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 "recovery.h"
  20 #include "replicas.h"
  21 #include "super-io.h"
  22
  23 #include <linux/random.h>
  24 #include <trace/events/bcachefs.h>
  25
  26 static void bch2_btree_insert_node(struct btree_update *, struct btree_trans *,
  27                                    struct btree_path *, struct btree *,
  28                                    struct keylist *, unsigned);
  29 static void bch2_btree_update_add_new_node(struct btree_update *, struct btree *);
  30
  31 /* Debug code: */
  32
  33 /*
  34  * Verify that child nodes correctly span parent node's range:
  35  */
  36 static void btree_node_interior_verify(struct bch_fs *c, struct btree *b)
  37 {
  38 #ifdef CONFIG_BCACHEFS_DEBUG
  39         struct bpos next_node = b->data->min_key;
  40         struct btree_node_iter iter;
  41         struct bkey_s_c k;
  42         struct bkey_s_c_btree_ptr_v2 bp;
  43         struct bkey unpacked;
  44         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
  45
  46         BUG_ON(!b->c.level);
  47
  48         if (!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))
  49                 return;
  50
  51         bch2_btree_node_iter_init_from_start(&iter, b);
  52
  53         while (1) {
  54                 k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked);
  55                 if (k.k->type != KEY_TYPE_btree_ptr_v2)
  56                         break;
  57                 bp = bkey_s_c_to_btree_ptr_v2(k);
  58
  59                 if (bpos_cmp(next_node, bp.v->min_key)) {
  60                         bch2_dump_btree_node(c, b);
  61                         bch2_bpos_to_text(&buf1, next_node);
  62                         bch2_bpos_to_text(&buf2, bp.v->min_key);
  63                         panic("expected next min_key %s got %s\n", buf1.buf, buf2.buf);
  64                 }
  65
  66                 bch2_btree_node_iter_advance(&iter, b);
  67
  68                 if (bch2_btree_node_iter_end(&iter)) {
  69                         if (bpos_cmp(k.k->p, b->key.k.p)) {
  70                                 bch2_dump_btree_node(c, b);
  71                                 bch2_bpos_to_text(&buf1, b->key.k.p);
  72                                 bch2_bpos_to_text(&buf2, k.k->p);
  73                                 panic("expected end %s got %s\n", buf1.buf, buf2.buf);
  74                         }
  75                         break;
  76                 }
  77
  78                 next_node = bpos_successor(k.k->p);
  79         }
  80 #endif
  81 }
  82
  83 /* Calculate ideal packed bkey format for new btree nodes: */
  84
  85 void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b)
  86 {
  87         struct bkey_packed *k;
  88         struct bset_tree *t;
  89         struct bkey uk;
  90
  91         for_each_bset(b, t)
  92                 bset_tree_for_each_key(b, t, k)
  93                         if (!bkey_deleted(k)) {
  94                                 uk = bkey_unpack_key(b, k);
  95                                 bch2_bkey_format_add_key(s, &uk);
  96                         }
  97 }
  98
  99 static struct bkey_format bch2_btree_calc_format(struct btree *b)
 100 {
 101         struct bkey_format_state s;
 102
 103         bch2_bkey_format_init(&s);
 104         bch2_bkey_format_add_pos(&s, b->data->min_key);
 105         bch2_bkey_format_add_pos(&s, b->data->max_key);
 106         __bch2_btree_calc_format(&s, b);
 107
 108         return bch2_bkey_format_done(&s);
 109 }
 110
 111 static size_t btree_node_u64s_with_format(struct btree *b,
 112                                           struct bkey_format *new_f)
 113 {
 114         struct bkey_format *old_f = &b->format;
 115
 116         /* stupid integer promotion rules */
 117         ssize_t delta =
 118             (((int) new_f->key_u64s - old_f->key_u64s) *
 119              (int) b->nr.packed_keys) +
 120             (((int) new_f->key_u64s - BKEY_U64s) *
 121              (int) b->nr.unpacked_keys);
 122
 123         BUG_ON(delta + b->nr.live_u64s < 0);
 124
 125         return b->nr.live_u64s + delta;
 126 }
 127
 128 /**
 129  * btree_node_format_fits - check if we could rewrite node with a new format
 130  *
 131  * This assumes all keys can pack with the new format -- it just checks if
 132  * the re-packed keys would fit inside the node itself.
 133  */
 134 bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
 135                                  struct bkey_format *new_f)
 136 {
 137         size_t u64s = btree_node_u64s_with_format(b, new_f);
 138
 139         return __vstruct_bytes(struct btree_node, u64s) < btree_bytes(c);
 140 }
 141
 142 /* Btree node freeing/allocation: */
 143
 144 static void __btree_node_free(struct bch_fs *c, struct btree *b)
 145 {
 146         trace_btree_node_free(c, b);
 147
 148         BUG_ON(btree_node_dirty(b));
 149         BUG_ON(btree_node_need_write(b));
 150         BUG_ON(b == btree_node_root(c, b));
 151         BUG_ON(b->ob.nr);
 152         BUG_ON(!list_empty(&b->write_blocked));
 153         BUG_ON(b->will_make_reachable);
 154
 155         clear_btree_node_noevict(b);
 156
 157         mutex_lock(&c->btree_cache.lock);
 158         list_move(&b->list, &c->btree_cache.freeable);
 159         mutex_unlock(&c->btree_cache.lock);
 160 }
 161
 162 static void bch2_btree_node_free_inmem(struct btree_trans *trans,
 163                                        struct btree *b)
 164 {
 165         struct bch_fs *c = trans->c;
 166         struct btree_path *path;
 167
 168         trans_for_each_path(trans, path)
 169                 BUG_ON(path->l[b->c.level].b == b &&
 170                        path->l[b->c.level].lock_seq == b->c.lock.state.seq);
 171
 172         six_lock_write(&b->c.lock, NULL, NULL);
 173
 174         bch2_btree_node_hash_remove(&c->btree_cache, b);
 175         __btree_node_free(c, b);
 176
 177         six_unlock_write(&b->c.lock);
 178         six_unlock_intent(&b->c.lock);
 179 }
 180
 181 static struct btree *__bch2_btree_node_alloc(struct bch_fs *c,
 182                                              struct disk_reservation *res,
 183                                              struct closure *cl,
 184                                              bool interior_node,
 185                                              unsigned flags)
 186 {
 187         struct write_point *wp;
 188         struct btree *b;
 189         __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
 190         struct open_buckets ob = { .nr = 0 };
 191         struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
 192         unsigned nr_reserve;
 193         enum alloc_reserve alloc_reserve;
 194
 195         if (flags & BTREE_INSERT_USE_RESERVE) {
 196                 nr_reserve      = 0;
 197                 alloc_reserve   = RESERVE_btree_movinggc;
 198         } else {
 199                 nr_reserve      = BTREE_NODE_RESERVE;
 200                 alloc_reserve   = RESERVE_btree;
 201         }
 202
 203         mutex_lock(&c->btree_reserve_cache_lock);
 204         if (c->btree_reserve_cache_nr > nr_reserve) {
 205                 struct btree_alloc *a =
 206                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
 207
 208                 ob = a->ob;
 209                 bkey_copy(&tmp.k, &a->k);
 210                 mutex_unlock(&c->btree_reserve_cache_lock);
 211                 goto mem_alloc;
 212         }
 213         mutex_unlock(&c->btree_reserve_cache_lock);
 214
 215 retry:
 216         wp = bch2_alloc_sectors_start(c,
 217                                       c->opts.metadata_target ?:
 218                                       c->opts.foreground_target,
 219                                       0,
 220                                       writepoint_ptr(&c->btree_write_point),
 221                                       &devs_have,
 222                                       res->nr_replicas,
 223                                       c->opts.metadata_replicas_required,
 224                                       alloc_reserve, 0, cl);
 225         if (IS_ERR(wp))
 226                 return ERR_CAST(wp);
 227
 228         if (wp->sectors_free < btree_sectors(c)) {
 229                 struct open_bucket *ob;
 230                 unsigned i;
 231
 232                 open_bucket_for_each(c, &wp->ptrs, ob, i)
 233                         if (ob->sectors_free < btree_sectors(c))
 234                                 ob->sectors_free = 0;
 235
 236                 bch2_alloc_sectors_done(c, wp);
 237                 goto retry;
 238         }
 239
 240         bkey_btree_ptr_v2_init(&tmp.k);
 241         bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, btree_sectors(c), false);
 242
 243         bch2_open_bucket_get(c, wp, &ob);
 244         bch2_alloc_sectors_done(c, wp);
 245 mem_alloc:
 246         b = bch2_btree_node_mem_alloc(c, interior_node);
 247         six_unlock_write(&b->c.lock);
 248         six_unlock_intent(&b->c.lock);
 249
 250         /* we hold cannibalize_lock: */
 251         BUG_ON(IS_ERR(b));
 252         BUG_ON(b->ob.nr);
 253
 254         bkey_copy(&b->key, &tmp.k);
 255         b->ob = ob;
 256
 257         return b;
 258 }
 259
 260 static struct btree *bch2_btree_node_alloc(struct btree_update *as, unsigned level)
 261 {
 262         struct bch_fs *c = as->c;
 263         struct btree *b;
 264         struct prealloc_nodes *p = &as->prealloc_nodes[!!level];
 265         int ret;
 266
 267         BUG_ON(level >= BTREE_MAX_DEPTH);
 268         BUG_ON(!p->nr);
 269
 270         b = p->b[--p->nr];
 271
 272         six_lock_intent(&b->c.lock, NULL, NULL);
 273         six_lock_write(&b->c.lock, NULL, NULL);
 274
 275         set_btree_node_accessed(b);
 276         set_btree_node_dirty_acct(c, b);
 277         set_btree_node_need_write(b);
 278
 279         bch2_bset_init_first(b, &b->data->keys);
 280         b->c.level      = level;
 281         b->c.btree_id   = as->btree_id;
 282         b->version_ondisk = c->sb.version;
 283
 284         memset(&b->nr, 0, sizeof(b->nr));
 285         b->data->magic = cpu_to_le64(bset_magic(c));
 286         memset(&b->data->_ptr, 0, sizeof(b->data->_ptr));
 287         b->data->flags = 0;
 288         SET_BTREE_NODE_ID(b->data, as->btree_id);
 289         SET_BTREE_NODE_LEVEL(b->data, level);
 290
 291         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 292                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
 293
 294                 bp->v.mem_ptr           = 0;
 295                 bp->v.seq               = b->data->keys.seq;
 296                 bp->v.sectors_written   = 0;
