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         ret = darray_make_room(trans->extra_journal_entries, as->journal_u64s);
 536         if (ret)
 537                 return ret;
 538
 539         memcpy(&darray_top(trans->extra_journal_entries),
 540                as->journal_entries,
 541                as->journal_u64s * sizeof(u64));
 542         trans->extra_journal_entries.nr += as->journal_u64s;
 543
 544         trans->journal_pin = &as->journal;
 545
 546         for_each_keylist_key(&as->new_keys, k) {
 547                 ret = bch2_trans_mark_new(trans, k, 0);
 548                 if (ret)
 549                         return ret;
 550         }
 551
 552         for_each_keylist_key(&as->old_keys, k) {
 553                 ret = bch2_trans_mark_old(trans, bkey_i_to_s_c(k), 0);
 554                 if (ret)
 555                         return ret;
 556         }
 557
 558         return 0;
 559 }
 560
 561 static void btree_update_nodes_written(struct btree_update *as)
 562 {
 563         struct bch_fs *c = as->c;
 564         struct btree *b = as->b;
 565         struct btree_trans trans;
 566         u64 journal_seq = 0;
 567         unsigned i;
 568         int ret;
 569
 570         /*
 571          * If we're already in an error state, it might be because a btree node
 572          * was never written, and we might be trying to free that same btree
 573          * node here, but it won't have been marked as allocated and we'll see
 574          * spurious disk usage inconsistencies in the transactional part below
 575          * if we don't skip it:
 576          */
 577         ret = bch2_journal_error(&c->journal);
 578         if (ret)
 579                 goto err;
 580
 581         /*
 582          * Wait for any in flight writes to finish before we free the old nodes
 583          * on disk:
 584          */
 585         for (i = 0; i < as->nr_old_nodes; i++) {
 586                 struct btree *old = as->old_nodes[i];
 587                 __le64 seq;
 588
 589                 six_lock_read(&old->c.lock, NULL, NULL);
 590                 seq = old->data ? old->data->keys.seq : 0;
 591                 six_unlock_read(&old->c.lock);
 592
 593                 if (seq == as->old_nodes_seq[i])
 594                         wait_on_bit_io(&old->flags, BTREE_NODE_write_in_flight_inner,
 595                                        TASK_UNINTERRUPTIBLE);
 596         }
 597
 598         /*
 599          * We did an update to a parent node where the pointers we added pointed
 600          * to child nodes that weren't written yet: now, the child nodes have
 601          * been written so we can write out the update to the interior node.
 602          */
 603
 604         /*
 605          * We can't call into journal reclaim here: we'd block on the journal
 606          * reclaim lock, but we may need to release the open buckets we have
 607          * pinned in order for other btree updates to make forward progress, and
 608          * journal reclaim does btree updates when flushing bkey_cached entries,
 609          * which may require allocations as well.
 610          */
 611         bch2_trans_init(&trans, c, 0, 512);
 612         ret = __bch2_trans_do(&trans, &as->disk_res, &journal_seq,
 613                               BTREE_INSERT_NOFAIL|
 614                               BTREE_INSERT_NOCHECK_RW|
 615                               BTREE_INSERT_JOURNAL_RECLAIM|
 616                               JOURNAL_WATERMARK_reserved,
 617                               btree_update_nodes_written_trans(&trans, as));
 618         bch2_trans_exit(&trans);
 619
 620         bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
 621                              "error %i in btree_update_nodes_written()", ret);
 622 err:
 623         if (b) {
 624                 /*
 625                  * @b is the node we did the final insert into:
 626                  *
 627                  * On failure to get a journal reservation, we still have to
 628                  * unblock the write and allow most of the write path to happen
 629                  * so that shutdown works, but the i->journal_seq mechanism
 630                  * won't work to prevent the btree write from being visible (we
 631                  * didn't get a journal sequence number) - instead
 632                  * __bch2_btree_node_write() doesn't do the actual write if
 633                  * we're in journal error state:
 634                  */
 635
 636                 six_lock_intent(&b->c.lock, NULL, NULL);
 637                 six_lock_write(&b->c.lock, NULL, NULL);
 638                 mutex_lock(&c->btree_interior_update_lock);
 639
 640                 list_del(&as->write_blocked_list);
 641                 if (list_empty(&b->write_blocked))
 642                         clear_btree_node_write_blocked(b);
 643
 644                 /*
 645                  * Node might have been freed, recheck under
 646                  * btree_interior_update_lock:
 647                  */
 648                 if (as->b == b) {
 649                         struct bset *i = btree_bset_last(b);
 650
 651                         BUG_ON(!b->c.level);
 652                         BUG_ON(!btree_node_dirty(b));
 653
 654                         if (!ret) {
 655                                 i->journal_seq = cpu_to_le64(
 656                                         max(journal_seq,
 657                                             le64_to_cpu(i->journal_seq)));
 658
 659                                 bch2_btree_add_journal_pin(c, b, journal_seq);
 660                         } else {
 661                                 /*
 662                                  * If we didn't get a journal sequence number we
 663                                  * can't write this btree node, because recovery
 664                                  * won't know to ignore this write:
 665                                  */
 666                                 set_btree_node_never_write(b);
 667                         }
 668                 }
 669
 670                 mutex_unlock(&c->btree_interior_update_lock);
 671                 six_unlock_write(&b->c.lock);
 672
 673                 btree_node_write_if_need(c, b, SIX_LOCK_intent);
 674                 six_unlock_intent(&b->c.lock);
 675         }
 676
 677         bch2_journal_pin_drop(&c->journal, &as->journal);
 678
 679         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 680
 681         mutex_lock(&c->btree_interior_update_lock);
 682         for (i = 0; i < as->nr_new_nodes; i++) {
 683                 b = as->new_nodes[i];
 684
 685                 BUG_ON(b->will_make_reachable != (unsigned long) as);
 686                 b->will_make_reachable = 0;
 687                 clear_btree_node_will_make_reachable(b);
 688         }
 689         mutex_unlock(&c->btree_interior_update_lock);
 690
 691         for (i = 0; i < as->nr_new_nodes; i++) {
 692                 b = as->new_nodes[i];
 693
 694                 six_lock_read(&b->c.lock, NULL, NULL);
 695                 btree_node_write_if_need(c, b, SIX_LOCK_read);
 696                 six_unlock_read(&b->c.lock);
 697         }
 698
 699         for (i = 0; i < as->nr_open_buckets; i++)
 700                 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
 701
 702         bch2_btree_update_free(as);
 703 }
 704
 705 static void btree_interior_update_work(struct work_struct *work)
 706 {
 707         struct bch_fs *c =
 708                 container_of(work, struct bch_fs, btree_interior_update_work);
 709         struct btree_update *as;
 710
 711         while (1) {
 712                 mutex_lock(&c->btree_interior_update_lock);
 713                 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
 714                                               struct btree_update, unwritten_list);
 715                 if (as && !as->nodes_written)
