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         struct printbuf buf = PRINTBUF;
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         if (bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1179                               btree_node_type(b), WRITE, &buf) ?:
1180             bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert), &buf)) {
1181                 printbuf_reset(&buf);
1182                 pr_buf(&buf, "inserting invalid bkey\n  ");
1183                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
1184                 pr_buf(&buf, "\n  ");
1185                 bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1186                                   btree_node_type(b), WRITE, &buf);
1187                 bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert), &buf);
1188
1189                 bch2_fs_inconsistent(c, "%s", buf.buf);
1190                 dump_stack();
1191         }
1192
1193         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1194                ARRAY_SIZE(as->journal_entries));
1195
1196         as->journal_u64s +=
1197                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1198                                   BCH_JSET_ENTRY_btree_keys,
1199                                   b->c.btree_id, b->c.level,
1200                                   insert, insert->k.u64s);
1201
1202         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1203                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1204                 bch2_btree_node_iter_advance(node_iter, b);
1205
1206         bch2_btree_bset_insert_key(trans, path, b, node_iter, insert);
1207         set_btree_node_dirty_acct(c, b);
1208         set_btree_node_need_write(b);
1209
1210         printbuf_exit(&buf);
1211 }
1212
1213 static void
1214 __bch2_btree_insert_keys_interior(struct btree_update *as,
1215                                   struct btree_trans *trans,
1216                                   struct btree_path *path,
1217                                   struct btree *b,
1218                                   struct btree_node_iter node_iter,
1219                                   struct keylist *keys)
1220 {
1221         struct bkey_i *insert = bch2_keylist_front(keys);
1222         struct bkey_packed *k;
1223
1224         BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1225
1226         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1227                (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1228                 ;
1229
1230         while (!bch2_keylist_empty(keys)) {
1231                 bch2_insert_fixup_btree_ptr(as, trans, path, b,
1232                                 &node_iter, bch2_keylist_front(keys));
1233                 bch2_keylist_pop_front(keys);
1234         }
1235 }
1236
1237 /*
1238  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1239  * node)
1240  */
1241 static struct btree *__btree_split_node(struct btree_update *as,
1242                                         struct btree *n1)
1243 {
1244         struct bkey_format_state s;
1245         size_t nr_packed = 0, nr_unpacked = 0;
1246         struct btree *n2;
1247         struct bset *set1, *set2;
1248         struct bkey_packed *k, *set2_start, *set2_end, *out, *prev = NULL;
1249         struct bpos n1_pos;
1250
1251         n2 = bch2_btree_node_alloc(as, n1->c.level);
1252         bch2_btree_update_add_new_node(as, n2);
1253
1254         n2->data->max_key       = n1->data->max_key;
1255         n2->data->format        = n1->format;
1256         SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data));
1257         n2->key.k.p = n1->key.k.p;
1258
1259         set1 = btree_bset_first(n1);
1260         set2 = btree_bset_first(n2);
1261
1262         /*
1263          * Has to be a linear search because we don't have an auxiliary
1264          * search tree yet
1265          */
1266         k = set1->start;
1267         while (1) {
1268                 struct bkey_packed *n = bkey_next(k);
1269
1270                 if (n == vstruct_last(set1))
1271                         break;
1272                 if (k->_data - set1->_data >= (le16_to_cpu(set1->u64s) * 3) / 5)
1273                         break;
1274
1275                 if (bkey_packed(k))
1276                         nr_packed++;
1277                 else
1278                         nr_unpacked++;
1279
1280                 prev = k;
1281                 k = n;
1282         }
1283
1284         BUG_ON(!prev);
1285         set2_start      = k;
1286         set2_end        = vstruct_last(set1);
1287
1288         set1->u64s = cpu_to_le16((u64 *) set2_start - set1->_data);
1289         set_btree_bset_end(n1, n1->set);
1290
1291         n1->nr.live_u64s        = le16_to_cpu(set1->u64s);
1292         n1->nr.bset_u64s[0]     = le16_to_cpu(set1->u64s);
1293         n1->nr.packed_keys      = nr_packed;
1294         n1->nr.unpacked_keys    = nr_unpacked;
1295
1296         n1_pos = bkey_unpack_pos(n1, prev);
