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                 as->prealloc_nodes[as->nr_prealloc_nodes++] = b;
 441         }
 442
 443         bch2_btree_cache_cannibalize_unlock(c);
 444         return 0;
 445 err_free:
 446         bch2_btree_cache_cannibalize_unlock(c);
 447         trace_btree_reserve_get_fail(c, nr_nodes, cl);
 448         return ret;
 449 }
 450
 451 /* Asynchronous interior node update machinery */
 452
 453 static void bch2_btree_update_free(struct btree_update *as)
 454 {
 455         struct bch_fs *c = as->c;
 456
 457         if (as->took_gc_lock)
 458                 up_read(&c->gc_lock);
 459         as->took_gc_lock = false;
 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         if (as->took_gc_lock)
 897                 up_read(&as->c->gc_lock);
 898         as->took_gc_lock = false;
 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_iter *iter, unsigned level,
 907                         unsigned nr_nodes, unsigned flags)
 908 {
 909         struct btree_trans *trans = iter->trans;
 910         struct bch_fs *c = trans->c;
 911         struct btree_update *as;
 912         struct closure cl;
 913         int disk_res_flags = (flags & BTREE_INSERT_NOFAIL)
 914                 ? BCH_DISK_RESERVATION_NOFAIL : 0;
 915         int journal_flags = 0;
 916         int ret = 0;
 917
 918         if (flags & BTREE_INSERT_JOURNAL_RESERVED)
 919                 journal_flags |= JOURNAL_RES_GET_RESERVED;
 920
 921         closure_init_stack(&cl);
 922 retry:
 923         /*
 924          * This check isn't necessary for correctness - it's just to potentially
 925          * prevent us from doing a lot of work that'll end up being wasted:
 926          */
 927         ret = bch2_journal_error(&c->journal);
 928         if (ret)
 929                 return ERR_PTR(ret);
 930
 931         /*
 932          * XXX: figure out how far we might need to split,
 933          * instead of locking/reserving all the way to the root:
 934          */
 935         if (!bch2_btree_iter_upgrade(iter, U8_MAX)) {
 936                 trace_trans_restart_iter_upgrade(trans->ip);
 937                 return ERR_PTR(-EINTR);
 938         }
 939
 940         if (flags & BTREE_INSERT_GC_LOCK_HELD)
 941                 lockdep_assert_held(&c->gc_lock);
 942         else if (!down_read_trylock(&c->gc_lock)) {
 943                 if (flags & BTREE_INSERT_NOUNLOCK)
 944                         return ERR_PTR(-EINTR);
 945
 946                 bch2_trans_unlock(trans);
 947                 down_read(&c->gc_lock);
 948                 if (!bch2_trans_relock(trans)) {
 949                         up_read(&c->gc_lock);
 950                         return ERR_PTR(-EINTR);
 951                 }
 952         }
 953
 954         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
 955         memset(as, 0, sizeof(*as));
 956         closure_init(&as->cl, NULL);
 957         as->c           = c;
 958         as->mode        = BTREE_INTERIOR_NO_UPDATE;
 959         as->took_gc_lock = !(flags & BTREE_INSERT_GC_LOCK_HELD);
 960         as->btree_id    = iter->btree_id;
 961         INIT_LIST_HEAD(&as->list);
 962         INIT_LIST_HEAD(&as->unwritten_list);
 963         INIT_LIST_HEAD(&as->write_blocked_list);
 964         bch2_keylist_init(&as->old_keys, as->_old_keys);
 965         bch2_keylist_init(&as->new_keys, as->_new_keys);
 966         bch2_keylist_init(&as->parent_keys, as->inline_keys);
 967
 968         ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
 969                                       BTREE_UPDATE_JOURNAL_RES,
 970                                       journal_flags|JOURNAL_RES_GET_NONBLOCK);
 971         if (ret == -EAGAIN) {
 972                 /*
 973                  * this would be cleaner if bch2_journal_preres_get() took a
 974                  * closure argument
 975                  */
 976                 if (flags & BTREE_INSERT_NOUNLOCK) {
 977                         trace_trans_restart_journal_preres_get(trans->ip);
 978                         ret = -EINTR;
 979                         goto err;
 980                 }
 981
 982                 bch2_trans_unlock(trans);
 983
 984                 if (flags & BTREE_INSERT_JOURNAL_RECLAIM) {
 985                         bch2_btree_update_free(as);
 986                         return ERR_PTR(ret);
 987                 }
 988
 989                 ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
 990                                 BTREE_UPDATE_JOURNAL_RES,
 991                                 journal_flags);
 992                 if (ret) {
 993                         trace_trans_restart_journal_preres_get(trans->ip);
 994                         goto err;
 995                 }
 996
 997                 if (!bch2_trans_relock(trans)) {
 998                         ret = -EINTR;
 999                         goto err;
1000                 }
1001         }