 716                         as = NULL;
 717                 mutex_unlock(&c->btree_interior_update_lock);
 718
 719                 if (!as)
 720                         break;
 721
 722                 btree_update_nodes_written(as);
 723         }
 724 }
 725
 726 static void btree_update_set_nodes_written(struct closure *cl)
 727 {
 728         struct btree_update *as = container_of(cl, struct btree_update, cl);
 729         struct bch_fs *c = as->c;
 730
 731         mutex_lock(&c->btree_interior_update_lock);
 732         as->nodes_written = true;
 733         mutex_unlock(&c->btree_interior_update_lock);
 734
 735         queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
 736 }
 737
 738 /*
 739  * We're updating @b with pointers to nodes that haven't finished writing yet:
 740  * block @b from being written until @as completes
 741  */
 742 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
 743 {
 744         struct bch_fs *c = as->c;
 745
 746         mutex_lock(&c->btree_interior_update_lock);
 747         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 748
 749         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 750         BUG_ON(!btree_node_dirty(b));
 751
 752         as->mode        = BTREE_INTERIOR_UPDATING_NODE;
 753         as->b           = b;
 754
 755         set_btree_node_write_blocked(b);
 756         list_add(&as->write_blocked_list, &b->write_blocked);
 757
 758         mutex_unlock(&c->btree_interior_update_lock);
 759 }
 760
 761 static void btree_update_reparent(struct btree_update *as,
 762                                   struct btree_update *child)
 763 {
 764         struct bch_fs *c = as->c;
 765
 766         lockdep_assert_held(&c->btree_interior_update_lock);
 767
 768         child->b = NULL;
 769         child->mode = BTREE_INTERIOR_UPDATING_AS;
 770
 771         bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL);
 772 }
 773
 774 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
 775 {
 776         struct bkey_i *insert = &b->key;
 777         struct bch_fs *c = as->c;
 778
 779         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 780
 781         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
 782                ARRAY_SIZE(as->journal_entries));
 783
 784         as->journal_u64s +=
 785                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
 786                                   BCH_JSET_ENTRY_btree_root,
 787                                   b->c.btree_id, b->c.level,
 788                                   insert, insert->k.u64s);
 789
 790         mutex_lock(&c->btree_interior_update_lock);
 791         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 792
 793         as->mode        = BTREE_INTERIOR_UPDATING_ROOT;
 794         mutex_unlock(&c->btree_interior_update_lock);
 795 }
 796
 797 /*
 798  * bch2_btree_update_add_new_node:
 799  *
 800  * This causes @as to wait on @b to be written, before it gets to
 801  * bch2_btree_update_nodes_written
 802  *
 803  * Additionally, it sets b->will_make_reachable to prevent any additional writes
 804  * to @b from happening besides the first until @b is reachable on disk
 805  *
 806  * And it adds @b to the list of @as's new nodes, so that we can update sector
 807  * counts in bch2_btree_update_nodes_written:
 808  */
 809 static void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
 810 {
 811         struct bch_fs *c = as->c;
 812
 813         closure_get(&as->cl);
 814
 815         mutex_lock(&c->btree_interior_update_lock);
 816         BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
 817         BUG_ON(b->will_make_reachable);
 818
 819         as->new_nodes[as->nr_new_nodes++] = b;
 820         b->will_make_reachable = 1UL|(unsigned long) as;
 821         set_btree_node_will_make_reachable(b);
 822
 823         mutex_unlock(&c->btree_interior_update_lock);
 824
 825         btree_update_will_add_key(as, &b->key);
 826 }
 827
 828 /*
 829  * returns true if @b was a new node
 830  */
 831 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
 832 {
 833         struct btree_update *as;
 834         unsigned long v;
 835         unsigned i;
 836
 837         mutex_lock(&c->btree_interior_update_lock);
 838         /*
 839          * When b->will_make_reachable != 0, it owns a ref on as->cl that's
 840          * dropped when it gets written by bch2_btree_complete_write - the
 841          * xchg() is for synchronization with bch2_btree_complete_write:
 842          */
 843         v = xchg(&b->will_make_reachable, 0);
 844         clear_btree_node_will_make_reachable(b);
 845         as = (struct btree_update *) (v & ~1UL);
 846
 847         if (!as) {
 848                 mutex_unlock(&c->btree_interior_update_lock);
 849                 return;
 850         }
 851
 852         for (i = 0; i < as->nr_new_nodes; i++)
 853                 if (as->new_nodes[i] == b)
 854                         goto found;
 855
 856         BUG();
 857 found:
 858         array_remove_item(as->new_nodes, as->nr_new_nodes, i);
 859         mutex_unlock(&c->btree_interior_update_lock);
 860
 861         if (v & 1)
 862                 closure_put(&as->cl);
 863 }
 864
 865 static void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
 866 {
 867         while (b->ob.nr)
 868                 as->open_buckets[as->nr_open_buckets++] =
 869                         b->ob.v[--b->ob.nr];
 870 }
 871
 872 /*
 873  * @b is being split/rewritten: it may have pointers to not-yet-written btree
 874  * nodes and thus outstanding btree_updates - redirect @b's
 875  * btree_updates to point to this btree_update:
 876  */
 877 static void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 878                                                struct btree *b)
 879 {
 880         struct bch_fs *c = as->c;
 881         struct btree_update *p, *n;
 882         struct btree_write *w;
 883
 884         set_btree_node_dying(b);
 885
 886         if (btree_node_fake(b))
 887                 return;
 888
 889         mutex_lock(&c->btree_interior_update_lock);
 890
 891         /*
 892          * Does this node have any btree_update operations preventing
 893          * it from being written?
 894          *
 895          * If so, redirect them to point to this btree_update: we can
 896          * write out our new nodes, but we won't make them visible until those
 897          * operations complete
 898          */
 899         list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
 900                 list_del_init(&p->write_blocked_list);
 901                 btree_update_reparent(as, p);
 902
 903                 /*
 904                  * for flush_held_btree_writes() waiting on updates to flush or
 905                  * nodes to be writeable:
 906                  */
 907                 closure_wake_up(&c->btree_interior_update_wait);
 908         }
 909
 910         clear_btree_node_dirty_acct(c, b);
 911         clear_btree_node_need_write(b);
 912
 913         /*
 914          * Does this node have unwritten data that has a pin on the journal?
 915          *
 916          * If so, transfer that pin to the btree_update operation -
 917          * note that if we're freeing multiple nodes, we only need to keep the
 918          * oldest pin of any of the nodes we're freeing. We'll release the pin
 919          * when the new nodes are persistent and reachable on disk:
 920          */
 921         w = btree_current_write(b);
 922         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 923         bch2_journal_pin_drop(&c->journal, &w->journal);
 924
 925         w = btree_prev_write(b);
 926         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 927         bch2_journal_pin_drop(&c->journal, &w->journal);
 928
 929         mutex_unlock(&c->btree_interior_update_lock);
 930
 931         /*
 932          * Is this a node that isn't reachable on disk yet?