1297         if (as->c->sb.version < bcachefs_metadata_version_snapshot)
1298                 n1_pos.snapshot = U32_MAX;
1299
1300         btree_set_max(n1, n1_pos);
1301         btree_set_min(n2, bpos_successor(n1->key.k.p));
1302
1303         bch2_bkey_format_init(&s);
1304         bch2_bkey_format_add_pos(&s, n2->data->min_key);
1305         bch2_bkey_format_add_pos(&s, n2->data->max_key);
1306
1307         for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1308                 struct bkey uk = bkey_unpack_key(n1, k);
1309                 bch2_bkey_format_add_key(&s, &uk);
1310         }
1311
1312         n2->data->format = bch2_bkey_format_done(&s);
1313         btree_node_set_format(n2, n2->data->format);
1314
1315         out = set2->start;
1316         memset(&n2->nr, 0, sizeof(n2->nr));
1317
1318         for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1319                 BUG_ON(!bch2_bkey_transform(&n2->format, out, bkey_packed(k)
1320                                        ? &n1->format : &bch2_bkey_format_current, k));
1321                 out->format = KEY_FORMAT_LOCAL_BTREE;
1322                 btree_keys_account_key_add(&n2->nr, 0, out);
1323                 out = bkey_next(out);
1324         }
1325
1326         set2->u64s = cpu_to_le16((u64 *) out - set2->_data);
1327         set_btree_bset_end(n2, n2->set);
1328
1329         BUG_ON(!set1->u64s);
1330         BUG_ON(!set2->u64s);
1331
1332         btree_node_reset_sib_u64s(n1);
1333         btree_node_reset_sib_u64s(n2);
1334
1335         bch2_verify_btree_nr_keys(n1);
1336         bch2_verify_btree_nr_keys(n2);
1337
1338         if (n1->c.level) {
1339                 btree_node_interior_verify(as->c, n1);
1340                 btree_node_interior_verify(as->c, n2);
1341         }
1342
1343         return n2;
1344 }
1345
1346 /*
1347  * For updates to interior nodes, we've got to do the insert before we split
1348  * because the stuff we're inserting has to be inserted atomically. Post split,
1349  * the keys might have to go in different nodes and the split would no longer be
1350  * atomic.
1351  *
1352  * Worse, if the insert is from btree node coalescing, if we do the insert after
1353  * we do the split (and pick the pivot) - the pivot we pick might be between
1354  * nodes that were coalesced, and thus in the middle of a child node post
1355  * coalescing:
1356  */
1357 static void btree_split_insert_keys(struct btree_update *as,
1358                                     struct btree_trans *trans,
1359                                     struct btree_path *path,
1360                                     struct btree *b,
1361                                     struct keylist *keys)
1362 {
1363         struct btree_node_iter node_iter;
1364         struct bkey_i *k = bch2_keylist_front(keys);
1365         struct bkey_packed *src, *dst, *n;
1366         struct bset *i;
1367
1368         bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
1369
1370         __bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
1371
1372         /*
1373          * We can't tolerate whiteouts here - with whiteouts there can be
1374          * duplicate keys, and it would be rather bad if we picked a duplicate
1375          * for the pivot:
1376          */
1377         i = btree_bset_first(b);
1378         src = dst = i->start;
1379         while (src != vstruct_last(i)) {
1380                 n = bkey_next(src);
1381                 if (!bkey_deleted(src)) {
1382                         memmove_u64s_down(dst, src, src->u64s);
1383                         dst = bkey_next(dst);
1384                 }
1385                 src = n;
1386         }
1387
1388         /* Also clear out the unwritten whiteouts area: */
1389         b->whiteout_u64s = 0;
1390
1391         i->u64s = cpu_to_le16((u64 *) dst - i->_data);
1392         set_btree_bset_end(b, b->set);
1393
1394         BUG_ON(b->nsets != 1 ||
1395                b->nr.live_u64s != le16_to_cpu(btree_bset_first(b)->u64s));
1396
1397         btree_node_interior_verify(as->c, b);
1398 }
1399
1400 static void btree_split(struct btree_update *as, struct btree_trans *trans,
1401                         struct btree_path *path, struct btree *b,
1402                         struct keylist *keys, unsigned flags)
1403 {
1404         struct bch_fs *c = as->c;
1405         struct btree *parent = btree_node_parent(path, b);
1406         struct btree *n1, *n2 = NULL, *n3 = NULL;
1407         u64 start_time = local_clock();
1408
1409         BUG_ON(!parent && (b != btree_node_root(c, b)));
1410         BUG_ON(!btree_node_intent_locked(path, btree_node_root(c, b)->c.level));
1411
1412         bch2_btree_interior_update_will_free_node(as, b);
1413
1414         n1 = bch2_btree_node_alloc_replacement(as, b);
1415         bch2_btree_update_add_new_node(as, n1);
1416
1417         if (keys)
1418                 btree_split_insert_keys(as, trans, path, n1, keys);
1419
1420         if (bset_u64s(&n1->set[0]) > BTREE_SPLIT_THRESHOLD(c)) {
1421                 trace_btree_split(c, b);
