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         memset(&b->data->_ptr, 0, sizeof(b->data->_ptr));
 290         b->data->flags = 0;
 291         SET_BTREE_NODE_ID(b->data, as->btree_id);
 292         SET_BTREE_NODE_LEVEL(b->data, level);
 293
 294         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 295                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
 296
 297                 bp->v.mem_ptr           = 0;
 298                 bp->v.seq               = b->data->keys.seq;
 299                 bp->v.sectors_written   = 0;
 300         }
 301
 302         SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
 303
 304         bch2_btree_build_aux_trees(b);
 305
 306         ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
 307         BUG_ON(ret);
 308
 309         trace_btree_node_alloc(c, b);
 310         return b;
 311 }
 312
 313 static void btree_set_min(struct btree *b, struct bpos pos)
 314 {
 315         if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
 316                 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
 317         b->data->min_key = pos;
 318 }
 319
 320 static void btree_set_max(struct btree *b, struct bpos pos)
 321 {
 322         b->key.k.p = pos;
 323         b->data->max_key = pos;
 324 }
 325
 326 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *as,
 327                                                   struct btree *b,
 328                                                   struct bkey_format format)
 329 {
 330         struct btree *n;
 331
 332         n = bch2_btree_node_alloc(as, b->c.level);
 333
 334         SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
 335
 336         btree_set_min(n, b->data->min_key);
 337         btree_set_max(n, b->data->max_key);
 338
 339         n->data->format         = format;
 340         btree_node_set_format(n, format);
 341
 342         bch2_btree_sort_into(as->c, n, b);
 343
 344         btree_node_reset_sib_u64s(n);
 345
 346         n->key.k.p = b->key.k.p;
 347         return n;
 348 }
 349
 350 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
 351                                                        struct btree *b)
 352 {
 353         struct bkey_format new_f = bch2_btree_calc_format(b);
 354
 355         /*
 356          * The keys might expand with the new format - if they wouldn't fit in
 357          * the btree node anymore, use the old format for now:
 358          */
 359         if (!bch2_btree_node_format_fits(as->c, b, &new_f))
 360                 new_f = b->format;
 361
 362         return __bch2_btree_node_alloc_replacement(as, b, new_f);
 363 }
 364
 365 static struct btree *__btree_root_alloc(struct btree_update *as, unsigned level)
 366 {
 367         struct btree *b = bch2_btree_node_alloc(as, level);
 368
 369         btree_set_min(b, POS_MIN);
 370         btree_set_max(b, POS_MAX);
 371         b->data->format = bch2_btree_calc_format(b);
 372
 373         btree_node_set_format(b, b->data->format);
 374         bch2_btree_build_aux_trees(b);
 375
 376         bch2_btree_update_add_new_node(as, b);
 377         six_unlock_write(&b->c.lock);
 378
 379         return b;
 380 }
 381
 382 static void bch2_btree_reserve_put(struct btree_update *as)
 383 {
 384         struct bch_fs *c = as->c;
 385
 386         mutex_lock(&c->btree_reserve_cache_lock);
 387
 388         while (as->nr_prealloc_nodes) {
 389                 struct btree *b = as->prealloc_nodes[--as->nr_prealloc_nodes];
 390
 391                 six_unlock_write(&b->c.lock);
 392
 393                 if (c->btree_reserve_cache_nr <
 394                     ARRAY_SIZE(c->btree_reserve_cache)) {
 395                         struct btree_alloc *a =
 396                                 &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
 397
 398                         a->ob = b->ob;
 399                         b->ob.nr = 0;
 400                         bkey_copy(&a->k, &b->key);
 401                 } else {
 402                         bch2_open_buckets_put(c, &b->ob);
 403                 }
 404
 405                 btree_node_lock_type(c, b, SIX_LOCK_write);
 406                 __btree_node_free(c, b);
 407                 six_unlock_write(&b->c.lock);
 408
 409                 six_unlock_intent(&b->c.lock);
 410         }
 411
 412         mutex_unlock(&c->btree_reserve_cache_lock);
 413 }
 414
 415 static int bch2_btree_reserve_get(struct btree_update *as, unsigned nr_nodes,
 416                                   unsigned flags, struct closure *cl)
 417 {
 418         struct bch_fs *c = as->c;
 419         struct btree *b;
 420         int ret;
 421
 422         BUG_ON(nr_nodes > BTREE_RESERVE_MAX);
 423
 424         /*
 425          * Protects reaping from the btree node cache and using the btree node
 426          * open bucket reserve:
 427          */
 428         ret = bch2_btree_cache_cannibalize_lock(c, cl);
 429         if (ret)
 430                 return ret;
 431
 432         while (as->nr_prealloc_nodes < nr_nodes) {
 433                 b = __bch2_btree_node_alloc(c, &as->disk_res,
 434                                             flags & BTREE_INSERT_NOWAIT
 435                                             ? NULL : cl, flags);
 436                 if (IS_ERR(b)) {
 437                         ret = PTR_ERR(b);
 438                         goto err_free;
 439                 }
 440
 441                 as->prealloc_nodes[as->nr_prealloc_nodes++] = b;
 442         }
 443
 444         bch2_btree_cache_cannibalize_unlock(c);
 445         return 0;
 446 err_free:
 447         bch2_btree_cache_cannibalize_unlock(c);
 448         trace_btree_reserve_get_fail(c, nr_nodes, cl);
 449         return ret;
 450 }
 451
 452 /* Asynchronous interior node update machinery */
 453
 454 static void bch2_btree_update_free(struct btree_update *as)
 455 {
 456         struct bch_fs *c = as->c;
 457
 458         if (as->took_gc_lock)
 459                 up_read(&c->gc_lock);
 460         as->took_gc_lock = false;
 461
 462         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 463
 464         bch2_journal_pin_drop(&c->journal, &as->journal);
 465         bch2_journal_pin_flush(&c->journal, &as->journal);
 466         bch2_disk_reservation_put(c, &as->disk_res);
 467         bch2_btree_reserve_put(as);
 468
 469         mutex_lock(&c->btree_interior_update_lock);
 470         list_del(&as->unwritten_list);
 471         list_del(&as->list);
 472         mutex_unlock(&c->btree_interior_update_lock);
 473
 474         closure_debug_destroy(&as->cl);
 475         mempool_free(as, &c->btree_interior_update_pool);
 476
 477         closure_wake_up(&c->btree_interior_update_wait);
 478 }
 479
 480 static void btree_update_will_delete_key(struct btree_update *as,
 481                                          struct bkey_i *k)
 482 {
 483         BUG_ON(bch2_keylist_u64s(&as->old_keys) + k->k.u64s >
 484                ARRAY_SIZE(as->_old_keys));
 485         bch2_keylist_add(&as->old_keys, k);
 486 }
 487
 488 static void btree_update_will_add_key(struct btree_update *as,
 489                                       struct bkey_i *k)
 490 {
 491         BUG_ON(bch2_keylist_u64s(&as->new_keys) + k->k.u64s >
 492                ARRAY_SIZE(as->_new_keys));
 493         bch2_keylist_add(&as->new_keys, k);
 494 }
 495
 496 /*
 497  * The transactional part of an interior btree node update, where we journal the
 498  * update we did to the interior node and update alloc info:
 499  */
 500 static int btree_update_nodes_written_trans(struct btree_trans *trans,
 501                                             struct btree_update *as)
 502 {
 503         struct bkey_i *k;
 504         int ret;
 505
 506         trans->extra_journal_entries = (void *) &as->journal_entries[0];
 507         trans->extra_journal_entry_u64s = as->journal_u64s;
 508         trans->journal_pin = &as->journal;
 509
 510         for_each_keylist_key(&as->new_keys, k) {
 511                 ret = bch2_trans_mark_key(trans,
 512                                           bkey_s_c_null,
 513                                           bkey_i_to_s_c(k),
 514                                           BTREE_TRIGGER_INSERT);
 515                 if (ret)
 516                         return ret;
 517         }
 518
 519         for_each_keylist_key(&as->old_keys, k) {
 520                 ret = bch2_trans_mark_key(trans,
 521                                           bkey_i_to_s_c(k),
 522                                           bkey_s_c_null,
 523                                           BTREE_TRIGGER_OVERWRITE);
 524                 if (ret)
 525                         return ret;
 526         }
 527
 528         return 0;
 529 }
 530
 531 static void btree_update_nodes_written(struct btree_update *as)
 532 {
 533         struct bch_fs *c = as->c;
 534         struct btree *b = as->b;
 535         struct btree_trans trans;
 536         u64 journal_seq = 0;
 537         unsigned i;
 538         int ret;
 539
 540         /*
 541          * If we're already in an error state, it might be because a btree node
 542          * was never written, and we might be trying to free that same btree
 543          * node here, but it won't have been marked as allocated and we'll see
 544          * spurious disk usage inconsistencies in the transactional part below
 545          * if we don't skip it:
 546          */
 547         ret = bch2_journal_error(&c->journal);
 548         if (ret)
 549                 goto err;
 550
 551         BUG_ON(!journal_pin_active(&as->journal));
 552
 553         /*
 554          * Wait for any in flight writes to finish before we free the old nodes
 555          * on disk:
 556          */
 557         for (i = 0; i < as->nr_old_nodes; i++) {
 558                 struct btree *old = as->old_nodes[i];
 559                 __le64 seq;
 560
 561                 six_lock_read(&old->c.lock, NULL, NULL);
 562                 seq = old->data ? old->data->keys.seq : 0;
 563                 six_unlock_read(&old->c.lock);
 564
 565                 if (seq == as->old_nodes_seq[i])
 566                         btree_node_wait_on_io(old);
 567         }
 568
 569         /*
 570          * We did an update to a parent node where the pointers we added pointed
 571          * to child nodes that weren't written yet: now, the child nodes have
 572          * been written so we can write out the update to the interior node.
