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_journal_iter.h"
   9 #include "btree_update.h"
  10 #include "btree_update_interior.h"
  11 #include "btree_io.h"
  12 #include "btree_iter.h"
  13 #include "btree_locking.h"
  14 #include "buckets.h"
  15 #include "clock.h"
  16 #include "error.h"
  17 #include "extents.h"
  18 #include "journal.h"
  19 #include "journal_reclaim.h"
  20 #include "keylist.h"
  21 #include "replicas.h"
  22 #include "super-io.h"
  23 #include "trace.h"
  24
  25 #include <linux/random.h>
  26
  27 static int bch2_btree_insert_node(struct btree_update *, struct btree_trans *,
  28                                   btree_path_idx_t, struct btree *,
  29                                   struct keylist *, unsigned);
  30 static void bch2_btree_update_add_new_node(struct btree_update *, struct btree *);
  31
  32 static btree_path_idx_t get_unlocked_mut_path(struct btree_trans *trans,
  33                                               enum btree_id btree_id,
  34                                               unsigned level,
  35                                               struct bpos pos)
  36 {
  37         btree_path_idx_t path_idx = bch2_path_get(trans, btree_id, pos, level + 1, level,
  38                              BTREE_ITER_NOPRESERVE|
  39                              BTREE_ITER_INTENT, _RET_IP_);
  40         path_idx = bch2_btree_path_make_mut(trans, path_idx, true, _RET_IP_);
  41
  42         struct btree_path *path = trans->paths + path_idx;
  43         bch2_btree_path_downgrade(trans, path);
  44         __bch2_btree_path_unlock(trans, path);
  45         return path_idx;
  46 }
  47
  48 /* Debug code: */
  49
  50 /*
  51  * Verify that child nodes correctly span parent node's range:
  52  */
  53 static void btree_node_interior_verify(struct bch_fs *c, struct btree *b)
  54 {
  55 #ifdef CONFIG_BCACHEFS_DEBUG
  56         struct bpos next_node = b->data->min_key;
  57         struct btree_node_iter iter;
  58         struct bkey_s_c k;
  59         struct bkey_s_c_btree_ptr_v2 bp;
  60         struct bkey unpacked;
  61         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
  62
  63         BUG_ON(!b->c.level);
  64
  65         if (!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))
  66                 return;
  67
  68         bch2_btree_node_iter_init_from_start(&iter, b);
  69
  70         while (1) {
  71                 k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked);
  72                 if (k.k->type != KEY_TYPE_btree_ptr_v2)
  73                         break;
  74                 bp = bkey_s_c_to_btree_ptr_v2(k);
  75
  76                 if (!bpos_eq(next_node, bp.v->min_key)) {
  77                         bch2_dump_btree_node(c, b);
  78                         bch2_bpos_to_text(&buf1, next_node);
  79                         bch2_bpos_to_text(&buf2, bp.v->min_key);
  80                         panic("expected next min_key %s got %s\n", buf1.buf, buf2.buf);
  81                 }
  82
  83                 bch2_btree_node_iter_advance(&iter, b);
  84
  85                 if (bch2_btree_node_iter_end(&iter)) {
  86                         if (!bpos_eq(k.k->p, b->key.k.p)) {
  87                                 bch2_dump_btree_node(c, b);
  88                                 bch2_bpos_to_text(&buf1, b->key.k.p);
  89                                 bch2_bpos_to_text(&buf2, k.k->p);
  90                                 panic("expected end %s got %s\n", buf1.buf, buf2.buf);
  91                         }
  92                         break;
  93                 }
  94
  95                 next_node = bpos_successor(k.k->p);
  96         }
  97 #endif
  98 }
  99
 100 /* Calculate ideal packed bkey format for new btree nodes: */
 101
 102 static void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b)
 103 {
 104         struct bkey_packed *k;
 105         struct bset_tree *t;
 106         struct bkey uk;
 107
 108         for_each_bset(b, t)
 109                 bset_tree_for_each_key(b, t, k)
 110                         if (!bkey_deleted(k)) {
 111                                 uk = bkey_unpack_key(b, k);
 112                                 bch2_bkey_format_add_key(s, &uk);
 113                         }
 114 }
 115
 116 static struct bkey_format bch2_btree_calc_format(struct btree *b)
 117 {
 118         struct bkey_format_state s;
 119
 120         bch2_bkey_format_init(&s);
 121         bch2_bkey_format_add_pos(&s, b->data->min_key);
 122         bch2_bkey_format_add_pos(&s, b->data->max_key);
 123         __bch2_btree_calc_format(&s, b);
 124
 125         return bch2_bkey_format_done(&s);
 126 }
 127
 128 static size_t btree_node_u64s_with_format(struct btree_nr_keys nr,
 129                                           struct bkey_format *old_f,
 130                                           struct bkey_format *new_f)
 131 {
 132         /* stupid integer promotion rules */
 133         ssize_t delta =
 134             (((int) new_f->key_u64s - old_f->key_u64s) *
 135              (int) nr.packed_keys) +
 136             (((int) new_f->key_u64s - BKEY_U64s) *
 137              (int) nr.unpacked_keys);
 138
 139         BUG_ON(delta + nr.live_u64s < 0);
 140
 141         return nr.live_u64s + delta;
 142 }
 143
 144 /**
 145  * bch2_btree_node_format_fits - check if we could rewrite node with a new format
 146  *
 147  * @c:          filesystem handle
 148  * @b:          btree node to rewrite
 149  * @nr:         number of keys for new node (i.e. b->nr)
 150  * @new_f:      bkey format to translate keys to
 151  *
 152  * Returns: true if all re-packed keys will be able to fit in a new node.
 153  *
 154  * Assumes all keys will successfully pack with the new format.
 155  */
 156 static bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
 157                                  struct btree_nr_keys nr,
 158                                  struct bkey_format *new_f)
 159 {
 160         size_t u64s = btree_node_u64s_with_format(nr, &b->format, new_f);
 161
 162         return __vstruct_bytes(struct btree_node, u64s) < btree_bytes(c);
 163 }
 164
 165 /* Btree node freeing/allocation: */
 166
 167 static void __btree_node_free(struct btree_trans *trans, struct btree *b)
 168 {
 169         struct bch_fs *c = trans->c;
 170
 171         trace_and_count(c, btree_node_free, trans, b);
 172
 173         BUG_ON(btree_node_write_blocked(b));
 174         BUG_ON(btree_node_dirty(b));
 175         BUG_ON(btree_node_need_write(b));
 176         BUG_ON(b == btree_node_root(c, b));
 177         BUG_ON(b->ob.nr);
 178         BUG_ON(!list_empty(&b->write_blocked));
 179         BUG_ON(b->will_make_reachable);
 180
 181         clear_btree_node_noevict(b);
 182
 183         mutex_lock(&c->btree_cache.lock);
 184         list_move(&b->list, &c->btree_cache.freeable);
 185         mutex_unlock(&c->btree_cache.lock);
 186 }
 187
 188 static void bch2_btree_node_free_inmem(struct btree_trans *trans,
 189                                        struct btree_path *path,
 190                                        struct btree *b)
 191 {
 192         struct bch_fs *c = trans->c;
 193         unsigned i, level = b->c.level;
 194
 195         bch2_btree_node_lock_write_nofail(trans, path, &b->c);
 196         bch2_btree_node_hash_remove(&c->btree_cache, b);
 197         __btree_node_free(trans, b);
 198         six_unlock_write(&b->c.lock);
 199         mark_btree_node_locked_noreset(path, level, BTREE_NODE_INTENT_LOCKED);
 200
 201         trans_for_each_path(trans, path, i)
 202                 if (path->l[level].b == b) {
 203                         btree_node_unlock(trans, path, level);
 204                         path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
 205                 }
 206 }
 207
 208 static void bch2_btree_node_free_never_used(struct btree_update *as,
 209                                             struct btree_trans *trans,
 210                                             struct btree *b)
 211 {
 212         struct bch_fs *c = as->c;
 213         struct prealloc_nodes *p = &as->prealloc_nodes[b->c.lock.readers != NULL];
 214         struct btree_path *path;
 215         unsigned i, level = b->c.level;
 216
 217         BUG_ON(!list_empty(&b->write_blocked));
 218         BUG_ON(b->will_make_reachable != (1UL|(unsigned long) as));
 219
 220         b->will_make_reachable = 0;
 221         closure_put(&as->cl);
 222
 223         clear_btree_node_will_make_reachable(b);
 224         clear_btree_node_accessed(b);
 225         clear_btree_node_dirty_acct(c, b);
 226         clear_btree_node_need_write(b);
 227
 228         mutex_lock(&c->btree_cache.lock);
 229         list_del_init(&b->list);
 230         bch2_btree_node_hash_remove(&c->btree_cache, b);
 231         mutex_unlock(&c->btree_cache.lock);
 232
 233         BUG_ON(p->nr >= ARRAY_SIZE(p->b));
 234         p->b[p->nr++] = b;
 235
 236         six_unlock_intent(&b->c.lock);
 237
 238         trans_for_each_path(trans, path, i)
 239                 if (path->l[level].b == b) {
 240                         btree_node_unlock(trans, path, level);
 241                         path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
 242                 }
 243 }
 244
 245 static struct btree *__bch2_btree_node_alloc(struct btree_trans *trans,
 246                                              struct disk_reservation *res,
 247                                              struct closure *cl,
 248                                              bool interior_node,
 249                                              unsigned flags)
 250 {
 251         struct bch_fs *c = trans->c;
 252         struct write_point *wp;
 253         struct btree *b;
 254         BKEY_PADDED_ONSTACK(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
 255         struct open_buckets obs = { .nr = 0 };
 256         struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
 257         enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
 258         unsigned nr_reserve = watermark > BCH_WATERMARK_reclaim
 259                 ? BTREE_NODE_RESERVE
 260                 : 0;
 261         int ret;
 262
 263         mutex_lock(&c->btree_reserve_cache_lock);
 264         if (c->btree_reserve_cache_nr > nr_reserve) {
 265                 struct btree_alloc *a =
 266                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
 267
 268                 obs = a->ob;
 269                 bkey_copy(&tmp.k, &a->k);
 270                 mutex_unlock(&c->btree_reserve_cache_lock);
 271                 goto mem_alloc;
 272         }
 273         mutex_unlock(&c->btree_reserve_cache_lock);
 274
 275 retry:
 276         ret = bch2_alloc_sectors_start_trans(trans,
 277                                       c->opts.metadata_target ?:
 278                                       c->opts.foreground_target,
 279                                       0,
 280                                       writepoint_ptr(&c->btree_write_point),
 281                                       &devs_have,
 282                                       res->nr_replicas,
 283                                       c->opts.metadata_replicas_required,
 284                                       watermark, 0, cl, &wp);
 285         if (unlikely(ret))
 286                 return ERR_PTR(ret);
 287
 288         if (wp->sectors_free < btree_sectors(c)) {
 289                 struct open_bucket *ob;
 290                 unsigned i;
 291
 292                 open_bucket_for_each(c, &wp->ptrs, ob, i)
 293                         if (ob->sectors_free < btree_sectors(c))
 294                                 ob->sectors_free = 0;
 295
 296                 bch2_alloc_sectors_done(c, wp);
 297                 goto retry;
 298         }
 299
 300         bkey_btree_ptr_v2_init(&tmp.k);
 301         bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, btree_sectors(c), false);
 302
 303         bch2_open_bucket_get(c, wp, &obs);
 304         bch2_alloc_sectors_done(c, wp);
 305 mem_alloc:
 306         b = bch2_btree_node_mem_alloc(trans, interior_node);
 307         six_unlock_write(&b->c.lock);
 308         six_unlock_intent(&b->c.lock);
 309
 310         /* we hold cannibalize_lock: */
