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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "alloc_foreground.h"
   5 #include "bkey_methods.h"
   6 #include "btree_cache.h"
   7 #include "btree_gc.h"
   8 #include "btree_update.h"
   9 #include "btree_update_interior.h"
  10 #include "btree_io.h"
  11 #include "btree_iter.h"
  12 #include "btree_locking.h"
  13 #include "buckets.h"
  14 #include "error.h"
  15 #include "extents.h"
  16 #include "journal.h"
  17 #include "journal_reclaim.h"
  18 #include "keylist.h"
  19 #include "replicas.h"
  20 #include "super-io.h"
  21
  22 #include <linux/random.h>
  23 #include <trace/events/bcachefs.h>
  24
  25 /* Debug code: */
  26
  27 /*
  28  * Verify that child nodes correctly span parent node's range:
  29  */
  30 static void btree_node_interior_verify(struct bch_fs *c, struct btree *b)
  31 {
  32 #ifdef CONFIG_BCACHEFS_DEBUG
  33         struct bpos next_node = b->data->min_key;
  34         struct btree_node_iter iter;
  35         struct bkey_s_c k;
  36         struct bkey_s_c_btree_ptr_v2 bp;
  37         struct bkey unpacked;
  38
  39         BUG_ON(!b->c.level);
  40
  41         if (!test_bit(BCH_FS_BTREE_INTERIOR_REPLAY_DONE, &c->flags))
  42                 return;
  43
  44         bch2_btree_node_iter_init_from_start(&iter, b);
  45
  46         while (1) {
  47                 k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked);
  48                 if (k.k->type != KEY_TYPE_btree_ptr_v2)
  49                         break;
  50                 bp = bkey_s_c_to_btree_ptr_v2(k);
  51
  52                 if (bkey_cmp(next_node, bp.v->min_key)) {
  53                         bch2_dump_btree_node(c, b);
  54                         panic("expected next min_key %llu:%llu got %llu:%llu\n",
  55                               next_node.inode,
  56                               next_node.offset,
  57                               bp.v->min_key.inode,
  58                               bp.v->min_key.offset);
  59                 }
  60
  61                 bch2_btree_node_iter_advance(&iter, b);
  62
  63                 if (bch2_btree_node_iter_end(&iter)) {
  64
  65                         if (bkey_cmp(k.k->p, b->key.k.p)) {
  66                                 bch2_dump_btree_node(c, b);
  67                                 panic("expected end %llu:%llu got %llu:%llu\n",
  68                                       b->key.k.p.inode,
  69                                       b->key.k.p.offset,
  70                                       k.k->p.inode,
  71                                       k.k->p.offset);
  72                         }
  73                         break;
  74                 }
  75
  76                 next_node = bkey_successor(k.k->p);
  77         }
  78 #endif
  79 }
  80
  81 /* Calculate ideal packed bkey format for new btree nodes: */
  82
  83 void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b)
  84 {
  85         struct bkey_packed *k;
  86         struct bset_tree *t;
  87         struct bkey uk;
  88
  89         bch2_bkey_format_add_pos(s, b->data->min_key);
  90
  91         for_each_bset(b, t)
  92                 bset_tree_for_each_key(b, t, k)
  93                         if (!bkey_deleted(k)) {
  94                                 uk = bkey_unpack_key(b, k);
  95                                 bch2_bkey_format_add_key(s, &uk);
  96                         }
  97 }
  98
  99 static struct bkey_format bch2_btree_calc_format(struct btree *b)
 100 {
 101         struct bkey_format_state s;
 102
 103         bch2_bkey_format_init(&s);
 104         __bch2_btree_calc_format(&s, b);
 105
 106         return bch2_bkey_format_done(&s);
 107 }
 108
 109 static size_t btree_node_u64s_with_format(struct btree *b,
 110                                           struct bkey_format *new_f)
 111 {
 112         struct bkey_format *old_f = &b->format;
 113
 114         /* stupid integer promotion rules */
 115         ssize_t delta =
 116             (((int) new_f->key_u64s - old_f->key_u64s) *
 117              (int) b->nr.packed_keys) +
 118             (((int) new_f->key_u64s - BKEY_U64s) *
 119              (int) b->nr.unpacked_keys);
 120
 121         BUG_ON(delta + b->nr.live_u64s < 0);
 122
 123         return b->nr.live_u64s + delta;
 124 }
 125
 126 /**
 127  * btree_node_format_fits - check if we could rewrite node with a new format
 128  *
 129  * This assumes all keys can pack with the new format -- it just checks if
 130  * the re-packed keys would fit inside the node itself.
 131  */
 132 bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
 133                                  struct bkey_format *new_f)
 134 {
 135         size_t u64s = btree_node_u64s_with_format(b, new_f);
 136
 137         return __vstruct_bytes(struct btree_node, u64s) < btree_bytes(c);
 138 }
 139
 140 /* Btree node freeing/allocation: */
 141
 142 static void __btree_node_free(struct bch_fs *c, struct btree *b)
 143 {
 144         trace_btree_node_free(c, b);
 145
 146         BUG_ON(btree_node_dirty(b));
 147         BUG_ON(btree_node_need_write(b));
 148         BUG_ON(b == btree_node_root(c, b));
 149         BUG_ON(b->ob.nr);
 150         BUG_ON(!list_empty(&b->write_blocked));
 151         BUG_ON(b->will_make_reachable);
 152
 153         clear_btree_node_noevict(b);
 154
 155         bch2_btree_node_hash_remove(&c->btree_cache, b);
 156
 157         mutex_lock(&c->btree_cache.lock);
 158         list_move(&b->list, &c->btree_cache.freeable);
 159         mutex_unlock(&c->btree_cache.lock);
 160 }
 161
 162 void bch2_btree_node_free_never_inserted(struct bch_fs *c, struct btree *b)
 163 {
 164         struct open_buckets ob = b->ob;
 165
 166         b->ob.nr = 0;
 167
 168         clear_btree_node_dirty(c, b);
 169
 170         btree_node_lock_type(c, b, SIX_LOCK_write);
 171         __btree_node_free(c, b);
 172         six_unlock_write(&b->c.lock);
 173
 174         bch2_open_buckets_put(c, &ob);
 175 }
 176
 177 void bch2_btree_node_free_inmem(struct bch_fs *c, struct btree *b,
 178                                 struct btree_iter *iter)
 179 {
 180         struct btree_iter *linked;
 181
 182         trans_for_each_iter(iter->trans, linked)
 183                 BUG_ON(linked->l[b->c.level].b == b);
 184
 185         six_lock_write(&b->c.lock, NULL, NULL);
 186         __btree_node_free(c, b);
 187         six_unlock_write(&b->c.lock);
 188         six_unlock_intent(&b->c.lock);
 189 }
 190
 191 static struct btree *__bch2_btree_node_alloc(struct bch_fs *c,
 192                                              struct disk_reservation *res,
 193                                              struct closure *cl,
 194                                              unsigned flags)
 195 {
 196         struct write_point *wp;
 197         struct btree *b;
 198         __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
 199         struct open_buckets ob = { .nr = 0 };
 200         struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
 201         unsigned nr_reserve;
 202         enum alloc_reserve alloc_reserve;
 203
 204         if (flags & BTREE_INSERT_USE_RESERVE) {
 205                 nr_reserve      = 0;
 206                 alloc_reserve   = RESERVE_BTREE_MOVINGGC;
 207         } else {
 208                 nr_reserve      = BTREE_NODE_RESERVE;
 209                 alloc_reserve   = RESERVE_BTREE;
 210         }
 211
 212         mutex_lock(&c->btree_reserve_cache_lock);
 213         if (c->btree_reserve_cache_nr > nr_reserve) {
 214                 struct btree_alloc *a =
 215                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
 216
 217                 ob = a->ob;
 218                 bkey_copy(&tmp.k, &a->k);
 219                 mutex_unlock(&c->btree_reserve_cache_lock);
 220                 goto mem_alloc;
 221         }
 222         mutex_unlock(&c->btree_reserve_cache_lock);
 223
 224 retry:
 225         wp = bch2_alloc_sectors_start(c,
 226                                       c->opts.metadata_target ?:
 227                                       c->opts.foreground_target,
 228                                       0,
 229                                       writepoint_ptr(&c->btree_write_point),
 230                                       &devs_have,
 231                                       res->nr_replicas,
 232                                       c->opts.metadata_replicas_required,
 233                                       alloc_reserve, 0, cl);
 234         if (IS_ERR(wp))
 235                 return ERR_CAST(wp);
 236
 237         if (wp->sectors_free < c->opts.btree_node_size) {
 238                 struct open_bucket *ob;
 239                 unsigned i;
 240
 241                 open_bucket_for_each(c, &wp->ptrs, ob, i)
 242                         if (ob->sectors_free < c->opts.btree_node_size)
 243                                 ob->sectors_free = 0;
 244
 245                 bch2_alloc_sectors_done(c, wp);
 246                 goto retry;
 247         }
 248
 249         if (c->sb.features & (1ULL << BCH_FEATURE_btree_ptr_v2))
 250                 bkey_btree_ptr_v2_init(&tmp.k);
 251         else
 252                 bkey_btree_ptr_init(&tmp.k);
 253
 254         bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, c->opts.btree_node_size);
 255
 256         bch2_open_bucket_get(c, wp, &ob);
 257         bch2_alloc_sectors_done(c, wp);
 258 mem_alloc:
 259         b = bch2_btree_node_mem_alloc(c);
 260
 261         /* we hold cannibalize_lock: */
 262         BUG_ON(IS_ERR(b));
 263         BUG_ON(b->ob.nr);
 264
 265         bkey_copy(&b->key, &tmp.k);
 266         b->ob = ob;
 267
 268         return b;
 269 }
 270
 271 static struct btree *bch2_btree_node_alloc(struct btree_update *as, unsigned level)
 272 {
 273         struct bch_fs *c = as->c;
 274         struct btree *b;
 275         int ret;
 276
 277         BUG_ON(level >= BTREE_MAX_DEPTH);
 278         BUG_ON(!as->nr_prealloc_nodes);
 279
 280         b = as->prealloc_nodes[--as->nr_prealloc_nodes];
 281
 282         set_btree_node_accessed(b);
 283         set_btree_node_dirty(c, b);