 573          */
 574
 575         /*
 576          * We can't call into journal reclaim here: we'd block on the journal
 577          * reclaim lock, but we may need to release the open buckets we have
 578          * pinned in order for other btree updates to make forward progress, and
 579          * journal reclaim does btree updates when flushing bkey_cached entries,
 580          * which may require allocations as well.
 581          */
 582         bch2_trans_init(&trans, c, 0, 512);
 583         ret = __bch2_trans_do(&trans, &as->disk_res, &journal_seq,
 584                               BTREE_INSERT_NOFAIL|
 585                               BTREE_INSERT_NOCHECK_RW|
 586                               BTREE_INSERT_JOURNAL_RECLAIM|
 587                               BTREE_INSERT_JOURNAL_RESERVED,
 588                               btree_update_nodes_written_trans(&trans, as));
 589         bch2_trans_exit(&trans);
 590
 591         bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
 592                              "error %i in btree_update_nodes_written()", ret);
 593 err:
 594         if (b) {
 595                 /*
 596                  * @b is the node we did the final insert into:
 597                  *
 598                  * On failure to get a journal reservation, we still have to
 599                  * unblock the write and allow most of the write path to happen
 600                  * so that shutdown works, but the i->journal_seq mechanism
 601                  * won't work to prevent the btree write from being visible (we
 602                  * didn't get a journal sequence number) - instead
 603                  * __bch2_btree_node_write() doesn't do the actual write if
 604                  * we're in journal error state:
 605                  */
 606
 607                 btree_node_lock_type(c, b, SIX_LOCK_intent);
 608                 btree_node_lock_type(c, b, SIX_LOCK_write);
 609                 mutex_lock(&c->btree_interior_update_lock);
 610
 611                 list_del(&as->write_blocked_list);
 612
 613                 /*
 614                  * Node might have been freed, recheck under
 615                  * btree_interior_update_lock:
 616                  */
 617                 if (as->b == b) {
 618                         struct bset *i = btree_bset_last(b);
 619
 620                         BUG_ON(!b->c.level);
 621                         BUG_ON(!btree_node_dirty(b));
 622
 623                         if (!ret) {
 624                                 i->journal_seq = cpu_to_le64(
 625                                         max(journal_seq,
 626                                             le64_to_cpu(i->journal_seq)));
 627
 628                                 bch2_btree_add_journal_pin(c, b, journal_seq);
 629                         } else {
 630                                 /*
 631                                  * If we didn't get a journal sequence number we
 632                                  * can't write this btree node, because recovery
 633                                  * won't know to ignore this write:
 634                                  */
 635                                 set_btree_node_never_write(b);
 636                         }
 637                 }
 638
 639                 mutex_unlock(&c->btree_interior_update_lock);
 640                 six_unlock_write(&b->c.lock);
 641
 642                 btree_node_write_if_need(c, b, SIX_LOCK_intent);
 643                 six_unlock_intent(&b->c.lock);
 644         }
 645
 646         bch2_journal_pin_drop(&c->journal, &as->journal);
 647
 648         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 649
 650         mutex_lock(&c->btree_interior_update_lock);
 651         for (i = 0; i < as->nr_new_nodes; i++) {
 652                 b = as->new_nodes[i];
 653
 654                 BUG_ON(b->will_make_reachable != (unsigned long) as);
 655                 b->will_make_reachable = 0;
 656         }
 657         mutex_unlock(&c->btree_interior_update_lock);
 658
 659         for (i = 0; i < as->nr_new_nodes; i++) {
 660                 b = as->new_nodes[i];
 661
 662                 btree_node_lock_type(c, b, SIX_LOCK_read);
 663                 btree_node_write_if_need(c, b, SIX_LOCK_read);
 664                 six_unlock_read(&b->c.lock);
 665         }
 666
 667         for (i = 0; i < as->nr_open_buckets; i++)
 668                 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
 669
 670         bch2_btree_update_free(as);
 671 }
 672
 673 static void btree_interior_update_work(struct work_struct *work)
 674 {
 675         struct bch_fs *c =
 676                 container_of(work, struct bch_fs, btree_interior_update_work);
 677         struct btree_update *as;
 678
 679         while (1) {
 680                 mutex_lock(&c->btree_interior_update_lock);
 681                 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
 682                                               struct btree_update, unwritten_list);
 683                 if (as && !as->nodes_written)
 684                         as = NULL;
 685                 mutex_unlock(&c->btree_interior_update_lock);
 686
 687                 if (!as)
 688                         break;
 689
 690                 btree_update_nodes_written(as);
 691         }
 692 }
 693
 694 static void btree_update_set_nodes_written(struct closure *cl)
 695 {
 696         struct btree_update *as = container_of(cl, struct btree_update, cl);
 697         struct bch_fs *c = as->c;
 698
 699         mutex_lock(&c->btree_interior_update_lock);
 700         as->nodes_written = true;
 701         mutex_unlock(&c->btree_interior_update_lock);
 702
 703         queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
 704 }
 705
 706 /*
 707  * We're updating @b with pointers to nodes that haven't finished writing yet:
 708  * block @b from being written until @as completes
 709  */
 710 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
 711 {
 712         struct bch_fs *c = as->c;
 713
 714         mutex_lock(&c->btree_interior_update_lock);
 715         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 716
 717         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 718         BUG_ON(!btree_node_dirty(b));
 719
 720         as->mode        = BTREE_INTERIOR_UPDATING_NODE;
 721         as->b           = b;
 722         list_add(&as->write_blocked_list, &b->write_blocked);
 723
 724         mutex_unlock(&c->btree_interior_update_lock);
 725 }
 726
 727 static void btree_update_reparent(struct btree_update *as,
 728                                   struct btree_update *child)
 729 {
 730         struct bch_fs *c = as->c;
 731
 732         lockdep_assert_held(&c->btree_interior_update_lock);
 733
 734         child->b = NULL;
 735         child->mode = BTREE_INTERIOR_UPDATING_AS;
 736
 737         bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL);
 738 }
 739
 740 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
 741 {
 742         struct bkey_i *insert = &b->key;
 743         struct bch_fs *c = as->c;
 744
 745         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 746
 747         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
 748                ARRAY_SIZE(as->journal_entries));
 749
 750         as->journal_u64s +=
 751                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
 752                                   BCH_JSET_ENTRY_btree_root,
 753                                   b->c.btree_id, b->c.level,
 754                                   insert, insert->k.u64s);
 755
 756         mutex_lock(&c->btree_interior_update_lock);
 757         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 758
 759         as->mode        = BTREE_INTERIOR_UPDATING_ROOT;
 760         mutex_unlock(&c->btree_interior_update_lock);
 761 }
 762
 763 /*
 764  * bch2_btree_update_add_new_node:
 765  *
 766  * This causes @as to wait on @b to be written, before it gets to
 767  * bch2_btree_update_nodes_written
 768  *
 769  * Additionally, it sets b->will_make_reachable to prevent any additional writes
 770  * to @b from happening besides the first until @b is reachable on disk
 771  *
 772  * And it adds @b to the list of @as's new nodes, so that we can update sector
 773  * counts in bch2_btree_update_nodes_written:
 774  */
 775 void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
 776 {
 777         struct bch_fs *c = as->c;
 778
 779         closure_get(&as->cl);
 780
 781         mutex_lock(&c->btree_interior_update_lock);
 782         BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
 783         BUG_ON(b->will_make_reachable);
 784
 785         as->new_nodes[as->nr_new_nodes++] = b;
 786         b->will_make_reachable = 1UL|(unsigned long) as;
 787
 788         mutex_unlock(&c->btree_interior_update_lock);
 789
 790         btree_update_will_add_key(as, &b->key);
 791 }
 792
 793 /*
 794  * returns true if @b was a new node
 795  */
 796 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
 797 {
 798         struct btree_update *as;