 297         }
 298
 299         SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
 300
 301         bch2_btree_build_aux_trees(b);
 302
 303         ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
 304         BUG_ON(ret);
 305
 306         trace_btree_node_alloc(c, b);
 307         return b;
 308 }
 309
 310 static void btree_set_min(struct btree *b, struct bpos pos)
 311 {
 312         if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
 313                 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
 314         b->data->min_key = pos;
 315 }
 316
 317 static void btree_set_max(struct btree *b, struct bpos pos)
 318 {
 319         b->key.k.p = pos;
 320         b->data->max_key = pos;
 321 }
 322
 323 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *as,
 324                                                   struct btree *b,
 325                                                   struct bkey_format format)
 326 {
 327         struct btree *n;
 328
 329         n = bch2_btree_node_alloc(as, b->c.level);
 330
 331         SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
 332
 333         btree_set_min(n, b->data->min_key);
 334         btree_set_max(n, b->data->max_key);
 335
 336         n->data->format         = format;
 337         btree_node_set_format(n, format);
 338
 339         bch2_btree_sort_into(as->c, n, b);
 340
 341         btree_node_reset_sib_u64s(n);
 342
 343         n->key.k.p = b->key.k.p;
 344         return n;
 345 }
 346
 347 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
 348                                                        struct btree *b)
 349 {
 350         struct bkey_format new_f = bch2_btree_calc_format(b);
 351
 352         /*
 353          * The keys might expand with the new format - if they wouldn't fit in
 354          * the btree node anymore, use the old format for now:
 355          */
 356         if (!bch2_btree_node_format_fits(as->c, b, &new_f))
 357                 new_f = b->format;
 358
 359         return __bch2_btree_node_alloc_replacement(as, b, new_f);
 360 }
 361
 362 static struct btree *__btree_root_alloc(struct btree_update *as, unsigned level)
 363 {
 364         struct btree *b = bch2_btree_node_alloc(as, level);
 365
 366         btree_set_min(b, POS_MIN);
 367         btree_set_max(b, SPOS_MAX);
 368         b->data->format = bch2_btree_calc_format(b);
 369
 370         btree_node_set_format(b, b->data->format);
 371         bch2_btree_build_aux_trees(b);
 372
 373         bch2_btree_update_add_new_node(as, b);
 374         six_unlock_write(&b->c.lock);
 375
 376         return b;
 377 }
 378
 379 static void bch2_btree_reserve_put(struct btree_update *as)
 380 {
 381         struct bch_fs *c = as->c;
 382         struct prealloc_nodes *p;
 383
 384         for (p = as->prealloc_nodes;
 385              p < as->prealloc_nodes + ARRAY_SIZE(as->prealloc_nodes);
 386              p++) {
 387                 while (p->nr) {
 388                         struct btree *b = p->b[--p->nr];
 389
 390                         mutex_lock(&c->btree_reserve_cache_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                         mutex_unlock(&c->btree_reserve_cache_lock);
 405
 406                         six_lock_intent(&b->c.lock, NULL, NULL);
 407                         six_lock_write(&b->c.lock, NULL, NULL);
 408                         __btree_node_free(c, b);
 409                         six_unlock_write(&b->c.lock);
 410                         six_unlock_intent(&b->c.lock);
 411                 }
 412         }
 413 }
 414
 415 static int bch2_btree_reserve_get(struct btree_update *as,
 416                                   unsigned nr_nodes[2],
 417                                   unsigned flags)
 418 {
 419         struct bch_fs *c = as->c;
 420         struct closure cl;
 421         struct btree *b;
 422         unsigned interior;
 423         int ret;
 424
 425         closure_init_stack(&cl);
 426 retry:
 427
 428         BUG_ON(nr_nodes[0] + nr_nodes[1] > BTREE_RESERVE_MAX);
 429
 430         /*
 431          * Protects reaping from the btree node cache and using the btree node
 432          * open bucket reserve:
 433          *
 434          * BTREE_INSERT_NOWAIT only applies to btree node allocation, not
 435          * blocking on this lock:
 436          */
 437         ret = bch2_btree_cache_cannibalize_lock(c, &cl);
 438         if (ret)
 439                 goto err;
 440
 441         for (interior = 0; interior < 2; interior++) {
 442                 struct prealloc_nodes *p = as->prealloc_nodes + interior;
 443
 444                 while (p->nr < nr_nodes[interior]) {
 445                         b = __bch2_btree_node_alloc(c, &as->disk_res,
 446                                                     flags & BTREE_INSERT_NOWAIT
 447                                                     ? NULL : &cl,
 448                                                     interior, flags);
 449                         if (IS_ERR(b)) {
 450                                 ret = PTR_ERR(b);
 451                                 goto err;
 452                         }
 453
 454                         p->b[p->nr++] = b;
 455                 }
 456         }
 457
 458         bch2_btree_cache_cannibalize_unlock(c);
 459         closure_sync(&cl);
 460         return 0;
 461 err:
 462         bch2_btree_cache_cannibalize_unlock(c);
 463         closure_sync(&cl);
 464
 465         if (ret == -EAGAIN)
 466                 goto retry;
 467
 468         trace_btree_reserve_get_fail(c, nr_nodes[0] + nr_nodes[1], &cl);
 469         return ret;
 470 }
 471
 472 /* Asynchronous interior node update machinery */
 473
 474 static void bch2_btree_update_free(struct btree_update *as)
 475 {
 476         struct bch_fs *c = as->c;
 477
 478         if (as->took_gc_lock)
 479                 up_read(&c->gc_lock);
 480         as->took_gc_lock = false;
 481
 482         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 483
 484         bch2_journal_pin_drop(&c->journal, &as->journal);
 485         bch2_journal_pin_flush(&c->journal, &as->journal);
 486         bch2_disk_reservation_put(c, &as->disk_res);
 487         bch2_btree_reserve_put(as);
 488
 489         bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_total],
 490                                as->start_time);
 491
 492         mutex_lock(&c->btree_interior_update_lock);
 493         list_del(&as->unwritten_list);
 494         list_del(&as->list);
 495
 496         closure_debug_destroy(&as->cl);
 497         mempool_free(as, &c->btree_interior_update_pool);
 498
 499         /*
 500          * Have to do the wakeup with btree_interior_update_lock still held,
 501          * since being on btree_interior_update_list is our ref on @c:
 502          */
 503         closure_wake_up(&c->btree_interior_update_wait);
 504
 505         mutex_unlock(&c->btree_interior_update_lock);
 506 }
 507
 508 static void btree_update_add_key(struct btree_update *as,
 509                                  struct keylist *keys, struct btree *b)
 510 {
 511         struct bkey_i *k = &b->key;
 512
 513         BUG_ON(bch2_keylist_u64s(keys) + k->k.u64s >
 514                ARRAY_SIZE(as->_old_keys));
 515
 516         bkey_copy(keys->top, k);
 517         bkey_i_to_btree_ptr_v2(keys->top)->v.mem_ptr = b->c.level + 1;
 518
 519         bch2_keylist_push(keys);
 520 }
 521
 522 /*
 523  * The transactional part of an interior btree node update, where we journal the
 524  * update we did to the interior node and update alloc info:
 525  */
 526 static int btree_update_nodes_written_trans(struct btree_trans *trans,
 527                                             struct btree_update *as)
 528 {
 529         struct bkey_i *k;
 530         int ret;
 531
 532         ret = darray_make_room(&trans->extra_journal_entries, as->journal_u64s);
 533         if (ret)
 534                 return ret;
 535
 536         memcpy(&darray_top(trans->extra_journal_entries),
 537                as->journal_entries,
 538                as->journal_u64s * sizeof(u64));
 539         trans->extra_journal_entries.nr += as->journal_u64s;
 540
 541         trans->journal_pin = &as->journal;
 542
 543         for_each_keylist_key(&as->old_keys, k) {
 544                 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
 545
 546                 ret = bch2_trans_mark_old(trans, as->btree_id, level, bkey_i_to_s_c(k), 0);
 547                 if (ret)
 548                         return ret;
 549         }
 550
 551         for_each_keylist_key(&as->new_keys, k) {
 552                 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
 553
 554                 ret = bch2_trans_mark_new(trans, as->btree_id, level, k, 0);
 555                 if (ret)
 556                         return ret;
 557         }
 558
 559         return 0;
 560 }
 561
 562 static void btree_update_nodes_written(struct btree_update *as)
 563 {
 564         struct bch_fs *c = as->c;
 565         struct btree *b = as->b;
 566         struct btree_trans trans;
 567         u64 journal_seq = 0;
 568         unsigned i;
 569         int ret;
 570
 571         /*
 572          * If we're already in an error state, it might be because a btree node
 573          * was never written, and we might be trying to free that same btree
 574          * node here, but it won't have been marked as allocated and we'll see
 575          * spurious disk usage inconsistencies in the transactional part below
 576          * if we don't skip it:
 577          */
 578         ret = bch2_journal_error(&c->journal);
 579         if (ret)
 580                 goto err;
 581
 582         /*
 583          * Wait for any in flight writes to finish before we free the old nodes
 584          * on disk:
 585          */
 586         for (i = 0; i < as->nr_old_nodes; i++) {
 587                 struct btree *old = as->old_nodes[i];
 588                 __le64 seq;
 589
 590                 six_lock_read(&old->c.lock, NULL, NULL);
 591                 seq = old->data ? old->data->keys.seq : 0;
 592                 six_unlock_read(&old->c.lock);
 593
 594                 if (seq == as->old_nodes_seq[i])
 595                         wait_on_bit_io(&old->flags, BTREE_NODE_write_in_flight_inner,
 596                                        TASK_UNINTERRUPTIBLE);
 597         }
 598
 599         /*
 600          * We did an update to a parent node where the pointers we added pointed
 601          * to child nodes that weren't written yet: now, the child nodes have
 602          * been written so we can write out the update to the interior node.