 284         set_btree_node_need_write(b);
 285
 286         bch2_bset_init_first(b, &b->data->keys);
 287         b->c.level      = level;
 288         b->c.btree_id   = as->btree_id;
 289
 290         memset(&b->nr, 0, sizeof(b->nr));
 291         b->data->magic = cpu_to_le64(bset_magic(c));
 292         b->data->flags = 0;
 293         SET_BTREE_NODE_ID(b->data, as->btree_id);
 294         SET_BTREE_NODE_LEVEL(b->data, level);
 295         b->data->ptr = bch2_bkey_ptrs_c(bkey_i_to_s_c(&b->key)).start->ptr;
 296
 297         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 298                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
 299
 300                 bp->v.mem_ptr           = 0;
 301                 bp->v.seq               = b->data->keys.seq;
 302                 bp->v.sectors_written   = 0;
 303         }
 304
 305         SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
 306
 307         bch2_btree_build_aux_trees(b);
 308
 309         ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
 310         BUG_ON(ret);
 311
 312         trace_btree_node_alloc(c, b);
 313         return b;
 314 }
 315
 316 static void btree_set_min(struct btree *b, struct bpos pos)
 317 {
 318         if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
 319                 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
 320         b->data->min_key = pos;
 321 }
 322
 323 static void btree_set_max(struct btree *b, struct bpos pos)
 324 {
 325         b->key.k.p = pos;
 326         b->data->max_key = pos;
 327 }
 328
 329 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *as,
 330                                                   struct btree *b,
 331                                                   struct bkey_format format)
 332 {
 333         struct btree *n;
 334
 335         n = bch2_btree_node_alloc(as, b->c.level);
 336
 337         SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
 338
 339         btree_set_min(n, b->data->min_key);
 340         btree_set_max(n, b->data->max_key);
 341
 342         n->data->format         = format;
 343         btree_node_set_format(n, format);
 344
 345         bch2_btree_sort_into(as->c, n, b);
 346
 347         btree_node_reset_sib_u64s(n);
 348
 349         n->key.k.p = b->key.k.p;
 350         return n;
 351 }
 352
 353 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
 354                                                        struct btree *b)
 355 {
 356         struct bkey_format new_f = bch2_btree_calc_format(b);
 357
 358         /*
 359          * The keys might expand with the new format - if they wouldn't fit in
 360          * the btree node anymore, use the old format for now:
 361          */
 362         if (!bch2_btree_node_format_fits(as->c, b, &new_f))
 363                 new_f = b->format;
 364
 365         return __bch2_btree_node_alloc_replacement(as, b, new_f);
 366 }
 367
 368 static struct btree *__btree_root_alloc(struct btree_update *as, unsigned level)
 369 {
 370         struct btree *b = bch2_btree_node_alloc(as, level);
 371
 372         btree_set_min(b, POS_MIN);
 373         btree_set_max(b, POS_MAX);
 374         b->data->format = bch2_btree_calc_format(b);
 375
 376         btree_node_set_format(b, b->data->format);
 377         bch2_btree_build_aux_trees(b);
 378
 379         bch2_btree_update_add_new_node(as, b);
 380         six_unlock_write(&b->c.lock);
 381
 382         return b;
 383 }
 384
 385 static void bch2_btree_reserve_put(struct btree_update *as)
 386 {
 387         struct bch_fs *c = as->c;
 388
 389         mutex_lock(&c->btree_reserve_cache_lock);
 390
 391         while (as->nr_prealloc_nodes) {
 392                 struct btree *b = as->prealloc_nodes[--as->nr_prealloc_nodes];
 393
 394                 six_unlock_write(&b->c.lock);
 395
 396                 if (c->btree_reserve_cache_nr <
 397                     ARRAY_SIZE(c->btree_reserve_cache)) {
 398                         struct btree_alloc *a =
 399                                 &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
 400
 401                         a->ob = b->ob;
 402                         b->ob.nr = 0;
 403                         bkey_copy(&a->k, &b->key);
 404                 } else {
 405                         bch2_open_buckets_put(c, &b->ob);
 406                 }
 407
 408                 btree_node_lock_type(c, b, SIX_LOCK_write);
 409                 __btree_node_free(c, b);
 410                 six_unlock_write(&b->c.lock);
 411
 412                 six_unlock_intent(&b->c.lock);
 413         }
 414
 415         mutex_unlock(&c->btree_reserve_cache_lock);
 416 }
 417
 418 static int bch2_btree_reserve_get(struct btree_update *as, unsigned nr_nodes,
 419                                   unsigned flags, struct closure *cl)
 420 {
 421         struct bch_fs *c = as->c;
 422         struct btree *b;
 423         int ret;
 424
 425         BUG_ON(nr_nodes > BTREE_RESERVE_MAX);
 426
 427         /*
 428          * Protects reaping from the btree node cache and using the btree node
 429          * open bucket reserve:
 430          */
 431         ret = bch2_btree_cache_cannibalize_lock(c, cl);
 432         if (ret)
 433                 return ret;
 434
 435         while (as->nr_prealloc_nodes < nr_nodes) {
 436                 b = __bch2_btree_node_alloc(c, &as->disk_res,
 437                                             flags & BTREE_INSERT_NOWAIT
 438                                             ? NULL : cl, flags);
 439                 if (IS_ERR(b)) {
 440                         ret = PTR_ERR(b);
 441                         goto err_free;
 442                 }
 443
 444                 ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(&b->key));
 445                 if (ret)
 446                         goto err_free;
 447
 448                 as->prealloc_nodes[as->nr_prealloc_nodes++] = b;
 449         }
 450
 451         bch2_btree_cache_cannibalize_unlock(c);
 452         return 0;
 453 err_free:
 454         bch2_btree_cache_cannibalize_unlock(c);
 455         trace_btree_reserve_get_fail(c, nr_nodes, cl);
 456         return ret;
 457 }
 458
 459 /* Asynchronous interior node update machinery */
 460
 461 static void bch2_btree_update_free(struct btree_update *as)
 462 {
 463         struct bch_fs *c = as->c;
 464
 465         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 466
 467         bch2_journal_pin_drop(&c->journal, &as->journal);
 468         bch2_journal_pin_flush(&c->journal, &as->journal);
 469         bch2_disk_reservation_put(c, &as->disk_res);
 470         bch2_btree_reserve_put(as);
 471
 472         mutex_lock(&c->btree_interior_update_lock);
 473         list_del(&as->unwritten_list);
 474         list_del(&as->list);
 475         mutex_unlock(&c->btree_interior_update_lock);
 476
 477         closure_debug_destroy(&as->cl);
 478         mempool_free(as, &c->btree_interior_update_pool);
 479
 480         closure_wake_up(&c->btree_interior_update_wait);
 481 }
 482
 483 static void btree_update_will_delete_key(struct btree_update *as,
 484                                          struct bkey_i *k)
 485 {
 486         BUG_ON(bch2_keylist_u64s(&as->old_keys) + k->k.u64s >
 487                ARRAY_SIZE(as->_old_keys));
 488         bch2_keylist_add(&as->old_keys, k);
 489 }
 490
 491 static void btree_update_will_add_key(struct btree_update *as,
 492                                       struct bkey_i *k)
 493 {
 494         BUG_ON(bch2_keylist_u64s(&as->new_keys) + k->k.u64s >
 495                ARRAY_SIZE(as->_new_keys));
 496         bch2_keylist_add(&as->new_keys, k);
 497 }
 498
 499 /*
 500  * The transactional part of an interior btree node update, where we journal the
 501  * update we did to the interior node and update alloc info:
 502  */
 503 static int btree_update_nodes_written_trans(struct btree_trans *trans,
 504                                             struct btree_update *as)
 505 {
 506         struct bkey_i *k;
 507         int ret;
 508
 509         trans->extra_journal_entries = (void *) &as->journal_entries[0];
 510         trans->extra_journal_entry_u64s = as->journal_u64s;
 511         trans->journal_pin = &as->journal;
 512
 513         for_each_keylist_key(&as->new_keys, k) {
 514                 ret = bch2_trans_mark_key(trans,
 515                                           bkey_s_c_null,
 516                                           bkey_i_to_s_c(k),
 517                                           0, 0, BTREE_TRIGGER_INSERT);
 518                 if (ret)
 519                         return ret;
 520         }
 521
 522         for_each_keylist_key(&as->old_keys, k) {
 523                 ret = bch2_trans_mark_key(trans,
 524                                           bkey_i_to_s_c(k),
 525                                           bkey_s_c_null,
 526                                           0, 0, BTREE_TRIGGER_OVERWRITE);
 527                 if (ret)
 528                         return ret;
 529         }
 530
 531         return 0;
 532 }
 533
 534 static void btree_update_nodes_written(struct btree_update *as)
 535 {
 536         struct bch_fs *c = as->c;
 537         struct btree *b = as->b;
 538         struct btree_trans trans;
 539         u64 journal_seq = 0;
 540         unsigned i;
 541         int ret;
 542
 543         /*
 544          * If we're already in an error state, it might be because a btree node
 545          * was never written, and we might be trying to free that same btree
 546          * node here, but it won't have been marked as allocated and we'll see
 547          * spurious disk usage inconsistencies in the transactional part below
 548          * if we don't skip it:
 549          */
 550         ret = bch2_journal_error(&c->journal);
 551         if (ret)
 552                 goto err;
 553
 554         BUG_ON(!journal_pin_active(&as->journal));
 555
 556         /*
 557          * We did an update to a parent node where the pointers we added pointed
 558          * to child nodes that weren't written yet: now, the child nodes have
 559          * been written so we can write out the update to the interior node.
