git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/btree_update_interior.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "alloc_foreground.h"
   5 #include "bkey_methods.h"
   6 #include "btree_cache.h"
   7 #include "btree_gc.h"
   8 #include "btree_journal_iter.h"
   9 #include "btree_update.h"
  10 #include "btree_update_interior.h"
  11 #include "btree_io.h"
  12 #include "btree_iter.h"
  13 #include "btree_locking.h"
  14 #include "buckets.h"
  15 #include "clock.h"
  16 #include "error.h"
  17 #include "extents.h"
  18 #include "journal.h"
  19 #include "journal_reclaim.h"
  20 #include "keylist.h"
  21 #include "replicas.h"
  22 #include "super-io.h"
  23 #include "trace.h"
  24
  25 #include <linux/random.h>
  26
  27 static int bch2_btree_insert_node(struct btree_update *, struct btree_trans *,
  28                                   struct btree_path *, struct btree *,
  29                                   struct keylist *, unsigned);
  30 static void bch2_btree_update_add_new_node(struct btree_update *, struct btree *);
  31
  32 static struct btree_path *get_unlocked_mut_path(struct btree_trans *trans,
  33                                                 enum btree_id btree_id,
  34                                                 unsigned level,
  35                                                 struct bpos pos)
  36 {
  37         struct btree_path *path;
  38
  39         path = bch2_path_get(trans, btree_id, pos, level + 1, level,
  40                              BTREE_ITER_NOPRESERVE|
  41                              BTREE_ITER_INTENT, _RET_IP_);
  42         path = bch2_btree_path_make_mut(trans, path, true, _RET_IP_);
  43         bch2_btree_path_downgrade(trans, path);
  44         __bch2_btree_path_unlock(trans, path);
  45         return path;
  46 }
  47
  48 /* Debug code: */
  49
  50 /*
  51  * Verify that child nodes correctly span parent node's range:
  52  */
  53 static void btree_node_interior_verify(struct bch_fs *c, struct btree *b)
  54 {
  55 #ifdef CONFIG_BCACHEFS_DEBUG
  56         struct bpos next_node = b->data->min_key;
  57         struct btree_node_iter iter;
  58         struct bkey_s_c k;
  59         struct bkey_s_c_btree_ptr_v2 bp;
  60         struct bkey unpacked;
  61         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
  62
  63         BUG_ON(!b->c.level);
  64
  65         if (!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))
  66                 return;
  67
  68         bch2_btree_node_iter_init_from_start(&iter, b);
  69
  70         while (1) {
  71                 k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked);
  72                 if (k.k->type != KEY_TYPE_btree_ptr_v2)
  73                         break;
  74                 bp = bkey_s_c_to_btree_ptr_v2(k);
  75
  76                 if (!bpos_eq(next_node, bp.v->min_key)) {
  77                         bch2_dump_btree_node(c, b);
  78                         bch2_bpos_to_text(&buf1, next_node);
  79                         bch2_bpos_to_text(&buf2, bp.v->min_key);
  80                         panic("expected next min_key %s got %s\n", buf1.buf, buf2.buf);
  81                 }
  82
  83                 bch2_btree_node_iter_advance(&iter, b);
  84
  85                 if (bch2_btree_node_iter_end(&iter)) {
  86                         if (!bpos_eq(k.k->p, b->key.k.p)) {
  87                                 bch2_dump_btree_node(c, b);
  88                                 bch2_bpos_to_text(&buf1, b->key.k.p);
  89                                 bch2_bpos_to_text(&buf2, k.k->p);
  90                                 panic("expected end %s got %s\n", buf1.buf, buf2.buf);
  91                         }
  92                         break;
  93                 }
  94
  95                 next_node = bpos_successor(k.k->p);
  96         }
  97 #endif
  98 }
  99
 100 /* Calculate ideal packed bkey format for new btree nodes: */
 101
 102 void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b)
 103 {
 104         struct bkey_packed *k;
 105         struct bset_tree *t;
 106         struct bkey uk;
 107
 108         for_each_bset(b, t)
 109                 bset_tree_for_each_key(b, t, k)
 110                         if (!bkey_deleted(k)) {
 111                                 uk = bkey_unpack_key(b, k);
 112                                 bch2_bkey_format_add_key(s, &uk);
 113                         }
 114 }
 115
 116 static struct bkey_format bch2_btree_calc_format(struct btree *b)
 117 {
 118         struct bkey_format_state s;
 119
 120         bch2_bkey_format_init(&s);
 121         bch2_bkey_format_add_pos(&s, b->data->min_key);
 122         bch2_bkey_format_add_pos(&s, b->data->max_key);
 123         __bch2_btree_calc_format(&s, b);
 124
 125         return bch2_bkey_format_done(&s);
 126 }
 127
 128 static size_t btree_node_u64s_with_format(struct btree *b,
 129                                           struct bkey_format *new_f)
 130 {
 131         struct bkey_format *old_f = &b->format;
 132
 133         /* stupid integer promotion rules */
 134         ssize_t delta =
 135             (((int) new_f->key_u64s - old_f->key_u64s) *
 136              (int) b->nr.packed_keys) +
 137             (((int) new_f->key_u64s - BKEY_U64s) *
 138              (int) b->nr.unpacked_keys);
 139
 140         BUG_ON(delta + b->nr.live_u64s < 0);
 141
 142         return b->nr.live_u64s + delta;
 143 }
 144
 145 /**
 146  * bch2_btree_node_format_fits - check if we could rewrite node with a new format
 147  *
 148  * @c:          filesystem handle
 149  * @b:          btree node to rewrite
 150  * @new_f:      bkey format to translate keys to
 151  *
 152  * Returns: true if all re-packed keys will be able to fit in a new node.
 153  *
 154  * Assumes all keys will successfully pack with the new format.
 155  */
 156 bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
 157                                  struct bkey_format *new_f)
 158 {
 159         size_t u64s = btree_node_u64s_with_format(b, new_f);
 160
 161         return __vstruct_bytes(struct btree_node, u64s) < btree_bytes(c);
 162 }
 163
 164 /* Btree node freeing/allocation: */
 165
 166 static void __btree_node_free(struct bch_fs *c, struct btree *b)
 167 {
 168         trace_and_count(c, btree_node_free, c, b);
 169
 170         BUG_ON(btree_node_write_blocked(b));
 171         BUG_ON(btree_node_dirty(b));
 172         BUG_ON(btree_node_need_write(b));
 173         BUG_ON(b == btree_node_root(c, b));
 174         BUG_ON(b->ob.nr);
 175         BUG_ON(!list_empty(&b->write_blocked));
 176         BUG_ON(b->will_make_reachable);
 177
 178         clear_btree_node_noevict(b);
 179
 180         mutex_lock(&c->btree_cache.lock);
 181         list_move(&b->list, &c->btree_cache.freeable);
 182         mutex_unlock(&c->btree_cache.lock);
 183 }
 184
 185 static void bch2_btree_node_free_inmem(struct btree_trans *trans,
 186                                        struct btree_path *path,
 187                                        struct btree *b)
 188 {
 189         struct bch_fs *c = trans->c;
 190         unsigned level = b->c.level;
 191
 192         bch2_btree_node_lock_write_nofail(trans, path, &b->c);
 193         bch2_btree_node_hash_remove(&c->btree_cache, b);
 194         __btree_node_free(c, b);
 195         six_unlock_write(&b->c.lock);
 196         mark_btree_node_locked_noreset(path, level, BTREE_NODE_INTENT_LOCKED);
 197
 198         trans_for_each_path(trans, path)
 199                 if (path->l[level].b == b) {
 200                         btree_node_unlock(trans, path, level);
 201                         path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
 202                 }
 203 }
 204
 205 static void bch2_btree_node_free_never_used(struct btree_update *as,
 206                                             struct btree_trans *trans,
 207                                             struct btree *b)
 208 {
 209         struct bch_fs *c = as->c;
 210         struct prealloc_nodes *p = &as->prealloc_nodes[b->c.lock.readers != NULL];
 211         struct btree_path *path;
 212         unsigned level = b->c.level;
 213
 214         BUG_ON(!list_empty(&b->write_blocked));
 215         BUG_ON(b->will_make_reachable != (1UL|(unsigned long) as));
 216
 217         b->will_make_reachable = 0;
 218         closure_put(&as->cl);
 219
 220         clear_btree_node_will_make_reachable(b);
 221         clear_btree_node_accessed(b);
 222         clear_btree_node_dirty_acct(c, b);
 223         clear_btree_node_need_write(b);
 224
 225         mutex_lock(&c->btree_cache.lock);
 226         list_del_init(&b->list);
 227         bch2_btree_node_hash_remove(&c->btree_cache, b);
 228         mutex_unlock(&c->btree_cache.lock);
 229
 230         BUG_ON(p->nr >= ARRAY_SIZE(p->b));
 231         p->b[p->nr++] = b;
 232
 233         six_unlock_intent(&b->c.lock);
 234
 235         trans_for_each_path(trans, path)
 236                 if (path->l[level].b == b) {
 237                         btree_node_unlock(trans, path, level);
 238                         path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
 239                 }
 240 }
 241
 242 static struct btree *__bch2_btree_node_alloc(struct btree_trans *trans,
 243                                              struct disk_reservation *res,
 244                                              struct closure *cl,
 245                                              bool interior_node,
 246                                              unsigned flags)
 247 {
 248         struct bch_fs *c = trans->c;
 249         struct write_point *wp;
 250         struct btree *b;
 251         BKEY_PADDED_ONSTACK(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