 311         BUG_ON(IS_ERR(b));
 312         BUG_ON(b->ob.nr);
 313
 314         bkey_copy(&b->key, &tmp.k);
 315         b->ob = obs;
 316
 317         return b;
 318 }
 319
 320 static struct btree *bch2_btree_node_alloc(struct btree_update *as,
 321                                            struct btree_trans *trans,
 322                                            unsigned level)
 323 {
 324         struct bch_fs *c = as->c;
 325         struct btree *b;
 326         struct prealloc_nodes *p = &as->prealloc_nodes[!!level];
 327         int ret;
 328
 329         BUG_ON(level >= BTREE_MAX_DEPTH);
 330         BUG_ON(!p->nr);
 331
 332         b = p->b[--p->nr];
 333
 334         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
 335         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
 336
 337         set_btree_node_accessed(b);
 338         set_btree_node_dirty_acct(c, b);
 339         set_btree_node_need_write(b);
 340
 341         bch2_bset_init_first(b, &b->data->keys);
 342         b->c.level      = level;
 343         b->c.btree_id   = as->btree_id;
 344         b->version_ondisk = c->sb.version;
 345
 346         memset(&b->nr, 0, sizeof(b->nr));
 347         b->data->magic = cpu_to_le64(bset_magic(c));
 348         memset(&b->data->_ptr, 0, sizeof(b->data->_ptr));
 349         b->data->flags = 0;
 350         SET_BTREE_NODE_ID(b->data, as->btree_id);
 351         SET_BTREE_NODE_LEVEL(b->data, level);
 352
 353         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 354                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
 355
 356                 bp->v.mem_ptr           = 0;
 357                 bp->v.seq               = b->data->keys.seq;
 358                 bp->v.sectors_written   = 0;
 359         }
 360
 361         SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
 362
 363         bch2_btree_build_aux_trees(b);
 364
 365         ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
 366         BUG_ON(ret);
 367
 368         trace_and_count(c, btree_node_alloc, trans, b);
 369         bch2_increment_clock(c, btree_sectors(c), WRITE);
 370         return b;
 371 }
 372
 373 static void btree_set_min(struct btree *b, struct bpos pos)
 374 {
 375         if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
 376                 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
 377         b->data->min_key = pos;
 378 }
 379
 380 static void btree_set_max(struct btree *b, struct bpos pos)
 381 {
 382         b->key.k.p = pos;
 383         b->data->max_key = pos;
 384 }
 385
 386 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
 387                                                        struct btree_trans *trans,
 388                                                        struct btree *b)
 389 {
 390         struct btree *n = bch2_btree_node_alloc(as, trans, b->c.level);
 391         struct bkey_format format = bch2_btree_calc_format(b);
 392
 393         /*
 394          * The keys might expand with the new format - if they wouldn't fit in
 395          * the btree node anymore, use the old format for now:
 396          */
 397         if (!bch2_btree_node_format_fits(as->c, b, b->nr, &format))
 398                 format = b->format;
 399
 400         SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
 401
 402         btree_set_min(n, b->data->min_key);
 403         btree_set_max(n, b->data->max_key);
 404
 405         n->data->format         = format;
 406         btree_node_set_format(n, format);
 407
 408         bch2_btree_sort_into(as->c, n, b);
 409
 410         btree_node_reset_sib_u64s(n);
 411         return n;
 412 }
 413
 414 static struct btree *__btree_root_alloc(struct btree_update *as,
 415                                 struct btree_trans *trans, unsigned level)
 416 {
 417         struct btree *b = bch2_btree_node_alloc(as, trans, level);
 418
 419         btree_set_min(b, POS_MIN);
 420         btree_set_max(b, SPOS_MAX);
 421         b->data->format = bch2_btree_calc_format(b);
 422
 423         btree_node_set_format(b, b->data->format);
 424         bch2_btree_build_aux_trees(b);
 425
 426         return b;
 427 }
 428
 429 static void bch2_btree_reserve_put(struct btree_update *as, struct btree_trans *trans)
 430 {
 431         struct bch_fs *c = as->c;
 432         struct prealloc_nodes *p;
 433
 434         for (p = as->prealloc_nodes;
 435              p < as->prealloc_nodes + ARRAY_SIZE(as->prealloc_nodes);
 436              p++) {
 437                 while (p->nr) {
 438                         struct btree *b = p->b[--p->nr];
 439
 440                         mutex_lock(&c->btree_reserve_cache_lock);
 441
 442                         if (c->btree_reserve_cache_nr <
 443                             ARRAY_SIZE(c->btree_reserve_cache)) {
 444                                 struct btree_alloc *a =
 445                                         &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
 446
 447                                 a->ob = b->ob;
 448                                 b->ob.nr = 0;
 449                                 bkey_copy(&a->k, &b->key);
 450                         } else {
 451                                 bch2_open_buckets_put(c, &b->ob);
 452                         }
 453
 454                         mutex_unlock(&c->btree_reserve_cache_lock);
 455
 456                         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
 457                         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
 458                         __btree_node_free(trans, b);
 459                         six_unlock_write(&b->c.lock);
 460                         six_unlock_intent(&b->c.lock);
 461                 }
 462         }
 463 }
 464
 465 static int bch2_btree_reserve_get(struct btree_trans *trans,
 466                                   struct btree_update *as,
 467                                   unsigned nr_nodes[2],
 468                                   unsigned flags,
 469                                   struct closure *cl)
 470 {
 471         struct btree *b;
 472         unsigned interior;
 473         int ret = 0;
 474
 475         BUG_ON(nr_nodes[0] + nr_nodes[1] > BTREE_RESERVE_MAX);
 476
 477         /*
 478          * Protects reaping from the btree node cache and using the btree node
 479          * open bucket reserve:
 480          */
 481         ret = bch2_btree_cache_cannibalize_lock(trans, 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, cl,
 490                                                     interior, flags);
 491                         if (IS_ERR(b)) {
 492                                 ret = PTR_ERR(b);
 493                                 goto err;
 494                         }
 495
 496                         p->b[p->nr++] = b;
 497                 }
 498         }
 499 err:
 500         bch2_btree_cache_cannibalize_unlock(trans);
 501         return ret;
 502 }
 503
 504 /* Asynchronous interior node update machinery */
 505
 506 static void bch2_btree_update_free(struct btree_update *as, struct btree_trans *trans)
 507 {
 508         struct bch_fs *c = as->c;
 509
 510         if (as->took_gc_lock)
 511                 up_read(&c->gc_lock);
 512         as->took_gc_lock = false;
 513
 514         bch2_journal_pin_drop(&c->journal, &as->journal);
 515         bch2_journal_pin_flush(&c->journal, &as->journal);
 516         bch2_disk_reservation_put(c, &as->disk_res);
 517         bch2_btree_reserve_put(as, trans);
 518
 519         bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_total],
 520                                as->start_time);
 521
 522         mutex_lock(&c->btree_interior_update_lock);
 523         list_del(&as->unwritten_list);
 524         list_del(&as->list);
 525
 526         closure_debug_destroy(&as->cl);
 527         mempool_free(as, &c->btree_interior_update_pool);
 528
 529         /*
 530          * Have to do the wakeup with btree_interior_update_lock still held,
 531          * since being on btree_interior_update_list is our ref on @c:
 532          */
 533         closure_wake_up(&c->btree_interior_update_wait);
 534
 535         mutex_unlock(&c->btree_interior_update_lock);
 536 }
 537
 538 static void btree_update_add_key(struct btree_update *as,
 539                                  struct keylist *keys, struct btree *b)
 540 {
 541         struct bkey_i *k = &b->key;
 542
 543         BUG_ON(bch2_keylist_u64s(keys) + k->k.u64s >
 544                ARRAY_SIZE(as->_old_keys));
 545
 546         bkey_copy(keys->top, k);
 547         bkey_i_to_btree_ptr_v2(keys->top)->v.mem_ptr = b->c.level + 1;
 548
 549         bch2_keylist_push(keys);
 550 }
 551
 552 /*
 553  * The transactional part of an interior btree node update, where we journal the
 554  * update we did to the interior node and update alloc info:
 555  */
 556 static int btree_update_nodes_written_trans(struct btree_trans *trans,
 557                                             struct btree_update *as)
 558 {
 559         struct jset_entry *e = bch2_trans_jset_entry_alloc(trans, as->journal_u64s);
 560         int ret = PTR_ERR_OR_ZERO(e);
 561         if (ret)
 562                 return ret;
 563
 564         memcpy(e, as->journal_entries, as->journal_u64s * sizeof(u64));
 565
 566         trans->journal_pin = &as->journal;
 567
 568         for_each_keylist_key(&as->old_keys, k) {
 569                 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
 570
 571                 ret = bch2_trans_mark_old(trans, as->btree_id, level, bkey_i_to_s_c(k), 0);
 572                 if (ret)
 573                         return ret;
 574         }
 575
 576         for_each_keylist_key(&as->new_keys, k) {
 577                 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
 578
 579                 ret = bch2_trans_mark_new(trans, as->btree_id, level, k, 0);
 580                 if (ret)
 581                         return ret;
 582         }
 583
 584         return 0;
 585 }
 586
 587 static void btree_update_nodes_written(struct btree_update *as)
 588 {
 589         struct bch_fs *c = as->c;
 590         struct btree *b;
 591         struct btree_trans *trans = bch2_trans_get(c);
 592         u64 journal_seq = 0;
 593         unsigned i;
 594         int ret;
 595
 596         /*
 597          * If we're already in an error state, it might be because a btree node
 598          * was never written, and we might be trying to free that same btree
 599          * node here, but it won't have been marked as allocated and we'll see
 600          * spurious disk usage inconsistencies in the transactional part below
 601          * if we don't skip it:
 602          */
 603         ret = bch2_journal_error(&c->journal);
 604         if (ret)
 605                 goto err;
 606
 607         /*
 608          * Wait for any in flight writes to finish before we free the old nodes
 609          * on disk:
 610          */
 611         for (i = 0; i < as->nr_old_nodes; i++) {
 612                 __le64 seq;
 613
 614                 b = as->old_nodes[i];
 615
 616                 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
 617                 seq = b->data ? b->data->keys.seq : 0;
 618                 six_unlock_read(&b->c.lock);
 619
 620                 if (seq == as->old_nodes_seq[i])
 621                         wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight_inner,
 622                                        TASK_UNINTERRUPTIBLE);
 623         }
 624
 625         /*
 626          * We did an update to a parent node where the pointers we added pointed
 627          * to child nodes that weren't written yet: now, the child nodes have
 628          * been written so we can write out the update to the interior node.
 629          */
 630
 631         /*
 632          * We can't call into journal reclaim here: we'd block on the journal
 633          * reclaim lock, but we may need to release the open buckets we have
 634          * pinned in order for other btree updates to make forward progress, and
 635          * journal reclaim does btree updates when flushing bkey_cached entries,
 636          * which may require allocations as well.