 560          */
 561
 562         /*
 563          * We can't call into journal reclaim here: we'd block on the journal
 564          * reclaim lock, but we may need to release the open buckets we have
 565          * pinned in order for other btree updates to make forward progress, and
 566          * journal reclaim does btree updates when flushing bkey_cached entries,
 567          * which may require allocations as well.
 568          */
 569         bch2_trans_init(&trans, c, 0, 512);
 570         ret = __bch2_trans_do(&trans, &as->disk_res, &journal_seq,
 571                               BTREE_INSERT_NOFAIL|
 572                               BTREE_INSERT_NOCHECK_RW|
 573                               BTREE_INSERT_JOURNAL_RECLAIM|
 574                               BTREE_INSERT_JOURNAL_RESERVED,
 575                               btree_update_nodes_written_trans(&trans, as));
 576         bch2_trans_exit(&trans);
 577
 578         bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
 579                              "error %i in btree_update_nodes_written()", ret);
 580 err:
 581         if (b) {
 582                 /*
 583                  * @b is the node we did the final insert into:
 584                  *
 585                  * On failure to get a journal reservation, we still have to
 586                  * unblock the write and allow most of the write path to happen
 587                  * so that shutdown works, but the i->journal_seq mechanism
 588                  * won't work to prevent the btree write from being visible (we
 589                  * didn't get a journal sequence number) - instead
 590                  * __bch2_btree_node_write() doesn't do the actual write if
 591                  * we're in journal error state:
 592                  */
 593
 594                 btree_node_lock_type(c, b, SIX_LOCK_intent);
 595                 btree_node_lock_type(c, b, SIX_LOCK_write);
 596                 mutex_lock(&c->btree_interior_update_lock);
 597
 598                 list_del(&as->write_blocked_list);
 599
 600                 /*
 601                  * Node might have been freed, recheck under
 602                  * btree_interior_update_lock:
 603                  */
 604                 if (as->b == b) {
 605                         struct bset *i = btree_bset_last(b);
 606
 607                         BUG_ON(!b->c.level);
 608                         BUG_ON(!btree_node_dirty(b));
 609
 610                         if (!ret) {
 611                                 i->journal_seq = cpu_to_le64(
 612                                         max(journal_seq,
 613                                             le64_to_cpu(i->journal_seq)));
 614
 615                                 bch2_btree_add_journal_pin(c, b, journal_seq);
 616                         } else {
 617                                 /*
 618                                  * If we didn't get a journal sequence number we
 619                                  * can't write this btree node, because recovery
 620                                  * won't know to ignore this write:
 621                                  */
 622                                 set_btree_node_never_write(b);
 623                         }
 624                 }
 625
 626                 mutex_unlock(&c->btree_interior_update_lock);
 627                 six_unlock_write(&b->c.lock);
 628
 629                 btree_node_write_if_need(c, b, SIX_LOCK_intent);
 630                 six_unlock_intent(&b->c.lock);
 631         }
 632
 633         bch2_journal_pin_drop(&c->journal, &as->journal);
 634
 635         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 636
 637         mutex_lock(&c->btree_interior_update_lock);
 638         for (i = 0; i < as->nr_new_nodes; i++) {
 639                 b = as->new_nodes[i];
 640
 641                 BUG_ON(b->will_make_reachable != (unsigned long) as);
 642                 b->will_make_reachable = 0;
 643         }
 644         mutex_unlock(&c->btree_interior_update_lock);
 645
 646         for (i = 0; i < as->nr_new_nodes; i++) {
 647                 b = as->new_nodes[i];
 648
 649                 btree_node_lock_type(c, b, SIX_LOCK_read);
 650                 btree_node_write_if_need(c, b, SIX_LOCK_read);
 651                 six_unlock_read(&b->c.lock);
 652         }
 653
 654         for (i = 0; i < as->nr_open_buckets; i++)
 655                 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
 656
 657         bch2_btree_update_free(as);
 658 }
 659
 660 static void btree_interior_update_work(struct work_struct *work)
 661 {
 662         struct bch_fs *c =
 663                 container_of(work, struct bch_fs, btree_interior_update_work);
 664         struct btree_update *as;
 665
 666         while (1) {
 667                 mutex_lock(&c->btree_interior_update_lock);
 668                 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
 669                                               struct btree_update, unwritten_list);
 670                 if (as && !as->nodes_written)
 671                         as = NULL;
 672                 mutex_unlock(&c->btree_interior_update_lock);
 673
 674                 if (!as)
 675                         break;
 676
 677                 btree_update_nodes_written(as);
 678         }
 679 }
 680
 681 static void btree_update_set_nodes_written(struct closure *cl)
 682 {
 683         struct btree_update *as = container_of(cl, struct btree_update, cl);
 684         struct bch_fs *c = as->c;
 685
 686         mutex_lock(&c->btree_interior_update_lock);
 687         as->nodes_written = true;
 688         mutex_unlock(&c->btree_interior_update_lock);
 689
 690         queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
 691 }
 692
 693 /*
 694  * We're updating @b with pointers to nodes that haven't finished writing yet:
 695  * block @b from being written until @as completes
 696  */
 697 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
 698 {
 699         struct bch_fs *c = as->c;
 700
 701         mutex_lock(&c->btree_interior_update_lock);
 702         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 703
 704         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 705         BUG_ON(!btree_node_dirty(b));
 706
 707         as->mode        = BTREE_INTERIOR_UPDATING_NODE;
 708         as->b           = b;
 709         list_add(&as->write_blocked_list, &b->write_blocked);
 710
 711         mutex_unlock(&c->btree_interior_update_lock);
 712 }
 713
 714 static void btree_update_reparent(struct btree_update *as,
 715                                   struct btree_update *child)
 716 {
 717         struct bch_fs *c = as->c;
 718
 719         lockdep_assert_held(&c->btree_interior_update_lock);
 720
 721         child->b = NULL;
 722         child->mode = BTREE_INTERIOR_UPDATING_AS;
 723
 724         bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL);
 725 }
 726
 727 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
 728 {
 729         struct bkey_i *insert = &b->key;
 730         struct bch_fs *c = as->c;
 731
 732         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 733
 734         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
 735                ARRAY_SIZE(as->journal_entries));
 736
 737         as->journal_u64s +=
 738                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
 739                                   BCH_JSET_ENTRY_btree_root,
 740                                   b->c.btree_id, b->c.level,
 741                                   insert, insert->k.u64s);
 742
 743         mutex_lock(&c->btree_interior_update_lock);
 744         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 745
 746         as->mode        = BTREE_INTERIOR_UPDATING_ROOT;
 747         mutex_unlock(&c->btree_interior_update_lock);
 748 }
 749
 750 /*
 751  * bch2_btree_update_add_new_node:
 752  *
 753  * This causes @as to wait on @b to be written, before it gets to
 754  * bch2_btree_update_nodes_written
 755  *
 756  * Additionally, it sets b->will_make_reachable to prevent any additional writes
 757  * to @b from happening besides the first until @b is reachable on disk
 758  *
 759  * And it adds @b to the list of @as's new nodes, so that we can update sector
 760  * counts in bch2_btree_update_nodes_written:
 761  */
 762 void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
 763 {
 764         struct bch_fs *c = as->c;
 765
 766         closure_get(&as->cl);
 767
 768         mutex_lock(&c->btree_interior_update_lock);
 769         BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
 770         BUG_ON(b->will_make_reachable);
 771
 772         as->new_nodes[as->nr_new_nodes++] = b;
 773         b->will_make_reachable = 1UL|(unsigned long) as;
 774
 775         mutex_unlock(&c->btree_interior_update_lock);
 776
 777         btree_update_will_add_key(as, &b->key);
 778 }
 779
 780 /*
 781  * returns true if @b was a new node
 782  */
 783 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
 784 {
 785         struct btree_update *as;
 786         unsigned long v;
 787         unsigned i;
 788
 789         mutex_lock(&c->btree_interior_update_lock);
 790         /*
 791          * When b->will_make_reachable != 0, it owns a ref on as->cl that's
 792          * dropped when it gets written by bch2_btree_complete_write - the
 793          * xchg() is for synchronization with bch2_btree_complete_write:
 794          */
 795         v = xchg(&b->will_make_reachable, 0);
 796         as = (struct btree_update *) (v & ~1UL);
 797
 798         if (!as) {
 799                 mutex_unlock(&c->btree_interior_update_lock);
 800                 return;
 801         }
 802
 803         for (i = 0; i < as->nr_new_nodes; i++)
 804                 if (as->new_nodes[i] == b)
 805                         goto found;
 806
 807         BUG();
 808 found:
 809         array_remove_item(as->new_nodes, as->nr_new_nodes, i);
 810         mutex_unlock(&c->btree_interior_update_lock);
 811
 812         if (v & 1)
 813                 closure_put(&as->cl);
 814 }
 815
 816 void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
 817 {
 818         while (b->ob.nr)
 819                 as->open_buckets[as->nr_open_buckets++] =
 820                         b->ob.v[--b->ob.nr];
 821 }
 822
 823 /*
 824  * @b is being split/rewritten: it may have pointers to not-yet-written btree
 825  * nodes and thus outstanding btree_updates - redirect @b's
 826  * btree_updates to point to this btree_update:
 827  */
 828 void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 829                                                struct btree *b)
 830 {
 831         struct bch_fs *c = as->c;
 832         struct btree_update *p, *n;
 833         struct btree_write *w;
 834
 835         set_btree_node_dying(b);
 836
 837         if (btree_node_fake(b))
 838                 return;
 839
 840         mutex_lock(&c->btree_interior_update_lock);
 841
 842         /*
 843          * Does this node have any btree_update operations preventing
 844          * it from being written?