 252         struct open_buckets obs = { .nr = 0 };
 253         struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
 254         enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
 255         unsigned nr_reserve = watermark > BCH_WATERMARK_reclaim
 256                 ? BTREE_NODE_RESERVE
 257                 : 0;
 258         int ret;
 259
 260         mutex_lock(&c->btree_reserve_cache_lock);
 261         if (c->btree_reserve_cache_nr > nr_reserve) {
 262                 struct btree_alloc *a =
 263                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
 264
 265                 obs = a->ob;
 266                 bkey_copy(&tmp.k, &a->k);
 267                 mutex_unlock(&c->btree_reserve_cache_lock);
 268                 goto mem_alloc;
 269         }
 270         mutex_unlock(&c->btree_reserve_cache_lock);
 271
 272 retry:
 273         ret = bch2_alloc_sectors_start_trans(trans,
 274                                       c->opts.metadata_target ?:
 275                                       c->opts.foreground_target,
 276                                       0,
 277                                       writepoint_ptr(&c->btree_write_point),
 278                                       &devs_have,
 279                                       res->nr_replicas,
 280                                       c->opts.metadata_replicas_required,
 281                                       watermark, 0, cl, &wp);
 282         if (unlikely(ret))
 283                 return ERR_PTR(ret);
 284
 285         if (wp->sectors_free < btree_sectors(c)) {
 286                 struct open_bucket *ob;
 287                 unsigned i;
 288
 289                 open_bucket_for_each(c, &wp->ptrs, ob, i)
 290                         if (ob->sectors_free < btree_sectors(c))
 291                                 ob->sectors_free = 0;
 292
 293                 bch2_alloc_sectors_done(c, wp);
 294                 goto retry;
 295         }
 296
 297         bkey_btree_ptr_v2_init(&tmp.k);
 298         bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, btree_sectors(c), false);
 299
 300         bch2_open_bucket_get(c, wp, &obs);
 301         bch2_alloc_sectors_done(c, wp);
 302 mem_alloc:
 303         b = bch2_btree_node_mem_alloc(trans, interior_node);
 304         six_unlock_write(&b->c.lock);
 305         six_unlock_intent(&b->c.lock);
 306
 307         /* we hold cannibalize_lock: */
 308         BUG_ON(IS_ERR(b));
 309         BUG_ON(b->ob.nr);
 310
 311         bkey_copy(&b->key, &tmp.k);
 312         b->ob = obs;
 313
 314         return b;
 315 }
 316
 317 static struct btree *bch2_btree_node_alloc(struct btree_update *as,
 318                                            struct btree_trans *trans,
 319                                            unsigned level)
 320 {
 321         struct bch_fs *c = as->c;
 322         struct btree *b;
 323         struct prealloc_nodes *p = &as->prealloc_nodes[!!level];
 324         int ret;
 325
 326         BUG_ON(level >= BTREE_MAX_DEPTH);
 327         BUG_ON(!p->nr);
 328
 329         b = p->b[--p->nr];
 330
 331         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
 332         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
 333
 334         set_btree_node_accessed(b);
 335         set_btree_node_dirty_acct(c, b);
 336         set_btree_node_need_write(b);
 337
 338         bch2_bset_init_first(b, &b->data->keys);
 339         b->c.level      = level;
 340         b->c.btree_id   = as->btree_id;
 341         b->version_ondisk = c->sb.version;
 342
 343         memset(&b->nr, 0, sizeof(b->nr));
 344         b->data->magic = cpu_to_le64(bset_magic(c));
 345         memset(&b->data->_ptr, 0, sizeof(b->data->_ptr));
 346         b->data->flags = 0;
 347         SET_BTREE_NODE_ID(b->data, as->btree_id);
 348         SET_BTREE_NODE_LEVEL(b->data, level);
 349
 350         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 351                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
 352
 353                 bp->v.mem_ptr           = 0;
 354                 bp->v.seq               = b->data->keys.seq;
 355                 bp->v.sectors_written   = 0;
 356         }
 357
 358         SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
 359
 360         bch2_btree_build_aux_trees(b);
 361
 362         ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
 363         BUG_ON(ret);
 364
 365         trace_and_count(c, btree_node_alloc, c, b);
 366         bch2_increment_clock(c, btree_sectors(c), WRITE);
 367         return b;
 368 }
 369
 370 static void btree_set_min(struct btree *b, struct bpos pos)
 371 {
 372         if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
 373                 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
 374         b->data->min_key = pos;
 375 }
 376
 377 static void btree_set_max(struct btree *b, struct bpos pos)
 378 {
 379         b->key.k.p = pos;
 380         b->data->max_key = pos;
 381 }
 382
 383 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
 384                                                        struct btree_trans *trans,
 385                                                        struct btree *b)
 386 {
 387         struct btree *n = bch2_btree_node_alloc(as, trans, b->c.level);
 388         struct bkey_format format = bch2_btree_calc_format(b);
 389
 390         /*
 391          * The keys might expand with the new format - if they wouldn't fit in
 392          * the btree node anymore, use the old format for now:
 393          */
 394         if (!bch2_btree_node_format_fits(as->c, b, &format))
 395                 format = b->format;
 396
 397         SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
 398
 399         btree_set_min(n, b->data->min_key);
 400         btree_set_max(n, b->data->max_key);
 401
 402         n->data->format         = format;
 403         btree_node_set_format(n, format);
 404
 405         bch2_btree_sort_into(as->c, n, b);
 406
 407         btree_node_reset_sib_u64s(n);
 408         return n;
 409 }
 410
 411 static struct btree *__btree_root_alloc(struct btree_update *as,
 412                                 struct btree_trans *trans, unsigned level)
 413 {
 414         struct btree *b = bch2_btree_node_alloc(as, trans, level);
 415
 416         btree_set_min(b, POS_MIN);
 417         btree_set_max(b, SPOS_MAX);
 418         b->data->format = bch2_btree_calc_format(b);
 419
 420         btree_node_set_format(b, b->data->format);
 421         bch2_btree_build_aux_trees(b);
 422
 423         return b;
 424 }
 425
 426 static void bch2_btree_reserve_put(struct btree_update *as, struct btree_trans *trans)
 427 {
 428         struct bch_fs *c = as->c;
 429         struct prealloc_nodes *p;
 430
 431         for (p = as->prealloc_nodes;
 432              p < as->prealloc_nodes + ARRAY_SIZE(as->prealloc_nodes);
 433              p++) {
 434                 while (p->nr) {
 435                         struct btree *b = p->b[--p->nr];
 436
 437                         mutex_lock(&c->btree_reserve_cache_lock);
 438
 439                         if (c->btree_reserve_cache_nr <
 440                             ARRAY_SIZE(c->btree_reserve_cache)) {
 441                                 struct btree_alloc *a =
 442                                         &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
 443
 444                                 a->ob = b->ob;
 445                                 b->ob.nr = 0;
 446                                 bkey_copy(&a->k, &b->key);
 447                         } else {
 448                                 bch2_open_buckets_put(c, &b->ob);
 449                         }
 450
 451                         mutex_unlock(&c->btree_reserve_cache_lock);
 452
 453                         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
 454                         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
 455                         __btree_node_free(c, b);
 456                         six_unlock_write(&b->c.lock);
 457                         six_unlock_intent(&b->c.lock);
 458                 }
 459         }
 460 }
 461
 462 static int bch2_btree_reserve_get(struct btree_trans *trans,
 463                                   struct btree_update *as,
 464                                   unsigned nr_nodes[2],
 465                                   unsigned flags,
 466                                   struct closure *cl)
 467 {
 468         struct bch_fs *c = as->c;
 469         struct btree *b;
 470         unsigned interior;
 471         int ret = 0;
 472
 473         BUG_ON(nr_nodes[0] + nr_nodes[1] > BTREE_RESERVE_MAX);
 474
 475         /*
 476          * Protects reaping from the btree node cache and using the btree node
 477          * open bucket reserve:
 478          */
 479         ret = bch2_btree_cache_cannibalize_lock(c, cl);
 480         if (ret)
 481                 return ret;
 482
 483         for (interior = 0; interior < 2; interior++) {
 484                 struct prealloc_nodes *p = as->prealloc_nodes + interior;
 485
 486                 while (p->nr < nr_nodes[interior]) {
 487                         b = __bch2_btree_node_alloc(trans, &as->disk_res, cl,
 488                                                     interior, flags);
 489                         if (IS_ERR(b)) {
 490                                 ret = PTR_ERR(b);
 491                                 goto err;
 492                         }
 493
 494                         p->b[p->nr++] = b;
 495                 }
 496         }
 497 err:
 498         bch2_btree_cache_cannibalize_unlock(c);
 499         return ret;
 500 }
 501
 502 /* Asynchronous interior node update machinery */
