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 int 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 static struct btree_path *get_unlocked_mut_path(struct btree_trans *trans,
  32                                                 enum btree_id btree_id,
  33                                                 unsigned level,
  34                                                 struct bpos pos)
  35 {
  36         struct btree_path *path;
  37
  38         path = bch2_path_get(trans, btree_id, pos, level + 1, level,
  39                              BTREE_ITER_NOPRESERVE|
  40                              BTREE_ITER_INTENT, _RET_IP_);
  41         path = bch2_btree_path_make_mut(trans, path, true, _RET_IP_);
  42         bch2_btree_path_downgrade(trans, path);
  43         __bch2_btree_path_unlock(trans, path);
  44         return path;
  45 }
  46
  47 /* Debug code: */
  48
  49 /*
  50  * Verify that child nodes correctly span parent node's range:
  51  */
  52 static void btree_node_interior_verify(struct bch_fs *c, struct btree *b)
  53 {
  54 #ifdef CONFIG_BCACHEFS_DEBUG
  55         struct bpos next_node = b->data->min_key;
  56         struct btree_node_iter iter;
  57         struct bkey_s_c k;
  58         struct bkey_s_c_btree_ptr_v2 bp;
  59         struct bkey unpacked;
  60         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
  61
  62         BUG_ON(!b->c.level);
  63
  64         if (!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))
  65                 return;
  66
  67         bch2_btree_node_iter_init_from_start(&iter, b);
  68
  69         while (1) {
  70                 k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked);
  71                 if (k.k->type != KEY_TYPE_btree_ptr_v2)
  72                         break;
  73                 bp = bkey_s_c_to_btree_ptr_v2(k);
  74
  75                 if (!bpos_eq(next_node, bp.v->min_key)) {
  76                         bch2_dump_btree_node(c, b);
  77                         bch2_bpos_to_text(&buf1, next_node);
  78                         bch2_bpos_to_text(&buf2, bp.v->min_key);
  79                         panic("expected next min_key %s got %s\n", buf1.buf, buf2.buf);
  80                 }
  81
  82                 bch2_btree_node_iter_advance(&iter, b);
  83
  84                 if (bch2_btree_node_iter_end(&iter)) {
  85                         if (!bpos_eq(k.k->p, b->key.k.p)) {
  86                                 bch2_dump_btree_node(c, b);
  87                                 bch2_bpos_to_text(&buf1, b->key.k.p);
  88                                 bch2_bpos_to_text(&buf2, k.k->p);
  89                                 panic("expected end %s got %s\n", buf1.buf, buf2.buf);
  90                         }
  91                         break;
  92                 }
  93
  94                 next_node = bpos_successor(k.k->p);
  95         }
  96 #endif
  97 }
  98
  99 /* Calculate ideal packed bkey format for new btree nodes: */
 100
 101 void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b)
 102 {
 103         struct bkey_packed *k;
 104         struct bset_tree *t;
 105         struct bkey uk;
 106
 107         for_each_bset(b, t)
 108                 bset_tree_for_each_key(b, t, k)
 109                         if (!bkey_deleted(k)) {
 110                                 uk = bkey_unpack_key(b, k);
 111                                 bch2_bkey_format_add_key(s, &uk);
 112                         }
 113 }
 114
 115 static struct bkey_format bch2_btree_calc_format(struct btree *b)
 116 {
 117         struct bkey_format_state s;
 118
 119         bch2_bkey_format_init(&s);
 120         bch2_bkey_format_add_pos(&s, b->data->min_key);
 121         bch2_bkey_format_add_pos(&s, b->data->max_key);
 122         __bch2_btree_calc_format(&s, b);
 123
 124         return bch2_bkey_format_done(&s);
 125 }
 126
 127 static size_t btree_node_u64s_with_format(struct btree *b,
 128                                           struct bkey_format *new_f)
 129 {
 130         struct bkey_format *old_f = &b->format;
 131
 132         /* stupid integer promotion rules */
 133         ssize_t delta =
 134             (((int) new_f->key_u64s - old_f->key_u64s) *
 135              (int) b->nr.packed_keys) +
 136             (((int) new_f->key_u64s - BKEY_U64s) *
 137              (int) b->nr.unpacked_keys);
 138
 139         BUG_ON(delta + b->nr.live_u64s < 0);
 140
 141         return b->nr.live_u64s + delta;
 142 }
 143
 144 /**
 145  * btree_node_format_fits - check if we could rewrite node with a new format
 146  *
 147  * This assumes all keys can pack with the new format -- it just checks if
 148  * the re-packed keys would fit inside the node itself.
 149  */
 150 bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
 151                                  struct bkey_format *new_f)
 152 {
 153         size_t u64s = btree_node_u64s_with_format(b, new_f);
 154
 155         return __vstruct_bytes(struct btree_node, u64s) < btree_bytes(c);
 156 }
 157
 158 /* Btree node freeing/allocation: */
 159
 160 static void __btree_node_free(struct bch_fs *c, struct btree *b)
 161 {
 162         trace_and_count(c, btree_node_free, c, b);
 163
 164         BUG_ON(btree_node_write_blocked(b));
 165         BUG_ON(btree_node_dirty(b));
 166         BUG_ON(btree_node_need_write(b));
 167         BUG_ON(b == btree_node_root(c, b));
 168         BUG_ON(b->ob.nr);
 169         BUG_ON(!list_empty(&b->write_blocked));
 170         BUG_ON(b->will_make_reachable);
 171
 172         clear_btree_node_noevict(b);
 173
 174         mutex_lock(&c->btree_cache.lock);
 175         list_move(&b->list, &c->btree_cache.freeable);
 176         mutex_unlock(&c->btree_cache.lock);
 177 }
 178
 179 static void bch2_btree_node_free_inmem(struct btree_trans *trans,
 180                                        struct btree_path *path,
 181                                        struct btree *b)
 182 {
 183         struct bch_fs *c = trans->c;
 184         unsigned level = b->c.level;
 185
 186         bch2_btree_node_lock_write_nofail(trans, path, &b->c);
 187         bch2_btree_node_hash_remove(&c->btree_cache, b);
 188         __btree_node_free(c, b);
 189         six_unlock_write(&b->c.lock);
 190         mark_btree_node_locked_noreset(path, level, SIX_LOCK_intent);
 191
 192         trans_for_each_path(trans, path)
 193                 if (path->l[level].b == b) {
 194                         btree_node_unlock(trans, path, level);
 195                         path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
 196                 }
 197 }
 198
 199 static void bch2_btree_node_free_never_used(struct btree_update *as,
 200                                             struct btree_trans *trans,
 201                                             struct btree *b)
 202 {
 203         struct bch_fs *c = as->c;
 204         struct prealloc_nodes *p = &as->prealloc_nodes[b->c.lock.readers != NULL];
 205         struct btree_path *path;
 206         unsigned level = b->c.level;
 207
 208         BUG_ON(!list_empty(&b->write_blocked));
 209         BUG_ON(b->will_make_reachable != (1UL|(unsigned long) as));
 210
 211         b->will_make_reachable = 0;
 212         closure_put(&as->cl);
 213
 214         clear_btree_node_will_make_reachable(b);
 215         clear_btree_node_accessed(b);
 216         clear_btree_node_dirty_acct(c, b);
 217         clear_btree_node_need_write(b);
 218
 219         mutex_lock(&c->btree_cache.lock);
 220         list_del_init(&b->list);
 221         bch2_btree_node_hash_remove(&c->btree_cache, b);
 222         mutex_unlock(&c->btree_cache.lock);
 223
 224         BUG_ON(p->nr >= ARRAY_SIZE(p->b));
 225         p->b[p->nr++] = b;
 226
 227         six_unlock_intent(&b->c.lock);
 228
 229         trans_for_each_path(trans, path)
 230                 if (path->l[level].b == b) {
 231                         btree_node_unlock(trans, path, level);
 232                         path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
 233                 }
 234 }
 235
 236 static struct btree *__bch2_btree_node_alloc(struct btree_trans *trans,
 237                                              struct disk_reservation *res,
 238                                              struct closure *cl,
 239                                              bool interior_node,
 240                                              unsigned flags)
 241 {
 242         struct bch_fs *c = trans->c;
 243         struct write_point *wp;
 244         struct btree *b;
 245         __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
 246         struct open_buckets ob = { .nr = 0 };
 247         struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
 248         unsigned nr_reserve;
 249         enum alloc_reserve alloc_reserve;
 250         int ret;
 251
 252         if (flags & BTREE_INSERT_USE_RESERVE) {
 253                 nr_reserve      = 0;
 254                 alloc_reserve   = RESERVE_btree_movinggc;
 255         } else {
 256                 nr_reserve      = BTREE_NODE_RESERVE;
 257                 alloc_reserve   = RESERVE_btree;
 258         }
 259
 260         mutex_lock(&c->btree_reserve_cache_lock);
 261         if (c->btree_reserve_cache_nr > nr_reserve) {
 262                 struct btree_alloc *a =
 263                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
 264
 265                 ob = a->ob;
 266                 bkey_copy(&tmp.k, &a->k);
 267                 mutex_unlock(&c->btree_reserve_cache_lock);
 268                 goto mem_alloc;
 269         }
 270         mutex_unlock(&c->btree_reserve_cache_lock);
 271
 272 retry:
 273         ret = bch2_alloc_sectors_start_trans(trans,
 274                                       c->opts.metadata_target ?:
 275                                       c->opts.foreground_target,
 276                                       0,
 277                                       writepoint_ptr(&c->btree_write_point),
 278                                       &devs_have,
 279                                       res->nr_replicas,
 280                                       c->opts.metadata_replicas_required,
 281                                       alloc_reserve, 0, cl, &wp);
 282         if (unlikely(ret))
 283                 return ERR_PTR(ret);
 284
 285         if (wp->sectors_free < btree_sectors(c)) {
 286                 struct open_bucket *ob;
 287                 unsigned i;
 288
 289                 open_bucket_for_each(c, &wp->ptrs, ob, i)
 290                         if (ob->sectors_free < btree_sectors(c))
 291                                 ob->sectors_free = 0;
 292
 293                 bch2_alloc_sectors_done(c, wp);
 294                 goto retry;
 295         }
 296
 297         bkey_btree_ptr_v2_init(&tmp.k);
 298         bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, btree_sectors(c), false);
 299
 300         bch2_open_bucket_get(c, wp, &ob);
 301         bch2_alloc_sectors_done(c, wp);
 302 mem_alloc:
 303         b = bch2_btree_node_mem_alloc(c, interior_node);
 304         six_unlock_write(&b->c.lock);
 305         six_unlock_intent(&b->c.lock);
 306
 307         /* we hold cannibalize_lock: */
 308         BUG_ON(IS_ERR(b));
 309         BUG_ON(b->ob.nr);
 310
 311         bkey_copy(&b->key, &tmp.k);
 312         b->ob = ob;
 313
 314         return b;
 315 }
 316
 317 static struct btree *bch2_btree_node_alloc(struct btree_update *as,
 318                                            struct btree_trans *trans,
 319                                            unsigned level)
 320 {
 321         struct bch_fs *c = as->c;
 322         struct btree *b;
 323         struct prealloc_nodes *p = &as->prealloc_nodes[!!level];
 324         int ret;
 325
 326         BUG_ON(level >= BTREE_MAX_DEPTH);
 327         BUG_ON(!p->nr);
 328