 845          *
 846          * If so, redirect them to point to this btree_update: we can
 847          * write out our new nodes, but we won't make them visible until those
 848          * operations complete
 849          */
 850         list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
 851                 list_del_init(&p->write_blocked_list);
 852                 btree_update_reparent(as, p);
 853
 854                 /*
 855                  * for flush_held_btree_writes() waiting on updates to flush or
 856                  * nodes to be writeable:
 857                  */
 858                 closure_wake_up(&c->btree_interior_update_wait);
 859         }
 860
 861         clear_btree_node_dirty(c, b);
 862         clear_btree_node_need_write(b);
 863
 864         /*
 865          * Does this node have unwritten data that has a pin on the journal?
 866          *
 867          * If so, transfer that pin to the btree_update operation -
 868          * note that if we're freeing multiple nodes, we only need to keep the
 869          * oldest pin of any of the nodes we're freeing. We'll release the pin
 870          * when the new nodes are persistent and reachable on disk:
 871          */
 872         w = btree_current_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         w = btree_prev_write(b);
 877         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 878         bch2_journal_pin_drop(&c->journal, &w->journal);
 879
 880         mutex_unlock(&c->btree_interior_update_lock);
 881
 882         /*
 883          * Is this a node that isn't reachable on disk yet?
 884          *
 885          * Nodes that aren't reachable yet have writes blocked until they're
 886          * reachable - now that we've cancelled any pending writes and moved
 887          * things waiting on that write to wait on this update, we can drop this
 888          * node from the list of nodes that the other update is making
 889          * reachable, prior to freeing it:
 890          */
 891         btree_update_drop_new_node(c, b);
 892
 893         btree_update_will_delete_key(as, &b->key);
 894 }
 895
 896 void bch2_btree_update_done(struct btree_update *as)
 897 {
 898         BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
 899
 900         bch2_btree_reserve_put(as);
 901
 902         continue_at(&as->cl, btree_update_set_nodes_written, system_freezable_wq);
 903 }
 904
 905 struct btree_update *
 906 bch2_btree_update_start(struct btree_trans *trans, enum btree_id id,
 907                         unsigned nr_nodes, unsigned flags,
 908                         struct closure *cl)
 909 {
 910         struct bch_fs *c = trans->c;
 911         struct btree_update *as;
 912         int disk_res_flags = (flags & BTREE_INSERT_NOFAIL)
 913                 ? BCH_DISK_RESERVATION_NOFAIL : 0;
 914         int journal_flags = (flags & BTREE_INSERT_JOURNAL_RESERVED)
 915                 ? JOURNAL_RES_GET_RECLAIM : 0;
 916         int ret = 0;
 917
 918         /*
 919          * This check isn't necessary for correctness - it's just to potentially
 920          * prevent us from doing a lot of work that'll end up being wasted:
 921          */
 922         ret = bch2_journal_error(&c->journal);
 923         if (ret)
 924                 return ERR_PTR(ret);
 925
 926         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
 927         memset(as, 0, sizeof(*as));
 928         closure_init(&as->cl, NULL);
 929         as->c           = c;
 930         as->mode        = BTREE_INTERIOR_NO_UPDATE;
 931         as->btree_id    = id;
 932         INIT_LIST_HEAD(&as->list);
 933         INIT_LIST_HEAD(&as->unwritten_list);
 934         INIT_LIST_HEAD(&as->write_blocked_list);
 935         bch2_keylist_init(&as->old_keys, as->_old_keys);
 936         bch2_keylist_init(&as->new_keys, as->_new_keys);
 937         bch2_keylist_init(&as->parent_keys, as->inline_keys);
 938
 939         ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
 940                                       BTREE_UPDATE_JOURNAL_RES,
 941                                       journal_flags|JOURNAL_RES_GET_NONBLOCK);
 942         if (ret == -EAGAIN) {
 943                 if (flags & BTREE_INSERT_NOUNLOCK)
 944                         return ERR_PTR(-EINTR);
 945
 946                 bch2_trans_unlock(trans);
 947
 948                 ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
 949                                 BTREE_UPDATE_JOURNAL_RES,
 950                                 journal_flags);
 951                 if (ret)
 952                         return ERR_PTR(ret);
 953
 954                 if (!bch2_trans_relock(trans)) {
 955                         ret = -EINTR;
 956                         goto err;
 957                 }
 958         }
 959
 960         ret = bch2_disk_reservation_get(c, &as->disk_res,
 961                         nr_nodes * c->opts.btree_node_size,
 962                         c->opts.metadata_replicas,
 963                         disk_res_flags);
 964         if (ret)
 965                 goto err;
 966
 967         ret = bch2_btree_reserve_get(as, nr_nodes, flags, cl);
 968         if (ret)
 969                 goto err;
 970
 971         bch2_journal_pin_add(&c->journal,
 972                              atomic64_read(&c->journal.seq),
 973                              &as->journal, NULL);
 974
 975         mutex_lock(&c->btree_interior_update_lock);
 976         list_add_tail(&as->list, &c->btree_interior_update_list);
 977         mutex_unlock(&c->btree_interior_update_lock);
 978
 979         return as;
 980 err:
 981         bch2_btree_update_free(as);
 982         return ERR_PTR(ret);
 983 }
 984
 985 /* Btree root updates: */
 986
 987 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
 988 {
 989         /* Root nodes cannot be reaped */
 990         mutex_lock(&c->btree_cache.lock);
 991         list_del_init(&b->list);
 992         mutex_unlock(&c->btree_cache.lock);
 993
 994         mutex_lock(&c->btree_root_lock);
 995         BUG_ON(btree_node_root(c, b) &&
 996                (b->c.level < btree_node_root(c, b)->c.level ||
 997                 !btree_node_dying(btree_node_root(c, b))));
 998
 999         btree_node_root(c, b) = b;
1000         mutex_unlock(&c->btree_root_lock);
1001
1002         bch2_recalc_btree_reserve(c);
1003 }
1004
1005 /**
1006  * bch_btree_set_root - update the root in memory and on disk
1007  *
1008  * To ensure forward progress, the current task must not be holding any
1009  * btree node write locks. However, you must hold an intent lock on the
1010  * old root.
1011  *
1012  * Note: This allocates a journal entry but doesn't add any keys to
1013  * it.  All the btree roots are part of every journal write, so there
1014  * is nothing new to be done.  This just guarantees that there is a
1015  * journal write.
1016  */
1017 static void bch2_btree_set_root(struct btree_update *as, struct btree *b,
1018                                 struct btree_iter *iter)
1019 {
1020         struct bch_fs *c = as->c;
1021         struct btree *old;
1022
1023         trace_btree_set_root(c, b);
1024         BUG_ON(!b->written &&
1025                !test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags));
1026
1027         old = btree_node_root(c, b);
1028
1029         /*
1030          * Ensure no one is using the old root while we switch to the
1031          * new root:
1032          */
1033         bch2_btree_node_lock_write(old, iter);
1034
1035         bch2_btree_set_root_inmem(c, b);
1036
1037         btree_update_updated_root(as, b);
1038
1039         /*
1040          * Unlock old root after new root is visible:
1041          *
1042          * The new root isn't persistent, but that's ok: we still have
1043          * an intent lock on the new root, and any updates that would
1044          * depend on the new root would have to update the new root.