 503
 504 static void bch2_btree_update_free(struct btree_update *as, struct btree_trans *trans)
 505 {
 506         struct bch_fs *c = as->c;
 507
 508         if (as->took_gc_lock)
 509                 up_read(&c->gc_lock);
 510         as->took_gc_lock = false;
 511
 512         bch2_journal_pin_drop(&c->journal, &as->journal);
 513         bch2_journal_pin_flush(&c->journal, &as->journal);
 514         bch2_disk_reservation_put(c, &as->disk_res);
 515         bch2_btree_reserve_put(as, trans);
 516
 517         bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_total],
 518                                as->start_time);
 519
 520         mutex_lock(&c->btree_interior_update_lock);
 521         list_del(&as->unwritten_list);
 522         list_del(&as->list);
 523
 524         closure_debug_destroy(&as->cl);
 525         mempool_free(as, &c->btree_interior_update_pool);
 526
 527         /*
 528          * Have to do the wakeup with btree_interior_update_lock still held,
 529          * since being on btree_interior_update_list is our ref on @c:
 530          */
 531         closure_wake_up(&c->btree_interior_update_wait);
 532
 533         mutex_unlock(&c->btree_interior_update_lock);
 534 }
 535
 536 static void btree_update_add_key(struct btree_update *as,
 537                                  struct keylist *keys, struct btree *b)
 538 {
 539         struct bkey_i *k = &b->key;
 540
 541         BUG_ON(bch2_keylist_u64s(keys) + k->k.u64s >
 542                ARRAY_SIZE(as->_old_keys));
 543
 544         bkey_copy(keys->top, k);
 545         bkey_i_to_btree_ptr_v2(keys->top)->v.mem_ptr = b->c.level + 1;
 546
 547         bch2_keylist_push(keys);
 548 }
 549
 550 /*
 551  * The transactional part of an interior btree node update, where we journal the
 552  * update we did to the interior node and update alloc info:
 553  */
 554 static int btree_update_nodes_written_trans(struct btree_trans *trans,
 555                                             struct btree_update *as)
 556 {
 557         struct bkey_i *k;
 558         int ret;
 559
 560         ret = darray_make_room(&trans->extra_journal_entries, as->journal_u64s);
 561         if (ret)
 562                 return ret;
 563
 564         memcpy(&darray_top(trans->extra_journal_entries),
 565                as->journal_entries,
 566                as->journal_u64s * sizeof(u64));
 567         trans->extra_journal_entries.nr += as->journal_u64s;
 568
 569         trans->journal_pin = &as->journal;
 570
 571         for_each_keylist_key(&as->old_keys, k) {
 572                 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
 573
 574                 ret = bch2_trans_mark_old(trans, as->btree_id, level, bkey_i_to_s_c(k), 0);
 575                 if (ret)
 576                         return ret;
 577         }
 578
 579         for_each_keylist_key(&as->new_keys, k) {
 580                 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
 581
 582                 ret = bch2_trans_mark_new(trans, as->btree_id, level, k, 0);
 583                 if (ret)
 584                         return ret;
 585         }
 586
 587         return 0;
 588 }
 589
 590 static void btree_update_nodes_written(struct btree_update *as)
 591 {
 592         struct bch_fs *c = as->c;
 593         struct btree *b;
 594         struct btree_trans *trans = bch2_trans_get(c);
 595         u64 journal_seq = 0;
 596         unsigned i;
 597         int ret;
 598
 599         /*
 600          * If we're already in an error state, it might be because a btree node
 601          * was never written, and we might be trying to free that same btree
 602          * node here, but it won't have been marked as allocated and we'll see
 603          * spurious disk usage inconsistencies in the transactional part below
 604          * if we don't skip it:
 605          */
 606         ret = bch2_journal_error(&c->journal);
 607         if (ret)
 608                 goto err;
 609
 610         /*
 611          * Wait for any in flight writes to finish before we free the old nodes
 612          * on disk:
 613          */
 614         for (i = 0; i < as->nr_old_nodes; i++) {
 615                 __le64 seq;
 616
 617                 b = as->old_nodes[i];
 618
 619                 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
 620                 seq = b->data ? b->data->keys.seq : 0;
 621                 six_unlock_read(&b->c.lock);
 622
 623                 if (seq == as->old_nodes_seq[i])
 624                         wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight_inner,
 625                                        TASK_UNINTERRUPTIBLE);
 626         }
 627
 628         /*
 629          * We did an update to a parent node where the pointers we added pointed
 630          * to child nodes that weren't written yet: now, the child nodes have
 631          * been written so we can write out the update to the interior node.
 632          */
 633
 634         /*
 635          * We can't call into journal reclaim here: we'd block on the journal
 636          * reclaim lock, but we may need to release the open buckets we have
 637          * pinned in order for other btree updates to make forward progress, and
 638          * journal reclaim does btree updates when flushing bkey_cached entries,
 639          * which may require allocations as well.
 640          */
 641         ret = commit_do(trans, &as->disk_res, &journal_seq,
 642                         BCH_WATERMARK_reclaim|
 643                         BCH_TRANS_COMMIT_no_enospc|
 644                         BCH_TRANS_COMMIT_no_check_rw|
 645                         BCH_TRANS_COMMIT_journal_reclaim,
 646                         btree_update_nodes_written_trans(trans, as));
 647         bch2_trans_unlock(trans);
 648
 649         bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
 650                              "%s(): error %s", __func__, bch2_err_str(ret));
 651 err:
 652         if (as->b) {
 653                 struct btree_path *path;
 654
 655                 b = as->b;
 656                 path = get_unlocked_mut_path(trans, as->btree_id, b->c.level, b->key.k.p);
 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, 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 void btree_update_set_nodes_written(struct closure *cl)
 778 {
 779         struct btree_update *as = container_of(cl, 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         unsigned journal_flags = 0;
1058         int ret = 0;
1059         u32 restart_count = trans->restart_count;
1060
1061         BUG_ON(!path->should_be_locked);
1062
1063         if (watermark == BCH_WATERMARK_copygc)
1064                 watermark = BCH_WATERMARK_btree_copygc;
1065         if (watermark < BCH_WATERMARK_btree)
1066                 watermark = BCH_WATERMARK_btree;
1067
1068         flags &= ~BCH_WATERMARK_MASK;
1069         flags |= watermark;
1070
1071         if (flags & BCH_TRANS_COMMIT_journal_reclaim)
1072                 journal_flags |= JOURNAL_RES_GET_NONBLOCK;
1073         journal_flags |= watermark;
1074
1075         while (1) {
1076                 nr_nodes[!!update_level] += 1 + split;
1077                 update_level++;
1078
1079                 ret = bch2_btree_path_upgrade(trans, path, update_level + 1);
1080                 if (ret)
1081                         return ERR_PTR(ret);
1082
1083                 if (!btree_path_node(path, update_level)) {
1084                         /* Allocating new root? */
1085                         nr_nodes[1] += split;
1086                         update_level = BTREE_MAX_DEPTH;
1087                         break;
1088                 }
1089
1090                 if (bch2_btree_node_insert_fits(c, path->l[update_level].b,
1091                                         BKEY_BTREE_PTR_U64s_MAX * (1 + split)))
1092                         break;
1093
1094                 split = path->l[update_level].b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c);
1095         }
1096
1097         if (!down_read_trylock(&c->gc_lock)) {
1098                 ret = drop_locks_do(trans, (down_read(&c->gc_lock), 0));
1099                 if (ret) {
1100                         up_read(&c->gc_lock);
1101                         return ERR_PTR(ret);
1102                 }
1103         }
1104
1105         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOFS);
1106         memset(as, 0, sizeof(*as));
1107         closure_init(&as->cl, NULL);
1108         as->c           = c;
1109         as->start_time  = start_time;
1110         as->mode        = BTREE_INTERIOR_NO_UPDATE;
1111         as->took_gc_lock = true;
1112         as->btree_id    = path->btree_id;
1113         as->update_level = update_level;
1114         INIT_LIST_HEAD(&as->list);
1115         INIT_LIST_HEAD(&as->unwritten_list);
1116         INIT_LIST_HEAD(&as->write_blocked_list);
1117         bch2_keylist_init(&as->old_keys, as->_old_keys);
1118         bch2_keylist_init(&as->new_keys, as->_new_keys);
1119         bch2_keylist_init(&as->parent_keys, as->inline_keys);
1120
1121         mutex_lock(&c->btree_interior_update_lock);
1122         list_add_tail(&as->list, &c->btree_interior_update_list);
1123         mutex_unlock(&c->btree_interior_update_lock);
1124
1125         /*
1126          * We don't want to allocate if we're in an error state, that can cause
1127          * deadlock on emergency shutdown due to open buckets getting stuck in
1128          * the btree_reserve_cache after allocator shutdown has cleared it out.
1129          * This check needs to come after adding us to the btree_interior_update
1130          * list but before calling bch2_btree_reserve_get, to synchronize with
1131          * __bch2_fs_read_only().