 329         b = p->b[--p->nr];
 330
 331         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
 332         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
 333
 334         set_btree_node_accessed(b);
 335         set_btree_node_dirty_acct(c, b);
 336         set_btree_node_need_write(b);
 337
 338         bch2_bset_init_first(b, &b->data->keys);
 339         b->c.level      = level;
 340         b->c.btree_id   = as->btree_id;
 341         b->version_ondisk = c->sb.version;
 342
 343         memset(&b->nr, 0, sizeof(b->nr));
 344         b->data->magic = cpu_to_le64(bset_magic(c));
 345         memset(&b->data->_ptr, 0, sizeof(b->data->_ptr));
 346         b->data->flags = 0;
 347         SET_BTREE_NODE_ID(b->data, as->btree_id);
 348         SET_BTREE_NODE_LEVEL(b->data, level);
 349
 350         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 351                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
 352
 353                 bp->v.mem_ptr           = 0;
 354                 bp->v.seq               = b->data->keys.seq;
 355                 bp->v.sectors_written   = 0;
 356         }
 357
 358         SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
 359
 360         bch2_btree_build_aux_trees(b);
 361
 362         ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
 363         BUG_ON(ret);
 364
 365         trace_and_count(c, btree_node_alloc, c, b);
 366         return b;
 367 }
 368
 369 static void btree_set_min(struct btree *b, struct bpos pos)
 370 {
 371         if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
 372                 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
 373         b->data->min_key = pos;
 374 }
 375
 376 static void btree_set_max(struct btree *b, struct bpos pos)
 377 {
 378         b->key.k.p = pos;
 379         b->data->max_key = pos;
 380 }
 381
 382 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
 383                                                        struct btree_trans *trans,
 384                                                        struct btree *b)
 385 {
 386         struct btree *n = bch2_btree_node_alloc(as, trans, b->c.level);
 387         struct bkey_format format = bch2_btree_calc_format(b);
 388
 389         /*
 390          * The keys might expand with the new format - if they wouldn't fit in
 391          * the btree node anymore, use the old format for now:
 392          */
 393         if (!bch2_btree_node_format_fits(as->c, b, &format))
 394                 format = b->format;
 395
 396         SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
 397
 398         btree_set_min(n, b->data->min_key);
 399         btree_set_max(n, b->data->max_key);
 400
 401         n->data->format         = format;
 402         btree_node_set_format(n, format);
 403
 404         bch2_btree_sort_into(as->c, n, b);
 405
 406         btree_node_reset_sib_u64s(n);
 407         return n;
 408 }
 409
 410 static struct btree *__btree_root_alloc(struct btree_update *as,
 411                                 struct btree_trans *trans, unsigned level)
 412 {
 413         struct btree *b = bch2_btree_node_alloc(as, trans, level);
 414
 415         btree_set_min(b, POS_MIN);
 416         btree_set_max(b, SPOS_MAX);
 417         b->data->format = bch2_btree_calc_format(b);
 418
 419         btree_node_set_format(b, b->data->format);
 420         bch2_btree_build_aux_trees(b);
 421
 422         return b;
 423 }
 424
 425 static void bch2_btree_reserve_put(struct btree_update *as, struct btree_trans *trans)
 426 {
 427         struct bch_fs *c = as->c;
 428         struct prealloc_nodes *p;
 429
 430         for (p = as->prealloc_nodes;
 431              p < as->prealloc_nodes + ARRAY_SIZE(as->prealloc_nodes);
 432              p++) {
 433                 while (p->nr) {
 434                         struct btree *b = p->b[--p->nr];
 435
 436                         mutex_lock(&c->btree_reserve_cache_lock);
 437
 438                         if (c->btree_reserve_cache_nr <
 439                             ARRAY_SIZE(c->btree_reserve_cache)) {
 440                                 struct btree_alloc *a =
 441                                         &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
 442
 443                                 a->ob = b->ob;
 444                                 b->ob.nr = 0;
 445                                 bkey_copy(&a->k, &b->key);
 446                         } else {
 447                                 bch2_open_buckets_put(c, &b->ob);
 448                         }
 449
 450                         mutex_unlock(&c->btree_reserve_cache_lock);
 451
 452                         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
 453                         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
 454                         __btree_node_free(c, b);
 455                         six_unlock_write(&b->c.lock);
 456                         six_unlock_intent(&b->c.lock);
 457                 }
 458         }
 459 }
 460
 461 static int bch2_btree_reserve_get(struct btree_trans *trans,
 462                                   struct btree_update *as,
 463                                   unsigned nr_nodes[2],
 464                                   unsigned flags,
 465                                   struct closure *cl)
 466 {
 467         struct bch_fs *c = as->c;
 468         struct btree *b;
 469         unsigned interior;
 470         int ret = 0;
 471
 472         BUG_ON(nr_nodes[0] + nr_nodes[1] > BTREE_RESERVE_MAX);
 473
 474         /*
 475          * Protects reaping from the btree node cache and using the btree node
 476          * open bucket reserve:
 477          *
 478          * BTREE_INSERT_NOWAIT only applies to btree node allocation, not
 479          * blocking on this lock:
 480          */
 481         ret = bch2_btree_cache_cannibalize_lock(c, cl);
 482         if (ret)
 483                 return ret;
 484
 485         for (interior = 0; interior < 2; interior++) {
 486                 struct prealloc_nodes *p = as->prealloc_nodes + interior;
 487
 488                 while (p->nr < nr_nodes[interior]) {
 489                         b = __bch2_btree_node_alloc(trans, &as->disk_res,
 490                                         flags & BTREE_INSERT_NOWAIT ? NULL : cl,
 491                                         interior, flags);
 492                         if (IS_ERR(b)) {
 493                                 ret = PTR_ERR(b);
 494                                 goto err;
 495                         }
 496
 497                         p->b[p->nr++] = b;
 498                 }
 499         }
 500 err:
 501         bch2_btree_cache_cannibalize_unlock(c);
 502         return ret;
 503 }
 504
 505 /* Asynchronous interior node update machinery */
 506
 507 static void bch2_btree_update_free(struct btree_update *as, struct btree_trans *trans)
 508 {
 509         struct bch_fs *c = as->c;
 510
 511         if (as->took_gc_lock)
 512                 up_read(&c->gc_lock);
 513         as->took_gc_lock = false;
 514
 515         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 516
 517         bch2_journal_pin_drop(&c->journal, &as->journal);
 518         bch2_journal_pin_flush(&c->journal, &as->journal);
 519         bch2_disk_reservation_put(c, &as->disk_res);
 520         bch2_btree_reserve_put(as, trans);
 521
 522         bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_total],
 523                                as->start_time);
 524
 525         mutex_lock(&c->btree_interior_update_lock);
 526         list_del(&as->unwritten_list);
 527         list_del(&as->list);
 528
 529         closure_debug_destroy(&as->cl);
 530         mempool_free(as, &c->btree_interior_update_pool);
 531
 532         /*
 533          * Have to do the wakeup with btree_interior_update_lock still held,
 534          * since being on btree_interior_update_list is our ref on @c:
 535          */
 536         closure_wake_up(&c->btree_interior_update_wait);
 537
 538         mutex_unlock(&c->btree_interior_update_lock);
 539 }
 540
 541 static void btree_update_add_key(struct btree_update *as,
 542                                  struct keylist *keys, struct btree *b)
 543 {
 544         struct bkey_i *k = &b->key;
 545
 546         BUG_ON(bch2_keylist_u64s(keys) + k->k.u64s >
 547                ARRAY_SIZE(as->_old_keys));
 548
 549         bkey_copy(keys->top, k);
 550         bkey_i_to_btree_ptr_v2(keys->top)->v.mem_ptr = b->c.level + 1;
 551
 552         bch2_keylist_push(keys);
 553 }
 554
 555 /*
 556  * The transactional part of an interior btree node update, where we journal the
 557  * update we did to the interior node and update alloc info:
 558  */
 559 static int btree_update_nodes_written_trans(struct btree_trans *trans,
 560                                             struct btree_update *as)
 561 {
 562         struct bkey_i *k;
 563         int ret;
 564
 565         ret = darray_make_room(&trans->extra_journal_entries, as->journal_u64s);
 566         if (ret)
 567                 return ret;
 568
 569         memcpy(&darray_top(trans->extra_journal_entries),
 570                as->journal_entries,
 571                as->journal_u64s * sizeof(u64));
 572         trans->extra_journal_entries.nr += as->journal_u64s;
 573
 574         trans->journal_pin = &as->journal;
 575
 576         for_each_keylist_key(&as->old_keys, k) {
 577                 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
 578
 579                 ret = bch2_trans_mark_old(trans, as->btree_id, level, bkey_i_to_s_c(k), 0);
 580                 if (ret)
 581                         return ret;
 582         }
 583
 584         for_each_keylist_key(&as->new_keys, k) {
 585                 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
 586
 587                 ret = bch2_trans_mark_new(trans, as->btree_id, level, k, 0);
 588                 if (ret)
 589                         return ret;
 590         }
 591
 592         return 0;
 593 }
 594
 595 static void btree_update_nodes_written(struct btree_update *as)
 596 {
 597         struct bch_fs *c = as->c;
 598         struct btree *b;
 599         struct btree_trans trans;
 600         u64 journal_seq = 0;
 601         unsigned i;
 602         int ret;
 603
 604         bch2_trans_init(&trans, c, 0, 512);
 605         /*
 606          * If we're already in an error state, it might be because a btree node
 607          * was never written, and we might be trying to free that same btree
 608          * node here, but it won't have been marked as allocated and we'll see
 609          * spurious disk usage inconsistencies in the transactional part below
 610          * if we don't skip it:
 611          */
 612         ret = bch2_journal_error(&c->journal);
 613         if (ret)
 614                 goto err;
 615
 616         /*
 617          * Wait for any in flight writes to finish before we free the old nodes
 618          * on disk:
 619          */
 620         for (i = 0; i < as->nr_old_nodes; i++) {
 621                 __le64 seq;
 622
 623                 b = as->old_nodes[i];
 624
 625                 btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
 626                 seq = b->data ? b->data->keys.seq : 0;
 627                 six_unlock_read(&b->c.lock);
 628
 629                 if (seq == as->old_nodes_seq[i])
 630                         wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight_inner,
 631                                        TASK_UNINTERRUPTIBLE);
 632         }
 633
 634         /*
 635          * We did an update to a parent node where the pointers we added pointed
 636          * to child nodes that weren't written yet: now, the child nodes have
 637          * been written so we can write out the update to the interior node.
