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 (bpos_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 (bpos_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 = bpos_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         for_each_bset(b, t)
  86                 bset_tree_for_each_key(b, t, k)
  87                         if (!bkey_deleted(k)) {
  88                                 uk = bkey_unpack_key(b, k);
  89                                 bch2_bkey_format_add_key(s, &uk);
  90                         }
  91 }
  92
  93 static struct bkey_format bch2_btree_calc_format(struct btree *b)
  94 {
  95         struct bkey_format_state s;
  96
  97         bch2_bkey_format_init(&s);
  98         bch2_bkey_format_add_pos(&s, b->data->min_key);
  99         bch2_bkey_format_add_pos(&s, b->data->max_key);
 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
 293         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 294                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
 295
 296                 bp->v.mem_ptr           = 0;
 297                 bp->v.seq               = b->data->keys.seq;
 298                 bp->v.sectors_written   = 0;
 299         }
 300
 301         SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
 302
 303         bch2_btree_build_aux_trees(b);
 304
 305         ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
 306         BUG_ON(ret);
 307
 308         trace_btree_node_alloc(c, b);
 309         return b;
 310 }
 311
 312 static void btree_set_min(struct btree *b, struct bpos pos)
 313 {
 314         if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
 315                 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
 316         b->data->min_key = pos;
 317 }
 318
 319 static void btree_set_max(struct btree *b, struct bpos pos)
 320 {
 321         b->key.k.p = pos;
 322         b->data->max_key = pos;
 323 }
 324
 325 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *as,
 326                                                   struct btree *b,
 327                                                   struct bkey_format format)
 328 {
 329         struct btree *n;
 330
 331         n = bch2_btree_node_alloc(as, b->c.level);
 332
 333         SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
 334
 335         btree_set_min(n, b->data->min_key);
 336         btree_set_max(n, b->data->max_key);
 337
 338         n->data->format         = format;
 339         btree_node_set_format(n, format);
 340
 341         bch2_btree_sort_into(as->c, n, b);
 342
 343         btree_node_reset_sib_u64s(n);
 344
 345         n->key.k.p = b->key.k.p;
 346         return n;
 347 }
 348
 349 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
 350                                                        struct btree *b)
 351 {
 352         struct bkey_format new_f = bch2_btree_calc_format(b);
 353
 354         /*
 355          * The keys might expand with the new format - if they wouldn't fit in
 356          * the btree node anymore, use the old format for now:
 357          */
 358         if (!bch2_btree_node_format_fits(as->c, b, &new_f))
 359                 new_f = b->format;
 360
 361         return __bch2_btree_node_alloc_replacement(as, b, new_f);
 362 }
 363
 364 static struct btree *__btree_root_alloc(struct btree_update *as, unsigned level)
 365 {
 366         struct btree *b = bch2_btree_node_alloc(as, level);
 367
 368         btree_set_min(b, POS_MIN);
 369         btree_set_max(b, POS_MAX);
 370         b->data->format = bch2_btree_calc_format(b);
 371
 372         btree_node_set_format(b, b->data->format);
 373         bch2_btree_build_aux_trees(b);
 374
 375         bch2_btree_update_add_new_node(as, b);
 376         six_unlock_write(&b->c.lock);
 377
 378         return b;
 379 }
 380
 381 static void bch2_btree_reserve_put(struct btree_update *as)
 382 {
 383         struct bch_fs *c = as->c;
 384
 385         mutex_lock(&c->btree_reserve_cache_lock);
 386
 387         while (as->nr_prealloc_nodes) {
 388                 struct btree *b = as->prealloc_nodes[--as->nr_prealloc_nodes];
 389
 390                 six_unlock_write(&b->c.lock);
 391
 392                 if (c->btree_reserve_cache_nr <
 393                     ARRAY_SIZE(c->btree_reserve_cache)) {
 394                         struct btree_alloc *a =
 395                                 &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
 396
 397                         a->ob = b->ob;
 398                         b->ob.nr = 0;
 399                         bkey_copy(&a->k, &b->key);
 400                 } else {
 401                         bch2_open_buckets_put(c, &b->ob);
 402                 }
 403
 404                 btree_node_lock_type(c, b, SIX_LOCK_write);
 405                 __btree_node_free(c, b);
 406                 six_unlock_write(&b->c.lock);
 407
 408                 six_unlock_intent(&b->c.lock);
 409         }
 410
 411         mutex_unlock(&c->btree_reserve_cache_lock);
 412 }
 413
 414 static int bch2_btree_reserve_get(struct btree_update *as, unsigned nr_nodes,
 415                                   unsigned flags, struct closure *cl)
 416 {
 417         struct bch_fs *c = as->c;
 418         struct btree *b;
 419         int ret;
 420
 421         BUG_ON(nr_nodes > BTREE_RESERVE_MAX);
 422
 423         /*
 424          * Protects reaping from the btree node cache and using the btree node
 425          * open bucket reserve:
 426          */
 427         ret = bch2_btree_cache_cannibalize_lock(c, cl);
 428         if (ret)
 429                 return ret;
 430
 431         while (as->nr_prealloc_nodes < nr_nodes) {
 432                 b = __bch2_btree_node_alloc(c, &as->disk_res,
 433                                             flags & BTREE_INSERT_NOWAIT
 434                                             ? NULL : cl, flags);
 435                 if (IS_ERR(b)) {
 436                         ret = PTR_ERR(b);
 437                         goto err_free;
 438                 }
 439
 440                 ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(&b->key));
 441                 if (ret)
 442                         goto err_free;
 443
 444                 as->prealloc_nodes[as->nr_prealloc_nodes++] = b;
 445         }
 446
 447         bch2_btree_cache_cannibalize_unlock(c);
 448         return 0;
 449 err_free:
 450         bch2_btree_cache_cannibalize_unlock(c);
 451         trace_btree_reserve_get_fail(c, nr_nodes, cl);
 452         return ret;
 453 }
 454
 455 /* Asynchronous interior node update machinery */
 456
 457 static void bch2_btree_update_free(struct btree_update *as)
 458 {
 459         struct bch_fs *c = as->c;
 460
 461         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 462
 463         bch2_journal_pin_drop(&c->journal, &as->journal);
 464         bch2_journal_pin_flush(&c->journal, &as->journal);
 465         bch2_disk_reservation_put(c, &as->disk_res);
 466         bch2_btree_reserve_put(as);
 467
 468         mutex_lock(&c->btree_interior_update_lock);
 469         list_del(&as->unwritten_list);
 470         list_del(&as->list);
 471         mutex_unlock(&c->btree_interior_update_lock);
 472
 473         closure_debug_destroy(&as->cl);
 474         mempool_free(as, &c->btree_interior_update_pool);
 475
 476         closure_wake_up(&c->btree_interior_update_wait);
 477 }
 478
 479 static void btree_update_will_delete_key(struct btree_update *as,
 480                                          struct bkey_i *k)
 481 {
 482         BUG_ON(bch2_keylist_u64s(&as->old_keys) + k->k.u64s >
 483                ARRAY_SIZE(as->_old_keys));
 484         bch2_keylist_add(&as->old_keys, k);
 485 }
 486
 487 static void btree_update_will_add_key(struct btree_update *as,
 488                                       struct bkey_i *k)
 489 {
 490         BUG_ON(bch2_keylist_u64s(&as->new_keys) + k->k.u64s >
 491                ARRAY_SIZE(as->_new_keys));
 492         bch2_keylist_add(&as->new_keys, k);
 493 }
 494
 495 /*
 496  * The transactional part of an interior btree node update, where we journal the
 497  * update we did to the interior node and update alloc info:
 498  */
 499 static int btree_update_nodes_written_trans(struct btree_trans *trans,
 500                                             struct btree_update *as)
 501 {
 502         struct bkey_i *k;
 503         int ret;
 504
 505         trans->extra_journal_entries = (void *) &as->journal_entries[0];
 506         trans->extra_journal_entry_u64s = as->journal_u64s;
 507         trans->journal_pin = &as->journal;
 508
 509         for_each_keylist_key(&as->new_keys, k) {
 510                 ret = bch2_trans_mark_key(trans,
 511                                           bkey_s_c_null,
 512                                           bkey_i_to_s_c(k),
 513                                           0, 0, BTREE_TRIGGER_INSERT);
 514                 if (ret)
 515                         return ret;
 516         }
 517
 518         for_each_keylist_key(&as->old_keys, k) {
 519                 ret = bch2_trans_mark_key(trans,
 520                                           bkey_i_to_s_c(k),
 521                                           bkey_s_c_null,
 522                                           0, 0, BTREE_TRIGGER_OVERWRITE);
 523                 if (ret)
 524                         return ret;
 525         }
 526
 527         return 0;
 528 }
 529
 530 static void btree_update_nodes_written(struct btree_update *as)
 531 {
 532         struct bch_fs *c = as->c;
 533         struct btree *b = as->b;
 534         struct btree_trans trans;
 535         u64 journal_seq = 0;
 536         unsigned i;
 537         int ret;
 538
 539         /*
 540          * If we're already in an error state, it might be because a btree node
 541          * was never written, and we might be trying to free that same btree
 542          * node here, but it won't have been marked as allocated and we'll see
 543          * spurious disk usage inconsistencies in the transactional part below
 544          * if we don't skip it:
 545          */
 546         ret = bch2_journal_error(&c->journal);
 547         if (ret)
 548                 goto err;
 549
 550         BUG_ON(!journal_pin_active(&as->journal));
 551
 552         /*
 553          * We did an update to a parent node where the pointers we added pointed
 554          * to child nodes that weren't written yet: now, the child nodes have
 555          * been written so we can write out the update to the interior node.