1045          */
1046         bch2_btree_node_unlock_write(old, iter);
1047 }
1048
1049 /* Interior node updates: */
1050
1051 static void bch2_insert_fixup_btree_ptr(struct btree_update *as, struct btree *b,
1052                                         struct btree_iter *iter,
1053                                         struct bkey_i *insert,
1054                                         struct btree_node_iter *node_iter)
1055 {
1056         struct bch_fs *c = as->c;
1057         struct bkey_packed *k;
1058         const char *invalid;
1059
1060         invalid = bch2_bkey_invalid(c, bkey_i_to_s_c(insert), btree_node_type(b)) ?:
1061                 bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert));
1062         if (invalid) {
1063                 char buf[160];
1064
1065                 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(insert));
1066                 bch2_fs_inconsistent(c, "inserting invalid bkey %s: %s", buf, invalid);
1067                 dump_stack();
1068         }
1069
1070         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1071                ARRAY_SIZE(as->journal_entries));
1072
1073         as->journal_u64s +=
1074                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1075                                   BCH_JSET_ENTRY_btree_keys,
1076                                   b->c.btree_id, b->c.level,
1077                                   insert, insert->k.u64s);
1078
1079         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1080                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1081                 bch2_btree_node_iter_advance(node_iter, b);
1082
1083         bch2_btree_bset_insert_key(iter, b, node_iter, insert);
1084         set_btree_node_dirty(c, b);
1085         set_btree_node_need_write(b);
1086 }
1087
1088 /*
1089  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1090  * node)
1091  */
1092 static struct btree *__btree_split_node(struct btree_update *as,
1093                                         struct btree *n1,
1094                                         struct btree_iter *iter)
1095 {
1096         size_t nr_packed = 0, nr_unpacked = 0;
1097         struct btree *n2;
1098         struct bset *set1, *set2;
1099         struct bkey_packed *k, *prev = NULL;
1100
1101         n2 = bch2_btree_node_alloc(as, n1->c.level);
1102         bch2_btree_update_add_new_node(as, n2);
1103
1104         n2->data->max_key       = n1->data->max_key;
1105         n2->data->format        = n1->format;
1106         SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data));
1107         n2->key.k.p = n1->key.k.p;
1108
1109         btree_node_set_format(n2, n2->data->format);
1110
1111         set1 = btree_bset_first(n1);
1112         set2 = btree_bset_first(n2);
1113
1114         /*
1115          * Has to be a linear search because we don't have an auxiliary
1116          * search tree yet
1117          */
1118         k = set1->start;
1119         while (1) {
1120                 struct bkey_packed *n = bkey_next_skip_noops(k, vstruct_last(set1));
1121
1122                 if (n == vstruct_last(set1))
1123                         break;
1124                 if (k->_data - set1->_data >= (le16_to_cpu(set1->u64s) * 3) / 5)
1125                         break;
1126
1127                 if (bkey_packed(k))
1128                         nr_packed++;
1129                 else
1130                         nr_unpacked++;
1131
1132                 prev = k;
1133                 k = n;
1134         }
1135
1136         BUG_ON(!prev);
1137
1138         btree_set_max(n1, bkey_unpack_pos(n1, prev));
1139         btree_set_min(n2, bkey_successor(n1->key.k.p));
1140
1141         set2->u64s = cpu_to_le16((u64 *) vstruct_end(set1) - (u64 *) k);
1142         set1->u64s = cpu_to_le16(le16_to_cpu(set1->u64s) - le16_to_cpu(set2->u64s));
1143
1144         set_btree_bset_end(n1, n1->set);
1145         set_btree_bset_end(n2, n2->set);
1146
1147         n2->nr.live_u64s        = le16_to_cpu(set2->u64s);
1148         n2->nr.bset_u64s[0]     = le16_to_cpu(set2->u64s);
1149         n2->nr.packed_keys      = n1->nr.packed_keys - nr_packed;
1150         n2->nr.unpacked_keys    = n1->nr.unpacked_keys - nr_unpacked;
1151
1152         n1->nr.live_u64s        = le16_to_cpu(set1->u64s);
1153         n1->nr.bset_u64s[0]     = le16_to_cpu(set1->u64s);
1154         n1->nr.packed_keys      = nr_packed;
1155         n1->nr.unpacked_keys    = nr_unpacked;
1156
1157         BUG_ON(!set1->u64s);
1158         BUG_ON(!set2->u64s);
1159
1160         memcpy_u64s(set2->start,
1161                     vstruct_end(set1),
1162                     le16_to_cpu(set2->u64s));
1163
1164         btree_node_reset_sib_u64s(n1);
1165         btree_node_reset_sib_u64s(n2);
1166
1167         bch2_verify_btree_nr_keys(n1);
1168         bch2_verify_btree_nr_keys(n2);
1169
1170         if (n1->c.level) {
1171                 btree_node_interior_verify(as->c, n1);
1172                 btree_node_interior_verify(as->c, n2);
1173         }
1174
1175         return n2;
1176 }
1177
1178 /*
1179  * For updates to interior nodes, we've got to do the insert before we split
1180  * because the stuff we're inserting has to be inserted atomically. Post split,
1181  * the keys might have to go in different nodes and the split would no longer be
1182  * atomic.
1183  *
1184  * Worse, if the insert is from btree node coalescing, if we do the insert after
1185  * we do the split (and pick the pivot) - the pivot we pick might be between
1186  * nodes that were coalesced, and thus in the middle of a child node post
1187  * coalescing:
1188  */
1189 static void btree_split_insert_keys(struct btree_update *as, struct btree *b,
1190                                     struct btree_iter *iter,
1191                                     struct keylist *keys)
1192 {
1193         struct btree_node_iter node_iter;
1194         struct bkey_i *k = bch2_keylist_front(keys);
1195         struct bkey_packed *src, *dst, *n;
1196         struct bset *i;
1197
1198         BUG_ON(btree_node_type(b) != BKEY_TYPE_BTREE);
1199
1200         bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
1201
1202         while (!bch2_keylist_empty(keys)) {
1203                 k = bch2_keylist_front(keys);
1204
1205                 bch2_insert_fixup_btree_ptr(as, b, iter, k, &node_iter);
1206                 bch2_keylist_pop_front(keys);
1207         }
1208
1209         /*
1210          * We can't tolerate whiteouts here - with whiteouts there can be
1211          * duplicate keys, and it would be rather bad if we picked a duplicate
1212          * for the pivot:
1213          */
1214         i = btree_bset_first(b);
1215         src = dst = i->start;
1216         while (src != vstruct_last(i)) {
1217                 n = bkey_next_skip_noops(src, vstruct_last(i));
1218                 if (!bkey_deleted(src)) {
1219                         memmove_u64s_down(dst, src, src->u64s);
1220                         dst = bkey_next(dst);
1221                 }
1222                 src = n;
1223         }
1224
1225         /* Also clear out the unwritten whiteouts area: */
1226         b->whiteout_u64s = 0;
1227
1228         i->u64s = cpu_to_le16((u64 *) dst - i->_data);
1229         set_btree_bset_end(b, b->set);
1230
1231         BUG_ON(b->nsets != 1 ||
1232                b->nr.live_u64s != le16_to_cpu(btree_bset_first(b)->u64s));
1233
1234         btree_node_interior_verify(as->c, b);
1235 }
1236
1237 static void btree_split(struct btree_update *as, struct btree *b,
1238                         struct btree_iter *iter, struct keylist *keys,
1239                         unsigned flags)
1240 {
1241         struct bch_fs *c = as->c;
1242         struct btree *parent = btree_node_parent(iter, b);
1243         struct btree *n1, *n2 = NULL, *n3 = NULL;
1244         u64 start_time = local_clock();
1245
1246         BUG_ON(!parent && (b != btree_node_root(c, b)));
1247         BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->c.level));
1248
1249         bch2_btree_interior_update_will_free_node(as, b);
1250
1251         n1 = bch2_btree_node_alloc_replacement(as, b);
1252         bch2_btree_update_add_new_node(as, n1);
1253
1254         if (keys)
1255                 btree_split_insert_keys(as, n1, iter, keys);
1256
1257         if (bset_u64s(&n1->set[0]) > BTREE_SPLIT_THRESHOLD(c)) {
1258                 trace_btree_split(c, b);
1259
1260                 n2 = __btree_split_node(as, n1, iter);
1261
1262                 bch2_btree_build_aux_trees(n2);
1263                 bch2_btree_build_aux_trees(n1);
1264                 six_unlock_write(&n2->c.lock);
1265                 six_unlock_write(&n1->c.lock);
1266
1267                 bch2_btree_node_write(c, n2, SIX_LOCK_intent);
1268
1269                 /*
1270                  * Note that on recursive parent_keys == keys, so we
1271                  * can't start adding new keys to parent_keys before emptying it
1272                  * out (which we did with btree_split_insert_keys() above)