 799         unsigned long v;
 800         unsigned i;
 801
 802         mutex_lock(&c->btree_interior_update_lock);
 803         /*
 804          * When b->will_make_reachable != 0, it owns a ref on as->cl that's
 805          * dropped when it gets written by bch2_btree_complete_write - the
 806          * xchg() is for synchronization with bch2_btree_complete_write:
 807          */
 808         v = xchg(&b->will_make_reachable, 0);
 809         as = (struct btree_update *) (v & ~1UL);
 810
 811         if (!as) {
 812                 mutex_unlock(&c->btree_interior_update_lock);
 813                 return;
 814         }
 815
 816         for (i = 0; i < as->nr_new_nodes; i++)
 817                 if (as->new_nodes[i] == b)
 818                         goto found;
 819
 820         BUG();
 821 found:
 822         array_remove_item(as->new_nodes, as->nr_new_nodes, i);
 823         mutex_unlock(&c->btree_interior_update_lock);
 824
 825         if (v & 1)
 826                 closure_put(&as->cl);
 827 }
 828
 829 void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
 830 {
 831         while (b->ob.nr)
 832                 as->open_buckets[as->nr_open_buckets++] =
 833                         b->ob.v[--b->ob.nr];
 834 }
 835
 836 /*
 837  * @b is being split/rewritten: it may have pointers to not-yet-written btree
 838  * nodes and thus outstanding btree_updates - redirect @b's
 839  * btree_updates to point to this btree_update:
 840  */
 841 void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 842                                                struct btree *b)
 843 {
 844         struct bch_fs *c = as->c;
 845         struct btree_update *p, *n;
 846         struct btree_write *w;
 847
 848         set_btree_node_dying(b);
 849
 850         if (btree_node_fake(b))
 851                 return;
 852
 853         mutex_lock(&c->btree_interior_update_lock);
 854
 855         /*
 856          * Does this node have any btree_update operations preventing
 857          * it from being written?
 858          *
 859          * If so, redirect them to point to this btree_update: we can
 860          * write out our new nodes, but we won't make them visible until those
 861          * operations complete
 862          */
 863         list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
 864                 list_del_init(&p->write_blocked_list);
 865                 btree_update_reparent(as, p);
 866
 867                 /*
 868                  * for flush_held_btree_writes() waiting on updates to flush or
 869                  * nodes to be writeable:
 870                  */
 871                 closure_wake_up(&c->btree_interior_update_wait);
 872         }
 873
 874         clear_btree_node_dirty(c, b);
 875         clear_btree_node_need_write(b);
 876
 877         /*
 878          * Does this node have unwritten data that has a pin on the journal?
 879          *
 880          * If so, transfer that pin to the btree_update operation -
 881          * note that if we're freeing multiple nodes, we only need to keep the
 882          * oldest pin of any of the nodes we're freeing. We'll release the pin
 883          * when the new nodes are persistent and reachable on disk:
 884          */
 885         w = btree_current_write(b);
 886         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 887         bch2_journal_pin_drop(&c->journal, &w->journal);
 888
 889         w = btree_prev_write(b);
 890         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 891         bch2_journal_pin_drop(&c->journal, &w->journal);
 892
 893         mutex_unlock(&c->btree_interior_update_lock);
 894
 895         /*
 896          * Is this a node that isn't reachable on disk yet?
 897          *
 898          * Nodes that aren't reachable yet have writes blocked until they're
 899          * reachable - now that we've cancelled any pending writes and moved
 900          * things waiting on that write to wait on this update, we can drop this
 901          * node from the list of nodes that the other update is making
 902          * reachable, prior to freeing it:
 903          */
 904         btree_update_drop_new_node(c, b);
 905
 906         btree_update_will_delete_key(as, &b->key);
 907
 908         as->old_nodes[as->nr_old_nodes] = b;
 909         as->old_nodes_seq[as->nr_old_nodes] = b->data->keys.seq;
 910         as->nr_old_nodes++;
 911 }
 912
 913 void bch2_btree_update_done(struct btree_update *as)
 914 {
 915         BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
 916
 917         if (as->took_gc_lock)
 918                 up_read(&as->c->gc_lock);
 919         as->took_gc_lock = false;
 920
 921         bch2_btree_reserve_put(as);
 922
 923         continue_at(&as->cl, btree_update_set_nodes_written,
 924                     as->c->btree_interior_update_worker);
 925 }
 926
 927 struct btree_update *
 928 bch2_btree_update_start(struct btree_iter *iter, unsigned level,
 929                         unsigned nr_nodes, unsigned flags)
 930 {
 931         struct btree_trans *trans = iter->trans;
 932         struct bch_fs *c = trans->c;
 933         struct btree_update *as;
 934         struct closure cl;
 935         int disk_res_flags = (flags & BTREE_INSERT_NOFAIL)
 936                 ? BCH_DISK_RESERVATION_NOFAIL : 0;
 937         int journal_flags = 0;
 938         int ret = 0;
 939
 940         BUG_ON(!iter->should_be_locked);
 941
 942         if (flags & BTREE_INSERT_JOURNAL_RESERVED)
 943                 journal_flags |= JOURNAL_RES_GET_RESERVED;
 944
 945         closure_init_stack(&cl);
 946 retry:
 947         /*
 948          * This check isn't necessary for correctness - it's just to potentially
 949          * prevent us from doing a lot of work that'll end up being wasted:
 950          */
 951         ret = bch2_journal_error(&c->journal);
 952         if (ret)
 953                 return ERR_PTR(ret);
 954
 955         /*
 956          * XXX: figure out how far we might need to split,
 957          * instead of locking/reserving all the way to the root:
 958          */
 959         if (!bch2_btree_iter_upgrade(iter, U8_MAX)) {
 960                 trace_trans_restart_iter_upgrade(trans->ip, _RET_IP_,
 961                                                  iter->btree_id,
 962                                                  &iter->real_pos);
 963                 return ERR_PTR(-EINTR);
 964         }
 965
 966         if (flags & BTREE_INSERT_GC_LOCK_HELD)
 967                 lockdep_assert_held(&c->gc_lock);
 968         else if (!down_read_trylock(&c->gc_lock)) {
 969                 if (flags & BTREE_INSERT_NOUNLOCK)
 970                         return ERR_PTR(-EINTR);
 971
 972                 bch2_trans_unlock(trans);
 973                 down_read(&c->gc_lock);
 974                 if (!bch2_trans_relock(trans)) {
 975                         up_read(&c->gc_lock);
 976                         return ERR_PTR(-EINTR);
 977                 }
 978         }
 979
 980         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
 981         memset(as, 0, sizeof(*as));
 982         closure_init(&as->cl, NULL);
 983         as->c           = c;
 984         as->mode        = BTREE_INTERIOR_NO_UPDATE;
 985         as->took_gc_lock = !(flags & BTREE_INSERT_GC_LOCK_HELD);
 986         as->btree_id    = iter->btree_id;
 987         INIT_LIST_HEAD(&as->list);
 988         INIT_LIST_HEAD(&as->unwritten_list);
 989         INIT_LIST_HEAD(&as->write_blocked_list);
 990         bch2_keylist_init(&as->old_keys, as->_old_keys);
 991         bch2_keylist_init(&as->new_keys, as->_new_keys);
 992         bch2_keylist_init(&as->parent_keys, as->inline_keys);
 993
 994         ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
 995                                       BTREE_UPDATE_JOURNAL_RES,
 996                                       journal_flags|JOURNAL_RES_GET_NONBLOCK);
 997         if (ret == -EAGAIN) {
 998                 /*
 999                  * this would be cleaner if bch2_journal_preres_get() took a
1000                  * closure argument
1001                  */
1002                 if (flags & BTREE_INSERT_NOUNLOCK) {
1003                         trace_trans_restart_journal_preres_get(trans->ip, _RET_IP_);
1004                         ret = -EINTR;
1005                         goto err;
1006                 }
1007
1008                 bch2_trans_unlock(trans);
1009
1010                 if (flags & BTREE_INSERT_JOURNAL_RECLAIM) {
1011                         bch2_btree_update_free(as);
1012                         return ERR_PTR(ret);
1013                 }
1014
1015                 ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
1016                                 BTREE_UPDATE_JOURNAL_RES,
1017                                 journal_flags);
1018                 if (ret) {
1019                         trace_trans_restart_journal_preres_get(trans->ip, _RET_IP_);
1020                         goto err;
1021                 }
1022
1023                 if (!bch2_trans_relock(trans)) {
1024                         ret = -EINTR;