 556          */
 557
 558         /*
 559          * We can't call into journal reclaim here: we'd block on the journal
 560          * reclaim lock, but we may need to release the open buckets we have
 561          * pinned in order for other btree updates to make forward progress, and
 562          * journal reclaim does btree updates when flushing bkey_cached entries,
 563          * which may require allocations as well.
 564          */
 565         bch2_trans_init(&trans, c, 0, 512);
 566         ret = __bch2_trans_do(&trans, &as->disk_res, &journal_seq,
 567                               BTREE_INSERT_NOFAIL|
 568                               BTREE_INSERT_NOCHECK_RW|
 569                               BTREE_INSERT_JOURNAL_RECLAIM|
 570                               BTREE_INSERT_JOURNAL_RESERVED,
 571                               btree_update_nodes_written_trans(&trans, as));
 572         bch2_trans_exit(&trans);
 573
 574         bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
 575                              "error %i in btree_update_nodes_written()", ret);
 576 err:
 577         if (b) {
 578                 /*
 579                  * @b is the node we did the final insert into:
 580                  *
 581                  * On failure to get a journal reservation, we still have to
 582                  * unblock the write and allow most of the write path to happen
 583                  * so that shutdown works, but the i->journal_seq mechanism
 584                  * won't work to prevent the btree write from being visible (we
 585                  * didn't get a journal sequence number) - instead
 586                  * __bch2_btree_node_write() doesn't do the actual write if
 587                  * we're in journal error state:
 588                  */
 589
 590                 btree_node_lock_type(c, b, SIX_LOCK_intent);
 591                 btree_node_lock_type(c, b, SIX_LOCK_write);
 592                 mutex_lock(&c->btree_interior_update_lock);
 593
 594                 list_del(&as->write_blocked_list);
 595
 596                 /*
 597                  * Node might have been freed, recheck under
 598                  * btree_interior_update_lock:
 599                  */
 600                 if (as->b == b) {
 601                         struct bset *i = btree_bset_last(b);
 602
 603                         BUG_ON(!b->c.level);
 604                         BUG_ON(!btree_node_dirty(b));
 605
 606                         if (!ret) {
 607                                 i->journal_seq = cpu_to_le64(
 608                                         max(journal_seq,
 609                                             le64_to_cpu(i->journal_seq)));
 610
 611                                 bch2_btree_add_journal_pin(c, b, journal_seq);
 612                         } else {
 613                                 /*
 614                                  * If we didn't get a journal sequence number we
 615                                  * can't write this btree node, because recovery
 616                                  * won't know to ignore this write:
 617                                  */
 618                                 set_btree_node_never_write(b);
 619                         }
 620                 }
 621
 622                 mutex_unlock(&c->btree_interior_update_lock);
 623                 six_unlock_write(&b->c.lock);
 624
 625                 btree_node_write_if_need(c, b, SIX_LOCK_intent);
 626                 six_unlock_intent(&b->c.lock);
 627         }
 628
 629         bch2_journal_pin_drop(&c->journal, &as->journal);
 630
 631         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 632
 633         mutex_lock(&c->btree_interior_update_lock);
 634         for (i = 0; i < as->nr_new_nodes; i++) {
 635                 b = as->new_nodes[i];
 636
 637                 BUG_ON(b->will_make_reachable != (unsigned long) as);
 638                 b->will_make_reachable = 0;
 639         }
 640         mutex_unlock(&c->btree_interior_update_lock);
 641
 642         for (i = 0; i < as->nr_new_nodes; i++) {
 643                 b = as->new_nodes[i];
 644
 645                 btree_node_lock_type(c, b, SIX_LOCK_read);
 646                 btree_node_write_if_need(c, b, SIX_LOCK_read);
 647                 six_unlock_read(&b->c.lock);
 648         }
 649
 650         for (i = 0; i < as->nr_open_buckets; i++)
 651                 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
 652
 653         bch2_btree_update_free(as);
 654 }
 655
 656 static void btree_interior_update_work(struct work_struct *work)
 657 {
 658         struct bch_fs *c =
 659                 container_of(work, struct bch_fs, btree_interior_update_work);
 660         struct btree_update *as;
 661
 662         while (1) {
 663                 mutex_lock(&c->btree_interior_update_lock);
 664                 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
 665                                               struct btree_update, unwritten_list);
 666                 if (as && !as->nodes_written)
 667                         as = NULL;
 668                 mutex_unlock(&c->btree_interior_update_lock);
 669
 670                 if (!as)
 671                         break;
 672
 673                 btree_update_nodes_written(as);
 674         }
 675 }
 676
 677 static void btree_update_set_nodes_written(struct closure *cl)
 678 {
 679         struct btree_update *as = container_of(cl, struct btree_update, cl);
 680         struct bch_fs *c = as->c;
 681
 682         mutex_lock(&c->btree_interior_update_lock);
 683         as->nodes_written = true;
 684         mutex_unlock(&c->btree_interior_update_lock);
 685
 686         queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
 687 }
 688
 689 /*
 690  * We're updating @b with pointers to nodes that haven't finished writing yet:
 691  * block @b from being written until @as completes
 692  */
 693 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
 694 {
 695         struct bch_fs *c = as->c;
 696
 697         mutex_lock(&c->btree_interior_update_lock);
 698         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 699
 700         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 701         BUG_ON(!btree_node_dirty(b));
 702
 703         as->mode        = BTREE_INTERIOR_UPDATING_NODE;
 704         as->b           = b;
 705         list_add(&as->write_blocked_list, &b->write_blocked);
 706
 707         mutex_unlock(&c->btree_interior_update_lock);
 708 }
 709
 710 static void btree_update_reparent(struct btree_update *as,
 711                                   struct btree_update *child)
 712 {
 713         struct bch_fs *c = as->c;
 714
 715         lockdep_assert_held(&c->btree_interior_update_lock);
 716
 717         child->b = NULL;
 718         child->mode = BTREE_INTERIOR_UPDATING_AS;
 719
 720         bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL);
 721 }
 722
 723 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
 724 {
 725         struct bkey_i *insert = &b->key;
 726         struct bch_fs *c = as->c;
 727
 728         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 729
 730         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
 731                ARRAY_SIZE(as->journal_entries));
 732
 733         as->journal_u64s +=
 734                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
 735                                   BCH_JSET_ENTRY_btree_root,
 736                                   b->c.btree_id, b->c.level,
 737                                   insert, insert->k.u64s);
 738
 739         mutex_lock(&c->btree_interior_update_lock);
 740         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 741
 742         as->mode        = BTREE_INTERIOR_UPDATING_ROOT;
 743         mutex_unlock(&c->btree_interior_update_lock);
 744 }
 745
 746 /*
 747  * bch2_btree_update_add_new_node:
 748  *
 749  * This causes @as to wait on @b to be written, before it gets to
 750  * bch2_btree_update_nodes_written
 751  *
 752  * Additionally, it sets b->will_make_reachable to prevent any additional writes
 753  * to @b from happening besides the first until @b is reachable on disk
 754  *
 755  * And it adds @b to the list of @as's new nodes, so that we can update sector
 756  * counts in bch2_btree_update_nodes_written:
 757  */
 758 void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
 759 {
 760         struct bch_fs *c = as->c;
 761
 762         closure_get(&as->cl);
 763
 764         mutex_lock(&c->btree_interior_update_lock);
 765         BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
 766         BUG_ON(b->will_make_reachable);
 767
 768         as->new_nodes[as->nr_new_nodes++] = b;
 769         b->will_make_reachable = 1UL|(unsigned long) as;
 770
 771         mutex_unlock(&c->btree_interior_update_lock);
 772
 773         btree_update_will_add_key(as, &b->key);
 774 }
 775
 776 /*
 777  * returns true if @b was a new node
 778  */
