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          * XXX: Waiting on io with btree node locks held, we don't want to be
 893          * doing this. We can't have btree writes happening after the space has
 894          * been freed, but we really only need to block before
 895          * btree_update_nodes_written_trans() happens.
 896          */
 897         btree_node_wait_on_io(b);
 898 }
 899
 900 void bch2_btree_update_done(struct btree_update *as)
 901 {
 902         BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
 903
 904         if (as->took_gc_lock)
 905                 up_read(&as->c->gc_lock);
 906         as->took_gc_lock = false;
 907
 908         bch2_btree_reserve_put(as);
 909
 910         continue_at(&as->cl, btree_update_set_nodes_written, system_freezable_wq);
 911 }
 912
 913 struct btree_update *
 914 bch2_btree_update_start(struct btree_iter *iter, unsigned level,
 915                         unsigned nr_nodes, unsigned flags)
 916 {
 917         struct btree_trans *trans = iter->trans;
 918         struct bch_fs *c = trans->c;
 919         struct btree_update *as;
 920         struct closure cl;
 921         int disk_res_flags = (flags & BTREE_INSERT_NOFAIL)
 922                 ? BCH_DISK_RESERVATION_NOFAIL : 0;
 923         int journal_flags = 0;
 924         int ret = 0;
 925
 926         if (flags & BTREE_INSERT_JOURNAL_RESERVED)
 927                 journal_flags |= JOURNAL_RES_GET_RESERVED;
 928
 929         closure_init_stack(&cl);
 930 retry:
 931         /*
 932          * This check isn't necessary for correctness - it's just to potentially
 933          * prevent us from doing a lot of work that'll end up being wasted:
 934          */
 935         ret = bch2_journal_error(&c->journal);
 936         if (ret)
 937                 return ERR_PTR(ret);
 938
 939         /*
 940          * XXX: figure out how far we might need to split,
 941          * instead of locking/reserving all the way to the root:
 942          */
 943         if (!bch2_btree_iter_upgrade(iter, U8_MAX)) {
 944                 trace_trans_restart_iter_upgrade(trans->ip);
 945                 return ERR_PTR(-EINTR);
 946         }
 947
 948         if (flags & BTREE_INSERT_GC_LOCK_HELD)
 949                 lockdep_assert_held(&c->gc_lock);
 950         else if (!down_read_trylock(&c->gc_lock)) {
 951                 if (flags & BTREE_INSERT_NOUNLOCK)
 952                         return ERR_PTR(-EINTR);
 953
 954                 bch2_trans_unlock(trans);
 955                 down_read(&c->gc_lock);
 956                 if (!bch2_trans_relock(trans)) {
 957                         up_read(&c->gc_lock);
 958                         return ERR_PTR(-EINTR);
 959                 }
 960         }
 961
 962         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
 963         memset(as, 0, sizeof(*as));
 964         closure_init(&as->cl, NULL);
 965         as->c           = c;
 966         as->mode        = BTREE_INTERIOR_NO_UPDATE;
 967         as->took_gc_lock = !(flags & BTREE_INSERT_GC_LOCK_HELD);
 968         as->btree_id    = iter->btree_id;
 969         INIT_LIST_HEAD(&as->list);
 970         INIT_LIST_HEAD(&as->unwritten_list);
 971         INIT_LIST_HEAD(&as->write_blocked_list);
 972         bch2_keylist_init(&as->old_keys, as->_old_keys);
 973         bch2_keylist_init(&as->new_keys, as->_new_keys);
 974         bch2_keylist_init(&as->parent_keys, as->inline_keys);
 975
 976         ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
 977                                       BTREE_UPDATE_JOURNAL_RES,
 978                                       journal_flags|JOURNAL_RES_GET_NONBLOCK);
 979         if (ret == -EAGAIN) {
 980                 /*
 981                  * this would be cleaner if bch2_journal_preres_get() took a
 982                  * closure argument
 983                  */
 984                 if (flags & BTREE_INSERT_NOUNLOCK) {
 985                         trace_trans_restart_journal_preres_get(trans->ip);
 986                         ret = -EINTR;
 987                         goto err;
 988                 }
 989
 990                 bch2_trans_unlock(trans);
 991
 992                 if (flags & BTREE_INSERT_JOURNAL_RECLAIM) {
 993                         bch2_btree_update_free(as);
 994                         return ERR_PTR(ret);
 995                 }
 996
 997                 ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
 998                                 BTREE_UPDATE_JOURNAL_RES,
 999                                 journal_flags);
1000                 if (ret) {
1001                         trace_trans_restart_journal_preres_get(trans->ip);
1002                         goto err;
1003                 }
1004
1005                 if (!bch2_trans_relock(trans)) {
1006                         ret = -EINTR;
1007                         goto err;
1008                 }
1009         }
1010
1011         ret = bch2_disk_reservation_get(c, &as->disk_res,
1012                         nr_nodes * c->opts.btree_node_size,
1013                         c->opts.metadata_replicas,
1014                         disk_res_flags);
1015         if (ret)