 933          *
 934          * Nodes that aren't reachable yet have writes blocked until they're
 935          * reachable - now that we've cancelled any pending writes and moved
 936          * things waiting on that write to wait on this update, we can drop this
 937          * node from the list of nodes that the other update is making
 938          * reachable, prior to freeing it:
 939          */
 940         btree_update_drop_new_node(c, b);
 941
 942         btree_update_will_delete_key(as, &b->key);
 943
 944         as->old_nodes[as->nr_old_nodes] = b;
 945         as->old_nodes_seq[as->nr_old_nodes] = b->data->keys.seq;
 946         as->nr_old_nodes++;
 947 }
 948
 949 static void bch2_btree_update_done(struct btree_update *as)
 950 {
 951         struct bch_fs *c = as->c;
 952         u64 start_time = as->start_time;
 953
 954         BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
 955
 956         if (as->took_gc_lock)
 957                 up_read(&as->c->gc_lock);
 958         as->took_gc_lock = false;
 959
 960         bch2_btree_reserve_put(as);
 961
 962         continue_at(&as->cl, btree_update_set_nodes_written,
 963                     as->c->btree_interior_update_worker);
 964
 965         bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_foreground],
 966                                start_time);
 967 }
 968
 969 static struct btree_update *
 970 bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
 971                         unsigned level, bool split, unsigned flags)
 972 {
 973         struct bch_fs *c = trans->c;
 974         struct btree_update *as;
 975         u64 start_time = local_clock();
 976         int disk_res_flags = (flags & BTREE_INSERT_NOFAIL)
 977                 ? BCH_DISK_RESERVATION_NOFAIL : 0;
 978         unsigned nr_nodes[2] = { 0, 0 };
 979         unsigned update_level = level;
 980         int journal_flags = flags & JOURNAL_WATERMARK_MASK;
 981         int ret = 0;
 982
 983         BUG_ON(!path->should_be_locked);
 984
 985         if (flags & BTREE_INSERT_JOURNAL_RECLAIM)
 986                 journal_flags |= JOURNAL_RES_GET_NONBLOCK;
 987
 988         while (1) {
 989                 nr_nodes[!!update_level] += 1 + split;
 990                 update_level++;
 991
 992                 if (!btree_path_node(path, update_level))
 993                         break;
 994
 995                 /*
 996                  * XXX: figure out how far we might need to split,
 997                  * instead of locking/reserving all the way to the root:
 998                  */
 999                 split = update_level + 1 < BTREE_MAX_DEPTH;
1000         }
1001
1002         /* Might have to allocate a new root: */
1003         if (update_level < BTREE_MAX_DEPTH)
1004                 nr_nodes[1] += 1;
1005
1006         if (!bch2_btree_path_upgrade(trans, path, U8_MAX)) {
1007                 trace_trans_restart_iter_upgrade(trans->fn, _RET_IP_,
1008                                                  path->btree_id, &path->pos);
1009                 ret = btree_trans_restart(trans);
1010                 return ERR_PTR(ret);
1011         }
1012
1013         if (flags & BTREE_INSERT_GC_LOCK_HELD)
1014                 lockdep_assert_held(&c->gc_lock);
1015         else if (!down_read_trylock(&c->gc_lock)) {
1016                 bch2_trans_unlock(trans);
1017                 down_read(&c->gc_lock);
1018                 if (!bch2_trans_relock(trans)) {
1019                         up_read(&c->gc_lock);
1020                         return ERR_PTR(-EINTR);
1021                 }
1022         }
1023
1024         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
1025         memset(as, 0, sizeof(*as));
1026         closure_init(&as->cl, NULL);
1027         as->c           = c;
1028         as->start_time  = start_time;
1029         as->mode        = BTREE_INTERIOR_NO_UPDATE;
1030         as->took_gc_lock = !(flags & BTREE_INSERT_GC_LOCK_HELD);
1031         as->btree_id    = path->btree_id;
1032         INIT_LIST_HEAD(&as->list);
1033         INIT_LIST_HEAD(&as->unwritten_list);
1034         INIT_LIST_HEAD(&as->write_blocked_list);
1035         bch2_keylist_init(&as->old_keys, as->_old_keys);
1036         bch2_keylist_init(&as->new_keys, as->_new_keys);
1037         bch2_keylist_init(&as->parent_keys, as->inline_keys);
1038
1039         mutex_lock(&c->btree_interior_update_lock);
1040         list_add_tail(&as->list, &c->btree_interior_update_list);
1041         mutex_unlock(&c->btree_interior_update_lock);
1042
1043         /*
1044          * We don't want to allocate if we're in an error state, that can cause
1045          * deadlock on emergency shutdown due to open buckets getting stuck in
1046          * the btree_reserve_cache after allocator shutdown has cleared it out.
1047          * This check needs to come after adding us to the btree_interior_update
1048          * list but before calling bch2_btree_reserve_get, to synchronize with
1049          * __bch2_fs_read_only().
1050          */
1051         ret = bch2_journal_error(&c->journal);
1052         if (ret)
1053                 goto err;
1054
1055         bch2_trans_unlock(trans);
1056
1057         ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
1058                                       BTREE_UPDATE_JOURNAL_RES,
1059                                       journal_flags);
1060         if (ret) {
1061                 bch2_btree_update_free(as);
1062                 trace_trans_restart_journal_preres_get(trans->fn, _RET_IP_);
1063                 btree_trans_restart(trans);
1064                 return ERR_PTR(ret);
1065         }
1066
1067         ret = bch2_disk_reservation_get(c, &as->disk_res,
1068                         (nr_nodes[0] + nr_nodes[1]) * btree_sectors(c),
1069                         c->opts.metadata_replicas,
1070                         disk_res_flags);
1071         if (ret)
1072                 goto err;
1073
1074         ret = bch2_btree_reserve_get(as, nr_nodes, flags);
1075         if (ret)
1076                 goto err;
1077
1078         if (!bch2_trans_relock(trans)) {
1079                 ret = -EINTR;
1080                 goto err;
1081         }
1082
1083         return as;
1084 err:
1085         bch2_btree_update_free(as);
1086         return ERR_PTR(ret);
1087 }
1088
1089 /* Btree root updates: */
1090
1091 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1092 {
1093         /* Root nodes cannot be reaped */
1094         mutex_lock(&c->btree_cache.lock);
1095         list_del_init(&b->list);
1096         mutex_unlock(&c->btree_cache.lock);
1097
1098         if (b->c.level)
1099                 six_lock_pcpu_alloc(&b->c.lock);
1100         else
1101                 six_lock_pcpu_free(&b->c.lock);
1102
1103         mutex_lock(&c->btree_root_lock);
1104         BUG_ON(btree_node_root(c, b) &&
1105                (b->c.level < btree_node_root(c, b)->c.level ||
1106                 !btree_node_dying(btree_node_root(c, b))));
1107
1108         btree_node_root(c, b) = b;
1109         mutex_unlock(&c->btree_root_lock);
1110
1111         bch2_recalc_btree_reserve(c);
1112 }
1113
1114 /**
1115  * bch_btree_set_root - update the root in memory and on disk
1116  *
1117  * To ensure forward progress, the current task must not be holding any
1118  * btree node write locks. However, you must hold an intent lock on the
1119  * old root.
1120  *
1121  * Note: This allocates a journal entry but doesn't add any keys to
1122  * it.  All the btree roots are part of every journal write, so there
1123  * is nothing new to be done.  This just guarantees that there is a
1124  * journal write.
1125  */
1126 static void bch2_btree_set_root(struct btree_update *as,
1127                                 struct btree_trans *trans,
1128                                 struct btree_path *path,
1129                                 struct btree *b)
1130 {
1131         struct bch_fs *c = as->c;
1132         struct btree *old;
1133
1134         trace_btree_set_root(c, b);
1135         BUG_ON(!b->written);
1136
1137         old = btree_node_root(c, b);
1138
1139         /*
1140          * Ensure no one is using the old root while we switch to the
1141          * new root:
1142          */
1143         bch2_btree_node_lock_write(trans, path, old);
1144
1145         bch2_btree_set_root_inmem(c, b);
1146
1147         btree_update_updated_root(as, b);
1148
1149         /*
1150          * Unlock old root after new root is visible:
1151          *
1152          * The new root isn't persistent, but that's ok: we still have
1153          * an intent lock on the new root, and any updates that would
1154          * depend on the new root would have to update the new root.