1025                         goto err;
1026                 }
1027         }
1028
1029         ret = bch2_disk_reservation_get(c, &as->disk_res,
1030                         nr_nodes * c->opts.btree_node_size,
1031                         c->opts.metadata_replicas,
1032                         disk_res_flags);
1033         if (ret)
1034                 goto err;
1035
1036         ret = bch2_btree_reserve_get(as, nr_nodes, flags,
1037                 !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1038         if (ret)
1039                 goto err;
1040
1041         bch2_journal_pin_add(&c->journal,
1042                              atomic64_read(&c->journal.seq),
1043                              &as->journal, NULL);
1044
1045         mutex_lock(&c->btree_interior_update_lock);
1046         list_add_tail(&as->list, &c->btree_interior_update_list);
1047         mutex_unlock(&c->btree_interior_update_lock);
1048
1049         return as;
1050 err:
1051         bch2_btree_update_free(as);
1052
1053         if (ret == -EAGAIN) {
1054                 BUG_ON(flags & BTREE_INSERT_NOUNLOCK);
1055
1056                 bch2_trans_unlock(trans);
1057                 closure_sync(&cl);
1058                 ret = -EINTR;
1059         }
1060
1061         if (ret == -EINTR && bch2_trans_relock(trans))
1062                 goto retry;
1063
1064         return ERR_PTR(ret);
1065 }
1066
1067 /* Btree root updates: */
1068
1069 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1070 {
1071         /* Root nodes cannot be reaped */
1072         mutex_lock(&c->btree_cache.lock);
1073         list_del_init(&b->list);
1074         mutex_unlock(&c->btree_cache.lock);
1075
1076         if (b->c.level)
1077                 six_lock_pcpu_alloc(&b->c.lock);
1078         else
1079                 six_lock_pcpu_free(&b->c.lock);
1080
1081         mutex_lock(&c->btree_root_lock);
1082         BUG_ON(btree_node_root(c, b) &&
1083                (b->c.level < btree_node_root(c, b)->c.level ||
1084                 !btree_node_dying(btree_node_root(c, b))));
1085
1086         btree_node_root(c, b) = b;
1087         mutex_unlock(&c->btree_root_lock);
1088
1089         bch2_recalc_btree_reserve(c);
1090 }
1091
1092 /**
1093  * bch_btree_set_root - update the root in memory and on disk
1094  *
1095  * To ensure forward progress, the current task must not be holding any
1096  * btree node write locks. However, you must hold an intent lock on the
1097  * old root.
1098  *
1099  * Note: This allocates a journal entry but doesn't add any keys to
1100  * it.  All the btree roots are part of every journal write, so there
1101  * is nothing new to be done.  This just guarantees that there is a
1102  * journal write.
1103  */
1104 static void bch2_btree_set_root(struct btree_update *as, struct btree *b,
1105                                 struct btree_iter *iter)
1106 {
1107         struct bch_fs *c = as->c;
1108         struct btree *old;
1109
1110         trace_btree_set_root(c, b);
1111         BUG_ON(!b->written &&
1112                !test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags));
1113
1114         old = btree_node_root(c, b);
1115
1116         /*
1117          * Ensure no one is using the old root while we switch to the
1118          * new root:
1119          */
1120         bch2_btree_node_lock_write(old, iter);
1121
1122         bch2_btree_set_root_inmem(c, b);
1123
1124         btree_update_updated_root(as, b);
1125
1126         /*
1127          * Unlock old root after new root is visible:
1128          *
1129          * The new root isn't persistent, but that's ok: we still have
1130          * an intent lock on the new root, and any updates that would
1131          * depend on the new root would have to update the new root.
1132          */
1133         bch2_btree_node_unlock_write(old, iter);
1134 }
1135
1136 /* Interior node updates: */
1137
1138 static void bch2_insert_fixup_btree_ptr(struct btree_update *as, struct btree *b,
1139                                         struct btree_iter *iter,
1140                                         struct bkey_i *insert,
1141                                         struct btree_node_iter *node_iter)
1142 {
1143         struct bch_fs *c = as->c;
1144         struct bkey_packed *k;
1145         const char *invalid;
1146
1147         invalid = bch2_bkey_invalid(c, bkey_i_to_s_c(insert), btree_node_type(b)) ?:
1148                 bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert));
1149         if (invalid) {
1150                 char buf[160];
1151
1152                 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(insert));
1153                 bch2_fs_inconsistent(c, "inserting invalid bkey %s: %s", buf, invalid);
1154                 dump_stack();
1155         }
1156
1157         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1158                ARRAY_SIZE(as->journal_entries));
1159
1160         as->journal_u64s +=
1161                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1162                                   BCH_JSET_ENTRY_btree_keys,
1163                                   b->c.btree_id, b->c.level,
1164                                   insert, insert->k.u64s);
1165
1166         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1167                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1168                 bch2_btree_node_iter_advance(node_iter, b);
1169
1170         bch2_btree_bset_insert_key(iter, b, node_iter, insert);
1171         set_btree_node_dirty(c, b);
1172         set_btree_node_need_write(b);
1173 }
1174
1175 static void
1176 __bch2_btree_insert_keys_interior(struct btree_update *as, struct btree *b,
1177                                   struct btree_iter *iter, struct keylist *keys,
1178                                   struct btree_node_iter node_iter)
1179 {
1180         struct bkey_i *insert = bch2_keylist_front(keys);
1181         struct bkey_packed *k;
1182
1183         BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1184
1185         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1186                (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1187                 ;
1188
1189         while (!bch2_keylist_empty(keys)) {
1190                 bch2_insert_fixup_btree_ptr(as, b, iter,
1191                                 bch2_keylist_front(keys), &node_iter);
1192                 bch2_keylist_pop_front(keys);
1193         }
1194 }
1195
1196 /*
1197  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1198  * node)
1199  */
1200 static struct btree *__btree_split_node(struct btree_update *as,
1201                                         struct btree *n1,
1202                                         struct btree_iter *iter)
1203 {
1204         struct bkey_format_state s;
1205         size_t nr_packed = 0, nr_unpacked = 0;
1206         struct btree *n2;
1207         struct bset *set1, *set2;
1208         struct bkey_packed *k, *set2_start, *set2_end, *out, *prev = NULL;
1209         struct bpos n1_pos;
1210
1211         n2 = bch2_btree_node_alloc(as, n1->c.level);
1212         bch2_btree_update_add_new_node(as, n2);
1213
1214         n2->data->max_key       = n1->data->max_key;
1215         n2->data->format        = n1->format;
1216         SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data));
1217         n2->key.k.p = n1->key.k.p;
1218
1219         set1 = btree_bset_first(n1);
1220         set2 = btree_bset_first(n2);
1221
1222         /*
1223          * Has to be a linear search because we don't have an auxiliary
1224          * search tree yet
1225          */
1226         k = set1->start;
1227         while (1) {
1228                 struct bkey_packed *n = bkey_next(k);
1229
1230                 if (n == vstruct_last(set1))
1231                         break;
1232                 if (k->_data - set1->_data >= (le16_to_cpu(set1->u64s) * 3) / 5)
1233                         break;
1234
1235                 if (bkey_packed(k))
1236                         nr_packed++;
1237                 else
1238                         nr_unpacked++;
1239
1240                 prev = k;
1241                 k = n;
1242         }
1243
1244         BUG_ON(!prev);
1245         set2_start      = k;
1246         set2_end        = vstruct_last(set1);
1247
1248         set1->u64s = cpu_to_le16((u64 *) set2_start - set1->_data);
1249         set_btree_bset_end(n1, n1->set);
1250
1251         n1->nr.live_u64s        = le16_to_cpu(set1->u64s);
1252         n1->nr.bset_u64s[0]     = le16_to_cpu(set1->u64s);
1253         n1->nr.packed_keys      = nr_packed;
1254         n1->nr.unpacked_keys    = nr_unpacked;
1255
1256         n1_pos = bkey_unpack_pos(n1, prev);
1257         if (as->c->sb.version < bcachefs_metadata_version_snapshot)
1258                 n1_pos.snapshot = U32_MAX;
1259
1260         btree_set_max(n1, n1_pos);
1261         btree_set_min(n2, bpos_successor(n1->key.k.p));
1262
1263         bch2_bkey_format_init(&s);
1264         bch2_bkey_format_add_pos(&s, n2->data->min_key);
1265         bch2_bkey_format_add_pos(&s, n2->data->max_key);