 638          */
 639
 640         /*
 641          * We can't call into journal reclaim here: we'd block on the journal
 642          * reclaim lock, but we may need to release the open buckets we have
 643          * pinned in order for other btree updates to make forward progress, and
 644          * journal reclaim does btree updates when flushing bkey_cached entries,
 645          * which may require allocations as well.
 646          */
 647         ret = commit_do(&trans, &as->disk_res, &journal_seq,
 648                         BTREE_INSERT_NOFAIL|
 649                         BTREE_INSERT_NOCHECK_RW|
 650                         BTREE_INSERT_USE_RESERVE|
 651                         BTREE_INSERT_JOURNAL_RECLAIM|
 652                         JOURNAL_WATERMARK_reserved,
 653                         btree_update_nodes_written_trans(&trans, as));
 654         bch2_trans_unlock(&trans);
 655
 656         bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
 657                              "%s(): error %s", __func__, bch2_err_str(ret));
 658 err:
 659         if (as->b) {
 660                 struct btree_path *path;
 661
 662                 b = as->b;
 663                 path = get_unlocked_mut_path(&trans, as->btree_id, b->c.level, b->key.k.p);
 664                 /*
 665                  * @b is the node we did the final insert into:
 666                  *
 667                  * On failure to get a journal reservation, we still have to
 668                  * unblock the write and allow most of the write path to happen
 669                  * so that shutdown works, but the i->journal_seq mechanism
 670                  * won't work to prevent the btree write from being visible (we
 671                  * didn't get a journal sequence number) - instead
 672                  * __bch2_btree_node_write() doesn't do the actual write if
 673                  * we're in journal error state:
 674                  */
 675
 676                 /*
 677                  * Ensure transaction is unlocked before using
 678                  * btree_node_lock_nopath() (the use of which is always suspect,
 679                  * we need to work on removing this in the future)
 680                  *
 681                  * It should be, but get_unlocked_mut_path() -> bch2_path_get()
 682                  * calls bch2_path_upgrade(), before we call path_make_mut(), so
 683                  * we may rarely end up with a locked path besides the one we
 684                  * have here:
 685                  */
 686                 bch2_trans_unlock(&trans);
 687                 btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_intent);
 688                 mark_btree_node_locked(&trans, path, b->c.level, SIX_LOCK_intent);
 689                 bch2_btree_path_level_init(&trans, path, b);
 690
 691                 bch2_btree_node_lock_write_nofail(&trans, path, &b->c);
 692
 693                 mutex_lock(&c->btree_interior_update_lock);
 694
 695                 list_del(&as->write_blocked_list);
 696                 if (list_empty(&b->write_blocked))
 697                         clear_btree_node_write_blocked(b);
 698
 699                 /*
 700                  * Node might have been freed, recheck under
 701                  * btree_interior_update_lock:
 702                  */
 703                 if (as->b == b) {
 704                         struct bset *i = btree_bset_last(b);
 705
 706                         BUG_ON(!b->c.level);
 707                         BUG_ON(!btree_node_dirty(b));
 708
 709                         if (!ret) {
 710                                 i->journal_seq = cpu_to_le64(
 711                                                              max(journal_seq,
 712                                                                  le64_to_cpu(i->journal_seq)));
 713
 714                                 bch2_btree_add_journal_pin(c, b, journal_seq);
 715                         } else {
 716                                 /*
 717                                  * If we didn't get a journal sequence number we
 718                                  * can't write this btree node, because recovery
 719                                  * won't know to ignore this write:
 720                                  */
 721                                 set_btree_node_never_write(b);
 722                         }
 723                 }
 724
 725                 mutex_unlock(&c->btree_interior_update_lock);
 726
 727                 mark_btree_node_locked_noreset(path, b->c.level, SIX_LOCK_intent);
 728                 six_unlock_write(&b->c.lock);
 729
 730                 btree_node_write_if_need(c, b, SIX_LOCK_intent);
 731                 btree_node_unlock(&trans, path, b->c.level);
 732                 bch2_path_put(&trans, path, true);
 733         }
 734
 735         bch2_journal_pin_drop(&c->journal, &as->journal);
 736
 737         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 738
 739         mutex_lock(&c->btree_interior_update_lock);
 740         for (i = 0; i < as->nr_new_nodes; i++) {
 741                 b = as->new_nodes[i];
 742
 743                 BUG_ON(b->will_make_reachable != (unsigned long) as);
 744                 b->will_make_reachable = 0;
 745                 clear_btree_node_will_make_reachable(b);
 746         }
 747         mutex_unlock(&c->btree_interior_update_lock);
 748
 749         for (i = 0; i < as->nr_new_nodes; i++) {
 750                 b = as->new_nodes[i];
 751
 752                 btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
 753                 btree_node_write_if_need(c, b, SIX_LOCK_read);
 754                 six_unlock_read(&b->c.lock);
 755         }
 756
 757         for (i = 0; i < as->nr_open_buckets; i++)
 758                 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
 759
 760         bch2_btree_update_free(as, &trans);
 761         bch2_trans_exit(&trans);
 762 }
 763
 764 static void btree_interior_update_work(struct work_struct *work)
 765 {
 766         struct bch_fs *c =
 767                 container_of(work, struct bch_fs, btree_interior_update_work);
 768         struct btree_update *as;
 769
 770         while (1) {
 771                 mutex_lock(&c->btree_interior_update_lock);
 772                 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
 773                                               struct btree_update, unwritten_list);
 774                 if (as && !as->nodes_written)
 775                         as = NULL;
 776                 mutex_unlock(&c->btree_interior_update_lock);
 777
 778                 if (!as)
 779                         break;
 780
 781                 btree_update_nodes_written(as);
 782         }
 783 }
 784
 785 static void btree_update_set_nodes_written(struct closure *cl)
 786 {
 787         struct btree_update *as = container_of(cl, struct btree_update, cl);
 788         struct bch_fs *c = as->c;
 789
 790         mutex_lock(&c->btree_interior_update_lock);
 791         as->nodes_written = true;
 792         mutex_unlock(&c->btree_interior_update_lock);
 793
 794         queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
 795 }
 796
 797 /*
 798  * We're updating @b with pointers to nodes that haven't finished writing yet:
 799  * block @b from being written until @as completes
 800  */
 801 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
 802 {
 803         struct bch_fs *c = as->c;
 804
 805         mutex_lock(&c->btree_interior_update_lock);
 806         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 807
 808         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 809         BUG_ON(!btree_node_dirty(b));
 810         BUG_ON(!b->c.level);
 811
 812         as->mode        = BTREE_INTERIOR_UPDATING_NODE;
 813         as->b           = b;
 814
 815         set_btree_node_write_blocked(b);
 816         list_add(&as->write_blocked_list, &b->write_blocked);
 817
 818         mutex_unlock(&c->btree_interior_update_lock);
 819 }
 820
 821 static void btree_update_reparent(struct btree_update *as,
 822                                   struct btree_update *child)
 823 {
 824         struct bch_fs *c = as->c;
 825
 826         lockdep_assert_held(&c->btree_interior_update_lock);
 827
 828         child->b = NULL;
 829         child->mode = BTREE_INTERIOR_UPDATING_AS;
 830
 831         bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL);
 832 }
 833
 834 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
 835 {
 836         struct bkey_i *insert = &b->key;
 837         struct bch_fs *c = as->c;
 838
 839         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 840
 841         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
 842                ARRAY_SIZE(as->journal_entries));
 843
 844         as->journal_u64s +=
 845                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
 846                                   BCH_JSET_ENTRY_btree_root,
 847                                   b->c.btree_id, b->c.level,
 848                                   insert, insert->k.u64s);
 849
 850         mutex_lock(&c->btree_interior_update_lock);
 851         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 852
 853         as->mode        = BTREE_INTERIOR_UPDATING_ROOT;
 854         mutex_unlock(&c->btree_interior_update_lock);
 855 }
 856
 857 /*
 858  * bch2_btree_update_add_new_node:
 859  *
 860  * This causes @as to wait on @b to be written, before it gets to
 861  * bch2_btree_update_nodes_written
 862  *
 863  * Additionally, it sets b->will_make_reachable to prevent any additional writes
 864  * to @b from happening besides the first until @b is reachable on disk
 865  *
 866  * And it adds @b to the list of @as's new nodes, so that we can update sector
 867  * counts in bch2_btree_update_nodes_written:
 868  */
 869 static void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
 870 {
 871         struct bch_fs *c = as->c;
 872
 873         closure_get(&as->cl);
 874
 875         mutex_lock(&c->btree_interior_update_lock);
 876         BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
 877         BUG_ON(b->will_make_reachable);
 878
 879         as->new_nodes[as->nr_new_nodes++] = b;
 880         b->will_make_reachable = 1UL|(unsigned long) as;
 881         set_btree_node_will_make_reachable(b);
 882
 883         mutex_unlock(&c->btree_interior_update_lock);
 884
 885         btree_update_add_key(as, &as->new_keys, b);
 886
 887         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 888                 unsigned bytes = vstruct_end(&b->data->keys) - (void *) b->data;
 889                 unsigned sectors = round_up(bytes, block_bytes(c)) >> 9;
 890
 891                 bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written =
 892                         cpu_to_le16(sectors);
 893         }
 894 }
 895
 896 /*
 897  * returns true if @b was a new node
 898  */
 899 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
 900 {
 901         struct btree_update *as;
 902         unsigned long v;
 903         unsigned i;
 904
 905         mutex_lock(&c->btree_interior_update_lock);
 906         /*
 907          * When b->will_make_reachable != 0, it owns a ref on as->cl that's
 908          * dropped when it gets written by bch2_btree_complete_write - the
 909          * xchg() is for synchronization with bch2_btree_complete_write:
 910          */
 911         v = xchg(&b->will_make_reachable, 0);
 912         clear_btree_node_will_make_reachable(b);
 913         as = (struct btree_update *) (v & ~1UL);
 914
 915         if (!as) {
 916                 mutex_unlock(&c->btree_interior_update_lock);
 917                 return;
 918         }
 919
 920         for (i = 0; i < as->nr_new_nodes; i++)
 921                 if (as->new_nodes[i] == b)
 922                         goto found;
 923
 924         BUG();
 925 found:
 926         array_remove_item(as->new_nodes, as->nr_new_nodes, i);
 927         mutex_unlock(&c->btree_interior_update_lock);
 928
 929         if (v & 1)
 930                 closure_put(&as->cl);
 931 }
 932
 933 static void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
 934 {
 935         while (b->ob.nr)
 936                 as->open_buckets[as->nr_open_buckets++] =
 937                         b->ob.v[--b->ob.nr];
 938 }
 939
 940 /*
 941  * @b is being split/rewritten: it may have pointers to not-yet-written btree
 942  * nodes and thus outstanding btree_updates - redirect @b's
 943  * btree_updates to point to this btree_update:
 944  */
 945 static void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 946                                                       struct btree *b)
 947 {
 948         struct bch_fs *c = as->c;
 949         struct btree_update *p, *n;
 950         struct btree_write *w;
 951
 952         set_btree_node_dying(b);
 953
 954         if (btree_node_fake(b))
 955                 return;
 956
 957         mutex_lock(&c->btree_interior_update_lock);
 958
 959         /*
 960          * Does this node have any btree_update operations preventing
 961          * it from being written?