 637          */
 638         ret = commit_do(trans, &as->disk_res, &journal_seq,
 639                         BCH_WATERMARK_reclaim|
 640                         BCH_TRANS_COMMIT_no_enospc|
 641                         BCH_TRANS_COMMIT_no_check_rw|
 642                         BCH_TRANS_COMMIT_journal_reclaim,
 643                         btree_update_nodes_written_trans(trans, as));
 644         bch2_trans_unlock(trans);
 645
 646         bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
 647                              "%s(): error %s", __func__, bch2_err_str(ret));
 648 err:
 649         if (as->b) {
 650
 651                 b = as->b;
 652                 btree_path_idx_t path_idx = get_unlocked_mut_path(trans,
 653                                                 as->btree_id, b->c.level, b->key.k.p);
 654                 struct btree_path *path = trans->paths + path_idx;
 655                 /*
 656                  * @b is the node we did the final insert into:
 657                  *
 658                  * On failure to get a journal reservation, we still have to
 659                  * unblock the write and allow most of the write path to happen
 660                  * so that shutdown works, but the i->journal_seq mechanism
 661                  * won't work to prevent the btree write from being visible (we
 662                  * didn't get a journal sequence number) - instead
 663                  * __bch2_btree_node_write() doesn't do the actual write if
 664                  * we're in journal error state:
 665                  */
 666
 667                 /*
 668                  * Ensure transaction is unlocked before using
 669                  * btree_node_lock_nopath() (the use of which is always suspect,
 670                  * we need to work on removing this in the future)
 671                  *
 672                  * It should be, but get_unlocked_mut_path() -> bch2_path_get()
 673                  * calls bch2_path_upgrade(), before we call path_make_mut(), so
 674                  * we may rarely end up with a locked path besides the one we
 675                  * have here:
 676                  */
 677                 bch2_trans_unlock(trans);
 678                 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
 679                 mark_btree_node_locked(trans, path, b->c.level, BTREE_NODE_INTENT_LOCKED);
 680                 path->l[b->c.level].lock_seq = six_lock_seq(&b->c.lock);
 681                 path->l[b->c.level].b = b;
 682
 683                 bch2_btree_node_lock_write_nofail(trans, path, &b->c);
 684
 685                 mutex_lock(&c->btree_interior_update_lock);
 686
 687                 list_del(&as->write_blocked_list);
 688                 if (list_empty(&b->write_blocked))
 689                         clear_btree_node_write_blocked(b);
 690
 691                 /*
 692                  * Node might have been freed, recheck under
 693                  * btree_interior_update_lock:
 694                  */
 695                 if (as->b == b) {
 696                         BUG_ON(!b->c.level);
 697                         BUG_ON(!btree_node_dirty(b));
 698
 699                         if (!ret) {
 700                                 struct bset *last = btree_bset_last(b);
 701
 702                                 last->journal_seq = cpu_to_le64(
 703                                                              max(journal_seq,
 704                                                                  le64_to_cpu(last->journal_seq)));
 705
 706                                 bch2_btree_add_journal_pin(c, b, journal_seq);
 707                         } else {
 708                                 /*
 709                                  * If we didn't get a journal sequence number we
 710                                  * can't write this btree node, because recovery
 711                                  * won't know to ignore this write:
 712                                  */
 713                                 set_btree_node_never_write(b);
 714                         }
 715                 }
 716
 717                 mutex_unlock(&c->btree_interior_update_lock);
 718
 719                 mark_btree_node_locked_noreset(path, b->c.level, BTREE_NODE_INTENT_LOCKED);
 720                 six_unlock_write(&b->c.lock);
 721
 722                 btree_node_write_if_need(c, b, SIX_LOCK_intent);
 723                 btree_node_unlock(trans, path, b->c.level);
 724                 bch2_path_put(trans, path_idx, true);
 725         }
 726
 727         bch2_journal_pin_drop(&c->journal, &as->journal);
 728
 729         mutex_lock(&c->btree_interior_update_lock);
 730         for (i = 0; i < as->nr_new_nodes; i++) {
 731                 b = as->new_nodes[i];
 732
 733                 BUG_ON(b->will_make_reachable != (unsigned long) as);
 734                 b->will_make_reachable = 0;
 735                 clear_btree_node_will_make_reachable(b);
 736         }
 737         mutex_unlock(&c->btree_interior_update_lock);
 738
 739         for (i = 0; i < as->nr_new_nodes; i++) {
 740                 b = as->new_nodes[i];
 741
 742                 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
 743                 btree_node_write_if_need(c, b, SIX_LOCK_read);
 744                 six_unlock_read(&b->c.lock);
 745         }
 746
 747         for (i = 0; i < as->nr_open_buckets; i++)
 748                 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
 749
 750         bch2_btree_update_free(as, trans);
 751         bch2_trans_put(trans);
 752 }
 753
 754 static void btree_interior_update_work(struct work_struct *work)
 755 {
 756         struct bch_fs *c =
 757                 container_of(work, struct bch_fs, btree_interior_update_work);
 758         struct btree_update *as;
 759
 760         while (1) {
 761                 mutex_lock(&c->btree_interior_update_lock);
 762                 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
 763                                               struct btree_update, unwritten_list);
 764                 if (as && !as->nodes_written)
 765                         as = NULL;
 766                 mutex_unlock(&c->btree_interior_update_lock);
 767
 768                 if (!as)
 769                         break;
 770
 771                 btree_update_nodes_written(as);
 772         }
 773 }
 774
 775 static CLOSURE_CALLBACK(btree_update_set_nodes_written)
 776 {
 777         closure_type(as, struct btree_update, cl);
 778         struct bch_fs *c = as->c;
 779
 780         mutex_lock(&c->btree_interior_update_lock);
 781         as->nodes_written = true;
 782         mutex_unlock(&c->btree_interior_update_lock);
 783
 784         queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
 785 }
 786
 787 /*
 788  * We're updating @b with pointers to nodes that haven't finished writing yet:
 789  * block @b from being written until @as completes
 790  */
 791 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
 792 {
 793         struct bch_fs *c = as->c;
 794
 795         mutex_lock(&c->btree_interior_update_lock);
 796         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 797
 798         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 799         BUG_ON(!btree_node_dirty(b));
 800         BUG_ON(!b->c.level);
 801
 802         as->mode        = BTREE_INTERIOR_UPDATING_NODE;
 803         as->b           = b;
 804
 805         set_btree_node_write_blocked(b);
 806         list_add(&as->write_blocked_list, &b->write_blocked);
 807
 808         mutex_unlock(&c->btree_interior_update_lock);
 809 }
 810
 811 static int bch2_update_reparent_journal_pin_flush(struct journal *j,
 812                                 struct journal_entry_pin *_pin, u64 seq)
 813 {
 814         return 0;
 815 }
 816
 817 static void btree_update_reparent(struct btree_update *as,
 818                                   struct btree_update *child)
 819 {
 820         struct bch_fs *c = as->c;
 821
 822         lockdep_assert_held(&c->btree_interior_update_lock);
 823
 824         child->b = NULL;
 825         child->mode = BTREE_INTERIOR_UPDATING_AS;
 826
 827         bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal,
 828                               bch2_update_reparent_journal_pin_flush);
 829 }
 830
 831 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
 832 {
 833         struct bkey_i *insert = &b->key;
 834         struct bch_fs *c = as->c;
 835
 836         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 837
 838         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
 839                ARRAY_SIZE(as->journal_entries));
 840
 841         as->journal_u64s +=
 842                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
 843                                   BCH_JSET_ENTRY_btree_root,
 844                                   b->c.btree_id, b->c.level,
 845                                   insert, insert->k.u64s);
 846
 847         mutex_lock(&c->btree_interior_update_lock);
 848         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 849
 850         as->mode        = BTREE_INTERIOR_UPDATING_ROOT;
 851         mutex_unlock(&c->btree_interior_update_lock);
 852 }
 853
 854 /*
 855  * bch2_btree_update_add_new_node:
 856  *
 857  * This causes @as to wait on @b to be written, before it gets to
 858  * bch2_btree_update_nodes_written
 859  *
 860  * Additionally, it sets b->will_make_reachable to prevent any additional writes
 861  * to @b from happening besides the first until @b is reachable on disk
 862  *
 863  * And it adds @b to the list of @as's new nodes, so that we can update sector
 864  * counts in bch2_btree_update_nodes_written:
 865  */
 866 static void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
 867 {
 868         struct bch_fs *c = as->c;
 869
 870         closure_get(&as->cl);
 871
 872         mutex_lock(&c->btree_interior_update_lock);
 873         BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
 874         BUG_ON(b->will_make_reachable);
 875
 876         as->new_nodes[as->nr_new_nodes++] = b;
 877         b->will_make_reachable = 1UL|(unsigned long) as;
 878         set_btree_node_will_make_reachable(b);
 879
 880         mutex_unlock(&c->btree_interior_update_lock);
 881
 882         btree_update_add_key(as, &as->new_keys, b);
 883
 884         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 885                 unsigned bytes = vstruct_end(&b->data->keys) - (void *) b->data;
 886                 unsigned sectors = round_up(bytes, block_bytes(c)) >> 9;
 887
 888                 bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written =
 889                         cpu_to_le16(sectors);
 890         }
 891 }
 892
 893 /*
 894  * returns true if @b was a new node
 895  */
 896 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
 897 {
 898         struct btree_update *as;
 899         unsigned long v;
 900         unsigned i;
 901
 902         mutex_lock(&c->btree_interior_update_lock);
 903         /*
 904          * When b->will_make_reachable != 0, it owns a ref on as->cl that's
 905          * dropped when it gets written by bch2_btree_complete_write - the
 906          * xchg() is for synchronization with bch2_btree_complete_write:
 907          */
 908         v = xchg(&b->will_make_reachable, 0);
 909         clear_btree_node_will_make_reachable(b);
 910         as = (struct btree_update *) (v & ~1UL);
 911
 912         if (!as) {
 913                 mutex_unlock(&c->btree_interior_update_lock);
 914                 return;
 915         }
 916
 917         for (i = 0; i < as->nr_new_nodes; i++)
 918                 if (as->new_nodes[i] == b)
 919                         goto found;
 920
 921         BUG();
 922 found:
 923         array_remove_item(as->new_nodes, as->nr_new_nodes, i);
 924         mutex_unlock(&c->btree_interior_update_lock);
 925
 926         if (v & 1)
 927                 closure_put(&as->cl);
 928 }
 929
 930 static void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
 931 {
 932         while (b->ob.nr)
 933                 as->open_buckets[as->nr_open_buckets++] =
 934                         b->ob.v[--b->ob.nr];
 935 }
 936
 937 static int bch2_btree_update_will_free_node_journal_pin_flush(struct journal *j,
 938                                 struct journal_entry_pin *_pin, u64 seq)
 939 {
 940         return 0;
 941 }
 942
 943 /*
 944  * @b is being split/rewritten: it may have pointers to not-yet-written btree
 945  * nodes and thus outstanding btree_updates - redirect @b's
 946  * btree_updates to point to this btree_update:
 947  */
 948 static void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 949                                                       struct btree *b)
 950 {
 951         struct bch_fs *c = as->c;
 952         struct btree_update *p, *n;
 953         struct btree_write *w;
 954
 955         set_btree_node_dying(b);
 956
 957         if (btree_node_fake(b))
 958                 return;
 959
 960         mutex_lock(&c->btree_interior_update_lock);
 961
 962         /*
 963          * Does this node have any btree_update operations preventing
 964          * it from being written?