1266
1267         for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1268                 struct bkey uk = bkey_unpack_key(n1, k);
1269                 bch2_bkey_format_add_key(&s, &uk);
1270         }
1271
1272         n2->data->format = bch2_bkey_format_done(&s);
1273         btree_node_set_format(n2, n2->data->format);
1274
1275         out = set2->start;
1276         memset(&n2->nr, 0, sizeof(n2->nr));
1277
1278         for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1279                 BUG_ON(!bch2_bkey_transform(&n2->format, out, bkey_packed(k)
1280                                        ? &n1->format : &bch2_bkey_format_current, k));
1281                 out->format = KEY_FORMAT_LOCAL_BTREE;
1282                 btree_keys_account_key_add(&n2->nr, 0, out);
1283                 out = bkey_next(out);
1284         }
1285
1286         set2->u64s = cpu_to_le16((u64 *) out - set2->_data);
1287         set_btree_bset_end(n2, n2->set);
1288
1289         BUG_ON(!set1->u64s);
1290         BUG_ON(!set2->u64s);
1291
1292         btree_node_reset_sib_u64s(n1);
1293         btree_node_reset_sib_u64s(n2);
1294
1295         bch2_verify_btree_nr_keys(n1);
1296         bch2_verify_btree_nr_keys(n2);
1297
1298         if (n1->c.level) {
1299                 btree_node_interior_verify(as->c, n1);
1300                 btree_node_interior_verify(as->c, n2);
1301         }
1302
1303         return n2;
1304 }
1305
1306 /*
1307  * For updates to interior nodes, we've got to do the insert before we split
1308  * because the stuff we're inserting has to be inserted atomically. Post split,
1309  * the keys might have to go in different nodes and the split would no longer be
1310  * atomic.
1311  *
1312  * Worse, if the insert is from btree node coalescing, if we do the insert after
1313  * we do the split (and pick the pivot) - the pivot we pick might be between
1314  * nodes that were coalesced, and thus in the middle of a child node post
1315  * coalescing:
1316  */
1317 static void btree_split_insert_keys(struct btree_update *as, struct btree *b,
1318                                     struct btree_iter *iter,
1319                                     struct keylist *keys)
1320 {
1321         struct btree_node_iter node_iter;
1322         struct bkey_i *k = bch2_keylist_front(keys);
1323         struct bkey_packed *src, *dst, *n;
1324         struct bset *i;
1325
1326         bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
1327
1328         __bch2_btree_insert_keys_interior(as, b, iter, keys, node_iter);
1329
1330         /*
1331          * We can't tolerate whiteouts here - with whiteouts there can be
1332          * duplicate keys, and it would be rather bad if we picked a duplicate
1333          * for the pivot:
1334          */
1335         i = btree_bset_first(b);
1336         src = dst = i->start;
1337         while (src != vstruct_last(i)) {
1338                 n = bkey_next(src);
1339                 if (!bkey_deleted(src)) {
1340                         memmove_u64s_down(dst, src, src->u64s);
1341                         dst = bkey_next(dst);
1342                 }
1343                 src = n;
1344         }
1345
1346         /* Also clear out the unwritten whiteouts area: */
1347         b->whiteout_u64s = 0;
1348
1349         i->u64s = cpu_to_le16((u64 *) dst - i->_data);
1350         set_btree_bset_end(b, b->set);
1351
1352         BUG_ON(b->nsets != 1 ||
1353                b->nr.live_u64s != le16_to_cpu(btree_bset_first(b)->u64s));
1354
1355         btree_node_interior_verify(as->c, b);
1356 }
1357
1358 static void btree_split(struct btree_update *as, struct btree *b,
1359                         struct btree_iter *iter, struct keylist *keys,
1360                         unsigned flags)
1361 {
1362         struct bch_fs *c = as->c;
1363         struct btree *parent = btree_node_parent(iter, b);
1364         struct btree *n1, *n2 = NULL, *n3 = NULL;
1365         u64 start_time = local_clock();
1366
1367         BUG_ON(!parent && (b != btree_node_root(c, b)));
1368         BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->c.level));
1369
1370         bch2_btree_interior_update_will_free_node(as, b);
1371
1372         n1 = bch2_btree_node_alloc_replacement(as, b);
1373         bch2_btree_update_add_new_node(as, n1);
1374
1375         if (keys)
1376                 btree_split_insert_keys(as, n1, iter, keys);
1377
1378         if (bset_u64s(&n1->set[0]) > BTREE_SPLIT_THRESHOLD(c)) {
1379                 trace_btree_split(c, b);
1380
1381                 n2 = __btree_split_node(as, n1, iter);
1382
1383                 bch2_btree_build_aux_trees(n2);
1384                 bch2_btree_build_aux_trees(n1);
1385                 six_unlock_write(&n2->c.lock);
1386                 six_unlock_write(&n1->c.lock);
1387
1388                 bch2_btree_node_write(c, n2, SIX_LOCK_intent);
1389
1390                 /*
1391                  * Note that on recursive parent_keys == keys, so we
1392                  * can't start adding new keys to parent_keys before emptying it
1393                  * out (which we did with btree_split_insert_keys() above)
1394                  */
1395                 bch2_keylist_add(&as->parent_keys, &n1->key);
1396                 bch2_keylist_add(&as->parent_keys, &n2->key);
1397
1398                 if (!parent) {
1399                         /* Depth increases, make a new root */
1400                         n3 = __btree_root_alloc(as, b->c.level + 1);
1401
1402                         n3->sib_u64s[0] = U16_MAX;
1403                         n3->sib_u64s[1] = U16_MAX;
1404
1405                         btree_split_insert_keys(as, n3, iter, &as->parent_keys);
1406
1407                         bch2_btree_node_write(c, n3, SIX_LOCK_intent);
1408                 }
1409         } else {
1410                 trace_btree_compact(c, b);
1411
1412                 bch2_btree_build_aux_trees(n1);
1413                 six_unlock_write(&n1->c.lock);
1414
1415                 if (parent)
1416                         bch2_keylist_add(&as->parent_keys, &n1->key);
1417         }
1418
1419         bch2_btree_node_write(c, n1, SIX_LOCK_intent);
1420
1421         /* New nodes all written, now make them visible: */
1422
1423         if (parent) {
1424                 /* Split a non root node */
1425                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1426         } else if (n3) {
1427                 bch2_btree_set_root(as, n3, iter);
1428         } else {
1429                 /* Root filled up but didn't need to be split */
1430                 bch2_btree_set_root(as, n1, iter);
1431         }
1432
1433         bch2_btree_update_get_open_buckets(as, n1);
1434         if (n2)
1435                 bch2_btree_update_get_open_buckets(as, n2);
1436         if (n3)
1437                 bch2_btree_update_get_open_buckets(as, n3);
1438
1439         /* Successful split, update the iterator to point to the new nodes: */
1440
1441         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1442         bch2_btree_iter_node_drop(iter, b);
1443         if (n3)
1444                 bch2_btree_iter_node_replace(iter, n3);
1445         if (n2)
1446                 bch2_btree_iter_node_replace(iter, n2);
1447         bch2_btree_iter_node_replace(iter, n1);
1448
1449         /*
1450          * The old node must be freed (in memory) _before_ unlocking the new
1451          * nodes - else another thread could re-acquire a read lock on the old
1452          * node after another thread has locked and updated the new node, thus
1453          * seeing stale data:
1454          */
1455         bch2_btree_node_free_inmem(c, b, iter);
1456
1457         if (n3)
1458                 six_unlock_intent(&n3->c.lock);
1459         if (n2)
1460                 six_unlock_intent(&n2->c.lock);
1461         six_unlock_intent(&n1->c.lock);
1462
1463         bch2_btree_trans_verify_locks(iter->trans);
1464
1465         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_split],
1466                                start_time);
1467 }
1468
1469 static void
1470 bch2_btree_insert_keys_interior(struct btree_update *as, struct btree *b,
1471                                 struct btree_iter *iter, struct keylist *keys)
1472 {
1473         struct btree_iter *linked;
1474
1475         __bch2_btree_insert_keys_interior(as, b, iter, keys, iter->l[b->c.level].iter);
1476
1477         btree_update_updated_node(as, b);
1478
1479         trans_for_each_iter_with_node(iter->trans, b, linked)
1480                 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1481
1482         bch2_btree_trans_verify_iters(iter->trans, b);
1483 }
1484
1485 /**
1486  * bch_btree_insert_node - insert bkeys into a given btree node
1487  *
1488  * @iter:               btree iterator
1489  * @keys:               list of keys to insert
1490  * @hook:               insert callback
1491  * @persistent:         if not null, @persistent will wait on journal write
1492  *
1493  * Inserts as many keys as it can into a given btree node, splitting it if full.