1016                 goto err;
1017
1018         ret = bch2_btree_reserve_get(as, nr_nodes, flags,
1019                 !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1020         if (ret)
1021                 goto err;
1022
1023         bch2_journal_pin_add(&c->journal,
1024                              atomic64_read(&c->journal.seq),
1025                              &as->journal, NULL);
1026
1027         mutex_lock(&c->btree_interior_update_lock);
1028         list_add_tail(&as->list, &c->btree_interior_update_list);
1029         mutex_unlock(&c->btree_interior_update_lock);
1030
1031         return as;
1032 err:
1033         bch2_btree_update_free(as);
1034
1035         if (ret == -EAGAIN) {
1036                 BUG_ON(flags & BTREE_INSERT_NOUNLOCK);
1037
1038                 bch2_trans_unlock(trans);
1039                 closure_sync(&cl);
1040                 ret = -EINTR;
1041         }
1042
1043         if (ret == -EINTR && bch2_trans_relock(trans))
1044                 goto retry;
1045
1046         return ERR_PTR(ret);
1047 }
1048
1049 /* Btree root updates: */
1050
1051 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1052 {
1053         /* Root nodes cannot be reaped */
1054         mutex_lock(&c->btree_cache.lock);
1055         list_del_init(&b->list);
1056         mutex_unlock(&c->btree_cache.lock);
1057
1058         if (b->c.level)
1059                 six_lock_pcpu_alloc(&b->c.lock);
1060         else
1061                 six_lock_pcpu_free(&b->c.lock);
1062
1063         mutex_lock(&c->btree_root_lock);
1064         BUG_ON(btree_node_root(c, b) &&
1065                (b->c.level < btree_node_root(c, b)->c.level ||
1066                 !btree_node_dying(btree_node_root(c, b))));
1067
1068         btree_node_root(c, b) = b;
1069         mutex_unlock(&c->btree_root_lock);
1070
1071         bch2_recalc_btree_reserve(c);
1072 }
1073
1074 /**
1075  * bch_btree_set_root - update the root in memory and on disk
1076  *
1077  * To ensure forward progress, the current task must not be holding any
1078  * btree node write locks. However, you must hold an intent lock on the
1079  * old root.
1080  *
1081  * Note: This allocates a journal entry but doesn't add any keys to
1082  * it.  All the btree roots are part of every journal write, so there
1083  * is nothing new to be done.  This just guarantees that there is a
1084  * journal write.
1085  */
1086 static void bch2_btree_set_root(struct btree_update *as, struct btree *b,
1087                                 struct btree_iter *iter)
1088 {
1089         struct bch_fs *c = as->c;
1090         struct btree *old;
1091
1092         trace_btree_set_root(c, b);
1093         BUG_ON(!b->written &&
1094                !test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags));
1095
1096         old = btree_node_root(c, b);
1097
1098         /*
1099          * Ensure no one is using the old root while we switch to the
1100          * new root:
1101          */
1102         bch2_btree_node_lock_write(old, iter);
1103
1104         bch2_btree_set_root_inmem(c, b);
1105
1106         btree_update_updated_root(as, b);
1107
1108         /*
1109          * Unlock old root after new root is visible:
1110          *
1111          * The new root isn't persistent, but that's ok: we still have
1112          * an intent lock on the new root, and any updates that would
1113          * depend on the new root would have to update the new root.
1114          */
1115         bch2_btree_node_unlock_write(old, iter);
1116 }
1117
1118 /* Interior node updates: */
1119
1120 static void bch2_insert_fixup_btree_ptr(struct btree_update *as, struct btree *b,
1121                                         struct btree_iter *iter,
1122                                         struct bkey_i *insert,
1123                                         struct btree_node_iter *node_iter)
1124 {
1125         struct bch_fs *c = as->c;
1126         struct bkey_packed *k;
1127         const char *invalid;
1128
1129         invalid = bch2_bkey_invalid(c, bkey_i_to_s_c(insert), btree_node_type(b)) ?:
1130                 bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert));
1131         if (invalid) {
1132                 char buf[160];
1133
1134                 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(insert));
1135                 bch2_fs_inconsistent(c, "inserting invalid bkey %s: %s", buf, invalid);
1136                 dump_stack();
1137         }
1138
1139         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1140                ARRAY_SIZE(as->journal_entries));
1141
1142         as->journal_u64s +=
1143                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1144                                   BCH_JSET_ENTRY_btree_keys,
1145                                   b->c.btree_id, b->c.level,
1146                                   insert, insert->k.u64s);
1147
1148         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1149                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1150                 bch2_btree_node_iter_advance(node_iter, b);
1151
1152         bch2_btree_bset_insert_key(iter, b, node_iter, insert);
