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         mutex_lock(&c->btree_reserve_cache_lock);
 385
 386         for (p = as->prealloc_nodes;
 387              p < as->prealloc_nodes + ARRAY_SIZE(as->prealloc_nodes);
 388              p++) {
 389                 while (p->nr) {
 390                         struct btree *b = p->b[--p->nr];
 391
 392                         six_lock_intent(&b->c.lock, NULL, NULL);
 393                         six_lock_write(&b->c.lock, NULL, NULL);
 394
 395                         if (c->btree_reserve_cache_nr <
 396                             ARRAY_SIZE(c->btree_reserve_cache)) {
 397                                 struct btree_alloc *a =
 398                                         &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
 399
 400                                 a->ob = b->ob;
 401                                 b->ob.nr = 0;
 402                                 bkey_copy(&a->k, &b->key);
 403                         } else {
 404                                 bch2_open_buckets_put(c, &b->ob);
 405                         }
 406
 407                         __btree_node_free(c, b);
 408                         six_unlock_write(&b->c.lock);
 409                         six_unlock_intent(&b->c.lock);
 410                 }
 411         }
 412
 413         mutex_unlock(&c->btree_reserve_cache_lock);
 414 }
 415
 416 static int bch2_btree_reserve_get(struct btree_update *as,
 417                                   unsigned nr_nodes[2],
 418                                   unsigned flags)
 419 {
 420         struct bch_fs *c = as->c;
 421         struct closure cl;
 422         struct btree *b;
 423         unsigned interior;
 424         int ret;
 425
 426         closure_init_stack(&cl);
 427 retry:
 428
 429         BUG_ON(nr_nodes[0] + nr_nodes[1] > BTREE_RESERVE_MAX);
 430
 431         /*
 432          * Protects reaping from the btree node cache and using the btree node
 433          * open bucket reserve:
 434          *
 435          * BTREE_INSERT_NOWAIT only applies to btree node allocation, not
 436          * blocking on this lock:
 437          */
 438         ret = bch2_btree_cache_cannibalize_lock(c, &cl);
 439         if (ret)
 440                 goto err;
 441
 442         for (interior = 0; interior < 2; interior++) {
 443                 struct prealloc_nodes *p = as->prealloc_nodes + interior;
 444
 445                 while (p->nr < nr_nodes[interior]) {
 446                         b = __bch2_btree_node_alloc(c, &as->disk_res,
 447                                                     flags & BTREE_INSERT_NOWAIT
 448                                                     ? NULL : &cl,
 449                                                     interior, flags);
 450                         if (IS_ERR(b)) {
 451                                 ret = PTR_ERR(b);
 452                                 goto err;
 453                         }
 454
 455                         p->b[p->nr++] = b;
 456                 }
 457         }
 458
 459         bch2_btree_cache_cannibalize_unlock(c);
 460         closure_sync(&cl);
 461         return 0;
 462 err:
 463         bch2_btree_cache_cannibalize_unlock(c);
 464         closure_sync(&cl);
 465
 466         if (ret == -EAGAIN)
 467                 goto retry;
 468
 469         trace_btree_reserve_get_fail(c, nr_nodes[0] + nr_nodes[1], &cl);
 470         return ret;
 471 }
 472
 473 /* Asynchronous interior node update machinery */
 474
 475 static void bch2_btree_update_free(struct btree_update *as)
 476 {
 477         struct bch_fs *c = as->c;
 478
 479         if (as->took_gc_lock)
 480                 up_read(&c->gc_lock);
 481         as->took_gc_lock = false;
 482
 483         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 484
 485         bch2_journal_pin_drop(&c->journal, &as->journal);
 486         bch2_journal_pin_flush(&c->journal, &as->journal);
 487         bch2_disk_reservation_put(c, &as->disk_res);
 488         bch2_btree_reserve_put(as);
 489
 490         bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_total],
 491                                as->start_time);
 492
 493         mutex_lock(&c->btree_interior_update_lock);
 494         list_del(&as->unwritten_list);
 495         list_del(&as->list);
 496
 497         closure_debug_destroy(&as->cl);
 498         mempool_free(as, &c->btree_interior_update_pool);
 499
 500         /*
 501          * Have to do the wakeup with btree_interior_update_lock still held,
 502          * since being on btree_interior_update_list is our ref on @c:
 503          */
 504         closure_wake_up(&c->btree_interior_update_wait);
 505
 506         mutex_unlock(&c->btree_interior_update_lock);
 507 }
 508
 509 static void btree_update_will_delete_key(struct btree_update *as,
 510                                          struct bkey_i *k)
 511 {
 512         BUG_ON(bch2_keylist_u64s(&as->old_keys) + k->k.u64s >
 513                ARRAY_SIZE(as->_old_keys));
 514         bch2_keylist_add(&as->old_keys, k);
 515 }
 516
 517 static void btree_update_will_add_key(struct btree_update *as,
 518                                       struct bkey_i *k)
 519 {
 520         BUG_ON(bch2_keylist_u64s(&as->new_keys) + k->k.u64s >
 521                ARRAY_SIZE(as->_new_keys));
 522         bch2_keylist_add(&as->new_keys, k);
 523 }
 524
 525 /*
 526  * The transactional part of an interior btree node update, where we journal the
 527  * update we did to the interior node and update alloc info:
 528  */
 529 static int btree_update_nodes_written_trans(struct btree_trans *trans,
 530                                             struct btree_update *as)
 531 {
 532         struct bkey_i *k;
 533         int ret;
 534
 535         trans->extra_journal_entries = (void *) &as->journal_entries[0];
 536         trans->extra_journal_entry_u64s = as->journal_u64s;
 537         trans->journal_pin = &as->journal;
 538
 539         for_each_keylist_key(&as->new_keys, k) {
 540                 ret = bch2_trans_mark_new(trans, k, 0);
 541                 if (ret)
 542                         return ret;
 543         }
 544
 545         for_each_keylist_key(&as->old_keys, k) {
 546                 ret = bch2_trans_mark_old(trans, bkey_i_to_s_c(k), 0);
 547                 if (ret)
 548                         return ret;
 549         }
 550
 551         return 0;
 552 }
 553
 554 static void btree_update_nodes_written(struct btree_update *as)
 555 {
 556         struct bch_fs *c = as->c;
 557         struct btree *b = as->b;
 558         struct btree_trans trans;
 559         u64 journal_seq = 0;
 560         unsigned i;
 561         int ret;
 562
 563         /*
 564          * If we're already in an error state, it might be because a btree node
 565          * was never written, and we might be trying to free that same btree
 566          * node here, but it won't have been marked as allocated and we'll see
 567          * spurious disk usage inconsistencies in the transactional part below
 568          * if we don't skip it:
 569          */
 570         ret = bch2_journal_error(&c->journal);
 571         if (ret)
 572                 goto err;
 573
 574         /*
 575          * Wait for any in flight writes to finish before we free the old nodes
 576          * on disk:
 577          */
 578         for (i = 0; i < as->nr_old_nodes; i++) {
 579                 struct btree *old = as->old_nodes[i];
 580                 __le64 seq;
 581
 582                 six_lock_read(&old->c.lock, NULL, NULL);
 583                 seq = old->data ? old->data->keys.seq : 0;
 584                 six_unlock_read(&old->c.lock);
 585
 586                 if (seq == as->old_nodes_seq[i])
 587                         wait_on_bit_io(&old->flags, BTREE_NODE_write_in_flight_inner,
 588                                        TASK_UNINTERRUPTIBLE);
 589         }
 590
 591         /*
 592          * We did an update to a parent node where the pointers we added pointed
 593          * to child nodes that weren't written yet: now, the child nodes have
 594          * been written so we can write out the update to the interior node.
 595          */
 596
 597         /*
 598          * We can't call into journal reclaim here: we'd block on the journal
 599          * reclaim lock, but we may need to release the open buckets we have
 600          * pinned in order for other btree updates to make forward progress, and
 601          * journal reclaim does btree updates when flushing bkey_cached entries,
 602          * which may require allocations as well.
