fs/bcachefs/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 *, struct keylist *);
  29 static void bch2_btree_update_add_new_node(struct btree_update *, struct btree *);
  30
  31 static btree_path_idx_t get_unlocked_mut_path(struct btree_trans *trans,
  32                                               enum btree_id btree_id,
  33                                               unsigned level,
  34                                               struct bpos pos)
  35 {
  36         btree_path_idx_t path_idx = bch2_path_get(trans, btree_id, pos, level + 1, level,
  37                              BTREE_ITER_NOPRESERVE|
  38                              BTREE_ITER_INTENT, _RET_IP_);
  39         path_idx = bch2_btree_path_make_mut(trans, path_idx, true, _RET_IP_);
  40
  41         struct btree_path *path = trans->paths + path_idx;
  42         bch2_btree_path_downgrade(trans, path);
  43         __bch2_btree_path_unlock(trans, path);
  44         return path_idx;
  45 }
  46
  47 /* Debug code: */
  48
  49 /*
  50  * Verify that child nodes correctly span parent node's range:
  51  */
  52 static void btree_node_interior_verify(struct bch_fs *c, struct btree *b)
  53 {
  54 #ifdef CONFIG_BCACHEFS_DEBUG
  55         struct bpos next_node = b->data->min_key;
  56         struct btree_node_iter iter;
  57         struct bkey_s_c k;
  58         struct bkey_s_c_btree_ptr_v2 bp;
  59         struct bkey unpacked;
  60         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
  61
  62         BUG_ON(!b->c.level);
  63
  64         if (!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))
  65                 return;
  66
  67         bch2_btree_node_iter_init_from_start(&iter, b);
  68
  69         while (1) {
  70                 k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked);
  71                 if (k.k->type != KEY_TYPE_btree_ptr_v2)
  72                         break;
  73                 bp = bkey_s_c_to_btree_ptr_v2(k);
  74
  75                 if (!bpos_eq(next_node, bp.v->min_key)) {
  76                         bch2_dump_btree_node(c, b);
  77                         bch2_bpos_to_text(&buf1, next_node);
  78                         bch2_bpos_to_text(&buf2, bp.v->min_key);
  79                         panic("expected next min_key %s got %s\n", buf1.buf, buf2.buf);
  80                 }
  81
  82                 bch2_btree_node_iter_advance(&iter, b);
  83
  84                 if (bch2_btree_node_iter_end(&iter)) {
  85                         if (!bpos_eq(k.k->p, b->key.k.p)) {
  86                                 bch2_dump_btree_node(c, b);
  87                                 bch2_bpos_to_text(&buf1, b->key.k.p);
  88                                 bch2_bpos_to_text(&buf2, k.k->p);
  89                                 panic("expected end %s got %s\n", buf1.buf, buf2.buf);
  90                         }
  91                         break;
  92                 }
  93
  94                 next_node = bpos_successor(k.k->p);
  95         }
  96 #endif
  97 }
  98
  99 /* Calculate ideal packed bkey format for new btree nodes: */
 100
 101 static void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b)
 102 {
 103         struct bkey_packed *k;
 104         struct bset_tree *t;
 105         struct bkey uk;
 106
 107         for_each_bset(b, t)
 108                 bset_tree_for_each_key(b, t, k)
 109                         if (!bkey_deleted(k)) {
 110                                 uk = bkey_unpack_key(b, k);
 111                                 bch2_bkey_format_add_key(s, &uk);
 112                         }
 113 }
 114
 115 static struct bkey_format bch2_btree_calc_format(struct btree *b)
 116 {
 117         struct bkey_format_state s;
 118
 119         bch2_bkey_format_init(&s);
 120         bch2_bkey_format_add_pos(&s, b->data->min_key);
 121         bch2_bkey_format_add_pos(&s, b->data->max_key);
 122         __bch2_btree_calc_format(&s, b);
 123
 124         return bch2_bkey_format_done(&s);
 125 }
 126
 127 static size_t btree_node_u64s_with_format(struct btree_nr_keys nr,
 128                                           struct bkey_format *old_f,
 129                                           struct bkey_format *new_f)
 130 {
 131         /* stupid integer promotion rules */
 132         ssize_t delta =
 133             (((int) new_f->key_u64s - old_f->key_u64s) *
 134              (int) nr.packed_keys) +
 135             (((int) new_f->key_u64s - BKEY_U64s) *
 136              (int) nr.unpacked_keys);
 137
 138         BUG_ON(delta + nr.live_u64s < 0);
 139
 140         return nr.live_u64s + delta;
 141 }
 142
 143 /**
 144  * bch2_btree_node_format_fits - check if we could rewrite node with a new format
 145  *
 146  * @c:          filesystem handle
 147  * @b:          btree node to rewrite
 148  * @nr:         number of keys for new node (i.e. b->nr)
 149  * @new_f:      bkey format to translate keys to
 150  *
 151  * Returns: true if all re-packed keys will be able to fit in a new node.
 152  *
 153  * Assumes all keys will successfully pack with the new format.
 154  */
 155 static bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
 156                                  struct btree_nr_keys nr,
 157                                  struct bkey_format *new_f)
 158 {
 159         size_t u64s = btree_node_u64s_with_format(nr, &b->format, new_f);
 160
 161         return __vstruct_bytes(struct btree_node, u64s) < btree_buf_bytes(b);
 162 }
 163
 164 /* Btree node freeing/allocation: */
 165
 166 static void __btree_node_free(struct btree_trans *trans, struct btree *b)
 167 {
 168         struct bch_fs *c = trans->c;
 169
 170         trace_and_count(c, btree_node_free, trans, b);
 171
 172         BUG_ON(btree_node_write_blocked(b));
 173         BUG_ON(btree_node_dirty(b));
 174         BUG_ON(btree_node_need_write(b));
 175         BUG_ON(b == btree_node_root(c, b));
 176         BUG_ON(b->ob.nr);
 177         BUG_ON(!list_empty(&b->write_blocked));
 178         BUG_ON(b->will_make_reachable);
 179
 180         clear_btree_node_noevict(b);
 181
 182         mutex_lock(&c->btree_cache.lock);
 183         list_move(&b->list, &c->btree_cache.freeable);
 184         mutex_unlock(&c->btree_cache.lock);
 185 }
 186
 187 static void bch2_btree_node_free_inmem(struct btree_trans *trans,
 188                                        struct btree_path *path,
 189                                        struct btree *b)
 190 {
 191         struct bch_fs *c = trans->c;
 192         unsigned i, level = b->c.level;
 193
 194         bch2_btree_node_lock_write_nofail(trans, path, &b->c);
 195         bch2_btree_node_hash_remove(&c->btree_cache, b);
 196         __btree_node_free(trans, b);
 197         six_unlock_write(&b->c.lock);
 198         mark_btree_node_locked_noreset(path, level, BTREE_NODE_INTENT_LOCKED);
 199
 200         trans_for_each_path(trans, path, i)
 201                 if (path->l[level].b == b) {
 202                         btree_node_unlock(trans, path, level);
 203                         path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
 204                 }
 205 }
 206
 207 static void bch2_btree_node_free_never_used(struct btree_update *as,
 208                                             struct btree_trans *trans,
 209                                             struct btree *b)
 210 {
 211         struct bch_fs *c = as->c;
 212         struct prealloc_nodes *p = &as->prealloc_nodes[b->c.lock.readers != NULL];
 213         struct btree_path *path;
 214         unsigned i, level = b->c.level;
 215
 216         BUG_ON(!list_empty(&b->write_blocked));
 217         BUG_ON(b->will_make_reachable != (1UL|(unsigned long) as));
 218
 219         b->will_make_reachable = 0;
 220         closure_put(&as->cl);
 221
 222         clear_btree_node_will_make_reachable(b);
 223         clear_btree_node_accessed(b);
 224         clear_btree_node_dirty_acct(c, b);
 225         clear_btree_node_need_write(b);
 226
 227         mutex_lock(&c->btree_cache.lock);
 228         list_del_init(&b->list);
 229         bch2_btree_node_hash_remove(&c->btree_cache, b);
 230         mutex_unlock(&c->btree_cache.lock);
 231
 232         BUG_ON(p->nr >= ARRAY_SIZE(p->b));
 233         p->b[p->nr++] = b;
 234
 235         six_unlock_intent(&b->c.lock);
 236
 237         trans_for_each_path(trans, path, i)
 238                 if (path->l[level].b == b) {
 239                         btree_node_unlock(trans, path, level);
 240                         path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
 241                 }
 242 }
 243
 244 static struct btree *__bch2_btree_node_alloc(struct btree_trans *trans,
 245                                              struct disk_reservation *res,
 246                                              struct closure *cl,
 247                                              bool interior_node,
 248                                              unsigned flags)
 249 {
 250         struct bch_fs *c = trans->c;
 251         struct write_point *wp;
 252         struct btree *b;
 253         BKEY_PADDED_ONSTACK(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
 254         struct open_buckets obs = { .nr = 0 };
 255         struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
 256         enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
 257         unsigned nr_reserve = watermark > BCH_WATERMARK_reclaim
 258                 ? BTREE_NODE_RESERVE
 259                 : 0;
 260         int ret;
 261
 262         mutex_lock(&c->btree_reserve_cache_lock);
 263         if (c->btree_reserve_cache_nr > nr_reserve) {
 264                 struct btree_alloc *a =
 265                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
 266
 267                 obs = a->ob;
 268                 bkey_copy(&tmp.k, &a->k);
 269                 mutex_unlock(&c->btree_reserve_cache_lock);
 270                 goto mem_alloc;
 271         }
 272         mutex_unlock(&c->btree_reserve_cache_lock);
 273
 274 retry:
 275         ret = bch2_alloc_sectors_start_trans(trans,
 276                                       c->opts.metadata_target ?:
 277                                       c->opts.foreground_target,
 278                                       0,
 279                                       writepoint_ptr(&c->btree_write_point),
 280                                       &devs_have,
 281                                       res->nr_replicas,
 282                                       min(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_key_trigger_old(trans, as->btree_id, level, bkey_i_to_s_c(k),
 572                                            BTREE_TRIGGER_TRANSACTIONAL);
 573                 if (ret)
 574                         return ret;
 575         }
 576
 577         for_each_keylist_key(&as->new_keys, k) {
 578                 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
 579
 580                 ret = bch2_key_trigger_new(trans, as->btree_id, level, bkey_i_to_s(k),
 581                                            BTREE_TRIGGER_TRANSACTIONAL);
 582                 if (ret)
 583                         return ret;
 584         }
 585
 586         return 0;
 587 }
 588
 589 static void btree_update_nodes_written(struct btree_update *as)
 590 {
 591         struct bch_fs *c = as->c;
 592         struct btree *b;
 593         struct btree_trans *trans = bch2_trans_get(c);
 594         u64 journal_seq = 0;
 595         unsigned i;
 596         int ret;
 597
 598         /*
 599          * If we're already in an error state, it might be because a btree node
 600          * was never written, and we might be trying to free that same btree
 601          * node here, but it won't have been marked as allocated and we'll see
 602          * spurious disk usage inconsistencies in the transactional part below
 603          * if we don't skip it:
 604          */
 605         ret = bch2_journal_error(&c->journal);
 606         if (ret)
 607                 goto err;
 608
 609         /*
 610          * Wait for any in flight writes to finish before we free the old nodes
 611          * on disk:
 612          */
 613         for (i = 0; i < as->nr_old_nodes; i++) {
 614                 __le64 seq;
 615
 616                 b = as->old_nodes[i];
 617
 618                 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
 619                 seq = b->data ? b->data->keys.seq : 0;
 620                 six_unlock_read(&b->c.lock);
 621
 622                 if (seq == as->old_nodes_seq[i])
 623                         wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight_inner,
 624                                        TASK_UNINTERRUPTIBLE);
 625         }
 626
 627         /*
 628          * We did an update to a parent node where the pointers we added pointed
 629          * to child nodes that weren't written yet: now, the child nodes have
 630          * been written so we can write out the update to the interior node.