 779 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
 780 {
 781         struct btree_update *as;
 782         unsigned long v;
 783         unsigned i;
 784
 785         mutex_lock(&c->btree_interior_update_lock);
 786         /*
 787          * When b->will_make_reachable != 0, it owns a ref on as->cl that's
 788          * dropped when it gets written by bch2_btree_complete_write - the
 789          * xchg() is for synchronization with bch2_btree_complete_write:
 790          */
 791         v = xchg(&b->will_make_reachable, 0);
 792         as = (struct btree_update *) (v & ~1UL);
 793
 794         if (!as) {
 795                 mutex_unlock(&c->btree_interior_update_lock);
 796                 return;
 797         }
 798
 799         for (i = 0; i < as->nr_new_nodes; i++)
 800                 if (as->new_nodes[i] == b)
 801                         goto found;
 802
 803         BUG();
 804 found:
 805         array_remove_item(as->new_nodes, as->nr_new_nodes, i);
 806         mutex_unlock(&c->btree_interior_update_lock);
 807
 808         if (v & 1)
 809                 closure_put(&as->cl);
 810 }
 811
 812 void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
 813 {
 814         while (b->ob.nr)
 815                 as->open_buckets[as->nr_open_buckets++] =
 816                         b->ob.v[--b->ob.nr];
 817 }
 818
 819 /*
 820  * @b is being split/rewritten: it may have pointers to not-yet-written btree
 821  * nodes and thus outstanding btree_updates - redirect @b's
 822  * btree_updates to point to this btree_update:
 823  */
 824 void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 825                                                struct btree *b)
 826 {
 827         struct bch_fs *c = as->c;
 828         struct btree_update *p, *n;
 829         struct btree_write *w;
 830
 831         set_btree_node_dying(b);
 832
 833         if (btree_node_fake(b))
 834                 return;
 835
 836         mutex_lock(&c->btree_interior_update_lock);
 837
 838         /*
 839          * Does this node have any btree_update operations preventing
 840          * it from being written?
 841          *
 842          * If so, redirect them to point to this btree_update: we can
 843          * write out our new nodes, but we won't make them visible until those
 844          * operations complete
 845          */
 846         list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
 847                 list_del_init(&p->write_blocked_list);
 848                 btree_update_reparent(as, p);
 849
 850                 /*
 851                  * for flush_held_btree_writes() waiting on updates to flush or
 852                  * nodes to be writeable:
 853                  */
 854                 closure_wake_up(&c->btree_interior_update_wait);
 855         }
 856
 857         clear_btree_node_dirty(c, b);
 858         clear_btree_node_need_write(b);
 859
 860         /*
 861          * Does this node have unwritten data that has a pin on the journal?
 862          *
 863          * If so, transfer that pin to the btree_update operation -
 864          * note that if we're freeing multiple nodes, we only need to keep the
 865          * oldest pin of any of the nodes we're freeing. We'll release the pin
 866          * when the new nodes are persistent and reachable on disk:
 867          */
 868         w = btree_current_write(b);
 869         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 870         bch2_journal_pin_drop(&c->journal, &w->journal);
 871
 872         w = btree_prev_write(b);
 873         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 874         bch2_journal_pin_drop(&c->journal, &w->journal);
 875
 876         mutex_unlock(&c->btree_interior_update_lock);
 877
 878         /*
 879          * Is this a node that isn't reachable on disk yet?
 880          *
 881          * Nodes that aren't reachable yet have writes blocked until they're
 882          * reachable - now that we've cancelled any pending writes and moved
 883          * things waiting on that write to wait on this update, we can drop this
 884          * node from the list of nodes that the other update is making
 885          * reachable, prior to freeing it:
 886          */
 887         btree_update_drop_new_node(c, b);
 888
 889         btree_update_will_delete_key(as, &b->key);
 890 }
 891
 892 void bch2_btree_update_done(struct btree_update *as)
 893 {
 894         BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
 895
 896         bch2_btree_reserve_put(as);
 897
 898         continue_at(&as->cl, btree_update_set_nodes_written, system_freezable_wq);
 899 }
 900
 901 struct btree_update *
 902 bch2_btree_update_start(struct btree_trans *trans, enum btree_id id,
 903                         unsigned nr_nodes, unsigned flags,
 904                         struct closure *cl)
 905 {
 906         struct bch_fs *c = trans->c;
 907         struct btree_update *as;
 908         int disk_res_flags = (flags & BTREE_INSERT_NOFAIL)
 909                 ? BCH_DISK_RESERVATION_NOFAIL : 0;
 910         int journal_flags = (flags & BTREE_INSERT_JOURNAL_RESERVED)
 911                 ? JOURNAL_RES_GET_RECLAIM : 0;
 912         int ret = 0;
 913
 914         /*
 915          * This check isn't necessary for correctness - it's just to potentially
 916          * prevent us from doing a lot of work that'll end up being wasted:
 917          */
 918         ret = bch2_journal_error(&c->journal);
 919         if (ret)
 920                 return ERR_PTR(ret);
 921
 922         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
 923         memset(as, 0, sizeof(*as));
 924         closure_init(&as->cl, NULL);
 925         as->c           = c;
 926         as->mode        = BTREE_INTERIOR_NO_UPDATE;
 927         as->btree_id    = id;
 928         INIT_LIST_HEAD(&as->list);
 929         INIT_LIST_HEAD(&as->unwritten_list);
 930         INIT_LIST_HEAD(&as->write_blocked_list);
 931         bch2_keylist_init(&as->old_keys, as->_old_keys);
 932         bch2_keylist_init(&as->new_keys, as->_new_keys);
 933         bch2_keylist_init(&as->parent_keys, as->inline_keys);
 934
 935         ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
 936                                       BTREE_UPDATE_JOURNAL_RES,
 937                                       journal_flags|JOURNAL_RES_GET_NONBLOCK);
 938         if (ret == -EAGAIN) {
 939                 if (flags & BTREE_INSERT_NOUNLOCK)
 940                         return ERR_PTR(-EINTR);
 941
 942                 bch2_trans_unlock(trans);
 943
 944                 ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
 945                                 BTREE_UPDATE_JOURNAL_RES,
 946                                 journal_flags);
 947                 if (ret)
 948                         return ERR_PTR(ret);
 949
 950                 if (!bch2_trans_relock(trans)) {
 951                         ret = -EINTR;
 952                         goto err;
 953                 }
 954         }
 955
 956         ret = bch2_disk_reservation_get(c, &as->disk_res,
 957                         nr_nodes * c->opts.btree_node_size,
 958                         c->opts.metadata_replicas,
 959                         disk_res_flags);
 960         if (ret)
 961                 goto err;
 962
 963         ret = bch2_btree_reserve_get(as, nr_nodes, flags, cl);
 964         if (ret)
 965                 goto err;
 966
 967         bch2_journal_pin_add(&c->journal,
 968                              atomic64_read(&c->journal.seq),
 969                              &as->journal, NULL);
 970
 971         mutex_lock(&c->btree_interior_update_lock);
 972         list_add_tail(&as->list, &c->btree_interior_update_list);
 973         mutex_unlock(&c->btree_interior_update_lock);
 974
 975         return as;
 976 err:
 977         bch2_btree_update_free(as);
 978         return ERR_PTR(ret);
 979 }
 980
 981 /* Btree root updates: */
 982
 983 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
 984 {
 985         /* Root nodes cannot be reaped */
 986         mutex_lock(&c->btree_cache.lock);
 987         list_del_init(&b->list);
 988         mutex_unlock(&c->btree_cache.lock);
 989
 990         if (b->c.level)
 991                 six_lock_pcpu_alloc(&b->c.lock);
 992         else
 993                 six_lock_pcpu_free(&b->c.lock);
 994
 995         mutex_lock(&c->btree_root_lock);
 996         BUG_ON(btree_node_root(c, b) &&
 997                (b->c.level < btree_node_root(c, b)->c.level ||
 998                 !btree_node_dying(btree_node_root(c, b))));
 999
1000         btree_node_root(c, b) = b;
1001         mutex_unlock(&c->btree_root_lock);
1002
1003         bch2_recalc_btree_reserve(c);
1004 }
1005
1006 /**
1007  * bch_btree_set_root - update the root in memory and on disk
1008  *
1009  * To ensure forward progress, the current task must not be holding any
1010  * btree node write locks. However, you must hold an intent lock on the
1011  * old root.