 603          */
 604         bch2_trans_init(&trans, c, 0, 512);
 605         ret = __bch2_trans_do(&trans, &as->disk_res, &journal_seq,
 606                               BTREE_INSERT_NOFAIL|
 607                               BTREE_INSERT_NOCHECK_RW|
 608                               BTREE_INSERT_JOURNAL_RECLAIM|
 609                               JOURNAL_WATERMARK_reserved,
 610                               btree_update_nodes_written_trans(&trans, as));
 611         bch2_trans_exit(&trans);
 612
 613         bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
 614                              "error %i in btree_update_nodes_written()", ret);
 615 err:
 616         if (b) {
 617                 /*
 618                  * @b is the node we did the final insert into:
 619                  *
 620                  * On failure to get a journal reservation, we still have to
 621                  * unblock the write and allow most of the write path to happen
 622                  * so that shutdown works, but the i->journal_seq mechanism
 623                  * won't work to prevent the btree write from being visible (we
 624                  * didn't get a journal sequence number) - instead
 625                  * __bch2_btree_node_write() doesn't do the actual write if
 626                  * we're in journal error state:
 627                  */
 628
 629                 six_lock_intent(&b->c.lock, NULL, NULL);
 630                 six_lock_write(&b->c.lock, NULL, NULL);
 631                 mutex_lock(&c->btree_interior_update_lock);
 632
 633                 list_del(&as->write_blocked_list);
 634                 if (list_empty(&b->write_blocked))
 635                         clear_btree_node_write_blocked(b);
 636
 637                 /*
 638                  * Node might have been freed, recheck under
 639                  * btree_interior_update_lock:
 640                  */
 641                 if (as->b == b) {
 642                         struct bset *i = btree_bset_last(b);
 643
 644                         BUG_ON(!b->c.level);
 645                         BUG_ON(!btree_node_dirty(b));
 646
 647                         if (!ret) {
 648                                 i->journal_seq = cpu_to_le64(
 649                                         max(journal_seq,
 650                                             le64_to_cpu(i->journal_seq)));
 651
 652                                 bch2_btree_add_journal_pin(c, b, journal_seq);
 653                         } else {
 654                                 /*
 655                                  * If we didn't get a journal sequence number we
 656                                  * can't write this btree node, because recovery
 657                                  * won't know to ignore this write:
 658                                  */
 659                                 set_btree_node_never_write(b);
 660                         }
 661                 }
 662
 663                 mutex_unlock(&c->btree_interior_update_lock);
 664                 six_unlock_write(&b->c.lock);
 665
 666                 btree_node_write_if_need(c, b, SIX_LOCK_intent);
 667                 six_unlock_intent(&b->c.lock);
 668         }
 669
 670         bch2_journal_pin_drop(&c->journal, &as->journal);
 671
 672         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 673
 674         mutex_lock(&c->btree_interior_update_lock);
 675         for (i = 0; i < as->nr_new_nodes; i++) {
 676                 b = as->new_nodes[i];
 677
 678                 BUG_ON(b->will_make_reachable != (unsigned long) as);
 679                 b->will_make_reachable = 0;
 680                 clear_btree_node_will_make_reachable(b);
 681         }
 682         mutex_unlock(&c->btree_interior_update_lock);
 683
 684         for (i = 0; i < as->nr_new_nodes; i++) {
 685                 b = as->new_nodes[i];
 686
 687                 six_lock_read(&b->c.lock, NULL, NULL);
 688                 btree_node_write_if_need(c, b, SIX_LOCK_read);
 689                 six_unlock_read(&b->c.lock);
 690         }
 691
 692         for (i = 0; i < as->nr_open_buckets; i++)
 693                 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
 694
 695         bch2_btree_update_free(as);
 696 }
 697
 698 static void btree_interior_update_work(struct work_struct *work)
 699 {
 700         struct bch_fs *c =
 701                 container_of(work, struct bch_fs, btree_interior_update_work);
 702         struct btree_update *as;
 703
 704         while (1) {
 705                 mutex_lock(&c->btree_interior_update_lock);
 706                 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
 707                                               struct btree_update, unwritten_list);
 708                 if (as && !as->nodes_written)
 709                         as = NULL;
 710                 mutex_unlock(&c->btree_interior_update_lock);
 711
 712                 if (!as)
 713                         break;
 714
 715                 btree_update_nodes_written(as);
 716         }
 717 }
 718
 719 static void btree_update_set_nodes_written(struct closure *cl)
 720 {
 721         struct btree_update *as = container_of(cl, struct btree_update, cl);
 722         struct bch_fs *c = as->c;
 723
 724         mutex_lock(&c->btree_interior_update_lock);
 725         as->nodes_written = true;
 726         mutex_unlock(&c->btree_interior_update_lock);
 727
 728         queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
 729 }
 730
 731 /*
 732  * We're updating @b with pointers to nodes that haven't finished writing yet:
 733  * block @b from being written until @as completes
 734  */
 735 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
 736 {
 737         struct bch_fs *c = as->c;
 738
 739         mutex_lock(&c->btree_interior_update_lock);
 740         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 741
 742         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 743         BUG_ON(!btree_node_dirty(b));
 744
 745         as->mode        = BTREE_INTERIOR_UPDATING_NODE;
 746         as->b           = b;
 747
 748         set_btree_node_write_blocked(b);
 749         list_add(&as->write_blocked_list, &b->write_blocked);
 750
 751         mutex_unlock(&c->btree_interior_update_lock);
 752 }
 753
 754 static void btree_update_reparent(struct btree_update *as,
 755                                   struct btree_update *child)
 756 {
 757         struct bch_fs *c = as->c;
 758
 759         lockdep_assert_held(&c->btree_interior_update_lock);
 760
 761         child->b = NULL;
 762         child->mode = BTREE_INTERIOR_UPDATING_AS;
 763
 764         bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL);
 765 }
 766
 767 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
 768 {
 769         struct bkey_i *insert = &b->key;
 770         struct bch_fs *c = as->c;
 771
 772         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 773
 774         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
 775                ARRAY_SIZE(as->journal_entries));
 776
 777         as->journal_u64s +=
 778                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
 779                                   BCH_JSET_ENTRY_btree_root,
 780                                   b->c.btree_id, b->c.level,
 781                                   insert, insert->k.u64s);
 782
 783         mutex_lock(&c->btree_interior_update_lock);
 784         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 785
 786         as->mode        = BTREE_INTERIOR_UPDATING_ROOT;
 787         mutex_unlock(&c->btree_interior_update_lock);
 788 }
 789
 790 /*
 791  * bch2_btree_update_add_new_node:
 792  *
 793  * This causes @as to wait on @b to be written, before it gets to
 794  * bch2_btree_update_nodes_written
 795  *
 796  * Additionally, it sets b->will_make_reachable to prevent any additional writes
 797  * to @b from happening besides the first until @b is reachable on disk
 798  *
 799  * And it adds @b to the list of @as's new nodes, so that we can update sector
 800  * counts in bch2_btree_update_nodes_written:
 801  */
 802 static void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
 803 {
 804         struct bch_fs *c = as->c;
 805
 806         closure_get(&as->cl);
 807
 808         mutex_lock(&c->btree_interior_update_lock);
 809         BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
 810         BUG_ON(b->will_make_reachable);
 811
 812         as->new_nodes[as->nr_new_nodes++] = b;
 813         b->will_make_reachable = 1UL|(unsigned long) as;
 814         set_btree_node_will_make_reachable(b);
 815
 816         mutex_unlock(&c->btree_interior_update_lock);
 817
 818         btree_update_will_add_key(as, &b->key);
 819 }
 820
 821 /*
 822  * returns true if @b was a new node
 823  */
 824 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
 825 {
 826         struct btree_update *as;
 827         unsigned long v;
 828         unsigned i;
 829
 830         mutex_lock(&c->btree_interior_update_lock);
 831         /*
 832          * When b->will_make_reachable != 0, it owns a ref on as->cl that's
 833          * dropped when it gets written by bch2_btree_complete_write - the
 834          * xchg() is for synchronization with bch2_btree_complete_write:
 835          */
 836         v = xchg(&b->will_make_reachable, 0);
 837         clear_btree_node_will_make_reachable(b);
 838         as = (struct btree_update *) (v & ~1UL);
 839
 840         if (!as) {
 841                 mutex_unlock(&c->btree_interior_update_lock);
 842                 return;
 843         }
 844
 845         for (i = 0; i < as->nr_new_nodes; i++)
 846                 if (as->new_nodes[i] == b)
 847                         goto found;
 848
 849         BUG();
 850 found:
 851         array_remove_item(as->new_nodes, as->nr_new_nodes, i);
 852         mutex_unlock(&c->btree_interior_update_lock);
 853
 854         if (v & 1)
 855                 closure_put(&as->cl);
 856 }
 857
 858 static void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
 859 {
 860         while (b->ob.nr)
 861                 as->open_buckets[as->nr_open_buckets++] =
 862                         b->ob.v[--b->ob.nr];
 863 }
 864
 865 /*
 866  * @b is being split/rewritten: it may have pointers to not-yet-written btree
 867  * nodes and thus outstanding btree_updates - redirect @b's
 868  * btree_updates to point to this btree_update:
 869  */
 870 static void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 871                                                struct btree *b)
 872 {
 873         struct bch_fs *c = as->c;
 874         struct btree_update *p, *n;
 875         struct btree_write *w;
 876
 877         set_btree_node_dying(b);
 878
 879         if (btree_node_fake(b))
 880                 return;
 881
 882         mutex_lock(&c->btree_interior_update_lock);
 883
 884         /*
 885          * Does this node have any btree_update operations preventing
 886          * it from being written?