 603          */
 604
 605         /*
 606          * We can't call into journal reclaim here: we'd block on the journal
 607          * reclaim lock, but we may need to release the open buckets we have
 608          * pinned in order for other btree updates to make forward progress, and
 609          * journal reclaim does btree updates when flushing bkey_cached entries,
 610          * which may require allocations as well.
 611          */
 612         bch2_trans_init(&trans, c, 0, 512);
 613         ret = __bch2_trans_do(&trans, &as->disk_res, &journal_seq,
 614                               BTREE_INSERT_NOFAIL|
 615                               BTREE_INSERT_NOCHECK_RW|
 616                               BTREE_INSERT_JOURNAL_RECLAIM|
 617                               JOURNAL_WATERMARK_reserved,
 618                               btree_update_nodes_written_trans(&trans, as));
 619         bch2_trans_exit(&trans);
 620
 621         bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
 622                              "error %i in btree_update_nodes_written()", ret);
 623 err:
 624         if (b) {
 625                 /*
 626                  * @b is the node we did the final insert into:
 627                  *
 628                  * On failure to get a journal reservation, we still have to
 629                  * unblock the write and allow most of the write path to happen
 630                  * so that shutdown works, but the i->journal_seq mechanism
 631                  * won't work to prevent the btree write from being visible (we
 632                  * didn't get a journal sequence number) - instead
 633                  * __bch2_btree_node_write() doesn't do the actual write if
 634                  * we're in journal error state:
 635                  */
 636
 637                 six_lock_intent(&b->c.lock, NULL, NULL);
 638                 six_lock_write(&b->c.lock, NULL, NULL);
 639                 mutex_lock(&c->btree_interior_update_lock);
 640
 641                 list_del(&as->write_blocked_list);
 642                 if (list_empty(&b->write_blocked))
 643                         clear_btree_node_write_blocked(b);
 644
 645                 /*
 646                  * Node might have been freed, recheck under
 647                  * btree_interior_update_lock:
 648                  */
 649                 if (as->b == b) {
 650                         struct bset *i = btree_bset_last(b);
 651
 652                         BUG_ON(!b->c.level);
 653                         BUG_ON(!btree_node_dirty(b));
 654
 655                         if (!ret) {
 656                                 i->journal_seq = cpu_to_le64(
 657                                                              max(journal_seq,
 658                                                                  le64_to_cpu(i->journal_seq)));
 659
 660                                 bch2_btree_add_journal_pin(c, b, journal_seq);
 661                         } else {
 662                                 /*
 663                                  * If we didn't get a journal sequence number we
 664                                  * can't write this btree node, because recovery
 665                                  * won't know to ignore this write:
 666                                  */
 667                                 set_btree_node_never_write(b);
 668                         }
 669                 }
 670
 671                 mutex_unlock(&c->btree_interior_update_lock);
 672                 six_unlock_write(&b->c.lock);
 673
 674                 btree_node_write_if_need(c, b, SIX_LOCK_intent);
 675                 six_unlock_intent(&b->c.lock);
 676         }
 677
 678         bch2_journal_pin_drop(&c->journal, &as->journal);
 679
 680         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 681
 682         mutex_lock(&c->btree_interior_update_lock);
 683         for (i = 0; i < as->nr_new_nodes; i++) {
 684                 b = as->new_nodes[i];
 685
 686                 BUG_ON(b->will_make_reachable != (unsigned long) as);
 687                 b->will_make_reachable = 0;
 688                 clear_btree_node_will_make_reachable(b);
 689         }
 690         mutex_unlock(&c->btree_interior_update_lock);
 691
 692         for (i = 0; i < as->nr_new_nodes; i++) {
 693                 b = as->new_nodes[i];
 694
 695                 six_lock_read(&b->c.lock, NULL, NULL);
 696                 btree_node_write_if_need(c, b, SIX_LOCK_read);
 697                 six_unlock_read(&b->c.lock);
 698         }
 699
 700         for (i = 0; i < as->nr_open_buckets; i++)
 701                 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
 702
 703         bch2_btree_update_free(as);
 704 }
 705
 706 static void btree_interior_update_work(struct work_struct *work)
 707 {
 708         struct bch_fs *c =
 709                 container_of(work, struct bch_fs, btree_interior_update_work);
 710         struct btree_update *as;
 711
 712         while (1) {
 713                 mutex_lock(&c->btree_interior_update_lock);
 714                 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
 715                                               struct btree_update, unwritten_list);
 716                 if (as && !as->nodes_written)
 717                         as = NULL;
 718                 mutex_unlock(&c->btree_interior_update_lock);
 719
 720                 if (!as)
 721                         break;
 722
 723                 btree_update_nodes_written(as);
 724         }
 725 }
 726
 727 static void btree_update_set_nodes_written(struct closure *cl)
 728 {
 729         struct btree_update *as = container_of(cl, struct btree_update, cl);
 730         struct bch_fs *c = as->c;
 731
 732         mutex_lock(&c->btree_interior_update_lock);
 733         as->nodes_written = true;
 734         mutex_unlock(&c->btree_interior_update_lock);
 735
 736         queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
 737 }
 738
 739 /*
 740  * We're updating @b with pointers to nodes that haven't finished writing yet:
 741  * block @b from being written until @as completes
 742  */
 743 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
 744 {
 745         struct bch_fs *c = as->c;
 746
 747         mutex_lock(&c->btree_interior_update_lock);
 748         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 749
 750         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 751         BUG_ON(!btree_node_dirty(b));
 752
 753         as->mode        = BTREE_INTERIOR_UPDATING_NODE;
 754         as->b           = b;
 755
 756         set_btree_node_write_blocked(b);
 757         list_add(&as->write_blocked_list, &b->write_blocked);
 758
 759         mutex_unlock(&c->btree_interior_update_lock);
 760 }
 761
 762 static void btree_update_reparent(struct btree_update *as,
 763                                   struct btree_update *child)
 764 {
 765         struct bch_fs *c = as->c;
 766
 767         lockdep_assert_held(&c->btree_interior_update_lock);
 768
 769         child->b = NULL;
 770         child->mode = BTREE_INTERIOR_UPDATING_AS;
 771
 772         bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL);
 773 }
 774
 775 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
 776 {
 777         struct bkey_i *insert = &b->key;
 778         struct bch_fs *c = as->c;
 779
 780         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 781
 782         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
 783                ARRAY_SIZE(as->journal_entries));
 784
 785         as->journal_u64s +=
 786                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
 787                                   BCH_JSET_ENTRY_btree_root,
 788                                   b->c.btree_id, b->c.level,
 789                                   insert, insert->k.u64s);
 790
 791         mutex_lock(&c->btree_interior_update_lock);
 792         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 793
 794         as->mode        = BTREE_INTERIOR_UPDATING_ROOT;
 795         mutex_unlock(&c->btree_interior_update_lock);
 796 }
 797
 798 /*
 799  * bch2_btree_update_add_new_node:
 800  *
 801  * This causes @as to wait on @b to be written, before it gets to
 802  * bch2_btree_update_nodes_written
 803  *
 804  * Additionally, it sets b->will_make_reachable to prevent any additional writes
 805  * to @b from happening besides the first until @b is reachable on disk
 806  *
 807  * And it adds @b to the list of @as's new nodes, so that we can update sector
 808  * counts in bch2_btree_update_nodes_written:
 809  */
 810 static void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
 811 {
 812         struct bch_fs *c = as->c;
 813
 814         closure_get(&as->cl);
 815
 816         mutex_lock(&c->btree_interior_update_lock);
 817         BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
 818         BUG_ON(b->will_make_reachable);
 819
 820         as->new_nodes[as->nr_new_nodes++] = b;
 821         b->will_make_reachable = 1UL|(unsigned long) as;
 822         set_btree_node_will_make_reachable(b);
 823
 824         mutex_unlock(&c->btree_interior_update_lock);
 825
 826         btree_update_add_key(as, &as->new_keys, b);
 827 }
 828
 829 /*
 830  * returns true if @b was a new node
 831  */
 832 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
 833 {
 834         struct btree_update *as;
 835         unsigned long v;
 836         unsigned i;
 837
 838         mutex_lock(&c->btree_interior_update_lock);
 839         /*
 840          * When b->will_make_reachable != 0, it owns a ref on as->cl that's
 841          * dropped when it gets written by bch2_btree_complete_write - the
 842          * xchg() is for synchronization with bch2_btree_complete_write:
 843          */
 844         v = xchg(&b->will_make_reachable, 0);
 845         clear_btree_node_will_make_reachable(b);
 846         as = (struct btree_update *) (v & ~1UL);
 847
 848         if (!as) {
 849                 mutex_unlock(&c->btree_interior_update_lock);
 850                 return;
 851         }
 852
 853         for (i = 0; i < as->nr_new_nodes; i++)
 854                 if (as->new_nodes[i] == b)
 855                         goto found;
 856
 857         BUG();
 858 found:
 859         array_remove_item(as->new_nodes, as->nr_new_nodes, i);
 860         mutex_unlock(&c->btree_interior_update_lock);
 861
 862         if (v & 1)
 863                 closure_put(&as->cl);
 864 }
 865
 866 static void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
 867 {
 868         while (b->ob.nr)
 869                 as->open_buckets[as->nr_open_buckets++] =
 870                         b->ob.v[--b->ob.nr];
 871 }
 872
 873 /*
 874  * @b is being split/rewritten: it may have pointers to not-yet-written btree
 875  * nodes and thus outstanding btree_updates - redirect @b's
 876  * btree_updates to point to this btree_update:
 877  */
 878 static void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 879                                                       struct btree *b)
 880 {
 881         struct bch_fs *c = as->c;
 882         struct btree_update *p, *n;
 883         struct btree_write *w;
 884
 885         set_btree_node_dying(b);
 886
 887         if (btree_node_fake(b))
 888                 return;
 889
 890         mutex_lock(&c->btree_interior_update_lock);
 891
 892         /*
 893          * Does this node have any btree_update operations preventing
 894          * it from being written?