1012  *
1013  * Note: This allocates a journal entry but doesn't add any keys to
1014  * it.  All the btree roots are part of every journal write, so there
1015  * is nothing new to be done.  This just guarantees that there is a
1016  * journal write.
1017  */
1018 static void bch2_btree_set_root(struct btree_update *as, struct btree *b,
1019                                 struct btree_iter *iter)
1020 {
1021         struct bch_fs *c = as->c;
1022         struct btree *old;
1023
1024         trace_btree_set_root(c, b);
1025         BUG_ON(!b->written &&
1026                !test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags));
1027
1028         old = btree_node_root(c, b);
1029
1030         /*
1031          * Ensure no one is using the old root while we switch to the
1032          * new root:
1033          */
1034         bch2_btree_node_lock_write(old, iter);
1035
1036         bch2_btree_set_root_inmem(c, b);
1037
1038         btree_update_updated_root(as, b);
1039
1040         /*
1041          * Unlock old root after new root is visible:
1042          *
1043          * The new root isn't persistent, but that's ok: we still have
1044          * an intent lock on the new root, and any updates that would
1045          * depend on the new root would have to update the new root.
1046          */
1047         bch2_btree_node_unlock_write(old, iter);
1048 }
1049
1050 /* Interior node updates: */
1051
1052 static void bch2_insert_fixup_btree_ptr(struct btree_update *as, struct btree *b,
1053                                         struct btree_iter *iter,
1054                                         struct bkey_i *insert,
1055                                         struct btree_node_iter *node_iter)
1056 {
1057         struct bch_fs *c = as->c;
1058         struct bkey_packed *k;
1059         const char *invalid;
1060
1061         invalid = bch2_bkey_invalid(c, bkey_i_to_s_c(insert), btree_node_type(b)) ?:
1062                 bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert));
1063         if (invalid) {
1064                 char buf[160];
1065
1066                 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(insert));
1067                 bch2_fs_inconsistent(c, "inserting invalid bkey %s: %s", buf, invalid);
1068                 dump_stack();
1069         }
1070
1071         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1072                ARRAY_SIZE(as->journal_entries));
1073
1074         as->journal_u64s +=
1075                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1076                                   BCH_JSET_ENTRY_btree_keys,
1077                                   b->c.btree_id, b->c.level,
1078                                   insert, insert->k.u64s);
1079
1080         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1081                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1082                 bch2_btree_node_iter_advance(node_iter, b);
1083
1084         bch2_btree_bset_insert_key(iter, b, node_iter, insert);
1085         set_btree_node_dirty(c, b);
1086         set_btree_node_need_write(b);
1087 }
1088
1089 /*
1090  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1091  * node)
1092  */
1093 static struct btree *__btree_split_node(struct btree_update *as,
1094                                         struct btree *n1,
1095                                         struct btree_iter *iter)
1096 {
1097         struct bkey_format_state s;
1098         size_t nr_packed = 0, nr_unpacked = 0;
1099         struct btree *n2;
1100         struct bset *set1, *set2;
1101         struct bkey_packed *k, *set2_start, *set2_end, *out, *prev = NULL;
1102         struct bpos n1_pos;
1103
1104         n2 = bch2_btree_node_alloc(as, n1->c.level);
1105         bch2_btree_update_add_new_node(as, n2);
1106
1107         n2->data->max_key       = n1->data->max_key;
1108         n2->data->format        = n1->format;
1109         SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data));
1110         n2->key.k.p = n1->key.k.p;
1111
1112         set1 = btree_bset_first(n1);
1113         set2 = btree_bset_first(n2);
1114
1115         /*
1116          * Has to be a linear search because we don't have an auxiliary
1117          * search tree yet
1118          */
1119         k = set1->start;
1120         while (1) {
1121                 struct bkey_packed *n = bkey_next(k);
1122
1123                 if (n == vstruct_last(set1))
1124                         break;
1125                 if (k->_data - set1->_data >= (le16_to_cpu(set1->u64s) * 3) / 5)
1126                         break;
1127
1128                 if (bkey_packed(k))
1129                         nr_packed++;
1130                 else
1131                         nr_unpacked++;
1132
1133                 prev = k;
1134                 k = n;
1135         }
1136
1137         BUG_ON(!prev);
1138         set2_start      = k;
1139         set2_end        = vstruct_last(set1);
1140
1141         set1->u64s = cpu_to_le16((u64 *) set2_start - set1->_data);
1142         set_btree_bset_end(n1, n1->set);
1143
1144         n1->nr.live_u64s        = le16_to_cpu(set1->u64s);
1145         n1->nr.bset_u64s[0]     = le16_to_cpu(set1->u64s);
1146         n1->nr.packed_keys      = nr_packed;
1147         n1->nr.unpacked_keys    = nr_unpacked;
1148
1149         n1_pos = bkey_unpack_pos(n1, prev);
1150         if (as->c->sb.version < bcachefs_metadata_version_snapshot)
1151                 n1_pos.snapshot = U32_MAX;
1152
1153         btree_set_max(n1, n1_pos);
1154         btree_set_min(n2, bpos_successor(n1->key.k.p));
1155
1156         bch2_bkey_format_init(&s);
1157         bch2_bkey_format_add_pos(&s, n2->data->min_key);
1158         bch2_bkey_format_add_pos(&s, n2->data->max_key);
1159
1160         for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1161                 struct bkey uk = bkey_unpack_key(n1, k);
1162                 bch2_bkey_format_add_key(&s, &uk);
1163         }
1164
1165         n2->data->format = bch2_bkey_format_done(&s);
1166         btree_node_set_format(n2, n2->data->format);
1167
1168         out = set2->start;
1169         memset(&n2->nr, 0, sizeof(n2->nr));
1170
1171         for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1172                 BUG_ON(!bch2_bkey_transform(&n2->format, out, bkey_packed(k)
1173                                        ? &n1->format : &bch2_bkey_format_current, k));
1174                 out->format = KEY_FORMAT_LOCAL_BTREE;
1175                 btree_keys_account_key_add(&n2->nr, 0, out);
1176                 out = bkey_next(out);
1177         }
1178
1179         set2->u64s = cpu_to_le16((u64 *) out - set2->_data);
1180         set_btree_bset_end(n2, n2->set);
1181
1182         BUG_ON(!set1->u64s);
1183         BUG_ON(!set2->u64s);
1184
1185         btree_node_reset_sib_u64s(n1);
1186         btree_node_reset_sib_u64s(n2);
1187
1188         bch2_verify_btree_nr_keys(n1);
1189         bch2_verify_btree_nr_keys(n2);
1190
1191         if (n1->c.level) {
1192                 btree_node_interior_verify(as->c, n1);
1193                 btree_node_interior_verify(as->c, n2);
1194         }
1195
1196         return n2;
1197 }
1198
1199 /*
1200  * For updates to interior nodes, we've got to do the insert before we split
1201  * because the stuff we're inserting has to be inserted atomically. Post split,
1202  * the keys might have to go in different nodes and the split would no longer be
1203  * atomic.