1155          */
1156         bch2_btree_node_unlock_write(trans, path, old);
1157 }
1158
1159 /* Interior node updates: */
1160
1161 static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
1162                                         struct btree_trans *trans,
1163                                         struct btree_path *path,
1164                                         struct btree *b,
1165                                         struct btree_node_iter *node_iter,
1166                                         struct bkey_i *insert)
1167 {
1168         struct bch_fs *c = as->c;
1169         struct bkey_packed *k;
1170         const char *invalid;
1171
1172         BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 &&
1173                !btree_ptr_sectors_written(insert));
1174
1175         if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
1176                 bch2_journal_key_overwritten(c, b->c.btree_id, b->c.level, insert->k.p);
1177
1178         invalid = bch2_bkey_invalid(c, bkey_i_to_s_c(insert), btree_node_type(b)) ?:
1179                 bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert));
1180         if (invalid) {
1181                 struct printbuf buf = PRINTBUF;
1182
1183                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
1184                 bch2_fs_inconsistent(c, "inserting invalid bkey %s: %s", buf.buf, invalid);
1185                 printbuf_exit(&buf);
1186                 dump_stack();
1187         }
1188
1189         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1190                ARRAY_SIZE(as->journal_entries));
1191
1192         as->journal_u64s +=
1193                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1194                                   BCH_JSET_ENTRY_btree_keys,
1195                                   b->c.btree_id, b->c.level,
1196                                   insert, insert->k.u64s);
1197
1198         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1199                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1200                 bch2_btree_node_iter_advance(node_iter, b);
1201
1202         bch2_btree_bset_insert_key(trans, path, b, node_iter, insert);
1203         set_btree_node_dirty_acct(c, b);
1204         set_btree_node_need_write(b);
1205 }
1206
1207 static void
1208 __bch2_btree_insert_keys_interior(struct btree_update *as,
1209                                   struct btree_trans *trans,
1210                                   struct btree_path *path,
1211                                   struct btree *b,
1212                                   struct btree_node_iter node_iter,
1213                                   struct keylist *keys)
1214 {
1215         struct bkey_i *insert = bch2_keylist_front(keys);
1216         struct bkey_packed *k;
1217
1218         BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1219
1220         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1221                (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1222                 ;
1223
1224         while (!bch2_keylist_empty(keys)) {
1225                 bch2_insert_fixup_btree_ptr(as, trans, path, b,
1226                                 &node_iter, bch2_keylist_front(keys));
1227                 bch2_keylist_pop_front(keys);
1228         }
1229 }
1230
1231 /*
1232  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1233  * node)
1234  */
1235 static struct btree *__btree_split_node(struct btree_update *as,
1236                                         struct btree *n1)
1237 {
1238         struct bkey_format_state s;
1239         size_t nr_packed = 0, nr_unpacked = 0;
1240         struct btree *n2;
1241         struct bset *set1, *set2;
1242         struct bkey_packed *k, *set2_start, *set2_end, *out, *prev = NULL;
1243         struct bpos n1_pos;
1244
1245         n2 = bch2_btree_node_alloc(as, n1->c.level);
1246         bch2_btree_update_add_new_node(as, n2);
1247
1248         n2->data->max_key       = n1->data->max_key;
1249         n2->data->format        = n1->format;
1250         SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data));
1251         n2->key.k.p = n1->key.k.p;
1252
1253         set1 = btree_bset_first(n1);
1254         set2 = btree_bset_first(n2);
1255
1256         /*
1257          * Has to be a linear search because we don't have an auxiliary
1258          * search tree yet
1259          */
1260         k = set1->start;
1261         while (1) {
1262                 struct bkey_packed *n = bkey_next(k);
1263
1264                 if (n == vstruct_last(set1))
1265                         break;
1266                 if (k->_data - set1->_data >= (le16_to_cpu(set1->u64s) * 3) / 5)
1267                         break;
1268
1269                 if (bkey_packed(k))
1270                         nr_packed++;
1271                 else
1272                         nr_unpacked++;
1273
1274                 prev = k;
1275                 k = n;
1276         }
1277
1278         BUG_ON(!prev);
1279         set2_start      = k;
1280         set2_end        = vstruct_last(set1);
1281
1282         set1->u64s = cpu_to_le16((u64 *) set2_start - set1->_data);
1283         set_btree_bset_end(n1, n1->set);
1284
1285         n1->nr.live_u64s        = le16_to_cpu(set1->u64s);
1286         n1->nr.bset_u64s[0]     = le16_to_cpu(set1->u64s);
1287         n1->nr.packed_keys      = nr_packed;
1288         n1->nr.unpacked_keys    = nr_unpacked;
1289
1290         n1_pos = bkey_unpack_pos(n1, prev);
1291         if (as->c->sb.version < bcachefs_metadata_version_snapshot)
1292                 n1_pos.snapshot = U32_MAX;
1293
1294         btree_set_max(n1, n1_pos);
1295         btree_set_min(n2, bpos_successor(n1->key.k.p));
1296
1297         bch2_bkey_format_init(&s);
1298         bch2_bkey_format_add_pos(&s, n2->data->min_key);
1299         bch2_bkey_format_add_pos(&s, n2->data->max_key);
1300
1301         for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1302                 struct bkey uk = bkey_unpack_key(n1, k);
1303                 bch2_bkey_format_add_key(&s, &uk);
1304         }
1305
1306         n2->data->format = bch2_bkey_format_done(&s);
1307         btree_node_set_format(n2, n2->data->format);
1308
1309         out = set2->start;
1310         memset(&n2->nr, 0, sizeof(n2->nr));
1311
1312         for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1313                 BUG_ON(!bch2_bkey_transform(&n2->format, out, bkey_packed(k)
1314                                        ? &n1->format : &bch2_bkey_format_current, k));
1315                 out->format = KEY_FORMAT_LOCAL_BTREE;
1316                 btree_keys_account_key_add(&n2->nr, 0, out);
1317                 out = bkey_next(out);
1318         }
1319
1320         set2->u64s = cpu_to_le16((u64 *) out - set2->_data);
1321         set_btree_bset_end(n2, n2->set);
1322
1323         BUG_ON(!set1->u64s);
1324         BUG_ON(!set2->u64s);
1325
1326         btree_node_reset_sib_u64s(n1);
1327         btree_node_reset_sib_u64s(n2);
1328
1329         bch2_verify_btree_nr_keys(n1);
1330         bch2_verify_btree_nr_keys(n2);
1331
1332         if (n1->c.level) {
1333                 btree_node_interior_verify(as->c, n1);
1334                 btree_node_interior_verify(as->c, n2);
1335         }
1336
1337         return n2;
1338 }
1339
1340 /*
1341  * For updates to interior nodes, we've got to do the insert before we split
1342  * because the stuff we're inserting has to be inserted atomically. Post split,
1343  * the keys might have to go in different nodes and the split would no longer be
1344  * atomic.