1153         set_btree_node_dirty(c, b);
1154         set_btree_node_need_write(b);
1155 }
1156
1157 static void
1158 __bch2_btree_insert_keys_interior(struct btree_update *as, struct btree *b,
1159                                   struct btree_iter *iter, struct keylist *keys,
1160                                   struct btree_node_iter node_iter)
1161 {
1162         struct bkey_i *insert = bch2_keylist_front(keys);
1163         struct bkey_packed *k;
1164
1165         BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1166
1167         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1168                (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1169                 ;
1170
1171         for_each_keylist_key(keys, insert)
1172                 bch2_insert_fixup_btree_ptr(as, b, iter, insert, &node_iter);
1173 }
1174
1175 /*
1176  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1177  * node)
1178  */
1179 static struct btree *__btree_split_node(struct btree_update *as,
1180                                         struct btree *n1,
1181                                         struct btree_iter *iter)
1182 {
1183         struct bkey_format_state s;
1184         size_t nr_packed = 0, nr_unpacked = 0;
1185         struct btree *n2;
1186         struct bset *set1, *set2;
1187         struct bkey_packed *k, *set2_start, *set2_end, *out, *prev = NULL;
1188         struct bpos n1_pos;
1189
1190         n2 = bch2_btree_node_alloc(as, n1->c.level);
1191         bch2_btree_update_add_new_node(as, n2);
1192
1193         n2->data->max_key       = n1->data->max_key;
1194         n2->data->format        = n1->format;
1195         SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data));
1196         n2->key.k.p = n1->key.k.p;
1197
1198         set1 = btree_bset_first(n1);
1199         set2 = btree_bset_first(n2);
1200
1201         /*
1202          * Has to be a linear search because we don't have an auxiliary
1203          * search tree yet
1204          */
1205         k = set1->start;
1206         while (1) {
1207                 struct bkey_packed *n = bkey_next(k);
1208
1209                 if (n == vstruct_last(set1))
1210                         break;
1211                 if (k->_data - set1->_data >= (le16_to_cpu(set1->u64s) * 3) / 5)
1212                         break;
1213
1214                 if (bkey_packed(k))
1215                         nr_packed++;
1216                 else
1217                         nr_unpacked++;
1218
1219                 prev = k;
1220                 k = n;
1221         }
1222
1223         BUG_ON(!prev);
1224         set2_start      = k;
1225         set2_end        = vstruct_last(set1);
1226
1227         set1->u64s = cpu_to_le16((u64 *) set2_start - set1->_data);
1228         set_btree_bset_end(n1, n1->set);
1229
1230         n1->nr.live_u64s        = le16_to_cpu(set1->u64s);
1231         n1->nr.bset_u64s[0]     = le16_to_cpu(set1->u64s);
1232         n1->nr.packed_keys      = nr_packed;
1233         n1->nr.unpacked_keys    = nr_unpacked;
1234
1235         n1_pos = bkey_unpack_pos(n1, prev);
1236         if (as->c->sb.version < bcachefs_metadata_version_snapshot)
1237                 n1_pos.snapshot = U32_MAX;
1238
1239         btree_set_max(n1, n1_pos);
1240         btree_set_min(n2, bpos_successor(n1->key.k.p));
1241
1242         bch2_bkey_format_init(&s);
1243         bch2_bkey_format_add_pos(&s, n2->data->min_key);
1244         bch2_bkey_format_add_pos(&s, n2->data->max_key);
1245
1246         for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1247                 struct bkey uk = bkey_unpack_key(n1, k);
1248                 bch2_bkey_format_add_key(&s, &uk);
1249         }
1250
1251         n2->data->format = bch2_bkey_format_done(&s);
1252         btree_node_set_format(n2, n2->data->format);
1253
1254         out = set2->start;
1255         memset(&n2->nr, 0, sizeof(n2->nr));
1256
1257         for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1258                 BUG_ON(!bch2_bkey_transform(&n2->format, out, bkey_packed(k)
1259                                        ? &n1->format : &bch2_bkey_format_current, k));
1260                 out->format = KEY_FORMAT_LOCAL_BTREE;
1261                 btree_keys_account_key_add(&n2->nr, 0, out);
1262                 out = bkey_next(out);
1263         }
1264
1265         set2->u64s = cpu_to_le16((u64 *) out - set2->_data);
1266         set_btree_bset_end(n2, n2->set);
1267
1268         BUG_ON(!set1->u64s);
1269         BUG_ON(!set2->u64s);
1270
1271         btree_node_reset_sib_u64s(n1);
1272         btree_node_reset_sib_u64s(n2);
1273
1274         bch2_verify_btree_nr_keys(n1);
1275         bch2_verify_btree_nr_keys(n2);
1276
1277         if (n1->c.level) {
1278                 btree_node_interior_verify(as->c, n1);
1279                 btree_node_interior_verify(as->c, n2);
1280         }
1281
1282         return n2;
1283 }
1284
1285 /*
1286  * For updates to interior nodes, we've got to do the insert before we split
1287  * because the stuff we're inserting has to be inserted atomically. Post split,
1288  * the keys might have to go in different nodes and the split would no longer be
1289  * atomic.