 631          */
 632
 633         /*
 634          * We can't call into journal reclaim here: we'd block on the journal
 635          * reclaim lock, but we may need to release the open buckets we have
 636          * pinned in order for other btree updates to make forward progress, and
 637          * journal reclaim does btree updates when flushing bkey_cached entries,
 638          * which may require allocations as well.
 639          */
 640         ret = commit_do(trans, &as->disk_res, &journal_seq,
 641                         BCH_WATERMARK_reclaim|
 642                         BCH_TRANS_COMMIT_no_enospc|
 643                         BCH_TRANS_COMMIT_no_check_rw|
 644                         BCH_TRANS_COMMIT_journal_reclaim,
 645                         btree_update_nodes_written_trans(trans, as));
 646         bch2_trans_unlock(trans);
 647
 648         bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
 649                              "%s", bch2_err_str(ret));
 650 err:
 651         if (as->b) {
 652
 653                 b = as->b;
 654                 btree_path_idx_t path_idx = get_unlocked_mut_path(trans,
 655                                                 as->btree_id, b->c.level, b->key.k.p);
 656                 struct btree_path *path = trans->paths + path_idx;
 657                 /*
 658                  * @b is the node we did the final insert into:
 659                  *
 660                  * On failure to get a journal reservation, we still have to
 661                  * unblock the write and allow most of the write path to happen
 662                  * so that shutdown works, but the i->journal_seq mechanism
 663                  * won't work to prevent the btree write from being visible (we
 664                  * didn't get a journal sequence number) - instead
 665                  * __bch2_btree_node_write() doesn't do the actual write if
 666                  * we're in journal error state:
 667                  */
 668
 669                 /*
 670                  * Ensure transaction is unlocked before using
 671                  * btree_node_lock_nopath() (the use of which is always suspect,
 672                  * we need to work on removing this in the future)
 673                  *
 674                  * It should be, but get_unlocked_mut_path() -> bch2_path_get()
 675                  * calls bch2_path_upgrade(), before we call path_make_mut(), so
 676                  * we may rarely end up with a locked path besides the one we
 677                  * have here:
 678                  */
 679                 bch2_trans_unlock(trans);
 680                 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
 681                 mark_btree_node_locked(trans, path, b->c.level, BTREE_NODE_INTENT_LOCKED);
 682                 path->l[b->c.level].lock_seq = six_lock_seq(&b->c.lock);
 683                 path->l[b->c.level].b = b;
 684
 685                 bch2_btree_node_lock_write_nofail(trans, path, &b->c);
 686
 687                 mutex_lock(&c->btree_interior_update_lock);
 688
 689                 list_del(&as->write_blocked_list);
 690                 if (list_empty(&b->write_blocked))
 691                         clear_btree_node_write_blocked(b);
 692
 693                 /*
 694                  * Node might have been freed, recheck under
 695                  * btree_interior_update_lock:
 696                  */
 697                 if (as->b == b) {
 698                         BUG_ON(!b->c.level);
 699                         BUG_ON(!btree_node_dirty(b));
 700
 701                         if (!ret) {
 702                                 struct bset *last = btree_bset_last(b);
 703
 704                                 last->journal_seq = cpu_to_le64(
 705                                                              max(journal_seq,
 706                                                                  le64_to_cpu(last->journal_seq)));
 707
 708                                 bch2_btree_add_journal_pin(c, b, journal_seq);
 709                         } else {
 710                                 /*
 711                                  * If we didn't get a journal sequence number we
 712                                  * can't write this btree node, because recovery
 713                                  * won't know to ignore this write:
 714                                  */
 715                                 set_btree_node_never_write(b);
 716                         }
 717                 }
 718
 719                 mutex_unlock(&c->btree_interior_update_lock);
 720
 721                 mark_btree_node_locked_noreset(path, b->c.level, BTREE_NODE_INTENT_LOCKED);
 722                 six_unlock_write(&b->c.lock);
 723
 724                 btree_node_write_if_need(c, b, SIX_LOCK_intent);
 725                 btree_node_unlock(trans, path, b->c.level);
 726                 bch2_path_put(trans, path_idx, true);
 727         }
 728
 729         bch2_journal_pin_drop(&c->journal, &as->journal);
 730
 731         mutex_lock(&c->btree_interior_update_lock);
 732         for (i = 0; i < as->nr_new_nodes; i++) {
 733                 b = as->new_nodes[i];
 734
 735                 BUG_ON(b->will_make_reachable != (unsigned long) as);
 736                 b->will_make_reachable = 0;
 737                 clear_btree_node_will_make_reachable(b);
 738         }
 739         mutex_unlock(&c->btree_interior_update_lock);
 740
 741         for (i = 0; i < as->nr_new_nodes; i++) {
 742                 b = as->new_nodes[i];
 743
 744                 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
 745                 btree_node_write_if_need(c, b, SIX_LOCK_read);
 746                 six_unlock_read(&b->c.lock);
 747         }
 748
 749         for (i = 0; i < as->nr_open_buckets; i++)
 750                 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
 751
 752         bch2_btree_update_free(as, trans);
 753         bch2_trans_put(trans);
 754 }
 755
 756 static void btree_interior_update_work(struct work_struct *work)
 757 {
 758         struct bch_fs *c =
 759                 container_of(work, struct bch_fs, btree_interior_update_work);
 760         struct btree_update *as;
 761
 762         while (1) {
 763                 mutex_lock(&c->btree_interior_update_lock);
 764                 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
 765                                               struct btree_update, unwritten_list);
 766                 if (as && !as->nodes_written)
 767                         as = NULL;
 768                 mutex_unlock(&c->btree_interior_update_lock);
 769
 770                 if (!as)
 771                         break;
 772
 773                 btree_update_nodes_written(as);
 774         }
 775 }
 776
 777 static CLOSURE_CALLBACK(btree_update_set_nodes_written)
 778 {
 779         closure_type(as, struct btree_update, cl);
 780         struct bch_fs *c = as->c;
 781
 782         mutex_lock(&c->btree_interior_update_lock);
 783         as->nodes_written = true;
 784         mutex_unlock(&c->btree_interior_update_lock);
 785
 786         queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
 787 }
 788
 789 /*
 790  * We're updating @b with pointers to nodes that haven't finished writing yet:
 791  * block @b from being written until @as completes
 792  */
 793 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
 794 {
 795         struct bch_fs *c = as->c;
 796
 797         mutex_lock(&c->btree_interior_update_lock);
 798         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 799
 800         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 801         BUG_ON(!btree_node_dirty(b));
 802         BUG_ON(!b->c.level);
 803
 804         as->mode        = BTREE_INTERIOR_UPDATING_NODE;
 805         as->b           = b;
 806
 807         set_btree_node_write_blocked(b);
 808         list_add(&as->write_blocked_list, &b->write_blocked);
 809
 810         mutex_unlock(&c->btree_interior_update_lock);
 811 }
 812
 813 static int bch2_update_reparent_journal_pin_flush(struct journal *j,
 814                                 struct journal_entry_pin *_pin, u64 seq)
 815 {
 816         return 0;
 817 }
 818
 819 static void btree_update_reparent(struct btree_update *as,
 820                                   struct btree_update *child)
 821 {
 822         struct bch_fs *c = as->c;
 823
 824         lockdep_assert_held(&c->btree_interior_update_lock);
 825
 826         child->b = NULL;
 827         child->mode = BTREE_INTERIOR_UPDATING_AS;
 828
 829         bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal,
 830                               bch2_update_reparent_journal_pin_flush);
 831 }
 832
 833 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
 834 {
 835         struct bkey_i *insert = &b->key;
 836         struct bch_fs *c = as->c;
 837
 838         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 839
 840         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
 841                ARRAY_SIZE(as->journal_entries));
 842
 843         as->journal_u64s +=
 844                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
 845                                   BCH_JSET_ENTRY_btree_root,
 846                                   b->c.btree_id, b->c.level,
 847                                   insert, insert->k.u64s);
 848
 849         mutex_lock(&c->btree_interior_update_lock);
 850         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 851
 852         as->mode        = BTREE_INTERIOR_UPDATING_ROOT;
 853         mutex_unlock(&c->btree_interior_update_lock);
 854 }
 855
 856 /*
 857  * bch2_btree_update_add_new_node:
 858  *
 859  * This causes @as to wait on @b to be written, before it gets to
 860  * bch2_btree_update_nodes_written
 861  *
 862  * Additionally, it sets b->will_make_reachable to prevent any additional writes
 863  * to @b from happening besides the first until @b is reachable on disk
 864  *
 865  * And it adds @b to the list of @as's new nodes, so that we can update sector
 866  * counts in bch2_btree_update_nodes_written:
 867  */
 868 static void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
 869 {
 870         struct bch_fs *c = as->c;
 871
 872         closure_get(&as->cl);
 873
 874         mutex_lock(&c->btree_interior_update_lock);
 875         BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
 876         BUG_ON(b->will_make_reachable);
 877
 878         as->new_nodes[as->nr_new_nodes++] = b;
 879         b->will_make_reachable = 1UL|(unsigned long) as;
 880         set_btree_node_will_make_reachable(b);
 881
 882         mutex_unlock(&c->btree_interior_update_lock);
 883
 884         btree_update_add_key(as, &as->new_keys, b);
 885
 886         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 887                 unsigned bytes = vstruct_end(&b->data->keys) - (void *) b->data;
 888                 unsigned sectors = round_up(bytes, block_bytes(c)) >> 9;
 889
 890                 bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written =
 891                         cpu_to_le16(sectors);
 892         }
 893 }
 894
 895 /*
 896  * returns true if @b was a new node
 897  */
 898 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
 899 {
 900         struct btree_update *as;
 901         unsigned long v;
 902         unsigned i;
 903
 904         mutex_lock(&c->btree_interior_update_lock);
 905         /*
 906          * When b->will_make_reachable != 0, it owns a ref on as->cl that's
 907          * dropped when it gets written by bch2_btree_complete_write - the
 908          * xchg() is for synchronization with bch2_btree_complete_write:
 909          */
 910         v = xchg(&b->will_make_reachable, 0);
 911         clear_btree_node_will_make_reachable(b);
 912         as = (struct btree_update *) (v & ~1UL);
 913
 914         if (!as) {
 915                 mutex_unlock(&c->btree_interior_update_lock);
 916                 return;
 917         }
 918
 919         for (i = 0; i < as->nr_new_nodes; i++)
 920                 if (as->new_nodes[i] == b)
 921                         goto found;
 922
 923         BUG();
 924 found:
 925         array_remove_item(as->new_nodes, as->nr_new_nodes, i);
 926         mutex_unlock(&c->btree_interior_update_lock);
 927
 928         if (v & 1)
 929                 closure_put(&as->cl);
 930 }
 931
 932 static void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
 933 {
 934         while (b->ob.nr)
 935                 as->open_buckets[as->nr_open_buckets++] =
 936                         b->ob.v[--b->ob.nr];
 937 }
 938
 939 static int bch2_btree_update_will_free_node_journal_pin_flush(struct journal *j,
 940                                 struct journal_entry_pin *_pin, u64 seq)
 941 {
 942         return 0;
 943 }
 944
 945 /*
 946  * @b is being split/rewritten: it may have pointers to not-yet-written btree
 947  * nodes and thus outstanding btree_updates - redirect @b's
 948  * btree_updates to point to this btree_update:
 949  */
 950 static void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 951                                                       struct btree *b)
 952 {
 953         struct bch_fs *c = as->c;
 954         struct btree_update *p, *n;
 955         struct btree_write *w;
 956
 957         set_btree_node_dying(b);
 958
 959         if (btree_node_fake(b))
 960                 return;
 961
 962         mutex_lock(&c->btree_interior_update_lock);
 963
 964         /*
 965          * Does this node have any btree_update operations preventing
 966          * it from being written?