1422
1423                 n2 = __btree_split_node(as, n1);
1424
1425                 bch2_btree_build_aux_trees(n2);
1426                 bch2_btree_build_aux_trees(n1);
1427                 six_unlock_write(&n2->c.lock);
1428                 six_unlock_write(&n1->c.lock);
1429
1430                 bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1431                 bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0);
1432
1433                 /*
1434                  * Note that on recursive parent_keys == keys, so we
1435                  * can't start adding new keys to parent_keys before emptying it
1436                  * out (which we did with btree_split_insert_keys() above)
1437                  */
1438                 bch2_keylist_add(&as->parent_keys, &n1->key);
1439                 bch2_keylist_add(&as->parent_keys, &n2->key);
1440
1441                 if (!parent) {
1442                         /* Depth increases, make a new root */
1443                         n3 = __btree_root_alloc(as, b->c.level + 1);
1444
1445                         n3->sib_u64s[0] = U16_MAX;
1446                         n3->sib_u64s[1] = U16_MAX;
1447
1448                         btree_split_insert_keys(as, trans, path, n3, &as->parent_keys);
1449
1450                         bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0);
1451                 }
1452         } else {
1453                 trace_btree_compact(c, b);
1454
1455                 bch2_btree_build_aux_trees(n1);
1456                 six_unlock_write(&n1->c.lock);
1457
1458                 bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1459
1460                 if (parent)
1461                         bch2_keylist_add(&as->parent_keys, &n1->key);
1462         }
1463
1464         /* New nodes all written, now make them visible: */
1465
1466         if (parent) {
1467                 /* Split a non root node */
1468                 bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1469         } else if (n3) {
1470                 bch2_btree_set_root(as, trans, path, n3);
1471         } else {
1472                 /* Root filled up but didn't need to be split */
1473                 bch2_btree_set_root(as, trans, path, n1);
1474         }
1475
1476         bch2_btree_update_get_open_buckets(as, n1);
1477         if (n2)
1478                 bch2_btree_update_get_open_buckets(as, n2);
1479         if (n3)
1480                 bch2_btree_update_get_open_buckets(as, n3);
1481
1482         /* Successful split, update the path to point to the new nodes: */
1483
1484         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1485         if (n3)
1486                 bch2_trans_node_add(trans, n3);
1487         if (n2)
1488                 bch2_trans_node_add(trans, n2);
1489         bch2_trans_node_add(trans, n1);
1490
1491         /*
1492          * The old node must be freed (in memory) _before_ unlocking the new
1493          * nodes - else another thread could re-acquire a read lock on the old
1494          * node after another thread has locked and updated the new node, thus
1495          * seeing stale data:
1496          */
1497         bch2_btree_node_free_inmem(trans, b);
1498
1499         if (n3)
1500                 six_unlock_intent(&n3->c.lock);
1501         if (n2)
1502                 six_unlock_intent(&n2->c.lock);
1503         six_unlock_intent(&n1->c.lock);
1504
1505         bch2_trans_verify_locks(trans);
1506
1507         bch2_time_stats_update(&c->times[n2
1508                                ? BCH_TIME_btree_node_split
1509                                : BCH_TIME_btree_node_compact],
1510                                start_time);
1511 }
1512
1513 static void
1514 bch2_btree_insert_keys_interior(struct btree_update *as,
1515                                 struct btree_trans *trans,
1516                                 struct btree_path *path,
1517                                 struct btree *b,
1518                                 struct keylist *keys)
1519 {
1520         struct btree_path *linked;
1521
1522         __bch2_btree_insert_keys_interior(as, trans, path, b,
1523                                           path->l[b->c.level].iter, keys);
1524
1525         btree_update_updated_node(as, b);
1526
1527         trans_for_each_path_with_node(trans, b, linked)
1528                 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1529
1530         bch2_trans_verify_paths(trans);
1531 }
1532
1533 /**
1534  * bch_btree_insert_node - insert bkeys into a given btree node
1535  *
1536  * @iter:               btree iterator
1537  * @keys:               list of keys to insert
1538  * @hook:               insert callback
1539  * @persistent:         if not null, @persistent will wait on journal write
1540  *
1541  * Inserts as many keys as it can into a given btree node, splitting it if full.
1542  * If a split occurred, this function will return early. This can only happen
1543  * for leaf nodes -- inserts into interior nodes have to be atomic.