1273                  */
1274                 bch2_keylist_add(&as->parent_keys, &n1->key);
1275                 bch2_keylist_add(&as->parent_keys, &n2->key);
1276
1277                 if (!parent) {
1278                         /* Depth increases, make a new root */
1279                         n3 = __btree_root_alloc(as, b->c.level + 1);
1280
1281                         n3->sib_u64s[0] = U16_MAX;
1282                         n3->sib_u64s[1] = U16_MAX;
1283
1284                         btree_split_insert_keys(as, n3, iter, &as->parent_keys);
1285
1286                         bch2_btree_node_write(c, n3, SIX_LOCK_intent);
1287                 }
1288         } else {
1289                 trace_btree_compact(c, b);
1290
1291                 bch2_btree_build_aux_trees(n1);
1292                 six_unlock_write(&n1->c.lock);
1293
1294                 if (parent)
1295                         bch2_keylist_add(&as->parent_keys, &n1->key);
1296         }
1297
1298         bch2_btree_node_write(c, n1, SIX_LOCK_intent);
1299
1300         /* New nodes all written, now make them visible: */
1301
1302         if (parent) {
1303                 /* Split a non root node */
1304                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1305         } else if (n3) {
1306                 bch2_btree_set_root(as, n3, iter);
1307         } else {
1308                 /* Root filled up but didn't need to be split */
1309                 bch2_btree_set_root(as, n1, iter);
1310         }
1311
1312         bch2_btree_update_get_open_buckets(as, n1);
1313         if (n2)
1314                 bch2_btree_update_get_open_buckets(as, n2);
1315         if (n3)
1316                 bch2_btree_update_get_open_buckets(as, n3);
1317
1318         /* Successful split, update the iterator to point to the new nodes: */
1319
1320         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1321         bch2_btree_iter_node_drop(iter, b);
1322         if (n3)
1323                 bch2_btree_iter_node_replace(iter, n3);
1324         if (n2)
1325                 bch2_btree_iter_node_replace(iter, n2);
1326         bch2_btree_iter_node_replace(iter, n1);
1327
1328         /*
1329          * The old node must be freed (in memory) _before_ unlocking the new
1330          * nodes - else another thread could re-acquire a read lock on the old
1331          * node after another thread has locked and updated the new node, thus
1332          * seeing stale data:
1333          */
1334         bch2_btree_node_free_inmem(c, b, iter);
1335
1336         if (n3)
1337                 six_unlock_intent(&n3->c.lock);
1338         if (n2)
1339                 six_unlock_intent(&n2->c.lock);
1340         six_unlock_intent(&n1->c.lock);
1341
1342         bch2_btree_trans_verify_locks(iter->trans);
1343
1344         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_split],
1345                                start_time);
1346 }
1347
1348 static void
1349 bch2_btree_insert_keys_interior(struct btree_update *as, struct btree *b,
1350                                 struct btree_iter *iter, struct keylist *keys)
1351 {
1352         struct btree_iter *linked;
1353         struct btree_node_iter node_iter;
1354         struct bkey_i *insert = bch2_keylist_front(keys);
1355         struct bkey_packed *k;
1356
1357         /* Don't screw up @iter's position: */
1358         node_iter = iter->l[b->c.level].iter;
1359
1360         /*
1361          * btree_split(), btree_gc_coalesce() will insert keys before
1362          * the iterator's current position - they know the keys go in
1363          * the node the iterator points to:
1364          */
1365         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1366                (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1367                 ;
1368
1369         for_each_keylist_key(keys, insert)
1370                 bch2_insert_fixup_btree_ptr(as, b, iter, insert, &node_iter);
1371
1372         btree_update_updated_node(as, b);
1373
1374         trans_for_each_iter_with_node(iter->trans, b, linked)
1375                 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1376
1377         bch2_btree_trans_verify_iters(iter->trans, b);
1378 }
1379
1380 /**
1381  * bch_btree_insert_node - insert bkeys into a given btree node
1382  *
1383  * @iter:               btree iterator
1384  * @keys:               list of keys to insert
1385  * @hook:               insert callback
1386  * @persistent:         if not null, @persistent will wait on journal write
1387  *
1388  * Inserts as many keys as it can into a given btree node, splitting it if full.
1389  * If a split occurred, this function will return early. This can only happen
1390  * for leaf nodes -- inserts into interior nodes have to be atomic.
1391  */
1392 void bch2_btree_insert_node(struct btree_update *as, struct btree *b,
1393                             struct btree_iter *iter, struct keylist *keys,
1394                             unsigned flags)
1395 {
1396         struct bch_fs *c = as->c;
1397         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1398         int old_live_u64s = b->nr.live_u64s;
1399         int live_u64s_added, u64s_added;
1400
1401         BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->c.level));
1402         BUG_ON(!b->c.level);
1403         BUG_ON(!as || as->b);
1404         bch2_verify_keylist_sorted(keys);
1405
1406         bch2_btree_node_lock_for_insert(c, b, iter);
1407
1408         if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1409                 bch2_btree_node_unlock_write(b, iter);
1410                 goto split;
1411         }
1412
1413         btree_node_interior_verify(c, b);
1414
1415         bch2_btree_insert_keys_interior(as, b, iter, keys);
1416
1417         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1418         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1419
1420         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1421                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1422         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1423                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1424
1425         if (u64s_added > live_u64s_added &&
1426             bch2_maybe_compact_whiteouts(c, b))
1427                 bch2_btree_iter_reinit_node(iter, b);
1428
1429         bch2_btree_node_unlock_write(b, iter);
1430
1431         btree_node_interior_verify(c, b);
1432
1433         /*
1434          * when called from the btree_split path the new nodes aren't added to
1435          * the btree iterator yet, so the merge path's unlock/wait/relock dance
1436          * won't work:
1437          */
1438         bch2_foreground_maybe_merge(c, iter, b->c.level,
1439                                     flags|BTREE_INSERT_NOUNLOCK);
1440         return;
1441 split:
1442         btree_split(as, b, iter, keys, flags);
1443 }
1444
1445 int bch2_btree_split_leaf(struct bch_fs *c, struct btree_iter *iter,
1446                           unsigned flags)
1447 {
1448         struct btree_trans *trans = iter->trans;
1449         struct btree *b = iter_l(iter)->b;
1450         struct btree_update *as;
1451         struct closure cl;
1452         int ret = 0;
1453
1454         closure_init_stack(&cl);
1455
1456         /* Hack, because gc and splitting nodes doesn't mix yet: */
1457         if (!(flags & BTREE_INSERT_GC_LOCK_HELD) &&
1458             !down_read_trylock(&c->gc_lock)) {
1459                 if (flags & BTREE_INSERT_NOUNLOCK) {
1460                         trace_transaction_restart_ip(trans->ip, _THIS_IP_);
1461                         return -EINTR;
1462                 }
1463
1464                 bch2_trans_unlock(trans);
1465                 down_read(&c->gc_lock);
1466
1467                 if (!bch2_trans_relock(trans))
1468                         ret = -EINTR;
1469         }
1470
1471         /*
1472          * XXX: figure out how far we might need to split,
1473          * instead of locking/reserving all the way to the root:
1474          */
1475         if (!bch2_btree_iter_upgrade(iter, U8_MAX)) {
1476                 trace_trans_restart_iter_upgrade(trans->ip);
1477                 ret = -EINTR;
1478                 goto out;
1479         }
1480
1481         as = bch2_btree_update_start(trans, iter->btree_id,
1482                 btree_update_reserve_required(c, b), flags,
1483                 !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1484         if (IS_ERR(as)) {
1485                 ret = PTR_ERR(as);
1486                 if (ret == -EAGAIN) {
1487                         BUG_ON(flags & BTREE_INSERT_NOUNLOCK);
1488                         bch2_trans_unlock(trans);
1489                         ret = -EINTR;
1490
1491                         trace_transaction_restart_ip(trans->ip, _THIS_IP_);
1492                 }
1493                 goto out;
1494         }
1495
1496         btree_split(as, b, iter, NULL, flags);
1497         bch2_btree_update_done(as);
1498