 887          *
 888          * If so, redirect them to point to this btree_update: we can
 889          * write out our new nodes, but we won't make them visible until those
 890          * operations complete
 891          */
 892         list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
 893                 list_del_init(&p->write_blocked_list);
 894                 btree_update_reparent(as, p);
 895
 896                 /*
 897                  * for flush_held_btree_writes() waiting on updates to flush or
 898                  * nodes to be writeable:
 899                  */
 900                 closure_wake_up(&c->btree_interior_update_wait);
 901         }
 902
 903         clear_btree_node_dirty_acct(c, b);
 904         clear_btree_node_need_write(b);
 905
 906         /*
 907          * Does this node have unwritten data that has a pin on the journal?
 908          *
 909          * If so, transfer that pin to the btree_update operation -
 910          * note that if we're freeing multiple nodes, we only need to keep the
 911          * oldest pin of any of the nodes we're freeing. We'll release the pin
 912          * when the new nodes are persistent and reachable on disk:
 913          */
 914         w = btree_current_write(b);
 915         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 916         bch2_journal_pin_drop(&c->journal, &w->journal);
 917
 918         w = btree_prev_write(b);
 919         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 920         bch2_journal_pin_drop(&c->journal, &w->journal);
 921
 922         mutex_unlock(&c->btree_interior_update_lock);
 923
 924         /*
 925          * Is this a node that isn't reachable on disk yet?
 926          *
 927          * Nodes that aren't reachable yet have writes blocked until they're
 928          * reachable - now that we've cancelled any pending writes and moved
 929          * things waiting on that write to wait on this update, we can drop this
 930          * node from the list of nodes that the other update is making
 931          * reachable, prior to freeing it:
 932          */
 933         btree_update_drop_new_node(c, b);
 934
 935         btree_update_will_delete_key(as, &b->key);
 936
 937         as->old_nodes[as->nr_old_nodes] = b;
 938         as->old_nodes_seq[as->nr_old_nodes] = b->data->keys.seq;
 939         as->nr_old_nodes++;
 940 }
 941
 942 static void bch2_btree_update_done(struct btree_update *as)
 943 {
 944         struct bch_fs *c = as->c;
 945         u64 start_time = as->start_time;
 946
 947         BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
 948
 949         if (as->took_gc_lock)
 950                 up_read(&as->c->gc_lock);
 951         as->took_gc_lock = false;
 952
 953         bch2_btree_reserve_put(as);
 954
 955         continue_at(&as->cl, btree_update_set_nodes_written,
 956                     as->c->btree_interior_update_worker);
 957
 958         bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_foreground],
 959                                start_time);
 960 }
 961
 962 static struct btree_update *
 963 bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
 964                         unsigned level, bool split, unsigned flags)
 965 {
 966         struct bch_fs *c = trans->c;
 967         struct btree_update *as;
 968         u64 start_time = local_clock();
 969         int disk_res_flags = (flags & BTREE_INSERT_NOFAIL)
 970                 ? BCH_DISK_RESERVATION_NOFAIL : 0;
 971         unsigned nr_nodes[2] = { 0, 0 };
 972         unsigned update_level = level;
 973         int journal_flags = flags & JOURNAL_WATERMARK_MASK;
 974         int ret = 0;
 975
 976         BUG_ON(!path->should_be_locked);
 977
 978         if (flags & BTREE_INSERT_JOURNAL_RECLAIM)
 979                 journal_flags |= JOURNAL_RES_GET_NONBLOCK;
 980
 981         while (1) {
 982                 nr_nodes[!!update_level] += 1 + split;
 983                 update_level++;
 984
 985                 if (!btree_path_node(path, update_level))
 986                         break;
 987
 988                 /*
 989                  * XXX: figure out how far we might need to split,
 990                  * instead of locking/reserving all the way to the root:
 991                  */
 992                 split = update_level + 1 < BTREE_MAX_DEPTH;
 993         }
 994
 995         /* Might have to allocate a new root: */
 996         if (update_level < BTREE_MAX_DEPTH)
 997                 nr_nodes[1] += 1;
 998
 999         if (!bch2_btree_path_upgrade(trans, path, U8_MAX)) {
1000                 trace_trans_restart_iter_upgrade(trans->fn, _RET_IP_,
1001                                                  path->btree_id, &path->pos);
1002                 ret = btree_trans_restart(trans);
1003                 return ERR_PTR(ret);
1004         }
1005
1006         if (flags & BTREE_INSERT_GC_LOCK_HELD)
1007                 lockdep_assert_held(&c->gc_lock);
1008         else if (!down_read_trylock(&c->gc_lock)) {
1009                 bch2_trans_unlock(trans);
1010                 down_read(&c->gc_lock);
1011                 if (!bch2_trans_relock(trans)) {
1012                         up_read(&c->gc_lock);
1013                         return ERR_PTR(-EINTR);
1014                 }
1015         }
1016
1017         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
1018         memset(as, 0, sizeof(*as));
1019         closure_init(&as->cl, NULL);
1020         as->c           = c;
1021         as->start_time  = start_time;
1022         as->mode        = BTREE_INTERIOR_NO_UPDATE;
1023         as->took_gc_lock = !(flags & BTREE_INSERT_GC_LOCK_HELD);
1024         as->btree_id    = path->btree_id;
1025         INIT_LIST_HEAD(&as->list);
1026         INIT_LIST_HEAD(&as->unwritten_list);
1027         INIT_LIST_HEAD(&as->write_blocked_list);
1028         bch2_keylist_init(&as->old_keys, as->_old_keys);
1029         bch2_keylist_init(&as->new_keys, as->_new_keys);
1030         bch2_keylist_init(&as->parent_keys, as->inline_keys);
1031
1032         mutex_lock(&c->btree_interior_update_lock);
1033         list_add_tail(&as->list, &c->btree_interior_update_list);
1034         mutex_unlock(&c->btree_interior_update_lock);
1035
1036         /*
1037          * We don't want to allocate if we're in an error state, that can cause
1038          * deadlock on emergency shutdown due to open buckets getting stuck in
1039          * the btree_reserve_cache after allocator shutdown has cleared it out.
1040          * This check needs to come after adding us to the btree_interior_update
1041          * list but before calling bch2_btree_reserve_get, to synchronize with
1042          * __bch2_fs_read_only().