1290  *
1291  * Worse, if the insert is from btree node coalescing, if we do the insert after
1292  * we do the split (and pick the pivot) - the pivot we pick might be between
1293  * nodes that were coalesced, and thus in the middle of a child node post
1294  * coalescing:
1295  */
1296 static void btree_split_insert_keys(struct btree_update *as, struct btree *b,
1297                                     struct btree_iter *iter,
1298                                     struct keylist *keys)
1299 {
1300         struct btree_node_iter node_iter;
1301         struct bkey_i *k = bch2_keylist_front(keys);
1302         struct bkey_packed *src, *dst, *n;
1303         struct bset *i;
1304
1305         bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
1306
1307         __bch2_btree_insert_keys_interior(as, b, iter, keys, node_iter);
1308
1309         /*
1310          * We can't tolerate whiteouts here - with whiteouts there can be
1311          * duplicate keys, and it would be rather bad if we picked a duplicate
1312          * for the pivot:
1313          */
1314         i = btree_bset_first(b);
1315         src = dst = i->start;
1316         while (src != vstruct_last(i)) {
1317                 n = bkey_next(src);
1318                 if (!bkey_deleted(src)) {
1319                         memmove_u64s_down(dst, src, src->u64s);
1320                         dst = bkey_next(dst);
1321                 }
1322                 src = n;
1323         }
1324
1325         /* Also clear out the unwritten whiteouts area: */
1326         b->whiteout_u64s = 0;
1327
1328         i->u64s = cpu_to_le16((u64 *) dst - i->_data);
1329         set_btree_bset_end(b, b->set);
1330
1331         BUG_ON(b->nsets != 1 ||
1332                b->nr.live_u64s != le16_to_cpu(btree_bset_first(b)->u64s));
1333
1334         btree_node_interior_verify(as->c, b);
1335 }
1336
1337 static void btree_split(struct btree_update *as, struct btree *b,
1338                         struct btree_iter *iter, struct keylist *keys,
1339                         unsigned flags)
1340 {
1341         struct bch_fs *c = as->c;
1342         struct btree *parent = btree_node_parent(iter, b);
1343         struct btree *n1, *n2 = NULL, *n3 = NULL;
1344         u64 start_time = local_clock();
1345
1346         BUG_ON(!parent && (b != btree_node_root(c, b)));
1347         BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->c.level));
1348
1349         bch2_btree_interior_update_will_free_node(as, b);
1350
1351         n1 = bch2_btree_node_alloc_replacement(as, b);
1352         bch2_btree_update_add_new_node(as, n1);
1353
1354         if (keys)
1355                 btree_split_insert_keys(as, n1, iter, keys);
1356
1357         if (bset_u64s(&n1->set[0]) > BTREE_SPLIT_THRESHOLD(c)) {
1358                 trace_btree_split(c, b);
1359
1360                 n2 = __btree_split_node(as, n1, iter);
1361
1362                 bch2_btree_build_aux_trees(n2);
1363                 bch2_btree_build_aux_trees(n1);
1364                 six_unlock_write(&n2->c.lock);
1365                 six_unlock_write(&n1->c.lock);
1366
1367                 bch2_btree_node_write(c, n2, SIX_LOCK_intent);
1368
1369                 /*
1370                  * Note that on recursive parent_keys == keys, so we
1371                  * can't start adding new keys to parent_keys before emptying it
1372                  * out (which we did with btree_split_insert_keys() above)
1373                  */
1374                 bch2_keylist_add(&as->parent_keys, &n1->key);
1375                 bch2_keylist_add(&as->parent_keys, &n2->key);
1376
1377                 if (!parent) {
1378                         /* Depth increases, make a new root */
1379                         n3 = __btree_root_alloc(as, b->c.level + 1);
1380
1381                         n3->sib_u64s[0] = U16_MAX;
1382                         n3->sib_u64s[1] = U16_MAX;
1383
1384                         btree_split_insert_keys(as, n3, iter, &as->parent_keys);
1385
1386                         bch2_btree_node_write(c, n3, SIX_LOCK_intent);
1387                 }
1388         } else {
1389                 trace_btree_compact(c, b);
1390
1391                 bch2_btree_build_aux_trees(n1);
1392                 six_unlock_write(&n1->c.lock);
1393
1394                 if (parent)
1395                         bch2_keylist_add(&as->parent_keys, &n1->key);
1396         }
1397
1398         bch2_btree_node_write(c, n1, SIX_LOCK_intent);
1399
1400         /* New nodes all written, now make them visible: */
1401
1402         if (parent) {
1403                 /* Split a non root node */
1404                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1405         } else if (n3) {
1406                 bch2_btree_set_root(as, n3, iter);
1407         } else {
1408                 /* Root filled up but didn't need to be split */
1409                 bch2_btree_set_root(as, n1, iter);
1410         }
1411
1412         bch2_btree_update_get_open_buckets(as, n1);
1413         if (n2)
1414                 bch2_btree_update_get_open_buckets(as, n2);
1415         if (n3)
1416                 bch2_btree_update_get_open_buckets(as, n3);
1417
1418         /* Successful split, update the iterator to point to the new nodes: */
1419
1420         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1421         bch2_btree_iter_node_drop(iter, b);
1422         if (n3)
1423                 bch2_btree_iter_node_replace(iter, n3);
1424         if (n2)
1425                 bch2_btree_iter_node_replace(iter, n2);
1426         bch2_btree_iter_node_replace(iter, n1);
1427
1428         /*
1429          * The old node must be freed (in memory) _before_ unlocking the new
1430          * nodes - else another thread could re-acquire a read lock on the old
1431          * node after another thread has locked and updated the new node, thus
1432          * seeing stale data:
1433          */
1434         bch2_btree_node_free_inmem(c, b, iter);
1435
1436         if (n3)
1437                 six_unlock_intent(&n3->c.lock);
1438         if (n2)
1439                 six_unlock_intent(&n2->c.lock);
1440         six_unlock_intent(&n1->c.lock);
1441
1442         bch2_btree_trans_verify_locks(iter->trans);
1443
1444         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_split],
1445                                start_time);
1446 }
1447
1448 static void
1449 bch2_btree_insert_keys_interior(struct btree_update *as, struct btree *b,
1450                                 struct btree_iter *iter, struct keylist *keys)
1451 {
1452         struct btree_iter *linked;
1453
1454         __bch2_btree_insert_keys_interior(as, b, iter, keys, iter->l[b->c.level].iter);
1455
1456         btree_update_updated_node(as, b);
1457
1458         trans_for_each_iter_with_node(iter->trans, b, linked)
1459                 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1460
1461         bch2_btree_trans_verify_iters(iter->trans, b);
1462 }
1463
1464 /**
1465  * bch_btree_insert_node - insert bkeys into a given btree node
1466  *
1467  * @iter:               btree iterator
1468  * @keys:               list of keys to insert
1469  * @hook:               insert callback
1470  * @persistent:         if not null, @persistent will wait on journal write
1471  *
1472  * Inserts as many keys as it can into a given btree node, splitting it if full.