1002
1003         ret = bch2_disk_reservation_get(c, &as->disk_res,
1004                         nr_nodes * c->opts.btree_node_size,
1005                         c->opts.metadata_replicas,
1006                         disk_res_flags);
1007         if (ret)
1008                 goto err;
1009
1010         ret = bch2_btree_reserve_get(as, nr_nodes, flags,
1011                 !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1012         if (ret)
1013                 goto err;
1014
1015         bch2_journal_pin_add(&c->journal,
1016                              atomic64_read(&c->journal.seq),
1017                              &as->journal, NULL);
1018
1019         mutex_lock(&c->btree_interior_update_lock);
1020         list_add_tail(&as->list, &c->btree_interior_update_list);
1021         mutex_unlock(&c->btree_interior_update_lock);
1022
1023         return as;
1024 err:
1025         bch2_btree_update_free(as);
1026
1027         if (ret == -EAGAIN) {
1028                 BUG_ON(flags & BTREE_INSERT_NOUNLOCK);
1029
1030                 bch2_trans_unlock(trans);
1031                 closure_sync(&cl);
1032                 ret = -EINTR;
1033         }
1034
1035         if (ret == -EINTR && bch2_trans_relock(trans))
1036                 goto retry;
1037
1038         return ERR_PTR(ret);
1039 }
1040
1041 /* Btree root updates: */
1042
1043 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1044 {
1045         /* Root nodes cannot be reaped */
1046         mutex_lock(&c->btree_cache.lock);
1047         list_del_init(&b->list);
1048         mutex_unlock(&c->btree_cache.lock);
1049
1050         if (b->c.level)
1051                 six_lock_pcpu_alloc(&b->c.lock);
1052         else
1053                 six_lock_pcpu_free(&b->c.lock);
1054
1055         mutex_lock(&c->btree_root_lock);
1056         BUG_ON(btree_node_root(c, b) &&
1057                (b->c.level < btree_node_root(c, b)->c.level ||
1058                 !btree_node_dying(btree_node_root(c, b))));
1059
1060         btree_node_root(c, b) = b;
1061         mutex_unlock(&c->btree_root_lock);
1062
1063         bch2_recalc_btree_reserve(c);
1064 }
1065
1066 /**
1067  * bch_btree_set_root - update the root in memory and on disk
1068  *
1069  * To ensure forward progress, the current task must not be holding any
1070  * btree node write locks. However, you must hold an intent lock on the
1071  * old root.
1072  *
1073  * Note: This allocates a journal entry but doesn't add any keys to
1074  * it.  All the btree roots are part of every journal write, so there
1075  * is nothing new to be done.  This just guarantees that there is a
1076  * journal write.
1077  */
1078 static void bch2_btree_set_root(struct btree_update *as, struct btree *b,
1079                                 struct btree_iter *iter)
1080 {
1081         struct bch_fs *c = as->c;
1082         struct btree *old;
1083
1084         trace_btree_set_root(c, b);
1085         BUG_ON(!b->written &&
1086                !test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags));
1087
1088         old = btree_node_root(c, b);
1089
1090         /*
1091          * Ensure no one is using the old root while we switch to the
1092          * new root:
1093          */
1094         bch2_btree_node_lock_write(old, iter);
1095
1096         bch2_btree_set_root_inmem(c, b);
1097
1098         btree_update_updated_root(as, b);
1099
1100         /*
1101          * Unlock old root after new root is visible:
1102          *
1103          * The new root isn't persistent, but that's ok: we still have
1104          * an intent lock on the new root, and any updates that would
1105          * depend on the new root would have to update the new root.
1106          */
1107         bch2_btree_node_unlock_write(old, iter);
1108 }
1109
1110 /* Interior node updates: */
1111
1112 static void bch2_insert_fixup_btree_ptr(struct btree_update *as, struct btree *b,
1113                                         struct btree_iter *iter,
1114                                         struct bkey_i *insert,
1115                                         struct btree_node_iter *node_iter)
1116 {
1117         struct bch_fs *c = as->c;
1118         struct bkey_packed *k;
1119         const char *invalid;
1120
1121         invalid = bch2_bkey_invalid(c, bkey_i_to_s_c(insert), btree_node_type(b)) ?:
1122                 bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert));
1123         if (invalid) {
1124                 char buf[160];
1125
1126                 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(insert));
1127                 bch2_fs_inconsistent(c, "inserting invalid bkey %s: %s", buf, invalid);
1128                 dump_stack();
1129         }
1130
1131         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1132                ARRAY_SIZE(as->journal_entries));
1133
1134         as->journal_u64s +=