 967          *
 968          * If so, redirect them to point to this btree_update: we can
 969          * write out our new nodes, but we won't make them visible until those
 970          * operations complete
 971          */
 972         list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
 973                 list_del_init(&p->write_blocked_list);
 974                 btree_update_reparent(as, p);
 975
 976                 /*
 977                  * for flush_held_btree_writes() waiting on updates to flush or
 978                  * nodes to be writeable:
 979                  */
 980                 closure_wake_up(&c->btree_interior_update_wait);
 981         }
 982
 983         clear_btree_node_dirty_acct(c, b);
 984         clear_btree_node_need_write(b);
 985         clear_btree_node_write_blocked(b);
 986
 987         /*
 988          * Does this node have unwritten data that has a pin on the journal?
 989          *
 990          * If so, transfer that pin to the btree_update operation -
 991          * note that if we're freeing multiple nodes, we only need to keep the
 992          * oldest pin of any of the nodes we're freeing. We'll release the pin
 993          * when the new nodes are persistent and reachable on disk:
 994          */
 995         w = btree_current_write(b);
 996         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal,
 997                               bch2_btree_update_will_free_node_journal_pin_flush);
 998         bch2_journal_pin_drop(&c->journal, &w->journal);
 999
1000         w = btree_prev_write(b);
1001         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal,
1002                               bch2_btree_update_will_free_node_journal_pin_flush);
1003         bch2_journal_pin_drop(&c->journal, &w->journal);
1004
1005         mutex_unlock(&c->btree_interior_update_lock);
1006
1007         /*
1008          * Is this a node that isn't reachable on disk yet?
1009          *
1010          * Nodes that aren't reachable yet have writes blocked until they're
1011          * reachable - now that we've cancelled any pending writes and moved
1012          * things waiting on that write to wait on this update, we can drop this
1013          * node from the list of nodes that the other update is making
1014          * reachable, prior to freeing it:
1015          */
1016         btree_update_drop_new_node(c, b);
1017
1018         btree_update_add_key(as, &as->old_keys, b);
1019
1020         as->old_nodes[as->nr_old_nodes] = b;
1021         as->old_nodes_seq[as->nr_old_nodes] = b->data->keys.seq;
1022         as->nr_old_nodes++;
1023 }
1024
1025 static void bch2_btree_update_done(struct btree_update *as, struct btree_trans *trans)
1026 {
1027         struct bch_fs *c = as->c;
1028         u64 start_time = as->start_time;
1029
1030         BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
1031
1032         if (as->took_gc_lock)
1033                 up_read(&as->c->gc_lock);
1034         as->took_gc_lock = false;
1035
1036         bch2_btree_reserve_put(as, trans);
1037
1038         continue_at(&as->cl, btree_update_set_nodes_written,
1039                     as->c->btree_interior_update_worker);
1040
1041         bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_foreground],
1042                                start_time);
1043 }
1044
1045 static struct btree_update *
1046 bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
1047                         unsigned level, bool split, unsigned flags)
1048 {
1049         struct bch_fs *c = trans->c;
1050         struct btree_update *as;
1051         u64 start_time = local_clock();
1052         int disk_res_flags = (flags & BCH_TRANS_COMMIT_no_enospc)
1053                 ? BCH_DISK_RESERVATION_NOFAIL : 0;
1054         unsigned nr_nodes[2] = { 0, 0 };
1055         unsigned update_level = level;
1056         enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
1057         int ret = 0;
1058         u32 restart_count = trans->restart_count;
1059
1060         BUG_ON(!path->should_be_locked);
1061
1062         if (watermark == BCH_WATERMARK_copygc)
1063                 watermark = BCH_WATERMARK_btree_copygc;
1064         if (watermark < BCH_WATERMARK_btree)
1065                 watermark = BCH_WATERMARK_btree;
1066
1067         flags &= ~BCH_WATERMARK_MASK;
1068         flags |= watermark;
1069
1070         if (watermark < c->journal.watermark) {
1071                 struct journal_res res = { 0 };
1072                 unsigned journal_flags = watermark|JOURNAL_RES_GET_CHECK;
1073
1074                 if ((flags & BCH_TRANS_COMMIT_journal_reclaim) &&
1075                     watermark != BCH_WATERMARK_reclaim)
1076                         journal_flags |= JOURNAL_RES_GET_NONBLOCK;
1077
1078                 ret = drop_locks_do(trans,
1079                         bch2_journal_res_get(&c->journal, &res, 1, journal_flags));
1080                 if (bch2_err_matches(ret, BCH_ERR_operation_blocked))
1081                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
1082                 if (ret)
1083                         return ERR_PTR(ret);
1084         }
1085
1086         while (1) {
1087                 nr_nodes[!!update_level] += 1 + split;
1088                 update_level++;
1089
1090                 ret = bch2_btree_path_upgrade(trans, path, update_level + 1);
1091                 if (ret)
1092                         return ERR_PTR(ret);
1093
1094                 if (!btree_path_node(path, update_level)) {
1095                         /* Allocating new root? */
1096                         nr_nodes[1] += split;
1097                         update_level = BTREE_MAX_DEPTH;
1098                         break;
1099                 }
1100
1101                 /*
1102                  * Always check for space for two keys, even if we won't have to
1103                  * split at prior level - it might have been a merge instead:
1104                  */
1105                 if (bch2_btree_node_insert_fits(path->l[update_level].b,
1106                                                 BKEY_BTREE_PTR_U64s_MAX * 2))
1107                         break;
1108
1109                 split = path->l[update_level].b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c);
1110         }
1111
1112         if (!down_read_trylock(&c->gc_lock)) {
1113                 ret = drop_locks_do(trans, (down_read(&c->gc_lock), 0));
1114                 if (ret) {
1115                         up_read(&c->gc_lock);
1116                         return ERR_PTR(ret);
1117                 }
1118         }
1119
1120         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOFS);
1121         memset(as, 0, sizeof(*as));
1122         closure_init(&as->cl, NULL);
1123         as->c           = c;
1124         as->start_time  = start_time;
1125         as->ip_started  = _RET_IP_;
1126         as->mode        = BTREE_INTERIOR_NO_UPDATE;
1127         as->took_gc_lock = true;
1128         as->btree_id    = path->btree_id;
1129         as->update_level = update_level;
1130         INIT_LIST_HEAD(&as->list);
1131         INIT_LIST_HEAD(&as->unwritten_list);
1132         INIT_LIST_HEAD(&as->write_blocked_list);
1133         bch2_keylist_init(&as->old_keys, as->_old_keys);
1134         bch2_keylist_init(&as->new_keys, as->_new_keys);
1135         bch2_keylist_init(&as->parent_keys, as->inline_keys);
1136
1137         mutex_lock(&c->btree_interior_update_lock);
1138         list_add_tail(&as->list, &c->btree_interior_update_list);
1139         mutex_unlock(&c->btree_interior_update_lock);
1140
1141         /*
1142          * We don't want to allocate if we're in an error state, that can cause
1143          * deadlock on emergency shutdown due to open buckets getting stuck in
1144          * the btree_reserve_cache after allocator shutdown has cleared it out.
1145          * This check needs to come after adding us to the btree_interior_update
1146          * list but before calling bch2_btree_reserve_get, to synchronize with
1147          * __bch2_fs_read_only().