1473  * If a split occurred, this function will return early. This can only happen
1474  * for leaf nodes -- inserts into interior nodes have to be atomic.
1475  */
1476 void bch2_btree_insert_node(struct btree_update *as, struct btree *b,
1477                             struct btree_iter *iter, struct keylist *keys,
1478                             unsigned flags)
1479 {
1480         struct bch_fs *c = as->c;
1481         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1482         int old_live_u64s = b->nr.live_u64s;
1483         int live_u64s_added, u64s_added;
1484
1485         lockdep_assert_held(&c->gc_lock);
1486         BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->c.level));
1487         BUG_ON(!b->c.level);
1488         BUG_ON(!as || as->b);
1489         bch2_verify_keylist_sorted(keys);
1490
1491         bch2_btree_node_lock_for_insert(c, b, iter);
1492
1493         if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1494                 bch2_btree_node_unlock_write(b, iter);
1495                 goto split;
1496         }
1497
1498         btree_node_interior_verify(c, b);
1499
1500         bch2_btree_insert_keys_interior(as, b, iter, keys);
1501
1502         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1503         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1504
1505         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1506                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1507         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1508                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1509
1510         if (u64s_added > live_u64s_added &&
1511             bch2_maybe_compact_whiteouts(c, b))
1512                 bch2_btree_iter_reinit_node(iter, b);
1513
1514         bch2_btree_node_unlock_write(b, iter);
1515
1516         btree_node_interior_verify(c, b);
1517         return;
1518 split:
1519         btree_split(as, b, iter, keys, flags);
1520 }
1521
1522 int bch2_btree_split_leaf(struct bch_fs *c, struct btree_iter *iter,
1523                           unsigned flags)
1524 {
1525         struct btree *b = iter_l(iter)->b;
1526         struct btree_update *as;
1527         unsigned l;
1528         int ret = 0;
1529
1530         as = bch2_btree_update_start(iter, iter->level,
1531                 btree_update_reserve_required(c, b), flags);
1532         if (IS_ERR(as))
1533                 return PTR_ERR(as);
1534
1535         btree_split(as, b, iter, NULL, flags);
1536         bch2_btree_update_done(as);
1537
1538         for (l = iter->level + 1; btree_iter_node(iter, l) && !ret; l++)
1539                 ret = bch2_foreground_maybe_merge(c, iter, l, flags);
1540
1541         return ret;
1542 }
1543
1544 int __bch2_foreground_maybe_merge(struct bch_fs *c,
1545                                   struct btree_iter *iter,
1546                                   unsigned level,
1547                                   unsigned flags,
1548                                   enum btree_node_sibling sib)
1549 {
1550         struct btree_trans *trans = iter->trans;
1551         struct btree_iter *sib_iter = NULL;
1552         struct btree_update *as;
1553         struct bkey_format_state new_s;
1554         struct bkey_format new_f;
1555         struct bkey_i delete;
1556         struct btree *b, *m, *n, *prev, *next, *parent;
1557         struct bpos sib_pos;
1558         size_t sib_u64s;
1559         int ret = 0, ret2 = 0;
1560
1561         BUG_ON(!btree_node_locked(iter, level));
1562 retry:
1563         ret = bch2_btree_iter_traverse(iter);
1564         if (ret)
1565                 goto err;
1566
1567         BUG_ON(!btree_node_locked(iter, level));
1568
1569         b = iter->l[level].b;
1570
1571         if ((sib == btree_prev_sib && !bpos_cmp(b->data->min_key, POS_MIN)) ||
1572             (sib == btree_next_sib && !bpos_cmp(b->data->max_key, POS_MAX))) {
1573                 b->sib_u64s[sib] = U16_MAX;
1574                 goto out;
1575         }
1576
1577         sib_pos = sib == btree_prev_sib
1578                 ? bpos_predecessor(b->data->min_key)
1579                 : bpos_successor(b->data->max_key);
1580
1581         sib_iter = bch2_trans_get_node_iter(trans, iter->btree_id,
1582                                             sib_pos, U8_MAX, level,
1583                                             BTREE_ITER_INTENT);
1584         ret = bch2_btree_iter_traverse(sib_iter);
1585         if (ret)
1586                 goto err;
1587
1588         m = sib_iter->l[level].b;
1589
1590         if (btree_node_parent(iter, b) !=
1591             btree_node_parent(sib_iter, m)) {
1592                 b->sib_u64s[sib] = U16_MAX;
1593                 goto out;
1594         }
1595
1596         if (sib == btree_prev_sib) {
1597                 prev = m;
1598                 next = b;
1599         } else {
1600                 prev = b;
1601                 next = m;
1602         }
1603
1604         if (bkey_cmp(bpos_successor(prev->data->max_key), next->data->min_key)) {
1605                 char buf1[100], buf2[100];
1606
1607                 bch2_bpos_to_text(&PBUF(buf1), prev->data->max_key);
1608                 bch2_bpos_to_text(&PBUF(buf2), next->data->min_key);
1609                 bch2_fs_inconsistent(c,
1610                                      "btree topology error in btree merge:\n"
1611                                      "prev ends at   %s\n"
1612                                      "next starts at %s\n",
1613                                      buf1, buf2);
1614                 ret = -EIO;