1135                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1136                                   BCH_JSET_ENTRY_btree_keys,
1137                                   b->c.btree_id, b->c.level,
1138                                   insert, insert->k.u64s);
1139
1140         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1141                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1142                 bch2_btree_node_iter_advance(node_iter, b);
1143
1144         bch2_btree_bset_insert_key(iter, b, node_iter, insert);
1145         set_btree_node_dirty(c, b);
1146         set_btree_node_need_write(b);
1147 }
1148
1149 /*
1150  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1151  * node)
1152  */
1153 static struct btree *__btree_split_node(struct btree_update *as,
1154                                         struct btree *n1,
1155                                         struct btree_iter *iter)
1156 {
1157         struct bkey_format_state s;
1158         size_t nr_packed = 0, nr_unpacked = 0;
1159         struct btree *n2;
1160         struct bset *set1, *set2;
1161         struct bkey_packed *k, *set2_start, *set2_end, *out, *prev = NULL;
1162         struct bpos n1_pos;
1163
1164         n2 = bch2_btree_node_alloc(as, n1->c.level);
1165         bch2_btree_update_add_new_node(as, n2);
1166
1167         n2->data->max_key       = n1->data->max_key;
1168         n2->data->format        = n1->format;
1169         SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data));
1170         n2->key.k.p = n1->key.k.p;
1171
1172         set1 = btree_bset_first(n1);
1173         set2 = btree_bset_first(n2);
1174
1175         /*
1176          * Has to be a linear search because we don't have an auxiliary
1177          * search tree yet
1178          */
1179         k = set1->start;
1180         while (1) {
1181                 struct bkey_packed *n = bkey_next(k);
1182
1183                 if (n == vstruct_last(set1))
1184                         break;
1185                 if (k->_data - set1->_data >= (le16_to_cpu(set1->u64s) * 3) / 5)
1186                         break;
1187
1188                 if (bkey_packed(k))
1189                         nr_packed++;
1190                 else
1191                         nr_unpacked++;
1192
1193                 prev = k;
1194                 k = n;
1195         }
1196
1197         BUG_ON(!prev);
1198         set2_start      = k;
1199         set2_end        = vstruct_last(set1);
1200
1201         set1->u64s = cpu_to_le16((u64 *) set2_start - set1->_data);
1202         set_btree_bset_end(n1, n1->set);
1203
1204         n1->nr.live_u64s        = le16_to_cpu(set1->u64s);
1205         n1->nr.bset_u64s[0]     = le16_to_cpu(set1->u64s);
1206         n1->nr.packed_keys      = nr_packed;
1207         n1->nr.unpacked_keys    = nr_unpacked;
1208
1209         n1_pos = bkey_unpack_pos(n1, prev);
1210         if (as->c->sb.version < bcachefs_metadata_version_snapshot)
1211                 n1_pos.snapshot = U32_MAX;
1212
1213         btree_set_max(n1, n1_pos);
1214         btree_set_min(n2, bpos_successor(n1->key.k.p));
1215
1216         bch2_bkey_format_init(&s);
1217         bch2_bkey_format_add_pos(&s, n2->data->min_key);
1218         bch2_bkey_format_add_pos(&s, n2->data->max_key);
1219
1220         for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1221                 struct bkey uk = bkey_unpack_key(n1, k);
1222                 bch2_bkey_format_add_key(&s, &uk);
1223         }
1224
1225         n2->data->format = bch2_bkey_format_done(&s);
1226         btree_node_set_format(n2, n2->data->format);
1227
1228         out = set2->start;
1229         memset(&n2->nr, 0, sizeof(n2->nr));
1230
1231         for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1232                 BUG_ON(!bch2_bkey_transform(&n2->format, out, bkey_packed(k)
1233                                        ? &n1->format : &bch2_bkey_format_current, k));
1234                 out->format = KEY_FORMAT_LOCAL_BTREE;
1235                 btree_keys_account_key_add(&n2->nr, 0, out);
1236                 out = bkey_next(out);
1237         }
1238
1239         set2->u64s = cpu_to_le16((u64 *) out - set2->_data);
1240         set_btree_bset_end(n2, n2->set);
1241
1242         BUG_ON(!set1->u64s);
1243         BUG_ON(!set2->u64s);
1244
1245         btree_node_reset_sib_u64s(n1);
1246         btree_node_reset_sib_u64s(n2);
1247
1248         bch2_verify_btree_nr_keys(n1);
1249         bch2_verify_btree_nr_keys(n2);
1250
1251         if (n1->c.level) {
1252                 btree_node_interior_verify(as->c, n1);
1253                 btree_node_interior_verify(as->c, n2);
1254         }
1255
1256         return n2;
1257 }
1258
1259 /*
1260  * For updates to interior nodes, we've got to do the insert before we split
1261  * because the stuff we're inserting has to be inserted atomically. Post split,
1262  * the keys might have to go in different nodes and the split would no longer be
1263  * atomic.