 895          *
 896          * If so, redirect them to point to this btree_update: we can
 897          * write out our new nodes, but we won't make them visible until those
 898          * operations complete
 899          */
 900         list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
 901                 list_del_init(&p->write_blocked_list);
 902                 btree_update_reparent(as, p);
 903
 904                 /*
 905                  * for flush_held_btree_writes() waiting on updates to flush or
 906                  * nodes to be writeable:
 907                  */
 908                 closure_wake_up(&c->btree_interior_update_wait);
 909         }
 910
 911         clear_btree_node_dirty_acct(c, b);
 912         clear_btree_node_need_write(b);
 913
 914         /*
 915          * Does this node have unwritten data that has a pin on the journal?
 916          *
 917          * If so, transfer that pin to the btree_update operation -
 918          * note that if we're freeing multiple nodes, we only need to keep the
 919          * oldest pin of any of the nodes we're freeing. We'll release the pin
 920          * when the new nodes are persistent and reachable on disk:
 921          */
 922         w = btree_current_write(b);
 923         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 924         bch2_journal_pin_drop(&c->journal, &w->journal);
 925
 926         w = btree_prev_write(b);
 927         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 928         bch2_journal_pin_drop(&c->journal, &w->journal);
 929
 930         mutex_unlock(&c->btree_interior_update_lock);
 931
 932         /*
 933          * Is this a node that isn't reachable on disk yet?
 934          *
 935          * Nodes that aren't reachable yet have writes blocked until they're
 936          * reachable - now that we've cancelled any pending writes and moved
 937          * things waiting on that write to wait on this update, we can drop this
 938          * node from the list of nodes that the other update is making
 939          * reachable, prior to freeing it:
 940          */
 941         btree_update_drop_new_node(c, b);
 942
 943         btree_update_add_key(as, &as->old_keys, b);
 944
 945         as->old_nodes[as->nr_old_nodes] = b;
 946         as->old_nodes_seq[as->nr_old_nodes] = b->data->keys.seq;
 947         as->nr_old_nodes++;
 948 }
 949
 950 static void bch2_btree_update_done(struct btree_update *as)
 951 {
 952         struct bch_fs *c = as->c;
 953         u64 start_time = as->start_time;
 954
 955         BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
 956
 957         if (as->took_gc_lock)
 958                 up_read(&as->c->gc_lock);
 959         as->took_gc_lock = false;
 960
 961         bch2_btree_reserve_put(as);
 962
 963         continue_at(&as->cl, btree_update_set_nodes_written,
 964                     as->c->btree_interior_update_worker);
 965
 966         bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_foreground],
 967                                start_time);
 968 }
 969
 970 static struct btree_update *
 971 bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
 972                         unsigned level, bool split, unsigned flags)
 973 {
 974         struct bch_fs *c = trans->c;
 975         struct btree_update *as;
 976         u64 start_time = local_clock();
 977         int disk_res_flags = (flags & BTREE_INSERT_NOFAIL)
 978                 ? BCH_DISK_RESERVATION_NOFAIL : 0;
 979         unsigned nr_nodes[2] = { 0, 0 };
 980         unsigned update_level = level;
 981         int journal_flags = flags & JOURNAL_WATERMARK_MASK;
 982         int ret = 0;
 983
 984         BUG_ON(!path->should_be_locked);
 985
 986         if (flags & BTREE_INSERT_JOURNAL_RECLAIM)
 987                 journal_flags |= JOURNAL_RES_GET_NONBLOCK;
 988
 989         while (1) {
 990                 nr_nodes[!!update_level] += 1 + split;
 991                 update_level++;
 992
 993                 if (!btree_path_node(path, update_level))
 994                         break;
 995
 996                 /*
 997                  * XXX: figure out how far we might need to split,
 998                  * instead of locking/reserving all the way to the root:
 999                  */
1000                 split = update_level + 1 < BTREE_MAX_DEPTH;
1001         }
1002
1003         /* Might have to allocate a new root: */
1004         if (update_level < BTREE_MAX_DEPTH)
1005                 nr_nodes[1] += 1;
1006
1007         if (!bch2_btree_path_upgrade(trans, path, U8_MAX)) {
1008                 trace_trans_restart_iter_upgrade(trans->fn, _RET_IP_,
1009                                                  path->btree_id, &path->pos);
1010                 ret = btree_trans_restart(trans);
1011                 return ERR_PTR(ret);
1012         }
1013
1014         if (flags & BTREE_INSERT_GC_LOCK_HELD)
1015                 lockdep_assert_held(&c->gc_lock);
1016         else if (!down_read_trylock(&c->gc_lock)) {
1017                 bch2_trans_unlock(trans);
1018                 down_read(&c->gc_lock);
1019                 if (!bch2_trans_relock(trans)) {
1020                         up_read(&c->gc_lock);
1021                         return ERR_PTR(-EINTR);
1022                 }
1023         }
1024
1025         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
1026         memset(as, 0, sizeof(*as));
1027         closure_init(&as->cl, NULL);
1028         as->c           = c;
1029         as->start_time  = start_time;
1030         as->mode        = BTREE_INTERIOR_NO_UPDATE;
1031         as->took_gc_lock = !(flags & BTREE_INSERT_GC_LOCK_HELD);
1032         as->btree_id    = path->btree_id;
1033         INIT_LIST_HEAD(&as->list);
1034         INIT_LIST_HEAD(&as->unwritten_list);
1035         INIT_LIST_HEAD(&as->write_blocked_list);
1036         bch2_keylist_init(&as->old_keys, as->_old_keys);
1037         bch2_keylist_init(&as->new_keys, as->_new_keys);
1038         bch2_keylist_init(&as->parent_keys, as->inline_keys);
1039
1040         mutex_lock(&c->btree_interior_update_lock);
1041         list_add_tail(&as->list, &c->btree_interior_update_list);
1042         mutex_unlock(&c->btree_interior_update_lock);
1043
1044         /*
1045          * We don't want to allocate if we're in an error state, that can cause
1046          * deadlock on emergency shutdown due to open buckets getting stuck in
1047          * the btree_reserve_cache after allocator shutdown has cleared it out.
1048          * This check needs to come after adding us to the btree_interior_update
1049          * list but before calling bch2_btree_reserve_get, to synchronize with
1050          * __bch2_fs_read_only().
1051          */
1052         ret = bch2_journal_error(&c->journal);
1053         if (ret)
1054                 goto err;
1055
1056         bch2_trans_unlock(trans);
1057
1058         ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
1059                                       BTREE_UPDATE_JOURNAL_RES,
1060                                       journal_flags);
1061         if (ret) {
1062                 bch2_btree_update_free(as);
1063                 trace_trans_restart_journal_preres_get(trans->fn, _RET_IP_);
1064                 btree_trans_restart(trans);
1065                 return ERR_PTR(ret);
1066         }
1067
1068         ret = bch2_disk_reservation_get(c, &as->disk_res,
1069                         (nr_nodes[0] + nr_nodes[1]) * btree_sectors(c),
1070                         c->opts.metadata_replicas,
1071                         disk_res_flags);
1072         if (ret)
1073                 goto err;
1074
1075         ret = bch2_btree_reserve_get(as, nr_nodes, flags);
1076         if (ret)
1077                 goto err;
1078
1079         if (!bch2_trans_relock(trans)) {
1080                 ret = -EINTR;
1081                 goto err;
1082         }
1083
1084         return as;
1085 err:
1086         bch2_btree_update_free(as);
1087         return ERR_PTR(ret);
1088 }
1089
1090 /* Btree root updates: */
1091
1092 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1093 {
1094         /* Root nodes cannot be reaped */
1095         mutex_lock(&c->btree_cache.lock);
1096         list_del_init(&b->list);
1097         mutex_unlock(&c->btree_cache.lock);
1098
1099         mutex_lock(&c->btree_root_lock);
1100         BUG_ON(btree_node_root(c, b) &&
1101                (b->c.level < btree_node_root(c, b)->c.level ||
1102                 !btree_node_dying(btree_node_root(c, b))));
1103
1104         btree_node_root(c, b) = b;
1105         mutex_unlock(&c->btree_root_lock);
1106
1107         bch2_recalc_btree_reserve(c);
1108 }
1109
1110 /**
1111  * bch_btree_set_root - update the root in memory and on disk
1112  *
1113  * To ensure forward progress, the current task must not be holding any
1114  * btree node write locks. However, you must hold an intent lock on the
1115  * old root.
1116  *
1117  * Note: This allocates a journal entry but doesn't add any keys to
1118  * it.  All the btree roots are part of every journal write, so there
1119  * is nothing new to be done.  This just guarantees that there is a
1120  * journal write.