 962          *
 963          * If so, redirect them to point to this btree_update: we can
 964          * write out our new nodes, but we won't make them visible until those
 965          * operations complete
 966          */
 967         list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
 968                 list_del_init(&p->write_blocked_list);
 969                 btree_update_reparent(as, p);
 970
 971                 /*
 972                  * for flush_held_btree_writes() waiting on updates to flush or
 973                  * nodes to be writeable:
 974                  */
 975                 closure_wake_up(&c->btree_interior_update_wait);
 976         }
 977
 978         clear_btree_node_dirty_acct(c, b);
 979         clear_btree_node_need_write(b);
 980         clear_btree_node_write_blocked(b);
 981
 982         /*
 983          * Does this node have unwritten data that has a pin on the journal?
 984          *
 985          * If so, transfer that pin to the btree_update operation -
 986          * note that if we're freeing multiple nodes, we only need to keep the
 987          * oldest pin of any of the nodes we're freeing. We'll release the pin
 988          * when the new nodes are persistent and reachable on disk:
 989          */
 990         w = btree_current_write(b);
 991         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 992         bch2_journal_pin_drop(&c->journal, &w->journal);
 993
 994         w = btree_prev_write(b);
 995         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 996         bch2_journal_pin_drop(&c->journal, &w->journal);
 997
 998         mutex_unlock(&c->btree_interior_update_lock);
 999
1000         /*
1001          * Is this a node that isn't reachable on disk yet?
1002          *
1003          * Nodes that aren't reachable yet have writes blocked until they're
1004          * reachable - now that we've cancelled any pending writes and moved
1005          * things waiting on that write to wait on this update, we can drop this
1006          * node from the list of nodes that the other update is making
1007          * reachable, prior to freeing it:
1008          */
1009         btree_update_drop_new_node(c, b);
1010
1011         btree_update_add_key(as, &as->old_keys, b);
1012
1013         as->old_nodes[as->nr_old_nodes] = b;
1014         as->old_nodes_seq[as->nr_old_nodes] = b->data->keys.seq;
1015         as->nr_old_nodes++;
1016 }
1017
1018 static void bch2_btree_update_done(struct btree_update *as, struct btree_trans *trans)
1019 {
1020         struct bch_fs *c = as->c;
1021         u64 start_time = as->start_time;
1022
1023         BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
1024
1025         if (as->took_gc_lock)
1026                 up_read(&as->c->gc_lock);
1027         as->took_gc_lock = false;
1028
1029         bch2_btree_reserve_put(as, trans);
1030
1031         continue_at(&as->cl, btree_update_set_nodes_written,
1032                     as->c->btree_interior_update_worker);
1033
1034         bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_foreground],
1035                                start_time);
1036 }
1037
1038 static struct btree_update *
1039 bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
1040                         unsigned level, bool split, unsigned flags)
1041 {
1042         struct bch_fs *c = trans->c;
1043         struct btree_update *as;
1044         u64 start_time = local_clock();
1045         int disk_res_flags = (flags & BTREE_INSERT_NOFAIL)
1046                 ? BCH_DISK_RESERVATION_NOFAIL : 0;
1047         unsigned nr_nodes[2] = { 0, 0 };
1048         unsigned update_level = level;
1049         int journal_flags = flags & JOURNAL_WATERMARK_MASK;
1050         int ret = 0;
1051         u32 restart_count = trans->restart_count;
1052
1053         BUG_ON(!path->should_be_locked);
1054
1055         if (flags & BTREE_INSERT_JOURNAL_RECLAIM)
1056                 journal_flags |= JOURNAL_RES_GET_NONBLOCK;
1057
1058         while (1) {
1059                 nr_nodes[!!update_level] += 1 + split;
1060                 update_level++;
1061
1062                 ret = bch2_btree_path_upgrade(trans, path, update_level + 1);
1063                 if (ret)
1064                         return ERR_PTR(ret);
1065
1066                 if (!btree_path_node(path, update_level)) {
1067                         /* Allocating new root? */
1068                         nr_nodes[1] += split;
1069                         update_level = BTREE_MAX_DEPTH;
1070                         break;
1071                 }
1072
1073                 if (bch2_btree_node_insert_fits(c, path->l[update_level].b,
1074                                         BKEY_BTREE_PTR_U64s_MAX * (1 + split)))
1075                         break;
1076
1077                 split = true;
1078         }
1079
1080         if (flags & BTREE_INSERT_GC_LOCK_HELD)
1081                 lockdep_assert_held(&c->gc_lock);
1082         else if (!down_read_trylock(&c->gc_lock)) {
1083                 bch2_trans_unlock(trans);
1084                 down_read(&c->gc_lock);
1085                 ret = bch2_trans_relock(trans);
1086                 if (ret) {
1087                         up_read(&c->gc_lock);
1088                         return ERR_PTR(ret);
1089                 }
1090         }
1091
1092         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
1093         memset(as, 0, sizeof(*as));
1094         closure_init(&as->cl, NULL);
1095         as->c           = c;
1096         as->start_time  = start_time;
1097         as->mode        = BTREE_INTERIOR_NO_UPDATE;
1098         as->took_gc_lock = !(flags & BTREE_INSERT_GC_LOCK_HELD);
1099         as->btree_id    = path->btree_id;
1100         as->update_level = update_level;
1101         INIT_LIST_HEAD(&as->list);
1102         INIT_LIST_HEAD(&as->unwritten_list);
1103         INIT_LIST_HEAD(&as->write_blocked_list);
1104         bch2_keylist_init(&as->old_keys, as->_old_keys);
1105         bch2_keylist_init(&as->new_keys, as->_new_keys);
1106         bch2_keylist_init(&as->parent_keys, as->inline_keys);
1107
1108         mutex_lock(&c->btree_interior_update_lock);
1109         list_add_tail(&as->list, &c->btree_interior_update_list);
1110         mutex_unlock(&c->btree_interior_update_lock);
1111
1112         /*
1113          * We don't want to allocate if we're in an error state, that can cause
1114          * deadlock on emergency shutdown due to open buckets getting stuck in
1115          * the btree_reserve_cache after allocator shutdown has cleared it out.
1116          * This check needs to come after adding us to the btree_interior_update
1117          * list but before calling bch2_btree_reserve_get, to synchronize with
1118          * __bch2_fs_read_only().
1119          */
1120         ret = bch2_journal_error(&c->journal);
1121         if (ret)
1122                 goto err;
1123
1124         ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
1125                                       BTREE_UPDATE_JOURNAL_RES,
1126                                       journal_flags|JOURNAL_RES_GET_NONBLOCK);
1127         if (ret) {
1128                 bch2_trans_unlock(trans);
1129
1130                 if (flags & BTREE_INSERT_JOURNAL_RECLAIM) {
1131                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
1132                         goto err;
1133                 }
1134
1135                 ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
1136                                               BTREE_UPDATE_JOURNAL_RES,
1137                                               journal_flags);
1138                 if (ret) {
1139                         trace_and_count(c, trans_restart_journal_preres_get, trans, _RET_IP_, journal_flags);
1140                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_journal_preres_get);
1141                         goto err;
1142                 }
1143
1144                 ret = bch2_trans_relock(trans);
1145                 if (ret)
1146                         goto err;
1147         }
1148
1149         ret = bch2_disk_reservation_get(c, &as->disk_res,
1150                         (nr_nodes[0] + nr_nodes[1]) * btree_sectors(c),
1151                         c->opts.metadata_replicas,
1152                         disk_res_flags);
1153         if (ret)
1154                 goto err;
1155
1156         ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, NULL);
1157         if (bch2_err_matches(ret, ENOSPC) ||
1158             bch2_err_matches(ret, ENOMEM)) {
1159                 struct closure cl;
1160
1161                 closure_init_stack(&cl);
1162
1163                 do {
1164                         ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, &cl);
1165
1166                         bch2_trans_unlock(trans);
1167                         closure_sync(&cl);
1168                 } while (bch2_err_matches(ret, BCH_ERR_operation_blocked));
1169         }
1170
1171         if (ret) {
1172                 trace_and_count(c, btree_reserve_get_fail, trans->fn,
1173                                 _RET_IP_, nr_nodes[0] + nr_nodes[1], ret);
1174                 goto err;
1175         }
1176
1177         ret = bch2_trans_relock(trans);
1178         if (ret)
1179                 goto err;
1180
1181         bch2_trans_verify_not_restarted(trans, restart_count);
1182         return as;
1183 err:
1184         bch2_btree_update_free(as, trans);
1185         return ERR_PTR(ret);
1186 }
1187
1188 /* Btree root updates: */
1189
1190 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1191 {
1192         /* Root nodes cannot be reaped */
1193         mutex_lock(&c->btree_cache.lock);
1194         list_del_init(&b->list);
1195         mutex_unlock(&c->btree_cache.lock);
1196
1197         mutex_lock(&c->btree_root_lock);
1198         BUG_ON(btree_node_root(c, b) &&
1199                (b->c.level < btree_node_root(c, b)->c.level ||
1200                 !btree_node_dying(btree_node_root(c, b))));
1201
1202         btree_node_root(c, b) = b;
1203         mutex_unlock(&c->btree_root_lock);
1204
1205         bch2_recalc_btree_reserve(c);
1206 }
1207
1208 /**
1209  * bch_btree_set_root - update the root in memory and on disk
1210  *
1211  * To ensure forward progress, the current task must not be holding any
1212  * btree node write locks. However, you must hold an intent lock on the
1213  * old root.
1214  *
1215  * Note: This allocates a journal entry but doesn't add any keys to
1216  * it.  All the btree roots are part of every journal write, so there
1217  * is nothing new to be done.  This just guarantees that there is a
1218  * journal write.