1264  *
1265  * Worse, if the insert is from btree node coalescing, if we do the insert after
1266  * we do the split (and pick the pivot) - the pivot we pick might be between
1267  * nodes that were coalesced, and thus in the middle of a child node post
1268  * coalescing:
1269  */
1270 static void btree_split_insert_keys(struct btree_update *as, struct btree *b,
1271                                     struct btree_iter *iter,
1272                                     struct keylist *keys)
1273 {
1274         struct btree_node_iter node_iter;
1275         struct bkey_i *k = bch2_keylist_front(keys);
1276         struct bkey_packed *src, *dst, *n;
1277         struct bset *i;
1278
1279         BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1280
1281         bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
1282
1283         while (!bch2_keylist_empty(keys)) {
1284                 k = bch2_keylist_front(keys);
1285
1286                 bch2_insert_fixup_btree_ptr(as, b, iter, k, &node_iter);
1287                 bch2_keylist_pop_front(keys);
1288         }
1289
1290         /*
1291          * We can't tolerate whiteouts here - with whiteouts there can be
1292          * duplicate keys, and it would be rather bad if we picked a duplicate
1293          * for the pivot:
1294          */
1295         i = btree_bset_first(b);
1296         src = dst = i->start;
1297         while (src != vstruct_last(i)) {
1298                 n = bkey_next(src);
1299                 if (!bkey_deleted(src)) {
1300                         memmove_u64s_down(dst, src, src->u64s);
1301                         dst = bkey_next(dst);
1302                 }
1303                 src = n;
1304         }
1305
1306         /* Also clear out the unwritten whiteouts area: */
1307         b->whiteout_u64s = 0;
1308
1309         i->u64s = cpu_to_le16((u64 *) dst - i->_data);
1310         set_btree_bset_end(b, b->set);
1311
1312         BUG_ON(b->nsets != 1 ||
1313                b->nr.live_u64s != le16_to_cpu(btree_bset_first(b)->u64s));
1314
1315         btree_node_interior_verify(as->c, b);
1316 }
1317
1318 static void btree_split(struct btree_update *as, struct btree *b,
1319                         struct btree_iter *iter, struct keylist *keys,
1320                         unsigned flags)
1321 {
1322         struct bch_fs *c = as->c;
1323         struct btree *parent = btree_node_parent(iter, b);
1324         struct btree *n1, *n2 = NULL, *n3 = NULL;
1325         u64 start_time = local_clock();
1326
1327         BUG_ON(!parent && (b != btree_node_root(c, b)));
1328         BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->c.level));
1329
1330         bch2_btree_interior_update_will_free_node(as, b);
1331
1332         n1 = bch2_btree_node_alloc_replacement(as, b);
1333         bch2_btree_update_add_new_node(as, n1);
1334
1335         if (keys)
1336                 btree_split_insert_keys(as, n1, iter, keys);
1337
1338         if (bset_u64s(&n1->set[0]) > BTREE_SPLIT_THRESHOLD(c)) {
1339                 trace_btree_split(c, b);
1340
1341                 n2 = __btree_split_node(as, n1, iter);
1342
1343                 bch2_btree_build_aux_trees(n2);
1344                 bch2_btree_build_aux_trees(n1);
1345                 six_unlock_write(&n2->c.lock);
1346                 six_unlock_write(&n1->c.lock);
1347
1348                 bch2_btree_node_write(c, n2, SIX_LOCK_intent);
1349
1350                 /*
1351                  * Note that on recursive parent_keys == keys, so we
1352                  * can't start adding new keys to parent_keys before emptying it
1353                  * out (which we did with btree_split_insert_keys() above)
1354                  */
1355                 bch2_keylist_add(&as->parent_keys, &n1->key);
1356                 bch2_keylist_add(&as->parent_keys, &n2->key);
1357
1358                 if (!parent) {
1359                         /* Depth increases, make a new root */
1360                         n3 = __btree_root_alloc(as, b->c.level + 1);
1361
1362                         n3->sib_u64s[0] = U16_MAX;
1363                         n3->sib_u64s[1] = U16_MAX;
1364
1365                         btree_split_insert_keys(as, n3, iter, &as->parent_keys);
1366
1367                         bch2_btree_node_write(c, n3, SIX_LOCK_intent);
1368                 }
1369         } else {
1370                 trace_btree_compact(c, b);
1371
1372                 bch2_btree_build_aux_trees(n1);
1373                 six_unlock_write(&n1->c.lock);
1374
1375                 if (parent)
1376                         bch2_keylist_add(&as->parent_keys, &n1->key);
1377         }
1378
1379         bch2_btree_node_write(c, n1, SIX_LOCK_intent);
1380
1381         /* New nodes all written, now make them visible: */
1382
1383         if (parent) {
1384                 /* Split a non root node */
1385                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1386         } else if (n3) {
1387                 bch2_btree_set_root(as, n3, iter);
1388         } else {
1389                 /* Root filled up but didn't need to be split */
1390                 bch2_btree_set_root(as, n1, iter);
1391         }
1392
1393         bch2_btree_update_get_open_buckets(as, n1);
1394         if (n2)
1395                 bch2_btree_update_get_open_buckets(as, n2);
1396         if (n3)
1397                 bch2_btree_update_get_open_buckets(as, n3);
1398
1399         /* Successful split, update the iterator to point to the new nodes: */
1400
1401         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1402         bch2_btree_iter_node_drop(iter, b);
1403         if (n3)
1404                 bch2_btree_iter_node_replace(iter, n3);
1405         if (n2)
1406                 bch2_btree_iter_node_replace(iter, n2);
1407         bch2_btree_iter_node_replace(iter, n1);
1408
1409         /*
1410          * The old node must be freed (in memory) _before_ unlocking the new
1411          * nodes - else another thread could re-acquire a read lock on the old
1412          * node after another thread has locked and updated the new node, thus
1413          * seeing stale data:
1414          */
1415         bch2_btree_node_free_inmem(c, b, iter);
1416
1417         if (n3)
1418                 six_unlock_intent(&n3->c.lock);
1419         if (n2)
1420                 six_unlock_intent(&n2->c.lock);
1421         six_unlock_intent(&n1->c.lock);
1422
1423         bch2_btree_trans_verify_locks(iter->trans);
1424
1425         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_split],
1426                                start_time);
1427 }
1428
1429 static void
1430 bch2_btree_insert_keys_interior(struct btree_update *as, struct btree *b,
1431                                 struct btree_iter *iter, struct keylist *keys)
1432 {
1433         struct btree_iter *linked;
1434         struct btree_node_iter node_iter;
1435         struct bkey_i *insert = bch2_keylist_front(keys);
1436         struct bkey_packed *k;
1437
1438         /* Don't screw up @iter's position: */
1439         node_iter = iter->l[b->c.level].iter;
1440
1441         /*
1442          * btree_split(), btree_gc_coalesce() will insert keys before
1443          * the iterator's current position - they know the keys go in
1444          * the node the iterator points to:
1445          */
1446         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1447                (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1448                 ;
1449
1450         for_each_keylist_key(keys, insert)
1451                 bch2_insert_fixup_btree_ptr(as, b, iter, insert, &node_iter);
1452
1453         btree_update_updated_node(as, b);
1454
1455         trans_for_each_iter_with_node(iter->trans, b, linked)
1456                 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1457
1458         bch2_btree_trans_verify_iters(iter->trans, b);
1459 }
1460
1461 /**
1462  * bch_btree_insert_node - insert bkeys into a given btree node
1463  *
1464  * @iter:               btree iterator
1465  * @keys:               list of keys to insert
1466  * @hook:               insert callback
1467  * @persistent:         if not null, @persistent will wait on journal write
1468  *
1469  * Inserts as many keys as it can into a given btree node, splitting it if full.