1204  *
1205  * Worse, if the insert is from btree node coalescing, if we do the insert after
1206  * we do the split (and pick the pivot) - the pivot we pick might be between
1207  * nodes that were coalesced, and thus in the middle of a child node post
1208  * coalescing:
1209  */
1210 static void btree_split_insert_keys(struct btree_update *as, struct btree *b,
1211                                     struct btree_iter *iter,
1212                                     struct keylist *keys)
1213 {
1214         struct btree_node_iter node_iter;
1215         struct bkey_i *k = bch2_keylist_front(keys);
1216         struct bkey_packed *src, *dst, *n;
1217         struct bset *i;
1218
1219         BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1220
1221         bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
1222
1223         while (!bch2_keylist_empty(keys)) {
1224                 k = bch2_keylist_front(keys);
1225
1226                 bch2_insert_fixup_btree_ptr(as, b, iter, k, &node_iter);
1227                 bch2_keylist_pop_front(keys);
1228         }
1229
1230         /*
1231          * We can't tolerate whiteouts here - with whiteouts there can be
1232          * duplicate keys, and it would be rather bad if we picked a duplicate
1233          * for the pivot:
1234          */
1235         i = btree_bset_first(b);
1236         src = dst = i->start;
1237         while (src != vstruct_last(i)) {
1238                 n = bkey_next(src);
1239                 if (!bkey_deleted(src)) {
1240                         memmove_u64s_down(dst, src, src->u64s);
1241                         dst = bkey_next(dst);
1242                 }
1243                 src = n;
1244         }
1245
1246         /* Also clear out the unwritten whiteouts area: */
1247         b->whiteout_u64s = 0;
1248
1249         i->u64s = cpu_to_le16((u64 *) dst - i->_data);
1250         set_btree_bset_end(b, b->set);
1251
1252         BUG_ON(b->nsets != 1 ||
1253                b->nr.live_u64s != le16_to_cpu(btree_bset_first(b)->u64s));
1254
1255         btree_node_interior_verify(as->c, b);
1256 }
1257
1258 static void btree_split(struct btree_update *as, struct btree *b,
1259                         struct btree_iter *iter, struct keylist *keys,
1260                         unsigned flags)
1261 {
1262         struct bch_fs *c = as->c;
1263         struct btree *parent = btree_node_parent(iter, b);
1264         struct btree *n1, *n2 = NULL, *n3 = NULL;
1265         u64 start_time = local_clock();
1266
1267         BUG_ON(!parent && (b != btree_node_root(c, b)));
1268         BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->c.level));
1269
1270         bch2_btree_interior_update_will_free_node(as, b);
1271
1272         n1 = bch2_btree_node_alloc_replacement(as, b);
1273         bch2_btree_update_add_new_node(as, n1);
1274
1275         if (keys)
1276                 btree_split_insert_keys(as, n1, iter, keys);
1277
1278         if (bset_u64s(&n1->set[0]) > BTREE_SPLIT_THRESHOLD(c)) {
1279                 trace_btree_split(c, b);
1280
1281                 n2 = __btree_split_node(as, n1, iter);
1282
1283                 bch2_btree_build_aux_trees(n2);
1284                 bch2_btree_build_aux_trees(n1);
1285                 six_unlock_write(&n2->c.lock);
1286                 six_unlock_write(&n1->c.lock);
1287
1288                 bch2_btree_node_write(c, n2, SIX_LOCK_intent);
1289
1290                 /*
1291                  * Note that on recursive parent_keys == keys, so we
1292                  * can't start adding new keys to parent_keys before emptying it
1293                  * out (which we did with btree_split_insert_keys() above)
1294                  */
1295                 bch2_keylist_add(&as->parent_keys, &n1->key);
1296                 bch2_keylist_add(&as->parent_keys, &n2->key);
1297
1298                 if (!parent) {
1299                         /* Depth increases, make a new root */
1300                         n3 = __btree_root_alloc(as, b->c.level + 1);
1301
1302                         n3->sib_u64s[0] = U16_MAX;
1303                         n3->sib_u64s[1] = U16_MAX;
1304
1305                         btree_split_insert_keys(as, n3, iter, &as->parent_keys);
1306
1307                         bch2_btree_node_write(c, n3, SIX_LOCK_intent);
1308                 }
1309         } else {
1310                 trace_btree_compact(c, b);
1311
1312                 bch2_btree_build_aux_trees(n1);
1313                 six_unlock_write(&n1->c.lock);
1314
1315                 if (parent)
1316                         bch2_keylist_add(&as->parent_keys, &n1->key);
1317         }
1318
1319         bch2_btree_node_write(c, n1, SIX_LOCK_intent);
1320
1321         /* New nodes all written, now make them visible: */
1322
1323         if (parent) {
1324                 /* Split a non root node */
1325                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1326         } else if (n3) {
1327                 bch2_btree_set_root(as, n3, iter);
1328         } else {
1329                 /* Root filled up but didn't need to be split */
1330                 bch2_btree_set_root(as, n1, iter);
1331         }
1332
1333         bch2_btree_update_get_open_buckets(as, n1);
1334         if (n2)
1335                 bch2_btree_update_get_open_buckets(as, n2);
1336         if (n3)
1337                 bch2_btree_update_get_open_buckets(as, n3);
1338
1339         /* Successful split, update the iterator to point to the new nodes: */
1340
1341         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1342         bch2_btree_iter_node_drop(iter, b);
1343         if (n3)
1344                 bch2_btree_iter_node_replace(iter, n3);
1345         if (n2)
1346                 bch2_btree_iter_node_replace(iter, n2);
1347         bch2_btree_iter_node_replace(iter, n1);
1348
1349         /*
1350          * The old node must be freed (in memory) _before_ unlocking the new
1351          * nodes - else another thread could re-acquire a read lock on the old
1352          * node after another thread has locked and updated the new node, thus
1353          * seeing stale data:
1354          */
1355         bch2_btree_node_free_inmem(c, b, iter);
1356
1357         if (n3)
1358                 six_unlock_intent(&n3->c.lock);
1359         if (n2)
1360                 six_unlock_intent(&n2->c.lock);
1361         six_unlock_intent(&n1->c.lock);
1362
1363         bch2_btree_trans_verify_locks(iter->trans);
1364
1365         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_split],
1366                                start_time);
1367 }
1368
1369 static void
1370 bch2_btree_insert_keys_interior(struct btree_update *as, struct btree *b,
1371                                 struct btree_iter *iter, struct keylist *keys)
1372 {
1373         struct btree_iter *linked;
1374         struct btree_node_iter node_iter;
1375         struct bkey_i *insert = bch2_keylist_front(keys);
1376         struct bkey_packed *k;
1377
1378         /* Don't screw up @iter's position: */
1379         node_iter = iter->l[b->c.level].iter;
1380
1381         /*
1382          * btree_split(), btree_gc_coalesce() will insert keys before
1383          * the iterator's current position - they know the keys go in
1384          * the node the iterator points to:
1385          */
1386         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1387                (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1388                 ;
1389
1390         for_each_keylist_key(keys, insert)
1391                 bch2_insert_fixup_btree_ptr(as, b, iter, insert, &node_iter);
1392
1393         btree_update_updated_node(as, b);
1394
1395         trans_for_each_iter_with_node(iter->trans, b, linked)
1396                 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1397
1398         bch2_btree_trans_verify_iters(iter->trans, b);
1399 }
1400
1401 /**
1402  * bch_btree_insert_node - insert bkeys into a given btree node
1403  *
1404  * @iter:               btree iterator
1405  * @keys:               list of keys to insert
1406  * @hook:               insert callback
1407  * @persistent:         if not null, @persistent will wait on journal write
1408  *
1409  * Inserts as many keys as it can into a given btree node, splitting it if full.
