;
2014
2015         mutex_lock(&c->btree_interior_update_lock);
2016         list_for_each_entry(as, &c->btree_interior_update_list, list)
2017                 pr_buf(out, "%p m %u w %u r %u j %llu\n",
2018                        as,
2019                        as->mode,
2020                        as->nodes_written,
2021                        atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2022                        as->journal.seq);
2023         mutex_unlock(&c->btree_interior_update_lock);
2024 }
2025
2026 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *c)
2027 {
2028         size_t ret = 0;
2029         struct list_head *i;
2030
2031         mutex_lock(&c->btree_interior_update_lock);
2032         list_for_each(i, &c->btree_interior_update_list)
2033                 ret++;
2034         mutex_unlock(&c->btree_interior_update_lock);
2035
2036         return ret;
2037 }
2038
2039 void bch2_journal_entries_to_btree_roots(struct bch_fs *c, struct jset *jset)
2040 {
2041         struct btree_root *r;
2042         struct jset_entry *entry;
2043
2044         mutex_lock(&c->btree_root_lock);
2045
2046         vstruct_for_each(jset, entry)
2047                 if (entry->type == BCH_JSET_ENTRY_btree_root) {
2048                         r = &c->btree_roots[entry->btree_id];
2049                         r->level = entry->level;
2050                         r->alive = true;
2051                         bkey_copy(&r->key, &entry->start[0]);
2052                 }
2053
2054         mutex_unlock(&c->btree_root_lock);
2055 }
2056
2057 struct jset_entry *
2058 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2059                                     struct jset_entry *start,
2060                                     struct jset_entry *end)
2061 {
2062         struct jset_entry *entry;
2063         unsigned long have = 0;
2064         unsigned i;
2065
2066         for (entry = start; entry < end; entry = vstruct_next(entry))
2067                 if (entry->type == BCH_JSET_ENTRY_btree_root)
2068                         __set_bit(entry->btree_id, &have);
2069
2070         mutex_lock(&c->btree_root_lock);
2071
2072         for (i = 0; i < BTREE_ID_NR; i++)
2073                 if (c->btree_roots[i].alive && !test_bit(i, &have)) {
2074                         journal_entry_set(end,
2075                                           BCH_JSET_ENTRY_btree_root,
2076                                           i, c->btree_roots[i].level,
2077                                           &c->btree_roots[i].key,
2078                                           c->btree_roots[i].key.u64s);
2079                         end = vstruct_next(end);
2080                 }
2081
2082         mutex_unlock(&c->btree_root_lock);
2083
2084         return end;
2085 }
2086
2087 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2088 {
2089         if (c->btree_interior_update_worker)
2090                 destroy_workqueue(c->btree_interior_update_worker);
2091         mempool_exit(&c->btree_interior_update_pool);
2092 }
2093
2094 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2095 {
2096         mutex_init(&c->btree_reserve_cache_lock);
2097         INIT_LIST_HEAD(&c->btree_interior_update_list);
2098         INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2099         mutex_init(&c->btree_interior_update_lock);
2100         INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2101
2102         c->btree_interior_update_worker =
2103                 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2104         if (!c->btree_interior_update_worker)
2105                 return -ENOMEM;
2106
2107         return mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2108                                          sizeof(struct btree_update));
2109 }