1470  * If a split occurred, this function will return early. This can only happen
1471  * for leaf nodes -- inserts into interior nodes have to be atomic.
1472  */
1473 void bch2_btree_insert_node(struct btree_update *as, struct btree *b,
1474                             struct btree_iter *iter, struct keylist *keys,
1475                             unsigned flags)
1476 {
1477         struct bch_fs *c = as->c;
1478         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1479         int old_live_u64s = b->nr.live_u64s;
1480         int live_u64s_added, u64s_added;
1481
1482         lockdep_assert_held(&c->gc_lock);
1483         BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->c.level));
1484         BUG_ON(!b->c.level);
1485         BUG_ON(!as || as->b);
1486         bch2_verify_keylist_sorted(keys);
1487
1488         bch2_btree_node_lock_for_insert(c, b, iter);
1489
1490         if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1491                 bch2_btree_node_unlock_write(b, iter);
1492                 goto split;
1493         }
1494
1495         btree_node_interior_verify(c, b);
1496
1497         bch2_btree_insert_keys_interior(as, b, iter, keys);
1498
1499         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1500         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1501
1502         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1503                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1504         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1505                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1506
1507         if (u64s_added > live_u64s_added &&
1508             bch2_maybe_compact_whiteouts(c, b))
1509                 bch2_btree_iter_reinit_node(iter, b);
1510
1511         bch2_btree_node_unlock_write(b, iter);
1512
1513         btree_node_interior_verify(c, b);
1514         return;
1515 split:
1516         btree_split(as, b, iter, keys, flags);
1517 }
1518
1519 int bch2_btree_split_leaf(struct bch_fs *c, struct btree_iter *iter,
1520                           unsigned flags)
1521 {
1522         struct btree *b = iter_l(iter)->b;
1523         struct btree_update *as;
1524         unsigned l;
1525         int ret = 0;
1526
1527         as = bch2_btree_update_start(iter, iter->level,
1528                 btree_update_reserve_required(c, b), flags);
1529         if (IS_ERR(as))
1530                 return PTR_ERR(as);
1531
1532         btree_split(as, b, iter, NULL, flags);
1533         bch2_btree_update_done(as);
1534
1535         for (l = iter->level + 1; btree_iter_node(iter, l) && !ret; l++)
1536                 ret = bch2_foreground_maybe_merge(c, iter, l, flags);
1537
1538         return ret;
1539 }
1540
1541 int __bch2_foreground_maybe_merge(struct bch_fs *c,
1542                                   struct btree_iter *iter,
1543                                   unsigned level,
1544                                   unsigned flags,
1545                                   enum btree_node_sibling sib)
1546 {
1547         struct btree_trans *trans = iter->trans;
1548         struct btree_iter *sib_iter = NULL;
1549         struct btree_update *as;
1550         struct bkey_format_state new_s;
1551         struct bkey_format new_f;
1552         struct bkey_i delete;
1553         struct btree *b, *m, *n, *prev, *next, *parent;
1554         struct bpos sib_pos;
1555         size_t sib_u64s;
1556         int ret = 0, ret2 = 0;
1557
1558         BUG_ON(!btree_node_locked(iter, level));
1559 retry:
1560         ret = bch2_btree_iter_traverse(iter);
1561         if (ret)
1562                 goto err;
1563
1564         BUG_ON(!btree_node_locked(iter, level));
1565
1566         b = iter->l[level].b;
1567
1568         if ((sib == btree_prev_sib && !bpos_cmp(b->data->min_key, POS_MIN)) ||
1569             (sib == btree_next_sib && !bpos_cmp(b->data->max_key, POS_MAX))) {
1570                 b->sib_u64s[sib] = U16_MAX;
1571                 goto out;
1572         }
1573
1574         sib_pos = sib == btree_prev_sib
1575                 ? bpos_predecessor(b->data->min_key)
1576                 : bpos_successor(b->data->max_key);
1577
1578         sib_iter = bch2_trans_get_node_iter(trans, iter->btree_id,
1579                                             sib_pos, U8_MAX, level,
1580                                             BTREE_ITER_INTENT);
1581         ret = bch2_btree_iter_traverse(sib_iter);
1582         if (ret)
1583                 goto err;
1584
1585         m = sib_iter->l[level].b;
1586
1587         if (btree_node_parent(iter, b) !=
1588             btree_node_parent(sib_iter, m)) {
1589                 b->sib_u64s[sib] = U16_MAX;
1590                 goto out;
1591         }
1592
1593         if (sib == btree_prev_sib) {
1594                 prev = m;
1595                 next = b;
1596         } else {
1597                 prev = b;
1598                 next = m;
1599         }
1600
1601         BUG_ON(bkey_cmp(bpos_successor(prev->data->max_key), next->data->min_key));
1602
1603         bch2_bkey_format_init(&new_s);
1604         bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
1605         __bch2_btree_calc_format(&new_s, prev);
1606         __bch2_btree_calc_format(&new_s, next);
1607         bch2_bkey_format_add_pos(&new_s, next->data->max_key);
1608         new_f = bch2_bkey_format_done(&new_s);
1609