1219  */
1220 static void bch2_btree_set_root(struct btree_update *as,
1221                                 struct btree_trans *trans,
1222                                 struct btree_path *path,
1223                                 struct btree *b)
1224 {
1225         struct bch_fs *c = as->c;
1226         struct btree *old;
1227
1228         trace_and_count(c, btree_node_set_root, c, b);
1229
1230         old = btree_node_root(c, b);
1231
1232         /*
1233          * Ensure no one is using the old root while we switch to the
1234          * new root:
1235          */
1236         bch2_btree_node_lock_write_nofail(trans, path, &old->c);
1237
1238         bch2_btree_set_root_inmem(c, b);
1239
1240         btree_update_updated_root(as, b);
1241
1242         /*
1243          * Unlock old root after new root is visible:
1244          *
1245          * The new root isn't persistent, but that's ok: we still have
1246          * an intent lock on the new root, and any updates that would
1247          * depend on the new root would have to update the new root.
1248          */
1249         bch2_btree_node_unlock_write(trans, path, old);
1250 }
1251
1252 /* Interior node updates: */
1253
1254 static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
1255                                         struct btree_trans *trans,
1256                                         struct btree_path *path,
1257                                         struct btree *b,
1258                                         struct btree_node_iter *node_iter,
1259                                         struct bkey_i *insert)
1260 {
1261         struct bch_fs *c = as->c;
1262         struct bkey_packed *k;
1263         struct printbuf buf = PRINTBUF;
1264         unsigned long old, new, v;
1265
1266         BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 &&
1267                !btree_ptr_sectors_written(insert));
1268
1269         if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
1270                 bch2_journal_key_overwritten(c, b->c.btree_id, b->c.level, insert->k.p);
1271
1272         if (bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1273                               btree_node_type(b), WRITE, &buf) ?:
1274             bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert), &buf)) {
1275                 printbuf_reset(&buf);
1276                 prt_printf(&buf, "inserting invalid bkey\n  ");
1277                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
1278                 prt_printf(&buf, "\n  ");
1279                 bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1280                                   btree_node_type(b), WRITE, &buf);
1281                 bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert), &buf);
1282
1283                 bch2_fs_inconsistent(c, "%s", buf.buf);
1284                 dump_stack();
1285         }
1286
1287         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1288                ARRAY_SIZE(as->journal_entries));
1289
1290         as->journal_u64s +=
1291                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1292                                   BCH_JSET_ENTRY_btree_keys,
1293                                   b->c.btree_id, b->c.level,
1294                                   insert, insert->k.u64s);
1295
1296         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1297                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1298                 bch2_btree_node_iter_advance(node_iter, b);
1299
1300         bch2_btree_bset_insert_key(trans, path, b, node_iter, insert);
1301         set_btree_node_dirty_acct(c, b);
1302
1303         v = READ_ONCE(b->flags);
1304         do {
1305                 old = new = v;
1306
1307                 new &= ~BTREE_WRITE_TYPE_MASK;
1308                 new |= BTREE_WRITE_interior;
1309                 new |= 1 << BTREE_NODE_need_write;
1310         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1311
1312         printbuf_exit(&buf);
1313 }
1314
1315 static void
1316 __bch2_btree_insert_keys_interior(struct btree_update *as,
1317                                   struct btree_trans *trans,
1318                                   struct btree_path *path,
1319                                   struct btree *b,
1320                                   struct btree_node_iter node_iter,
1321                                   struct keylist *keys)
1322 {
1323         struct bkey_i *insert = bch2_keylist_front(keys);
1324         struct bkey_packed *k;
1325
1326         BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1327
1328         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1329                (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1330                 ;
1331
1332         while (!bch2_keylist_empty(keys)) {
1333                 struct bkey_i *k = bch2_keylist_front(keys);
1334
1335                 if (bpos_gt(k->k.p, b->key.k.p))
1336                         break;
1337
1338                 bch2_insert_fixup_btree_ptr(as, trans, path, b, &node_iter, k);
1339                 bch2_keylist_pop_front(keys);
1340         }
1341 }
1342
1343 /*
1344  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1345  * node)
1346  */
1347 static void __btree_split_node(struct btree_update *as,
1348                                struct btree_trans *trans,
1349                                struct btree *b,
1350                                struct btree *n[2])
1351 {
1352         struct bkey_packed *k;
1353         struct bpos n1_pos = POS_MIN;
1354         struct btree_node_iter iter;
1355         struct bset *bsets[2];
1356         struct bkey_format_state format[2];
1357         struct bkey_packed *out[2];
1358         struct bkey uk;
1359         unsigned u64s, n1_u64s = (b->nr.live_u64s * 3) / 5;
1360         int i;
1361
1362         for (i = 0; i < 2; i++) {
1363                 BUG_ON(n[i]->nsets != 1);
1364
1365                 bsets[i] = btree_bset_first(n[i]);
1366                 out[i] = bsets[i]->start;
1367
1368                 SET_BTREE_NODE_SEQ(n[i]->data, BTREE_NODE_SEQ(b->data) + 1);
1369                 bch2_bkey_format_init(&format[i]);
1370         }
1371
1372         u64s = 0;
1373         for_each_btree_node_key(b, k, &iter) {
1374                 if (bkey_deleted(k))
1375                         continue;
1376
1377                 i = u64s >= n1_u64s;
1378                 u64s += k->u64s;
1379                 uk = bkey_unpack_key(b, k);
1380                 if (!i)
1381                         n1_pos = uk.p;
1382                 bch2_bkey_format_add_key(&format[i], &uk);
1383         }
1384
1385         btree_set_min(n[0], b->data->min_key);
1386         btree_set_max(n[0], n1_pos);
1387         btree_set_min(n[1], bpos_successor(n1_pos));
1388         btree_set_max(n[1], b->data->max_key);
1389
1390         for (i = 0; i < 2; i++) {
1391                 bch2_bkey_format_add_pos(&format[i], n[i]->data->min_key);
1392                 bch2_bkey_format_add_pos(&format[i], n[i]->data->max_key);
1393
1394                 n[i]->data->format = bch2_bkey_format_done(&format[i]);
1395                 btree_node_set_format(n[i], n[i]->data->format);
1396         }
1397
1398         u64s = 0;
1399         for_each_btree_node_key(b, k, &iter) {
1400                 if (bkey_deleted(k))
1401                         continue;
1402
1403                 i = u64s >= n1_u64s;
1404                 u64s += k->u64s;
1405
1406                 if (bch2_bkey_transform(&n[i]->format, out[i], bkey_packed(k)
1407                                         ? &b->format: &bch2_bkey_format_current, k))
1408                         out[i]->format = KEY_FORMAT_LOCAL_BTREE;
1409                 else
1410                         bch2_bkey_unpack(b, (void *) out[i], k);
1411
1412                 out[i]->needs_whiteout = false;
1413
1414                 btree_keys_account_key_add(&n[i]->nr, 0, out[i]);
1415                 out[i] = bkey_next(out[i]);
1416         }
1417
1418         for (i = 0; i < 2; i++) {
1419                 bsets[i]->u64s = cpu_to_le16((u64 *) out[i] - bsets[i]->_data);
1420
1421                 BUG_ON(!bsets[i]->u64s);
1422
1423                 set_btree_bset_end(n[i], n[i]->set);
1424
1425                 btree_node_reset_sib_u64s(n[i]);
1426
1427                 bch2_verify_btree_nr_keys(n[i]);
1428
1429                 if (b->c.level)
1430                         btree_node_interior_verify(as->c, n[i]);
1431         }
1432 }
1433
1434 /*
1435  * For updates to interior nodes, we've got to do the insert before we split
1436  * because the stuff we're inserting has to be inserted atomically. Post split,
1437  * the keys might have to go in different nodes and the split would no longer be
1438  * atomic.
1439  *
1440  * Worse, if the insert is from btree node coalescing, if we do the insert after
1441  * we do the split (and pick the pivot) - the pivot we pick might be between
1442  * nodes that were coalesced, and thus in the middle of a child node post
1443  * coalescing:
1444  */
1445 static void btree_split_insert_keys(struct btree_update *as,
1446                                     struct btree_trans *trans,
1447                                     struct btree_path *path,
1448                                     struct btree *b,
1449                                     struct keylist *keys)
1450 {
1451         if (!bch2_keylist_empty(keys) &&
1452             bpos_le(bch2_keylist_front(keys)->k.p, b->data->max_key)) {
1453                 struct btree_node_iter node_iter;
1454
1455                 bch2_btree_node_iter_init(&node_iter, b, &bch2_keylist_front(keys)->k.p);
1456
1457                 __bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
1458
1459                 btree_node_interior_verify(as->c, b);
1460         }
1461 }
1462
1463 static int btree_split(struct btree_update *as, struct btree_trans *trans,
1464                        struct btree_path *path, struct btree *b,
1465                        struct keylist *keys, unsigned flags)
1466 {
1467         struct bch_fs *c = as->c;
1468         struct btree *parent = btree_node_parent(path, b);
1469         struct btree *n1, *n2 = NULL, *n3 = NULL;
1470         struct btree_path *path1 = NULL, *path2 = NULL;
1471         u64 start_time = local_clock();
1472         int ret = 0;
1473
1474         BUG_ON(!parent && (b != btree_node_root(c, b)));
1475         BUG_ON(parent && !btree_node_intent_locked(path, b->c.level + 1));
1476
1477         bch2_btree_interior_update_will_free_node(as, b);
1478
1479         if (b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c)) {
1480                 struct btree *n[2];
1481
1482                 trace_and_count(c, btree_node_split, c, b);
1483
1484                 n[0] = n1 = bch2_btree_node_alloc(as, trans, b->c.level);
1485                 n[1] = n2 = bch2_btree_node_alloc(as, trans, b->c.level);
1486
1487                 __btree_split_node(as, trans, b, n);
1488
1489                 if (keys) {
1490                         btree_split_insert_keys(as, trans, path, n1, keys);
1491                         btree_split_insert_keys(as, trans, path, n2, keys);
1492                         BUG_ON(!bch2_keylist_empty(keys));
1493                 }
1494
1495                 bch2_btree_build_aux_trees(n2);
1496                 bch2_btree_build_aux_trees(n1);
1497
1498                 bch2_btree_update_add_new_node(as, n1);
1499                 bch2_btree_update_add_new_node(as, n2);
1500                 six_unlock_write(&n2->c.lock);
1501                 six_unlock_write(&n1->c.lock);
1502
1503                 path1 = get_unlocked_mut_path(trans, path->btree_id, n1->c.level, n1->key.k.p);
1504                 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1505                 mark_btree_node_locked(trans, path1, n1->c.level, SIX_LOCK_intent);
1506                 bch2_btree_path_level_init(trans, path1, n1);
1507