 965          *
 966          * If so, redirect them to point to this btree_update: we can
 967          * write out our new nodes, but we won't make them visible until those
 968          * operations complete
 969          */
 970         list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
 971                 list_del_init(&p->write_blocked_list);
 972                 btree_update_reparent(as, p);
 973
 974                 /*
 975                  * for flush_held_btree_writes() waiting on updates to flush or
 976                  * nodes to be writeable:
 977                  */
 978                 closure_wake_up(&c->btree_interior_update_wait);
 979         }
 980
 981         clear_btree_node_dirty_acct(c, b);
 982         clear_btree_node_need_write(b);
 983         clear_btree_node_write_blocked(b);
 984
 985         /*
 986          * Does this node have unwritten data that has a pin on the journal?
 987          *
 988          * If so, transfer that pin to the btree_update operation -
 989          * note that if we're freeing multiple nodes, we only need to keep the
 990          * oldest pin of any of the nodes we're freeing. We'll release the pin
 991          * when the new nodes are persistent and reachable on disk:
 992          */
 993         w = btree_current_write(b);
 994         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal,
 995                               bch2_btree_update_will_free_node_journal_pin_flush);
 996         bch2_journal_pin_drop(&c->journal, &w->journal);
 997
 998         w = btree_prev_write(b);
 999         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal,
1000                               bch2_btree_update_will_free_node_journal_pin_flush);
1001         bch2_journal_pin_drop(&c->journal, &w->journal);
1002
1003         mutex_unlock(&c->btree_interior_update_lock);
1004
1005         /*
1006          * Is this a node that isn't reachable on disk yet?
1007          *
1008          * Nodes that aren't reachable yet have writes blocked until they're
1009          * reachable - now that we've cancelled any pending writes and moved
1010          * things waiting on that write to wait on this update, we can drop this
1011          * node from the list of nodes that the other update is making
1012          * reachable, prior to freeing it:
1013          */
1014         btree_update_drop_new_node(c, b);
1015
1016         btree_update_add_key(as, &as->old_keys, b);
1017
1018         as->old_nodes[as->nr_old_nodes] = b;
1019         as->old_nodes_seq[as->nr_old_nodes] = b->data->keys.seq;
1020         as->nr_old_nodes++;
1021 }
1022
1023 static void bch2_btree_update_done(struct btree_update *as, struct btree_trans *trans)
1024 {
1025         struct bch_fs *c = as->c;
1026         u64 start_time = as->start_time;
1027
1028         BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
1029
1030         if (as->took_gc_lock)
1031                 up_read(&as->c->gc_lock);
1032         as->took_gc_lock = false;
1033
1034         bch2_btree_reserve_put(as, trans);
1035
1036         continue_at(&as->cl, btree_update_set_nodes_written,
1037                     as->c->btree_interior_update_worker);
1038
1039         bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_foreground],
1040                                start_time);
1041 }
1042
1043 static struct btree_update *
1044 bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
1045                         unsigned level, bool split, unsigned flags)
1046 {
1047         struct bch_fs *c = trans->c;
1048         struct btree_update *as;
1049         u64 start_time = local_clock();
1050         int disk_res_flags = (flags & BCH_TRANS_COMMIT_no_enospc)
1051                 ? BCH_DISK_RESERVATION_NOFAIL : 0;
1052         unsigned nr_nodes[2] = { 0, 0 };
1053         unsigned update_level = level;
1054         enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
1055         int ret = 0;
1056         u32 restart_count = trans->restart_count;
1057
1058         BUG_ON(!path->should_be_locked);
1059
1060         if (watermark == BCH_WATERMARK_copygc)
1061                 watermark = BCH_WATERMARK_btree_copygc;
1062         if (watermark < BCH_WATERMARK_btree)
1063                 watermark = BCH_WATERMARK_btree;
1064
1065         flags &= ~BCH_WATERMARK_MASK;
1066         flags |= watermark;
1067
1068         if (!(flags & BCH_TRANS_COMMIT_journal_reclaim) &&
1069             watermark < c->journal.watermark) {
1070                 struct journal_res res = { 0 };
1071
1072                 ret = drop_locks_do(trans,
1073                         bch2_journal_res_get(&c->journal, &res, 1,
1074                                              watermark|JOURNAL_RES_GET_CHECK));
1075                 if (ret)
1076                         return ERR_PTR(ret);
1077         }
1078
1079         while (1) {
1080                 nr_nodes[!!update_level] += 1 + split;
1081                 update_level++;
1082
1083                 ret = bch2_btree_path_upgrade(trans, path, update_level + 1);
1084                 if (ret)
1085                         return ERR_PTR(ret);
1086
1087                 if (!btree_path_node(path, update_level)) {
1088                         /* Allocating new root? */
1089                         nr_nodes[1] += split;
1090                         update_level = BTREE_MAX_DEPTH;
1091                         break;
1092                 }
1093
1094                 /*
1095                  * Always check for space for two keys, even if we won't have to
1096                  * split at prior level - it might have been a merge instead:
1097                  */
1098                 if (bch2_btree_node_insert_fits(c, path->l[update_level].b,
1099                                                 BKEY_BTREE_PTR_U64s_MAX * 2))
1100                         break;
1101
1102                 split = path->l[update_level].b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c);
1103         }
1104
1105         if (!down_read_trylock(&c->gc_lock)) {
1106                 ret = drop_locks_do(trans, (down_read(&c->gc_lock), 0));
1107                 if (ret) {
1108                         up_read(&c->gc_lock);
1109                         return ERR_PTR(ret);
1110                 }
1111         }
1112
1113         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOFS);
1114         memset(as, 0, sizeof(*as));
1115         closure_init(&as->cl, NULL);
1116         as->c           = c;
1117         as->start_time  = start_time;
1118         as->mode        = BTREE_INTERIOR_NO_UPDATE;
1119         as->took_gc_lock = true;
1120         as->btree_id    = path->btree_id;
1121         as->update_level = update_level;
1122         INIT_LIST_HEAD(&as->list);
1123         INIT_LIST_HEAD(&as->unwritten_list);
1124         INIT_LIST_HEAD(&as->write_blocked_list);
1125         bch2_keylist_init(&as->old_keys, as->_old_keys);
1126         bch2_keylist_init(&as->new_keys, as->_new_keys);
1127         bch2_keylist_init(&as->parent_keys, as->inline_keys);
1128
1129         mutex_lock(&c->btree_interior_update_lock);
1130         list_add_tail(&as->list, &c->btree_interior_update_list);
1131         mutex_unlock(&c->btree_interior_update_lock);
1132
1133         /*
1134          * We don't want to allocate if we're in an error state, that can cause
1135          * deadlock on emergency shutdown due to open buckets getting stuck in
1136          * the btree_reserve_cache after allocator shutdown has cleared it out.
1137          * This check needs to come after adding us to the btree_interior_update
1138          * list but before calling bch2_btree_reserve_get, to synchronize with
1139          * __bch2_fs_read_only().
1140          */
1141         ret = bch2_journal_error(&c->journal);
1142         if (ret)
1143                 goto err;
1144
1145         ret = bch2_disk_reservation_get(c, &as->disk_res,
1146                         (nr_nodes[0] + nr_nodes[1]) * btree_sectors(c),
1147                         c->opts.metadata_replicas,
1148                         disk_res_flags);
1149         if (ret)
1150                 goto err;
1151
1152         ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, NULL);
1153         if (bch2_err_matches(ret, ENOSPC) ||
1154             bch2_err_matches(ret, ENOMEM)) {
1155                 struct closure cl;
1156
1157                 /*
1158                  * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
1159                  * flag
1160                  */
1161                 if (bch2_err_matches(ret, ENOSPC) &&
1162                     (flags & BCH_TRANS_COMMIT_journal_reclaim) &&
1163                     watermark != BCH_WATERMARK_reclaim) {
1164                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
1165                         goto err;
1166                 }
1167
1168                 closure_init_stack(&cl);
1169
1170                 do {
1171                         ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, &cl);
1172
1173                         bch2_trans_unlock(trans);
1174                         closure_sync(&cl);
1175                 } while (bch2_err_matches(ret, BCH_ERR_operation_blocked));
1176         }
1177
1178         if (ret) {
1179                 trace_and_count(c, btree_reserve_get_fail, trans->fn,
1180                                 _RET_IP_, nr_nodes[0] + nr_nodes[1], ret);
1181                 goto err;
1182         }
1183
1184         ret = bch2_trans_relock(trans);
1185         if (ret)
1186                 goto err;
1187
1188         bch2_trans_verify_not_restarted(trans, restart_count);
1189         return as;
1190 err:
1191         bch2_btree_update_free(as, trans);
1192         if (!bch2_err_matches(ret, ENOSPC) &&
1193             !bch2_err_matches(ret, EROFS))
1194                 bch_err_fn_ratelimited(c, ret);
1195         return ERR_PTR(ret);
1196 }
1197
1198 /* Btree root updates: */
1199
1200 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1201 {
1202         /* Root nodes cannot be reaped */
1203         mutex_lock(&c->btree_cache.lock);
1204         list_del_init(&b->list);
1205         mutex_unlock(&c->btree_cache.lock);
1206
1207         mutex_lock(&c->btree_root_lock);
1208         BUG_ON(btree_node_root(c, b) &&
1209                (b->c.level < btree_node_root(c, b)->c.level ||
1210                 !btree_node_dying(btree_node_root(c, b))));
1211
1212         bch2_btree_id_root(c, b->c.btree_id)->b = b;
1213         mutex_unlock(&c->btree_root_lock);
1214
1215         bch2_recalc_btree_reserve(c);
1216 }
1217
1218 static void bch2_btree_set_root(struct btree_update *as,
1219                                 struct btree_trans *trans,
1220                                 struct btree_path *path,
1221                                 struct btree *b)
1222 {
1223         struct bch_fs *c = as->c;
1224         struct btree *old;
1225
1226         trace_and_count(c, btree_node_set_root, trans, b);
1227
1228         old = btree_node_root(c, b);
1229
1230         /*
1231          * Ensure no one is using the old root while we switch to the
1232          * new root:
1233          */
1234         bch2_btree_node_lock_write_nofail(trans, path, &old->c);
1235
1236         bch2_btree_set_root_inmem(c, b);
1237
1238         btree_update_updated_root(as, b);
1239
1240         /*
1241          * Unlock old root after new root is visible:
1242          *
1243          * The new root isn't persistent, but that's ok: we still have
1244          * an intent lock on the new root, and any updates that would
1245          * depend on the new root would have to update the new root.