1132          */
1133         ret = bch2_journal_error(&c->journal);
1134         if (ret)
1135                 goto err;
1136
1137         ret = bch2_disk_reservation_get(c, &as->disk_res,
1138                         (nr_nodes[0] + nr_nodes[1]) * btree_sectors(c),
1139                         c->opts.metadata_replicas,
1140                         disk_res_flags);
1141         if (ret)
1142                 goto err;
1143
1144         ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, NULL);
1145         if (bch2_err_matches(ret, ENOSPC) ||
1146             bch2_err_matches(ret, ENOMEM)) {
1147                 struct closure cl;
1148
1149                 /*
1150                  * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
1151                  * flag
1152                  */
1153                 if (bch2_err_matches(ret, ENOSPC) &&
1154                     (flags & BCH_TRANS_COMMIT_journal_reclaim) &&
1155                     watermark != BCH_WATERMARK_reclaim) {
1156                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
1157                         goto err;
1158                 }
1159
1160                 closure_init_stack(&cl);
1161
1162                 do {
1163                         ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, &cl);
1164
1165                         bch2_trans_unlock(trans);
1166                         closure_sync(&cl);
1167                 } while (bch2_err_matches(ret, BCH_ERR_operation_blocked));
1168         }
1169
1170         if (ret) {
1171                 trace_and_count(c, btree_reserve_get_fail, trans->fn,
1172                                 _RET_IP_, nr_nodes[0] + nr_nodes[1], ret);
1173                 goto err;
1174         }
1175
1176         ret = bch2_trans_relock(trans);
1177         if (ret)
1178                 goto err;
1179
1180         bch2_trans_verify_not_restarted(trans, restart_count);
1181         return as;
1182 err:
1183         bch2_btree_update_free(as, trans);
1184         return ERR_PTR(ret);
1185 }
1186
1187 /* Btree root updates: */
1188
1189 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1190 {
1191         /* Root nodes cannot be reaped */
1192         mutex_lock(&c->btree_cache.lock);
1193         list_del_init(&b->list);
1194         mutex_unlock(&c->btree_cache.lock);
1195
1196         mutex_lock(&c->btree_root_lock);
1197         BUG_ON(btree_node_root(c, b) &&
1198                (b->c.level < btree_node_root(c, b)->c.level ||
1199                 !btree_node_dying(btree_node_root(c, b))));
1200
1201         bch2_btree_id_root(c, b->c.btree_id)->b = b;
1202         mutex_unlock(&c->btree_root_lock);
1203
1204         bch2_recalc_btree_reserve(c);
1205 }
1206
1207 static void bch2_btree_set_root(struct btree_update *as,
1208                                 struct btree_trans *trans,
1209                                 struct btree_path *path,
1210                                 struct btree *b)
1211 {
1212         struct bch_fs *c = as->c;
1213         struct btree *old;
1214
1215         trace_and_count(c, btree_node_set_root, c, b);
1216
1217         old = btree_node_root(c, b);
1218
1219         /*
1220          * Ensure no one is using the old root while we switch to the
1221          * new root:
1222          */
1223         bch2_btree_node_lock_write_nofail(trans, path, &old->c);
1224
1225         bch2_btree_set_root_inmem(c, b);
1226
1227         btree_update_updated_root(as, b);
1228
1229         /*
1230          * Unlock old root after new root is visible:
1231          *
1232          * The new root isn't persistent, but that's ok: we still have
1233          * an intent lock on the new root, and any updates that would
1234          * depend on the new root would have to update the new root.
1235          */
1236         bch2_btree_node_unlock_write(trans, path, old);
1237 }
1238
1239 /* Interior node updates: */
1240
1241 static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
1242                                         struct btree_trans *trans,
1243                                         struct btree_path *path,
1244                                         struct btree *b,
1245                                         struct btree_node_iter *node_iter,
1246                                         struct bkey_i *insert)
1247 {
1248         struct bch_fs *c = as->c;
1249         struct bkey_packed *k;
1250         struct printbuf buf = PRINTBUF;
1251         unsigned long old, new, v;
1252
1253         BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 &&
1254                !btree_ptr_sectors_written(insert));
1255
1256         if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
1257                 bch2_journal_key_overwritten(c, b->c.btree_id, b->c.level, insert->k.p);
1258
1259         if (bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1260                               btree_node_type(b), WRITE, &buf) ?:
1261             bch2_bkey_in_btree_node(c, b, bkey_i_to_s_c(insert), &buf)) {
1262                 printbuf_reset(&buf);
1263                 prt_printf(&buf, "inserting invalid bkey\n  ");
1264                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
1265                 prt_printf(&buf, "\n  ");
1266                 bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1267                                   btree_node_type(b), WRITE, &buf);
1268                 bch2_bkey_in_btree_node(c, b, bkey_i_to_s_c(insert), &buf);
1269
1270                 bch2_fs_inconsistent(c, "%s", buf.buf);
1271                 dump_stack();
1272         }
1273
1274         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1275                ARRAY_SIZE(as->journal_entries));
1276
1277         as->journal_u64s +=
1278                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1279                                   BCH_JSET_ENTRY_btree_keys,
1280                                   b->c.btree_id, b->c.level,
1281                                   insert, insert->k.u64s);
1282
1283         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1284                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1285                 bch2_btree_node_iter_advance(node_iter, b);
1286
1287         bch2_btree_bset_insert_key(trans, path, b, node_iter, insert);
1288         set_btree_node_dirty_acct(c, b);
1289
1290         v = READ_ONCE(b->flags);
1291         do {
1292                 old = new = v;
1293
1294                 new &= ~BTREE_WRITE_TYPE_MASK;
1295                 new |= BTREE_WRITE_interior;
1296                 new |= 1 << BTREE_NODE_need_write;
1297         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1298
1299         printbuf_exit(&buf);
1300 }
1301
1302 static void
1303 __bch2_btree_insert_keys_interior(struct btree_update *as,
1304                                   struct btree_trans *trans,
1305                                   struct btree_path *path,
1306                                   struct btree *b,
1307                                   struct btree_node_iter node_iter,
1308                                   struct keylist *keys)
1309 {
1310         struct bkey_i *insert = bch2_keylist_front(keys);
1311         struct bkey_packed *k;
1312
1313         BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1314
1315         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1316                (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1317                 ;
1318
1319         while (!bch2_keylist_empty(keys)) {
1320                 insert = bch2_keylist_front(keys);
1321
1322                 if (bpos_gt(insert->k.p, b->key.k.p))
1323                         break;
1324
1325                 bch2_insert_fixup_btree_ptr(as, trans, path, b, &node_iter, insert);
1326                 bch2_keylist_pop_front(keys);
1327         }
1328 }
1329
1330 /*
1331  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1332  * node)
1333  */
1334 static void __btree_split_node(struct btree_update *as,
1335                                struct btree_trans *trans,
1336                                struct btree *b,
1337                                struct btree *n[2])
1338 {
1339         struct bkey_packed *k;
1340         struct bpos n1_pos = POS_MIN;
1341         struct btree_node_iter iter;
1342         struct bset *bsets[2];
1343         struct bkey_format_state format[2];
1344         struct bkey_packed *out[2];
1345         struct bkey uk;
1346         unsigned u64s, n1_u64s = (b->nr.live_u64s * 3) / 5;
1347         int i;
1348
1349         for (i = 0; i < 2; i++) {
1350                 BUG_ON(n[i]->nsets != 1);
1351
1352                 bsets[i] = btree_bset_first(n[i]);
1353                 out[i] = bsets[i]->start;
1354
1355                 SET_BTREE_NODE_SEQ(n[i]->data, BTREE_NODE_SEQ(b->data) + 1);
1356                 bch2_bkey_format_init(&format[i]);
1357         }
1358
1359         u64s = 0;
1360         for_each_btree_node_key(b, k, &iter) {
1361                 if (bkey_deleted(k))
1362                         continue;
1363
1364                 i = u64s >= n1_u64s;
1365                 u64s += k->u64s;
1366                 uk = bkey_unpack_key(b, k);
1367                 if (!i)
1368                         n1_pos = uk.p;
1369                 bch2_bkey_format_add_key(&format[i], &uk);
1370         }
1371
1372         btree_set_min(n[0], b->data->min_key);
1373         btree_set_max(n[0], n1_pos);
1374         btree_set_min(n[1], bpos_successor(n1_pos));
1375         btree_set_max(n[1], b->data->max_key);
1376
1377         for (i = 0; i < 2; i++) {
1378                 bch2_bkey_format_add_pos(&format[i], n[i]->data->min_key);
1379                 bch2_bkey_format_add_pos(&format[i], n[i]->data->max_key);
1380
1381                 n[i]->data->format = bch2_bkey_format_done(&format[i]);
1382                 btree_node_set_format(n[i], n[i]->data->format);
1383         }
1384
1385         u64s = 0;
1386         for_each_btree_node_key(b, k, &iter) {
1387                 if (bkey_deleted(k))
1388                         continue;
1389
1390                 i = u64s >= n1_u64s;
1391                 u64s += k->u64s;
1392
1393                 if (bch2_bkey_transform(&n[i]->format, out[i], bkey_packed(k)
1394                                         ? &b->format: &bch2_bkey_format_current, k))
1395                         out[i]->format = KEY_FORMAT_LOCAL_BTREE;
1396                 else
1397                         bch2_bkey_unpack(b, (void *) out[i], k);
1398
1399                 out[i]->needs_whiteout = false;
1400
1401                 btree_keys_account_key_add(&n[i]->nr, 0, out[i]);
1402                 out[i] = bkey_p_next(out[i]);
1403         }
1404
1405         for (i = 0; i < 2; i++) {
1406                 bsets[i]->u64s = cpu_to_le16((u64 *) out[i] - bsets[i]->_data);
1407
1408                 BUG_ON(!bsets[i]->u64s);
1409
1410                 set_btree_bset_end(n[i], n[i]->set);
1411
1412                 btree_node_reset_sib_u64s(n[i]);
1413
1414                 bch2_verify_btree_nr_keys(n[i]);
1415
1416                 if (b->c.level)
1417                         btree_node_interior_verify(as->c, n[i]);
1418         }
1419 }
1420
1421 /*
1422  * For updates to interior nodes, we've got to do the insert before we split
1423  * because the stuff we're inserting has to be inserted atomically. Post split,
1424  * the keys might have to go in different nodes and the split would no longer be
1425  * atomic.