1148          */
1149         ret = bch2_journal_error(&c->journal);
1150         if (ret)
1151                 goto err;
1152
1153         ret = bch2_disk_reservation_get(c, &as->disk_res,
1154                         (nr_nodes[0] + nr_nodes[1]) * btree_sectors(c),
1155                         c->opts.metadata_replicas,
1156                         disk_res_flags);
1157         if (ret)
1158                 goto err;
1159
1160         ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, NULL);
1161         if (bch2_err_matches(ret, ENOSPC) ||
1162             bch2_err_matches(ret, ENOMEM)) {
1163                 struct closure cl;
1164
1165                 /*
1166                  * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
1167                  * flag
1168                  */
1169                 if (bch2_err_matches(ret, ENOSPC) &&
1170                     (flags & BCH_TRANS_COMMIT_journal_reclaim) &&
1171                     watermark != BCH_WATERMARK_reclaim) {
1172                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
1173                         goto err;
1174                 }
1175
1176                 closure_init_stack(&cl);
1177
1178                 do {
1179                         ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, &cl);
1180
1181                         bch2_trans_unlock(trans);
1182                         closure_sync(&cl);
1183                 } while (bch2_err_matches(ret, BCH_ERR_operation_blocked));
1184         }
1185
1186         if (ret) {
1187                 trace_and_count(c, btree_reserve_get_fail, trans->fn,
1188                                 _RET_IP_, nr_nodes[0] + nr_nodes[1], ret);
1189                 goto err;
1190         }
1191
1192         ret = bch2_trans_relock(trans);
1193         if (ret)
1194                 goto err;
1195
1196         bch2_trans_verify_not_restarted(trans, restart_count);
1197         return as;
1198 err:
1199         bch2_btree_update_free(as, trans);
1200         if (!bch2_err_matches(ret, ENOSPC) &&
1201             !bch2_err_matches(ret, EROFS) &&
1202             ret != -BCH_ERR_journal_reclaim_would_deadlock)
1203                 bch_err_fn_ratelimited(c, ret);
1204         return ERR_PTR(ret);
1205 }
1206
1207 /* Btree root updates: */
1208
1209 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1210 {
1211         /* Root nodes cannot be reaped */
1212         mutex_lock(&c->btree_cache.lock);
1213         list_del_init(&b->list);
1214         mutex_unlock(&c->btree_cache.lock);
1215
1216         mutex_lock(&c->btree_root_lock);
1217         bch2_btree_id_root(c, b->c.btree_id)->b = b;
1218         mutex_unlock(&c->btree_root_lock);
1219
1220         bch2_recalc_btree_reserve(c);
1221 }
1222
1223 static void bch2_btree_set_root(struct btree_update *as,
1224                                 struct btree_trans *trans,
1225                                 struct btree_path *path,
1226                                 struct btree *b)
1227 {
1228         struct bch_fs *c = as->c;
1229         struct btree *old;
1230
1231         trace_and_count(c, btree_node_set_root, trans, b);
1232
1233         old = btree_node_root(c, b);
1234
1235         /*
1236          * Ensure no one is using the old root while we switch to the
1237          * new root:
1238          */
1239         bch2_btree_node_lock_write_nofail(trans, path, &old->c);
1240
1241         bch2_btree_set_root_inmem(c, b);
1242
1243         btree_update_updated_root(as, b);
1244
1245         /*
1246          * Unlock old root after new root is visible:
1247          *
1248          * The new root isn't persistent, but that's ok: we still have
1249          * an intent lock on the new root, and any updates that would
1250          * depend on the new root would have to update the new root.
1251          */
1252         bch2_btree_node_unlock_write(trans, path, old);
1253 }
1254
1255 /* Interior node updates: */
1256
1257 static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
1258                                         struct btree_trans *trans,
1259                                         struct btree_path *path,
1260                                         struct btree *b,
1261                                         struct btree_node_iter *node_iter,
1262                                         struct bkey_i *insert)
1263 {
1264         struct bch_fs *c = as->c;
1265         struct bkey_packed *k;
1266         struct printbuf buf = PRINTBUF;
1267         unsigned long old, new, v;
1268
1269         BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 &&
1270                !btree_ptr_sectors_written(insert));
1271
1272         if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
1273                 bch2_journal_key_overwritten(c, b->c.btree_id, b->c.level, insert->k.p);
1274
1275         if (bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1276                               btree_node_type(b), WRITE, &buf) ?:
1277             bch2_bkey_in_btree_node(c, b, bkey_i_to_s_c(insert), &buf)) {
1278                 printbuf_reset(&buf);
1279                 prt_printf(&buf, "inserting invalid bkey\n  ");
1280                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
1281                 prt_printf(&buf, "\n  ");
1282                 bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1283                                   btree_node_type(b), WRITE, &buf);
1284                 bch2_bkey_in_btree_node(c, b, bkey_i_to_s_c(insert), &buf);
1285
1286                 bch2_fs_inconsistent(c, "%s", buf.buf);
1287                 dump_stack();
1288         }
1289
1290         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1291                ARRAY_SIZE(as->journal_entries));
1292
1293         as->journal_u64s +=
1294                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1295                                   BCH_JSET_ENTRY_btree_keys,
1296                                   b->c.btree_id, b->c.level,
1297                                   insert, insert->k.u64s);
1298
1299         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1300                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1301                 bch2_btree_node_iter_advance(node_iter, b);
1302
1303         bch2_btree_bset_insert_key(trans, path, b, node_iter, insert);
1304         set_btree_node_dirty_acct(c, b);
1305
1306         v = READ_ONCE(b->flags);
1307         do {
1308                 old = new = v;
1309
1310                 new &= ~BTREE_WRITE_TYPE_MASK;
1311                 new |= BTREE_WRITE_interior;
1312                 new |= 1 << BTREE_NODE_need_write;
1313         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1314
1315         printbuf_exit(&buf);
1316 }
1317
1318 static void
1319 __bch2_btree_insert_keys_interior(struct btree_update *as,
1320                                   struct btree_trans *trans,
1321                                   struct btree_path *path,
1322                                   struct btree *b,
1323                                   struct btree_node_iter node_iter,
1324                                   struct keylist *keys)
1325 {
1326         struct bkey_i *insert = bch2_keylist_front(keys);
1327         struct bkey_packed *k;
1328
1329         BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1330
1331         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1332                (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1333                 ;
1334
1335         while (!bch2_keylist_empty(keys)) {
1336                 insert = bch2_keylist_front(keys);
1337
1338                 if (bpos_gt(insert->k.p, b->key.k.p))
1339                         break;
1340
1341                 bch2_insert_fixup_btree_ptr(as, trans, path, b, &node_iter, insert);
1342                 bch2_keylist_pop_front(keys);
1343         }
1344 }
1345
1346 /*
1347  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1348  * node)
1349  */
1350 static void __btree_split_node(struct btree_update *as,
1351                                struct btree_trans *trans,
1352                                struct btree *b,
1353                                struct btree *n[2])
1354 {
1355         struct bkey_packed *k;
1356         struct bpos n1_pos = POS_MIN;
1357         struct btree_node_iter iter;
1358         struct bset *bsets[2];
1359         struct bkey_format_state format[2];
1360         struct bkey_packed *out[2];
1361         struct bkey uk;
1362         unsigned u64s, n1_u64s = (b->nr.live_u64s * 3) / 5;
1363         struct { unsigned nr_keys, val_u64s; } nr_keys[2];
1364         int i;
1365
1366         memset(&nr_keys, 0, sizeof(nr_keys));
1367
1368         for (i = 0; i < 2; i++) {
1369                 BUG_ON(n[i]->nsets != 1);
1370
1371                 bsets[i] = btree_bset_first(n[i]);
1372                 out[i] = bsets[i]->start;
1373
1374                 SET_BTREE_NODE_SEQ(n[i]->data, BTREE_NODE_SEQ(b->data) + 1);
1375                 bch2_bkey_format_init(&format[i]);
1376         }
1377
1378         u64s = 0;
1379         for_each_btree_node_key(b, k, &iter) {
1380                 if (bkey_deleted(k))
1381                         continue;
1382
1383                 i = u64s >= n1_u64s;
1384                 u64s += k->u64s;
1385                 uk = bkey_unpack_key(b, k);
1386                 if (!i)
1387                         n1_pos = uk.p;
1388                 bch2_bkey_format_add_key(&format[i], &uk);
1389
1390                 nr_keys[i].nr_keys++;
1391                 nr_keys[i].val_u64s += bkeyp_val_u64s(&b->format, k);
1392         }
1393
1394         btree_set_min(n[0], b->data->min_key);
1395         btree_set_max(n[0], n1_pos);
1396         btree_set_min(n[1], bpos_successor(n1_pos));
1397         btree_set_max(n[1], b->data->max_key);
1398
1399         for (i = 0; i < 2; i++) {
1400                 bch2_bkey_format_add_pos(&format[i], n[i]->data->min_key);
1401                 bch2_bkey_format_add_pos(&format[i], n[i]->data->max_key);
1402
1403                 n[i]->data->format = bch2_bkey_format_done(&format[i]);
1404
1405                 unsigned u64s = nr_keys[i].nr_keys * n[i]->data->format.key_u64s +
1406                         nr_keys[i].val_u64s;
1407                 if (__vstruct_bytes(struct btree_node, u64s) > btree_buf_bytes(b))
1408                         n[i]->data->format = b->format;
1409
1410                 btree_node_set_format(n[i], n[i]->data->format);
1411         }
1412
1413         u64s = 0;
1414         for_each_btree_node_key(b, k, &iter) {
1415                 if (bkey_deleted(k))
1416                         continue;
1417
1418                 i = u64s >= n1_u64s;
1419                 u64s += k->u64s;
1420
1421                 if (bch2_bkey_transform(&n[i]->format, out[i], bkey_packed(k)
1422                                         ? &b->format: &bch2_bkey_format_current, k))
1423                         out[i]->format = KEY_FORMAT_LOCAL_BTREE;
1424                 else
1425                         bch2_bkey_unpack(b, (void *) out[i], k);
1426
1427                 out[i]->needs_whiteout = false;
1428
1429                 btree_keys_account_key_add(&n[i]->nr, 0, out[i]);
1430                 out[i] = bkey_p_next(out[i]);
1431         }
1432
1433         for (i = 0; i < 2; i++) {
1434                 bsets[i]->u64s = cpu_to_le16((u64 *) out[i] - bsets[i]->_data);
1435
1436                 BUG_ON(!bsets[i]->u64s);
1437
1438                 set_btree_bset_end(n[i], n[i]->set);
1439
1440                 btree_node_reset_sib_u64s(n[i]);
1441
1442                 bch2_verify_btree_nr_keys(n[i]);
1443
1444                 if (b->c.level)
1445                         btree_node_interior_verify(as->c, n[i]);
1446         }
1447 }
1448
1449 /*
1450  * For updates to interior nodes, we've got to do the insert before we split
1451  * because the stuff we're inserting has to be inserted atomically. Post split,
1452  * the keys might have to go in different nodes and the split would no longer be
1453  * atomic.