1246          */
1247         bch2_btree_node_unlock_write(trans, path, old);
1248 }
1249
1250 /* Interior node updates: */
1251
1252 static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
1253                                         struct btree_trans *trans,
1254                                         struct btree_path *path,
1255                                         struct btree *b,
1256                                         struct btree_node_iter *node_iter,
1257                                         struct bkey_i *insert)
1258 {
1259         struct bch_fs *c = as->c;
1260         struct bkey_packed *k;
1261         struct printbuf buf = PRINTBUF;
1262         unsigned long old, new, v;
1263
1264         BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 &&
1265                !btree_ptr_sectors_written(insert));
1266
1267         if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
1268                 bch2_journal_key_overwritten(c, b->c.btree_id, b->c.level, insert->k.p);
1269
1270         if (bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1271                               btree_node_type(b), WRITE, &buf) ?:
1272             bch2_bkey_in_btree_node(c, b, bkey_i_to_s_c(insert), &buf)) {
1273                 printbuf_reset(&buf);
1274                 prt_printf(&buf, "inserting invalid bkey\n  ");
1275                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
1276                 prt_printf(&buf, "\n  ");
1277                 bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1278                                   btree_node_type(b), WRITE, &buf);
1279                 bch2_bkey_in_btree_node(c, b, bkey_i_to_s_c(insert), &buf);
1280
1281                 bch2_fs_inconsistent(c, "%s", buf.buf);
1282                 dump_stack();
1283         }
1284
1285         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1286                ARRAY_SIZE(as->journal_entries));
1287
1288         as->journal_u64s +=
1289                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1290                                   BCH_JSET_ENTRY_btree_keys,
1291                                   b->c.btree_id, b->c.level,
1292                                   insert, insert->k.u64s);
1293
1294         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1295                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1296                 bch2_btree_node_iter_advance(node_iter, b);
1297
1298         bch2_btree_bset_insert_key(trans, path, b, node_iter, insert);
1299         set_btree_node_dirty_acct(c, b);
1300
1301         v = READ_ONCE(b->flags);
1302         do {
1303                 old = new = v;
1304
1305                 new &= ~BTREE_WRITE_TYPE_MASK;
1306                 new |= BTREE_WRITE_interior;
1307                 new |= 1 << BTREE_NODE_need_write;
1308         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1309
1310         printbuf_exit(&buf);
1311 }
1312
1313 static void
1314 __bch2_btree_insert_keys_interior(struct btree_update *as,
1315                                   struct btree_trans *trans,
1316                                   struct btree_path *path,
1317                                   struct btree *b,
1318                                   struct btree_node_iter node_iter,
1319                                   struct keylist *keys)
1320 {
1321         struct bkey_i *insert = bch2_keylist_front(keys);
1322         struct bkey_packed *k;
1323
1324         BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1325
1326         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1327                (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1328                 ;
1329
1330         while (!bch2_keylist_empty(keys)) {
1331                 insert = bch2_keylist_front(keys);
1332
1333                 if (bpos_gt(insert->k.p, b->key.k.p))
1334                         break;
1335
1336                 bch2_insert_fixup_btree_ptr(as, trans, path, b, &node_iter, insert);
1337                 bch2_keylist_pop_front(keys);
1338         }
1339 }
1340
1341 /*
1342  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1343  * node)
1344  */
1345 static void __btree_split_node(struct btree_update *as,
1346                                struct btree_trans *trans,
1347                                struct btree *b,
1348                                struct btree *n[2])
1349 {
1350         struct bkey_packed *k;
1351         struct bpos n1_pos = POS_MIN;
1352         struct btree_node_iter iter;
1353         struct bset *bsets[2];
1354         struct bkey_format_state format[2];
1355         struct bkey_packed *out[2];
1356         struct bkey uk;
1357         unsigned u64s, n1_u64s = (b->nr.live_u64s * 3) / 5;
1358         struct { unsigned nr_keys, val_u64s; } nr_keys[2];
1359         int i;
1360
1361         memset(&nr_keys, 0, sizeof(nr_keys));
1362
1363         for (i = 0; i < 2; i++) {
1364                 BUG_ON(n[i]->nsets != 1);
1365
1366                 bsets[i] = btree_bset_first(n[i]);
1367                 out[i] = bsets[i]->start;
1368
1369                 SET_BTREE_NODE_SEQ(n[i]->data, BTREE_NODE_SEQ(b->data) + 1);
1370                 bch2_bkey_format_init(&format[i]);
1371         }
1372
1373         u64s = 0;
1374         for_each_btree_node_key(b, k, &iter) {
1375                 if (bkey_deleted(k))
1376                         continue;
1377
1378                 i = u64s >= n1_u64s;
1379                 u64s += k->u64s;
1380                 uk = bkey_unpack_key(b, k);
1381                 if (!i)
1382                         n1_pos = uk.p;
1383                 bch2_bkey_format_add_key(&format[i], &uk);
1384
1385                 nr_keys[i].nr_keys++;
1386                 nr_keys[i].val_u64s += bkeyp_val_u64s(&b->format, k);
1387         }
1388
1389         btree_set_min(n[0], b->data->min_key);
1390         btree_set_max(n[0], n1_pos);
1391         btree_set_min(n[1], bpos_successor(n1_pos));
1392         btree_set_max(n[1], b->data->max_key);
1393
1394         for (i = 0; i < 2; i++) {
1395                 bch2_bkey_format_add_pos(&format[i], n[i]->data->min_key);
1396                 bch2_bkey_format_add_pos(&format[i], n[i]->data->max_key);
1397
1398                 n[i]->data->format = bch2_bkey_format_done(&format[i]);
1399
1400                 unsigned u64s = nr_keys[i].nr_keys * n[i]->data->format.key_u64s +
1401                         nr_keys[i].val_u64s;
1402                 if (__vstruct_bytes(struct btree_node, u64s) > btree_bytes(as->c))
1403                         n[i]->data->format = b->format;
1404
1405                 btree_node_set_format(n[i], n[i]->data->format);
1406         }
1407
1408         u64s = 0;
1409         for_each_btree_node_key(b, k, &iter) {
1410                 if (bkey_deleted(k))
1411                         continue;
1412
1413                 i = u64s >= n1_u64s;
1414                 u64s += k->u64s;
1415
1416                 if (bch2_bkey_transform(&n[i]->format, out[i], bkey_packed(k)
1417                                         ? &b->format: &bch2_bkey_format_current, k))
1418                         out[i]->format = KEY_FORMAT_LOCAL_BTREE;
1419                 else
1420                         bch2_bkey_unpack(b, (void *) out[i], k);
1421
1422                 out[i]->needs_whiteout = false;
1423
1424                 btree_keys_account_key_add(&n[i]->nr, 0, out[i]);
1425                 out[i] = bkey_p_next(out[i]);
1426         }
1427
1428         for (i = 0; i < 2; i++) {
1429                 bsets[i]->u64s = cpu_to_le16((u64 *) out[i] - bsets[i]->_data);
1430
1431                 BUG_ON(!bsets[i]->u64s);
1432
1433                 set_btree_bset_end(n[i], n[i]->set);
1434
1435                 btree_node_reset_sib_u64s(n[i]);
1436
1437                 bch2_verify_btree_nr_keys(n[i]);
1438
1439                 if (b->c.level)
1440                         btree_node_interior_verify(as->c, n[i]);
1441         }
1442 }
1443
1444 /*
1445  * For updates to interior nodes, we've got to do the insert before we split
1446  * because the stuff we're inserting has to be inserted atomically. Post split,
1447  * the keys might have to go in different nodes and the split would no longer be
1448  * atomic.
1449  *
1450  * Worse, if the insert is from btree node coalescing, if we do the insert after
1451  * we do the split (and pick the pivot) - the pivot we pick might be between
1452  * nodes that were coalesced, and thus in the middle of a child node post
1453  * coalescing:
1454  */
1455 static void btree_split_insert_keys(struct btree_update *as,
1456                                     struct btree_trans *trans,
1457                                     btree_path_idx_t path_idx,
1458                                     struct btree *b,
1459                                     struct keylist *keys)
1460 {
1461         struct btree_path *path = trans->paths + path_idx;
1462
1463         if (!bch2_keylist_empty(keys) &&
1464             bpos_le(bch2_keylist_front(keys)->k.p, b->data->max_key)) {
1465                 struct btree_node_iter node_iter;
1466
1467                 bch2_btree_node_iter_init(&node_iter, b, &bch2_keylist_front(keys)->k.p);
1468
1469                 __bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
1470
1471                 btree_node_interior_verify(as->c, b);
1472         }
1473 }
1474
1475 static int btree_split(struct btree_update *as, struct btree_trans *trans,
1476                        btree_path_idx_t path, struct btree *b,
1477                        struct keylist *keys, unsigned flags)
1478 {
1479         struct bch_fs *c = as->c;
1480         struct btree *parent = btree_node_parent(trans->paths + path, b);
1481         struct btree *n1, *n2 = NULL, *n3 = NULL;
1482         btree_path_idx_t path1 = 0, path2 = 0;
1483         u64 start_time = local_clock();
1484         int ret = 0;
1485
1486         BUG_ON(!parent && (b != btree_node_root(c, b)));
1487         BUG_ON(parent && !btree_node_intent_locked(trans->paths + path, b->c.level + 1));
1488
1489         bch2_btree_interior_update_will_free_node(as, b);
1490
1491         if (b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c)) {
1492                 struct btree *n[2];
1493
1494                 trace_and_count(c, btree_node_split, trans, b);
1495
1496                 n[0] = n1 = bch2_btree_node_alloc(as, trans, b->c.level);
1497                 n[1] = n2 = bch2_btree_node_alloc(as, trans, b->c.level);
1498
1499                 __btree_split_node(as, trans, b, n);
1500
1501                 if (keys) {
1502                         btree_split_insert_keys(as, trans, path, n1, keys);
1503                         btree_split_insert_keys(as, trans, path, n2, keys);
1504                         BUG_ON(!bch2_keylist_empty(keys));
1505                 }
1506
1507                 bch2_btree_build_aux_trees(n2);
1508                 bch2_btree_build_aux_trees(n1);
1509
1510                 bch2_btree_update_add_new_node(as, n1);
1511                 bch2_btree_update_add_new_node(as, n2);
1512                 six_unlock_write(&n2->c.lock);
1513                 six_unlock_write(&n1->c.lock);
1514
1515                 path1 = get_unlocked_mut_path(trans, as->btree_id, n1->c.level, n1->key.k.p);
1516                 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1517                 mark_btree_node_locked(trans, trans->paths + path1, n1->c.level, BTREE_NODE_INTENT_LOCKED);
1518                 bch2_btree_path_level_init(trans, trans->paths + path1, n1);
1519
1520                 path2 = get_unlocked_mut_path(trans, as->btree_id, n2->c.level, n2->key.k.p);
1521                 six_lock_increment(&n2->c.lock, SIX_LOCK_intent);
1522                 mark_btree_node_locked(trans, trans->paths + path2, n2->c.level, BTREE_NODE_INTENT_LOCKED);
1523                 bch2_btree_path_level_init(trans, trans->paths + path2, n2);
1524
1525                 /*
1526                  * Note that on recursive parent_keys == keys, so we
1527                  * can't start adding new keys to parent_keys before emptying it
1528                  * out (which we did with btree_split_insert_keys() above)
1529                  */
1530                 bch2_keylist_add(&as->parent_keys, &n1->key);
1531                 bch2_keylist_add(&as->parent_keys, &n2->key);
1532
1533                 if (!parent) {
1534                         /* Depth increases, make a new root */
1535                         n3 = __btree_root_alloc(as, trans, b->c.level + 1);
1536
1537                         bch2_btree_update_add_new_node(as, n3);