1615                 goto err;
1616         }
1617
1618         bch2_bkey_format_init(&new_s);
1619         bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
1620         __bch2_btree_calc_format(&new_s, prev);
1621         __bch2_btree_calc_format(&new_s, next);
1622         bch2_bkey_format_add_pos(&new_s, next->data->max_key);
1623         new_f = bch2_bkey_format_done(&new_s);
1624
1625         sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1626                 btree_node_u64s_with_format(m, &new_f);
1627
1628         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1629                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1630                 sib_u64s /= 2;
1631                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1632         }
1633
1634         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1635         sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
1636         b->sib_u64s[sib] = sib_u64s;
1637
1638         if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
1639                 goto out;
1640
1641         parent = btree_node_parent(iter, b);
1642         as = bch2_btree_update_start(iter, level,
1643                          btree_update_reserve_required(c, parent) + 1,
1644                          flags|
1645                          BTREE_INSERT_NOFAIL|
1646                          BTREE_INSERT_USE_RESERVE);
1647         ret = PTR_ERR_OR_ZERO(as);
1648         if (ret)
1649                 goto err;
1650
1651         trace_btree_merge(c, b);
1652
1653         bch2_btree_interior_update_will_free_node(as, b);
1654         bch2_btree_interior_update_will_free_node(as, m);
1655
1656         n = bch2_btree_node_alloc(as, b->c.level);
1657         bch2_btree_update_add_new_node(as, n);
1658
1659         btree_set_min(n, prev->data->min_key);
1660         btree_set_max(n, next->data->max_key);
1661         n->data->format         = new_f;
1662
1663         btree_node_set_format(n, new_f);
1664
1665         bch2_btree_sort_into(c, n, prev);
1666         bch2_btree_sort_into(c, n, next);
1667
1668         bch2_btree_build_aux_trees(n);
1669         six_unlock_write(&n->c.lock);
1670
1671         bkey_init(&delete.k);
1672         delete.k.p = prev->key.k.p;
1673         bch2_keylist_add(&as->parent_keys, &delete);
1674         bch2_keylist_add(&as->parent_keys, &n->key);
1675
1676         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1677
1678         bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1679
1680         bch2_btree_update_get_open_buckets(as, n);
1681
1682         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1683         six_lock_increment(&m->c.lock, SIX_LOCK_intent);
1684         bch2_btree_iter_node_drop(iter, b);
1685         bch2_btree_iter_node_drop(iter, m);
1686
1687         bch2_btree_iter_node_replace(iter, n);
1688
1689         bch2_btree_trans_verify_iters(trans, n);
1690
1691         bch2_btree_node_free_inmem(c, b, iter);
1692         bch2_btree_node_free_inmem(c, m, iter);
1693
1694         six_unlock_intent(&n->c.lock);
1695
1696         bch2_btree_update_done(as);
1697 out:
1698         bch2_btree_trans_verify_locks(trans);
1699         bch2_trans_iter_free(trans, sib_iter);
1700
1701         /*
1702          * Don't downgrade locks here: we're called after successful insert,
1703          * and the caller will downgrade locks after a successful insert
1704          * anyways (in case e.g. a split was required first)
1705          *
1706          * And we're also called when inserting into interior nodes in the
1707          * split path, and downgrading to read locks in there is potentially
1708          * confusing:
1709          */
1710         return ret ?: ret2;
1711 err:
1712         bch2_trans_iter_put(trans, sib_iter);
1713         sib_iter = NULL;
1714
1715         if (ret == -EINTR && bch2_trans_relock(trans))
1716                 goto retry;
1717
1718         if (ret == -EINTR && !(flags & BTREE_INSERT_NOUNLOCK)) {
1719                 ret2 = ret;
1720                 ret = bch2_btree_iter_traverse_all(trans);
1721                 if (!ret)
1722                         goto retry;
1723         }
1724
1725         goto out;
1726 }
1727
1728 /**
1729  * bch_btree_node_rewrite - Rewrite/move a btree node
1730  */
1731 int bch2_btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
1732                             __le64 seq, unsigned flags)
1733 {
1734         struct btree *b, *n, *parent;
1735         struct btree_update *as;
1736         int ret;
1737
1738         flags |= BTREE_INSERT_NOFAIL;
1739 retry:
1740         ret = bch2_btree_iter_traverse(iter);
1741         if (ret)
1742                 goto out;
1743
1744         b = bch2_btree_iter_peek_node(iter);
1745         if (!b || b->data->keys.seq != seq)
1746                 goto out;
1747
1748         parent = btree_node_parent(iter, b);
1749         as = bch2_btree_update_start(iter, b->c.level,
1750                 (parent
1751                  ? btree_update_reserve_required(c, parent)
1752                  : 0) + 1,
1753                 flags);
1754         ret = PTR_ERR_OR_ZERO(as);
1755         if (ret == -EINTR)
1756                 goto retry;
1757         if (ret) {
1758                 trace_btree_gc_rewrite_node_fail(c, b);
1759                 goto out;
1760         }
1761
1762         bch2_btree_interior_update_will_free_node(as, b);
1763
1764         n = bch2_btree_node_alloc_replacement(as, b);
1765         bch2_btree_update_add_new_node(as, n);
1766
1767         bch2_btree_build_aux_trees(n);