1454  *
1455  * Worse, if the insert is from btree node coalescing, if we do the insert after
1456  * we do the split (and pick the pivot) - the pivot we pick might be between
1457  * nodes that were coalesced, and thus in the middle of a child node post
1458  * coalescing:
1459  */
1460 static void btree_split_insert_keys(struct btree_update *as,
1461                                     struct btree_trans *trans,
1462                                     btree_path_idx_t path_idx,
1463                                     struct btree *b,
1464                                     struct keylist *keys)
1465 {
1466         struct btree_path *path = trans->paths + path_idx;
1467
1468         if (!bch2_keylist_empty(keys) &&
1469             bpos_le(bch2_keylist_front(keys)->k.p, b->data->max_key)) {
1470                 struct btree_node_iter node_iter;
1471
1472                 bch2_btree_node_iter_init(&node_iter, b, &bch2_keylist_front(keys)->k.p);
1473
1474                 __bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
1475
1476                 btree_node_interior_verify(as->c, b);
1477         }
1478 }
1479
1480 static int btree_split(struct btree_update *as, struct btree_trans *trans,
1481                        btree_path_idx_t path, struct btree *b,
1482                        struct keylist *keys)
1483 {
1484         struct bch_fs *c = as->c;
1485         struct btree *parent = btree_node_parent(trans->paths + path, b);
1486         struct btree *n1, *n2 = NULL, *n3 = NULL;
1487         btree_path_idx_t path1 = 0, path2 = 0;
1488         u64 start_time = local_clock();
1489         int ret = 0;
1490
1491         BUG_ON(!parent && (b != btree_node_root(c, b)));
1492         BUG_ON(parent && !btree_node_intent_locked(trans->paths + path, b->c.level + 1));
1493
1494         bch2_btree_interior_update_will_free_node(as, b);
1495
1496         if (b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c)) {
1497                 struct btree *n[2];
1498
1499                 trace_and_count(c, btree_node_split, trans, b);
1500
1501                 n[0] = n1 = bch2_btree_node_alloc(as, trans, b->c.level);
1502                 n[1] = n2 = bch2_btree_node_alloc(as, trans, b->c.level);
1503
1504                 __btree_split_node(as, trans, b, n);
1505
1506                 if (keys) {
1507                         btree_split_insert_keys(as, trans, path, n1, keys);
1508                         btree_split_insert_keys(as, trans, path, n2, keys);
1509                         BUG_ON(!bch2_keylist_empty(keys));
1510                 }
1511
1512                 bch2_btree_build_aux_trees(n2);
1513                 bch2_btree_build_aux_trees(n1);
1514
1515                 bch2_btree_update_add_new_node(as, n1);
1516                 bch2_btree_update_add_new_node(as, n2);
1517                 six_unlock_write(&n2->c.lock);
1518                 six_unlock_write(&n1->c.lock);
1519
1520                 path1 = get_unlocked_mut_path(trans, as->btree_id, n1->c.level, n1->key.k.p);
1521                 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1522                 mark_btree_node_locked(trans, trans->paths + path1, n1->c.level, BTREE_NODE_INTENT_LOCKED);
1523                 bch2_btree_path_level_init(trans, trans->paths + path1, n1);
1524
1525                 path2 = get_unlocked_mut_path(trans, as->btree_id, n2->c.level, n2->key.k.p);
1526                 six_lock_increment(&n2->c.lock, SIX_LOCK_intent);
1527                 mark_btree_node_locked(trans, trans->paths + path2, n2->c.level, BTREE_NODE_INTENT_LOCKED);
1528                 bch2_btree_path_level_init(trans, trans->paths + path2, n2);
1529
1530                 /*
1531                  * Note that on recursive parent_keys == keys, so we
1532                  * can't start adding new keys to parent_keys before emptying it
1533                  * out (which we did with btree_split_insert_keys() above)
1534                  */
1535                 bch2_keylist_add(&as->parent_keys, &n1->key);
1536                 bch2_keylist_add(&as->parent_keys, &n2->key);
1537
1538                 if (!parent) {
1539                         /* Depth increases, make a new root */
1540                         n3 = __btree_root_alloc(as, trans, b->c.level + 1);
1541
1542                         bch2_btree_update_add_new_node(as, n3);
1543                         six_unlock_write(&n3->c.lock);
1544
1545                         trans->paths[path2].locks_want++;
1546                         BUG_ON(btree_node_locked(trans->paths + path2, n3->c.level));
1547                         six_lock_increment(&n3->c.lock, SIX_LOCK_intent);
1548                         mark_btree_node_locked(trans, trans->paths + path2, n3->c.level, BTREE_NODE_INTENT_LOCKED);
1549                         bch2_btree_path_level_init(trans, trans->paths + path2, n3);
1550
1551                         n3->sib_u64s[0] = U16_MAX;
1552                         n3->sib_u64s[1] = U16_MAX;
1553
1554                         btree_split_insert_keys(as, trans, path, n3, &as->parent_keys);
1555                 }
1556         } else {
1557                 trace_and_count(c, btree_node_compact, trans, b);
1558
1559                 n1 = bch2_btree_node_alloc_replacement(as, trans, b);
1560
1561                 if (keys) {
1562                         btree_split_insert_keys(as, trans, path, n1, keys);
1563                         BUG_ON(!bch2_keylist_empty(keys));
1564                 }
1565
1566                 bch2_btree_build_aux_trees(n1);
1567                 bch2_btree_update_add_new_node(as, n1);
1568                 six_unlock_write(&n1->c.lock);
1569
1570                 path1 = get_unlocked_mut_path(trans, as->btree_id, n1->c.level, n1->key.k.p);
1571                 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1572                 mark_btree_node_locked(trans, trans->paths + path1, n1->c.level, BTREE_NODE_INTENT_LOCKED);
1573                 bch2_btree_path_level_init(trans, trans->paths + path1, n1);
1574
1575                 if (parent)
1576                         bch2_keylist_add(&as->parent_keys, &n1->key);
1577         }
1578
1579         /* New nodes all written, now make them visible: */
1580
1581         if (parent) {
1582                 /* Split a non root node */
1583                 ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys);
1584                 if (ret)
1585                         goto err;
1586         } else if (n3) {
1587                 bch2_btree_set_root(as, trans, trans->paths + path, n3);
1588         } else {
1589                 /* Root filled up but didn't need to be split */
1590                 bch2_btree_set_root(as, trans, trans->paths + path, n1);
1591         }
1592
1593         if (n3) {
1594                 bch2_btree_update_get_open_buckets(as, n3);
1595                 bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0);
1596         }
1597         if (n2) {
1598                 bch2_btree_update_get_open_buckets(as, n2);
1599                 bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0);
1600         }
1601         bch2_btree_update_get_open_buckets(as, n1);
1602         bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1603
1604         /*
1605          * The old node must be freed (in memory) _before_ unlocking the new
1606          * nodes - else another thread could re-acquire a read lock on the old
1607          * node after another thread has locked and updated the new node, thus
1608          * seeing stale data:
1609          */
1610         bch2_btree_node_free_inmem(trans, trans->paths + path, b);
1611
1612         if (n3)
1613                 bch2_trans_node_add(trans, trans->paths + path, n3);
1614         if (n2)
1615                 bch2_trans_node_add(trans, trans->paths + path2, n2);
1616         bch2_trans_node_add(trans, trans->paths + path1, n1);
1617
1618         if (n3)
1619                 six_unlock_intent(&n3->c.lock);
1620         if (n2)
1621                 six_unlock_intent(&n2->c.lock);
1622         six_unlock_intent(&n1->c.lock);
1623 out:
1624         if (path2) {
1625                 __bch2_btree_path_unlock(trans, trans->paths + path2);
1626                 bch2_path_put(trans, path2, true);
1627         }
1628         if (path1) {
1629                 __bch2_btree_path_unlock(trans, trans->paths + path1);
1630                 bch2_path_put(trans, path1, true);
1631         }
1632
1633         bch2_trans_verify_locks(trans);
1634
1635         bch2_time_stats_update(&c->times[n2
1636                                ? BCH_TIME_btree_node_split
1637                                : BCH_TIME_btree_node_compact],
1638                                start_time);
1639         return ret;
1640 err:
1641         if (n3)
1642                 bch2_btree_node_free_never_used(as, trans, n3);
1643         if (n2)
1644                 bch2_btree_node_free_never_used(as, trans, n2);
1645         bch2_btree_node_free_never_used(as, trans, n1);
1646         goto out;
1647 }
1648
1649 static void
1650 bch2_btree_insert_keys_interior(struct btree_update *as,
1651                                 struct btree_trans *trans,
1652                                 struct btree_path *path,
1653                                 struct btree *b,
1654                                 struct keylist *keys)
1655 {
1656         struct btree_path *linked;
1657         unsigned i;
1658
1659         __bch2_btree_insert_keys_interior(as, trans, path, b,
1660                                           path->l[b->c.level].iter, keys);
1661
1662         btree_update_updated_node(as, b);
1663
1664         trans_for_each_path_with_node(trans, b, linked, i)
1665                 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1666
1667         bch2_trans_verify_paths(trans);
1668 }
1669
1670 /**
1671  * bch2_btree_insert_node - insert bkeys into a given btree node
1672  *
1673  * @as:                 btree_update object
1674  * @trans:              btree_trans object
1675  * @path_idx:           path that points to current node
1676  * @b:                  node to insert keys into
1677  * @keys:               list of keys to insert
1678  *
1679  * Returns: 0 on success, typically transaction restart error on failure
1680  *
1681  * Inserts as many keys as it can into a given btree node, splitting it if full.
1682  * If a split occurred, this function will return early. This can only happen
1683  * for leaf nodes -- inserts into interior nodes have to be atomic.