1121  */
1122 static void bch2_btree_set_root(struct btree_update *as,
1123                                 struct btree_trans *trans,
1124                                 struct btree_path *path,
1125                                 struct btree *b)
1126 {
1127         struct bch_fs *c = as->c;
1128         struct btree *old;
1129
1130         trace_btree_set_root(c, b);
1131         BUG_ON(!b->written);
1132
1133         old = btree_node_root(c, b);
1134
1135         /*
1136          * Ensure no one is using the old root while we switch to the
1137          * new root:
1138          */
1139         bch2_btree_node_lock_write(trans, path, old);
1140
1141         bch2_btree_set_root_inmem(c, b);
1142
1143         btree_update_updated_root(as, b);
1144
1145         /*
1146          * Unlock old root after new root is visible:
1147          *
1148          * The new root isn't persistent, but that's ok: we still have
1149          * an intent lock on the new root, and any updates that would
1150          * depend on the new root would have to update the new root.
1151          */
1152         bch2_btree_node_unlock_write(trans, path, old);
1153 }
1154
1155 /* Interior node updates: */
1156
1157 static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
1158                                         struct btree_trans *trans,
1159                                         struct btree_path *path,
1160                                         struct btree *b,
1161                                         struct btree_node_iter *node_iter,
1162                                         struct bkey_i *insert)
1163 {
1164         struct bch_fs *c = as->c;
1165         struct bkey_packed *k;
1166         struct printbuf buf = PRINTBUF;
1167
1168         BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 &&
1169                !btree_ptr_sectors_written(insert));
1170
1171         if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
1172                 bch2_journal_key_overwritten(c, b->c.btree_id, b->c.level, insert->k.p);
1173
1174         if (bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1175                               btree_node_type(b), WRITE, &buf) ?:
1176             bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert), &buf)) {
1177                 printbuf_reset(&buf);
1178                 pr_buf(&buf, "inserting invalid bkey\n  ");
1179                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
1180                 pr_buf(&buf, "\n  ");
1181                 bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1182                                   btree_node_type(b), WRITE, &buf);
1183                 bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert), &buf);
1184
1185                 bch2_fs_inconsistent(c, "%s", buf.buf);
1186                 dump_stack();
1187         }
1188
1189         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1190                ARRAY_SIZE(as->journal_entries));
1191
1192         as->journal_u64s +=
1193                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1194                                   BCH_JSET_ENTRY_btree_keys,
1195                                   b->c.btree_id, b->c.level,
1196                                   insert, insert->k.u64s);
1197
1198         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1199                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1200                 bch2_btree_node_iter_advance(node_iter, b);
1201
1202         bch2_btree_bset_insert_key(trans, path, b, node_iter, insert);
1203         set_btree_node_dirty_acct(c, b);
1204         set_btree_node_need_write(b);
1205
1206         printbuf_exit(&buf);
1207 }
1208
1209 static void
1210 __bch2_btree_insert_keys_interior(struct btree_update *as,
1211                                   struct btree_trans *trans,
1212                                   struct btree_path *path,
1213                                   struct btree *b,
1214                                   struct btree_node_iter node_iter,
1215                                   struct keylist *keys)
1216 {
1217         struct bkey_i *insert = bch2_keylist_front(keys);
1218         struct bkey_packed *k;
1219
1220         BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1221
1222         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1223                (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1224                 ;
1225
1226         while (!bch2_keylist_empty(keys)) {
1227                 bch2_insert_fixup_btree_ptr(as, trans, path, b,
1228                                 &node_iter, bch2_keylist_front(keys));
1229                 bch2_keylist_pop_front(keys);
1230         }
1231 }
1232
1233 /*
1234  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1235  * node)
1236  */
1237 static struct btree *__btree_split_node(struct btree_update *as,
1238                                         struct btree *n1)
1239 {
1240         struct bkey_format_state s;
1241         size_t nr_packed = 0, nr_unpacked = 0;
1242         struct btree *n2;
1243         struct bset *set1, *set2;
1244         struct bkey_packed *k, *set2_start, *set2_end, *out, *prev = NULL;
1245         struct bpos n1_pos;
1246
1247         n2 = bch2_btree_node_alloc(as, n1->c.level);
1248
1249         n2->data->max_key       = n1->data->max_key;
1250         n2->data->format        = n1->format;
1251         SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data));
1252         n2->key.k.p = n1->key.k.p;
1253
1254         bch2_btree_update_add_new_node(as, n2);
1255
1256         set1 = btree_bset_first(n1);
1257         set2 = btree_bset_first(n2);
1258
1259         /*
1260          * Has to be a linear search because we don't have an auxiliary
1261          * search tree yet
1262          */
1263         k = set1->start;
1264         while (1) {
1265                 struct bkey_packed *n = bkey_next(k);
1266
1267                 if (n == vstruct_last(set1))
1268                         break;
1269                 if (k->_data - set1->_data >= (le16_to_cpu(set1->u64s) * 3) / 5)
1270                         break;
1271
1272                 if (bkey_packed(k))
1273                         nr_packed++;
1274                 else
1275                         nr_unpacked++;
1276
1277                 prev = k;
1278                 k = n;
1279         }
1280
1281         BUG_ON(!prev);
1282         set2_start      = k;
1283         set2_end        = vstruct_last(set1);
1284
1285         set1->u64s = cpu_to_le16((u64 *) set2_start - set1->_data);
1286         set_btree_bset_end(n1, n1->set);
1287
1288         n1->nr.live_u64s        = le16_to_cpu(set1->u64s);
1289         n1->nr.bset_u64s[0]     = le16_to_cpu(set1->u64s);
1290         n1->nr.packed_keys      = nr_packed;
1291         n1->nr.unpacked_keys    = nr_unpacked;
1292
1293         n1_pos = bkey_unpack_pos(n1, prev);
1294         if (as->c->sb.version < bcachefs_metadata_version_snapshot)
1295                 n1_pos.snapshot = U32_MAX;
1296
1297         btree_set_max(n1, n1_pos);
1298         btree_set_min(n2, bpos_successor(n1->key.k.p));
1299
1300         bch2_bkey_format_init(&s);
1301         bch2_bkey_format_add_pos(&s, n2->data->min_key);
1302         bch2_bkey_format_add_pos(&s, n2->data->max_key);
1303
1304         for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1305                 struct bkey uk = bkey_unpack_key(n1, k);
1306                 bch2_bkey_format_add_key(&s, &uk);
1307         }
1308
1309         n2->data->format = bch2_bkey_format_done(&s);
1310         btree_node_set_format(n2, n2->data->format);
1311
1312         out = set2->start;
1313         memset(&n2->nr, 0, sizeof(n2->nr));
1314
1315         for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1316                 BUG_ON(!bch2_bkey_transform(&n2->format, out, bkey_packed(k)
1317                                        ? &n1->format : &bch2_bkey_format_current, k));
1318                 out->format = KEY_FORMAT_LOCAL_BTREE;
1319                 btree_keys_account_key_add(&n2->nr, 0, out);
1320                 out = bkey_next(out);
1321         }
1322
1323         set2->u64s = cpu_to_le16((u64 *) out - set2->_data);
1324         set_btree_bset_end(n2, n2->set);
1325
1326         BUG_ON(!set1->u64s);
1327         BUG_ON(!set2->u64s);
1328
1329         btree_node_reset_sib_u64s(n1);
1330         btree_node_reset_sib_u64s(n2);
1331
1332         bch2_verify_btree_nr_keys(n1);
1333         bch2_verify_btree_nr_keys(n2);
1334
1335         if (n1->c.level) {
1336                 btree_node_interior_verify(as->c, n1);
1337                 btree_node_interior_verify(as->c, n2);
1338         }
1339
1340         return n2;
1341 }
1342
1343 /*
1344  * For updates to interior nodes, we've got to do the insert before we split
1345  * because the stuff we're inserting has to be inserted atomically. Post split,
1346  * the keys might have to go in different nodes and the split would no longer be
1347  * atomic.