1768         six_unlock_write(&n->c.lock);
1769
1770         trace_btree_gc_rewrite_node(c, b);
1771
1772         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1773
1774         if (parent) {
1775                 bch2_keylist_add(&as->parent_keys, &n->key);
1776                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1777         } else {
1778                 bch2_btree_set_root(as, n, iter);
1779         }
1780
1781         bch2_btree_update_get_open_buckets(as, n);
1782
1783         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1784         bch2_btree_iter_node_drop(iter, b);
1785         bch2_btree_iter_node_replace(iter, n);
1786         bch2_btree_node_free_inmem(c, b, iter);
1787         six_unlock_intent(&n->c.lock);
1788
1789         bch2_btree_update_done(as);
1790 out:
1791         bch2_btree_iter_downgrade(iter);
1792         return ret;
1793 }
1794
1795 static void __bch2_btree_node_update_key(struct bch_fs *c,
1796                                          struct btree_update *as,
1797                                          struct btree_iter *iter,
1798                                          struct btree *b, struct btree *new_hash,
1799                                          struct bkey_i *new_key)
1800 {
1801         struct btree *parent;
1802         int ret;
1803
1804         btree_update_will_delete_key(as, &b->key);
1805         btree_update_will_add_key(as, new_key);
1806
1807         parent = btree_node_parent(iter, b);
1808         if (parent) {
1809                 if (new_hash) {
1810                         bkey_copy(&new_hash->key, new_key);
1811                         ret = bch2_btree_node_hash_insert(&c->btree_cache,
1812                                         new_hash, b->c.level, b->c.btree_id);
1813                         BUG_ON(ret);
1814                 }
1815
1816                 bch2_keylist_add(&as->parent_keys, new_key);
1817                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, 0);
1818
1819                 if (new_hash) {
1820                         mutex_lock(&c->btree_cache.lock);
1821                         bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
1822
1823                         bch2_btree_node_hash_remove(&c->btree_cache, b);
1824
1825                         bkey_copy(&b->key, new_key);
1826                         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1827                         BUG_ON(ret);
1828                         mutex_unlock(&c->btree_cache.lock);
1829                 } else {
1830                         bkey_copy(&b->key, new_key);
1831                 }
1832         } else {
1833                 BUG_ON(btree_node_root(c, b) != b);
1834
1835                 bch2_btree_node_lock_write(b, iter);
1836                 bkey_copy(&b->key, new_key);
1837
1838                 if (btree_ptr_hash_val(&b->key) != b->hash_val) {
1839                         mutex_lock(&c->btree_cache.lock);
1840                         bch2_btree_node_hash_remove(&c->btree_cache, b);
1841
1842                         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1843                         BUG_ON(ret);
1844                         mutex_unlock(&c->btree_cache.lock);
1845                 }
1846
1847                 btree_update_updated_root(as, b);
1848                 bch2_btree_node_unlock_write(b, iter);
1849         }
1850
1851         bch2_btree_update_done(as);
1852 }
1853
1854 int bch2_btree_node_update_key(struct bch_fs *c, struct btree_iter *iter,
1855                                struct btree *b,
1856                                struct bkey_i *new_key)
1857 {
1858         struct btree *parent = btree_node_parent(iter, b);
1859         struct btree_update *as = NULL;
1860         struct btree *new_hash = NULL;
1861         struct closure cl;
1862         int ret = 0;
1863
1864         closure_init_stack(&cl);
1865
1866         /*
1867          * check btree_ptr_hash_val() after @b is locked by
1868          * btree_iter_traverse():
1869          */
1870         if (btree_ptr_hash_val(new_key) != b->hash_val) {
1871                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1872                 if (ret) {
1873                         bch2_trans_unlock(iter->trans);
1874                         closure_sync(&cl);
1875                         if (!bch2_trans_relock(iter->trans))
1876                                 return -EINTR;
1877                 }
1878
1879                 new_hash = bch2_btree_node_mem_alloc(c);
1880         }
1881
1882         as = bch2_btree_update_start(iter, b->c.level,
1883                 parent ? btree_update_reserve_required(c, parent) : 0,
1884                 BTREE_INSERT_NOFAIL);
1885         if (IS_ERR(as)) {
1886                 ret = PTR_ERR(as);
1887                 goto err;
1888         }
1889
1890         __bch2_btree_node_update_key(c, as, iter, b, new_hash, new_key);
1891
1892         bch2_btree_iter_downgrade(iter);
1893 err:
1894         if (new_hash) {
1895                 mutex_lock(&c->btree_cache.lock);
1896                 list_move(&new_hash->list, &c->btree_cache.freeable);
1897                 mutex_unlock(&c->btree_cache.lock);
1898
1899                 six_unlock_write(&new_hash->c.lock);
1900                 six_unlock_intent(&new_hash->c.lock);
1901         }
1902         closure_sync(&cl);
1903         bch2_btree_cache_cannibalize_unlock(c);
1904         return ret;
1905 }
1906
1907 /* Init code: */
1908
1909 /*
1910  * Only for filesystem bringup, when first reading the btree roots or allocating