1538                         six_unlock_write(&n3->c.lock);
1539
1540                         trans->paths[path2].locks_want++;
1541                         BUG_ON(btree_node_locked(trans->paths + path2, n3->c.level));
1542                         six_lock_increment(&n3->c.lock, SIX_LOCK_intent);
1543                         mark_btree_node_locked(trans, trans->paths + path2, n3->c.level, BTREE_NODE_INTENT_LOCKED);
1544                         bch2_btree_path_level_init(trans, trans->paths + path2, n3);
1545
1546                         n3->sib_u64s[0] = U16_MAX;
1547                         n3->sib_u64s[1] = U16_MAX;
1548
1549                         btree_split_insert_keys(as, trans, path, n3, &as->parent_keys);
1550                 }
1551         } else {
1552                 trace_and_count(c, btree_node_compact, trans, b);
1553
1554                 n1 = bch2_btree_node_alloc_replacement(as, trans, b);
1555
1556                 if (keys) {
1557                         btree_split_insert_keys(as, trans, path, n1, keys);
1558                         BUG_ON(!bch2_keylist_empty(keys));
1559                 }
1560
1561                 bch2_btree_build_aux_trees(n1);
1562                 bch2_btree_update_add_new_node(as, n1);
1563                 six_unlock_write(&n1->c.lock);
1564
1565                 path1 = get_unlocked_mut_path(trans, as->btree_id, n1->c.level, n1->key.k.p);
1566                 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1567                 mark_btree_node_locked(trans, trans->paths + path1, n1->c.level, BTREE_NODE_INTENT_LOCKED);
1568                 bch2_btree_path_level_init(trans, trans->paths + path1, n1);
1569
1570                 if (parent)
1571                         bch2_keylist_add(&as->parent_keys, &n1->key);
1572         }
1573
1574         /* New nodes all written, now make them visible: */
1575
1576         if (parent) {
1577                 /* Split a non root node */
1578                 ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1579                 if (ret)
1580                         goto err;
1581         } else if (n3) {
1582                 bch2_btree_set_root(as, trans, trans->paths + path, n3);
1583         } else {
1584                 /* Root filled up but didn't need to be split */
1585                 bch2_btree_set_root(as, trans, trans->paths + path, n1);
1586         }
1587
1588         if (n3) {
1589                 bch2_btree_update_get_open_buckets(as, n3);
1590                 bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0);
1591         }
1592         if (n2) {
1593                 bch2_btree_update_get_open_buckets(as, n2);
1594                 bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0);
1595         }
1596         bch2_btree_update_get_open_buckets(as, n1);
1597         bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1598
1599         /*
1600          * The old node must be freed (in memory) _before_ unlocking the new
1601          * nodes - else another thread could re-acquire a read lock on the old
1602          * node after another thread has locked and updated the new node, thus
1603          * seeing stale data:
1604          */
1605         bch2_btree_node_free_inmem(trans, trans->paths + path, b);
1606
1607         if (n3)
1608                 bch2_trans_node_add(trans, trans->paths + path, n3);
1609         if (n2)
1610                 bch2_trans_node_add(trans, trans->paths + path2, n2);
1611         bch2_trans_node_add(trans, trans->paths + path1, n1);
1612
1613         if (n3)
1614                 six_unlock_intent(&n3->c.lock);
1615         if (n2)
1616                 six_unlock_intent(&n2->c.lock);
1617         six_unlock_intent(&n1->c.lock);
1618 out:
1619         if (path2) {
1620                 __bch2_btree_path_unlock(trans, trans->paths + path2);
1621                 bch2_path_put(trans, path2, true);
1622         }
1623         if (path1) {
1624                 __bch2_btree_path_unlock(trans, trans->paths + path1);
1625                 bch2_path_put(trans, path1, true);
1626         }
1627
1628         bch2_trans_verify_locks(trans);
1629
1630         bch2_time_stats_update(&c->times[n2
1631                                ? BCH_TIME_btree_node_split
1632                                : BCH_TIME_btree_node_compact],
1633                                start_time);
1634         return ret;
1635 err:
1636         if (n3)
1637                 bch2_btree_node_free_never_used(as, trans, n3);
1638         if (n2)
1639                 bch2_btree_node_free_never_used(as, trans, n2);
1640         bch2_btree_node_free_never_used(as, trans, n1);
1641         goto out;
1642 }
1643
1644 static void
1645 bch2_btree_insert_keys_interior(struct btree_update *as,
1646                                 struct btree_trans *trans,
1647                                 struct btree_path *path,
1648                                 struct btree *b,
1649                                 struct keylist *keys)
1650 {
1651         struct btree_path *linked;
1652         unsigned i;
1653
1654         __bch2_btree_insert_keys_interior(as, trans, path, b,
1655                                           path->l[b->c.level].iter, keys);
1656
1657         btree_update_updated_node(as, b);
1658
1659         trans_for_each_path_with_node(trans, b, linked, i)
1660                 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1661
1662         bch2_trans_verify_paths(trans);
1663 }
1664
1665 /**
1666  * bch2_btree_insert_node - insert bkeys into a given btree node
1667  *
1668  * @as:                 btree_update object
1669  * @trans:              btree_trans object
1670  * @path_idx:           path that points to current node
1671  * @b:                  node to insert keys into
1672  * @keys:               list of keys to insert
1673  * @flags:              transaction commit flags
1674  *
1675  * Returns: 0 on success, typically transaction restart error on failure
1676  *
1677  * Inserts as many keys as it can into a given btree node, splitting it if full.
1678  * If a split occurred, this function will return early. This can only happen
1679  * for leaf nodes -- inserts into interior nodes have to be atomic.
1680  */
1681 static int bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans,
1682                                   btree_path_idx_t path_idx, struct btree *b,
1683                                   struct keylist *keys, unsigned flags)
1684 {
1685         struct bch_fs *c = as->c;
1686         struct btree_path *path = trans->paths + path_idx;
1687         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1688         int old_live_u64s = b->nr.live_u64s;
1689         int live_u64s_added, u64s_added;
1690         int ret;
1691
1692         lockdep_assert_held(&c->gc_lock);
1693         BUG_ON(!btree_node_intent_locked(path, b->c.level));
1694         BUG_ON(!b->c.level);
1695         BUG_ON(!as || as->b);
1696         bch2_verify_keylist_sorted(keys);
1697
1698         ret = bch2_btree_node_lock_write(trans, path, &b->c);
1699         if (ret)
1700                 return ret;
1701
1702         bch2_btree_node_prep_for_write(trans, path, b);
1703
1704         if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1705                 bch2_btree_node_unlock_write(trans, path, b);
1706                 goto split;
1707         }
1708
1709         btree_node_interior_verify(c, b);
1710
1711         bch2_btree_insert_keys_interior(as, trans, path, b, keys);
1712
1713         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1714         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1715
1716         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1717                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1718         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1719                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1720
1721         if (u64s_added > live_u64s_added &&
1722             bch2_maybe_compact_whiteouts(c, b))
1723                 bch2_trans_node_reinit_iter(trans, b);
1724
1725         bch2_btree_node_unlock_write(trans, path, b);
1726
1727         btree_node_interior_verify(c, b);
1728         return 0;
1729 split:
1730         /*
1731          * We could attempt to avoid the transaction restart, by calling
1732          * bch2_btree_path_upgrade() and allocating more nodes:
1733          */
1734         if (b->c.level >= as->update_level) {
1735                 trace_and_count(c, trans_restart_split_race, trans, _THIS_IP_, b);
1736                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_split_race);
1737         }
1738
1739         return btree_split(as, trans, path_idx, b, keys, flags);
1740 }
1741
1742 int bch2_btree_split_leaf(struct btree_trans *trans,
1743                           btree_path_idx_t path,
1744                           unsigned flags)
1745 {
1746         /* btree_split & merge may both cause paths array to be reallocated */
1747
1748         struct btree *b = path_l(trans->paths + path)->b;
1749         struct btree_update *as;
1750         unsigned l;
1751         int ret = 0;
1752
1753         as = bch2_btree_update_start(trans, trans->paths + path,
1754                                      trans->paths[path].level,
1755                                      true, flags);
1756         if (IS_ERR(as))
1757                 return PTR_ERR(as);
1758
1759         ret = btree_split(as, trans, path, b, NULL, flags);
1760         if (ret) {
1761                 bch2_btree_update_free(as, trans);
1762                 return ret;
1763         }
1764
1765         bch2_btree_update_done(as, trans);
1766
1767         for (l = trans->paths[path].level + 1;
1768              btree_node_intent_locked(&trans->paths[path], l) && !ret;
1769              l++)
1770                 ret = bch2_foreground_maybe_merge(trans, path, l, flags);
1771
1772         return ret;
1773 }
1774
1775 int __bch2_foreground_maybe_merge(struct btree_trans *trans,
1776                                   btree_path_idx_t path,
1777                                   unsigned level,
1778                                   unsigned flags,
1779                                   enum btree_node_sibling sib)
1780 {
1781         struct bch_fs *c = trans->c;
1782         struct btree_update *as;
1783         struct bkey_format_state new_s;
1784         struct bkey_format new_f;
1785         struct bkey_i delete;
1786         struct btree *b, *m, *n, *prev, *next, *parent;
1787         struct bpos sib_pos;
1788         size_t sib_u64s;
1789         enum btree_id btree = trans->paths[path].btree_id;
1790         btree_path_idx_t sib_path = 0, new_path = 0;
1791         u64 start_time = local_clock();
1792         int ret = 0;
1793
1794         BUG_ON(!trans->paths[path].should_be_locked);
1795         BUG_ON(!btree_node_locked(&trans->paths[path], level));
1796
1797         b = trans->paths[path].l[level].b;
1798
1799         if ((sib == btree_prev_sib && bpos_eq(b->data->min_key, POS_MIN)) ||
1800             (sib == btree_next_sib && bpos_eq(b->data->max_key, SPOS_MAX))) {
1801                 b->sib_u64s[sib] = U16_MAX;
1802                 return 0;
1803         }
1804
1805         sib_pos = sib == btree_prev_sib
1806                 ? bpos_predecessor(b->data->min_key)
1807                 : bpos_successor(b->data->max_key);
1808
1809         sib_path = bch2_path_get(trans, btree, sib_pos,
1810                                  U8_MAX, level, BTREE_ITER_INTENT, _THIS_IP_);
1811         ret = bch2_btree_path_traverse(trans, sib_path, false);
1812         if (ret)
1813                 goto err;
1814
1815         btree_path_set_should_be_locked(trans->paths + sib_path);
1816
1817         m = trans->paths[sib_path].l[level].b;
1818
1819         if (btree_node_parent(trans->paths + path, b) !=
1820             btree_node_parent(trans->paths + sib_path, m)) {
1821                 b->sib_u64s[sib] = U16_MAX;
1822                 goto out;
1823         }
1824
1825         if (sib == btree_prev_sib) {
1826                 prev = m;
1827                 next = b;
1828         } else {
1829                 prev = b;
1830                 next = m;
1831         }
1832
1833         if (!bpos_eq(bpos_successor(prev->data->max_key), next->data->min_key)) {
1834                 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1835
1836                 bch2_bpos_to_text(&buf1, prev->data->max_key);
1837                 bch2_bpos_to_text(&buf2, next->data->min_key);
1838                 bch_err(c,