1043          */
1044         ret = bch2_journal_error(&c->journal);
1045         if (ret)
1046                 goto err;
1047
1048         bch2_trans_unlock(trans);
1049
1050         ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
1051                                       BTREE_UPDATE_JOURNAL_RES,
1052                                       journal_flags);
1053         if (ret) {
1054                 bch2_btree_update_free(as);
1055                 trace_trans_restart_journal_preres_get(trans->fn, _RET_IP_);
1056                 btree_trans_restart(trans);
1057                 return ERR_PTR(ret);
1058         }
1059
1060         ret = bch2_disk_reservation_get(c, &as->disk_res,
1061                         (nr_nodes[0] + nr_nodes[1]) * btree_sectors(c),
1062                         c->opts.metadata_replicas,
1063                         disk_res_flags);
1064         if (ret)
1065                 goto err;
1066
1067         ret = bch2_btree_reserve_get(as, nr_nodes, flags);
1068         if (ret)
1069                 goto err;
1070
1071         if (!bch2_trans_relock(trans)) {
1072                 ret = -EINTR;
1073                 goto err;
1074         }
1075
1076         return as;
1077 err:
1078         bch2_btree_update_free(as);
1079         return ERR_PTR(ret);
1080 }
1081
1082 /* Btree root updates: */
1083
1084 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1085 {
1086         /* Root nodes cannot be reaped */
1087         mutex_lock(&c->btree_cache.lock);
1088         list_del_init(&b->list);
1089         mutex_unlock(&c->btree_cache.lock);
1090
1091         if (b->c.level)
1092                 six_lock_pcpu_alloc(&b->c.lock);
1093         else
1094                 six_lock_pcpu_free(&b->c.lock);
1095
1096         mutex_lock(&c->btree_root_lock);
1097         BUG_ON(btree_node_root(c, b) &&
1098                (b->c.level < btree_node_root(c, b)->c.level ||
1099                 !btree_node_dying(btree_node_root(c, b))));
1100
1101         btree_node_root(c, b) = b;
1102         mutex_unlock(&c->btree_root_lock);
1103
1104         bch2_recalc_btree_reserve(c);
1105 }
1106
1107 /**
1108  * bch_btree_set_root - update the root in memory and on disk
1109  *
1110  * To ensure forward progress, the current task must not be holding any
1111  * btree node write locks. However, you must hold an intent lock on the
1112  * old root.
1113  *
1114  * Note: This allocates a journal entry but doesn't add any keys to
1115  * it.  All the btree roots are part of every journal write, so there
1116  * is nothing new to be done.  This just guarantees that there is a
1117  * journal write.
1118  */
1119 static void bch2_btree_set_root(struct btree_update *as,
1120                                 struct btree_trans *trans,
1121                                 struct btree_path *path,
1122                                 struct btree *b)
1123 {
1124         struct bch_fs *c = as->c;
1125         struct btree *old;
1126
1127         trace_btree_set_root(c, b);
1128         BUG_ON(!b->written);
1129
1130         old = btree_node_root(c, b);
1131
1132         /*
1133          * Ensure no one is using the old root while we switch to the
1134          * new root:
1135          */
1136         bch2_btree_node_lock_write(trans, path, old);
1137
1138         bch2_btree_set_root_inmem(c, b);
1139
1140         btree_update_updated_root(as, b);
1141
1142         /*
1143          * Unlock old root after new root is visible:
1144          *
1145          * The new root isn't persistent, but that's ok: we still have
1146          * an intent lock on the new root, and any updates that would
1147          * depend on the new root would have to update the new root.
1148          */
1149         bch2_btree_node_unlock_write(trans, path, old);
1150 }
1151
1152 /* Interior node updates: */
1153
1154 static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
1155                                         struct btree_trans *trans,
1156                                         struct btree_path *path,
1157                                         struct btree *b,
1158                                         struct btree_node_iter *node_iter,
1159                                         struct bkey_i *insert)
1160 {
1161         struct bch_fs *c = as->c;
1162         struct bkey_packed *k;
1163         const char *invalid;
1164
1165         BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 &&
1166                !btree_ptr_sectors_written(insert));
1167
1168         if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
1169                 bch2_journal_key_overwritten(c, b->c.btree_id, b->c.level, insert->k.p);
1170
1171         invalid = bch2_bkey_invalid(c, bkey_i_to_s_c(insert), btree_node_type(b)) ?:
1172                 bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert));
1173         if (invalid) {
1174                 struct printbuf buf = PRINTBUF;
1175
1176                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
1177                 bch2_fs_inconsistent(c, "inserting invalid bkey %s: %s", buf.buf, invalid);
1178                 printbuf_exit(&buf);
1179                 dump_stack();
1180         }
1181
1182         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1183                ARRAY_SIZE(as->journal_entries));
1184
1185         as->journal_u64s +=
1186                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1187                                   BCH_JSET_ENTRY_btree_keys,
1188                                   b->c.btree_id, b->c.level,
1189                                   insert, insert->k.u64s);
1190
1191         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1192                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1193                 bch2_btree_node_iter_advance(node_iter, b);
1194
1195         bch2_btree_bset_insert_key(trans, path, b, node_iter, insert);
1196         set_btree_node_dirty_acct(c, b);
1197         set_btree_node_need_write(b);
1198 }
1199
1200 static void
1201 __bch2_btree_insert_keys_interior(struct btree_update *as,
1202                                   struct btree_trans *trans,
1203                                   struct btree_path *path,
1204                                   struct btree *b,
1205                                   struct btree_node_iter node_iter,
1206                                   struct keylist *keys)
1207 {
1208         struct bkey_i *insert = bch2_keylist_front(keys);
1209         struct bkey_packed *k;
1210
1211         BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1212
1213         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1214                (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1215                 ;
1216
1217         while (!bch2_keylist_empty(keys)) {
1218                 bch2_insert_fixup_btree_ptr(as, trans, path, b,
1219                                 &node_iter, bch2_keylist_front(keys));
1220                 bch2_keylist_pop_front(keys);
1221         }
1222 }
1223
1224 /*
1225  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1226  * node)
1227  */
1228 static struct btree *__btree_split_node(struct btree_update *as,
1229                                         struct btree *n1)
1230 {
1231         struct bkey_format_state s;
1232         size_t nr_packed = 0, nr_unpacked = 0;
1233         struct btree *n2;
1234         struct bset *set1, *set2;
1235         struct bkey_packed *k, *set2_start, *set2_end, *out, *prev = NULL;
1236         struct bpos n1_pos;
1237
1238         n2 = bch2_btree_node_alloc(as, n1->c.level);
1239         bch2_btree_update_add_new_node(as, n2);
1240
1241         n2->data->max_key       = n1->data->max_key;
1242         n2->data->format        = n1->format;
1243         SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data));
1244         n2->key.k.p = n1->key.k.p;
1245
1246         set1 = btree_bset_first(n1);
1247         set2 = btree_bset_first(n2);
1248
1249         /*
1250          * Has to be a linear search because we don't have an auxiliary
1251          * search tree yet
1252          */
1253         k = set1->start;
1254         while (1) {
1255                 struct bkey_packed *n = bkey_next(k);
1256
1257                 if (n == vstruct_last(set1))
1258                         break;
1259                 if (k->_data - set1->_data >= (le16_to_cpu(set1->u64s) * 3) / 5)
1260                         break;
1261
1262                 if (bkey_packed(k))
1263                         nr_packed++;
1264                 else
1265                         nr_unpacked++;
1266
1267                 prev = k;
1268                 k = n;
1269         }
1270
1271         BUG_ON(!prev);
1272         set2_start      = k;
1273         set2_end        = vstruct_last(set1);
1274
1275         set1->u64s = cpu_to_le16((u64 *) set2_start - set1->_data);
1276         set_btree_bset_end(n1, n1->set);
1277
1278         n1->nr.live_u64s        = le16_to_cpu(set1->u64s);
1279         n1->nr.bset_u64s[0]     = le16_to_cpu(set1->u64s);
1280         n1->nr.packed_keys      = nr_packed;
1281         n1->nr.unpacked_keys    = nr_unpacked;
1282
1283         n1_pos = bkey_unpack_pos(n1, prev);
1284         if (as->c->sb.version < bcachefs_metadata_version_snapshot)
1285                 n1_pos.snapshot = U32_MAX;
1286
1287         btree_set_max(n1, n1_pos);
1288         btree_set_min(n2, bpos_successor(n1->key.k.p));
1289
1290         bch2_bkey_format_init(&s);
1291         bch2_bkey_format_add_pos(&s, n2->data->min_key);
1292         bch2_bkey_format_add_pos(&s, n2->data->max_key);
1293
1294         for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1295                 struct bkey uk = bkey_unpack_key(n1, k);
1296                 bch2_bkey_format_add_key(&s, &uk);
1297         }
1298
1299         n2->data->format = bch2_bkey_format_done(&s);
1300         btree_node_set_format(n2, n2->data->format);
1301
1302         out = set2->start;
1303         memset(&n2->nr, 0, sizeof(n2->nr));
1304
1305         for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1306                 BUG_ON(!bch2_bkey_transform(&n2->format, out, bkey_packed(k)
1307                                        ? &n1->format : &bch2_bkey_format_current, k));
1308                 out->format = KEY_FORMAT_LOCAL_BTREE;
1309                 btree_keys_account_key_add(&n2->nr, 0, out);
1310                 out = bkey_next(out);
1311         }
1312
1313         set2->u64s = cpu_to_le16((u64 *) out - set2->_data);
1314         set_btree_bset_end(n2, n2->set);
1315
1316         BUG_ON(!set1->u64s);
1317         BUG_ON(!set2->u64s);
1318
1319         btree_node_reset_sib_u64s(n1);
1320         btree_node_reset_sib_u64s(n2);
1321
1322         bch2_verify_btree_nr_keys(n1);
1323         bch2_verify_btree_nr_keys(n2);
1324
1325         if (n1->c.level) {
1326                 btree_node_interior_verify(as->c, n1);
1327                 btree_node_interior_verify(as->c, n2);
1328         }
1329
1330         return n2;
1331 }
1332
1333 /*
1334  * For updates to interior nodes, we've got to do the insert before we split
1335  * because the stuff we're inserting has to be inserted atomically. Post split,
1336  * the keys might have to go in different nodes and the split would no longer be
1337  * atomic.