1348  *
1349  * Worse, if the insert is from btree node coalescing, if we do the insert after
1350  * we do the split (and pick the pivot) - the pivot we pick might be between
1351  * nodes that were coalesced, and thus in the middle of a child node post
1352  * coalescing:
1353  */
1354 static void btree_split_insert_keys(struct btree_update *as,
1355                                     struct btree_trans *trans,
1356                                     struct btree_path *path,
1357                                     struct btree *b,
1358                                     struct keylist *keys)
1359 {
1360         struct btree_node_iter node_iter;
1361         struct bkey_i *k = bch2_keylist_front(keys);
1362         struct bkey_packed *src, *dst, *n;
1363         struct bset *i;
1364
1365         bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
1366
1367         __bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
1368
1369         /*
1370          * We can't tolerate whiteouts here - with whiteouts there can be
1371          * duplicate keys, and it would be rather bad if we picked a duplicate
1372          * for the pivot:
1373          */
1374         i = btree_bset_first(b);
1375         src = dst = i->start;
1376         while (src != vstruct_last(i)) {
1377                 n = bkey_next(src);
1378                 if (!bkey_deleted(src)) {
1379                         memmove_u64s_down(dst, src, src->u64s);
1380                         dst = bkey_next(dst);
1381                 }
1382                 src = n;
1383         }
1384
1385         /* Also clear out the unwritten whiteouts area: */
1386         b->whiteout_u64s = 0;
1387
1388         i->u64s = cpu_to_le16((u64 *) dst - i->_data);
1389         set_btree_bset_end(b, b->set);
1390
1391         BUG_ON(b->nsets != 1 ||
1392                b->nr.live_u64s != le16_to_cpu(btree_bset_first(b)->u64s));
1393
1394         btree_node_interior_verify(as->c, b);
1395 }
1396
1397 static void btree_split(struct btree_update *as, struct btree_trans *trans,
1398                         struct btree_path *path, struct btree *b,
1399                         struct keylist *keys, unsigned flags)
1400 {
1401         struct bch_fs *c = as->c;
1402         struct btree *parent = btree_node_parent(path, b);
1403         struct btree *n1, *n2 = NULL, *n3 = NULL;
1404         u64 start_time = local_clock();
1405
1406         BUG_ON(!parent && (b != btree_node_root(c, b)));
1407         BUG_ON(!btree_node_intent_locked(path, btree_node_root(c, b)->c.level));
1408
1409         bch2_btree_interior_update_will_free_node(as, b);
1410
1411         n1 = bch2_btree_node_alloc_replacement(as, b);
1412
1413         if (keys)
1414                 btree_split_insert_keys(as, trans, path, n1, keys);
1415
1416         if (bset_u64s(&n1->set[0]) > BTREE_SPLIT_THRESHOLD(c)) {
1417                 trace_btree_split(c, b);
1418
1419                 n2 = __btree_split_node(as, n1);
1420
1421                 bch2_btree_build_aux_trees(n2);
1422                 bch2_btree_build_aux_trees(n1);
1423                 six_unlock_write(&n2->c.lock);
1424                 six_unlock_write(&n1->c.lock);
1425
1426                 bch2_btree_update_add_new_node(as, n1);
1427
1428                 bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1429                 bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0);
1430
1431                 /*
1432                  * Note that on recursive parent_keys == keys, so we
1433                  * can't start adding new keys to parent_keys before emptying it
1434                  * out (which we did with btree_split_insert_keys() above)
1435                  */
1436                 bch2_keylist_add(&as->parent_keys, &n1->key);
1437                 bch2_keylist_add(&as->parent_keys, &n2->key);
1438
1439                 if (!parent) {
1440                         /* Depth increases, make a new root */
1441                         n3 = __btree_root_alloc(as, b->c.level + 1);
1442
1443                         n3->sib_u64s[0] = U16_MAX;
1444                         n3->sib_u64s[1] = U16_MAX;
1445
1446                         btree_split_insert_keys(as, trans, path, n3, &as->parent_keys);
1447
1448                         bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0);
1449                 }
1450         } else {
1451                 trace_btree_compact(c, b);
1452
1453                 bch2_btree_build_aux_trees(n1);
1454                 six_unlock_write(&n1->c.lock);
1455
1456                 bch2_btree_update_add_new_node(as, n1);
1457
1458                 bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1459
1460                 if (parent)
1461                         bch2_keylist_add(&as->parent_keys, &n1->key);
1462         }
1463
1464         /* New nodes all written, now make them visible: */
1465
1466         if (parent) {
1467                 /* Split a non root node */
1468                 bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1469         } else if (n3) {
1470                 bch2_btree_set_root(as, trans, path, n3);
1471         } else {
1472                 /* Root filled up but didn't need to be split */
1473                 bch2_btree_set_root(as, trans, path, n1);
1474         }
1475
1476         bch2_btree_update_get_open_buckets(as, n1);
1477         if (n2)
1478                 bch2_btree_update_get_open_buckets(as, n2);
1479         if (n3)
1480                 bch2_btree_update_get_open_buckets(as, n3);
1481
1482         /* Successful split, update the path to point to the new nodes: */
1483
1484         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1485         if (n3)
1486                 bch2_trans_node_add(trans, n3);
1487         if (n2)
1488                 bch2_trans_node_add(trans, n2);
1489         bch2_trans_node_add(trans, n1);
1490
1491         /*
1492          * The old node must be freed (in memory) _before_ unlocking the new
1493          * nodes - else another thread could re-acquire a read lock on the old
1494          * node after another thread has locked and updated the new node, thus
1495          * seeing stale data:
1496          */
1497         bch2_btree_node_free_inmem(trans, b);
1498
1499         if (n3)
1500                 six_unlock_intent(&n3->c.lock);
1501         if (n2)
1502                 six_unlock_intent(&n2->c.lock);
1503         six_unlock_intent(&n1->c.lock);
1504
1505         bch2_trans_verify_locks(trans);
1506
1507         bch2_time_stats_update(&c->times[n2
1508                                ? BCH_TIME_btree_node_split
1509                                : BCH_TIME_btree_node_compact],
1510                                start_time);
1511 }
1512
1513 static void
1514 bch2_btree_insert_keys_interior(struct btree_update *as,
1515                                 struct btree_trans *trans,
1516                                 struct btree_path *path,
1517                                 struct btree *b,
1518                                 struct keylist *keys)
1519 {
1520         struct btree_path *linked;
1521
1522         __bch2_btree_insert_keys_interior(as, trans, path, b,
1523                                           path->l[b->c.level].iter, keys);
1524
1525         btree_update_updated_node(as, b);
1526
1527         trans_for_each_path_with_node(trans, b, linked)
1528                 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1529
1530         bch2_trans_verify_paths(trans);
1531 }
1532
1533 /**
1534  * bch_btree_insert_node - insert bkeys into a given btree node
1535  *
1536  * @iter:               btree iterator
1537  * @keys:               list of keys to insert
1538  * @hook:               insert callback
1539  * @persistent:         if not null, @persistent will wait on journal write
1540  *
1541  * Inserts as many keys as it can into a given btree node, splitting it if full.
1542  * If a split occurred, this function will return early. This can only happen
1543  * for leaf nodes -- inserts into interior nodes have to be atomic.
1544  */
1545 static void bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans,
1546                                    struct btree_path *path, struct btree *b,
1547                                    struct keylist *keys, unsigned flags)
1548 {
1549         struct bch_fs *c = as->c;
1550         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1551         int old_live_u64s = b->nr.live_u64s;
1552         int live_u64s_added, u64s_added;
1553
1554         lockdep_assert_held(&c->gc_lock);
1555         BUG_ON(!btree_node_intent_locked(path, btree_node_root(c, b)->c.level));
1556         BUG_ON(!b->c.level);
1557         BUG_ON(!as || as->b);
1558         bch2_verify_keylist_sorted(keys);
1559
1560         bch2_btree_node_lock_for_insert(trans, path, b);
1561
1562         if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1563                 bch2_btree_node_unlock_write(trans, path, b);
1564                 goto split;
1565         }
1566
1567         btree_node_interior_verify(c, b);
1568
1569         bch2_btree_insert_keys_interior(as, trans, path, b, keys);
1570
1571         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1572         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1573
1574         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1575                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1576         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1577                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1578
1579         if (u64s_added > live_u64s_added &&
1580             bch2_maybe_compact_whiteouts(c, b))
1581                 bch2_trans_node_reinit_iter(trans, b);
1582
1583         bch2_btree_node_unlock_write(trans, path, b);
1584
1585         btree_node_interior_verify(c, b);
1586         return;
1587 split:
1588         btree_split(as, trans, path, b, keys, flags);
1589 }
1590
1591 int bch2_btree_split_leaf(struct btree_trans *trans,
1592                           struct btree_path *path,
1593                           unsigned flags)
1594 {
1595         struct btree *b = path_l(path)->b;
1596         struct btree_update *as;
1597         unsigned l;
1598         int ret = 0;
1599
1600         as = bch2_btree_update_start(trans, path, path->level,
1601                                      true, flags);
1602         if (IS_ERR(as))
1603                 return PTR_ERR(as);
1604
1605         btree_split(as, trans, path, b, NULL, flags);
1606         bch2_btree_update_done(as);
1607
1608         for (l = path->level + 1; btree_path_node(path, l) && !ret; l++)
1609                 ret = bch2_foreground_maybe_merge(trans, path, l, flags);
1610
1611         return ret;
1612 }
1613
1614 int __bch2_foreground_maybe_merge(struct btree_trans *trans,
1615                                   struct btree_path *path,
1616                                   unsigned level,
1617                                   unsigned flags,
1618                                   enum btree_node_sibling sib)
1619 {
1620         struct bch_fs *c = trans->c;
1621         struct btree_path *sib_path = NULL;
1622         struct btree_update *as;
1623         struct bkey_format_state new_s;
1624         struct bkey_format new_f;
1625         struct bkey_i delete;
1626         struct btree *b, *m, *n, *prev, *next, *parent;
1627         struct bpos sib_pos;
1628         size_t sib_u64s;
1629         u64 start_time = local_clock();
1630         int ret = 0;
1631
1632         BUG_ON(!path->should_be_locked);
1633         BUG_ON(!btree_node_locked(path, level));