1544  */
1545 static void bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans,
1546                                    struct btree_path *path, struct btree *b,
1547                                    struct keylist *keys, unsigned flags)
1548 {
1549         struct bch_fs *c = as->c;
1550         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1551         int old_live_u64s = b->nr.live_u64s;
1552         int live_u64s_added, u64s_added;
1553
1554         lockdep_assert_held(&c->gc_lock);
1555         BUG_ON(!btree_node_intent_locked(path, btree_node_root(c, b)->c.level));
1556         BUG_ON(!b->c.level);
1557         BUG_ON(!as || as->b);
1558         bch2_verify_keylist_sorted(keys);
1559
1560         bch2_btree_node_lock_for_insert(trans, path, b);
1561
1562         if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1563                 bch2_btree_node_unlock_write(trans, path, b);
1564                 goto split;
1565         }
1566
1567         btree_node_interior_verify(c, b);
1568
1569         bch2_btree_insert_keys_interior(as, trans, path, b, keys);
1570
1571         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1572         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1573
1574         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1575                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1576         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1577                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1578
1579         if (u64s_added > live_u64s_added &&
1580             bch2_maybe_compact_whiteouts(c, b))
1581                 bch2_trans_node_reinit_iter(trans, b);
1582
1583         bch2_btree_node_unlock_write(trans, path, b);
1584
1585         btree_node_interior_verify(c, b);
1586         return;
1587 split:
1588         btree_split(as, trans, path, b, keys, flags);
1589 }
1590
1591 int bch2_btree_split_leaf(struct btree_trans *trans,
1592                           struct btree_path *path,
1593                           unsigned flags)
1594 {
1595         struct btree *b = path_l(path)->b;
1596         struct btree_update *as;
1597         unsigned l;
1598         int ret = 0;
1599
1600         as = bch2_btree_update_start(trans, path, path->level,
1601                                      true, flags);
1602         if (IS_ERR(as))
1603                 return PTR_ERR(as);
1604
1605         btree_split(as, trans, path, b, NULL, flags);
1606         bch2_btree_update_done(as);
1607
1608         for (l = path->level + 1; btree_path_node(path, l) && !ret; l++)
1609                 ret = bch2_foreground_maybe_merge(trans, path, l, flags);
1610
1611         return ret;
1612 }
1613
1614 int __bch2_foreground_maybe_merge(struct btree_trans *trans,
1615                                   struct btree_path *path,
1616                                   unsigned level,
1617                                   unsigned flags,
1618                                   enum btree_node_sibling sib)
1619 {
1620         struct bch_fs *c = trans->c;
1621         struct btree_path *sib_path = NULL;
1622         struct btree_update *as;
1623         struct bkey_format_state new_s;
1624         struct bkey_format new_f;
1625         struct bkey_i delete;
1626         struct btree *b, *m, *n, *prev, *next, *parent;
1627         struct bpos sib_pos;
1628         size_t sib_u64s;
1629         u64 start_time = local_clock();
1630         int ret = 0;
1631
1632         BUG_ON(!path->should_be_locked);
1633         BUG_ON(!btree_node_locked(path, level));
1634
1635         b = path->l[level].b;
1636
1637         if ((sib == btree_prev_sib && !bpos_cmp(b->data->min_key, POS_MIN)) ||
1638             (sib == btree_next_sib && !bpos_cmp(b->data->max_key, SPOS_MAX))) {
1639                 b->sib_u64s[sib] = U16_MAX;
1640                 return 0;
1641         }
1642
1643         sib_pos = sib == btree_prev_sib
1644                 ? bpos_predecessor(b->data->min_key)
1645                 : bpos_successor(b->data->max_key);
1646
1647         sib_path = bch2_path_get(trans, path->btree_id, sib_pos,
1648                                  U8_MAX, level, BTREE_ITER_INTENT, _THIS_IP_);
1649         ret = bch2_btree_path_traverse(trans, sib_path, false);
1650         if (ret)
1651                 goto err;
1652
1653         sib_path->should_be_locked = true;
1654
1655         m = sib_path->l[level].b;
1656
1657         if (btree_node_parent(path, b) !=
1658             btree_node_parent(sib_path, m)) {
1659                 b->sib_u64s[sib] = U16_MAX;
1660                 goto out;
1661         }
1662
1663         if (sib == btree_prev_sib) {
1664                 prev = m;
1665                 next = b;
1666         } else {
1667                 prev = b;
1668                 next = m;
1669         }
1670
1671         if (bkey_cmp(bpos_successor(prev->data->max_key), next->data->min_key)) {
1672                 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1673
1674                 bch2_bpos_to_text(&buf1, prev->data->max_key);
1675                 bch2_bpos_to_text(&buf2, next->data->min_key);
1676                 bch_err(c,
1677                         "btree topology error in btree merge:\n"
1678                         "  prev ends at   %s\n"