1839                         "%s(): btree topology error:\n"
1840                         "  prev ends at   %s\n"
1841                         "  next starts at %s",
1842                         __func__, buf1.buf, buf2.buf);
1843                 printbuf_exit(&buf1);
1844                 printbuf_exit(&buf2);
1845                 bch2_topology_error(c);
1846                 ret = -EIO;
1847                 goto err;
1848         }
1849
1850         bch2_bkey_format_init(&new_s);
1851         bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
1852         __bch2_btree_calc_format(&new_s, prev);
1853         __bch2_btree_calc_format(&new_s, next);
1854         bch2_bkey_format_add_pos(&new_s, next->data->max_key);
1855         new_f = bch2_bkey_format_done(&new_s);
1856
1857         sib_u64s = btree_node_u64s_with_format(b->nr, &b->format, &new_f) +
1858                 btree_node_u64s_with_format(m->nr, &m->format, &new_f);
1859
1860         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1861                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1862                 sib_u64s /= 2;
1863                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1864         }
1865
1866         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1867         sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
1868         b->sib_u64s[sib] = sib_u64s;
1869
1870         if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
1871                 goto out;
1872
1873         parent = btree_node_parent(trans->paths + path, b);
1874         as = bch2_btree_update_start(trans, trans->paths + path, level, false,
1875                                      BCH_TRANS_COMMIT_no_enospc|flags);
1876         ret = PTR_ERR_OR_ZERO(as);
1877         if (ret)
1878                 goto err;
1879
1880         trace_and_count(c, btree_node_merge, trans, b);
1881
1882         bch2_btree_interior_update_will_free_node(as, b);
1883         bch2_btree_interior_update_will_free_node(as, m);
1884
1885         n = bch2_btree_node_alloc(as, trans, b->c.level);
1886
1887         SET_BTREE_NODE_SEQ(n->data,
1888                            max(BTREE_NODE_SEQ(b->data),
1889                                BTREE_NODE_SEQ(m->data)) + 1);
1890
1891         btree_set_min(n, prev->data->min_key);
1892         btree_set_max(n, next->data->max_key);
1893
1894         n->data->format  = new_f;
1895         btree_node_set_format(n, new_f);
1896
1897         bch2_btree_sort_into(c, n, prev);
1898         bch2_btree_sort_into(c, n, next);
1899
1900         bch2_btree_build_aux_trees(n);
1901         bch2_btree_update_add_new_node(as, n);
1902         six_unlock_write(&n->c.lock);
1903
1904         new_path = get_unlocked_mut_path(trans, btree, n->c.level, n->key.k.p);
1905         six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1906         mark_btree_node_locked(trans, trans->paths + new_path, n->c.level, BTREE_NODE_INTENT_LOCKED);
1907         bch2_btree_path_level_init(trans, trans->paths + new_path, n);
1908
1909         bkey_init(&delete.k);
1910         delete.k.p = prev->key.k.p;
1911         bch2_keylist_add(&as->parent_keys, &delete);
1912         bch2_keylist_add(&as->parent_keys, &n->key);
1913
1914         bch2_trans_verify_paths(trans);
1915
1916         ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1917         if (ret)
1918                 goto err_free_update;
1919
1920         bch2_trans_verify_paths(trans);
1921
1922         bch2_btree_update_get_open_buckets(as, n);
1923         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1924
1925         bch2_btree_node_free_inmem(trans, trans->paths + path, b);
1926         bch2_btree_node_free_inmem(trans, trans->paths + sib_path, m);
1927
1928         bch2_trans_node_add(trans, trans->paths + path, n);
1929
1930         bch2_trans_verify_paths(trans);
1931
1932         six_unlock_intent(&n->c.lock);
1933
1934         bch2_btree_update_done(as, trans);
1935
1936         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time);
1937 out:
1938 err:
1939         if (new_path)
1940                 bch2_path_put(trans, new_path, true);
1941         bch2_path_put(trans, sib_path, true);
1942         bch2_trans_verify_locks(trans);
1943         return ret;
1944 err_free_update:
1945         bch2_btree_node_free_never_used(as, trans, n);
1946         bch2_btree_update_free(as, trans);
1947         goto out;
1948 }
1949
1950 int bch2_btree_node_rewrite(struct btree_trans *trans,
1951                             struct btree_iter *iter,
1952                             struct btree *b,
1953                             unsigned flags)
1954 {
1955         struct bch_fs *c = trans->c;
1956         struct btree *n, *parent;
1957         struct btree_update *as;
1958         btree_path_idx_t new_path = 0;
1959         int ret;
1960
1961         flags |= BCH_TRANS_COMMIT_no_enospc;
1962
1963         struct btree_path *path = btree_iter_path(trans, iter);
1964         parent = btree_node_parent(path, b);
1965         as = bch2_btree_update_start(trans, path, b->c.level, false, flags);
1966         ret = PTR_ERR_OR_ZERO(as);
1967         if (ret)
1968                 goto out;
1969
1970         bch2_btree_interior_update_will_free_node(as, b);
1971
1972         n = bch2_btree_node_alloc_replacement(as, trans, b);
1973
1974         bch2_btree_build_aux_trees(n);
1975         bch2_btree_update_add_new_node(as, n);
1976         six_unlock_write(&n->c.lock);
1977
1978         new_path = get_unlocked_mut_path(trans, iter->btree_id, n->c.level, n->key.k.p);
1979         six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1980         mark_btree_node_locked(trans, trans->paths + new_path, n->c.level, BTREE_NODE_INTENT_LOCKED);
1981         bch2_btree_path_level_init(trans, trans->paths + new_path, n);
1982
1983         trace_and_count(c, btree_node_rewrite, trans, b);
1984
1985         if (parent) {
1986                 bch2_keylist_add(&as->parent_keys, &n->key);
1987                 ret = bch2_btree_insert_node(as, trans, iter->path,
1988                                              parent, &as->parent_keys, flags);
1989                 if (ret)
1990                         goto err;
1991         } else {
1992                 bch2_btree_set_root(as, trans, btree_iter_path(trans, iter), n);
1993         }
1994
1995         bch2_btree_update_get_open_buckets(as, n);
1996         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1997
1998         bch2_btree_node_free_inmem(trans, btree_iter_path(trans, iter), b);
1999
2000         bch2_trans_node_add(trans, trans->paths + iter->path, n);
2001         six_unlock_intent(&n->c.lock);
2002
2003         bch2_btree_update_done(as, trans);
2004 out:
2005         if (new_path)
2006                 bch2_path_put(trans, new_path, true);
2007         bch2_trans_downgrade(trans);
2008         return ret;
2009 err:
2010         bch2_btree_node_free_never_used(as, trans, n);
2011         bch2_btree_update_free(as, trans);
2012         goto out;
2013 }
2014
2015 struct async_btree_rewrite {
2016         struct bch_fs           *c;
2017         struct work_struct      work;
2018         struct list_head        list;
2019         enum btree_id           btree_id;
2020         unsigned                level;
2021         struct bpos             pos;
2022         __le64                  seq;
2023 };
2024
2025 static int async_btree_node_rewrite_trans(struct btree_trans *trans,
2026                                           struct async_btree_rewrite *a)
2027 {
2028         struct bch_fs *c = trans->c;
2029         struct btree_iter iter;
2030         struct btree *b;
2031         int ret;
2032
2033         bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos,
2034                                   BTREE_MAX_DEPTH, a->level, 0);
2035         b = bch2_btree_iter_peek_node(&iter);
2036         ret = PTR_ERR_OR_ZERO(b);
2037         if (ret)
2038                 goto out;
2039
2040         if (!b || b->data->keys.seq != a->seq) {
2041                 struct printbuf buf = PRINTBUF;
2042
2043                 if (b)
2044                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
2045                 else
2046                         prt_str(&buf, "(null");
2047                 bch_info(c, "%s: node to rewrite not found:, searching for seq %llu, got\n%s",
2048                          __func__, a->seq, buf.buf);
2049                 printbuf_exit(&buf);
2050                 goto out;
2051         }
2052
2053         ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
2054 out:
2055         bch2_trans_iter_exit(trans, &iter);
2056
2057         return ret;
2058 }
2059
2060 static void async_btree_node_rewrite_work(struct work_struct *work)
2061 {
2062         struct async_btree_rewrite *a =
2063                 container_of(work, struct async_btree_rewrite, work);
2064         struct bch_fs *c = a->c;
2065         int ret;
2066
2067         ret = bch2_trans_do(c, NULL, NULL, 0,
2068                       async_btree_node_rewrite_trans(trans, a));
2069         bch_err_fn(c, ret);
2070         bch2_write_ref_put(c, BCH_WRITE_REF_node_rewrite);
2071         kfree(a);
2072 }
2073
2074 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
2075 {
2076         struct async_btree_rewrite *a;
2077         int ret;
2078
2079         a = kmalloc(sizeof(*a), GFP_NOFS);
2080         if (!a) {
2081                 bch_err(c, "%s: error allocating memory", __func__);
2082                 return;
2083         }
2084
2085         a->c            = c;
2086         a->btree_id     = b->c.btree_id;
2087         a->level        = b->c.level;
2088         a->pos          = b->key.k.p;
2089         a->seq          = b->data->keys.seq;
2090         INIT_WORK(&a->work, async_btree_node_rewrite_work);
2091
2092         if (unlikely(!test_bit(BCH_FS_may_go_rw, &c->flags))) {
2093                 mutex_lock(&c->pending_node_rewrites_lock);
2094                 list_add(&a->list, &c->pending_node_rewrites);
2095                 mutex_unlock(&c->pending_node_rewrites_lock);
2096                 return;
2097         }
2098
2099         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_node_rewrite)) {
2100                 if (test_bit(BCH_FS_started, &c->flags)) {
2101                         bch_err(c, "%s: error getting c->writes ref", __func__);
2102                         kfree(a);
2103                         return;
2104                 }
2105
2106                 ret = bch2_fs_read_write_early(c);
2107                 bch_err_msg(c, ret, "going read-write");
2108                 if (ret) {
2109                         kfree(a);
2110                         return;
2111                 }
2112
2113                 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2114         }
2115
2116         queue_work(c->btree_interior_update_worker, &a->work);
2117 }
2118
2119 void bch2_do_pending_node_rewrites(struct bch_fs *c)
2120 {
2121         struct async_btree_rewrite *a, *n;
2122
2123         mutex_lock(&c->pending_node_rewrites_lock);
2124         list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2125                 list_del(&a->list);
2126
2127                 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2128                 queue_work(c->btree_interior_update_worker, &a->work);
2129         }
2130         mutex_unlock(&c->pending_node_rewrites_lock);
2131 }
2132
2133 void bch2_free_pending_node_rewrites(struct bch_fs *c)
2134 {
2135         struct async_btree_rewrite *a, *n;
2136
2137         mutex_lock(&c->pending_node_rewrites_lock);
2138         list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2139                 list_del(&a->list);
2140
2141                 kfree(a);
2142         }
2143         mutex_unlock(&c->pending_node_rewrites_lock);
2144 }
2145
2146 static int __bch2_btree_node_update_key(struct btree_trans *trans,
2147                                         struct btree_iter *iter,
2148                                         struct btree *b, struct btree *new_hash,
2149                                         struct bkey_i *new_key,
2150                                         unsigned commit_flags,
2151                                         bool skip_triggers)
2152 {
2153         struct bch_fs *c = trans->c;
2154         struct btree_iter iter2 = { NULL };
2155         struct btree *parent;
2156         int ret;
2157
2158         if (!skip_triggers) {
2159                 ret = bch2_trans_mark_old(trans, b->c.btree_id, b->c.level + 1,