1426  *
1427  * Worse, if the insert is from btree node coalescing, if we do the insert after
1428  * we do the split (and pick the pivot) - the pivot we pick might be between
1429  * nodes that were coalesced, and thus in the middle of a child node post
1430  * coalescing:
1431  */
1432 static void btree_split_insert_keys(struct btree_update *as,
1433                                     struct btree_trans *trans,
1434                                     struct btree_path *path,
1435                                     struct btree *b,
1436                                     struct keylist *keys)
1437 {
1438         if (!bch2_keylist_empty(keys) &&
1439             bpos_le(bch2_keylist_front(keys)->k.p, b->data->max_key)) {
1440                 struct btree_node_iter node_iter;
1441
1442                 bch2_btree_node_iter_init(&node_iter, b, &bch2_keylist_front(keys)->k.p);
1443
1444                 __bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
1445
1446                 btree_node_interior_verify(as->c, b);
1447         }
1448 }
1449
1450 static int btree_split(struct btree_update *as, struct btree_trans *trans,
1451                        struct btree_path *path, struct btree *b,
1452                        struct keylist *keys, unsigned flags)
1453 {
1454         struct bch_fs *c = as->c;
1455         struct btree *parent = btree_node_parent(path, b);
1456         struct btree *n1, *n2 = NULL, *n3 = NULL;
1457         struct btree_path *path1 = NULL, *path2 = NULL;
1458         u64 start_time = local_clock();
1459         int ret = 0;
1460
1461         BUG_ON(!parent && (b != btree_node_root(c, b)));
1462         BUG_ON(parent && !btree_node_intent_locked(path, b->c.level + 1));
1463
1464         bch2_btree_interior_update_will_free_node(as, b);
1465
1466         if (b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c)) {
1467                 struct btree *n[2];
1468
1469                 trace_and_count(c, btree_node_split, c, b);
1470
1471                 n[0] = n1 = bch2_btree_node_alloc(as, trans, b->c.level);
1472                 n[1] = n2 = bch2_btree_node_alloc(as, trans, b->c.level);
1473
1474                 __btree_split_node(as, trans, b, n);
1475
1476                 if (keys) {
1477                         btree_split_insert_keys(as, trans, path, n1, keys);
1478                         btree_split_insert_keys(as, trans, path, n2, keys);
1479                         BUG_ON(!bch2_keylist_empty(keys));
1480                 }
1481
1482                 bch2_btree_build_aux_trees(n2);
1483                 bch2_btree_build_aux_trees(n1);
1484
1485                 bch2_btree_update_add_new_node(as, n1);
1486                 bch2_btree_update_add_new_node(as, n2);
1487                 six_unlock_write(&n2->c.lock);
1488                 six_unlock_write(&n1->c.lock);
1489
1490                 path1 = get_unlocked_mut_path(trans, path->btree_id, n1->c.level, n1->key.k.p);
1491                 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1492                 mark_btree_node_locked(trans, path1, n1->c.level, BTREE_NODE_INTENT_LOCKED);
1493                 bch2_btree_path_level_init(trans, path1, n1);
1494
1495                 path2 = get_unlocked_mut_path(trans, path->btree_id, n2->c.level, n2->key.k.p);
1496                 six_lock_increment(&n2->c.lock, SIX_LOCK_intent);
1497                 mark_btree_node_locked(trans, path2, n2->c.level, BTREE_NODE_INTENT_LOCKED);
1498                 bch2_btree_path_level_init(trans, path2, n2);
1499
1500                 /*
1501                  * Note that on recursive parent_keys == keys, so we
1502                  * can't start adding new keys to parent_keys before emptying it
1503                  * out (which we did with btree_split_insert_keys() above)
1504                  */
1505                 bch2_keylist_add(&as->parent_keys, &n1->key);
1506                 bch2_keylist_add(&as->parent_keys, &n2->key);
1507
1508                 if (!parent) {
1509                         /* Depth increases, make a new root */
1510                         n3 = __btree_root_alloc(as, trans, b->c.level + 1);
1511
1512                         bch2_btree_update_add_new_node(as, n3);
1513                         six_unlock_write(&n3->c.lock);
1514
1515                         path2->locks_want++;
1516                         BUG_ON(btree_node_locked(path2, n3->c.level));
1517                         six_lock_increment(&n3->c.lock, SIX_LOCK_intent);
1518                         mark_btree_node_locked(trans, path2, n3->c.level, BTREE_NODE_INTENT_LOCKED);
1519                         bch2_btree_path_level_init(trans, path2, n3);
1520
1521                         n3->sib_u64s[0] = U16_MAX;
1522                         n3->sib_u64s[1] = U16_MAX;
1523
1524                         btree_split_insert_keys(as, trans, path, n3, &as->parent_keys);
1525                 }
1526         } else {
1527                 trace_and_count(c, btree_node_compact, c, b);
1528
1529                 n1 = bch2_btree_node_alloc_replacement(as, trans, b);
1530
1531                 if (keys) {
1532                         btree_split_insert_keys(as, trans, path, n1, keys);
1533                         BUG_ON(!bch2_keylist_empty(keys));
1534                 }
1535
1536                 bch2_btree_build_aux_trees(n1);
1537                 bch2_btree_update_add_new_node(as, n1);
1538                 six_unlock_write(&n1->c.lock);
1539
1540                 path1 = get_unlocked_mut_path(trans, path->btree_id, n1->c.level, n1->key.k.p);
1541                 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1542                 mark_btree_node_locked(trans, path1, n1->c.level, BTREE_NODE_INTENT_LOCKED);
1543                 bch2_btree_path_level_init(trans, path1, n1);
1544
1545                 if (parent)
1546                         bch2_keylist_add(&as->parent_keys, &n1->key);
1547         }
1548
1549         /* New nodes all written, now make them visible: */
1550
1551         if (parent) {
1552                 /* Split a non root node */
1553                 ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1554                 if (ret)
1555                         goto err;
1556         } else if (n3) {
1557                 bch2_btree_set_root(as, trans, path, n3);
1558         } else {
1559                 /* Root filled up but didn't need to be split */
1560                 bch2_btree_set_root(as, trans, path, n1);
1561         }
1562
1563         if (n3) {
1564                 bch2_btree_update_get_open_buckets(as, n3);
1565                 bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0);
1566         }
1567         if (n2) {
1568                 bch2_btree_update_get_open_buckets(as, n2);
1569                 bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0);
1570         }
1571         bch2_btree_update_get_open_buckets(as, n1);
1572         bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1573
1574         /*
1575          * The old node must be freed (in memory) _before_ unlocking the new
1576          * nodes - else another thread could re-acquire a read lock on the old
1577          * node after another thread has locked and updated the new node, thus
1578          * seeing stale data:
1579          */
1580         bch2_btree_node_free_inmem(trans, path, b);
1581
1582         if (n3)
1583                 bch2_trans_node_add(trans, n3);
1584         if (n2)
1585                 bch2_trans_node_add(trans, n2);
1586         bch2_trans_node_add(trans, n1);
1587
1588         if (n3)
1589                 six_unlock_intent(&n3->c.lock);
1590         if (n2)
1591                 six_unlock_intent(&n2->c.lock);
1592         six_unlock_intent(&n1->c.lock);
1593 out:
1594         if (path2) {
1595                 __bch2_btree_path_unlock(trans, path2);
1596                 bch2_path_put(trans, path2, true);
1597         }
1598         if (path1) {
1599                 __bch2_btree_path_unlock(trans, path1);
1600                 bch2_path_put(trans, path1, true);
1601         }
1602
1603         bch2_trans_verify_locks(trans);
1604
1605         bch2_time_stats_update(&c->times[n2
1606                                ? BCH_TIME_btree_node_split
1607                                : BCH_TIME_btree_node_compact],
1608                                start_time);
1609         return ret;
1610 err:
1611         if (n3)
1612                 bch2_btree_node_free_never_used(as, trans, n3);
1613         if (n2)
1614                 bch2_btree_node_free_never_used(as, trans, n2);
1615         bch2_btree_node_free_never_used(as, trans, n1);
1616         goto out;
1617 }
1618
1619 static void
1620 bch2_btree_insert_keys_interior(struct btree_update *as,
1621                                 struct btree_trans *trans,
1622                                 struct btree_path *path,
1623                                 struct btree *b,
1624                                 struct keylist *keys)
1625 {
1626         struct btree_path *linked;
1627
1628         __bch2_btree_insert_keys_interior(as, trans, path, b,
1629                                           path->l[b->c.level].iter, keys);
1630
1631         btree_update_updated_node(as, b);
1632
1633         trans_for_each_path_with_node(trans, b, linked)
1634                 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1635
1636         bch2_trans_verify_paths(trans);
1637 }
1638
1639 /**
1640  * bch2_btree_insert_node - insert bkeys into a given btree node
1641  *
1642  * @as:                 btree_update object
1643  * @trans:              btree_trans object
1644  * @path:               path that points to current node
1645  * @b:                  node to insert keys into
1646  * @keys:               list of keys to insert
1647  * @flags:              transaction commit flags
1648  *
1649  * Returns: 0 on success, typically transaction restart error on failure
1650  *
1651  * Inserts as many keys as it can into a given btree node, splitting it if full.