1684  */
1685 static int bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans,
1686                                   btree_path_idx_t path_idx, struct btree *b,
1687                                   struct keylist *keys)
1688 {
1689         struct bch_fs *c = as->c;
1690         struct btree_path *path = trans->paths + path_idx;
1691         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1692         int old_live_u64s = b->nr.live_u64s;
1693         int live_u64s_added, u64s_added;
1694         int ret;
1695
1696         lockdep_assert_held(&c->gc_lock);
1697         BUG_ON(!btree_node_intent_locked(path, b->c.level));
1698         BUG_ON(!b->c.level);
1699         BUG_ON(!as || as->b);
1700         bch2_verify_keylist_sorted(keys);
1701
1702         ret = bch2_btree_node_lock_write(trans, path, &b->c);
1703         if (ret)
1704                 return ret;
1705
1706         bch2_btree_node_prep_for_write(trans, path, b);
1707
1708         if (!bch2_btree_node_insert_fits(b, bch2_keylist_u64s(keys))) {
1709                 bch2_btree_node_unlock_write(trans, path, b);
1710                 goto split;
1711         }
1712
1713         btree_node_interior_verify(c, b);
1714
1715         bch2_btree_insert_keys_interior(as, trans, path, b, keys);
1716
1717         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1718         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1719
1720         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1721                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1722         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1723                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1724
1725         if (u64s_added > live_u64s_added &&
1726             bch2_maybe_compact_whiteouts(c, b))
1727                 bch2_trans_node_reinit_iter(trans, b);
1728
1729         bch2_btree_node_unlock_write(trans, path, b);
1730
1731         btree_node_interior_verify(c, b);
1732         return 0;
1733 split:
1734         /*
1735          * We could attempt to avoid the transaction restart, by calling
1736          * bch2_btree_path_upgrade() and allocating more nodes:
1737          */
1738         if (b->c.level >= as->update_level) {
1739                 trace_and_count(c, trans_restart_split_race, trans, _THIS_IP_, b);
1740                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_split_race);
1741         }
1742
1743         return btree_split(as, trans, path_idx, b, keys);
1744 }
1745
1746 int bch2_btree_split_leaf(struct btree_trans *trans,
1747                           btree_path_idx_t path,
1748                           unsigned flags)
1749 {
1750         /* btree_split & merge may both cause paths array to be reallocated */
1751         struct btree *b = path_l(trans->paths + path)->b;
1752         struct btree_update *as;
1753         unsigned l;
1754         int ret = 0;
1755
1756         as = bch2_btree_update_start(trans, trans->paths + path,
1757                                      trans->paths[path].level,
1758                                      true, flags);
1759         if (IS_ERR(as))
1760                 return PTR_ERR(as);
1761
1762         ret = btree_split(as, trans, path, b, NULL);
1763         if (ret) {
1764                 bch2_btree_update_free(as, trans);
1765                 return ret;
1766         }
1767
1768         bch2_btree_update_done(as, trans);
1769
1770         for (l = trans->paths[path].level + 1;
1771              btree_node_intent_locked(&trans->paths[path], l) && !ret;
1772              l++)
1773                 ret = bch2_foreground_maybe_merge(trans, path, l, flags);
1774
1775         return ret;
1776 }
1777
1778 static void __btree_increase_depth(struct btree_update *as, struct btree_trans *trans,
1779                                    btree_path_idx_t path_idx)
1780 {
1781         struct bch_fs *c = as->c;
1782         struct btree_path *path = trans->paths + path_idx;
1783         struct btree *n, *b = bch2_btree_id_root(c, path->btree_id)->b;
1784
1785         BUG_ON(!btree_node_locked(path, b->c.level));
1786
1787         n = __btree_root_alloc(as, trans, b->c.level + 1);
1788
1789         bch2_btree_update_add_new_node(as, n);
1790         six_unlock_write(&n->c.lock);
1791
1792         path->locks_want++;
1793         BUG_ON(btree_node_locked(path, n->c.level));
1794         six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1795         mark_btree_node_locked(trans, path, n->c.level, BTREE_NODE_INTENT_LOCKED);
1796         bch2_btree_path_level_init(trans, path, n);
1797
1798         n->sib_u64s[0] = U16_MAX;
1799         n->sib_u64s[1] = U16_MAX;
1800
1801         bch2_keylist_add(&as->parent_keys, &b->key);
1802         btree_split_insert_keys(as, trans, path_idx, n, &as->parent_keys);
1803
1804         bch2_btree_set_root(as, trans, path, n);
1805         bch2_btree_update_get_open_buckets(as, n);
1806         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1807         bch2_trans_node_add(trans, path, n);
1808         six_unlock_intent(&n->c.lock);
1809
1810         mutex_lock(&c->btree_cache.lock);
1811         list_add_tail(&b->list, &c->btree_cache.live);
1812         mutex_unlock(&c->btree_cache.lock);
1813
1814         bch2_trans_verify_locks(trans);
1815 }
1816
1817 int bch2_btree_increase_depth(struct btree_trans *trans, btree_path_idx_t path, unsigned flags)
1818 {
1819         struct bch_fs *c = trans->c;
1820         struct btree *b = bch2_btree_id_root(c, trans->paths[path].btree_id)->b;
1821         struct btree_update *as =
1822                 bch2_btree_update_start(trans, trans->paths + path,
1823                                         b->c.level, true, flags);
1824         if (IS_ERR(as))
1825                 return PTR_ERR(as);
1826
1827         __btree_increase_depth(as, trans, path);
1828         bch2_btree_update_done(as, trans);
1829         return 0;
1830 }
1831
1832 int __bch2_foreground_maybe_merge(struct btree_trans *trans,
1833                                   btree_path_idx_t path,
1834                                   unsigned level,
1835                                   unsigned flags,
1836                                   enum btree_node_sibling sib)
1837 {
1838         struct bch_fs *c = trans->c;
1839         struct btree_update *as;
1840         struct bkey_format_state new_s;
1841         struct bkey_format new_f;
1842         struct bkey_i delete;
1843         struct btree *b, *m, *n, *prev, *next, *parent;
1844         struct bpos sib_pos;
1845         size_t sib_u64s;
1846         enum btree_id btree = trans->paths[path].btree_id;
1847         btree_path_idx_t sib_path = 0, new_path = 0;
1848         u64 start_time = local_clock();
1849         int ret = 0;
1850
1851         BUG_ON(!trans->paths[path].should_be_locked);
1852         BUG_ON(!btree_node_locked(&trans->paths[path], level));
1853
1854         b = trans->paths[path].l[level].b;
1855
1856         if ((sib == btree_prev_sib && bpos_eq(b->data->min_key, POS_MIN)) ||
1857             (sib == btree_next_sib && bpos_eq(b->data->max_key, SPOS_MAX))) {
1858                 b->sib_u64s[sib] = U16_MAX;
1859                 return 0;
1860         }
1861
1862         sib_pos = sib == btree_prev_sib
1863                 ? bpos_predecessor(b->data->min_key)
1864                 : bpos_successor(b->data->max_key);
1865
1866         sib_path = bch2_path_get(trans, btree, sib_pos,
1867                                  U8_MAX, level, BTREE_ITER_INTENT, _THIS_IP_);
1868         ret = bch2_btree_path_traverse(trans, sib_path, false);
1869         if (ret)
1870                 goto err;
1871
1872         btree_path_set_should_be_locked(trans->paths + sib_path);
1873
1874         m = trans->paths[sib_path].l[level].b;
1875
1876         if (btree_node_parent(trans->paths + path, b) !=
1877             btree_node_parent(trans->paths + sib_path, m)) {
1878                 b->sib_u64s[sib] = U16_MAX;
1879                 goto out;
1880         }
1881
1882         if (sib == btree_prev_sib) {
1883                 prev = m;
1884                 next = b;
1885         } else {
1886                 prev = b;
1887                 next = m;
1888         }
1889
1890         if (!bpos_eq(bpos_successor(prev->data->max_key), next->data->min_key)) {
1891                 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1892
1893                 bch2_bpos_to_text(&buf1, prev->data->max_key);
1894                 bch2_bpos_to_text(&buf2, next->data->min_key);
1895                 bch_err(c,
1896                         "%s(): btree topology error:\n"
1897                         "  prev ends at   %s\n"
1898                         "  next starts at %s",
1899                         __func__, buf1.buf, buf2.buf);
1900                 printbuf_exit(&buf1);
1901                 printbuf_exit(&buf2);
1902                 ret = bch2_topology_error(c);
1903                 goto err;
1904         }
1905
1906         bch2_bkey_format_init(&new_s);
1907         bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
1908         __bch2_btree_calc_format(&new_s, prev);
1909         __bch2_btree_calc_format(&new_s, next);
1910         bch2_bkey_format_add_pos(&new_s, next->data->max_key);
1911         new_f = bch2_bkey_format_done(&new_s);
1912
1913         sib_u64s = btree_node_u64s_with_format(b->nr, &b->format, &new_f) +
1914                 btree_node_u64s_with_format(m->nr, &m->format, &new_f);
1915
1916         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1917                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1918                 sib_u64s /= 2;
1919                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1920         }
1921
1922         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1923         sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
1924         b->sib_u64s[sib] = sib_u64s;
1925
1926         if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
1927                 goto out;
1928
1929         parent = btree_node_parent(trans->paths + path, b);
1930         as = bch2_btree_update_start(trans, trans->paths + path, level, false,
1931                                      BCH_TRANS_COMMIT_no_enospc|flags);
1932         ret = PTR_ERR_OR_ZERO(as);
1933         if (ret)
1934                 goto err;
1935
1936         trace_and_count(c, btree_node_merge, trans, b);
1937
1938         bch2_btree_interior_update_will_free_node(as, b);
1939         bch2_btree_interior_update_will_free_node(as, m);
1940
1941         n = bch2_btree_node_alloc(as, trans, b->c.level);
1942
1943         SET_BTREE_NODE_SEQ(n->data,
1944                            max(BTREE_NODE_SEQ(b->data),
1945                                BTREE_NODE_SEQ(m->data)) + 1);
1946
1947         btree_set_min(n, prev->data->min_key);
1948         btree_set_max(n, next->data->max_key);
1949
1950         n->data->format  = new_f;
1951         btree_node_set_format(n, new_f);
1952
1953         bch2_btree_sort_into(c, n, prev);
1954         bch2_btree_sort_into(c, n, next);
1955
1956         bch2_btree_build_aux_trees(n);
1957         bch2_btree_update_add_new_node(as, n);
1958         six_unlock_write(&n->c.lock);
1959
1960         new_path = get_unlocked_mut_path(trans, btree, n->c.level, n->key.k.p);