1508                 path2 = get_unlocked_mut_path(trans, path->btree_id, n2->c.level, n2->key.k.p);
1509                 six_lock_increment(&n2->c.lock, SIX_LOCK_intent);
1510                 mark_btree_node_locked(trans, path2, n2->c.level, SIX_LOCK_intent);
1511                 bch2_btree_path_level_init(trans, path2, n2);
1512
1513                 /*
1514                  * Note that on recursive parent_keys == keys, so we
1515                  * can't start adding new keys to parent_keys before emptying it
1516                  * out (which we did with btree_split_insert_keys() above)
1517                  */
1518                 bch2_keylist_add(&as->parent_keys, &n1->key);
1519                 bch2_keylist_add(&as->parent_keys, &n2->key);
1520
1521                 if (!parent) {
1522                         /* Depth increases, make a new root */
1523                         n3 = __btree_root_alloc(as, trans, b->c.level + 1);
1524
1525                         bch2_btree_update_add_new_node(as, n3);
1526                         six_unlock_write(&n3->c.lock);
1527
1528                         path2->locks_want++;
1529                         BUG_ON(btree_node_locked(path2, n3->c.level));
1530                         six_lock_increment(&n3->c.lock, SIX_LOCK_intent);
1531                         mark_btree_node_locked(trans, path2, n3->c.level, SIX_LOCK_intent);
1532                         bch2_btree_path_level_init(trans, path2, n3);
1533
1534                         n3->sib_u64s[0] = U16_MAX;
1535                         n3->sib_u64s[1] = U16_MAX;
1536
1537                         btree_split_insert_keys(as, trans, path, n3, &as->parent_keys);
1538                 }
1539         } else {
1540                 trace_and_count(c, btree_node_compact, c, b);
1541
1542                 n1 = bch2_btree_node_alloc_replacement(as, trans, b);
1543
1544                 if (keys) {
1545                         btree_split_insert_keys(as, trans, path, n1, keys);
1546                         BUG_ON(!bch2_keylist_empty(keys));
1547                 }
1548
1549                 bch2_btree_build_aux_trees(n1);
1550                 bch2_btree_update_add_new_node(as, n1);
1551                 six_unlock_write(&n1->c.lock);
1552
1553                 path1 = get_unlocked_mut_path(trans, path->btree_id, n1->c.level, n1->key.k.p);
1554                 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1555                 mark_btree_node_locked(trans, path1, n1->c.level, SIX_LOCK_intent);
1556                 bch2_btree_path_level_init(trans, path1, n1);
1557
1558                 if (parent)
1559                         bch2_keylist_add(&as->parent_keys, &n1->key);
1560         }
1561
1562         /* New nodes all written, now make them visible: */
1563
1564         if (parent) {
1565                 /* Split a non root node */
1566                 ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1567                 if (ret)
1568                         goto err;
1569         } else if (n3) {
1570                 bch2_btree_set_root(as, trans, path, n3);
1571         } else {
1572                 /* Root filled up but didn't need to be split */
1573                 bch2_btree_set_root(as, trans, path, n1);
1574         }
1575
1576         if (n3) {
1577                 bch2_btree_update_get_open_buckets(as, n3);
1578                 bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0);
1579         }
1580         if (n2) {
1581                 bch2_btree_update_get_open_buckets(as, n2);
1582                 bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0);
1583         }
1584         bch2_btree_update_get_open_buckets(as, n1);
1585         bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1586
1587         /*
1588          * The old node must be freed (in memory) _before_ unlocking the new
1589          * nodes - else another thread could re-acquire a read lock on the old
1590          * node after another thread has locked and updated the new node, thus
1591          * seeing stale data:
1592          */
1593         bch2_btree_node_free_inmem(trans, path, b);
1594
1595         if (n3)
1596                 bch2_trans_node_add(trans, n3);
1597         if (n2)
1598                 bch2_trans_node_add(trans, n2);
1599         bch2_trans_node_add(trans, n1);
1600
1601         if (n3)
1602                 six_unlock_intent(&n3->c.lock);
1603         if (n2)
1604                 six_unlock_intent(&n2->c.lock);
1605         six_unlock_intent(&n1->c.lock);
1606 out:
1607         if (path2) {
1608                 __bch2_btree_path_unlock(trans, path2);
1609                 bch2_path_put(trans, path2, true);
1610         }
1611         if (path1) {
1612                 __bch2_btree_path_unlock(trans, path1);
1613                 bch2_path_put(trans, path1, true);
1614         }
1615
1616         bch2_trans_verify_locks(trans);
1617
1618         bch2_time_stats_update(&c->times[n2
1619                                ? BCH_TIME_btree_node_split
1620                                : BCH_TIME_btree_node_compact],
1621                                start_time);
1622         return ret;
1623 err:
1624         if (n3)
1625                 bch2_btree_node_free_never_used(as, trans, n3);
1626         if (n2)
1627                 bch2_btree_node_free_never_used(as, trans, n2);
1628         bch2_btree_node_free_never_used(as, trans, n1);
1629         goto out;
1630 }
1631
1632 static void
1633 bch2_btree_insert_keys_interior(struct btree_update *as,
1634                                 struct btree_trans *trans,
1635                                 struct btree_path *path,
1636                                 struct btree *b,
1637                                 struct keylist *keys)
1638 {
1639         struct btree_path *linked;
1640
1641         __bch2_btree_insert_keys_interior(as, trans, path, b,
1642                                           path->l[b->c.level].iter, keys);
1643
1644         btree_update_updated_node(as, b);
1645
1646         trans_for_each_path_with_node(trans, b, linked)
1647                 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1648
1649         bch2_trans_verify_paths(trans);
1650 }
1651
1652 /**
1653  * bch_btree_insert_node - insert bkeys into a given btree node
1654  *
1655  * @iter:               btree iterator
1656  * @keys:               list of keys to insert
1657  * @hook:               insert callback
1658  * @persistent:         if not null, @persistent will wait on journal write
1659  *
1660  * Inserts as many keys as it can into a given btree node, splitting it if full.
1661  * If a split occurred, this function will return early. This can only happen
1662  * for leaf nodes -- inserts into interior nodes have to be atomic.
1663  */
1664 static int bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans,
1665                                   struct btree_path *path, struct btree *b,
1666                                   struct keylist *keys, unsigned flags)
1667 {
1668         struct bch_fs *c = as->c;
1669         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1670         int old_live_u64s = b->nr.live_u64s;
1671         int live_u64s_added, u64s_added;
1672         int ret;
1673
1674         lockdep_assert_held(&c->gc_lock);
1675         BUG_ON(!btree_node_intent_locked(path, b->c.level));
1676         BUG_ON(!b->c.level);
1677         BUG_ON(!as || as->b);
1678         bch2_verify_keylist_sorted(keys);
1679
1680         if (!(local_clock() & 63))
1681                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_split_race);
1682
1683         ret = bch2_btree_node_lock_write(trans, path, &b->c);
1684         if (ret)
1685                 return ret;
1686
1687         bch2_btree_node_prep_for_write(trans, path, b);
1688
1689         if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1690                 bch2_btree_node_unlock_write(trans, path, b);
1691                 goto split;
1692         }
1693
1694         btree_node_interior_verify(c, b);
1695
1696         bch2_btree_insert_keys_interior(as, trans, path, b, keys);
1697
1698         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1699         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1700
1701         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1702                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1703         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1704                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1705
1706         if (u64s_added > live_u64s_added &&
1707             bch2_maybe_compact_whiteouts(c, b))
1708                 bch2_trans_node_reinit_iter(trans, b);
1709
1710         bch2_btree_node_unlock_write(trans, path, b);
1711
1712         btree_node_interior_verify(c, b);
1713         return 0;
1714 split:
1715         /*
1716          * We could attempt to avoid the transaction restart, by calling
1717          * bch2_btree_path_upgrade() and allocating more nodes:
1718          */
1719         if (b->c.level >= as->update_level)
1720                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_split_race);
1721
1722         return btree_split(as, trans, path, b, keys, flags);
1723 }
1724
1725 int bch2_btree_split_leaf(struct btree_trans *trans,
1726                           struct btree_path *path,
1727                           unsigned flags)
1728 {
1729         struct btree *b = path_l(path)->b;
1730         struct btree_update *as;
1731         unsigned l;
1732         int ret = 0;
1733
1734         as = bch2_btree_update_start(trans, path, path->level,
1735                                      true, flags);
1736         if (IS_ERR(as))
1737                 return PTR_ERR(as);
1738
1739         ret = btree_split(as, trans, path, b, NULL, flags);
1740         if (ret) {
1741                 bch2_btree_update_free(as, trans);
1742                 return ret;
1743         }
1744
1745         bch2_btree_update_done(as, trans);
1746
1747         for (l = path->level + 1; btree_node_intent_locked(path, l) && !ret; l++)
1748                 ret = bch2_foreground_maybe_merge(trans, path, l, flags);
1749
1750         return ret;
1751 }
1752
1753 int __bch2_foreground_maybe_merge(struct btree_trans *trans,
1754                                   struct btree_path *path,
1755                                   unsigned level,
1756                                   unsigned flags,
1757                                   enum btree_node_sibling sib)
1758 {
1759         struct bch_fs *c = trans->c;
1760         struct btree_path *sib_path = NULL, *new_path = NULL;
1761         struct btree_update *as;
1762         struct bkey_format_state new_s;
1763         struct bkey_format new_f;
1764         struct bkey_i delete;
1765         struct btree *b, *m, *n, *prev, *next, *parent;
1766         struct bpos sib_pos;
1767         size_t sib_u64s;
1768         u64 start_time = local_clock();
1769         int ret = 0;
1770
1771         BUG_ON(!path->should_be_locked);
1772         BUG_ON(!btree_node_locked(path, level));
1773
1774         b = path->l[level].b;
1775
1776         if ((sib == btree_prev_sib && bpos_eq(b->data->min_key, POS_MIN)) ||
1777             (sib == btree_next_sib && bpos_eq(b->data->max_key, SPOS_MAX))) {
1778                 b->sib_u64s[sib] = U16_MAX;
1779                 return 0;
1780         }
1781
1782         sib_pos = sib == btree_prev_sib
1783                 ? bpos_predecessor(b->data->min_key)
1784                 : bpos_successor(b->data->max_key);
1785
1786         sib_path = bch2_path_get(trans, path->btree_id, sib_pos,
1787                                  U8_MAX, level, BTREE_ITER_INTENT, _THIS_IP_);
1788         ret = bch2_btree_path_traverse(trans, sib_path, false);
1789         if (ret)
1790                 goto err;
1791
1792         btree_path_set_should_be_locked(sib_path);
1793
1794         m = sib_path->l[level].b;
1795
1796         if (btree_node_parent(path, b) !=
1797             btree_node_parent(sib_path, m)) {