1652  * If a split occurred, this function will return early. This can only happen
1653  * for leaf nodes -- inserts into interior nodes have to be atomic.
1654  */
1655 static int bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans,
1656                                   struct btree_path *path, struct btree *b,
1657                                   struct keylist *keys, unsigned flags)
1658 {
1659         struct bch_fs *c = as->c;
1660         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1661         int old_live_u64s = b->nr.live_u64s;
1662         int live_u64s_added, u64s_added;
1663         int ret;
1664
1665         lockdep_assert_held(&c->gc_lock);
1666         BUG_ON(!btree_node_intent_locked(path, b->c.level));
1667         BUG_ON(!b->c.level);
1668         BUG_ON(!as || as->b);
1669         bch2_verify_keylist_sorted(keys);
1670
1671         ret = bch2_btree_node_lock_write(trans, path, &b->c);
1672         if (ret)
1673                 return ret;
1674
1675         bch2_btree_node_prep_for_write(trans, path, b);
1676
1677         if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1678                 bch2_btree_node_unlock_write(trans, path, b);
1679                 goto split;
1680         }
1681
1682         btree_node_interior_verify(c, b);
1683
1684         bch2_btree_insert_keys_interior(as, trans, path, b, keys);
1685
1686         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1687         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1688
1689         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1690                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1691         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1692                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1693
1694         if (u64s_added > live_u64s_added &&
1695             bch2_maybe_compact_whiteouts(c, b))
1696                 bch2_trans_node_reinit_iter(trans, b);
1697
1698         bch2_btree_node_unlock_write(trans, path, b);
1699
1700         btree_node_interior_verify(c, b);
1701         return 0;
1702 split:
1703         /*
1704          * We could attempt to avoid the transaction restart, by calling
1705          * bch2_btree_path_upgrade() and allocating more nodes:
1706          */
1707         if (b->c.level >= as->update_level) {
1708                 trace_and_count(c, trans_restart_split_race, trans, _THIS_IP_, b);
1709                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_split_race);
1710         }
1711
1712         return btree_split(as, trans, path, b, keys, flags);
1713 }
1714
1715 int bch2_btree_split_leaf(struct btree_trans *trans,
1716                           struct btree_path *path,
1717                           unsigned flags)
1718 {
1719         struct btree *b = path_l(path)->b;
1720         struct btree_update *as;
1721         unsigned l;
1722         int ret = 0;
1723
1724         as = bch2_btree_update_start(trans, path, path->level,
1725                                      true, flags);
1726         if (IS_ERR(as))
1727                 return PTR_ERR(as);
1728
1729         ret = btree_split(as, trans, path, b, NULL, flags);
1730         if (ret) {
1731                 bch2_btree_update_free(as, trans);
1732                 return ret;
1733         }
1734
1735         bch2_btree_update_done(as, trans);
1736
1737         for (l = path->level + 1; btree_node_intent_locked(path, l) && !ret; l++)
1738                 ret = bch2_foreground_maybe_merge(trans, path, l, flags);
1739
1740         return ret;
1741 }
1742
1743 int __bch2_foreground_maybe_merge(struct btree_trans *trans,
1744                                   struct btree_path *path,
1745                                   unsigned level,
1746                                   unsigned flags,
1747                                   enum btree_node_sibling sib)
1748 {
1749         struct bch_fs *c = trans->c;
1750         struct btree_path *sib_path = NULL, *new_path = NULL;
1751         struct btree_update *as;
1752         struct bkey_format_state new_s;
1753         struct bkey_format new_f;
1754         struct bkey_i delete;
1755         struct btree *b, *m, *n, *prev, *next, *parent;
1756         struct bpos sib_pos;
1757         size_t sib_u64s;
1758         u64 start_time = local_clock();
1759         int ret = 0;
1760
1761         BUG_ON(!path->should_be_locked);
1762         BUG_ON(!btree_node_locked(path, level));
1763
1764         b = path->l[level].b;
1765
1766         if ((sib == btree_prev_sib && bpos_eq(b->data->min_key, POS_MIN)) ||
1767             (sib == btree_next_sib && bpos_eq(b->data->max_key, SPOS_MAX))) {
1768                 b->sib_u64s[sib] = U16_MAX;
1769                 return 0;
1770         }
1771
1772         sib_pos = sib == btree_prev_sib
1773                 ? bpos_predecessor(b->data->min_key)
1774                 : bpos_successor(b->data->max_key);
1775
1776         sib_path = bch2_path_get(trans, path->btree_id, sib_pos,
1777                                  U8_MAX, level, BTREE_ITER_INTENT, _THIS_IP_);
1778         ret = bch2_btree_path_traverse(trans, sib_path, false);
1779         if (ret)
1780                 goto err;
1781
1782         btree_path_set_should_be_locked(sib_path);
1783
1784         m = sib_path->l[level].b;
1785
1786         if (btree_node_parent(path, b) !=
1787             btree_node_parent(sib_path, m)) {
1788                 b->sib_u64s[sib] = U16_MAX;
1789                 goto out;
1790         }
1791
1792         if (sib == btree_prev_sib) {
1793                 prev = m;
1794                 next = b;
1795         } else {
1796                 prev = b;
1797                 next = m;
1798         }
1799
1800         if (!bpos_eq(bpos_successor(prev->data->max_key), next->data->min_key)) {
1801                 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1802
1803                 bch2_bpos_to_text(&buf1, prev->data->max_key);
1804                 bch2_bpos_to_text(&buf2, next->data->min_key);
1805                 bch_err(c,
1806                         "%s(): btree topology error:\n"
1807                         "  prev ends at   %s\n"
1808                         "  next starts at %s",
1809                         __func__, buf1.buf, buf2.buf);
1810                 printbuf_exit(&buf1);
1811                 printbuf_exit(&buf2);
1812                 bch2_topology_error(c);
1813                 ret = -EIO;
1814                 goto err;
1815         }
1816
1817         bch2_bkey_format_init(&new_s);
1818         bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
1819         __bch2_btree_calc_format(&new_s, prev);
1820         __bch2_btree_calc_format(&new_s, next);
1821         bch2_bkey_format_add_pos(&new_s, next->data->max_key);
1822         new_f = bch2_bkey_format_done(&new_s);
1823
1824         sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1825                 btree_node_u64s_with_format(m, &new_f);
1826
1827         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1828                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1829                 sib_u64s /= 2;
1830                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1831         }
1832
1833         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1834         sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
1835         b->sib_u64s[sib] = sib_u64s;
1836
1837         if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
1838                 goto out;
1839
1840         parent = btree_node_parent(path, b);
1841         as = bch2_btree_update_start(trans, path, level, false,
1842                                      BCH_TRANS_COMMIT_no_enospc|flags);
1843         ret = PTR_ERR_OR_ZERO(as);
1844         if (ret)
1845                 goto err;
1846
1847         trace_and_count(c, btree_node_merge, c, b);
1848
1849         bch2_btree_interior_update_will_free_node(as, b);
1850         bch2_btree_interior_update_will_free_node(as, m);
1851
1852         n = bch2_btree_node_alloc(as, trans, b->c.level);
1853
1854         SET_BTREE_NODE_SEQ(n->data,
1855                            max(BTREE_NODE_SEQ(b->data),
1856                                BTREE_NODE_SEQ(m->data)) + 1);
1857
1858         btree_set_min(n, prev->data->min_key);
1859         btree_set_max(n, next->data->max_key);
1860
1861         n->data->format  = new_f;
1862         btree_node_set_format(n, new_f);
1863
1864         bch2_btree_sort_into(c, n, prev);
1865         bch2_btree_sort_into(c, n, next);
1866
1867         bch2_btree_build_aux_trees(n);
1868         bch2_btree_update_add_new_node(as, n);
1869         six_unlock_write(&n->c.lock);
1870
1871         new_path = get_unlocked_mut_path(trans, path->btree_id, n->c.level, n->key.k.p);
1872         six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1873         mark_btree_node_locked(trans, new_path, n->c.level, BTREE_NODE_INTENT_LOCKED);
1874         bch2_btree_path_level_init(trans, new_path, n);
1875
1876         bkey_init(&delete.k);
1877         delete.k.p = prev->key.k.p;
1878         bch2_keylist_add(&as->parent_keys, &delete);
1879         bch2_keylist_add(&as->parent_keys, &n->key);
1880
1881         bch2_trans_verify_paths(trans);
1882
1883         ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1884         if (ret)
1885                 goto err_free_update;
1886