1679                         "  next starts at %s",
1680                         buf1.buf, buf2.buf);
1681                 printbuf_exit(&buf1);
1682                 printbuf_exit(&buf2);
1683                 bch2_topology_error(c);
1684                 ret = -EIO;
1685                 goto err;
1686         }
1687
1688         bch2_bkey_format_init(&new_s);
1689         bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
1690         __bch2_btree_calc_format(&new_s, prev);
1691         __bch2_btree_calc_format(&new_s, next);
1692         bch2_bkey_format_add_pos(&new_s, next->data->max_key);
1693         new_f = bch2_bkey_format_done(&new_s);
1694
1695         sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1696                 btree_node_u64s_with_format(m, &new_f);
1697
1698         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1699                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1700                 sib_u64s /= 2;
1701                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1702         }
1703
1704         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1705         sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
1706         b->sib_u64s[sib] = sib_u64s;
1707
1708         if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
1709                 goto out;
1710
1711         parent = btree_node_parent(path, b);
1712         as = bch2_btree_update_start(trans, path, level, false,
1713                          BTREE_INSERT_NOFAIL|
1714                          BTREE_INSERT_USE_RESERVE|
1715                          flags);
1716         ret = PTR_ERR_OR_ZERO(as);
1717         if (ret)
1718                 goto err;
1719
1720         trace_btree_merge(c, b);
1721
1722         bch2_btree_interior_update_will_free_node(as, b);
1723         bch2_btree_interior_update_will_free_node(as, m);
1724
1725         n = bch2_btree_node_alloc(as, b->c.level);
1726         bch2_btree_update_add_new_node(as, n);
1727
1728         SET_BTREE_NODE_SEQ(n->data,
1729                            max(BTREE_NODE_SEQ(b->data),
1730                                BTREE_NODE_SEQ(m->data)) + 1);
1731
1732         btree_set_min(n, prev->data->min_key);
1733         btree_set_max(n, next->data->max_key);
1734         n->data->format         = new_f;
1735
1736         btree_node_set_format(n, new_f);
1737
1738         bch2_btree_sort_into(c, n, prev);
1739         bch2_btree_sort_into(c, n, next);
1740
1741         bch2_btree_build_aux_trees(n);
1742         six_unlock_write(&n->c.lock);
1743
1744         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1745
1746         bkey_init(&delete.k);
1747         delete.k.p = prev->key.k.p;
1748         bch2_keylist_add(&as->parent_keys, &delete);
1749         bch2_keylist_add(&as->parent_keys, &n->key);
1750
1751         bch2_trans_verify_paths(trans);
1752
1753         bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1754
1755         bch2_trans_verify_paths(trans);
1756
1757         bch2_btree_update_get_open_buckets(as, n);
1758
1759         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1760         six_lock_increment(&m->c.lock, SIX_LOCK_intent);
1761
1762         bch2_trans_node_add(trans, n);
1763
1764         bch2_trans_verify_paths(trans);
1765
1766         bch2_btree_node_free_inmem(trans, b);
1767         bch2_btree_node_free_inmem(trans, m);
1768
1769         six_unlock_intent(&n->c.lock);
1770
1771         bch2_btree_update_done(as);
1772
1773         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time);
1774 out:
1775 err:
1776         bch2_path_put(trans, sib_path, true);
1777         bch2_trans_verify_locks(trans);
1778         return ret;
1779 }
1780
1781 /**
1782  * bch_btree_node_rewrite - Rewrite/move a btree node
1783  */
1784 int bch2_btree_node_rewrite(struct btree_trans *trans,
1785                             struct btree_iter *iter,
1786                             struct btree *b,
1787                             unsigned flags)
1788 {
1789         struct bch_fs *c = trans->c;
1790         struct btree *n, *parent;
1791         struct btree_update *as;
1792         int ret;
1793
1794         flags |= BTREE_INSERT_NOFAIL;
1795
1796         parent = btree_node_parent(iter->path, b);
1797         as = bch2_btree_update_start(trans, iter->path, b->c.level,
1798                                      false, flags);
1799         ret = PTR_ERR_OR_ZERO(as);
1800         if (ret) {
1801                 trace_btree_gc_rewrite_node_fail(c, b);
1802                 goto out;
1803         }
1804
1805         bch2_btree_interior_update_will_free_node(as, b);
1806
1807         n = bch2_btree_node_alloc_replacement(as, b);
1808         bch2_btree_update_add_new_node(as, n);
1809
1810         bch2_btree_build_aux_trees(n);
1811         six_unlock_write(&n->c.lock);
1812
1813         trace_btree_gc_rewrite_node(c, b);
1814
1815         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1816
1817         if (parent) {
1818                 bch2_keylist_add(&as->parent_keys, &n->key);
1819                 bch2_btree_insert_node(as, trans, iter->path, parent,
1820                                        &as->parent_keys, flags);
1821         } else {
1822                 bch2_btree_set_root(as, trans, iter->path, n);