1610         sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1611                 btree_node_u64s_with_format(m, &new_f);
1612
1613         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1614                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1615                 sib_u64s /= 2;
1616                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1617         }
1618
1619         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1620         sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
1621         b->sib_u64s[sib] = sib_u64s;
1622
1623         if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
1624                 goto out;
1625
1626         parent = btree_node_parent(iter, b);
1627         as = bch2_btree_update_start(iter, level,
1628                          btree_update_reserve_required(c, parent) + 1,
1629                          flags|
1630                          BTREE_INSERT_NOFAIL|
1631                          BTREE_INSERT_USE_RESERVE);
1632         ret = PTR_ERR_OR_ZERO(as);
1633         if (ret)
1634                 goto err;
1635
1636         trace_btree_merge(c, b);
1637
1638         bch2_btree_interior_update_will_free_node(as, b);
1639         bch2_btree_interior_update_will_free_node(as, m);
1640
1641         n = bch2_btree_node_alloc(as, b->c.level);
1642         bch2_btree_update_add_new_node(as, n);
1643
1644         btree_set_min(n, prev->data->min_key);
1645         btree_set_max(n, next->data->max_key);
1646         n->data->format         = new_f;
1647
1648         btree_node_set_format(n, new_f);
1649
1650         bch2_btree_sort_into(c, n, prev);
1651         bch2_btree_sort_into(c, n, next);
1652
1653         bch2_btree_build_aux_trees(n);
1654         six_unlock_write(&n->c.lock);
1655
1656         bkey_init(&delete.k);
1657         delete.k.p = prev->key.k.p;
1658         bch2_keylist_add(&as->parent_keys, &delete);
1659         bch2_keylist_add(&as->parent_keys, &n->key);
1660
1661         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1662
1663         bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1664
1665         bch2_btree_update_get_open_buckets(as, n);
1666
1667         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1668         six_lock_increment(&m->c.lock, SIX_LOCK_intent);
1669         bch2_btree_iter_node_drop(iter, b);
1670         bch2_btree_iter_node_drop(iter, m);
1671
1672         bch2_btree_iter_node_replace(iter, n);
1673
1674         bch2_btree_trans_verify_iters(trans, n);
1675
1676         bch2_btree_node_free_inmem(c, b, iter);
1677         bch2_btree_node_free_inmem(c, m, iter);
1678
1679         six_unlock_intent(&n->c.lock);
1680
1681         bch2_btree_update_done(as);
1682 out:
1683         bch2_btree_trans_verify_locks(trans);
1684         bch2_trans_iter_free(trans, sib_iter);
1685
1686         /*
1687          * Don't downgrade locks here: we're called after successful insert,
1688          * and the caller will downgrade locks after a successful insert
1689          * anyways (in case e.g. a split was required first)
1690          *
1691          * And we're also called when inserting into interior nodes in the
1692          * split path, and downgrading to read locks in there is potentially
1693          * confusing:
1694          */
1695         return ret ?: ret2;
1696 err:
1697         bch2_trans_iter_put(trans, sib_iter);
1698         sib_iter = NULL;
1699
1700         if (ret == -EINTR && bch2_trans_relock(trans))
1701                 goto retry;
1702
1703         if (ret == -EINTR && !(flags & BTREE_INSERT_NOUNLOCK)) {
1704                 ret2 = ret;
1705                 ret = bch2_btree_iter_traverse_all(trans);
1706                 if (!ret)
1707                         goto retry;
1708         }
1709
1710         goto out;
1711 }
1712
1713 /**
1714  * bch_btree_node_rewrite - Rewrite/move a btree node
1715  */
1716 int bch2_btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
1717                             __le64 seq, unsigned flags)
1718 {
1719         struct btree *b, *n, *parent;
1720         struct btree_update *as;
1721         int ret;
1722
1723         flags |= BTREE_INSERT_NOFAIL;
1724 retry:
1725         ret = bch2_btree_iter_traverse(iter);
1726         if (ret)
1727                 goto out;
1728
1729         b = bch2_btree_iter_peek_node(iter);
1730         if (!b || b->data->keys.seq != seq)
1731                 goto out;
1732
1733         parent = btree_node_parent(iter, b);
1734         as = bch2_btree_update_start(iter, b->c.level,
1735                 (parent
1736                  ? btree_update_reserve_required(c, parent)
1737                  : 0) + 1,
1738                 flags);
1739         ret = PTR_ERR_OR_ZERO(as);
1740         if (ret == -EINTR)
1741                 goto retry;
1742         if (ret) {
1743                 trace_btree_gc_rewrite_node_fail(c, b);
1744                 goto out;
1745         }
1746
1747         bch2_btree_interior_update_will_free_node(as, b);
1748
1749         n = bch2_btree_node_alloc_replacement(as, b);
1750         bch2_btree_update_add_new_node(as, n);
1751
1752         bch2_btree_build_aux_trees(n);
1753         six_unlock_write(&n->c.lock);
1754
1755         trace_btree_gc_rewrite_node(c, b);
1756
1757         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1758
1759         if (parent) {
1760                 bch2_keylist_add(&as->parent_keys, &n->key);