1887         bch2_trans_verify_paths(trans);
1888
1889         bch2_btree_update_get_open_buckets(as, n);
1890         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1891
1892         bch2_btree_node_free_inmem(trans, path, b);
1893         bch2_btree_node_free_inmem(trans, sib_path, m);
1894
1895         bch2_trans_node_add(trans, n);
1896
1897         bch2_trans_verify_paths(trans);
1898
1899         six_unlock_intent(&n->c.lock);
1900
1901         bch2_btree_update_done(as, trans);
1902
1903         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time);
1904 out:
1905 err:
1906         if (new_path)
1907                 bch2_path_put(trans, new_path, true);
1908         bch2_path_put(trans, sib_path, true);
1909         bch2_trans_verify_locks(trans);
1910         return ret;
1911 err_free_update:
1912         bch2_btree_node_free_never_used(as, trans, n);
1913         bch2_btree_update_free(as, trans);
1914         goto out;
1915 }
1916
1917 int bch2_btree_node_rewrite(struct btree_trans *trans,
1918                             struct btree_iter *iter,
1919                             struct btree *b,
1920                             unsigned flags)
1921 {
1922         struct bch_fs *c = trans->c;
1923         struct btree_path *new_path = NULL;
1924         struct btree *n, *parent;
1925         struct btree_update *as;
1926         int ret;
1927
1928         flags |= BCH_TRANS_COMMIT_no_enospc;
1929
1930         parent = btree_node_parent(iter->path, b);
1931         as = bch2_btree_update_start(trans, iter->path, b->c.level,
1932                                      false, flags);
1933         ret = PTR_ERR_OR_ZERO(as);
1934         if (ret)
1935                 goto out;
1936
1937         bch2_btree_interior_update_will_free_node(as, b);
1938
1939         n = bch2_btree_node_alloc_replacement(as, trans, b);
1940
1941         bch2_btree_build_aux_trees(n);
1942         bch2_btree_update_add_new_node(as, n);
1943         six_unlock_write(&n->c.lock);
1944
1945         new_path = get_unlocked_mut_path(trans, iter->btree_id, n->c.level, n->key.k.p);
1946         six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1947         mark_btree_node_locked(trans, new_path, n->c.level, BTREE_NODE_INTENT_LOCKED);
1948         bch2_btree_path_level_init(trans, new_path, n);
1949
1950         trace_and_count(c, btree_node_rewrite, c, b);
1951
1952         if (parent) {
1953                 bch2_keylist_add(&as->parent_keys, &n->key);
1954                 ret = bch2_btree_insert_node(as, trans, iter->path, parent,
1955                                              &as->parent_keys, flags);
1956                 if (ret)
1957                         goto err;
1958         } else {
1959                 bch2_btree_set_root(as, trans, iter->path, n);
1960         }
1961
1962         bch2_btree_update_get_open_buckets(as, n);
1963         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1964
1965         bch2_btree_node_free_inmem(trans, iter->path, b);
1966
1967         bch2_trans_node_add(trans, n);
1968         six_unlock_intent(&n->c.lock);
1969
1970         bch2_btree_update_done(as, trans);
1971 out:
1972         if (new_path)
1973                 bch2_path_put(trans, new_path, true);
1974         bch2_trans_downgrade(trans);
1975         return ret;
1976 err:
1977         bch2_btree_node_free_never_used(as, trans, n);
1978         bch2_btree_update_free(as, trans);
1979         goto out;
1980 }
1981
1982 struct async_btree_rewrite {
1983         struct bch_fs           *c;
1984         struct work_struct      work;
1985         struct list_head        list;
1986         enum btree_id           btree_id;
1987         unsigned                level;
1988         struct bpos             pos;
1989         __le64                  seq;
1990 };
1991
1992 static int async_btree_node_rewrite_trans(struct btree_trans *trans,
1993                                           struct async_btree_rewrite *a)
1994 {
1995         struct bch_fs *c = trans->c;
1996         struct btree_iter iter;
1997         struct btree *b;
1998         int ret;
1999
2000         bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos,
2001                                   BTREE_MAX_DEPTH, a->level, 0);
2002         b = bch2_btree_iter_peek_node(&iter);
2003         ret = PTR_ERR_OR_ZERO(b);
2004         if (ret)
2005                 goto out;
2006
2007         if (!b || b->data->keys.seq != a->seq) {
2008                 struct printbuf buf = PRINTBUF;
2009
2010                 if (b)
2011                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
2012                 else
2013                         prt_str(&buf, "(null");
2014                 bch_info(c, "%s: node to rewrite not found:, searching for seq %llu, got\n%s",
2015                          __func__, a->seq, buf.buf);
2016                 printbuf_exit(&buf);
2017                 goto out;
2018         }
2019
2020         ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
2021 out:
2022         bch2_trans_iter_exit(trans, &iter);
2023
2024         return ret;
2025 }
2026
2027 static void async_btree_node_rewrite_work(struct work_struct *work)
2028 {
2029         struct async_btree_rewrite *a =
2030                 container_of(work, struct async_btree_rewrite, work);
2031         struct bch_fs *c = a->c;
2032         int ret;
2033
2034         ret = bch2_trans_do(c, NULL, NULL, 0,
2035                       async_btree_node_rewrite_trans(trans, a));
2036         if (ret)
2037                 bch_err_fn(c, ret);
2038         bch2_write_ref_put(c, BCH_WRITE_REF_node_rewrite);
2039         kfree(a);
2040 }
2041
2042 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
2043 {
2044         struct async_btree_rewrite *a;
2045         int ret;
2046
2047         a = kmalloc(sizeof(*a), GFP_NOFS);
2048         if (!a) {
2049                 bch_err(c, "%s: error allocating memory", __func__);
2050                 return;
2051         }
2052
2053         a->c            = c;
2054         a->btree_id     = b->c.btree_id;
2055         a->level        = b->c.level;
2056         a->pos          = b->key.k.p;
2057         a->seq          = b->data->keys.seq;
2058         INIT_WORK(&a->work, async_btree_node_rewrite_work);
2059
2060         if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
2061                 mutex_lock(&c->pending_node_rewrites_lock);
2062                 list_add(&a->list, &c->pending_node_rewrites);
2063                 mutex_unlock(&c->pending_node_rewrites_lock);
2064                 return;
2065         }
2066
2067         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_node_rewrite)) {
2068                 if (test_bit(BCH_FS_STARTED, &c->flags)) {
2069                         bch_err(c, "%s: error getting c->writes ref", __func__);
2070                         kfree(a);
2071                         return;
2072                 }
2073
2074                 ret = bch2_fs_read_write_early(c);
2075                 if (ret) {
2076                         bch_err_msg(c, ret, "going read-write");
2077                         kfree(a);
2078                         return;
2079                 }
2080
2081                 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2082         }
2083
2084         queue_work(c->btree_interior_update_worker, &a->work);
2085 }
2086
2087 void bch2_do_pending_node_rewrites(struct bch_fs *c)
2088 {
2089         struct async_btree_rewrite *a, *n;
2090
2091         mutex_lock(&c->pending_node_rewrites_lock);
2092         list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2093                 list_del(&a->list);
2094
2095                 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2096                 queue_work(c->btree_interior_update_worker, &a->work);
2097         }
2098         mutex_unlock(&c->pending_node_rewrites_lock);
2099 }
2100
2101 void bch2_free_pending_node_rewrites(struct bch_fs *c)
2102 {
2103         struct async_btree_rewrite *a, *n;
2104
2105         mutex_lock(&c->pending_node_rewrites_lock);
2106         list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2107                 list_del(&a->list);
2108
2109                 kfree(a);
2110         }
2111         mutex_unlock(&c->pending_node_rewrites_lock);
2112 }
2113
2114 static int __bch2_btree_node_update_key(struct btree_trans *trans,
2115                                         struct btree_iter *iter,
2116                                         struct btree *b, struct btree *new_hash,
2117                                         struct bkey_i *new_key,
2118                                         unsigned commit_flags,
2119                                         bool skip_triggers)
2120 {
2121         struct bch_fs *c = trans->c;
2122         struct btree_iter iter2 = { NULL };
2123         struct btree *parent;
2124         int ret;
2125
2126         if (!skip_triggers) {
2127                 ret = bch2_trans_mark_old(trans, b->c.btree_id, b->c.level + 1,
2128                                           bkey_i_to_s_c(&b->key), 0);
2129                 if (ret)
2130                         return ret;
2131
2132                 ret = bch2_trans_mark_new(trans, b->c.btree_id, b->c.level + 1,
2133                                           new_key, 0);
2134                 if (ret)
2135                         return ret;
2136         }
2137
2138         if (new_hash) {