1961         six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1962         mark_btree_node_locked(trans, trans->paths + new_path, n->c.level, BTREE_NODE_INTENT_LOCKED);
1963         bch2_btree_path_level_init(trans, trans->paths + new_path, n);
1964
1965         bkey_init(&delete.k);
1966         delete.k.p = prev->key.k.p;
1967         bch2_keylist_add(&as->parent_keys, &delete);
1968         bch2_keylist_add(&as->parent_keys, &n->key);
1969
1970         bch2_trans_verify_paths(trans);
1971
1972         ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys);
1973         if (ret)
1974                 goto err_free_update;
1975
1976         bch2_trans_verify_paths(trans);
1977
1978         bch2_btree_update_get_open_buckets(as, n);
1979         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1980
1981         bch2_btree_node_free_inmem(trans, trans->paths + path, b);
1982         bch2_btree_node_free_inmem(trans, trans->paths + sib_path, m);
1983
1984         bch2_trans_node_add(trans, trans->paths + path, n);
1985
1986         bch2_trans_verify_paths(trans);
1987
1988         six_unlock_intent(&n->c.lock);
1989
1990         bch2_btree_update_done(as, trans);
1991
1992         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time);
1993 out:
1994 err:
1995         if (new_path)
1996                 bch2_path_put(trans, new_path, true);
1997         bch2_path_put(trans, sib_path, true);
1998         bch2_trans_verify_locks(trans);
1999         return ret;
2000 err_free_update:
2001         bch2_btree_node_free_never_used(as, trans, n);
2002         bch2_btree_update_free(as, trans);
2003         goto out;
2004 }
2005
2006 int bch2_btree_node_rewrite(struct btree_trans *trans,
2007                             struct btree_iter *iter,
2008                             struct btree *b,
2009                             unsigned flags)
2010 {
2011         struct bch_fs *c = trans->c;
2012         struct btree *n, *parent;
2013         struct btree_update *as;
2014         btree_path_idx_t new_path = 0;
2015         int ret;
2016
2017         flags |= BCH_TRANS_COMMIT_no_enospc;
2018
2019         struct btree_path *path = btree_iter_path(trans, iter);
2020         parent = btree_node_parent(path, b);
2021         as = bch2_btree_update_start(trans, path, b->c.level, false, flags);
2022         ret = PTR_ERR_OR_ZERO(as);
2023         if (ret)
2024                 goto out;
2025
2026         bch2_btree_interior_update_will_free_node(as, b);
2027
2028         n = bch2_btree_node_alloc_replacement(as, trans, b);
2029
2030         bch2_btree_build_aux_trees(n);
2031         bch2_btree_update_add_new_node(as, n);
2032         six_unlock_write(&n->c.lock);
2033
2034         new_path = get_unlocked_mut_path(trans, iter->btree_id, n->c.level, n->key.k.p);
2035         six_lock_increment(&n->c.lock, SIX_LOCK_intent);
2036         mark_btree_node_locked(trans, trans->paths + new_path, n->c.level, BTREE_NODE_INTENT_LOCKED);
2037         bch2_btree_path_level_init(trans, trans->paths + new_path, n);
2038
2039         trace_and_count(c, btree_node_rewrite, trans, b);
2040
2041         if (parent) {
2042                 bch2_keylist_add(&as->parent_keys, &n->key);
2043                 ret = bch2_btree_insert_node(as, trans, iter->path, parent, &as->parent_keys);
2044                 if (ret)
2045                         goto err;
2046         } else {
2047                 bch2_btree_set_root(as, trans, btree_iter_path(trans, iter), n);
2048         }
2049
2050         bch2_btree_update_get_open_buckets(as, n);
2051         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
2052
2053         bch2_btree_node_free_inmem(trans, btree_iter_path(trans, iter), b);
2054
2055         bch2_trans_node_add(trans, trans->paths + iter->path, n);
2056         six_unlock_intent(&n->c.lock);
2057
2058         bch2_btree_update_done(as, trans);
2059 out:
2060         if (new_path)
2061                 bch2_path_put(trans, new_path, true);
2062         bch2_trans_downgrade(trans);
2063         return ret;
2064 err:
2065         bch2_btree_node_free_never_used(as, trans, n);
2066         bch2_btree_update_free(as, trans);
2067         goto out;
2068 }
2069
2070 struct async_btree_rewrite {
2071         struct bch_fs           *c;
2072         struct work_struct      work;
2073         struct list_head        list;
2074         enum btree_id           btree_id;
2075         unsigned                level;
2076         struct bpos             pos;
2077         __le64                  seq;
2078 };
2079
2080 static int async_btree_node_rewrite_trans(struct btree_trans *trans,
2081                                           struct async_btree_rewrite *a)
2082 {
2083         struct bch_fs *c = trans->c;
2084         struct btree_iter iter;
2085         struct btree *b;
2086         int ret;
2087
2088         bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos,
2089                                   BTREE_MAX_DEPTH, a->level, 0);
2090         b = bch2_btree_iter_peek_node(&iter);
2091         ret = PTR_ERR_OR_ZERO(b);
2092         if (ret)
2093                 goto out;
2094
2095         if (!b || b->data->keys.seq != a->seq) {
2096                 struct printbuf buf = PRINTBUF;
2097
2098                 if (b)
2099                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
2100                 else
2101                         prt_str(&buf, "(null");
2102                 bch_info(c, "%s: node to rewrite not found:, searching for seq %llu, got\n%s",
2103                          __func__, a->seq, buf.buf);
2104                 printbuf_exit(&buf);
2105                 goto out;
2106         }
2107
2108         ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
2109 out:
2110         bch2_trans_iter_exit(trans, &iter);
2111
2112         return ret;
2113 }
2114
2115 static void async_btree_node_rewrite_work(struct work_struct *work)
2116 {
2117         struct async_btree_rewrite *a =
2118                 container_of(work, struct async_btree_rewrite, work);
2119         struct bch_fs *c = a->c;
2120         int ret;
2121
2122         ret = bch2_trans_do(c, NULL, NULL, 0,
2123                       async_btree_node_rewrite_trans(trans, a));
2124         bch_err_fn_ratelimited(c, ret);
2125         bch2_write_ref_put(c, BCH_WRITE_REF_node_rewrite);
2126         kfree(a);
2127 }
2128
2129 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
2130 {
2131         struct async_btree_rewrite *a;
2132         int ret;
2133
2134         a = kmalloc(sizeof(*a), GFP_NOFS);
2135         if (!a) {
2136                 bch_err(c, "%s: error allocating memory", __func__);
2137                 return;
2138         }
2139
2140         a->c            = c;
2141         a->btree_id     = b->c.btree_id;
2142         a->level        = b->c.level;
2143         a->pos          = b->key.k.p;
2144         a->seq          = b->data->keys.seq;
2145         INIT_WORK(&a->work, async_btree_node_rewrite_work);
2146
2147         if (unlikely(!test_bit(BCH_FS_may_go_rw, &c->flags))) {
2148                 mutex_lock(&c->pending_node_rewrites_lock);
2149                 list_add(&a->list, &c->pending_node_rewrites);
2150                 mutex_unlock(&c->pending_node_rewrites_lock);
2151                 return;
2152         }
2153
2154         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_node_rewrite)) {
2155                 if (test_bit(BCH_FS_started, &c->flags)) {
2156                         bch_err(c, "%s: error getting c->writes ref", __func__);
2157                         kfree(a);
2158                         return;
2159                 }
2160
2161                 ret = bch2_fs_read_write_early(c);
2162                 bch_err_msg(c, ret, "going read-write");
2163                 if (ret) {
2164                         kfree(a);
2165                         return;
2166                 }
2167
2168                 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2169         }
2170
2171         queue_work(c->btree_node_rewrite_worker, &a->work);
2172 }
2173
2174 void bch2_do_pending_node_rewrites(struct bch_fs *c)
2175 {
2176         struct async_btree_rewrite *a, *n;
2177
2178         mutex_lock(&c->pending_node_rewrites_lock);
2179         list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2180                 list_del(&a->list);
2181
2182                 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2183                 queue_work(c->btree_node_rewrite_worker, &a->work);
2184         }
2185         mutex_unlock(&c->pending_node_rewrites_lock);
2186 }
2187
2188 void bch2_free_pending_node_rewrites(struct bch_fs *c)
2189 {
2190         struct async_btree_rewrite *a, *n;
2191
2192         mutex_lock(&c->pending_node_rewrites_lock);
2193         list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2194                 list_del(&a->list);
2195
2196                 kfree(a);
2197         }
2198         mutex_unlock(&c->pending_node_rewrites_lock);
2199 }
2200
2201 static int __bch2_btree_node_update_key(struct btree_trans *trans,
2202                                         struct btree_iter *iter,
2203                                         struct btree *b, struct btree *new_hash,
2204                                         struct bkey_i *new_key,
2205                                         unsigned commit_flags,
2206                                         bool skip_triggers)
2207 {
2208         struct bch_fs *c = trans->c;
2209         struct btree_iter iter2 = { NULL };
2210         struct btree *parent;
2211         int ret;
2212
2213         if (!skip_triggers) {
2214                 ret   = bch2_key_trigger_old(trans, b->c.btree_id, b->c.level + 1,
2215                                              bkey_i_to_s_c(&b->key),
2216                                              BTREE_TRIGGER_TRANSACTIONAL) ?:
2217                         bch2_key_trigger_new(trans, b->c.btree_id, b->c.level + 1,
2218                                              bkey_i_to_s(new_key),
2219                                              BTREE_TRIGGER_TRANSACTIONAL);
2220                 if (ret)
2221                         return ret;
2222         }
2223
2224         if (new_hash) {
2225                 bkey_copy(&new_hash->key, new_key);
2226                 ret = bch2_btree_node_hash_insert(&c->btree_cache,
2227                                 new_hash, b->c.level, b->c.btree_id);
2228                 BUG_ON(ret);
2229         }
2230
2231         parent = btree_node_parent(btree_iter_path(trans, iter), b);
2232         if (parent) {
2233                 bch2_trans_copy_iter(&iter2, iter);
2234
2235                 iter2.path = bch2_btree_path_make_mut(trans, iter2.path,
2236                                 iter2.flags & BTREE_ITER_INTENT,
2237                                 _THIS_IP_);
2238
2239                 struct btree_path *path2 = btree_iter_path(trans, &iter2);
2240                 BUG_ON(path2->level != b->c.level);
2241                 BUG_ON(!bpos_eq(path2->pos, new_key->k.p));
2242
2243                 btree_path_set_level_up(trans, path2);
2244
2245                 trans->paths_sorted = false;
2246
2247                 ret   = bch2_btree_iter_traverse(&iter2) ?:
2248                         bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_NORUN);
2249                 if (ret)
2250                         goto err;
2251         } else {
2252                 BUG_ON(btree_node_root(c, b) != b);