1345  *
1346  * Worse, if the insert is from btree node coalescing, if we do the insert after
1347  * we do the split (and pick the pivot) - the pivot we pick might be between
1348  * nodes that were coalesced, and thus in the middle of a child node post
1349  * coalescing:
1350  */
1351 static void btree_split_insert_keys(struct btree_update *as,
1352                                     struct btree_trans *trans,
1353                                     struct btree_path *path,
1354                                     struct btree *b,
1355                                     struct keylist *keys)
1356 {
1357         struct btree_node_iter node_iter;
1358         struct bkey_i *k = bch2_keylist_front(keys);
1359         struct bkey_packed *src, *dst, *n;
1360         struct bset *i;
1361
1362         bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
1363
1364         __bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
1365
1366         /*
1367          * We can't tolerate whiteouts here - with whiteouts there can be
1368          * duplicate keys, and it would be rather bad if we picked a duplicate
1369          * for the pivot:
1370          */
1371         i = btree_bset_first(b);
1372         src = dst = i->start;
1373         while (src != vstruct_last(i)) {
1374                 n = bkey_next(src);
1375                 if (!bkey_deleted(src)) {
1376                         memmove_u64s_down(dst, src, src->u64s);
1377                         dst = bkey_next(dst);
1378                 }
1379                 src = n;
1380         }
1381
1382         /* Also clear out the unwritten whiteouts area: */
1383         b->whiteout_u64s = 0;
1384
1385         i->u64s = cpu_to_le16((u64 *) dst - i->_data);
1386         set_btree_bset_end(b, b->set);
1387
1388         BUG_ON(b->nsets != 1 ||
1389                b->nr.live_u64s != le16_to_cpu(btree_bset_first(b)->u64s));
1390
1391         btree_node_interior_verify(as->c, b);
1392 }
1393
1394 static void btree_split(struct btree_update *as, struct btree_trans *trans,
1395                         struct btree_path *path, struct btree *b,
1396                         struct keylist *keys, unsigned flags)
1397 {
1398         struct bch_fs *c = as->c;
1399         struct btree *parent = btree_node_parent(path, b);
1400         struct btree *n1, *n2 = NULL, *n3 = NULL;
1401         u64 start_time = local_clock();
1402
1403         BUG_ON(!parent && (b != btree_node_root(c, b)));
1404         BUG_ON(!btree_node_intent_locked(path, btree_node_root(c, b)->c.level));
1405
1406         bch2_btree_interior_update_will_free_node(as, b);
1407
1408         n1 = bch2_btree_node_alloc_replacement(as, b);
1409         bch2_btree_update_add_new_node(as, n1);
1410
1411         if (keys)
1412                 btree_split_insert_keys(as, trans, path, n1, keys);
1413
1414         if (bset_u64s(&n1->set[0]) > BTREE_SPLIT_THRESHOLD(c)) {
1415                 trace_btree_split(c, b);
1416
1417                 n2 = __btree_split_node(as, n1);
1418
1419                 bch2_btree_build_aux_trees(n2);
1420                 bch2_btree_build_aux_trees(n1);
1421                 six_unlock_write(&n2->c.lock);
1422                 six_unlock_write(&n1->c.lock);
1423
1424                 bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1425                 bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0);
1426
1427                 /*
1428                  * Note that on recursive parent_keys == keys, so we
1429                  * can't start adding new keys to parent_keys before emptying it
1430                  * out (which we did with btree_split_insert_keys() above)
1431                  */
1432                 bch2_keylist_add(&as->parent_keys, &n1->key);
1433                 bch2_keylist_add(&as->parent_keys, &n2->key);
1434
1435                 if (!parent) {
1436                         /* Depth increases, make a new root */
1437                         n3 = __btree_root_alloc(as, b->c.level + 1);
1438
1439                         n3->sib_u64s[0] = U16_MAX;
1440                         n3->sib_u64s[1] = U16_MAX;
1441
1442                         btree_split_insert_keys(as, trans, path, n3, &as->parent_keys);
1443
1444                         bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0);
1445                 }
1446         } else {
1447                 trace_btree_compact(c, b);
1448
1449                 bch2_btree_build_aux_trees(n1);
1450                 six_unlock_write(&n1->c.lock);
1451
1452                 bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1453
1454                 if (parent)
1455                         bch2_keylist_add(&as->parent_keys, &n1->key);
1456         }
1457
1458         /* New nodes all written, now make them visible: */
1459
1460         if (parent) {
1461                 /* Split a non root node */
1462                 bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1463         } else if (n3) {
1464                 bch2_btree_set_root(as, trans, path, n3);
1465         } else {
1466                 /* Root filled up but didn't need to be split */
1467                 bch2_btree_set_root(as, trans, path, n1);
1468         }
1469
1470         bch2_btree_update_get_open_buckets(as, n1);
1471         if (n2)
1472                 bch2_btree_update_get_open_buckets(as, n2);
1473         if (n3)
1474                 bch2_btree_update_get_open_buckets(as, n3);
1475
1476         /* Successful split, update the path to point to the new nodes: */
1477
1478         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1479         if (n3)
1480                 bch2_trans_node_add(trans, n3);
1481         if (n2)
1482                 bch2_trans_node_add(trans, n2);
1483         bch2_trans_node_add(trans, n1);
1484
1485         /*
1486          * The old node must be freed (in memory) _before_ unlocking the new
1487          * nodes - else another thread could re-acquire a read lock on the old
1488          * node after another thread has locked and updated the new node, thus
1489          * seeing stale data:
1490          */
1491         bch2_btree_node_free_inmem(trans, b);
1492
1493         if (n3)
1494                 six_unlock_intent(&n3->c.lock);
1495         if (n2)
1496                 six_unlock_intent(&n2->c.lock);
1497         six_unlock_intent(&n1->c.lock);
1498
1499         bch2_trans_verify_locks(trans);
1500
1501         bch2_time_stats_update(&c->times[n2
1502                                ? BCH_TIME_btree_node_split
1503                                : BCH_TIME_btree_node_compact],
1504                                start_time);
1505 }
1506
1507 static void
1508 bch2_btree_insert_keys_interior(struct btree_update *as,
1509                                 struct btree_trans *trans,
1510                                 struct btree_path *path,
1511                                 struct btree *b,
1512                                 struct keylist *keys)
1513 {
1514         struct btree_path *linked;
1515
1516         __bch2_btree_insert_keys_interior(as, trans, path, b,
1517                                           path->l[b->c.level].iter, keys);
1518
1519         btree_update_updated_node(as, b);
1520
1521         trans_for_each_path_with_node(trans, b, linked)
1522                 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1523
1524         bch2_trans_verify_paths(trans);
1525 }
1526
1527 /**
1528  * bch_btree_insert_node - insert bkeys into a given btree node
1529  *
1530  * @iter:               btree iterator
1531  * @keys:               list of keys to insert
1532  * @hook:               insert callback
1533  * @persistent:         if not null, @persistent will wait on journal write
1534  *
1535  * Inserts as many keys as it can into a given btree node, splitting it if full.
1536  * If a split occurred, this function will return early. This can only happen
1537  * for leaf nodes -- inserts into interior nodes have to be atomic.