1634
1635         b = path->l[level].b;
1636
1637         if ((sib == btree_prev_sib && !bpos_cmp(b->data->min_key, POS_MIN)) ||
1638             (sib == btree_next_sib && !bpos_cmp(b->data->max_key, SPOS_MAX))) {
1639                 b->sib_u64s[sib] = U16_MAX;
1640                 return 0;
1641         }
1642
1643         sib_pos = sib == btree_prev_sib
1644                 ? bpos_predecessor(b->data->min_key)
1645                 : bpos_successor(b->data->max_key);
1646
1647         sib_path = bch2_path_get(trans, path->btree_id, sib_pos,
1648                                  U8_MAX, level, BTREE_ITER_INTENT, _THIS_IP_);
1649         ret = bch2_btree_path_traverse(trans, sib_path, false);
1650         if (ret)
1651                 goto err;
1652
1653         sib_path->should_be_locked = true;
1654
1655         m = sib_path->l[level].b;
1656
1657         if (btree_node_parent(path, b) !=
1658             btree_node_parent(sib_path, m)) {
1659                 b->sib_u64s[sib] = U16_MAX;
1660                 goto out;
1661         }
1662
1663         if (sib == btree_prev_sib) {
1664                 prev = m;
1665                 next = b;
1666         } else {
1667                 prev = b;
1668                 next = m;
1669         }
1670
1671         if (bkey_cmp(bpos_successor(prev->data->max_key), next->data->min_key)) {
1672                 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1673
1674                 bch2_bpos_to_text(&buf1, prev->data->max_key);
1675                 bch2_bpos_to_text(&buf2, next->data->min_key);
1676                 bch_err(c,
1677                         "btree topology error in btree merge:\n"
1678                         "  prev ends at   %s\n"
1679                         "  next starts at %s",
1680                         buf1.buf, buf2.buf);
1681                 printbuf_exit(&buf1);
1682                 printbuf_exit(&buf2);
1683                 bch2_topology_error(c);
1684                 ret = -EIO;
1685                 goto err;
1686         }
1687
1688         bch2_bkey_format_init(&new_s);
1689         bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
1690         __bch2_btree_calc_format(&new_s, prev);
1691         __bch2_btree_calc_format(&new_s, next);
1692         bch2_bkey_format_add_pos(&new_s, next->data->max_key);
1693         new_f = bch2_bkey_format_done(&new_s);
1694
1695         sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1696                 btree_node_u64s_with_format(m, &new_f);
1697
1698         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1699                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1700                 sib_u64s /= 2;
1701                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1702         }
1703
1704         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1705         sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
1706         b->sib_u64s[sib] = sib_u64s;
1707
1708         if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
1709                 goto out;
1710
1711         parent = btree_node_parent(path, b);
1712         as = bch2_btree_update_start(trans, path, level, false,
1713                          BTREE_INSERT_NOFAIL|
1714                          BTREE_INSERT_USE_RESERVE|
1715                          flags);
1716         ret = PTR_ERR_OR_ZERO(as);
1717         if (ret)
1718                 goto err;
1719
1720         trace_btree_merge(c, b);
1721
1722         bch2_btree_interior_update_will_free_node(as, b);
1723         bch2_btree_interior_update_will_free_node(as, m);
1724
1725         n = bch2_btree_node_alloc(as, b->c.level);
1726
1727         SET_BTREE_NODE_SEQ(n->data,
1728                            max(BTREE_NODE_SEQ(b->data),
1729                                BTREE_NODE_SEQ(m->data)) + 1);
1730
1731         btree_set_min(n, prev->data->min_key);
1732         btree_set_max(n, next->data->max_key);
1733
1734         bch2_btree_update_add_new_node(as, n);
1735
1736         n->data->format  = new_f;
1737         btree_node_set_format(n, new_f);
1738
1739         bch2_btree_sort_into(c, n, prev);
1740         bch2_btree_sort_into(c, n, next);
1741
1742         bch2_btree_build_aux_trees(n);
1743         six_unlock_write(&n->c.lock);
1744
1745         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1746
1747         bkey_init(&delete.k);
1748         delete.k.p = prev->key.k.p;
1749         bch2_keylist_add(&as->parent_keys, &delete);
1750         bch2_keylist_add(&as->parent_keys, &n->key);
1751
1752         bch2_trans_verify_paths(trans);
1753
1754         bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1755
1756         bch2_trans_verify_paths(trans);
1757
1758         bch2_btree_update_get_open_buckets(as, n);
1759
1760         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1761         six_lock_increment(&m->c.lock, SIX_LOCK_intent);
1762
1763         bch2_trans_node_add(trans, n);
1764
1765         bch2_trans_verify_paths(trans);
1766
1767         bch2_btree_node_free_inmem(trans, b);
1768         bch2_btree_node_free_inmem(trans, m);
1769
1770         six_unlock_intent(&n->c.lock);
1771
1772         bch2_btree_update_done(as);
1773
1774         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time);
1775 out:
1776 err:
1777         bch2_path_put(trans, sib_path, true);
1778         bch2_trans_verify_locks(trans);
1779         return ret;
1780 }
1781
1782 /**
1783  * bch_btree_node_rewrite - Rewrite/move a btree node
1784  */
1785 int bch2_btree_node_rewrite(struct btree_trans *trans,
1786                             struct btree_iter *iter,
1787                             struct btree *b,
1788                             unsigned flags)
1789 {
1790         struct bch_fs *c = trans->c;
1791         struct btree *n, *parent;
1792         struct btree_update *as;
1793         int ret;
1794
1795         flags |= BTREE_INSERT_NOFAIL;
1796
1797         parent = btree_node_parent(iter->path, b);
1798         as = bch2_btree_update_start(trans, iter->path, b->c.level,
1799                                      false, flags);
1800         ret = PTR_ERR_OR_ZERO(as);
1801         if (ret)
1802                 goto out;
1803
1804         bch2_btree_interior_update_will_free_node(as, b);
1805
1806         n = bch2_btree_node_alloc_replacement(as, b);
1807         bch2_btree_update_add_new_node(as, n);
1808
1809         bch2_btree_build_aux_trees(n);
1810         six_unlock_write(&n->c.lock);
1811
1812         trace_btree_rewrite(c, b);
1813
1814         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1815
1816         if (parent) {
1817                 bch2_keylist_add(&as->parent_keys, &n->key);
1818                 bch2_btree_insert_node(as, trans, iter->path, parent,
1819                                        &as->parent_keys, flags);
1820         } else {
1821                 bch2_btree_set_root(as, trans, iter->path, n);
1822         }
1823
1824         bch2_btree_update_get_open_buckets(as, n);
1825
1826         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1827         bch2_trans_node_add(trans, n);
1828         bch2_btree_node_free_inmem(trans, b);
1829         six_unlock_intent(&n->c.lock);
1830
1831         bch2_btree_update_done(as);
1832 out:
1833         bch2_btree_path_downgrade(iter->path);
1834         return ret;
1835 }
1836
1837 struct async_btree_rewrite {
1838         struct bch_fs           *c;
1839         struct work_struct      work;
1840         enum btree_id           btree_id;
1841         unsigned                level;
1842         struct bpos             pos;
1843         __le64                  seq;
1844 };
1845
1846 static int async_btree_node_rewrite_trans(struct btree_trans *trans,
1847                                           struct async_btree_rewrite *a)
1848 {
1849         struct btree_iter iter;
1850         struct btree *b;
1851         int ret;
1852
1853         bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos,
1854                                   BTREE_MAX_DEPTH, a->level, 0);
1855         b = bch2_btree_iter_peek_node(&iter);
1856         ret = PTR_ERR_OR_ZERO(b);
1857         if (ret)
1858                 goto out;
1859
1860         if (!b || b->data->keys.seq != a->seq)
1861                 goto out;
1862
1863         ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
1864 out :
1865         bch2_trans_iter_exit(trans, &iter);
1866
1867         return ret;
1868 }
1869
1870 void async_btree_node_rewrite_work(struct work_struct *work)
1871 {
1872         struct async_btree_rewrite *a =
1873                 container_of(work, struct async_btree_rewrite, work);
1874         struct bch_fs *c = a->c;
1875
1876         bch2_trans_do(c, NULL, NULL, 0,
1877                       async_btree_node_rewrite_trans(&trans, a));
1878         percpu_ref_put(&c->writes);
1879         kfree(a);
1880 }
1881
1882 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
1883 {
1884         struct async_btree_rewrite *a;
1885
1886         if (!percpu_ref_tryget(&c->writes))
1887                 return;
1888
1889         a = kmalloc(sizeof(*a), GFP_NOFS);
1890         if (!a) {
1891                 percpu_ref_put(&c->writes);
1892                 return;
1893         }
1894
1895         a->c            = c;
1896         a->btree_id     = b->c.btree_id;
1897         a->level        = b->c.level;
1898         a->pos          = b->key.k.p;
1899         a->seq          = b->data->keys.seq;
1900
1901         INIT_WORK(&a->work, async_btree_node_rewrite_work);
1902         queue_work(c->btree_interior_update_worker, &a->work);
1903 }
1904
1905 static int __bch2_btree_node_update_key(struct btree_trans *trans,
1906                                         struct btree_iter *iter,
1907                                         struct btree *b, struct btree *new_hash,
1908                                         struct bkey_i *new_key,
1909                                         bool skip_triggers)
1910 {
1911         struct bch_fs *c = trans->c;
1912         struct btree_iter iter2 = { NULL };
1913         struct btree *parent;
1914         int ret;
1915
1916         if (!skip_triggers) {
1917                 ret = bch2_trans_mark_old(trans, b->c.btree_id, b->c.level + 1,
1918                                           bkey_i_to_s_c(&b->key), 0);
1919                 if (ret)
1920                         return ret;
1921
1922                 ret = bch2_trans_mark_new(trans, b->c.btree_id, b->c.level + 1,
1923                                           new_key, 0);
1924                 if (ret)
1925                         return ret;
1926         }
1927
1928         if (new_hash) {
1929                 bkey_copy(&new_hash->key, new_key);
1930                 ret = bch2_btree_node_hash_insert(&c->btree_cache,
1931                                 new_hash, b->c.level, b->c.btree_id);
1932                 BUG_ON(ret);
1933         }
1934
1935         parent = btree_node_parent(iter->path, b);
1936         if (parent) {
1937                 bch2_trans_copy_iter(&iter2, iter);
1938
1939                 iter2.path = bch2_btree_path_make_mut(trans, iter2.path,
1940                                 iter2.flags & BTREE_ITER_INTENT,
1941                                 _THIS_IP_);
1942
1943                 BUG_ON(iter2.path->level != b->c.level);
1944                 BUG_ON(bpos_cmp(iter2.path->pos, new_key->k.p));
1945
1946                 btree_node_unlock(iter2.path, iter2.path->level);
1947                 path_l(iter2.path)->b = BTREE_ITER_NO_NODE_UP;
1948                 iter2.path->level++;