1410  * If a split occurred, this function will return early. This can only happen
1411  * for leaf nodes -- inserts into interior nodes have to be atomic.
1412  */
1413 void bch2_btree_insert_node(struct btree_update *as, struct btree *b,
1414                             struct btree_iter *iter, struct keylist *keys,
1415                             unsigned flags)
1416 {
1417         struct bch_fs *c = as->c;
1418         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1419         int old_live_u64s = b->nr.live_u64s;
1420         int live_u64s_added, u64s_added;
1421
1422         BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->c.level));
1423         BUG_ON(!b->c.level);
1424         BUG_ON(!as || as->b);
1425         bch2_verify_keylist_sorted(keys);
1426
1427         bch2_btree_node_lock_for_insert(c, b, iter);
1428
1429         if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1430                 bch2_btree_node_unlock_write(b, iter);
1431                 goto split;
1432         }
1433
1434         btree_node_interior_verify(c, b);
1435
1436         bch2_btree_insert_keys_interior(as, b, iter, keys);
1437
1438         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1439         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1440
1441         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1442                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1443         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1444                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1445
1446         if (u64s_added > live_u64s_added &&
1447             bch2_maybe_compact_whiteouts(c, b))
1448                 bch2_btree_iter_reinit_node(iter, b);
1449
1450         bch2_btree_node_unlock_write(b, iter);
1451
1452         btree_node_interior_verify(c, b);
1453
1454         /*
1455          * when called from the btree_split path the new nodes aren't added to
1456          * the btree iterator yet, so the merge path's unlock/wait/relock dance
1457          * won't work:
1458          */
1459         bch2_foreground_maybe_merge(c, iter, b->c.level,
1460                                     flags|BTREE_INSERT_NOUNLOCK);
1461         return;
1462 split:
1463         btree_split(as, b, iter, keys, flags);
1464 }
1465
1466 int bch2_btree_split_leaf(struct bch_fs *c, struct btree_iter *iter,
1467                           unsigned flags)
1468 {
1469         struct btree_trans *trans = iter->trans;
1470         struct btree *b = iter_l(iter)->b;
1471         struct btree_update *as;
1472         struct closure cl;
1473         int ret = 0;
1474
1475         closure_init_stack(&cl);
1476
1477         /* Hack, because gc and splitting nodes doesn't mix yet: */
1478         if (!(flags & BTREE_INSERT_GC_LOCK_HELD) &&
1479             !down_read_trylock(&c->gc_lock)) {
1480                 if (flags & BTREE_INSERT_NOUNLOCK) {
1481                         trace_transaction_restart_ip(trans->ip, _THIS_IP_);
1482                         return -EINTR;
1483                 }
1484
1485                 bch2_trans_unlock(trans);
1486                 down_read(&c->gc_lock);
1487
1488                 if (!bch2_trans_relock(trans))
1489                         ret = -EINTR;
1490         }
1491
1492         /*
1493          * XXX: figure out how far we might need to split,
1494          * instead of locking/reserving all the way to the root:
1495          */
1496         if (!bch2_btree_iter_upgrade(iter, U8_MAX)) {
1497                 trace_trans_restart_iter_upgrade(trans->ip);
1498                 ret = -EINTR;
1499                 goto out;
1500         }
1501
1502         as = bch2_btree_update_start(trans, iter->btree_id,
1503                 btree_update_reserve_required(c, b), flags,
1504                 !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1505         if (IS_ERR(as)) {
1506                 ret = PTR_ERR(as);
1507                 if (ret == -EAGAIN) {
1508                         BUG_ON(flags & BTREE_INSERT_NOUNLOCK);
1509                         bch2_trans_unlock(trans);
1510                         ret = -EINTR;
1511
1512                         trace_transaction_restart_ip(trans->ip, _THIS_IP_);
1513                 }
1514                 goto out;
1515         }
1516
1517         btree_split(as, b, iter, NULL, flags);
1518         bch2_btree_update_done(as);
1519
1520         /*
1521          * We haven't successfully inserted yet, so don't downgrade all the way
1522          * back to read locks;
1523          */
1524         __bch2_btree_iter_downgrade(iter, 1);
1525 out:
1526         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1527                 up_read(&c->gc_lock);
1528         closure_sync(&cl);
1529         return ret;
1530 }
1531
1532 void __bch2_foreground_maybe_merge(struct bch_fs *c,
1533                                    struct btree_iter *iter,
1534                                    unsigned level,
1535                                    unsigned flags,
1536                                    enum btree_node_sibling sib)
1537 {
1538         struct btree_trans *trans = iter->trans;
1539         struct btree_update *as;
1540         struct bkey_format_state new_s;
1541         struct bkey_format new_f;
1542         struct bkey_i delete;
1543         struct btree *b, *m, *n, *prev, *next, *parent;
1544         struct closure cl;
1545         size_t sib_u64s;
1546         int ret = 0;
1547
1548         BUG_ON(!btree_node_locked(iter, level));
1549
1550         closure_init_stack(&cl);
1551 retry:
1552         BUG_ON(!btree_node_locked(iter, level));
1553
1554         b = iter->l[level].b;
1555
1556         parent = btree_node_parent(iter, b);
1557         if (!parent)
1558                 goto out;
1559
1560         if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c))
1561                 goto out;
1562
1563         /* XXX: can't be holding read locks */
1564         m = bch2_btree_node_get_sibling(c, iter, b, sib);
1565         if (IS_ERR(m)) {
1566                 ret = PTR_ERR(m);
1567                 goto err;
1568         }
1569
1570         /* NULL means no sibling: */
1571         if (!m) {
1572                 b->sib_u64s[sib] = U16_MAX;
1573                 goto out;
1574         }
1575
1576         if (sib == btree_prev_sib) {
1577                 prev = m;
1578                 next = b;
1579         } else {
1580                 prev = b;
1581                 next = m;
1582         }
1583
1584         bch2_bkey_format_init(&new_s);
1585         bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
1586         __bch2_btree_calc_format(&new_s, prev);
1587         __bch2_btree_calc_format(&new_s, next);
1588         bch2_bkey_format_add_pos(&new_s, next->data->max_key);
1589         new_f = bch2_bkey_format_done(&new_s);
1590
1591         sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1592                 btree_node_u64s_with_format(m, &new_f);
1593
1594         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1595                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1596                 sib_u64s /= 2;
1597                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1598         }
1599
1600         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1601         b->sib_u64s[sib] = sib_u64s;
1602
1603         if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c)) {
1604                 six_unlock_intent(&m->c.lock);
1605                 goto out;
1606         }
1607
1608         /* We're changing btree topology, doesn't mix with gc: */
1609         if (!(flags & BTREE_INSERT_GC_LOCK_HELD) &&
1610             !down_read_trylock(&c->gc_lock))
1611                 goto err_cycle_gc_lock;
1612
1613         if (!bch2_btree_iter_upgrade(iter, U8_MAX)) {
1614                 ret = -EINTR;
1615                 goto err_unlock;
1616         }
1617
1618         as = bch2_btree_update_start(trans, iter->btree_id,
1619                          btree_update_reserve_required(c, parent) + 1,
1620                          flags|
1621                          BTREE_INSERT_NOFAIL|
1622                          BTREE_INSERT_USE_RESERVE,
1623                          !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1624         if (IS_ERR(as)) {
1625                 ret = PTR_ERR(as);
1626                 goto err_unlock;
1627         }
1628
1629         trace_btree_merge(c, b);
1630
1631         bch2_btree_interior_update_will_free_node(as, b);
1632         bch2_btree_interior_update_will_free_node(as, m);
1633
1634         n = bch2_btree_node_alloc(as, b->c.level);
1635         bch2_btree_update_add_new_node(as, n);
1636
1637         btree_set_min(n, prev->data->min_key);
1638         btree_set_max(n, next->data->max_key);
1639         n->data->format         = new_f;
1640
1641         btree_node_set_format(n, new_f);
1642
1643         bch2_btree_sort_into(c, n, prev);