1494  * If a split occurred, this function will return early. This can only happen
1495  * for leaf nodes -- inserts into interior nodes have to be atomic.
1496  */
1497 void bch2_btree_insert_node(struct btree_update *as, struct btree *b,
1498                             struct btree_iter *iter, struct keylist *keys,
1499                             unsigned flags)
1500 {
1501         struct bch_fs *c = as->c;
1502         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1503         int old_live_u64s = b->nr.live_u64s;
1504         int live_u64s_added, u64s_added;
1505
1506         lockdep_assert_held(&c->gc_lock);
1507         BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->c.level));
1508         BUG_ON(!b->c.level);
1509         BUG_ON(!as || as->b);
1510         bch2_verify_keylist_sorted(keys);
1511
1512         bch2_btree_node_lock_for_insert(c, b, iter);
1513
1514         if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1515                 bch2_btree_node_unlock_write(b, iter);
1516                 goto split;
1517         }
1518
1519         btree_node_interior_verify(c, b);
1520
1521         bch2_btree_insert_keys_interior(as, b, iter, keys);
1522
1523         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1524         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1525
1526         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1527                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1528         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1529                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1530
1531         if (u64s_added > live_u64s_added &&
1532             bch2_maybe_compact_whiteouts(c, b))
1533                 bch2_btree_iter_reinit_node(iter, b);
1534
1535         bch2_btree_node_unlock_write(b, iter);
1536
1537         btree_node_interior_verify(c, b);
1538         return;
1539 split:
1540         btree_split(as, b, iter, keys, flags);
1541 }
1542
1543 int bch2_btree_split_leaf(struct bch_fs *c, struct btree_iter *iter,
1544                           unsigned flags)
1545 {
1546         struct btree *b = iter_l(iter)->b;
1547         struct btree_update *as;
1548         unsigned l;
1549         int ret = 0;
1550
1551         as = bch2_btree_update_start(iter, iter->level,
1552                 btree_update_reserve_required(c, b), flags);
1553         if (IS_ERR(as))
1554                 return PTR_ERR(as);
1555
1556         btree_split(as, b, iter, NULL, flags);
1557         bch2_btree_update_done(as);
1558
1559         for (l = iter->level + 1; btree_iter_node(iter, l) && !ret; l++)
1560                 ret = bch2_foreground_maybe_merge(c, iter, l, flags);
1561
1562         return ret;
1563 }
1564
1565 int __bch2_foreground_maybe_merge(struct bch_fs *c,
1566                                   struct btree_iter *iter,
1567                                   unsigned level,
1568                                   unsigned flags,
1569                                   enum btree_node_sibling sib)
1570 {
1571         struct btree_trans *trans = iter->trans;
1572         struct btree_iter *sib_iter = NULL;
1573         struct btree_update *as;
1574         struct bkey_format_state new_s;
1575         struct bkey_format new_f;
1576         struct bkey_i delete;
1577         struct btree *b, *m, *n, *prev, *next, *parent;
1578         struct bpos sib_pos;
1579         size_t sib_u64s;
1580         int ret = 0, ret2 = 0;
1581
1582         BUG_ON(!btree_node_locked(iter, level));
1583 retry:
1584         ret = bch2_btree_iter_traverse(iter);
1585         if (ret)
1586                 goto err;
1587
1588         BUG_ON(!btree_node_locked(iter, level));
1589
1590         b = iter->l[level].b;
1591
1592         if ((sib == btree_prev_sib && !bpos_cmp(b->data->min_key, POS_MIN)) ||
1593             (sib == btree_next_sib && !bpos_cmp(b->data->max_key, POS_MAX))) {
1594                 b->sib_u64s[sib] = U16_MAX;
1595                 goto out;
1596         }
1597
1598         sib_pos = sib == btree_prev_sib
1599                 ? bpos_predecessor(b->data->min_key)
1600                 : bpos_successor(b->data->max_key);
1601
1602         sib_iter = bch2_trans_get_node_iter(trans, iter->btree_id,
1603                                             sib_pos, U8_MAX, level,
1604                                             BTREE_ITER_INTENT);
1605         ret = bch2_btree_iter_traverse(sib_iter);
1606         if (ret)
1607                 goto err;
1608
1609         m = sib_iter->l[level].b;
1610
1611         if (btree_node_parent(iter, b) !=
1612             btree_node_parent(sib_iter, m)) {
1613                 b->sib_u64s[sib] = U16_MAX;
1614                 goto out;
1615         }
1616
1617         if (sib == btree_prev_sib) {
1618                 prev = m;
1619                 next = b;
1620         } else {
1621                 prev = b;
1622                 next = m;
1623         }
1624
1625         if (bkey_cmp(bpos_successor(prev->data->max_key), next->data->min_key)) {
1626                 char buf1[100], buf2[100];
1627
1628                 bch2_bpos_to_text(&PBUF(buf1), prev->data->max_key);
1629                 bch2_bpos_to_text(&PBUF(buf2), next->data->min_key);
1630                 bch_err(c,
1631                         "btree topology error in btree merge:\n"
1632                         "  prev ends at   %s\n"
1633                         "  next starts at %s",
1634                         buf1, buf2);
1635                 bch2_topology_error(c);
1636                 ret = -EIO;
1637                 goto err;
1638         }
1639
1640         bch2_bkey_format_init(&new_s);
1641         bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
1642         __bch2_btree_calc_format(&new_s, prev);
1643         __bch2_btree_calc_format(&new_s, next);
1644         bch2_bkey_format_add_pos(&new_s, next->data->max_key);
1645         new_f = bch2_bkey_format_done(&new_s);
1646
1647         sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1648                 btree_node_u64s_with_format(m, &new_f);
1649
1650         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1651                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1652                 sib_u64s /= 2;
1653                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1654         }
1655
1656         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1657         sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
1658         b->sib_u64s[sib] = sib_u64s;
1659
1660         if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
1661                 goto out;
1662
1663         parent = btree_node_parent(iter, b);
1664         as = bch2_btree_update_start(iter, level,
1665                          btree_update_reserve_required(c, parent) + 1,
1666                          flags|
1667                          BTREE_INSERT_NOFAIL|
1668                          BTREE_INSERT_USE_RESERVE);
1669         ret = PTR_ERR_OR_ZERO(as);
1670         if (ret)
1671                 goto err;
1672
1673         trace_btree_merge(c, b);
1674
1675         bch2_btree_interior_update_will_free_node(as, b);
1676         bch2_btree_interior_update_will_free_node(as, m);
1677
1678         n = bch2_btree_node_alloc(as, b->c.level);
1679         bch2_btree_update_add_new_node(as, n);
1680
1681         btree_set_min(n, prev->data->min_key);
1682         btree_set_max(n, next->data->max_key);
1683         n->data->format         = new_f;
1684
1685         btree_node_set_format(n, new_f);
1686
1687         bch2_btree_sort_into(c, n, prev);
1688         bch2_btree_sort_into(c, n, next);
1689
1690         bch2_btree_build_aux_trees(n);
1691         six_unlock_write(&n->c.lock);
1692
1693         bkey_init(&delete.k);
1694         delete.k.p = prev->key.k.p;
1695         bch2_keylist_add(&as->parent_keys, &delete);
1696         bch2_keylist_add(&as->parent_keys, &n->key);
1697
1698         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1699