1338  *
1339  * Worse, if the insert is from btree node coalescing, if we do the insert after
1340  * we do the split (and pick the pivot) - the pivot we pick might be between
1341  * nodes that were coalesced, and thus in the middle of a child node post
1342  * coalescing:
1343  */
1344 static void btree_split_insert_keys(struct btree_update *as,
1345                                     struct btree_trans *trans,
1346                                     struct btree_path *path,
1347                                     struct btree *b,
1348                                     struct keylist *keys)
1349 {
1350         struct btree_node_iter node_iter;
1351         struct bkey_i *k = bch2_keylist_front(keys);
1352         struct bkey_packed *src, *dst, *n;
1353         struct bset *i;
1354
1355         bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
1356
1357         __bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
1358
1359         /*
1360          * We can't tolerate whiteouts here - with whiteouts there can be
1361          * duplicate keys, and it would be rather bad if we picked a duplicate
1362          * for the pivot:
1363          */
1364         i = btree_bset_first(b);
1365         src = dst = i->start;
1366         while (src != vstruct_last(i)) {
1367                 n = bkey_next(src);
1368                 if (!bkey_deleted(src)) {
1369                         memmove_u64s_down(dst, src, src->u64s);
1370                         dst = bkey_next(dst);
1371                 }
1372                 src = n;
1373         }
1374
1375         /* Also clear out the unwritten whiteouts area: */
1376         b->whiteout_u64s = 0;
1377
1378         i->u64s = cpu_to_le16((u64 *) dst - i->_data);
1379         set_btree_bset_end(b, b->set);
1380
1381         BUG_ON(b->nsets != 1 ||
1382                b->nr.live_u64s != le16_to_cpu(btree_bset_first(b)->u64s));
1383
1384         btree_node_interior_verify(as->c, b);
1385 }
1386
1387 static void btree_split(struct btree_update *as, struct btree_trans *trans,
1388                         struct btree_path *path, struct btree *b,
1389                         struct keylist *keys, unsigned flags)
1390 {
1391         struct bch_fs *c = as->c;
1392         struct btree *parent = btree_node_parent(path, b);
1393         struct btree *n1, *n2 = NULL, *n3 = NULL;
1394         u64 start_time = local_clock();
1395
1396         BUG_ON(!parent && (b != btree_node_root(c, b)));
1397         BUG_ON(!btree_node_intent_locked(path, btree_node_root(c, b)->c.level));
1398
1399         bch2_btree_interior_update_will_free_node(as, b);
1400
1401         n1 = bch2_btree_node_alloc_replacement(as, b);
1402         bch2_btree_update_add_new_node(as, n1);
1403
1404         if (keys)
1405                 btree_split_insert_keys(as, trans, path, n1, keys);
1406
1407         if (bset_u64s(&n1->set[0]) > BTREE_SPLIT_THRESHOLD(c)) {
1408                 trace_btree_split(c, b);
1409
1410                 n2 = __btree_split_node(as, n1);
1411
1412                 bch2_btree_build_aux_trees(n2);
1413                 bch2_btree_build_aux_trees(n1);
1414                 six_unlock_write(&n2->c.lock);
1415                 six_unlock_write(&n1->c.lock);
1416
1417                 bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1418                 bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0);
1419
1420                 /*
1421                  * Note that on recursive parent_keys == keys, so we
1422                  * can't start adding new keys to parent_keys before emptying it
1423                  * out (which we did with btree_split_insert_keys() above)
1424                  */
1425                 bch2_keylist_add(&as->parent_keys, &n1->key);
1426                 bch2_keylist_add(&as->parent_keys, &n2->key);
1427
1428                 if (!parent) {
1429                         /* Depth increases, make a new root */
1430                         n3 = __btree_root_alloc(as, b->c.level + 1);
1431
1432                         n3->sib_u64s[0] = U16_MAX;
1433                         n3->sib_u64s[1] = U16_MAX;
1434
1435                         btree_split_insert_keys(as, trans, path, n3, &as->parent_keys);
1436
1437                         bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0);
1438                 }
1439         } else {
1440                 trace_btree_compact(c, b);
1441
1442                 bch2_btree_build_aux_trees(n1);
1443                 six_unlock_write(&n1->c.lock);
1444
1445                 bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1446
1447                 if (parent)
1448                         bch2_keylist_add(&as->parent_keys, &n1->key);
1449         }
1450
1451         /* New nodes all written, now make them visible: */
1452
1453         if (parent) {
1454                 /* Split a non root node */
1455                 bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1456         } else if (n3) {
1457                 bch2_btree_set_root(as, trans, path, n3);
1458         } else {
1459                 /* Root filled up but didn't need to be split */
1460                 bch2_btree_set_root(as, trans, path, n1);
1461         }
1462
1463         bch2_btree_update_get_open_buckets(as, n1);
1464         if (n2)
1465                 bch2_btree_update_get_open_buckets(as, n2);
1466         if (n3)
1467                 bch2_btree_update_get_open_buckets(as, n3);
1468
1469         /* Successful split, update the path to point to the new nodes: */
1470
1471         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1472         if (n3)
1473                 bch2_trans_node_add(trans, n3);
1474         if (n2)
1475                 bch2_trans_node_add(trans, n2);
1476         bch2_trans_node_add(trans, n1);
1477
1478         /*
1479          * The old node must be freed (in memory) _before_ unlocking the new
1480          * nodes - else another thread could re-acquire a read lock on the old
1481          * node after another thread has locked and updated the new node, thus
1482          * seeing stale data:
1483          */
1484         bch2_btree_node_free_inmem(trans, b);
1485
1486         if (n3)
1487                 six_unlock_intent(&n3->c.lock);
1488         if (n2)
1489                 six_unlock_intent(&n2->c.lock);
1490         six_unlock_intent(&n1->c.lock);
1491
1492         bch2_trans_verify_locks(trans);
1493
1494         bch2_time_stats_update(&c->times[n2
1495                                ? BCH_TIME_btree_node_split
1496                                : BCH_TIME_btree_node_compact],
1497                                start_time);
1498 }
1499
1500 static void
1501 bch2_btree_insert_keys_interior(struct btree_update *as,
1502                                 struct btree_trans *trans,
1503                                 struct btree_path *path,
1504                                 struct btree *b,
1505                                 struct keylist *keys)
1506 {
1507         struct btree_path *linked;
1508
1509         __bch2_btree_insert_keys_interior(as, trans, path, b,
1510                                           path->l[b->c.level].iter, keys);
1511
1512         btree_update_updated_node(as, b);
1513
1514         trans_for_each_path_with_node(trans, b, linked)
1515                 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1516
1517         bch2_trans_verify_paths(trans);
1518 }
1519
1520 /**
1521  * bch_btree_insert_node - insert bkeys into a given btree node
1522  *
1523  * @iter:               btree iterator
1524  * @keys:               list of keys to insert
1525  * @hook:               insert callback
1526  * @persistent:         if not null, @persistent will wait on journal write
1527  *
1528  * Inserts as many keys as it can into a given btree node, splitting it if full.