1538  */
1539 static void bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans,
1540                                    struct btree_path *path, struct btree *b,
1541                                    struct keylist *keys, unsigned flags)
1542 {
1543         struct bch_fs *c = as->c;
1544         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1545         int old_live_u64s = b->nr.live_u64s;
1546         int live_u64s_added, u64s_added;
1547
1548         lockdep_assert_held(&c->gc_lock);
1549         BUG_ON(!btree_node_intent_locked(path, btree_node_root(c, b)->c.level));
1550         BUG_ON(!b->c.level);
1551         BUG_ON(!as || as->b);
1552         bch2_verify_keylist_sorted(keys);
1553
1554         bch2_btree_node_lock_for_insert(trans, path, b);
1555
1556         if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1557                 bch2_btree_node_unlock_write(trans, path, b);
1558                 goto split;
1559         }
1560
1561         btree_node_interior_verify(c, b);
1562
1563         bch2_btree_insert_keys_interior(as, trans, path, b, keys);
1564
1565         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1566         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1567
1568         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1569                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1570         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1571                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1572
1573         if (u64s_added > live_u64s_added &&
1574             bch2_maybe_compact_whiteouts(c, b))
1575                 bch2_trans_node_reinit_iter(trans, b);
1576
1577         bch2_btree_node_unlock_write(trans, path, b);
1578
1579         btree_node_interior_verify(c, b);
1580         return;
1581 split:
1582         btree_split(as, trans, path, b, keys, flags);
1583 }
1584
1585 int bch2_btree_split_leaf(struct btree_trans *trans,
1586                           struct btree_path *path,
1587                           unsigned flags)
1588 {
1589         struct btree *b = path_l(path)->b;
1590         struct btree_update *as;
1591         unsigned l;
1592         int ret = 0;
1593
1594         as = bch2_btree_update_start(trans, path, path->level,
1595                                      true, flags);
1596         if (IS_ERR(as))
1597                 return PTR_ERR(as);
1598
1599         btree_split(as, trans, path, b, NULL, flags);
1600         bch2_btree_update_done(as);
1601
1602         for (l = path->level + 1; btree_path_node(path, l) && !ret; l++)
1603                 ret = bch2_foreground_maybe_merge(trans, path, l, flags);
1604
1605         return ret;
1606 }
1607
1608 int __bch2_foreground_maybe_merge(struct btree_trans *trans,
1609                                   struct btree_path *path,
1610                                   unsigned level,
1611                                   unsigned flags,
1612                                   enum btree_node_sibling sib)
1613 {
1614         struct bch_fs *c = trans->c;
1615         struct btree_path *sib_path = NULL;
1616         struct btree_update *as;
1617         struct bkey_format_state new_s;
1618         struct bkey_format new_f;
1619         struct bkey_i delete;
1620         struct btree *b, *m, *n, *prev, *next, *parent;
1621         struct bpos sib_pos;
1622         size_t sib_u64s;
1623         u64 start_time = local_clock();
1624         int ret = 0;
1625
1626         BUG_ON(!path->should_be_locked);
1627         BUG_ON(!btree_node_locked(path, level));
1628
1629         b = path->l[level].b;
1630
1631         if ((sib == btree_prev_sib && !bpos_cmp(b->data->min_key, POS_MIN)) ||
1632             (sib == btree_next_sib && !bpos_cmp(b->data->max_key, SPOS_MAX))) {
1633                 b->sib_u64s[sib] = U16_MAX;
1634                 return 0;
1635         }
1636
1637         sib_pos = sib == btree_prev_sib
1638                 ? bpos_predecessor(b->data->min_key)
1639                 : bpos_successor(b->data->max_key);
1640
1641         sib_path = bch2_path_get(trans, path->btree_id, sib_pos,
1642                                  U8_MAX, level, BTREE_ITER_INTENT, _THIS_IP_);
1643         ret = bch2_btree_path_traverse(trans, sib_path, false);
1644         if (ret)
1645                 goto err;
1646
1647         sib_path->should_be_locked = true;
1648
1649         m = sib_path->l[level].b;
1650
1651         if (btree_node_parent(path, b) !=
1652             btree_node_parent(sib_path, m)) {
1653                 b->sib_u64s[sib] = U16_MAX;
1654                 goto out;
1655         }
1656
1657         if (sib == btree_prev_sib) {
1658                 prev = m;
1659                 next = b;
1660         } else {
1661                 prev = b;
1662                 next = m;
1663         }
1664
1665         if (bkey_cmp(bpos_successor(prev->data->max_key), next->data->min_key)) {
1666                 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1667
1668                 bch2_bpos_to_text(&buf1, prev->data->max_key);
1669                 bch2_bpos_to_text(&buf2, next->data->min_key);
1670                 bch_err(c,
1671                         "btree topology error in btree merge:\n"
1672                         "  prev ends at   %s\n"
1673                         "  next starts at %s",
1674                         buf1.buf, buf2.buf);
1675                 printbuf_exit(&buf1);
1676                 printbuf_exit(&buf2);
1677                 bch2_topology_error(c);
1678                 ret = -EIO;
1679                 goto err;
1680         }
1681
1682         bch2_bkey_format_init(&new_s);
1683         bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
1684         __bch2_btree_calc_format(&new_s, prev);
1685         __bch2_btree_calc_format(&new_s, next);
1686         bch2_bkey_format_add_pos(&new_s, next->data->max_key);
1687         new_f = bch2_bkey_format_done(&new_s);
1688
1689         sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1690                 btree_node_u64s_with_format(m, &new_f);
1691
1692         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1693                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1694                 sib_u64s /= 2;
1695                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1696         }
1697
1698         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1699         sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
1700         b->sib_u64s[sib] = sib_u64s;
1701
1702         if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
1703                 goto out;
1704
1705         parent = btree_node_parent(path, b);
1706         as = bch2_btree_update_start(trans, path, level, false,
1707                          BTREE_INSERT_NOFAIL|
1708                          BTREE_INSERT_USE_RESERVE|
1709                          flags);
1710         ret = PTR_ERR_OR_ZERO(as);
1711         if (ret)
1712                 goto err;
1713
1714         trace_btree_merge(c, b);
1715
1716         bch2_btree_interior_update_will_free_node(as, b);
1717         bch2_btree_interior_update_will_free_node(as, m);
1718
1719         n = bch2_btree_node_alloc(as, b->c.level);
1720         bch2_btree_update_add_new_node(as, n);
1721
1722         SET_BTREE_NODE_SEQ(n->data,
1723                            max(BTREE_NODE_SEQ(b->data),
1724                                BTREE_NODE_SEQ(m->data)) + 1);
1725
1726         btree_set_min(n, prev->data->min_key);
1727         btree_set_max(n, next->data->max_key);
1728         n->data->format         = new_f;
1729
1730         btree_node_set_format(n, new_f);
1731
1732         bch2_btree_sort_into(c, n, prev);
1733         bch2_btree_sort_into(c, n, next);
1734
1735         bch2_btree_build_aux_trees(n);
1736         six_unlock_write(&n->c.lock);
1737
1738         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1739
1740         bkey_init(&delete.k);
1741         delete.k.p = prev->key.k.p;
1742         bch2_keylist_add(&as->parent_keys, &delete);
1743         bch2_keylist_add(&as->parent_keys, &n->key);
1744
1745         bch2_trans_verify_paths(trans);
1746
1747         bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1748
1749         bch2_trans_verify_paths(trans);
1750
1751         bch2_btree_update_get_open_buckets(as, n);
1752
1753         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1754         six_lock_increment(&m->c.lock, SIX_LOCK_intent);
1755
1756         bch2_trans_node_add(trans, n);
1757
1758         bch2_trans_verify_paths(trans);
1759
1760         bch2_btree_node_free_inmem(trans, b);
1761         bch2_btree_node_free_inmem(trans, m);
1762
1763         six_unlock_intent(&n->c.lock);
1764
1765         bch2_btree_update_done(as);
1766