2160                                           bkey_i_to_s_c(&b->key), 0);
2161                 if (ret)
2162                         return ret;
2163
2164                 ret = bch2_trans_mark_new(trans, b->c.btree_id, b->c.level + 1,
2165                                           new_key, 0);
2166                 if (ret)
2167                         return ret;
2168         }
2169
2170         if (new_hash) {
2171                 bkey_copy(&new_hash->key, new_key);
2172                 ret = bch2_btree_node_hash_insert(&c->btree_cache,
2173                                 new_hash, b->c.level, b->c.btree_id);
2174                 BUG_ON(ret);
2175         }
2176
2177         parent = btree_node_parent(btree_iter_path(trans, iter), b);
2178         if (parent) {
2179                 bch2_trans_copy_iter(&iter2, iter);
2180
2181                 iter2.path = bch2_btree_path_make_mut(trans, iter2.path,
2182                                 iter2.flags & BTREE_ITER_INTENT,
2183                                 _THIS_IP_);
2184
2185                 struct btree_path *path2 = btree_iter_path(trans, &iter2);
2186                 BUG_ON(path2->level != b->c.level);
2187                 BUG_ON(!bpos_eq(path2->pos, new_key->k.p));
2188
2189                 btree_path_set_level_up(trans, path2);
2190
2191                 trans->paths_sorted = false;
2192
2193                 ret   = bch2_btree_iter_traverse(&iter2) ?:
2194                         bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_NORUN);
2195                 if (ret)
2196                         goto err;
2197         } else {
2198                 BUG_ON(btree_node_root(c, b) != b);
2199
2200                 struct jset_entry *e = bch2_trans_jset_entry_alloc(trans,
2201                                        jset_u64s(new_key->k.u64s));
2202                 ret = PTR_ERR_OR_ZERO(e);
2203                 if (ret)
2204                         return ret;
2205
2206                 journal_entry_set(e,
2207                                   BCH_JSET_ENTRY_btree_root,
2208                                   b->c.btree_id, b->c.level,
2209                                   new_key, new_key->k.u64s);
2210         }
2211
2212         ret = bch2_trans_commit(trans, NULL, NULL, commit_flags);
2213         if (ret)
2214                 goto err;
2215
2216         bch2_btree_node_lock_write_nofail(trans, btree_iter_path(trans, iter), &b->c);
2217
2218         if (new_hash) {
2219                 mutex_lock(&c->btree_cache.lock);
2220                 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
2221                 bch2_btree_node_hash_remove(&c->btree_cache, b);
2222
2223                 bkey_copy(&b->key, new_key);
2224                 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
2225                 BUG_ON(ret);
2226                 mutex_unlock(&c->btree_cache.lock);
2227         } else {
2228                 bkey_copy(&b->key, new_key);
2229         }
2230
2231         bch2_btree_node_unlock_write(trans, btree_iter_path(trans, iter), b);
2232 out:
2233         bch2_trans_iter_exit(trans, &iter2);
2234         return ret;
2235 err:
2236         if (new_hash) {
2237                 mutex_lock(&c->btree_cache.lock);
2238                 bch2_btree_node_hash_remove(&c->btree_cache, b);
2239                 mutex_unlock(&c->btree_cache.lock);
2240         }
2241         goto out;
2242 }
2243
2244 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
2245                                struct btree *b, struct bkey_i *new_key,
2246                                unsigned commit_flags, bool skip_triggers)
2247 {
2248         struct bch_fs *c = trans->c;
2249         struct btree *new_hash = NULL;
2250         struct btree_path *path = btree_iter_path(trans, iter);
2251         struct closure cl;
2252         int ret = 0;
2253
2254         ret = bch2_btree_path_upgrade(trans, path, b->c.level + 1);
2255         if (ret)
2256                 return ret;
2257
2258         closure_init_stack(&cl);
2259
2260         /*
2261          * check btree_ptr_hash_val() after @b is locked by
2262          * btree_iter_traverse():
2263          */
2264         if (btree_ptr_hash_val(new_key) != b->hash_val) {
2265                 ret = bch2_btree_cache_cannibalize_lock(trans, &cl);
2266                 if (ret) {
2267                         ret = drop_locks_do(trans, (closure_sync(&cl), 0));
2268                         if (ret)
2269                                 return ret;
2270                 }
2271
2272                 new_hash = bch2_btree_node_mem_alloc(trans, false);
2273         }
2274
2275         path->intent_ref++;
2276         ret = __bch2_btree_node_update_key(trans, iter, b, new_hash, new_key,
2277                                            commit_flags, skip_triggers);
2278         --path->intent_ref;
2279
2280         if (new_hash) {
2281                 mutex_lock(&c->btree_cache.lock);
2282                 list_move(&new_hash->list, &c->btree_cache.freeable);
2283                 mutex_unlock(&c->btree_cache.lock);
2284
2285                 six_unlock_write(&new_hash->c.lock);
2286                 six_unlock_intent(&new_hash->c.lock);
2287         }
2288         closure_sync(&cl);
2289         bch2_btree_cache_cannibalize_unlock(trans);
2290         return ret;
2291 }
2292
2293 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
2294                                         struct btree *b, struct bkey_i *new_key,
2295                                         unsigned commit_flags, bool skip_triggers)
2296 {
2297         struct btree_iter iter;
2298         int ret;
2299
2300         bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p,
2301                                   BTREE_MAX_DEPTH, b->c.level,
2302                                   BTREE_ITER_INTENT);
2303         ret = bch2_btree_iter_traverse(&iter);
2304         if (ret)
2305                 goto out;
2306
2307         /* has node been freed? */
2308         if (btree_iter_path(trans, &iter)->l[b->c.level].b != b) {
2309                 /* node has been freed: */
2310                 BUG_ON(!btree_node_dying(b));
2311                 goto out;
2312         }
2313
2314         BUG_ON(!btree_node_hashed(b));
2315
2316         struct bch_extent_ptr *ptr;
2317         bch2_bkey_drop_ptrs(bkey_i_to_s(new_key), ptr,
2318                             !bch2_bkey_has_device(bkey_i_to_s(&b->key), ptr->dev));
2319
2320         ret = bch2_btree_node_update_key(trans, &iter, b, new_key,
2321                                          commit_flags, skip_triggers);
2322 out:
2323         bch2_trans_iter_exit(trans, &iter);
2324         return ret;
2325 }
2326
2327 /* Init code: */
2328
2329 /*
2330  * Only for filesystem bringup, when first reading the btree roots or allocating
2331  * btree roots when initializing a new filesystem:
2332  */
2333 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2334 {
2335         BUG_ON(btree_node_root(c, b));
2336
2337         bch2_btree_set_root_inmem(c, b);
2338 }
2339
2340 static int __bch2_btree_root_alloc(struct btree_trans *trans, enum btree_id id)
2341 {
2342         struct bch_fs *c = trans->c;
2343         struct closure cl;
2344         struct btree *b;
2345         int ret;
2346
2347         closure_init_stack(&cl);
2348
2349         do {
2350                 ret = bch2_btree_cache_cannibalize_lock(trans, &cl);
2351                 closure_sync(&cl);
2352         } while (ret);
2353
2354         b = bch2_btree_node_mem_alloc(trans, false);
2355         bch2_btree_cache_cannibalize_unlock(trans);
2356
2357         set_btree_node_fake(b);
2358         set_btree_node_need_rewrite(b);
2359         b->c.level      = 0;
2360         b->c.btree_id   = id;
2361
2362         bkey_btree_ptr_init(&b->key);
2363         b->key.k.p = SPOS_MAX;
2364         *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2365
2366         bch2_bset_init_first(b, &b->data->keys);
2367         bch2_btree_build_aux_trees(b);
2368
2369         b->data->flags = 0;
2370         btree_set_min(b, POS_MIN);
2371         btree_set_max(b, SPOS_MAX);
2372         b->data->format = bch2_btree_calc_format(b);
2373         btree_node_set_format(b, b->data->format);
2374
2375         ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2376                                           b->c.level, b->c.btree_id);
2377         BUG_ON(ret);
2378
2379         bch2_btree_set_root_inmem(c, b);
2380
2381         six_unlock_write(&b->c.lock);
2382         six_unlock_intent(&b->c.lock);
2383         return 0;
2384 }
2385
2386 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2387 {
2388         bch2_trans_run(c, __bch2_btree_root_alloc(trans, id));
2389 }
2390
2391 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2392 {
2393         struct btree_update *as;
2394
2395         mutex_lock(&c->btree_interior_update_lock);
2396         list_for_each_entry(as, &c->btree_interior_update_list, list)
2397                 prt_printf(out, "%p m %u w %u r %u j %llu\n",
2398                        as,
2399                        as->mode,
2400                        as->nodes_written,
2401                        closure_nr_remaining(&as->cl),
2402                        as->journal.seq);
2403         mutex_unlock(&c->btree_interior_update_lock);
2404 }
2405
2406 static bool bch2_btree_interior_updates_pending(struct bch_fs *c)
2407 {
2408         bool ret;
2409
2410         mutex_lock(&c->btree_interior_update_lock);
2411         ret = !list_empty(&c->btree_interior_update_list);
2412         mutex_unlock(&c->btree_interior_update_lock);
2413
2414         return ret;
2415 }
2416
2417 bool bch2_btree_interior_updates_flush(struct bch_fs *c)
2418 {
2419         bool ret = bch2_btree_interior_updates_pending(c);
2420
2421         if (ret)
2422                 closure_wait_event(&c->btree_interior_update_wait,
2423                                    !bch2_btree_interior_updates_pending(c));
2424         return ret;
2425 }
2426
2427 void bch2_journal_entry_to_btree_root(struct bch_fs *c, struct jset_entry *entry)
2428 {
2429         struct btree_root *r = bch2_btree_id_root(c, entry->btree_id);
2430
2431         mutex_lock(&c->btree_root_lock);
2432
2433         r->level = entry->level;
2434         r->alive = true;
2435         bkey_copy(&r->key, (struct bkey_i *) entry->start);
2436
2437         mutex_unlock(&c->btree_root_lock);
2438 }
2439
2440 struct jset_entry *
2441 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2442                                     struct jset_entry *end,
2443                                     unsigned long skip)
2444 {
2445         unsigned i;
2446
2447         mutex_lock(&c->btree_root_lock);
2448
2449         for (i = 0; i < btree_id_nr_alive(c); i++) {
2450                 struct btree_root *r = bch2_btree_id_root(c, i);
2451
2452                 if (r->alive && !test_bit(i, &skip)) {
2453                         journal_entry_set(end, BCH_JSET_ENTRY_btree_root,
2454                                           i, r->level, &r->key, r->key.k.u64s);
2455                         end = vstruct_next(end);
2456                 }
2457         }
2458
2459         mutex_unlock(&c->btree_root_lock);
2460
2461         return end;
2462 }
2463
2464 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2465 {
2466         if (c->btree_interior_update_worker)
2467                 destroy_workqueue(c->btree_interior_update_worker);
2468         mempool_exit(&c->btree_interior_update_pool);
2469 }
2470
2471 void bch2_fs_btree_interior_update_init_early(struct bch_fs *c)
2472 {
2473         mutex_init(&c->btree_reserve_cache_lock);
2474         INIT_LIST_HEAD(&c->btree_interior_update_list);
2475         INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2476         mutex_init(&c->btree_interior_update_lock);
2477         INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2478
2479         INIT_LIST_HEAD(&c->pending_node_rewrites);
2480         mutex_init(&c->pending_node_rewrites_lock);
2481 }
2482
2483 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2484 {
2485         c->btree_interior_update_worker =
2486                 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2487         if (!c->btree_interior_update_worker)
2488                 return -BCH_ERR_ENOMEM_btree_interior_update_worker_init;
2489
2490         if (mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2491                                       sizeof(struct btree_update)))
2492                 return -BCH_ERR_ENOMEM_btree_interior_update_pool_init;
2493
2494         return 0;
2495 }