1499         /*
1500          * We haven't successfully inserted yet, so don't downgrade all the way
1501          * back to read locks;
1502          */
1503         __bch2_btree_iter_downgrade(iter, 1);
1504 out:
1505         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1506                 up_read(&c->gc_lock);
1507         closure_sync(&cl);
1508         return ret;
1509 }
1510
1511 void __bch2_foreground_maybe_merge(struct bch_fs *c,
1512                                    struct btree_iter *iter,
1513                                    unsigned level,
1514                                    unsigned flags,
1515                                    enum btree_node_sibling sib)
1516 {
1517         struct btree_trans *trans = iter->trans;
1518         struct btree_update *as;
1519         struct bkey_format_state new_s;
1520         struct bkey_format new_f;
1521         struct bkey_i delete;
1522         struct btree *b, *m, *n, *prev, *next, *parent;
1523         struct closure cl;
1524         size_t sib_u64s;
1525         int ret = 0;
1526
1527         BUG_ON(!btree_node_locked(iter, level));
1528
1529         closure_init_stack(&cl);
1530 retry:
1531         BUG_ON(!btree_node_locked(iter, level));
1532
1533         b = iter->l[level].b;
1534
1535         parent = btree_node_parent(iter, b);
1536         if (!parent)
1537                 goto out;
1538
1539         if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c))
1540                 goto out;
1541
1542         /* XXX: can't be holding read locks */
1543         m = bch2_btree_node_get_sibling(c, iter, b, sib);
1544         if (IS_ERR(m)) {
1545                 ret = PTR_ERR(m);
1546                 goto err;
1547         }
1548
1549         /* NULL means no sibling: */
1550         if (!m) {
1551                 b->sib_u64s[sib] = U16_MAX;
1552                 goto out;
1553         }
1554
1555         if (sib == btree_prev_sib) {
1556                 prev = m;
1557                 next = b;
1558         } else {
1559                 prev = b;
1560                 next = m;
1561         }
1562
1563         bch2_bkey_format_init(&new_s);
1564         __bch2_btree_calc_format(&new_s, b);
1565         __bch2_btree_calc_format(&new_s, m);
1566         new_f = bch2_bkey_format_done(&new_s);
1567
1568         sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1569                 btree_node_u64s_with_format(m, &new_f);
1570
1571         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1572                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1573                 sib_u64s /= 2;
1574                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1575         }
1576
1577         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1578         b->sib_u64s[sib] = sib_u64s;
1579
1580         if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c)) {
1581                 six_unlock_intent(&m->c.lock);
1582                 goto out;
1583         }
1584
1585         /* We're changing btree topology, doesn't mix with gc: */
1586         if (!(flags & BTREE_INSERT_GC_LOCK_HELD) &&
1587             !down_read_trylock(&c->gc_lock))
1588                 goto err_cycle_gc_lock;
1589
1590         if (!bch2_btree_iter_upgrade(iter, U8_MAX)) {
1591                 ret = -EINTR;
1592                 goto err_unlock;
1593         }
1594
1595         as = bch2_btree_update_start(trans, iter->btree_id,
1596                          btree_update_reserve_required(c, parent) + 1,
1597                          flags|
1598                          BTREE_INSERT_NOFAIL|
1599                          BTREE_INSERT_USE_RESERVE,
1600                          !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1601         if (IS_ERR(as)) {
1602                 ret = PTR_ERR(as);
1603                 goto err_unlock;
1604         }
1605
1606         trace_btree_merge(c, b);
1607
1608         bch2_btree_interior_update_will_free_node(as, b);
1609         bch2_btree_interior_update_will_free_node(as, m);
1610
1611         n = bch2_btree_node_alloc(as, b->c.level);
1612         bch2_btree_update_add_new_node(as, n);
1613
1614         btree_set_min(n, prev->data->min_key);
1615         btree_set_max(n, next->data->max_key);
1616         n->data->format         = new_f;
1617
1618         btree_node_set_format(n, new_f);
1619
1620         bch2_btree_sort_into(c, n, prev);
1621         bch2_btree_sort_into(c, n, next);
1622
1623         bch2_btree_build_aux_trees(n);
1624         six_unlock_write(&n->c.lock);
1625
1626         bkey_init(&delete.k);
1627         delete.k.p = prev->key.k.p;
1628         bch2_keylist_add(&as->parent_keys, &delete);
1629         bch2_keylist_add(&as->parent_keys, &n->key);
1630
1631         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1632
1633         bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1634
1635         bch2_btree_update_get_open_buckets(as, n);
1636
1637         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1638         bch2_btree_iter_node_drop(iter, b);
1639         bch2_btree_iter_node_drop(iter, m);
1640
1641         bch2_btree_iter_node_replace(iter, n);
1642
1643         bch2_btree_trans_verify_iters(trans, n);
1644
1645         bch2_btree_node_free_inmem(c, b, iter);
1646         bch2_btree_node_free_inmem(c, m, iter);
1647
1648         six_unlock_intent(&n->c.lock);
1649
1650         bch2_btree_update_done(as);
1651
1652         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1653                 up_read(&c->gc_lock);
1654 out:
1655         bch2_btree_trans_verify_locks(trans);
1656
1657         /*
1658          * Don't downgrade locks here: we're called after successful insert,
1659          * and the caller will downgrade locks after a successful insert
1660          * anyways (in case e.g. a split was required first)
1661          *
1662          * And we're also called when inserting into interior nodes in the
1663          * split path, and downgrading to read locks in there is potentially
1664          * confusing:
1665          */
1666         closure_sync(&cl);
1667         return;
1668
1669 err_cycle_gc_lock:
1670         six_unlock_intent(&m->c.lock);
1671
1672         if (flags & BTREE_INSERT_NOUNLOCK)
1673                 goto out;
1674
1675         bch2_trans_unlock(trans);
1676
1677         down_read(&c->gc_lock);
1678         up_read(&c->gc_lock);
1679         ret = -EINTR;
1680         goto err;
1681
1682 err_unlock:
1683         six_unlock_intent(&m->c.lock);
1684         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1685                 up_read(&c->gc_lock);
1686 err:
1687         BUG_ON(ret == -EAGAIN && (flags & BTREE_INSERT_NOUNLOCK));
1688
1689         if ((ret == -EAGAIN || ret == -EINTR) &&
1690             !(flags & BTREE_INSERT_NOUNLOCK)) {
1691                 bch2_trans_unlock(trans);
1692                 closure_sync(&cl);
1693                 ret = bch2_btree_iter_traverse(iter);
1694                 if (ret)
1695                         goto out;
1696
1697                 goto retry;
1698         }
1699
1700         goto out;
1701 }
1702
1703 static int __btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
1704                                 struct btree *b, unsigned flags,
1705                                 struct closure *cl)
1706 {
1707         struct btree *n, *parent = btree_node_parent(iter, b);
1708         struct btree_update *as;
1709
1710         as = bch2_btree_update_start(iter->trans, iter->btree_id,
1711                 (parent
1712                  ? btree_update_reserve_required(c, parent)
1713                  : 0) + 1,
1714                 flags, cl);
1715         if (IS_ERR(as)) {
1716                 trace_btree_gc_rewrite_node_fail(c, b);
1717                 return PTR_ERR(as);
1718         }
1719
1720         bch2_btree_interior_update_will_free_node(as, b);
1721
1722         n = bch2_btree_node_alloc_replacement(as, b);
1723         bch2_btree_update_add_new_node(as, n);
1724
1725         bch2_btree_build_aux_trees(n);
1726         six_unlock_write(&n->c.lock);
1727
1728         trace_btree_gc_rewrite_node(c, b);
1729
1730         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1731
1732         if (parent) {
1733                 bch2_keylist_add(&as->parent_keys, &n->key);
1734                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1735         } else {
1736                 bch2_btree_set_root(as, n, iter);
1737         }
1738
1739         bch2_btree_update_get_open_buckets(as, n);
1740
1741         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1742         bch2_btree_iter_node_drop(iter, b);
1743         bch2_btree_iter_node_replace(iter, n);
1744         bch2_btree_node_free_inmem(c, b, iter);
1745         six_unlock_intent(&n->c.lock);
1746
1747         bch2_btree_update_done(as);
1748         return 0;
1749 }
1750
1751 /**
1752  * bch_btree_node_rewrite - Rewrite/move a btree node
1753  *
1754  * Returns 0 on success, -EINTR or -EAGAIN on failure (i.e.