1644         bch2_btree_sort_into(c, n, next);
1645
1646         bch2_btree_build_aux_trees(n);
1647         six_unlock_write(&n->c.lock);
1648
1649         bkey_init(&delete.k);
1650         delete.k.p = prev->key.k.p;
1651         bch2_keylist_add(&as->parent_keys, &delete);
1652         bch2_keylist_add(&as->parent_keys, &n->key);
1653
1654         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1655
1656         bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1657
1658         bch2_btree_update_get_open_buckets(as, n);
1659
1660         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1661         bch2_btree_iter_node_drop(iter, b);
1662         bch2_btree_iter_node_drop(iter, m);
1663
1664         bch2_btree_iter_node_replace(iter, n);
1665
1666         bch2_btree_trans_verify_iters(trans, n);
1667
1668         bch2_btree_node_free_inmem(c, b, iter);
1669         bch2_btree_node_free_inmem(c, m, iter);
1670
1671         six_unlock_intent(&n->c.lock);
1672
1673         bch2_btree_update_done(as);
1674
1675         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1676                 up_read(&c->gc_lock);
1677 out:
1678         bch2_btree_trans_verify_locks(trans);
1679
1680         /*
1681          * Don't downgrade locks here: we're called after successful insert,
1682          * and the caller will downgrade locks after a successful insert
1683          * anyways (in case e.g. a split was required first)
1684          *
1685          * And we're also called when inserting into interior nodes in the
1686          * split path, and downgrading to read locks in there is potentially
1687          * confusing:
1688          */
1689         closure_sync(&cl);
1690         return;
1691
1692 err_cycle_gc_lock:
1693         six_unlock_intent(&m->c.lock);
1694
1695         if (flags & BTREE_INSERT_NOUNLOCK)
1696                 goto out;
1697
1698         bch2_trans_unlock(trans);
1699
1700         down_read(&c->gc_lock);
1701         up_read(&c->gc_lock);
1702         ret = -EINTR;
1703         goto err;
1704
1705 err_unlock:
1706         six_unlock_intent(&m->c.lock);
1707         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1708                 up_read(&c->gc_lock);
1709 err:
1710         BUG_ON(ret == -EAGAIN && (flags & BTREE_INSERT_NOUNLOCK));
1711
1712         if ((ret == -EAGAIN || ret == -EINTR) &&
1713             !(flags & BTREE_INSERT_NOUNLOCK)) {
1714                 bch2_trans_unlock(trans);
1715                 closure_sync(&cl);
1716                 ret = bch2_btree_iter_traverse(iter);
1717                 if (ret)
1718                         goto out;
1719
1720                 goto retry;
1721         }
1722
1723         goto out;
1724 }
1725
1726 static int __btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
1727                                 struct btree *b, unsigned flags,
1728                                 struct closure *cl)
1729 {
1730         struct btree *n, *parent = btree_node_parent(iter, b);
1731         struct btree_update *as;
1732
1733         as = bch2_btree_update_start(iter->trans, iter->btree_id,
1734                 (parent
1735                  ? btree_update_reserve_required(c, parent)
1736                  : 0) + 1,
1737                 flags, cl);
1738         if (IS_ERR(as)) {
1739                 trace_btree_gc_rewrite_node_fail(c, b);
1740                 return PTR_ERR(as);
1741         }
1742
1743         bch2_btree_interior_update_will_free_node(as, b);
1744
1745         n = bch2_btree_node_alloc_replacement(as, b);
1746         bch2_btree_update_add_new_node(as, n);
1747
1748         bch2_btree_build_aux_trees(n);
1749         six_unlock_write(&n->c.lock);
1750
1751         trace_btree_gc_rewrite_node(c, b);
1752
1753         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1754
1755         if (parent) {
1756                 bch2_keylist_add(&as->parent_keys, &n->key);
1757                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1758         } else {
1759                 bch2_btree_set_root(as, n, iter);
1760         }
1761
1762         bch2_btree_update_get_open_buckets(as, n);
1763
1764         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1765         bch2_btree_iter_node_drop(iter, b);
1766         bch2_btree_iter_node_replace(iter, n);
1767         bch2_btree_node_free_inmem(c, b, iter);
1768         six_unlock_intent(&n->c.lock);
1769
1770         bch2_btree_update_done(as);
1771         return 0;
1772 }
1773
1774 /**
1775  * bch_btree_node_rewrite - Rewrite/move a btree node
1776  *
1777  * Returns 0 on success, -EINTR or -EAGAIN on failure (i.e.
1778  * btree_check_reserve() has to wait)
1779  */
1780 int bch2_btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
1781                             __le64 seq, unsigned flags)
1782 {
1783         struct btree_trans *trans = iter->trans;
1784         struct closure cl;
1785         struct btree *b;
1786         int ret;
1787
1788         flags |= BTREE_INSERT_NOFAIL;
1789
1790         closure_init_stack(&cl);
1791
1792         bch2_btree_iter_upgrade(iter, U8_MAX);
1793
1794         if (!(flags & BTREE_INSERT_GC_LOCK_HELD)) {
1795                 if (!down_read_trylock(&c->gc_lock)) {
1796                         bch2_trans_unlock(trans);
1797                         down_read(&c->gc_lock);
1798                 }
1799         }
1800
1801         while (1) {
1802                 ret = bch2_btree_iter_traverse(iter);
1803                 if (ret)
1804                         break;
1805
1806                 b = bch2_btree_iter_peek_node(iter);
1807                 if (!b || b->data->keys.seq != seq)
1808                         break;
1809
1810                 ret = __btree_node_rewrite(c, iter, b, flags, &cl);
1811                 if (ret != -EAGAIN &&
1812                     ret != -EINTR)
1813                         break;
1814
1815                 bch2_trans_unlock(trans);
1816                 closure_sync(&cl);
1817         }
1818
1819         bch2_btree_iter_downgrade(iter);
1820
1821         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1822                 up_read(&c->gc_lock);
1823
1824         closure_sync(&cl);
1825         return ret;
1826 }
1827
1828 static void __bch2_btree_node_update_key(struct bch_fs *c,
1829                                          struct btree_update *as,
1830                                          struct btree_iter *iter,
1831                                          struct btree *b, struct btree *new_hash,
1832                                          struct bkey_i *new_key)
1833 {
1834         struct btree *parent;
1835         int ret;
1836
1837         btree_update_will_delete_key(as, &b->key);
1838         btree_update_will_add_key(as, new_key);
1839
1840         parent = btree_node_parent(iter, b);
1841         if (parent) {
1842                 if (new_hash) {
1843                         bkey_copy(&new_hash->key, new_key);
1844                         ret = bch2_btree_node_hash_insert(&c->btree_cache,
1845                                         new_hash, b->c.level, b->c.btree_id);
1846                         BUG_ON(ret);
1847                 }
1848
1849                 bch2_keylist_add(&as->parent_keys, new_key);
1850                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, 0);
1851
1852                 if (new_hash) {
1853                         mutex_lock(&c->btree_cache.lock);
1854                         bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
1855
1856                         bch2_btree_node_hash_remove(&c->btree_cache, b);
1857
1858                         bkey_copy(&b->key, new_key);
1859                         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1860                         BUG_ON(ret);
1861                         mutex_unlock(&c->btree_cache.lock);
1862                 } else {
1863                         bkey_copy(&b->key, new_key);
1864                 }
1865         } else {
1866                 BUG_ON(btree_node_root(c, b) != b);
1867
1868                 bch2_btree_node_lock_write(b, iter);
1869                 bkey_copy(&b->key, new_key);
1870
1871                 if (btree_ptr_hash_val(&b->key) != b->hash_val) {
1872                         mutex_lock(&c->btree_cache.lock);
1873                         bch2_btree_node_hash_remove(&c->btree_cache, b);
1874
1875                         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1876                         BUG_ON(ret);
1877                         mutex_unlock(&c->btree_cache.lock);
1878                 }
1879
1880                 btree_update_updated_root(as, b);
1881                 bch2_btree_node_unlock_write(b, iter);
1882         }
1883
1884         bch2_btree_update_done(as);
1885 }
1886
1887 int bch2_btree_node_update_key(struct bch_fs *c, struct btree_iter *iter,
1888                                struct btree *b,