1767         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time);
1768 out:
1769 err:
1770         bch2_path_put(trans, sib_path, true);
1771         bch2_trans_verify_locks(trans);
1772         return ret;
1773 }
1774
1775 /**
1776  * bch_btree_node_rewrite - Rewrite/move a btree node
1777  */
1778 int bch2_btree_node_rewrite(struct btree_trans *trans,
1779                             struct btree_iter *iter,
1780                             struct btree *b,
1781                             unsigned flags)
1782 {
1783         struct bch_fs *c = trans->c;
1784         struct btree *n, *parent;
1785         struct btree_update *as;
1786         int ret;
1787
1788         flags |= BTREE_INSERT_NOFAIL;
1789
1790         parent = btree_node_parent(iter->path, b);
1791         as = bch2_btree_update_start(trans, iter->path, b->c.level,
1792                                      false, flags);
1793         ret = PTR_ERR_OR_ZERO(as);
1794         if (ret) {
1795                 trace_btree_gc_rewrite_node_fail(c, b);
1796                 goto out;
1797         }
1798
1799         bch2_btree_interior_update_will_free_node(as, b);
1800
1801         n = bch2_btree_node_alloc_replacement(as, b);
1802         bch2_btree_update_add_new_node(as, n);
1803
1804         bch2_btree_build_aux_trees(n);
1805         six_unlock_write(&n->c.lock);
1806
1807         trace_btree_gc_rewrite_node(c, b);
1808
1809         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1810
1811         if (parent) {
1812                 bch2_keylist_add(&as->parent_keys, &n->key);
1813                 bch2_btree_insert_node(as, trans, iter->path, parent,
1814                                        &as->parent_keys, flags);
1815         } else {
1816                 bch2_btree_set_root(as, trans, iter->path, n);
1817         }
1818
1819         bch2_btree_update_get_open_buckets(as, n);
1820
1821         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1822         bch2_trans_node_add(trans, n);
1823         bch2_btree_node_free_inmem(trans, b);
1824         six_unlock_intent(&n->c.lock);
1825
1826         bch2_btree_update_done(as);
1827 out:
1828         bch2_btree_path_downgrade(iter->path);
1829         return ret;
1830 }
1831
1832 struct async_btree_rewrite {
1833         struct bch_fs           *c;
1834         struct work_struct      work;
1835         enum btree_id           btree_id;
1836         unsigned                level;
1837         struct bpos             pos;
1838         __le64                  seq;
1839 };
1840
1841 static int async_btree_node_rewrite_trans(struct btree_trans *trans,
1842                                           struct async_btree_rewrite *a)
1843 {
1844         struct btree_iter iter;
1845         struct btree *b;
1846         int ret;
1847
1848         bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos,
1849                                   BTREE_MAX_DEPTH, a->level, 0);
1850         b = bch2_btree_iter_peek_node(&iter);
1851         ret = PTR_ERR_OR_ZERO(b);
1852         if (ret)
1853                 goto out;
1854
1855         if (!b || b->data->keys.seq != a->seq)
1856                 goto out;
1857
1858         ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
1859 out :
1860         bch2_trans_iter_exit(trans, &iter);
1861
1862         return ret;
1863 }
1864
1865 void async_btree_node_rewrite_work(struct work_struct *work)
1866 {
1867         struct async_btree_rewrite *a =
1868                 container_of(work, struct async_btree_rewrite, work);
1869         struct bch_fs *c = a->c;
1870
1871         bch2_trans_do(c, NULL, NULL, 0,
1872                       async_btree_node_rewrite_trans(&trans, a));
1873         percpu_ref_put(&c->writes);
1874         kfree(a);
1875 }
1876
1877 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
1878 {
1879         struct async_btree_rewrite *a;
1880
1881         if (!percpu_ref_tryget(&c->writes))
1882                 return;
1883
1884         a = kmalloc(sizeof(*a), GFP_NOFS);
1885         if (!a) {
1886                 percpu_ref_put(&c->writes);
1887                 return;
1888         }
1889
1890         a->c            = c;
1891         a->btree_id     = b->c.btree_id;
1892         a->level        = b->c.level;
1893         a->pos          = b->key.k.p;
1894         a->seq          = b->data->keys.seq;
1895
1896         INIT_WORK(&a->work, async_btree_node_rewrite_work);
1897         queue_work(c->btree_interior_update_worker, &a->work);
1898 }
1899
1900 static int __bch2_btree_node_update_key(struct btree_trans *trans,
1901                                         struct btree_iter *iter,
1902                                         struct btree *b, struct btree *new_hash,
1903                                         struct bkey_i *new_key,
1904                                         bool skip_triggers)
1905 {
1906         struct bch_fs *c = trans->c;
1907         struct btree_iter iter2 = { NULL };
1908         struct btree *parent;
1909         int ret;
1910
1911         if (!skip_triggers) {
1912                 ret = bch2_trans_mark_new(trans, new_key, 0);
1913                 if (ret)
1914                         return ret;
1915
1916                 ret = bch2_trans_mark_old(trans, bkey_i_to_s_c(&b->key), 0);
1917                 if (ret)
1918                         return ret;
1919         }
1920
1921         if (new_hash) {
1922                 bkey_copy(&new_hash->key, new_key);
1923                 ret = bch2_btree_node_hash_insert(&c->btree_cache,
1924                                 new_hash, b->c.level, b->c.btree_id);
1925                 BUG_ON(ret);
1926         }
1927
1928         parent = btree_node_parent(iter->path, b);
1929         if (parent) {
1930                 bch2_trans_copy_iter(&iter2, iter);
1931
1932                 iter2.path = bch2_btree_path_make_mut(trans, iter2.path,
1933                                 iter2.flags & BTREE_ITER_INTENT,
1934                                 _THIS_IP_);
1935
1936                 BUG_ON(iter2.path->level != b->c.level);
1937                 BUG_ON(bpos_cmp(iter2.path->pos, new_key->k.p));
1938
1939                 btree_node_unlock(iter2.path, iter2.path->level);
1940                 path_l(iter2.path)->b = BTREE_ITER_NO_NODE_UP;
1941                 iter2.path->level++;
1942                 btree_path_set_dirty(iter2.path, BTREE_ITER_NEED_TRAVERSE);
1943
1944                 bch2_btree_path_check_sort(trans, iter2.path, 0);
1945
1946                 ret   = bch2_btree_iter_traverse(&iter2) ?:
1947                         bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_NORUN);
1948                 if (ret)
1949                         goto err;
1950         } else {
1951                 BUG_ON(btree_node_root(c, b) != b);
1952
1953                 ret = darray_make_room(trans->extra_journal_entries,
1954                                        jset_u64s(new_key->k.u64s));
1955                 if (ret)
1956                         return ret;
1957
1958                 journal_entry_set((void *) &darray_top(trans->extra_journal_entries),
1959                                   BCH_JSET_ENTRY_btree_root,
1960                                   b->c.btree_id, b->c.level,
1961                                   new_key, new_key->k.u64s);
1962                 trans->extra_journal_entries.nr += jset_u64s(new_key->k.u64s);
1963         }
1964
1965         ret = bch2_trans_commit(trans, NULL, NULL,
1966                                 BTREE_INSERT_NOFAIL|
1967                                 BTREE_INSERT_NOCHECK_RW|
1968                                 BTREE_INSERT_USE_RESERVE|
1969                                 BTREE_INSERT_JOURNAL_RECLAIM|
1970                                 JOURNAL_WATERMARK_reserved);
1971         if (ret)
1972                 goto err;
1973
1974         bch2_btree_node_lock_write(trans, iter->path, b);
1975
1976         if (new_hash) {
1977                 mutex_lock(&c->btree_cache.lock);
1978                 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
1979                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1980
1981                 bkey_copy(&b->key, new_key);
1982                 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1983                 BUG_ON(ret);
1984                 mutex_unlock(&c->btree_cache.lock);
1985         } else {
1986                 bkey_copy(&b->key, new_key);
1987         }
1988
1989         bch2_btree_node_unlock_write(trans, iter->path, b);
1990 out:
1991         bch2_trans_iter_exit(trans, &iter2);
1992         return ret;
1993 err:
1994         if (new_hash) {
1995                 mutex_lock(&c->btree_cache.lock);
1996                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1997                 mutex_unlock(&c->btree_cache.lock);