1755  * btree_check_reserve() has to wait)
1756  */
1757 int bch2_btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
1758                             __le64 seq, unsigned flags)
1759 {
1760         struct btree_trans *trans = iter->trans;
1761         struct closure cl;
1762         struct btree *b;
1763         int ret;
1764
1765         flags |= BTREE_INSERT_NOFAIL;
1766
1767         closure_init_stack(&cl);
1768
1769         bch2_btree_iter_upgrade(iter, U8_MAX);
1770
1771         if (!(flags & BTREE_INSERT_GC_LOCK_HELD)) {
1772                 if (!down_read_trylock(&c->gc_lock)) {
1773                         bch2_trans_unlock(trans);
1774                         down_read(&c->gc_lock);
1775                 }
1776         }
1777
1778         while (1) {
1779                 ret = bch2_btree_iter_traverse(iter);
1780                 if (ret)
1781                         break;
1782
1783                 b = bch2_btree_iter_peek_node(iter);
1784                 if (!b || b->data->keys.seq != seq)
1785                         break;
1786
1787                 ret = __btree_node_rewrite(c, iter, b, flags, &cl);
1788                 if (ret != -EAGAIN &&
1789                     ret != -EINTR)
1790                         break;
1791
1792                 bch2_trans_unlock(trans);
1793                 closure_sync(&cl);
1794         }
1795
1796         bch2_btree_iter_downgrade(iter);
1797
1798         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1799                 up_read(&c->gc_lock);
1800
1801         closure_sync(&cl);
1802         return ret;
1803 }
1804
1805 static void __bch2_btree_node_update_key(struct bch_fs *c,
1806                                          struct btree_update *as,
1807                                          struct btree_iter *iter,
1808                                          struct btree *b, struct btree *new_hash,
1809                                          struct bkey_i *new_key)
1810 {
1811         struct btree *parent;
1812         int ret;
1813
1814         btree_update_will_delete_key(as, &b->key);
1815         btree_update_will_add_key(as, new_key);
1816
1817         parent = btree_node_parent(iter, b);
1818         if (parent) {
1819                 if (new_hash) {
1820                         bkey_copy(&new_hash->key, new_key);
1821                         ret = bch2_btree_node_hash_insert(&c->btree_cache,
1822                                         new_hash, b->c.level, b->c.btree_id);
1823                         BUG_ON(ret);
1824                 }
1825
1826                 bch2_keylist_add(&as->parent_keys, new_key);
1827                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, 0);
1828
1829                 if (new_hash) {
1830                         mutex_lock(&c->btree_cache.lock);
1831                         bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
1832
1833                         bch2_btree_node_hash_remove(&c->btree_cache, b);
1834
1835                         bkey_copy(&b->key, new_key);
1836                         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1837                         BUG_ON(ret);
1838                         mutex_unlock(&c->btree_cache.lock);
1839                 } else {
1840                         bkey_copy(&b->key, new_key);
1841                 }
1842         } else {
1843                 BUG_ON(btree_node_root(c, b) != b);
1844
1845                 bch2_btree_node_lock_write(b, iter);
1846                 bkey_copy(&b->key, new_key);
1847
1848                 if (btree_ptr_hash_val(&b->key) != b->hash_val) {
1849                         mutex_lock(&c->btree_cache.lock);
1850                         bch2_btree_node_hash_remove(&c->btree_cache, b);
1851
1852                         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1853                         BUG_ON(ret);
1854                         mutex_unlock(&c->btree_cache.lock);
1855                 }
1856
1857                 btree_update_updated_root(as, b);
1858                 bch2_btree_node_unlock_write(b, iter);
1859         }
1860
1861         bch2_btree_update_done(as);
1862 }
1863
1864 int bch2_btree_node_update_key(struct bch_fs *c, struct btree_iter *iter,
1865                                struct btree *b,
1866                                struct bkey_i *new_key)
1867 {
1868         struct btree *parent = btree_node_parent(iter, b);
1869         struct btree_update *as = NULL;
1870         struct btree *new_hash = NULL;
1871         struct closure cl;
1872         int ret;
1873
1874         closure_init_stack(&cl);
1875
1876         if (!bch2_btree_iter_upgrade(iter, U8_MAX))
1877                 return -EINTR;
1878
1879         if (!down_read_trylock(&c->gc_lock)) {
1880                 bch2_trans_unlock(iter->trans);
1881                 down_read(&c->gc_lock);
1882
1883                 if (!bch2_trans_relock(iter->trans)) {
1884                         ret = -EINTR;
1885                         goto err;
1886                 }
1887         }
1888
1889         /*
1890          * check btree_ptr_hash_val() after @b is locked by
1891          * btree_iter_traverse():
1892          */
1893         if (btree_ptr_hash_val(new_key) != b->hash_val) {
1894                 /* bch2_btree_reserve_get will unlock */
1895                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1896                 if (ret) {
1897                         bch2_trans_unlock(iter->trans);
1898                         up_read(&c->gc_lock);
1899                         closure_sync(&cl);
1900                         down_read(&c->gc_lock);
1901
1902                         if (!bch2_trans_relock(iter->trans)) {
1903                                 ret = -EINTR;
1904                                 goto err;
1905                         }
1906                 }
1907
1908                 new_hash = bch2_btree_node_mem_alloc(c);
1909         }
1910 retry:
1911         as = bch2_btree_update_start(iter->trans, iter->btree_id,
1912                 parent ? btree_update_reserve_required(c, parent) : 0,
1913                 BTREE_INSERT_NOFAIL, &cl);
1914
1915         if (IS_ERR(as)) {
1916                 ret = PTR_ERR(as);
1917                 if (ret == -EAGAIN)
1918                         ret = -EINTR;
1919
1920                 if (ret == -EINTR) {
1921                         bch2_trans_unlock(iter->trans);
1922                         up_read(&c->gc_lock);
1923                         closure_sync(&cl);
1924                         down_read(&c->gc_lock);
1925
1926                         if (bch2_trans_relock(iter->trans))
1927                                 goto retry;
1928                 }
1929
1930                 goto err;
1931         }
1932
1933         ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(new_key));
1934         if (ret)
1935                 goto err_free_update;
1936
1937         __bch2_btree_node_update_key(c, as, iter, b, new_hash, new_key);
1938
1939         bch2_btree_iter_downgrade(iter);
1940 err:
1941         if (new_hash) {
1942                 mutex_lock(&c->btree_cache.lock);
1943                 list_move(&new_hash->list, &c->btree_cache.freeable);
1944                 mutex_unlock(&c->btree_cache.lock);
1945
1946                 six_unlock_write(&new_hash->c.lock);
1947                 six_unlock_intent(&new_hash->c.lock);
1948         }
1949         up_read(&c->gc_lock);
1950         closure_sync(&cl);
1951         return ret;
1952 err_free_update:
1953         bch2_btree_update_free(as);
1954         goto err;
1955 }
1956
1957 /* Init code: */
1958
1959 /*
1960  * Only for filesystem bringup, when first reading the btree roots or allocating
1961  * btree roots when initializing a new filesystem:
1962  */
1963 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
1964 {
1965         BUG_ON(btree_node_root(c, b));
1966
1967         bch2_btree_set_root_inmem(c, b);
1968 }
1969
1970 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
1971 {
1972         struct closure cl;
1973         struct btree *b;
1974         int ret;
1975
1976         closure_init_stack(&cl);
1977
1978         do {
1979                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1980                 closure_sync(&cl);
1981         } while (ret);
1982
1983         b = bch2_btree_node_mem_alloc(c);
1984         bch2_btree_cache_cannibalize_unlock(c);
1985
1986         set_btree_node_fake(b);
1987         set_btree_node_need_rewrite(b);
1988         b->c.level      = 0;
1989         b->c.btree_id   = id;
1990
1991         bkey_btree_ptr_init(&b->key);
1992         b->key.k.p = POS_MAX;
1993         *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
1994
1995         bch2_bset_init_first(b, &b->data->keys);
1996         bch2_btree_build_aux_trees(b);
1997
1998         b->data->flags = 0;
1999         btree_set_min(b, POS_MIN);
2000         btree_set_max(b, POS_MAX);
2001         b->data->format = bch2_btree_calc_format(b);
2002         btree_node_set_format(b, b->data->format);
2003
2004         ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2005                                           b->c.level, b->c.btree_id);
2006         BUG_ON(ret);
2007
2008         bch2_btree_set_root_inmem(c, b);
2009
2010         six_unlock_write(&b->c.lock);
2011         six_unlock_intent(&b->c.lock);
2012 }
2013
2014 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2015 {
2016         struct btree_update *as;
2017
2018         mutex_lock(&c->btree_interior_update_lock);
2019         list_for_each_entry(as, &c->btree_interior_update_list, list)
2020                 pr_buf(out, "%p m %u w %u r %u j %llu\n",
2021                        as,
2022                        as->mode,
2023                        as->nodes_written,
2024                        atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2025                        as->journal.seq);
2026         mutex_unlock(&c->btree_interior_update_lock);
2027 }
2028
2029 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *c)
2030 {
2031         size_t ret = 0;
2032         struct list_head *i;
2033
2034         mutex_lock(&c->btree_interior_update_lock);
2035         list_for_each(i, &c->btree_interior_update_list)
2036                 ret++;
2037         mutex_unlock(&c->btree_interior_update_lock);
2038
2039         return ret;
2040 }
2041
2042 void bch2_journal_entries_to_btree_roots(struct bch_fs *c, struct jset *jset)
2043 {
2044         struct btree_root *r;
2045         struct jset_entry *entry;
2046
2047         mutex_lock(&c->btree_root_lock);
2048
2049         vstruct_for_each(jset, entry)
2050                 if (entry->type == BCH_JSET_ENTRY_btree_root) {
2051                         r = &c->btree_roots[entry->btree_id];
2052                         r->level = entry->level;
2053                         r->alive = true;
2054                         bkey_copy(&r->key, &entry->start[0]);
2055                 }
2056
2057         mutex_unlock(&c->btree_root_lock);
2058 }
2059
2060 struct jset_entry *
2061 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2062                                     struct jset_entry *start,
2063                                     struct jset_entry *end)
2064 {
2065         struct jset_entry *entry;
2066         unsigned long have = 0;
2067         unsigned i;
2068
2069         for (entry = start; entry < end; entry = vstruct_next(entry))
2070                 if (entry->type == BCH_JSET_ENTRY_btree_root)
2071                         __set_bit(entry->btree_id, &have);
2072
2073         mutex_lock(&c->btree_root_lock);
2074
2075         for (i = 0; i < BTREE_ID_NR; i++)
2076                 if (c->btree_roots[i].alive && !test_bit(i, &have)) {
2077                         journal_entry_set(end,
2078                                           BCH_JSET_ENTRY_btree_root,
2079                                           i, c->btree_roots[i].level,
2080                                           &c->btree_roots[i].key,
2081                                           c->btree_roots[i].key.u64s);
2082                         end = vstruct_next(end);
2083                 }
2084
2085         mutex_unlock(&c->btree_root_lock);
2086
2087         return end;
2088 }
2089
2090 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2091 {
2092         if (c->btree_interior_update_worker)
2093                 destroy_workqueue(c->btree_interior_update_worker);
2094         mempool_exit(&c->btree_interior_update_pool);
2095 }
2096
2097 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2098 {
2099         mutex_init(&c->btree_reserve_cache_lock);
2100         INIT_LIST_HEAD(&c->btree_interior_update_list);
2101         INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2102         mutex_init(&c->btree_interior_update_lock);
2103         INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2104
2105         c->btree_interior_update_worker =
2106                 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2107         if (!c->btree_interior_update_worker)
2108                 return -ENOMEM;
2109
2110         return mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2111                                          sizeof(struct btree_update));
2112 }