1949                 btree_path_set_dirty(iter2.path, BTREE_ITER_NEED_TRAVERSE);
1950
1951                 bch2_btree_path_check_sort(trans, iter2.path, 0);
1952
1953                 ret   = bch2_btree_iter_traverse(&iter2) ?:
1954                         bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_NORUN);
1955                 if (ret)
1956                         goto err;
1957         } else {
1958                 BUG_ON(btree_node_root(c, b) != b);
1959
1960                 ret = darray_make_room(&trans->extra_journal_entries,
1961                                        jset_u64s(new_key->k.u64s));
1962                 if (ret)
1963                         return ret;
1964
1965                 journal_entry_set((void *) &darray_top(trans->extra_journal_entries),
1966                                   BCH_JSET_ENTRY_btree_root,
1967                                   b->c.btree_id, b->c.level,
1968                                   new_key, new_key->k.u64s);
1969                 trans->extra_journal_entries.nr += jset_u64s(new_key->k.u64s);
1970         }
1971
1972         ret = bch2_trans_commit(trans, NULL, NULL,
1973                                 BTREE_INSERT_NOFAIL|
1974                                 BTREE_INSERT_NOCHECK_RW|
1975                                 BTREE_INSERT_USE_RESERVE|
1976                                 BTREE_INSERT_JOURNAL_RECLAIM|
1977                                 JOURNAL_WATERMARK_reserved);
1978         if (ret)
1979                 goto err;
1980
1981         bch2_btree_node_lock_write(trans, iter->path, b);
1982
1983         if (new_hash) {
1984                 mutex_lock(&c->btree_cache.lock);
1985                 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
1986                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1987
1988                 bkey_copy(&b->key, new_key);
1989                 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1990                 BUG_ON(ret);
1991                 mutex_unlock(&c->btree_cache.lock);
1992         } else {
1993                 bkey_copy(&b->key, new_key);
1994         }
1995
1996         bch2_btree_node_unlock_write(trans, iter->path, b);
1997 out:
1998         bch2_trans_iter_exit(trans, &iter2);
1999         return ret;
2000 err:
2001         if (new_hash) {
2002                 mutex_lock(&c->btree_cache.lock);
2003                 bch2_btree_node_hash_remove(&c->btree_cache, b);
2004                 mutex_unlock(&c->btree_cache.lock);
2005         }
2006         goto out;
2007 }
2008
2009 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
2010                                struct btree *b, struct bkey_i *new_key,
2011                                bool skip_triggers)
2012 {
2013         struct bch_fs *c = trans->c;
2014         struct btree *new_hash = NULL;
2015         struct btree_path *path = iter->path;
2016         struct closure cl;
2017         int ret = 0;
2018
2019         if (!btree_node_intent_locked(path, b->c.level) &&
2020             !bch2_btree_path_upgrade(trans, path, b->c.level + 1)) {
2021                 btree_trans_restart(trans);
2022                 return -EINTR;
2023         }
2024
2025         closure_init_stack(&cl);
2026
2027         /*
2028          * check btree_ptr_hash_val() after @b is locked by
2029          * btree_iter_traverse():
2030          */
2031         if (btree_ptr_hash_val(new_key) != b->hash_val) {
2032                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2033                 if (ret) {
2034                         bch2_trans_unlock(trans);
2035                         closure_sync(&cl);
2036                         if (!bch2_trans_relock(trans))
2037                                 return -EINTR;
2038                 }
2039
2040                 new_hash = bch2_btree_node_mem_alloc(c, false);
2041         }
2042
2043         path->intent_ref++;
2044         ret = __bch2_btree_node_update_key(trans, iter, b, new_hash,
2045                                            new_key, skip_triggers);
2046         --path->intent_ref;
2047
2048         if (new_hash) {
2049                 mutex_lock(&c->btree_cache.lock);
2050                 list_move(&new_hash->list, &c->btree_cache.freeable);
2051                 mutex_unlock(&c->btree_cache.lock);
2052
2053                 six_unlock_write(&new_hash->c.lock);
2054                 six_unlock_intent(&new_hash->c.lock);
2055         }
2056         closure_sync(&cl);
2057         bch2_btree_cache_cannibalize_unlock(c);
2058         return ret;
2059 }
2060
2061 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
2062                                         struct btree *b, struct bkey_i *new_key,
2063                                         bool skip_triggers)
2064 {
2065         struct btree_iter iter;
2066         int ret;
2067
2068         bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p,
2069                                   BTREE_MAX_DEPTH, b->c.level,
2070                                   BTREE_ITER_INTENT);
2071         ret = bch2_btree_iter_traverse(&iter);
2072         if (ret)
2073                 goto out;
2074
2075         /* has node been freed? */
2076         if (iter.path->l[b->c.level].b != b) {
2077                 /* node has been freed: */
2078                 BUG_ON(!btree_node_dying(b));
2079                 goto out;
2080         }
2081
2082         BUG_ON(!btree_node_hashed(b));
2083
2084         ret = bch2_btree_node_update_key(trans, &iter, b, new_key, skip_triggers);
2085 out:
2086         bch2_trans_iter_exit(trans, &iter);
2087         return ret;
2088 }
2089
2090 /* Init code: */
2091
2092 /*
2093  * Only for filesystem bringup, when first reading the btree roots or allocating
2094  * btree roots when initializing a new filesystem:
2095  */
2096 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2097 {
2098         BUG_ON(btree_node_root(c, b));
2099
2100         bch2_btree_set_root_inmem(c, b);
2101 }
2102
2103 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2104 {
2105         struct closure cl;
2106         struct btree *b;
2107         int ret;
2108
2109         closure_init_stack(&cl);
2110
2111         do {
2112                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2113                 closure_sync(&cl);
2114         } while (ret);
2115
2116         b = bch2_btree_node_mem_alloc(c, false);
2117         bch2_btree_cache_cannibalize_unlock(c);
2118
2119         set_btree_node_fake(b);
2120         set_btree_node_need_rewrite(b);
2121         b->c.level      = 0;
2122         b->c.btree_id   = id;
2123
2124         bkey_btree_ptr_init(&b->key);
2125         b->key.k.p = SPOS_MAX;
2126         *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2127
2128         bch2_bset_init_first(b, &b->data->keys);
2129         bch2_btree_build_aux_trees(b);
2130
2131         b->data->flags = 0;
2132         btree_set_min(b, POS_MIN);
2133         btree_set_max(b, SPOS_MAX);
2134         b->data->format = bch2_btree_calc_format(b);
2135         btree_node_set_format(b, b->data->format);
2136
2137         ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2138                                           b->c.level, b->c.btree_id);
2139         BUG_ON(ret);
2140
2141         bch2_btree_set_root_inmem(c, b);
2142
2143         six_unlock_write(&b->c.lock);
2144         six_unlock_intent(&b->c.lock);
2145 }
2146
2147 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2148 {
2149         struct btree_update *as;
2150
2151         mutex_lock(&c->btree_interior_update_lock);
2152         list_for_each_entry(as, &c->btree_interior_update_list, list)
2153                 pr_buf(out, "%p m %u w %u r %u j %llu\n",
2154                        as,
2155                        as->mode,
2156                        as->nodes_written,
2157                        atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2158                        as->journal.seq);
2159         mutex_unlock(&c->btree_interior_update_lock);
2160 }
2161
2162 static bool bch2_btree_interior_updates_pending(struct bch_fs *c)
2163 {
2164         bool ret;
2165
2166         mutex_lock(&c->btree_interior_update_lock);
2167         ret = !list_empty(&c->btree_interior_update_list);
2168         mutex_unlock(&c->btree_interior_update_lock);
2169
2170         return ret;
2171 }
2172
2173 bool bch2_btree_interior_updates_flush(struct bch_fs *c)
2174 {
2175         bool ret = bch2_btree_interior_updates_pending(c);
2176
2177         if (ret)
2178                 closure_wait_event(&c->btree_interior_update_wait,
2179                                    !bch2_btree_interior_updates_pending(c));
2180         return ret;
2181 }
2182
2183 void bch2_journal_entries_to_btree_roots(struct bch_fs *c, struct jset *jset)
2184 {
2185         struct btree_root *r;
2186         struct jset_entry *entry;
2187
2188         mutex_lock(&c->btree_root_lock);
2189
2190         vstruct_for_each(jset, entry)
2191                 if (entry->type == BCH_JSET_ENTRY_btree_root) {
2192                         r = &c->btree_roots[entry->btree_id];
2193                         r->level = entry->level;
2194                         r->alive = true;
2195                         bkey_copy(&r->key, &entry->start[0]);
2196                 }
2197
2198         mutex_unlock(&c->btree_root_lock);
2199 }
2200
2201 struct jset_entry *
2202 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2203                                     struct jset_entry *start,
2204                                     struct jset_entry *end)
2205 {
2206         struct jset_entry *entry;
2207         unsigned long have = 0;
2208         unsigned i;
2209
2210         for (entry = start; entry < end; entry = vstruct_next(entry))
2211                 if (entry->type == BCH_JSET_ENTRY_btree_root)
2212                         __set_bit(entry->btree_id, &have);
2213
2214         mutex_lock(&c->btree_root_lock);
2215
2216         for (i = 0; i < BTREE_ID_NR; i++)
2217                 if (c->btree_roots[i].alive && !test_bit(i, &have)) {
2218                         journal_entry_set(end,
2219                                           BCH_JSET_ENTRY_btree_root,
2220                                           i, c->btree_roots[i].level,
2221                                           &c->btree_roots[i].key,
2222                                           c->btree_roots[i].key.u64s);
2223                         end = vstruct_next(end);
2224                 }
2225
2226         mutex_unlock(&c->btree_root_lock);
2227
2228         return end;
2229 }
2230
2231 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2232 {
2233         if (c->btree_interior_update_worker)
2234                 destroy_workqueue(c->btree_interior_update_worker);
2235         mempool_exit(&c->btree_interior_update_pool);
2236 }
2237
2238 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2239 {
2240         mutex_init(&c->btree_reserve_cache_lock);
2241         INIT_LIST_HEAD(&c->btree_interior_update_list);
2242         INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2243         mutex_init(&c->btree_interior_update_lock);
2244         INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2245
2246         c->btree_interior_update_worker =
2247                 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2248         if (!c->btree_interior_update_worker)
2249                 return -ENOMEM;
2250
2251         return mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2252                                          sizeof(struct btree_update));
2253 }