1998         }
1999         goto out;
2000 }
2001
2002 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
2003                                struct btree *b, struct bkey_i *new_key,
2004                                bool skip_triggers)
2005 {
2006         struct bch_fs *c = trans->c;
2007         struct btree *new_hash = NULL;
2008         struct btree_path *path = iter->path;
2009         struct closure cl;
2010         int ret = 0;
2011
2012         if (!btree_node_intent_locked(path, b->c.level) &&
2013             !bch2_btree_path_upgrade(trans, path, b->c.level + 1)) {
2014                 btree_trans_restart(trans);
2015                 return -EINTR;
2016         }
2017
2018         closure_init_stack(&cl);
2019
2020         /*
2021          * check btree_ptr_hash_val() after @b is locked by
2022          * btree_iter_traverse():
2023          */
2024         if (btree_ptr_hash_val(new_key) != b->hash_val) {
2025                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2026                 if (ret) {
2027                         bch2_trans_unlock(trans);
2028                         closure_sync(&cl);
2029                         if (!bch2_trans_relock(trans))
2030                                 return -EINTR;
2031                 }
2032
2033                 new_hash = bch2_btree_node_mem_alloc(c, false);
2034         }
2035
2036         path->intent_ref++;
2037         ret = __bch2_btree_node_update_key(trans, iter, b, new_hash,
2038                                            new_key, skip_triggers);
2039         --path->intent_ref;
2040
2041         if (new_hash) {
2042                 mutex_lock(&c->btree_cache.lock);
2043                 list_move(&new_hash->list, &c->btree_cache.freeable);
2044                 mutex_unlock(&c->btree_cache.lock);
2045
2046                 six_unlock_write(&new_hash->c.lock);
2047                 six_unlock_intent(&new_hash->c.lock);
2048         }
2049         closure_sync(&cl);
2050         bch2_btree_cache_cannibalize_unlock(c);
2051         return ret;
2052 }
2053
2054 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
2055                                         struct btree *b, struct bkey_i *new_key,
2056                                         bool skip_triggers)
2057 {
2058         struct btree_iter iter;
2059         int ret;
2060
2061         bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p,
2062                                   BTREE_MAX_DEPTH, b->c.level,
2063                                   BTREE_ITER_INTENT);
2064         ret = bch2_btree_iter_traverse(&iter);
2065         if (ret)
2066                 goto out;
2067
2068         /* has node been freed? */
2069         if (iter.path->l[b->c.level].b != b) {
2070                 /* node has been freed: */
2071                 BUG_ON(!btree_node_dying(b));
2072                 goto out;
2073         }
2074
2075         BUG_ON(!btree_node_hashed(b));
2076
2077         ret = bch2_btree_node_update_key(trans, &iter, b, new_key, skip_triggers);
2078 out:
2079         bch2_trans_iter_exit(trans, &iter);
2080         return ret;
2081 }
2082
2083 /* Init code: */
2084
2085 /*
2086  * Only for filesystem bringup, when first reading the btree roots or allocating
2087  * btree roots when initializing a new filesystem:
2088  */
2089 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2090 {
2091         BUG_ON(btree_node_root(c, b));
2092
2093         bch2_btree_set_root_inmem(c, b);
2094 }
2095
2096 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2097 {
2098         struct closure cl;
2099         struct btree *b;
2100         int ret;
2101
2102         closure_init_stack(&cl);
2103
2104         do {
2105                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2106                 closure_sync(&cl);
2107         } while (ret);
2108
2109         b = bch2_btree_node_mem_alloc(c, false);
2110         bch2_btree_cache_cannibalize_unlock(c);
2111
2112         set_btree_node_fake(b);
2113         set_btree_node_need_rewrite(b);
2114         b->c.level      = 0;
2115         b->c.btree_id   = id;
2116
2117         bkey_btree_ptr_init(&b->key);
2118         b->key.k.p = SPOS_MAX;
2119         *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2120
2121         bch2_bset_init_first(b, &b->data->keys);
2122         bch2_btree_build_aux_trees(b);
2123
2124         b->data->flags = 0;
2125         btree_set_min(b, POS_MIN);
2126         btree_set_max(b, SPOS_MAX);
2127         b->data->format = bch2_btree_calc_format(b);
2128         btree_node_set_format(b, b->data->format);
2129
2130         ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2131                                           b->c.level, b->c.btree_id);
2132         BUG_ON(ret);
2133
2134         bch2_btree_set_root_inmem(c, b);
2135
2136         six_unlock_write(&b->c.lock);
2137         six_unlock_intent(&b->c.lock);
2138 }
2139
2140 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2141 {
2142         struct btree_update *as;
2143
2144         mutex_lock(&c->btree_interior_update_lock);
2145         list_for_each_entry(as, &c->btree_interior_update_list, list)
2146                 pr_buf(out, "%p m %u w %u r %u j %llu\n",
2147                        as,
2148                        as->mode,
2149                        as->nodes_written,
2150                        atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2151                        as->journal.seq);
2152         mutex_unlock(&c->btree_interior_update_lock);
2153 }
2154
2155 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *c)
2156 {
2157         size_t ret = 0;
2158         struct list_head *i;
2159
2160         mutex_lock(&c->btree_interior_update_lock);
2161         list_for_each(i, &c->btree_interior_update_list)
2162                 ret++;
2163         mutex_unlock(&c->btree_interior_update_lock);
2164
2165         return ret;
2166 }
2167
2168 void bch2_journal_entries_to_btree_roots(struct bch_fs *c, struct jset *jset)
2169 {
2170         struct btree_root *r;
2171         struct jset_entry *entry;
2172
2173         mutex_lock(&c->btree_root_lock);
2174
2175         vstruct_for_each(jset, entry)
2176                 if (entry->type == BCH_JSET_ENTRY_btree_root) {
2177                         r = &c->btree_roots[entry->btree_id];
2178                         r->level = entry->level;
2179                         r->alive = true;
2180                         bkey_copy(&r->key, &entry->start[0]);
2181                 }
2182
2183         mutex_unlock(&c->btree_root_lock);
2184 }
2185
2186 struct jset_entry *
2187 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2188                                     struct jset_entry *start,
2189                                     struct jset_entry *end)
2190 {
2191         struct jset_entry *entry;
2192         unsigned long have = 0;
2193         unsigned i;
2194
2195         for (entry = start; entry < end; entry = vstruct_next(entry))
2196                 if (entry->type == BCH_JSET_ENTRY_btree_root)
2197                         __set_bit(entry->btree_id, &have);
2198
2199         mutex_lock(&c->btree_root_lock);
2200
2201         for (i = 0; i < BTREE_ID_NR; i++)
2202                 if (c->btree_roots[i].alive && !test_bit(i, &have)) {
2203                         journal_entry_set(end,
2204                                           BCH_JSET_ENTRY_btree_root,
2205                                           i, c->btree_roots[i].level,
2206                                           &c->btree_roots[i].key,
2207                                           c->btree_roots[i].key.u64s);
2208                         end = vstruct_next(end);
2209                 }
2210
2211         mutex_unlock(&c->btree_root_lock);
2212
2213         return end;
2214 }
2215
2216 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2217 {
2218         if (c->btree_interior_update_worker)
2219                 destroy_workqueue(c->btree_interior_update_worker);
2220         mempool_exit(&c->btree_interior_update_pool);
2221 }
2222
2223 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2224 {
2225         mutex_init(&c->btree_reserve_cache_lock);
2226         INIT_LIST_HEAD(&c->btree_interior_update_list);
2227         INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2228         mutex_init(&c->btree_interior_update_lock);
2229         INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2230
2231         c->btree_interior_update_worker =
2232                 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2233         if (!c->btree_interior_update_worker)
2234                 return -ENOMEM;
2235
2236         return mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2237                                          sizeof(struct btree_update));
2238 }