1889                                struct bkey_i *new_key)
1890 {
1891         struct btree *parent = btree_node_parent(iter, b);
1892         struct btree_update *as = NULL;
1893         struct btree *new_hash = NULL;
1894         struct closure cl;
1895         int ret;
1896
1897         closure_init_stack(&cl);
1898
1899         if (!bch2_btree_iter_upgrade(iter, U8_MAX))
1900                 return -EINTR;
1901
1902         if (!down_read_trylock(&c->gc_lock)) {
1903                 bch2_trans_unlock(iter->trans);
1904                 down_read(&c->gc_lock);
1905
1906                 if (!bch2_trans_relock(iter->trans)) {
1907                         ret = -EINTR;
1908                         goto err;
1909                 }
1910         }
1911
1912         /*
1913          * check btree_ptr_hash_val() after @b is locked by
1914          * btree_iter_traverse():
1915          */
1916         if (btree_ptr_hash_val(new_key) != b->hash_val) {
1917                 /* bch2_btree_reserve_get will unlock */
1918                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1919                 if (ret) {
1920                         bch2_trans_unlock(iter->trans);
1921                         up_read(&c->gc_lock);
1922                         closure_sync(&cl);
1923                         down_read(&c->gc_lock);
1924
1925                         if (!bch2_trans_relock(iter->trans)) {
1926                                 ret = -EINTR;
1927                                 goto err;
1928                         }
1929                 }
1930
1931                 new_hash = bch2_btree_node_mem_alloc(c);
1932         }
1933 retry:
1934         as = bch2_btree_update_start(iter->trans, iter->btree_id,
1935                 parent ? btree_update_reserve_required(c, parent) : 0,
1936                 BTREE_INSERT_NOFAIL, &cl);
1937
1938         if (IS_ERR(as)) {
1939                 ret = PTR_ERR(as);
1940                 if (ret == -EAGAIN)
1941                         ret = -EINTR;
1942
1943                 if (ret == -EINTR) {
1944                         bch2_trans_unlock(iter->trans);
1945                         up_read(&c->gc_lock);
1946                         closure_sync(&cl);
1947                         down_read(&c->gc_lock);
1948
1949                         if (bch2_trans_relock(iter->trans))
1950                                 goto retry;
1951                 }
1952
1953                 goto err;
1954         }
1955
1956         ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(new_key));
1957         if (ret)
1958                 goto err_free_update;
1959
1960         __bch2_btree_node_update_key(c, as, iter, b, new_hash, new_key);
1961
1962         bch2_btree_iter_downgrade(iter);
1963 err:
1964         if (new_hash) {
1965                 mutex_lock(&c->btree_cache.lock);
1966                 list_move(&new_hash->list, &c->btree_cache.freeable);
1967                 mutex_unlock(&c->btree_cache.lock);
1968
1969                 six_unlock_write(&new_hash->c.lock);
1970                 six_unlock_intent(&new_hash->c.lock);
1971         }
1972         up_read(&c->gc_lock);
1973         closure_sync(&cl);
1974         return ret;
1975 err_free_update:
1976         bch2_btree_update_free(as);
1977         goto err;
1978 }
1979
1980 /* Init code: */
1981
1982 /*
1983  * Only for filesystem bringup, when first reading the btree roots or allocating
1984  * btree roots when initializing a new filesystem:
1985  */
1986 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
1987 {
1988         BUG_ON(btree_node_root(c, b));
1989
1990         bch2_btree_set_root_inmem(c, b);
1991 }
1992
1993 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
1994 {
1995         struct closure cl;
1996         struct btree *b;
1997         int ret;
1998
1999         closure_init_stack(&cl);
2000
2001         do {
2002                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2003                 closure_sync(&cl);
2004         } while (ret);
2005
2006         b = bch2_btree_node_mem_alloc(c);
2007         bch2_btree_cache_cannibalize_unlock(c);
2008
2009         set_btree_node_fake(b);
2010         set_btree_node_need_rewrite(b);
2011         b->c.level      = 0;
2012         b->c.btree_id   = id;
2013
2014         bkey_btree_ptr_init(&b->key);
2015         b->key.k.p = POS_MAX;
2016         *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2017
2018         bch2_bset_init_first(b, &b->data->keys);
2019         bch2_btree_build_aux_trees(b);
2020
2021         b->data->flags = 0;
2022         btree_set_min(b, POS_MIN);
2023         btree_set_max(b, POS_MAX);
2024         b->data->format = bch2_btree_calc_format(b);
2025         btree_node_set_format(b, b->data->format);
2026
2027         ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2028                                           b->c.level, b->c.btree_id);
2029         BUG_ON(ret);
2030
2031         bch2_btree_set_root_inmem(c, b);
2032
2033         six_unlock_write(&b->c.lock);
2034         six_unlock_intent(&b->c.lock);
2035 }
2036
2037 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2038 {
2039         struct btree_update *as;
2040
2041         mutex_lock(&c->btree_interior_update_lock);
2042         list_for_each_entry(as, &c->btree_interior_update_list, list)
2043                 pr_buf(out, "%p m %u w %u r %u j %llu\n",
2044                        as,
2045                        as->mode,
2046                        as->nodes_written,
2047                        atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2048                        as->journal.seq);
2049         mutex_unlock(&c->btree_interior_update_lock);
2050 }
2051
2052 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *c)
2053 {
2054         size_t ret = 0;
2055         struct list_head *i;
2056
2057         mutex_lock(&c->btree_interior_update_lock);
2058         list_for_each(i, &c->btree_interior_update_list)
2059                 ret++;
2060         mutex_unlock(&c->btree_interior_update_lock);
2061
2062         return ret;
2063 }
2064
2065 void bch2_journal_entries_to_btree_roots(struct bch_fs *c, struct jset *jset)
2066 {
2067         struct btree_root *r;
2068         struct jset_entry *entry;
2069
2070         mutex_lock(&c->btree_root_lock);
2071
2072         vstruct_for_each(jset, entry)
2073                 if (entry->type == BCH_JSET_ENTRY_btree_root) {
2074                         r = &c->btree_roots[entry->btree_id];
2075                         r->level = entry->level;
2076                         r->alive = true;
2077                         bkey_copy(&r->key, &entry->start[0]);
2078                 }
2079
2080         mutex_unlock(&c->btree_root_lock);
2081 }
2082
2083 struct jset_entry *
2084 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2085                                     struct jset_entry *start,
2086                                     struct jset_entry *end)
2087 {
2088         struct jset_entry *entry;
2089         unsigned long have = 0;
2090         unsigned i;
2091
2092         for (entry = start; entry < end; entry = vstruct_next(entry))
2093                 if (entry->type == BCH_JSET_ENTRY_btree_root)
2094                         __set_bit(entry->btree_id, &have);
2095
2096         mutex_lock(&c->btree_root_lock);
2097
2098         for (i = 0; i < BTREE_ID_NR; i++)
2099                 if (c->btree_roots[i].alive && !test_bit(i, &have)) {
2100                         journal_entry_set(end,
2101                                           BCH_JSET_ENTRY_btree_root,
2102                                           i, c->btree_roots[i].level,
2103                                           &c->btree_roots[i].key,
2104                                           c->btree_roots[i].key.u64s);
2105                         end = vstruct_next(end);
2106                 }
2107
2108         mutex_unlock(&c->btree_root_lock);
2109
2110         return end;
2111 }
2112
2113 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2114 {
2115         if (c->btree_interior_update_worker)
2116                 destroy_workqueue(c->btree_interior_update_worker);
2117         mempool_exit(&c->btree_interior_update_pool);
2118 }
2119
2120 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2121 {
2122         mutex_init(&c->btree_reserve_cache_lock);
2123         INIT_LIST_HEAD(&c->btree_interior_update_list);
2124         INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2125         mutex_init(&c->btree_interior_update_lock);
2126         INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2127
2128         c->btree_interior_update_worker =
2129                 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2130         if (!c->btree_interior_update_worker)
2131                 return -ENOMEM;
2132
2133         return mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2134                                          sizeof(struct btree_update));
2135 }