1823         }
1824
1825         bch2_btree_update_get_open_buckets(as, n);
1826
1827         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1828         bch2_trans_node_add(trans, n);
1829         bch2_btree_node_free_inmem(trans, b);
1830         six_unlock_intent(&n->c.lock);
1831
1832         bch2_btree_update_done(as);
1833 out:
1834         bch2_btree_path_downgrade(iter->path);
1835         return ret;
1836 }
1837
1838 struct async_btree_rewrite {
1839         struct bch_fs           *c;
1840         struct work_struct      work;
1841         enum btree_id           btree_id;
1842         unsigned                level;
1843         struct bpos             pos;
1844         __le64                  seq;
1845 };
1846
1847 static int async_btree_node_rewrite_trans(struct btree_trans *trans,
1848                                           struct async_btree_rewrite *a)
1849 {
1850         struct btree_iter iter;
1851         struct btree *b;
1852         int ret;
1853
1854         bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos,
1855                                   BTREE_MAX_DEPTH, a->level, 0);
1856         b = bch2_btree_iter_peek_node(&iter);
1857         ret = PTR_ERR_OR_ZERO(b);
1858         if (ret)
1859                 goto out;
1860
1861         if (!b || b->data->keys.seq != a->seq)
1862                 goto out;
1863
1864         ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
1865 out :
1866         bch2_trans_iter_exit(trans, &iter);
1867
1868         return ret;
1869 }
1870
1871 void async_btree_node_rewrite_work(struct work_struct *work)
1872 {
1873         struct async_btree_rewrite *a =
1874                 container_of(work, struct async_btree_rewrite, work);
1875         struct bch_fs *c = a->c;
1876
1877         bch2_trans_do(c, NULL, NULL, 0,
1878                       async_btree_node_rewrite_trans(&trans, a));
1879         percpu_ref_put(&c->writes);
1880         kfree(a);
1881 }
1882
1883 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
1884 {
1885         struct async_btree_rewrite *a;
1886
1887         if (!percpu_ref_tryget(&c->writes))
1888                 return;
1889
1890         a = kmalloc(sizeof(*a), GFP_NOFS);
1891         if (!a) {
1892                 percpu_ref_put(&c->writes);
1893                 return;
1894         }
1895
1896         a->c            = c;
1897         a->btree_id     = b->c.btree_id;
1898         a->level        = b->c.level;
1899         a->pos          = b->key.k.p;
1900         a->seq          = b->data->keys.seq;
1901
1902         INIT_WORK(&a->work, async_btree_node_rewrite_work);
1903         queue_work(c->btree_interior_update_worker, &a->work);
1904 }
1905
1906 static int __bch2_btree_node_update_key(struct btree_trans *trans,
1907                                         struct btree_iter *iter,
1908                                         struct btree *b, struct btree *new_hash,
1909                                         struct bkey_i *new_key,
1910                                         bool skip_triggers)
1911 {
1912         struct bch_fs *c = trans->c;
1913         struct btree_iter iter2 = { NULL };
1914         struct btree *parent;
1915         int ret;
1916
1917         if (!skip_triggers) {
1918                 ret = bch2_trans_mark_new(trans, new_key, 0);
1919                 if (ret)
1920                         return ret;
1921
1922                 ret = bch2_trans_mark_old(trans, bkey_i_to_s_c(&b->key), 0);
1923                 if (ret)
1924                         return ret;
1925         }
1926
1927         if (new_hash) {
1928                 bkey_copy(&new_hash->key, new_key);
1929                 ret = bch2_btree_node_hash_insert(&c->btree_cache,
1930                                 new_hash, b->c.level, b->c.btree_id);
1931                 BUG_ON(ret);
1932         }
1933
1934         parent = btree_node_parent(iter->path, b);
1935         if (parent) {
1936                 bch2_trans_copy_iter(&iter2, iter);
1937
1938                 iter2.path = bch2_btree_path_make_mut(trans, iter2.path,
1939                                 iter2.flags & BTREE_ITER_INTENT,
1940                                 _THIS_IP_);
1941
1942                 BUG_ON(iter2.path->level != b->c.level);
1943                 BUG_ON(bpos_cmp(iter2.path->pos, new_key->k.p));
1944
1945                 btree_node_unlock(iter2.path, iter2.path->level);
1946                 path_l(iter2.path)->b = BTREE_ITER_NO_NODE_UP;
1947                 iter2.path->level++;
1948                 btree_path_set_dirty(iter2.path, BTREE_ITER_NEED_TRAVERSE);
1949
1950                 bch2_btree_path_check_sort(trans, iter2.path, 0);
1951
1952                 ret   = bch2_btree_iter_traverse(&iter2) ?:
1953                         bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_NORUN);
1954                 if (ret)
1955                         goto err;
1956         } else {
1957                 BUG_ON(btree_node_root(c, b) != b);
1958
1959                 ret = darray_make_room(trans->extra_journal_entries,
1960                                        jset_u64s(new_key->k.u64s));
1961                 if (ret)
1962                         return ret;
1963
1964                 journal_entry_set((void *) &darray_top(trans->extra_journal_entries),