1529  * If a split occurred, this function will return early. This can only happen
1530  * for leaf nodes -- inserts into interior nodes have to be atomic.
1531  */
1532 static void bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans,
1533                                    struct btree_path *path, struct btree *b,
1534                                    struct keylist *keys, unsigned flags)
1535 {
1536         struct bch_fs *c = as->c;
1537         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1538         int old_live_u64s = b->nr.live_u64s;
1539         int live_u64s_added, u64s_added;
1540
1541         lockdep_assert_held(&c->gc_lock);
1542         BUG_ON(!btree_node_intent_locked(path, btree_node_root(c, b)->c.level));
1543         BUG_ON(!b->c.level);
1544         BUG_ON(!as || as->b);
1545         bch2_verify_keylist_sorted(keys);
1546
1547         bch2_btree_node_lock_for_insert(trans, path, b);
1548
1549         if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1550                 bch2_btree_node_unlock_write(trans, path, b);
1551                 goto split;
1552         }
1553
1554         btree_node_interior_verify(c, b);
1555
1556         bch2_btree_insert_keys_interior(as, trans, path, b, keys);
1557
1558         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1559         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1560
1561         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1562                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1563         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1564                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1565
1566         if (u64s_added > live_u64s_added &&
1567             bch2_maybe_compact_whiteouts(c, b))
1568                 bch2_trans_node_reinit_iter(trans, b);
1569
1570         bch2_btree_node_unlock_write(trans, path, b);
1571
1572         btree_node_interior_verify(c, b);
1573         return;
1574 split:
1575         btree_split(as, trans, path, b, keys, flags);
1576 }
1577
1578 int bch2_btree_split_leaf(struct btree_trans *trans,
1579                           struct btree_path *path,
1580                           unsigned flags)
1581 {
1582         struct btree *b = path_l(path)->b;
1583         struct btree_update *as;
1584         unsigned l;
1585         int ret = 0;
1586
1587         as = bch2_btree_update_start(trans, path, path->level,
1588                                      true, flags);
1589         if (IS_ERR(as))
1590                 return PTR_ERR(as);
1591
1592         btree_split(as, trans, path, b, NULL, flags);
1593         bch2_btree_update_done(as);
1594
1595         for (l = path->level + 1; btree_path_node(path, l) && !ret; l++)
1596                 ret = bch2_foreground_maybe_merge(trans, path, l, flags);
1597
1598         return ret;
1599 }
1600
1601 int __bch2_foreground_maybe_merge(struct btree_trans *trans,
1602                                   struct btree_path *path,
1603                                   unsigned level,
1604                                   unsigned flags,
1605                                   enum btree_node_sibling sib)
1606 {
1607         struct bch_fs *c = trans->c;
1608         struct btree_path *sib_path = NULL;
1609         struct btree_update *as;
1610         struct bkey_format_state new_s;
1611         struct bkey_format new_f;
1612         struct bkey_i delete;
1613         struct btree *b, *m, *n, *prev, *next, *parent;
1614         struct bpos sib_pos;
1615         size_t sib_u64s;
1616         u64 start_time = local_clock();
1617         int ret = 0;
1618
1619         BUG_ON(!path->should_be_locked);
1620         BUG_ON(!btree_node_locked(path, level));
1621
1622         b = path->l[level].b;
1623
1624         if ((sib == btree_prev_sib && !bpos_cmp(b->data->min_key, POS_MIN)) ||
1625             (sib == btree_next_sib && !bpos_cmp(b->data->max_key, SPOS_MAX))) {
1626                 b->sib_u64s[sib] = U16_MAX;
1627                 return 0;
1628         }
1629
1630         sib_pos = sib == btree_prev_sib
1631                 ? bpos_predecessor(b->data->min_key)
1632                 : bpos_successor(b->data->max_key);
1633
1634         sib_path = bch2_path_get(trans, path->btree_id, sib_pos,
1635                                  U8_MAX, level, BTREE_ITER_INTENT, _THIS_IP_);
1636         ret = bch2_btree_path_traverse(trans, sib_path, false);
1637         if (ret)
1638                 goto err;
1639
1640         sib_path->should_be_locked = true;
1641
1642         m = sib_path->l[level].b;
1643
1644         if (btree_node_parent(path, b) !=
1645             btree_node_parent(sib_path, m)) {
1646                 b->sib_u64s[sib] = U16_MAX;
1647                 goto out;
1648         }
1649
1650         if (sib == btree_prev_sib) {
1651                 prev = m;
1652                 next = b;
1653         } else {
1654                 prev = b;
1655                 next = m;
1656         }
1657
1658         if (bkey_cmp(bpos_successor(prev->data->max_key), next->data->min_key)) {
1659                 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1660
1661                 bch2_bpos_to_text(&buf1, prev->data->max_key);
1662                 bch2_bpos_to_text(&buf2, next->data->min_key);
1663                 bch_err(c,
1664                         "btree topology error in btree merge:\n"
1665                         "  prev ends at   %s\n"
1666                         "  next starts at %s",
1667                         buf1.buf, buf2.buf);
1668                 printbuf_exit(&buf1);
1669                 printbuf_exit(&buf2);
1670                 bch2_topology_error(c);
1671                 ret = -EIO;
1672                 goto err;
1673         }
1674
1675         bch2_bkey_format_init(&new_s);
1676         bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
1677         __bch2_btree_calc_format(&new_s, prev);
1678         __bch2_btree_calc_format(&new_s, next);
1679         bch2_bkey_format_add_pos(&new_s, next->data->max_key);
1680         new_f = bch2_bkey_format_done(&new_s);
1681
1682         sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1683                 btree_node_u64s_with_format(m, &new_f);
1684
1685         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1686                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1687                 sib_u64s /= 2;
1688                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1689         }
1690
1691         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1692         sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
1693         b->sib_u64s[sib] = sib_u64s;
1694
1695         if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
1696                 goto out;
1697
1698         parent = btree_node_parent(path, b);
1699         as = bch2_btree_update_start(trans, path, level, false,
1700                          BTREE_INSERT_NOFAIL|
1701                          BTREE_INSERT_USE_RESERVE|
1702                          flags);
1703         ret = PTR_ERR_OR_ZERO(as);
1704         if (ret)
1705                 goto err;
1706
1707         trace_btree_merge(c, b);
1708
1709         bch2_btree_interior_update_will_free_node(as, b);
1710         bch2_btree_interior_update_will_free_node(as, m);
1711
1712         n = bch2_btree_node_alloc(as, b->c.level);
1713         bch2_btree_update_add_new_node(as, n);
1714
1715         SET_BTREE_NODE_SEQ(n->data,
1716                            max(BTREE_NODE_SEQ(b->data),
1717                                BTREE_NODE_SEQ(m->data)) + 1);
1718
1719         btree_set_min(n, prev->data->min_key);
1720         btree_set_max(n, next->data->max_key);
1721         n->data->format         = new_f;
1722
1723         btree_node_set_format(n, new_f);
1724
1725         bch2_btree_sort_into(c, n, prev);
1726         bch2_btree_sort_into(c, n, next);
1727
1728         bch2_btree_build_aux_trees(n);
1729         six_unlock_write(&n->c.lock);
1730
1731         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1732
1733         bkey_init(&delete.k);
1734         delete.k.p = prev->key.k.p;
1735         bch2_keylist_add(&as->parent_keys, &delete);
1736         bch2_keylist_add(&as->parent_keys, &n->key);
1737
1738         bch2_trans_verify_paths(trans);
1739
1740         bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1741
1742         bch2_trans_verify_paths(trans);
1743
1744         bch2_btree_update_get_open_buckets(as, n);
1745
1746         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1747         six_lock_increment(&m->c.lock, SIX_LOCK_intent);
1748
1749         bch2_trans_node_add(trans, n);
1750
1751         bch2_trans_verify_paths(trans);
1752
1753         bch2_btree_node_free_inmem(trans, b);