1798                 b->sib_u64s[sib] = U16_MAX;
1799                 goto out;
1800         }
1801
1802         if (sib == btree_prev_sib) {
1803                 prev = m;
1804                 next = b;
1805         } else {
1806                 prev = b;
1807                 next = m;
1808         }
1809
1810         if (!bpos_eq(bpos_successor(prev->data->max_key), next->data->min_key)) {
1811                 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1812
1813                 bch2_bpos_to_text(&buf1, prev->data->max_key);
1814                 bch2_bpos_to_text(&buf2, next->data->min_key);
1815                 bch_err(c,
1816                         "%s(): btree topology error:\n"
1817                         "  prev ends at   %s\n"
1818                         "  next starts at %s",
1819                         __func__, buf1.buf, buf2.buf);
1820                 printbuf_exit(&buf1);
1821                 printbuf_exit(&buf2);
1822                 bch2_topology_error(c);
1823                 ret = -EIO;
1824                 goto err;
1825         }
1826
1827         bch2_bkey_format_init(&new_s);
1828         bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
1829         __bch2_btree_calc_format(&new_s, prev);
1830         __bch2_btree_calc_format(&new_s, next);
1831         bch2_bkey_format_add_pos(&new_s, next->data->max_key);
1832         new_f = bch2_bkey_format_done(&new_s);
1833
1834         sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1835                 btree_node_u64s_with_format(m, &new_f);
1836
1837         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1838                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1839                 sib_u64s /= 2;
1840                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1841         }
1842
1843         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1844         sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
1845         b->sib_u64s[sib] = sib_u64s;
1846
1847         if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
1848                 goto out;
1849
1850         parent = btree_node_parent(path, b);
1851         as = bch2_btree_update_start(trans, path, level, false,
1852                          BTREE_INSERT_NOFAIL|
1853                          BTREE_INSERT_USE_RESERVE|
1854                          flags);
1855         ret = PTR_ERR_OR_ZERO(as);
1856         if (ret)
1857                 goto err;
1858
1859         trace_and_count(c, btree_node_merge, c, b);
1860
1861         bch2_btree_interior_update_will_free_node(as, b);
1862         bch2_btree_interior_update_will_free_node(as, m);
1863
1864         n = bch2_btree_node_alloc(as, trans, b->c.level);
1865
1866         SET_BTREE_NODE_SEQ(n->data,
1867                            max(BTREE_NODE_SEQ(b->data),
1868                                BTREE_NODE_SEQ(m->data)) + 1);
1869
1870         btree_set_min(n, prev->data->min_key);
1871         btree_set_max(n, next->data->max_key);
1872
1873         n->data->format  = new_f;
1874         btree_node_set_format(n, new_f);
1875
1876         bch2_btree_sort_into(c, n, prev);
1877         bch2_btree_sort_into(c, n, next);
1878
1879         bch2_btree_build_aux_trees(n);
1880         bch2_btree_update_add_new_node(as, n);
1881         six_unlock_write(&n->c.lock);
1882
1883         new_path = get_unlocked_mut_path(trans, path->btree_id, n->c.level, n->key.k.p);
1884         six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1885         mark_btree_node_locked(trans, new_path, n->c.level, SIX_LOCK_intent);
1886         bch2_btree_path_level_init(trans, new_path, n);
1887
1888         bkey_init(&delete.k);
1889         delete.k.p = prev->key.k.p;
1890         bch2_keylist_add(&as->parent_keys, &delete);
1891         bch2_keylist_add(&as->parent_keys, &n->key);
1892
1893         bch2_trans_verify_paths(trans);
1894
1895         ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1896         if (ret)
1897                 goto err_free_update;
1898
1899         bch2_trans_verify_paths(trans);
1900
1901         bch2_btree_update_get_open_buckets(as, n);
1902         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1903
1904         bch2_btree_node_free_inmem(trans, path, b);
1905         bch2_btree_node_free_inmem(trans, sib_path, m);
1906
1907         bch2_trans_node_add(trans, n);
1908
1909         bch2_trans_verify_paths(trans);
1910
1911         six_unlock_intent(&n->c.lock);
1912
1913         bch2_btree_update_done(as, trans);
1914
1915         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time);
1916 out:
1917 err:
1918         if (new_path)
1919                 bch2_path_put(trans, new_path, true);
1920         bch2_path_put(trans, sib_path, true);
1921         bch2_trans_verify_locks(trans);
1922         return ret;
1923 err_free_update:
1924         bch2_btree_node_free_never_used(as, trans, n);
1925         bch2_btree_update_free(as, trans);
1926         goto out;
1927 }
1928
1929 /**
1930  * bch_btree_node_rewrite - Rewrite/move a btree node
1931  */
1932 int bch2_btree_node_rewrite(struct btree_trans *trans,
1933                             struct btree_iter *iter,
1934                             struct btree *b,
1935                             unsigned flags)
1936 {
1937         struct bch_fs *c = trans->c;
1938         struct btree_path *new_path = NULL;
1939         struct btree *n, *parent;
1940         struct btree_update *as;
1941         int ret;
1942
1943         flags |= BTREE_INSERT_NOFAIL;
1944
1945         parent = btree_node_parent(iter->path, b);
1946         as = bch2_btree_update_start(trans, iter->path, b->c.level,
1947                                      false, flags);
1948         ret = PTR_ERR_OR_ZERO(as);
1949         if (ret)
1950                 goto out;
1951
1952         bch2_btree_interior_update_will_free_node(as, b);
1953
1954         n = bch2_btree_node_alloc_replacement(as, trans, b);
1955
1956         bch2_btree_build_aux_trees(n);
1957         bch2_btree_update_add_new_node(as, n);
1958         six_unlock_write(&n->c.lock);
1959
1960         new_path = get_unlocked_mut_path(trans, iter->btree_id, n->c.level, n->key.k.p);
1961         six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1962         mark_btree_node_locked(trans, new_path, n->c.level, SIX_LOCK_intent);
1963         bch2_btree_path_level_init(trans, new_path, n);
1964
1965         trace_and_count(c, btree_node_rewrite, c, b);
1966
1967         if (parent) {
1968                 bch2_keylist_add(&as->parent_keys, &n->key);
1969                 ret = bch2_btree_insert_node(as, trans, iter->path, parent,
1970                                              &as->parent_keys, flags);
1971                 if (ret)
1972                         goto err;
1973         } else {
1974                 bch2_btree_set_root(as, trans, iter->path, n);
1975         }
1976
1977         bch2_btree_update_get_open_buckets(as, n);
1978         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1979
1980         bch2_btree_node_free_inmem(trans, iter->path, b);
1981
1982         bch2_trans_node_add(trans, n);
1983         six_unlock_intent(&n->c.lock);
1984
1985         bch2_btree_update_done(as, trans);
1986 out:
1987         if (new_path)
1988                 bch2_path_put(trans, new_path, true);
1989         bch2_btree_path_downgrade(trans, iter->path);
1990         return ret;
1991 err:
1992         bch2_btree_node_free_never_used(as, trans, n);
1993         bch2_btree_update_free(as, trans);
1994         goto out;
1995 }
1996
1997 struct async_btree_rewrite {
1998         struct bch_fs           *c;
1999         struct work_struct      work;
2000         enum btree_id           btree_id;
2001         unsigned                level;
2002         struct bpos             pos;
2003         __le64                  seq;
2004 };
2005
2006 static int async_btree_node_rewrite_trans(struct btree_trans *trans,
2007                                           struct async_btree_rewrite *a)
2008 {
2009         struct bch_fs *c = trans->c;
2010         struct btree_iter iter;
2011         struct btree *b;
2012         int ret;
2013
2014         bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos,
2015                                   BTREE_MAX_DEPTH, a->level, 0);
2016         b = bch2_btree_iter_peek_node(&iter);
2017         ret = PTR_ERR_OR_ZERO(b);
2018         if (ret)
2019                 goto out;
2020
2021         if (!b || b->data->keys.seq != a->seq) {
2022                 struct printbuf buf = PRINTBUF;
2023
2024                 if (b)
2025                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
2026                 else
2027                         prt_str(&buf, "(null");
2028                 bch_info(c, "%s: node to rewrite not found:, searching for seq %llu, got\n%s",
2029                          __func__, a->seq, buf.buf);
2030                 printbuf_exit(&buf);
2031                 goto out;
2032         }
2033
2034         ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
2035 out:
2036         bch2_trans_iter_exit(trans, &iter);
2037
2038         return ret;
2039 }
2040
2041 void async_btree_node_rewrite_work(struct work_struct *work)
2042 {
2043         struct async_btree_rewrite *a =
2044                 container_of(work, struct async_btree_rewrite, work);
2045         struct bch_fs *c = a->c;
2046         int ret;
2047
2048         ret = bch2_trans_do(c, NULL, NULL, 0,
2049                       async_btree_node_rewrite_trans(&trans, a));
2050         if (ret)
2051                 bch_err(c, "%s: error %s", __func__, bch2_err_str(ret));
2052         bch2_write_ref_put(c, BCH_WRITE_REF_node_rewrite);
2053         kfree(a);
2054 }
2055
2056 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
2057 {
2058         struct async_btree_rewrite *a;
2059
2060         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_node_rewrite)) {
2061                 bch_err(c, "%s: error getting c->writes ref", __func__);
2062                 return;
2063         }
2064
2065         a = kmalloc(sizeof(*a), GFP_NOFS);
2066         if (!a) {
2067                 bch2_write_ref_put(c, BCH_WRITE_REF_node_rewrite);
2068                 bch_err(c, "%s: error allocating memory", __func__);
2069                 return;
2070         }
2071
2072         a->c            = c;
2073         a->btree_id     = b->c.btree_id;
2074         a->level        = b->c.level;
2075         a->pos          = b->key.k.p;
2076         a->seq          = b->data->keys.seq;
2077
2078         INIT_WORK(&a->work, async_btree_node_rewrite_work);
2079         queue_work(c->btree_interior_update_worker, &a->work);
2080 }
2081
2082 static int __bch2_btree_node_update_key(struct btree_trans *trans,
2083                                         struct btree_iter *iter,
2084                                         struct btree *b, struct btree *new_hash,
2085                                         struct bkey_i *new_key,
2086                                         bool skip_triggers)
2087 {
2088         struct bch_fs *c = trans->c;
2089         struct btree_iter iter2 = { NULL };
2090         struct btree *parent;
2091         int ret;
2092
2093         if (!skip_triggers) {
2094                 ret = bch2_trans_mark_old(trans, b->c.btree_id, b->c.level + 1,
2095                                           bkey_i_to_s_c(&b->key), 0);
2096                 if (ret)
2097                         return ret;
2098
2099                 ret = bch2_trans_mark_new(trans, b->c.btree_id, b->c.level + 1,
2100                                           new_key, 0);
2101                 if (ret)
2102                         return ret;
2103         }
2104
2105         if (new_hash) {
2106                 bkey_copy(&new_hash->key, new_key);
2107                 ret = bch2_btree_node_hash_insert(&c->btree_cache,
2108                                 new_hash, b->c.level, b->c.btree_id);
2109                 BUG_ON(ret);