1911  * btree roots when initializing a new filesystem:
1912  */
1913 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
1914 {
1915         BUG_ON(btree_node_root(c, b));
1916
1917         bch2_btree_set_root_inmem(c, b);
1918 }
1919
1920 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
1921 {
1922         struct closure cl;
1923         struct btree *b;
1924         int ret;
1925
1926         closure_init_stack(&cl);
1927
1928         do {
1929                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1930                 closure_sync(&cl);
1931         } while (ret);
1932
1933         b = bch2_btree_node_mem_alloc(c);
1934         bch2_btree_cache_cannibalize_unlock(c);
1935
1936         set_btree_node_fake(b);
1937         set_btree_node_need_rewrite(b);
1938         b->c.level      = 0;
1939         b->c.btree_id   = id;
1940
1941         bkey_btree_ptr_init(&b->key);
1942         b->key.k.p = POS_MAX;
1943         *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
1944
1945         bch2_bset_init_first(b, &b->data->keys);
1946         bch2_btree_build_aux_trees(b);
1947
1948         b->data->flags = 0;
1949         btree_set_min(b, POS_MIN);
1950         btree_set_max(b, POS_MAX);
1951         b->data->format = bch2_btree_calc_format(b);
1952         btree_node_set_format(b, b->data->format);
1953
1954         ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
1955                                           b->c.level, b->c.btree_id);
1956         BUG_ON(ret);
1957
1958         bch2_btree_set_root_inmem(c, b);
1959
1960         six_unlock_write(&b->c.lock);
1961         six_unlock_intent(&b->c.lock);
1962 }
1963
1964 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
1965 {
1966         struct btree_update *as;
1967
1968         mutex_lock(&c->btree_interior_update_lock);
1969         list_for_each_entry(as, &c->btree_interior_update_list, list)
1970                 pr_buf(out, "%p m %u w %u r %u j %llu\n",
1971                        as,
1972                        as->mode,
1973                        as->nodes_written,
1974                        atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
1975                        as->journal.seq);
1976         mutex_unlock(&c->btree_interior_update_lock);
1977 }
1978
1979 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *c)
1980 {
1981         size_t ret = 0;
1982         struct list_head *i;
1983
1984         mutex_lock(&c->btree_interior_update_lock);
1985         list_for_each(i, &c->btree_interior_update_list)
1986                 ret++;
1987         mutex_unlock(&c->btree_interior_update_lock);
1988
1989         return ret;
1990 }
1991
1992 void bch2_journal_entries_to_btree_roots(struct bch_fs *c, struct jset *jset)
1993 {
1994         struct btree_root *r;
1995         struct jset_entry *entry;
1996
1997         mutex_lock(&c->btree_root_lock);
1998
1999         vstruct_for_each(jset, entry)
2000                 if (entry->type == BCH_JSET_ENTRY_btree_root) {
2001                         r = &c->btree_roots[entry->btree_id];
2002                         r->level = entry->level;
2003                         r->alive = true;
2004                         bkey_copy(&r->key, &entry->start[0]);
2005                 }
2006
2007         mutex_unlock(&c->btree_root_lock);
2008 }
2009
2010 struct jset_entry *
2011 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2012                                     struct jset_entry *start,
2013                                     struct jset_entry *end)
2014 {
2015         struct jset_entry *entry;
2016         unsigned long have = 0;
2017         unsigned i;
2018
2019         for (entry = start; entry < end; entry = vstruct_next(entry))
2020                 if (entry->type == BCH_JSET_ENTRY_btree_root)
2021                         __set_bit(entry->btree_id, &have);
2022
2023         mutex_lock(&c->btree_root_lock);
2024
2025         for (i = 0; i < BTREE_ID_NR; i++)
2026                 if (c->btree_roots[i].alive && !test_bit(i, &have)) {
2027                         journal_entry_set(end,
2028                                           BCH_JSET_ENTRY_btree_root,
2029                                           i, c->btree_roots[i].level,
2030                                           &c->btree_roots[i].key,
2031                                           c->btree_roots[i].key.u64s);
2032                         end = vstruct_next(end);
2033                 }
2034
2035         mutex_unlock(&c->btree_root_lock);
2036
2037         return end;
2038 }
2039
2040 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2041 {
2042         if (c->btree_interior_update_worker)
2043                 destroy_workqueue(c->btree_interior_update_worker);
2044         mempool_exit(&c->btree_interior_update_pool);
2045 }
2046
2047 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2048 {
2049         mutex_init(&c->btree_reserve_cache_lock);
2050         INIT_LIST_HEAD(&c->btree_interior_update_list);
2051         INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2052         mutex_init(&c->btree_interior_update_lock);
2053         INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2054
2055         c->btree_interior_update_worker =
2056                 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2057         if (!c->btree_interior_update_worker)
2058                 return -ENOMEM;
2059
2060         return mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2061                                          sizeof(struct btree_update));
2062 }