1700         bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1701
1702         bch2_btree_update_get_open_buckets(as, n);
1703
1704         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1705         six_lock_increment(&m->c.lock, SIX_LOCK_intent);
1706         bch2_btree_iter_node_drop(iter, b);
1707         bch2_btree_iter_node_drop(iter, m);
1708
1709         bch2_btree_iter_node_replace(iter, n);
1710
1711         bch2_btree_trans_verify_iters(trans, n);
1712
1713         bch2_btree_node_free_inmem(c, b, iter);
1714         bch2_btree_node_free_inmem(c, m, iter);
1715
1716         six_unlock_intent(&n->c.lock);
1717
1718         bch2_btree_update_done(as);
1719 out:
1720         bch2_btree_trans_verify_locks(trans);
1721         bch2_trans_iter_free(trans, sib_iter);
1722
1723         /*
1724          * Don't downgrade locks here: we're called after successful insert,
1725          * and the caller will downgrade locks after a successful insert
1726          * anyways (in case e.g. a split was required first)
1727          *
1728          * And we're also called when inserting into interior nodes in the
1729          * split path, and downgrading to read locks in there is potentially
1730          * confusing:
1731          */
1732         return ret ?: ret2;
1733 err:
1734         bch2_trans_iter_put(trans, sib_iter);
1735         sib_iter = NULL;
1736
1737         if (ret == -EINTR && bch2_trans_relock(trans))
1738                 goto retry;
1739
1740         if (ret == -EINTR && !(flags & BTREE_INSERT_NOUNLOCK)) {
1741                 ret2 = ret;
1742                 ret = bch2_btree_iter_traverse_all(trans);
1743                 if (!ret)
1744                         goto retry;
1745         }
1746
1747         goto out;
1748 }
1749
1750 /**
1751  * bch_btree_node_rewrite - Rewrite/move a btree node
1752  */
1753 int bch2_btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
1754                             __le64 seq, unsigned flags)
1755 {
1756         struct btree *b, *n, *parent;
1757         struct btree_update *as;
1758         int ret;
1759
1760         flags |= BTREE_INSERT_NOFAIL;
1761 retry:
1762         ret = bch2_btree_iter_traverse(iter);
1763         if (ret)
1764                 goto out;
1765
1766         b = bch2_btree_iter_peek_node(iter);
1767         if (!b || b->data->keys.seq != seq)
1768                 goto out;
1769
1770         parent = btree_node_parent(iter, b);
1771         as = bch2_btree_update_start(iter, b->c.level,
1772                 (parent
1773                  ? btree_update_reserve_required(c, parent)
1774                  : 0) + 1,
1775                 flags);
1776         ret = PTR_ERR_OR_ZERO(as);
1777         if (ret == -EINTR)
1778                 goto retry;
1779         if (ret) {
1780                 trace_btree_gc_rewrite_node_fail(c, b);
1781                 goto out;
1782         }
1783
1784         bch2_btree_interior_update_will_free_node(as, b);
1785
1786         n = bch2_btree_node_alloc_replacement(as, b);
1787         bch2_btree_update_add_new_node(as, n);
1788
1789         bch2_btree_build_aux_trees(n);
1790         six_unlock_write(&n->c.lock);
1791
1792         trace_btree_gc_rewrite_node(c, b);
1793
1794         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1795
1796         if (parent) {
1797                 bch2_keylist_add(&as->parent_keys, &n->key);
1798                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1799         } else {
1800                 bch2_btree_set_root(as, n, iter);
1801         }
1802
1803         bch2_btree_update_get_open_buckets(as, n);
1804
1805         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1806         bch2_btree_iter_node_drop(iter, b);
1807         bch2_btree_iter_node_replace(iter, n);
1808         bch2_btree_node_free_inmem(c, b, iter);
1809         six_unlock_intent(&n->c.lock);
1810
1811         bch2_btree_update_done(as);
1812 out:
1813         bch2_btree_iter_downgrade(iter);
1814         return ret;
1815 }
1816
1817 struct async_btree_rewrite {
1818         struct bch_fs           *c;
1819         struct work_struct      work;
1820         enum btree_id           btree_id;
1821         unsigned                level;
1822         struct bpos             pos;
1823         __le64                  seq;
1824 };
1825
1826 void async_btree_node_rewrite_work(struct work_struct *work)
1827 {
1828         struct async_btree_rewrite *a =
1829                 container_of(work, struct async_btree_rewrite, work);
1830         struct bch_fs *c = a->c;
1831         struct btree_trans trans;
1832         struct btree_iter *iter;
1833
1834         bch2_trans_init(&trans, c, 0, 0);
1835         iter = bch2_trans_get_node_iter(&trans, a->btree_id, a->pos,
1836                                         BTREE_MAX_DEPTH, a->level, 0);
1837         bch2_btree_node_rewrite(c, iter, a->seq, 0);
1838         bch2_trans_iter_put(&trans, iter);
1839         bch2_trans_exit(&trans);
1840         percpu_ref_put(&c->writes);
1841         kfree(a);
1842 }
1843
1844 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
1845 {
1846         struct async_btree_rewrite *a;
1847
1848         if (!test_bit(BCH_FS_BTREE_INTERIOR_REPLAY_DONE, &c->flags))
1849                 return;
1850
1851         if (!percpu_ref_tryget(&c->writes))
1852                 return;
1853
1854         a = kmalloc(sizeof(*a), GFP_NOFS);
1855         if (!a) {
1856                 percpu_ref_put(&c->writes);
1857                 return;
1858         }
1859
1860         a->c            = c;
1861         a->btree_id     = b->c.btree_id;
1862         a->level        = b->c.level;
1863         a->pos          = b->key.k.p;
1864         a->seq          = b->data->keys.seq;
1865
1866         INIT_WORK(&a->work, async_btree_node_rewrite_work);
1867         queue_work(c->btree_interior_update_worker, &a->work);
1868 }
1869
1870 static void __bch2_btree_node_update_key(struct bch_fs *c,
1871                                          struct btree_update *as,
1872                                          struct btree_iter *iter,
1873                                          struct btree *b, struct btree *new_hash,
1874                                          struct bkey_i *new_key)
1875 {
1876         struct btree *parent;
1877         int ret;
1878
1879         btree_update_will_delete_key(as, &b->key);
1880         btree_update_will_add_key(as, new_key);
1881
1882         parent = btree_node_parent(iter, b);
1883         if (parent) {
1884                 if (new_hash) {
1885                         bkey_copy(&new_hash->key, new_key);
1886                         ret = bch2_btree_node_hash_insert(&c->btree_cache,
1887                                         new_hash, b->c.level, b->c.btree_id);
1888                         BUG_ON(ret);
1889                 }
1890
1891                 bch2_keylist_add(&as->parent_keys, new_key);
1892                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, 0);
1893
1894                 if (new_hash) {
1895                         mutex_lock(&c->btree_cache.lock);
1896                         bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
1897
1898                         bch2_btree_node_hash_remove(&c->btree_cache, b);
1899
1900                         bkey_copy(&b->key, new_key);
1901                         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1902                         BUG_ON(ret);
1903                         mutex_unlock(&c->btree_cache.lock);
1904                 } else {
1905                         bkey_copy(&b->key, new_key);
1906                 }
1907         } else {
1908                 BUG_ON(btree_node_root(c, b) != b);
1909
1910                 bch2_btree_node_lock_write(b, iter);
1911                 bkey_copy(&b->key, new_key);
1912
1913                 if (btree_ptr_hash_val(&b->key) != b->hash_val) {
1914                         mutex_lock(&c->btree_cache.lock);
1915                         bch2_btree_node_hash_remove(&c->btree_cache, b);
1916