1761                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1762         } else {
1763                 bch2_btree_set_root(as, n, iter);
1764         }
1765
1766         bch2_btree_update_get_open_buckets(as, n);
1767
1768         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1769         bch2_btree_iter_node_drop(iter, b);
1770         bch2_btree_iter_node_replace(iter, n);
1771         bch2_btree_node_free_inmem(c, b, iter);
1772         six_unlock_intent(&n->c.lock);
1773
1774         bch2_btree_update_done(as);
1775 out:
1776         bch2_btree_iter_downgrade(iter);
1777         return ret;
1778 }
1779
1780 static void __bch2_btree_node_update_key(struct bch_fs *c,
1781                                          struct btree_update *as,
1782                                          struct btree_iter *iter,
1783                                          struct btree *b, struct btree *new_hash,
1784                                          struct bkey_i *new_key)
1785 {
1786         struct btree *parent;
1787         int ret;
1788
1789         btree_update_will_delete_key(as, &b->key);
1790         btree_update_will_add_key(as, new_key);
1791
1792         parent = btree_node_parent(iter, b);
1793         if (parent) {
1794                 if (new_hash) {
1795                         bkey_copy(&new_hash->key, new_key);
1796                         ret = bch2_btree_node_hash_insert(&c->btree_cache,
1797                                         new_hash, b->c.level, b->c.btree_id);
1798                         BUG_ON(ret);
1799                 }
1800
1801                 bch2_keylist_add(&as->parent_keys, new_key);
1802                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, 0);
1803
1804                 if (new_hash) {
1805                         mutex_lock(&c->btree_cache.lock);
1806                         bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
1807
1808                         bch2_btree_node_hash_remove(&c->btree_cache, b);
1809
1810                         bkey_copy(&b->key, new_key);
1811                         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1812                         BUG_ON(ret);
1813                         mutex_unlock(&c->btree_cache.lock);
1814                 } else {
1815                         bkey_copy(&b->key, new_key);
1816                 }
1817         } else {
1818                 BUG_ON(btree_node_root(c, b) != b);
1819
1820                 bch2_btree_node_lock_write(b, iter);
1821                 bkey_copy(&b->key, new_key);
1822
1823                 if (btree_ptr_hash_val(&b->key) != b->hash_val) {
1824                         mutex_lock(&c->btree_cache.lock);
1825                         bch2_btree_node_hash_remove(&c->btree_cache, b);
1826
1827                         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1828                         BUG_ON(ret);
1829                         mutex_unlock(&c->btree_cache.lock);
1830                 }
1831
1832                 btree_update_updated_root(as, b);
1833                 bch2_btree_node_unlock_write(b, iter);
1834         }
1835
1836         bch2_btree_update_done(as);
1837 }
1838
1839 int bch2_btree_node_update_key(struct bch_fs *c, struct btree_iter *iter,
1840                                struct btree *b,
1841                                struct bkey_i *new_key)
1842 {
1843         struct btree *parent = btree_node_parent(iter, b);
1844         struct btree_update *as = NULL;
1845         struct btree *new_hash = NULL;
1846         struct closure cl;
1847         int ret = 0;
1848
1849         closure_init_stack(&cl);
1850
1851         /*
1852          * check btree_ptr_hash_val() after @b is locked by
1853          * btree_iter_traverse():
1854          */
1855         if (btree_ptr_hash_val(new_key) != b->hash_val) {
1856                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1857                 if (ret) {
1858                         bch2_trans_unlock(iter->trans);
1859                         closure_sync(&cl);
1860                         if (!bch2_trans_relock(iter->trans))
1861                                 return -EINTR;
1862                 }
1863
1864                 new_hash = bch2_btree_node_mem_alloc(c);
1865         }
1866
1867         as = bch2_btree_update_start(iter, b->c.level,
1868                 parent ? btree_update_reserve_required(c, parent) : 0,
1869                 BTREE_INSERT_NOFAIL);
1870         if (IS_ERR(as)) {
1871                 ret = PTR_ERR(as);
1872                 goto err;
1873         }
1874
1875         __bch2_btree_node_update_key(c, as, iter, b, new_hash, new_key);
1876
1877         bch2_btree_iter_downgrade(iter);
1878 err:
1879         if (new_hash) {
1880                 mutex_lock(&c->btree_cache.lock);
1881                 list_move(&new_hash->list, &c->btree_cache.freeable);
1882                 mutex_unlock(&c->btree_cache.lock);
1883
1884                 six_unlock_write(&new_hash->c.lock);
1885                 six_unlock_intent(&new_hash->c.lock);
1886         }
1887         closure_sync(&cl);
1888         bch2_btree_cache_cannibalize_unlock(c);
1889         return ret;
1890 }
1891
1892 /* Init code: */
1893
1894 /*
1895  * Only for filesystem bringup, when first reading the btree roots or allocating
1896  * btree roots when initializing a new filesystem:
1897  */
1898 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
1899 {