1754         bch2_btree_node_free_inmem(trans, m);
1755
1756         six_unlock_intent(&n->c.lock);
1757
1758         bch2_btree_update_done(as);
1759
1760         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time);
1761 out:
1762 err:
1763         bch2_path_put(trans, sib_path, true);
1764         bch2_trans_verify_locks(trans);
1765         return ret;
1766 }
1767
1768 /**
1769  * bch_btree_node_rewrite - Rewrite/move a btree node
1770  */
1771 int bch2_btree_node_rewrite(struct btree_trans *trans,
1772                             struct btree_iter *iter,
1773                             struct btree *b,
1774                             unsigned flags)
1775 {
1776         struct bch_fs *c = trans->c;
1777         struct btree *n, *parent;
1778         struct btree_update *as;
1779         int ret;
1780
1781         flags |= BTREE_INSERT_NOFAIL;
1782
1783         parent = btree_node_parent(iter->path, b);
1784         as = bch2_btree_update_start(trans, iter->path, b->c.level,
1785                                      false, flags);
1786         ret = PTR_ERR_OR_ZERO(as);
1787         if (ret) {
1788                 trace_btree_gc_rewrite_node_fail(c, b);
1789                 goto out;
1790         }
1791
1792         bch2_btree_interior_update_will_free_node(as, b);
1793
1794         n = bch2_btree_node_alloc_replacement(as, b);
1795         bch2_btree_update_add_new_node(as, n);
1796
1797         bch2_btree_build_aux_trees(n);
1798         six_unlock_write(&n->c.lock);
1799
1800         trace_btree_gc_rewrite_node(c, b);
1801
1802         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1803
1804         if (parent) {
1805                 bch2_keylist_add(&as->parent_keys, &n->key);
1806                 bch2_btree_insert_node(as, trans, iter->path, parent,
1807                                        &as->parent_keys, flags);
1808         } else {
1809                 bch2_btree_set_root(as, trans, iter->path, n);
1810         }
1811
1812         bch2_btree_update_get_open_buckets(as, n);
1813
1814         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1815         bch2_trans_node_add(trans, n);
1816         bch2_btree_node_free_inmem(trans, b);
1817         six_unlock_intent(&n->c.lock);
1818
1819         bch2_btree_update_done(as);
1820 out:
1821         bch2_btree_path_downgrade(iter->path);
1822         return ret;
1823 }
1824
1825 struct async_btree_rewrite {
1826         struct bch_fs           *c;
1827         struct work_struct      work;
1828         enum btree_id           btree_id;
1829         unsigned                level;
1830         struct bpos             pos;
1831         __le64                  seq;
1832 };
1833
1834 static int async_btree_node_rewrite_trans(struct btree_trans *trans,
1835                                           struct async_btree_rewrite *a)
1836 {
1837         struct btree_iter iter;
1838         struct btree *b;
1839         int ret;
1840
1841         bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos,
1842                                   BTREE_MAX_DEPTH, a->level, 0);
1843         b = bch2_btree_iter_peek_node(&iter);
1844         ret = PTR_ERR_OR_ZERO(b);
1845         if (ret)
1846                 goto out;
1847
1848         if (!b || b->data->keys.seq != a->seq)
1849                 goto out;
1850
1851         ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
1852 out :
1853         bch2_trans_iter_exit(trans, &iter);
1854
1855         return ret;
1856 }
1857
1858 void async_btree_node_rewrite_work(struct work_struct *work)
1859 {
1860         struct async_btree_rewrite *a =
1861                 container_of(work, struct async_btree_rewrite, work);
1862         struct bch_fs *c = a->c;
1863
1864         bch2_trans_do(c, NULL, NULL, 0,
1865                       async_btree_node_rewrite_trans(&trans, a));
1866         percpu_ref_put(&c->writes);
1867         kfree(a);
1868 }
1869
1870 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
1871 {
1872         struct async_btree_rewrite *a;
1873
1874         if (!percpu_ref_tryget(&c->writes))
1875                 return;
1876
1877         a = kmalloc(sizeof(*a), GFP_NOFS);
1878         if (!a) {
1879                 percpu_ref_put(&c->writes);
1880                 return;
1881         }
1882
1883         a->c            = c;
1884         a->btree_id     = b->c.btree_id;
1885         a->level        = b->c.level;
1886         a->pos          = b->key.k.p;
1887         a->seq          = b->data->keys.seq;
1888
1889         INIT_WORK(&a->work, async_btree_node_rewrite_work);
1890         queue_work(c->btree_interior_update_worker, &a->work);
1891 }
1892
1893 static int __bch2_btree_node_update_key(struct btree_trans *trans,
1894                                         struct btree_iter *iter,
1895                                         struct btree *b, struct btree *new_hash,
1896                                         struct bkey_i *new_key,
1897                                         bool skip_triggers)
1898 {
1899         struct bch_fs *c = trans->c;
1900         struct btree_iter iter2 = { NULL };
1901         struct btree *parent;
1902         u64 journal_entries[BKEY_BTREE_PTR_U64s_MAX];
1903         int ret;
1904
1905         if (!skip_triggers) {
1906                 ret = bch2_trans_mark_new(trans, new_key, 0);
1907                 if (ret)
1908                         return ret;
1909
1910                 ret = bch2_trans_mark_old(trans, bkey_i_to_s_c(&b->key), 0);
1911                 if (ret)
1912                         return ret;
1913         }
1914
1915         if (new_hash) {
1916                 bkey_copy(&new_hash->key, new_key);
1917                 ret = bch2_btree_node_hash_insert(&c->btree_cache,
1918                                 new_hash, b->c.level, b->c.btree_id);
1919                 BUG_ON(ret);
1920         }
1921
1922         parent = btree_node_parent(iter->path, b);
1923         if (parent) {
1924                 bch2_trans_copy_iter(&iter2, iter);
1925
1926                 iter2.path = bch2_btree_path_make_mut(trans, iter2.path,
1927                                 iter2.flags & BTREE_ITER_INTENT,
1928                                 _THIS_IP_);
1929
1930                 BUG_ON(iter2.path->level != b->c.level);
1931                 BUG_ON(bpos_cmp(iter2.path->pos, new_key->k.p));
1932
1933                 btree_node_unlock(iter2.path, iter2.path->level);
1934                 path_l(iter2.path)->b = BTREE_ITER_NO_NODE_UP;
1935                 iter2.path->level++;
1936
1937                 bch2_btree_path_check_sort(trans, iter2.path, 0);
1938
1939                 ret   = bch2_btree_iter_traverse(&iter2) ?:
1940                         bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_NORUN);
1941                 if (ret)
1942                         goto err;
1943         } else {
1944                 BUG_ON(btree_node_root(c, b) != b);
1945
1946                 trans->extra_journal_entries = (void *) &journal_entries[0];
1947                 trans->extra_journal_entry_u64s =
1948                         journal_entry_set((void *) &journal_entries[0],
1949                                           BCH_JSET_ENTRY_btree_root,
1950                                           b->c.btree_id, b->c.level,
1951                                           new_key, new_key->k.u64s);
1952         }
1953
1954         ret = bch2_trans_commit(trans, NULL, NULL,
1955                                 BTREE_INSERT_NOFAIL|
1956                                 BTREE_INSERT_NOCHECK_RW|
1957                                 BTREE_INSERT_USE_RESERVE|
1958                                 BTREE_INSERT_JOURNAL_RECLAIM|
1959                                 JOURNAL_WATERMARK_reserved);
1960         if (ret)
1961                 goto err;
1962
1963         bch2_btree_node_lock_write(trans, iter->path, b);
1964
1965         if (new_hash) {
1966                 mutex_lock(&c->btree_cache.lock);
1967                 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
1968                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1969
1970                 bkey_copy(&b->key, new_key);
1971                 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1972                 BUG_ON(ret);
1973                 mutex_unlock(&c->btree_cache.lock);
1974         } else {
1975                 bkey_copy(&b->key, new_key);
1976         }
1977
1978         bch2_btree_node_unlock_write(trans, iter->path, b);
1979 out:
1980         bch2_trans_iter_exit(trans, &iter2);
1981         return ret;
1982 err:
1983         if (new_hash) {
1984                 mutex_lock(&c->btree_cache.lock);
1985                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1986                 mutex_unlock(&c->btree_cache.lock);
1987         }