2110         }
2111
2112         parent = btree_node_parent(iter->path, b);
2113         if (parent) {
2114                 bch2_trans_copy_iter(&iter2, iter);
2115
2116                 iter2.path = bch2_btree_path_make_mut(trans, iter2.path,
2117                                 iter2.flags & BTREE_ITER_INTENT,
2118                                 _THIS_IP_);
2119
2120                 BUG_ON(iter2.path->level != b->c.level);
2121                 BUG_ON(!bpos_eq(iter2.path->pos, new_key->k.p));
2122
2123                 btree_path_set_level_up(trans, iter2.path);
2124
2125                 trans->paths_sorted = false;
2126
2127                 ret   = bch2_btree_iter_traverse(&iter2) ?:
2128                         bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_NORUN);
2129                 if (ret)
2130                         goto err;
2131         } else {
2132                 BUG_ON(btree_node_root(c, b) != b);
2133
2134                 ret = darray_make_room(&trans->extra_journal_entries,
2135                                        jset_u64s(new_key->k.u64s));
2136                 if (ret)
2137                         return ret;
2138
2139                 journal_entry_set((void *) &darray_top(trans->extra_journal_entries),
2140                                   BCH_JSET_ENTRY_btree_root,
2141                                   b->c.btree_id, b->c.level,
2142                                   new_key, new_key->k.u64s);
2143                 trans->extra_journal_entries.nr += jset_u64s(new_key->k.u64s);
2144         }
2145
2146         ret = bch2_trans_commit(trans, NULL, NULL,
2147                                 BTREE_INSERT_NOFAIL|
2148                                 BTREE_INSERT_NOCHECK_RW|
2149                                 BTREE_INSERT_USE_RESERVE|
2150                                 BTREE_INSERT_JOURNAL_RECLAIM|
2151                                 JOURNAL_WATERMARK_reserved);
2152         if (ret)
2153                 goto err;
2154
2155         bch2_btree_node_lock_write_nofail(trans, iter->path, &b->c);
2156
2157         if (new_hash) {
2158                 mutex_lock(&c->btree_cache.lock);
2159                 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
2160                 bch2_btree_node_hash_remove(&c->btree_cache, b);
2161
2162                 bkey_copy(&b->key, new_key);
2163                 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
2164                 BUG_ON(ret);
2165                 mutex_unlock(&c->btree_cache.lock);
2166         } else {
2167                 bkey_copy(&b->key, new_key);
2168         }
2169
2170         bch2_btree_node_unlock_write(trans, iter->path, b);
2171 out:
2172         bch2_trans_iter_exit(trans, &iter2);
2173         return ret;
2174 err:
2175         if (new_hash) {
2176                 mutex_lock(&c->btree_cache.lock);
2177                 bch2_btree_node_hash_remove(&c->btree_cache, b);
2178                 mutex_unlock(&c->btree_cache.lock);
2179         }
2180         goto out;
2181 }
2182
2183 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
2184                                struct btree *b, struct bkey_i *new_key,
2185                                bool skip_triggers)
2186 {
2187         struct bch_fs *c = trans->c;
2188         struct btree *new_hash = NULL;
2189         struct btree_path *path = iter->path;
2190         struct closure cl;
2191         int ret = 0;
2192
2193         ret = bch2_btree_path_upgrade(trans, path, b->c.level + 1);
2194         if (ret)
2195                 return ret;
2196
2197         closure_init_stack(&cl);
2198
2199         /*
2200          * check btree_ptr_hash_val() after @b is locked by
2201          * btree_iter_traverse():
2202          */
2203         if (btree_ptr_hash_val(new_key) != b->hash_val) {
2204                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2205                 if (ret) {
2206                         bch2_trans_unlock(trans);
2207                         closure_sync(&cl);
2208                         ret = bch2_trans_relock(trans);
2209                         if (ret)
2210                                 return ret;
2211                 }
2212
2213                 new_hash = bch2_btree_node_mem_alloc(c, false);
2214         }
2215
2216         path->intent_ref++;
2217         ret = __bch2_btree_node_update_key(trans, iter, b, new_hash,
2218                                            new_key, skip_triggers);
2219         --path->intent_ref;
2220
2221         if (new_hash) {
2222                 mutex_lock(&c->btree_cache.lock);
2223                 list_move(&new_hash->list, &c->btree_cache.freeable);
2224                 mutex_unlock(&c->btree_cache.lock);
2225
2226                 six_unlock_write(&new_hash->c.lock);
2227                 six_unlock_intent(&new_hash->c.lock);
2228         }
2229         closure_sync(&cl);
2230         bch2_btree_cache_cannibalize_unlock(c);
2231         return ret;
2232 }
2233
2234 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
2235                                         struct btree *b, struct bkey_i *new_key,
2236                                         bool skip_triggers)
2237 {
2238         struct btree_iter iter;
2239         int ret;
2240
2241         bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p,
2242                                   BTREE_MAX_DEPTH, b->c.level,
2243                                   BTREE_ITER_INTENT);
2244         ret = bch2_btree_iter_traverse(&iter);
2245         if (ret)
2246                 goto out;
2247
2248         /* has node been freed? */
2249         if (iter.path->l[b->c.level].b != b) {
2250                 /* node has been freed: */
2251                 BUG_ON(!btree_node_dying(b));
2252                 goto out;
2253         }
2254
2255         BUG_ON(!btree_node_hashed(b));
2256
2257         ret = bch2_btree_node_update_key(trans, &iter, b, new_key, skip_triggers);
2258 out:
2259         bch2_trans_iter_exit(trans, &iter);
2260         return ret;
2261 }
2262
2263 /* Init code: */
2264
2265 /*
2266  * Only for filesystem bringup, when first reading the btree roots or allocating
2267  * btree roots when initializing a new filesystem:
2268  */
2269 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2270 {
2271         BUG_ON(btree_node_root(c, b));
2272
2273         bch2_btree_set_root_inmem(c, b);
2274 }
2275
2276 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2277 {
2278         struct closure cl;
2279         struct btree *b;
2280         int ret;
2281
2282         closure_init_stack(&cl);
2283
2284         do {
2285                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2286                 closure_sync(&cl);
2287         } while (ret);
2288
2289         b = bch2_btree_node_mem_alloc(c, false);
2290         bch2_btree_cache_cannibalize_unlock(c);
2291
2292         set_btree_node_fake(b);
2293         set_btree_node_need_rewrite(b);
2294         b->c.level      = 0;
2295         b->c.btree_id   = id;
2296
2297         bkey_btree_ptr_init(&b->key);
2298         b->key.k.p = SPOS_MAX;
2299         *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2300
2301         bch2_bset_init_first(b, &b->data->keys);
2302         bch2_btree_build_aux_trees(b);
2303
2304         b->data->flags = 0;
2305         btree_set_min(b, POS_MIN);
2306         btree_set_max(b, SPOS_MAX);
2307         b->data->format = bch2_btree_calc_format(b);
2308         btree_node_set_format(b, b->data->format);
2309
2310         ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2311                                           b->c.level, b->c.btree_id);
2312         BUG_ON(ret);
2313
2314         bch2_btree_set_root_inmem(c, b);
2315
2316         six_unlock_write(&b->c.lock);
2317         six_unlock_intent(&b->c.lock);
2318 }
2319
2320 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2321 {
2322         struct btree_update *as;
2323
2324         mutex_lock(&c->btree_interior_update_lock);
2325         list_for_each_entry(as, &c->btree_interior_update_list, list)
2326                 prt_printf(out, "%p m %u w %u r %u j %llu\n",
2327                        as,
2328                        as->mode,
2329                        as->nodes_written,
2330                        atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2331                        as->journal.seq);
2332         mutex_unlock(&c->btree_interior_update_lock);
2333 }
2334
2335 static bool bch2_btree_interior_updates_pending(struct bch_fs *c)
2336 {
2337         bool ret;
2338
2339         mutex_lock(&c->btree_interior_update_lock);
2340         ret = !list_empty(&c->btree_interior_update_list);
2341         mutex_unlock(&c->btree_interior_update_lock);
2342
2343         return ret;
2344 }
2345
2346 bool bch2_btree_interior_updates_flush(struct bch_fs *c)
2347 {
2348         bool ret = bch2_btree_interior_updates_pending(c);
2349
2350         if (ret)
2351                 closure_wait_event(&c->btree_interior_update_wait,
2352                                    !bch2_btree_interior_updates_pending(c));
2353         return ret;
2354 }
2355
2356 void bch2_journal_entries_to_btree_roots(struct bch_fs *c, struct jset *jset)
2357 {
2358         struct btree_root *r;
2359         struct jset_entry *entry;
2360
2361         mutex_lock(&c->btree_root_lock);
2362
2363         vstruct_for_each(jset, entry)
2364                 if (entry->type == BCH_JSET_ENTRY_btree_root) {
2365                         r = &c->btree_roots[entry->btree_id];
2366                         r->level = entry->level;
2367                         r->alive = true;
2368                         bkey_copy(&r->key, &entry->start[0]);
2369                 }
2370
2371         mutex_unlock(&c->btree_root_lock);
2372 }
2373
2374 struct jset_entry *
2375 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2376                                     struct jset_entry *start,
2377                                     struct jset_entry *end)
2378 {
2379         struct jset_entry *entry;
2380         unsigned long have = 0;
2381         unsigned i;
2382
2383         for (entry = start; entry < end; entry = vstruct_next(entry))
2384                 if (entry->type == BCH_JSET_ENTRY_btree_root)
2385                         __set_bit(entry->btree_id, &have);
2386
2387         mutex_lock(&c->btree_root_lock);
2388
2389         for (i = 0; i < BTREE_ID_NR; i++)
2390                 if (c->btree_roots[i].alive && !test_bit(i, &have)) {
2391                         journal_entry_set(end,
2392                                           BCH_JSET_ENTRY_btree_root,
2393                                           i, c->btree_roots[i].level,
2394                                           &c->btree_roots[i].key,
2395                                           c->btree_roots[i].key.u64s);
2396                         end = vstruct_next(end);
2397                 }
2398
2399         mutex_unlock(&c->btree_root_lock);
2400
2401         return end;
2402 }
2403
2404 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2405 {
2406         if (c->btree_interior_update_worker)
2407                 destroy_workqueue(c->btree_interior_update_worker);
2408         mempool_exit(&c->btree_interior_update_pool);
2409 }
2410
2411 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2412 {
2413         mutex_init(&c->btree_reserve_cache_lock);
2414         INIT_LIST_HEAD(&c->btree_interior_update_list);
2415         INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2416         mutex_init(&c->btree_interior_update_lock);
2417         INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2418
2419         c->btree_interior_update_worker =
2420                 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2421         if (!c->btree_interior_update_worker)
2422                 return -ENOMEM;
2423
2424         return mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2425                                          sizeof(struct btree_update));
2426 }