1900         BUG_ON(btree_node_root(c, b));
1901
1902         bch2_btree_set_root_inmem(c, b);
1903 }
1904
1905 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
1906 {
1907         struct closure cl;
1908         struct btree *b;
1909         int ret;
1910
1911         closure_init_stack(&cl);
1912
1913         do {
1914                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1915                 closure_sync(&cl);
1916         } while (ret);
1917
1918         b = bch2_btree_node_mem_alloc(c);
1919         bch2_btree_cache_cannibalize_unlock(c);
1920
1921         set_btree_node_fake(b);
1922         set_btree_node_need_rewrite(b);
1923         b->c.level      = 0;
1924         b->c.btree_id   = id;
1925
1926         bkey_btree_ptr_init(&b->key);
1927         b->key.k.p = POS_MAX;
1928         *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
1929
1930         bch2_bset_init_first(b, &b->data->keys);
1931         bch2_btree_build_aux_trees(b);
1932
1933         b->data->flags = 0;
1934         btree_set_min(b, POS_MIN);
1935         btree_set_max(b, POS_MAX);
1936         b->data->format = bch2_btree_calc_format(b);
1937         btree_node_set_format(b, b->data->format);
1938
1939         ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
1940                                           b->c.level, b->c.btree_id);
1941         BUG_ON(ret);
1942
1943         bch2_btree_set_root_inmem(c, b);
1944
1945         six_unlock_write(&b->c.lock);
1946         six_unlock_intent(&b->c.lock);
1947 }
1948
1949 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
1950 {
1951         struct btree_update *as;
1952
1953         mutex_lock(&c->btree_interior_update_lock);
1954         list_for_each_entry(as, &c->btree_interior_update_list, list)
1955                 pr_buf(out, "%p m %u w %u r %u j %llu\n",
1956                        as,
1957                        as->mode,
1958                        as->nodes_written,
1959                        atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
1960                        as->journal.seq);
1961         mutex_unlock(&c->btree_interior_update_lock);
1962 }
1963
1964 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *c)
1965 {
1966         size_t ret = 0;
1967         struct list_head *i;
1968
1969         mutex_lock(&c->btree_interior_update_lock);
1970         list_for_each(i, &c->btree_interior_update_list)
1971                 ret++;
1972         mutex_unlock(&c->btree_interior_update_lock);
1973
1974         return ret;
1975 }
1976
1977 void bch2_journal_entries_to_btree_roots(struct bch_fs *c, struct jset *jset)
1978 {
1979         struct btree_root *r;
1980         struct jset_entry *entry;
1981
1982         mutex_lock(&c->btree_root_lock);
1983
1984         vstruct_for_each(jset, entry)
1985                 if (entry->type == BCH_JSET_ENTRY_btree_root) {
1986                         r = &c->btree_roots[entry->btree_id];
1987                         r->level = entry->level;
1988                         r->alive = true;
1989                         bkey_copy(&r->key, &entry->start[0]);
1990                 }
1991
1992         mutex_unlock(&c->btree_root_lock);
1993 }
1994
1995 struct jset_entry *
1996 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
1997                                     struct jset_entry *start,
1998                                     struct jset_entry *end)
1999 {
2000         struct jset_entry *entry;
2001         unsigned long have = 0;
2002         unsigned i;
2003
2004         for (entry = start; entry < end; entry = vstruct_next(entry))
2005                 if (entry->type == BCH_JSET_ENTRY_btree_root)
2006                         __set_bit(entry->btree_id, &have);
2007
2008         mutex_lock(&c->btree_root_lock);
2009
2010         for (i = 0; i < BTREE_ID_NR; i++)
2011                 if (c->btree_roots[i].alive && !test_bit(i, &have)) {
2012                         journal_entry_set(end,
2013                                           BCH_JSET_ENTRY_btree_root,
2014                                           i, c->btree_roots[i].level,
2015                                           &c->btree_roots[i].key,
2016                                           c->btree_roots[i].key.u64s);
2017                         end = vstruct_next(end);
2018                 }
2019
2020         mutex_unlock(&c->btree_root_lock);
2021
2022         return end;
2023 }
2024
2025 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2026 {
2027         if (c->btree_interior_update_worker)
2028                 destroy_workqueue(c->btree_interior_update_worker);
2029         mempool_exit(&c->btree_interior_update_pool);
2030 }
2031
2032 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2033 {
2034         mutex_init(&c->btree_reserve_cache_lock);
2035         INIT_LIST_HEAD(&c->btree_interior_update_list);
2036         INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2037         mutex_init(&c->btree_interior_update_lock);
2038         INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2039
2040         c->btree_interior_update_worker =
2041                 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2042         if (!c->btree_interior_update_worker)
2043                 return -ENOMEM;
2044
2045         return mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2046                                          sizeof(struct btree_update));
2047 }