1988         goto out;
1989 }
1990
1991 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
1992                                struct btree *b, struct bkey_i *new_key,
1993                                bool skip_triggers)
1994 {
1995         struct bch_fs *c = trans->c;
1996         struct btree *new_hash = NULL;
1997         struct btree_path *path = iter->path;
1998         struct closure cl;
1999         int ret = 0;
2000
2001         if (!btree_node_intent_locked(path, b->c.level) &&
2002             !bch2_btree_path_upgrade(trans, path, b->c.level + 1)) {
2003                 btree_trans_restart(trans);
2004                 return -EINTR;
2005         }
2006
2007         closure_init_stack(&cl);
2008
2009         /*
2010          * check btree_ptr_hash_val() after @b is locked by
2011          * btree_iter_traverse():
2012          */
2013         if (btree_ptr_hash_val(new_key) != b->hash_val) {
2014                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2015                 if (ret) {
2016                         bch2_trans_unlock(trans);
2017                         closure_sync(&cl);
2018                         if (!bch2_trans_relock(trans))
2019                                 return -EINTR;
2020                 }
2021
2022                 new_hash = bch2_btree_node_mem_alloc(c, false);
2023         }
2024
2025         path->intent_ref++;
2026         ret = __bch2_btree_node_update_key(trans, iter, b, new_hash,
2027                                            new_key, skip_triggers);
2028         --path->intent_ref;
2029
2030         if (new_hash) {
2031                 mutex_lock(&c->btree_cache.lock);
2032                 list_move(&new_hash->list, &c->btree_cache.freeable);
2033                 mutex_unlock(&c->btree_cache.lock);
2034
2035                 six_unlock_write(&new_hash->c.lock);
2036                 six_unlock_intent(&new_hash->c.lock);
2037         }
2038         closure_sync(&cl);
2039         bch2_btree_cache_cannibalize_unlock(c);
2040         return ret;
2041 }
2042
2043 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
2044                                         struct btree *b, struct bkey_i *new_key,
2045                                         bool skip_triggers)
2046 {
2047         struct btree_iter iter;
2048         int ret;
2049
2050         bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p,
2051                                   BTREE_MAX_DEPTH, b->c.level,
2052                                   BTREE_ITER_INTENT);
2053         ret = bch2_btree_iter_traverse(&iter);
2054         if (ret)
2055                 goto out;
2056
2057         /* has node been freed? */
2058         if (iter.path->l[b->c.level].b != b) {
2059                 /* node has been freed: */
2060                 BUG_ON(!btree_node_dying(b));
2061                 goto out;
2062         }
2063
2064         BUG_ON(!btree_node_hashed(b));
2065
2066         ret = bch2_btree_node_update_key(trans, &iter, b, new_key, skip_triggers);
2067 out:
2068         bch2_trans_iter_exit(trans, &iter);
2069         return ret;
2070 }
2071
2072 /* Init code: */
2073
2074 /*
2075  * Only for filesystem bringup, when first reading the btree roots or allocating
2076  * btree roots when initializing a new filesystem:
2077  */
2078 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2079 {
2080         BUG_ON(btree_node_root(c, b));
2081
2082         bch2_btree_set_root_inmem(c, b);
2083 }
2084
2085 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2086 {
2087         struct closure cl;
2088         struct btree *b;
2089         int ret;
2090
2091         closure_init_stack(&cl);
2092
2093         do {
2094                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2095                 closure_sync(&cl);
2096         } while (ret);
2097
2098         b = bch2_btree_node_mem_alloc(c, false);
2099         bch2_btree_cache_cannibalize_unlock(c);
2100
2101         set_btree_node_fake(b);
2102         set_btree_node_need_rewrite(b);
2103         b->c.level      = 0;
2104         b->c.btree_id   = id;
2105
2106         bkey_btree_ptr_init(&b->key);
2107         b->key.k.p = SPOS_MAX;
2108         *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2109
2110         bch2_bset_init_first(b, &b->data->keys);
2111         bch2_btree_build_aux_trees(b);
2112
2113         b->data->flags = 0;
2114         btree_set_min(b, POS_MIN);
2115         btree_set_max(b, SPOS_MAX);
2116         b->data->format = bch2_btree_calc_format(b);
2117         btree_node_set_format(b, b->data->format);
2118
2119         ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2120                                           b->c.level, b->c.btree_id);
2121         BUG_ON(ret);
2122
2123         bch2_btree_set_root_inmem(c, b);
2124
2125         six_unlock_write(&b->c.lock);
2126         six_unlock_intent(&b->c.lock);
2127 }
2128
2129 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2130 {
2131         struct btree_update *as;
2132
2133         mutex_lock(&c->btree_interior_update_lock);
2134         list_for_each_entry(as, &c->btree_interior_update_list, list)
2135                 pr_buf(out, "%p m %u w %u r %u j %llu\n",
2136                        as,
2137                        as->mode,
2138                        as->nodes_written,
2139                        atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2140                        as->journal.seq);
2141         mutex_unlock(&c->btree_interior_update_lock);
2142 }
2143
2144 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *c)
2145 {
2146         size_t ret = 0;
2147         struct list_head *i;
2148
2149         mutex_lock(&c->btree_interior_update_lock);
2150         list_for_each(i, &c->btree_interior_update_list)
2151                 ret++;
2152         mutex_unlock(&c->btree_interior_update_lock);
2153
2154         return ret;
2155 }
2156
2157 void bch2_journal_entries_to_btree_roots(struct bch_fs *c, struct jset *jset)
2158 {
2159         struct btree_root *r;
2160         struct jset_entry *entry;
2161
2162         mutex_lock(&c->btree_root_lock);
2163
2164         vstruct_for_each(jset, entry)
2165                 if (entry->type == BCH_JSET_ENTRY_btree_root) {
2166                         r = &c->btree_roots[entry->btree_id];
2167                         r->level = entry->level;
2168                         r->alive = true;
2169                         bkey_copy(&r->key, &entry->start[0]);
2170                 }
2171
2172         mutex_unlock(&c->btree_root_lock);
2173 }
2174
2175 struct jset_entry *
2176 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2177                                     struct jset_entry *start,
2178                                     struct jset_entry *end)
2179 {
2180         struct jset_entry *entry;
2181         unsigned long have = 0;
2182         unsigned i;
2183
2184         for (entry = start; entry < end; entry = vstruct_next(entry))
2185                 if (entry->type == BCH_JSET_ENTRY_btree_root)
2186                         __set_bit(entry->btree_id, &have);
2187
2188         mutex_lock(&c->btree_root_lock);
2189
2190         for (i = 0; i < BTREE_ID_NR; i++)
2191                 if (c->btree_roots[i].alive && !test_bit(i, &have)) {
2192                         journal_entry_set(end,
2193                                           BCH_JSET_ENTRY_btree_root,
2194                                           i, c->btree_roots[i].level,
2195                                           &c->btree_roots[i].key,
2196                                           c->btree_roots[i].key.u64s);
2197                         end = vstruct_next(end);
2198                 }
2199
2200         mutex_unlock(&c->btree_root_lock);
2201
2202         return end;
2203 }
2204
2205 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2206 {
2207         if (c->btree_interior_update_worker)
2208                 destroy_workqueue(c->btree_interior_update_worker);
2209         mempool_exit(&c->btree_interior_update_pool);
2210 }
2211
2212 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2213 {
2214         mutex_init(&c->btree_reserve_cache_lock);
2215         INIT_LIST_HEAD(&c->btree_interior_update_list);
2216         INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2217         mutex_init(&c->btree_interior_update_lock);
2218         INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2219
2220         c->btree_interior_update_worker =
2221                 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2222         if (!c->btree_interior_update_worker)
2223                 return -ENOMEM;
2224
2225         return mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2226                                          sizeof(struct btree_update));
2227 }