2253
2254                 struct jset_entry *e = bch2_trans_jset_entry_alloc(trans,
2255                                        jset_u64s(new_key->k.u64s));
2256                 ret = PTR_ERR_OR_ZERO(e);
2257                 if (ret)
2258                         return ret;
2259
2260                 journal_entry_set(e,
2261                                   BCH_JSET_ENTRY_btree_root,
2262                                   b->c.btree_id, b->c.level,
2263                                   new_key, new_key->k.u64s);
2264         }
2265
2266         ret = bch2_trans_commit(trans, NULL, NULL, commit_flags);
2267         if (ret)
2268                 goto err;
2269
2270         bch2_btree_node_lock_write_nofail(trans, btree_iter_path(trans, iter), &b->c);
2271
2272         if (new_hash) {
2273                 mutex_lock(&c->btree_cache.lock);
2274                 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
2275                 bch2_btree_node_hash_remove(&c->btree_cache, b);
2276
2277                 bkey_copy(&b->key, new_key);
2278                 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
2279                 BUG_ON(ret);
2280                 mutex_unlock(&c->btree_cache.lock);
2281         } else {
2282                 bkey_copy(&b->key, new_key);
2283         }
2284
2285         bch2_btree_node_unlock_write(trans, btree_iter_path(trans, iter), b);
2286 out:
2287         bch2_trans_iter_exit(trans, &iter2);
2288         return ret;
2289 err:
2290         if (new_hash) {
2291                 mutex_lock(&c->btree_cache.lock);
2292                 bch2_btree_node_hash_remove(&c->btree_cache, b);
2293                 mutex_unlock(&c->btree_cache.lock);
2294         }
2295         goto out;
2296 }
2297
2298 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
2299                                struct btree *b, struct bkey_i *new_key,
2300                                unsigned commit_flags, bool skip_triggers)
2301 {
2302         struct bch_fs *c = trans->c;
2303         struct btree *new_hash = NULL;
2304         struct btree_path *path = btree_iter_path(trans, iter);
2305         struct closure cl;
2306         int ret = 0;
2307
2308         ret = bch2_btree_path_upgrade(trans, path, b->c.level + 1);
2309         if (ret)
2310                 return ret;
2311
2312         closure_init_stack(&cl);
2313
2314         /*
2315          * check btree_ptr_hash_val() after @b is locked by
2316          * btree_iter_traverse():
2317          */
2318         if (btree_ptr_hash_val(new_key) != b->hash_val) {
2319                 ret = bch2_btree_cache_cannibalize_lock(trans, &cl);
2320                 if (ret) {
2321                         ret = drop_locks_do(trans, (closure_sync(&cl), 0));
2322                         if (ret)
2323                                 return ret;
2324                 }
2325
2326                 new_hash = bch2_btree_node_mem_alloc(trans, false);
2327         }
2328
2329         path->intent_ref++;
2330         ret = __bch2_btree_node_update_key(trans, iter, b, new_hash, new_key,
2331                                            commit_flags, skip_triggers);
2332         --path->intent_ref;
2333
2334         if (new_hash) {
2335                 mutex_lock(&c->btree_cache.lock);
2336                 list_move(&new_hash->list, &c->btree_cache.freeable);
2337                 mutex_unlock(&c->btree_cache.lock);
2338
2339                 six_unlock_write(&new_hash->c.lock);
2340                 six_unlock_intent(&new_hash->c.lock);
2341         }
2342         closure_sync(&cl);
2343         bch2_btree_cache_cannibalize_unlock(trans);
2344         return ret;
2345 }
2346
2347 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
2348                                         struct btree *b, struct bkey_i *new_key,
2349                                         unsigned commit_flags, bool skip_triggers)
2350 {
2351         struct btree_iter iter;
2352         int ret;
2353
2354         bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p,
2355                                   BTREE_MAX_DEPTH, b->c.level,
2356                                   BTREE_ITER_INTENT);
2357         ret = bch2_btree_iter_traverse(&iter);
2358         if (ret)
2359                 goto out;
2360
2361         /* has node been freed? */
2362         if (btree_iter_path(trans, &iter)->l[b->c.level].b != b) {
2363                 /* node has been freed: */
2364                 BUG_ON(!btree_node_dying(b));
2365                 goto out;
2366         }
2367
2368         BUG_ON(!btree_node_hashed(b));
2369
2370         struct bch_extent_ptr *ptr;
2371         bch2_bkey_drop_ptrs(bkey_i_to_s(new_key), ptr,
2372                             !bch2_bkey_has_device(bkey_i_to_s(&b->key), ptr->dev));
2373
2374         ret = bch2_btree_node_update_key(trans, &iter, b, new_key,
2375                                          commit_flags, skip_triggers);
2376 out:
2377         bch2_trans_iter_exit(trans, &iter);
2378         return ret;
2379 }
2380
2381 /* Init code: */
2382
2383 /*
2384  * Only for filesystem bringup, when first reading the btree roots or allocating
2385  * btree roots when initializing a new filesystem:
2386  */
2387 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2388 {
2389         BUG_ON(btree_node_root(c, b));
2390
2391         bch2_btree_set_root_inmem(c, b);
2392 }
2393
2394 static int __bch2_btree_root_alloc(struct btree_trans *trans, enum btree_id id)
2395 {
2396         struct bch_fs *c = trans->c;
2397         struct closure cl;
2398         struct btree *b;
2399         int ret;
2400
2401         closure_init_stack(&cl);
2402
2403         do {
2404                 ret = bch2_btree_cache_cannibalize_lock(trans, &cl);
2405                 closure_sync(&cl);
2406         } while (ret);
2407
2408         b = bch2_btree_node_mem_alloc(trans, false);
2409         bch2_btree_cache_cannibalize_unlock(trans);
2410
2411         set_btree_node_fake(b);
2412         set_btree_node_need_rewrite(b);
2413         b->c.level      = 0;
2414         b->c.btree_id   = id;
2415
2416         bkey_btree_ptr_init(&b->key);
2417         b->key.k.p = SPOS_MAX;
2418         *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2419
2420         bch2_bset_init_first(b, &b->data->keys);
2421         bch2_btree_build_aux_trees(b);
2422
2423         b->data->flags = 0;
2424         btree_set_min(b, POS_MIN);
2425         btree_set_max(b, SPOS_MAX);
2426         b->data->format = bch2_btree_calc_format(b);
2427         btree_node_set_format(b, b->data->format);
2428
2429         ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2430                                           b->c.level, b->c.btree_id);
2431         BUG_ON(ret);
2432
2433         bch2_btree_set_root_inmem(c, b);
2434
2435         six_unlock_write(&b->c.lock);
2436         six_unlock_intent(&b->c.lock);
2437         return 0;
2438 }
2439
2440 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2441 {
2442         bch2_trans_run(c, __bch2_btree_root_alloc(trans, id));
2443 }
2444
2445 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2446 {
2447         struct btree_update *as;
2448
2449         mutex_lock(&c->btree_interior_update_lock);
2450         list_for_each_entry(as, &c->btree_interior_update_list, list)
2451                 prt_printf(out, "%ps: mode=%u nodes_written=%u cl.remaining=%u journal_seq=%llu\n",
2452                            (void *) as->ip_started,
2453                            as->mode,
2454                            as->nodes_written,
2455                            closure_nr_remaining(&as->cl),
2456                            as->journal.seq);
2457         mutex_unlock(&c->btree_interior_update_lock);
2458 }
2459
2460 static bool bch2_btree_interior_updates_pending(struct bch_fs *c)
2461 {
2462         bool ret;
2463
2464         mutex_lock(&c->btree_interior_update_lock);
2465         ret = !list_empty(&c->btree_interior_update_list);
2466         mutex_unlock(&c->btree_interior_update_lock);
2467
2468         return ret;
2469 }
2470
2471 bool bch2_btree_interior_updates_flush(struct bch_fs *c)
2472 {
2473         bool ret = bch2_btree_interior_updates_pending(c);
2474
2475         if (ret)
2476                 closure_wait_event(&c->btree_interior_update_wait,
2477                                    !bch2_btree_interior_updates_pending(c));
2478         return ret;
2479 }
2480
2481 void bch2_journal_entry_to_btree_root(struct bch_fs *c, struct jset_entry *entry)
2482 {
2483         struct btree_root *r = bch2_btree_id_root(c, entry->btree_id);
2484
2485         mutex_lock(&c->btree_root_lock);
2486
2487         r->level = entry->level;
2488         r->alive = true;
2489         bkey_copy(&r->key, (struct bkey_i *) entry->start);
2490
2491         mutex_unlock(&c->btree_root_lock);
2492 }
2493
2494 struct jset_entry *
2495 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2496                                     struct jset_entry *end,
2497                                     unsigned long skip)
2498 {
2499         unsigned i;
2500
2501         mutex_lock(&c->btree_root_lock);
2502
2503         for (i = 0; i < btree_id_nr_alive(c); i++) {
2504                 struct btree_root *r = bch2_btree_id_root(c, i);
2505
2506                 if (r->alive && !test_bit(i, &skip)) {
2507                         journal_entry_set(end, BCH_JSET_ENTRY_btree_root,
2508                                           i, r->level, &r->key, r->key.k.u64s);
2509                         end = vstruct_next(end);
2510                 }
2511         }
2512
2513         mutex_unlock(&c->btree_root_lock);
2514
2515         return end;
2516 }
2517
2518 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2519 {
2520         if (c->btree_node_rewrite_worker)
2521                 destroy_workqueue(c->btree_node_rewrite_worker);
2522         if (c->btree_interior_update_worker)
2523                 destroy_workqueue(c->btree_interior_update_worker);
2524         mempool_exit(&c->btree_interior_update_pool);
2525 }
2526
2527 void bch2_fs_btree_interior_update_init_early(struct bch_fs *c)
2528 {
2529         mutex_init(&c->btree_reserve_cache_lock);
2530         INIT_LIST_HEAD(&c->btree_interior_update_list);
2531         INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2532         mutex_init(&c->btree_interior_update_lock);
2533         INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2534
2535         INIT_LIST_HEAD(&c->pending_node_rewrites);
2536         mutex_init(&c->pending_node_rewrites_lock);
2537 }
2538
2539 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2540 {
2541         c->btree_interior_update_worker =
2542                 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 8);
2543         if (!c->btree_interior_update_worker)
2544                 return -BCH_ERR_ENOMEM_btree_interior_update_worker_init;
2545
2546         c->btree_node_rewrite_worker =
2547                 alloc_ordered_workqueue("btree_node_rewrite", WQ_UNBOUND);
2548         if (!c->btree_node_rewrite_worker)
2549                 return -BCH_ERR_ENOMEM_btree_interior_update_worker_init;
2550
2551         if (mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2552                                       sizeof(struct btree_update)))
2553                 return -BCH_ERR_ENOMEM_btree_interior_update_pool_init;
2554
2555         return 0;
2556 }