1965                                   BCH_JSET_ENTRY_btree_root,
1966                                   b->c.btree_id, b->c.level,
1967                                   new_key, new_key->k.u64s);
1968                 trans->extra_journal_entries.nr += jset_u64s(new_key->k.u64s);
1969         }
1970
1971         ret = bch2_trans_commit(trans, NULL, NULL,
1972                                 BTREE_INSERT_NOFAIL|
1973                                 BTREE_INSERT_NOCHECK_RW|
1974                                 BTREE_INSERT_USE_RESERVE|
1975                                 BTREE_INSERT_JOURNAL_RECLAIM|
1976                                 JOURNAL_WATERMARK_reserved);
1977         if (ret)
1978                 goto err;
1979
1980         bch2_btree_node_lock_write(trans, iter->path, b);
1981
1982         if (new_hash) {
1983                 mutex_lock(&c->btree_cache.lock);
1984                 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
1985                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1986
1987                 bkey_copy(&b->key, new_key);
1988                 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1989                 BUG_ON(ret);
1990                 mutex_unlock(&c->btree_cache.lock);
1991         } else {
1992                 bkey_copy(&b->key, new_key);
1993         }
1994
1995         bch2_btree_node_unlock_write(trans, iter->path, b);
1996 out:
1997         bch2_trans_iter_exit(trans, &iter2);
1998         return ret;
1999 err:
2000         if (new_hash) {
2001                 mutex_lock(&c->btree_cache.lock);
2002                 bch2_btree_node_hash_remove(&c->btree_cache, b);
2003                 mutex_unlock(&c->btree_cache.lock);
2004         }
2005         goto out;
2006 }
2007
2008 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
2009                                struct btree *b, struct bkey_i *new_key,
2010                                bool skip_triggers)
2011 {
2012         struct bch_fs *c = trans->c;
2013         struct btree *new_hash = NULL;
2014         struct btree_path *path = iter->path;
2015         struct closure cl;
2016         int ret = 0;
2017
2018         if (!btree_node_intent_locked(path, b->c.level) &&
2019             !bch2_btree_path_upgrade(trans, path, b->c.level + 1)) {
2020                 btree_trans_restart(trans);
2021                 return -EINTR;
2022         }
2023
2024         closure_init_stack(&cl);
2025
2026         /*
2027          * check btree_ptr_hash_val() after @b is locked by
2028          * btree_iter_traverse():
2029          */
2030         if (btree_ptr_hash_val(new_key) != b->hash_val) {
2031                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2032                 if (ret) {
2033                         bch2_trans_unlock(trans);
2034                         closure_sync(&cl);
2035                         if (!bch2_trans_relock(trans))
2036                                 return -EINTR;
2037                 }
2038
2039                 new_hash = bch2_btree_node_mem_alloc(c, false);
2040         }
2041
2042         path->intent_ref++;
2043         ret = __bch2_btree_node_update_key(trans, iter, b, new_hash,
2044                                            new_key, skip_triggers);
2045         --path->intent_ref;
2046
2047         if (new_hash) {
2048                 mutex_lock(&c->btree_cache.lock);
2049                 list_move(&new_hash->list, &c->btree_cache.freeable);
2050                 mutex_unlock(&c->btree_cache.lock);
2051
2052                 six_unlock_write(&new_hash->c.lock);
2053                 six_unlock_intent(&new_hash->c.lock);
2054         }
2055         closure_sync(&cl);
2056         bch2_btree_cache_cannibalize_unlock(c);
2057         return ret;
2058 }
2059
2060 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
2061                                         struct btree *b, struct bkey_i *new_key,
2062                                         bool skip_triggers)
2063 {
2064         struct btree_iter iter;
2065         int ret;
2066
2067         bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p,
2068                                   BTREE_MAX_DEPTH, b->c.level,
2069                                   BTREE_ITER_INTENT);
2070         ret = bch2_btree_iter_traverse(&iter);
2071         if (ret)
2072                 goto out;
2073
2074         /* has node been freed? */
2075         if (iter.path->l[b->c.level].b != b) {
2076                 /* node has been freed: */
2077                 BUG_ON(!btree_node_dying(b));
2078                 goto out;
2079         }
2080
2081         BUG_ON(!btree_node_hashed(b));
2082
2083         ret = bch2_btree_node_update_key(trans, &iter, b, new_key, skip_triggers);
2084 out:
2085         bch2_trans_iter_exit(trans, &iter);
2086         return ret;
2087 }
2088
2089 /* Init code: */
2090
2091 /*
2092  * Only for filesystem bringup, when first reading the btree roots or allocating
2093  * btree roots when initializing a new filesystem:
2094  */
2095 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2096 {
2097         BUG_ON(btree_node_root(c, b));
2098
2099         bch2_btree_set_root_inmem(c, b);
2100 }
2101
2102 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2103 {