1917                         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1918                         BUG_ON(ret);
1919                         mutex_unlock(&c->btree_cache.lock);
1920                 }
1921
1922                 btree_update_updated_root(as, b);
1923                 bch2_btree_node_unlock_write(b, iter);
1924         }
1925
1926         bch2_btree_update_done(as);
1927 }
1928
1929 int bch2_btree_node_update_key(struct bch_fs *c, struct btree_iter *iter,
1930                                struct btree *b,
1931                                struct bkey_i *new_key)
1932 {
1933         struct btree *parent = btree_node_parent(iter, b);
1934         struct btree_update *as = NULL;
1935         struct btree *new_hash = NULL;
1936         struct closure cl;
1937         int ret = 0;
1938
1939         closure_init_stack(&cl);
1940
1941         /*
1942          * check btree_ptr_hash_val() after @b is locked by
1943          * btree_iter_traverse():
1944          */
1945         if (btree_ptr_hash_val(new_key) != b->hash_val) {
1946                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1947                 if (ret) {
1948                         bch2_trans_unlock(iter->trans);
1949                         closure_sync(&cl);
1950                         if (!bch2_trans_relock(iter->trans))
1951                                 return -EINTR;
1952                 }
1953
1954                 new_hash = bch2_btree_node_mem_alloc(c);
1955         }
1956
1957         as = bch2_btree_update_start(iter, b->c.level,
1958                 parent ? btree_update_reserve_required(c, parent) : 0,
1959                 BTREE_INSERT_NOFAIL);
1960         if (IS_ERR(as)) {
1961                 ret = PTR_ERR(as);
1962                 goto err;
1963         }
1964
1965         __bch2_btree_node_update_key(c, as, iter, b, new_hash, new_key);
1966
1967         bch2_btree_iter_downgrade(iter);
1968 err:
1969         if (new_hash) {
1970                 mutex_lock(&c->btree_cache.lock);
1971                 list_move(&new_hash->list, &c->btree_cache.freeable);
1972                 mutex_unlock(&c->btree_cache.lock);
1973
1974                 six_unlock_write(&new_hash->c.lock);
1975                 six_unlock_intent(&new_hash->c.lock);
1976         }
1977         closure_sync(&cl);
1978         bch2_btree_cache_cannibalize_unlock(c);
1979         return ret;
1980 }
1981
1982 /* Init code: */
1983
1984 /*
1985  * Only for filesystem bringup, when first reading the btree roots or allocating
1986  * btree roots when initializing a new filesystem:
1987  */
1988 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
1989 {
1990         BUG_ON(btree_node_root(c, b));
1991
1992         bch2_btree_set_root_inmem(c, b);
1993 }
1994
1995 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
1996 {
1997         struct closure cl;
1998         struct btree *b;
1999         int ret;
2000
2001         closure_init_stack(&cl);
2002
2003         do {
2004                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2005                 closure_sync(&cl);
2006         } while (ret);
2007
2008         b = bch2_btree_node_mem_alloc(c);
2009         bch2_btree_cache_cannibalize_unlock(c);
2010
2011         set_btree_node_fake(b);
2012         set_btree_node_need_rewrite(b);
2013         b->c.level      = 0;
2014         b->c.btree_id   = id;
2015
2016         bkey_btree_ptr_init(&b->key);
2017         b->key.k.p = POS_MAX;
2018         *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2019
2020         bch2_bset_init_first(b, &b->data->keys);
2021         bch2_btree_build_aux_trees(b);
2022
2023         b->data->flags = 0;
2024         btree_set_min(b, POS_MIN);
2025         btree_set_max(b, POS_MAX);
2026         b->data->format = bch2_btree_calc_format(b);
2027         btree_node_set_format(b, b->data->format);
2028
2029         ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2030                                           b->c.level, b->c.btree_id);
2031         BUG_ON(ret);
2032
2033         bch2_btree_set_root_inmem(c, b);
2034
2035         six_unlock_write(&b->c.lock);
2036         six_unlock_intent(&b->c.lock);
2037 }
2038
2039 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2040 {
2041         struct btree_update *as;
2042
2043         mutex_lock(&c->btree_interior_update_lock);
2044         list_for_each_entry(as, &c->btree_interior_update_list, list)
2045                 pr_buf(out, "%p m %u w %u r %u j %llu\n",
2046                        as,
2047                        as->mode,
2048                        as->nodes_written,
2049                        atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2050                        as->journal.seq);
2051         mutex_unlock(&c->btree_interior_update_lock);
2052 }
2053
2054 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *c)
2055 {
2056         size_t ret = 0;
2057         struct list_head *i;
2058
2059         mutex_lock(&c->btree_interior_update_lock);
2060         list_for_each(i, &c->btree_interior_update_list)
2061                 ret++;
2062         mutex_unlock(&c->btree_interior_update_lock);
2063
2064         return ret;
2065 }
2066
2067 void bch2_journal_entries_to_btree_roots(struct bch_fs *c, struct jset *jset)
2068 {
2069         struct btree_root *r;
2070         struct jset_entry *entry;
2071
2072         mutex_lock(&c->btree_root_lock);
2073
2074         vstruct_for_each(jset, entry)
2075                 if (entry->type == BCH_JSET_ENTRY_btree_root) {
2076                         r = &c->btree_roots[entry->btree_id];
2077                         r->level = entry->level;
2078                         r->alive = true;
2079                         bkey_copy(&r->key, &entry->start[0]);
2080                 }
2081
2082         mutex_unlock(&c->btree_root_lock);
2083 }
2084
2085 struct jset_entry *
2086 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2087                                     struct jset_entry *start,
2088                                     struct jset_entry *end)
2089 {
2090         struct jset_entry *entry;
2091         unsigned long have = 0;
2092         unsigned i;
2093
2094         for (entry = start; entry < end; entry = vstruct_next(entry))
2095                 if (entry->type == BCH_JSET_ENTRY_btree_root)
2096                         __set_bit(entry->btree_id, &have);
2097
2098         mutex_lock(&c->btree_root_lock);
2099
2100         for (i = 0; i < BTREE_ID_NR; i++)
2101                 if (c->btree_roots[i].alive && !test_bit(i, &have)) {
2102                         journal_entry_set(end,
2103                                           BCH_JSET_ENTRY_btree_root,
2104                                           i, c->btree_roots[i].level,
2105                                           &c->btree_roots[i].key,
2106                                           c->btree_roots[i].key.u64s);
2107                         end = vstruct_next(end);
2108                 }
2109
2110         mutex_unlock(&c->btree_root_lock);
2111
2112         return end;
2113 }
2114
2115 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2116 {
2117         if (c->btree_interior_update_worker)
2118                 destroy_workqueue(c->btree_interior_update_worker);
2119         mempool_exit(&c->btree_interior_update_pool);
2120 }
2121
2122 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2123 {
2124         mutex_init(&c->btree_reserve_cache_lock);
2125         INIT_LIST_HEAD(&c->btree_interior_update_list);
2126         INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2127         mutex_init(&c->btree_interior_update_lock);
2128         INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2129
2130         c->btree_interior_update_worker =
2131                 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2132         if (!c->btree_interior_update_worker)
2133                 return -ENOMEM;
2134
2135         return mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2136                                          sizeof(struct btree_update));
2137 }