2104         struct closure cl;
2105         struct btree *b;
2106         int ret;
2107
2108         closure_init_stack(&cl);
2109
2110         do {
2111                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2112                 closure_sync(&cl);
2113         } while (ret);
2114
2115         b = bch2_btree_node_mem_alloc(c, false);
2116         bch2_btree_cache_cannibalize_unlock(c);
2117
2118         set_btree_node_fake(b);
2119         set_btree_node_need_rewrite(b);
2120         b->c.level      = 0;
2121         b->c.btree_id   = id;
2122
2123         bkey_btree_ptr_init(&b->key);
2124         b->key.k.p = SPOS_MAX;
2125         *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2126
2127         bch2_bset_init_first(b, &b->data->keys);
2128         bch2_btree_build_aux_trees(b);
2129
2130         b->data->flags = 0;
2131         btree_set_min(b, POS_MIN);
2132         btree_set_max(b, SPOS_MAX);
2133         b->data->format = bch2_btree_calc_format(b);
2134         btree_node_set_format(b, b->data->format);
2135
2136         ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2137                                           b->c.level, b->c.btree_id);
2138         BUG_ON(ret);
2139
2140         bch2_btree_set_root_inmem(c, b);
2141
2142         six_unlock_write(&b->c.lock);
2143         six_unlock_intent(&b->c.lock);
2144 }
2145
2146 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2147 {
2148         struct btree_update *as;
2149
2150         mutex_lock(&c->btree_interior_update_lock);
2151         list_for_each_entry(as, &c->btree_interior_update_list, list)
2152                 pr_buf(out, "%p m %u w %u r %u j %llu\n",
2153                        as,
2154                        as->mode,
2155                        as->nodes_written,
2156                        atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2157                        as->journal.seq);
2158         mutex_unlock(&c->btree_interior_update_lock);
2159 }
2160
2161 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *c)
2162 {
2163         size_t ret = 0;
2164         struct list_head *i;
2165
2166         mutex_lock(&c->btree_interior_update_lock);
2167         list_for_each(i, &c->btree_interior_update_list)
2168                 ret++;
2169         mutex_unlock(&c->btree_interior_update_lock);
2170
2171         return ret;
2172 }
2173
2174 void bch2_journal_entries_to_btree_roots(struct bch_fs *c, struct jset *jset)
2175 {
2176         struct btree_root *r;
2177         struct jset_entry *entry;
2178
2179         mutex_lock(&c->btree_root_lock);
2180
2181         vstruct_for_each(jset, entry)
2182                 if (entry->type == BCH_JSET_ENTRY_btree_root) {
2183                         r = &c->btree_roots[entry->btree_id];
2184                         r->level = entry->level;
2185                         r->alive = true;
2186                         bkey_copy(&r->key, &entry->start[0]);
2187                 }
2188
2189         mutex_unlock(&c->btree_root_lock);
2190 }
2191
2192 struct jset_entry *
2193 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2194                                     struct jset_entry *start,
2195                                     struct jset_entry *end)
2196 {
2197         struct jset_entry *entry;
2198         unsigned long have = 0;
2199         unsigned i;
2200
2201         for (entry = start; entry < end; entry = vstruct_next(entry))
2202                 if (entry->type == BCH_JSET_ENTRY_btree_root)
2203                         __set_bit(entry->btree_id, &have);
2204
2205         mutex_lock(&c->btree_root_lock);
2206
2207         for (i = 0; i < BTREE_ID_NR; i++)
2208                 if (c->btree_roots[i].alive && !test_bit(i, &have)) {
2209                         journal_entry_set(end,
2210                                           BCH_JSET_ENTRY_btree_root,
2211                                           i, c->btree_roots[i].level,
2212                                           &c->btree_roots[i].key,
2213                                           c->btree_roots[i].key.u64s);
2214                         end = vstruct_next(end);
2215                 }
2216
2217         mutex_unlock(&c->btree_root_lock);
2218
2219         return end;
2220 }
2221
2222 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2223 {
2224         if (c->btree_interior_update_worker)
2225                 destroy_workqueue(c->btree_interior_update_worker);
2226         mempool_exit(&c->btree_interior_update_pool);
2227 }
2228
2229 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2230 {
2231         mutex_init(&c->btree_reserve_cache_lock);
2232         INIT_LIST_HEAD(&c->btree_interior_update_list);
2233         INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2234         mutex_init(&c->btree_interior_update_lock);
2235         INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2236
2237         c->btree_interior_update_worker =
2238                 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2239         if (!c->btree_interior_update_worker)
2240                 return -ENOMEM;
2241
2242         return mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2243                                          sizeof(struct btree_update));
2244 }