2139                 bkey_copy(&new_hash->key, new_key);
2140                 ret = bch2_btree_node_hash_insert(&c->btree_cache,
2141                                 new_hash, b->c.level, b->c.btree_id);
2142                 BUG_ON(ret);
2143         }
2144
2145         parent = btree_node_parent(iter->path, b);
2146         if (parent) {
2147                 bch2_trans_copy_iter(&iter2, iter);
2148
2149                 iter2.path = bch2_btree_path_make_mut(trans, iter2.path,
2150                                 iter2.flags & BTREE_ITER_INTENT,
2151                                 _THIS_IP_);
2152
2153                 BUG_ON(iter2.path->level != b->c.level);
2154                 BUG_ON(!bpos_eq(iter2.path->pos, new_key->k.p));
2155
2156                 btree_path_set_level_up(trans, iter2.path);
2157
2158                 trans->paths_sorted = false;
2159
2160                 ret   = bch2_btree_iter_traverse(&iter2) ?:
2161                         bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_NORUN);
2162                 if (ret)
2163                         goto err;
2164         } else {
2165                 BUG_ON(btree_node_root(c, b) != b);
2166
2167                 ret = darray_make_room(&trans->extra_journal_entries,
2168                                        jset_u64s(new_key->k.u64s));
2169                 if (ret)
2170                         return ret;
2171
2172                 journal_entry_set((void *) &darray_top(trans->extra_journal_entries),
2173                                   BCH_JSET_ENTRY_btree_root,
2174                                   b->c.btree_id, b->c.level,
2175                                   new_key, new_key->k.u64s);
2176                 trans->extra_journal_entries.nr += jset_u64s(new_key->k.u64s);
2177         }
2178
2179         ret = bch2_trans_commit(trans, NULL, NULL, commit_flags);
2180         if (ret)
2181                 goto err;
2182
2183         bch2_btree_node_lock_write_nofail(trans, iter->path, &b->c);
2184
2185         if (new_hash) {
2186                 mutex_lock(&c->btree_cache.lock);
2187                 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
2188                 bch2_btree_node_hash_remove(&c->btree_cache, b);
2189
2190                 bkey_copy(&b->key, new_key);
2191                 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
2192                 BUG_ON(ret);
2193                 mutex_unlock(&c->btree_cache.lock);
2194         } else {
2195                 bkey_copy(&b->key, new_key);
2196         }
2197
2198         bch2_btree_node_unlock_write(trans, iter->path, b);
2199 out:
2200         bch2_trans_iter_exit(trans, &iter2);
2201         return ret;
2202 err:
2203         if (new_hash) {
2204                 mutex_lock(&c->btree_cache.lock);
2205                 bch2_btree_node_hash_remove(&c->btree_cache, b);
2206                 mutex_unlock(&c->btree_cache.lock);
2207         }
2208         goto out;
2209 }
2210
2211 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
2212                                struct btree *b, struct bkey_i *new_key,
2213                                unsigned commit_flags, bool skip_triggers)
2214 {
2215         struct bch_fs *c = trans->c;
2216         struct btree *new_hash = NULL;
2217         struct btree_path *path = iter->path;
2218         struct closure cl;
2219         int ret = 0;
2220
2221         ret = bch2_btree_path_upgrade(trans, path, b->c.level + 1);
2222         if (ret)
2223                 return ret;
2224
2225         closure_init_stack(&cl);
2226
2227         /*
2228          * check btree_ptr_hash_val() after @b is locked by
2229          * btree_iter_traverse():
2230          */
2231         if (btree_ptr_hash_val(new_key) != b->hash_val) {
2232                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2233                 if (ret) {
2234                         ret = drop_locks_do(trans, (closure_sync(&cl), 0));
2235                         if (ret)
2236                                 return ret;
2237                 }
2238
2239                 new_hash = bch2_btree_node_mem_alloc(trans, false);
2240         }
2241
2242         path->intent_ref++;
2243         ret = __bch2_btree_node_update_key(trans, iter, b, new_hash, new_key,
2244                                            commit_flags, skip_triggers);
2245         --path->intent_ref;
2246
2247         if (new_hash) {
2248                 mutex_lock(&c->btree_cache.lock);
2249                 list_move(&new_hash->list, &c->btree_cache.freeable);
2250                 mutex_unlock(&c->btree_cache.lock);
2251
2252                 six_unlock_write(&new_hash->c.lock);
2253                 six_unlock_intent(&new_hash->c.lock);
2254         }
2255         closure_sync(&cl);
2256         bch2_btree_cache_cannibalize_unlock(c);
2257         return ret;
2258 }
2259
2260 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
2261                                         struct btree *b, struct bkey_i *new_key,
2262                                         unsigned commit_flags, bool skip_triggers)
2263 {
2264         struct btree_iter iter;
2265         int ret;
2266
2267         bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p,
2268                                   BTREE_MAX_DEPTH, b->c.level,
2269                                   BTREE_ITER_INTENT);
2270         ret = bch2_btree_iter_traverse(&iter);
2271         if (ret)
2272                 goto out;
2273
2274         /* has node been freed? */
2275         if (iter.path->l[b->c.level].b != b) {
2276                 /* node has been freed: */
2277                 BUG_ON(!btree_node_dying(b));
2278                 goto out;
2279         }
2280
2281         BUG_ON(!btree_node_hashed(b));
2282
2283         ret = bch2_btree_node_update_key(trans, &iter, b, new_key,
2284                                          commit_flags, skip_triggers);
2285 out:
2286         bch2_trans_iter_exit(trans, &iter);
2287         return ret;
2288 }
2289
2290 /* Init code: */
2291
2292 /*
2293  * Only for filesystem bringup, when first reading the btree roots or allocating
2294  * btree roots when initializing a new filesystem:
2295  */
2296 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2297 {
2298         BUG_ON(btree_node_root(c, b));
2299
2300         bch2_btree_set_root_inmem(c, b);
2301 }
2302
2303 static int __bch2_btree_root_alloc(struct btree_trans *trans, enum btree_id id)
2304 {
2305         struct bch_fs *c = trans->c;
2306         struct closure cl;
2307         struct btree *b;
2308         int ret;
2309
2310         closure_init_stack(&cl);
2311
2312         do {
2313                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2314                 closure_sync(&cl);
2315         } while (ret);
2316
2317         b = bch2_btree_node_mem_alloc(trans, false);
2318         bch2_btree_cache_cannibalize_unlock(c);
2319
2320         set_btree_node_fake(b);
2321         set_btree_node_need_rewrite(b);
2322         b->c.level      = 0;
2323         b->c.btree_id   = id;
2324
2325         bkey_btree_ptr_init(&b->key);
2326         b->key.k.p = SPOS_MAX;
2327         *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2328
2329         bch2_bset_init_first(b, &b->data->keys);
2330         bch2_btree_build_aux_trees(b);
2331
2332         b->data->flags = 0;
2333         btree_set_min(b, POS_MIN);
2334         btree_set_max(b, SPOS_MAX);
2335         b->data->format = bch2_btree_calc_format(b);
2336         btree_node_set_format(b, b->data->format);
2337
2338         ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2339                                           b->c.level, b->c.btree_id);
2340         BUG_ON(ret);
2341
2342         bch2_btree_set_root_inmem(c, b);
2343
2344         six_unlock_write(&b->c.lock);
2345         six_unlock_intent(&b->c.lock);
2346         return 0;
2347 }
2348
2349 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2350 {
2351         bch2_trans_run(c, __bch2_btree_root_alloc(trans, id));
2352 }
2353
2354 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2355 {
2356         struct btree_update *as;
2357
2358         mutex_lock(&c->btree_interior_update_lock);
2359         list_for_each_entry(as, &c->btree_interior_update_list, list)
2360                 prt_printf(out, "%p m %u w %u r %u j %llu\n",
2361                        as,
2362                        as->mode,
2363                        as->nodes_written,
2364                        closure_nr_remaining(&as->cl),
2365                        as->journal.seq);
2366         mutex_unlock(&c->btree_interior_update_lock);
2367 }
2368
2369 static bool bch2_btree_interior_updates_pending(struct bch_fs *c)
2370 {
2371         bool ret;
2372
2373         mutex_lock(&c->btree_interior_update_lock);
2374         ret = !list_empty(&c->btree_interior_update_list);
2375         mutex_unlock(&c->btree_interior_update_lock);
2376
2377         return ret;
2378 }
2379
2380 bool bch2_btree_interior_updates_flush(struct bch_fs *c)
2381 {
2382         bool ret = bch2_btree_interior_updates_pending(c);
2383
2384         if (ret)
2385                 closure_wait_event(&c->btree_interior_update_wait,
2386                                    !bch2_btree_interior_updates_pending(c));
2387         return ret;
2388 }
2389
2390 void bch2_journal_entry_to_btree_root(struct bch_fs *c, struct jset_entry *entry)
2391 {
2392         struct btree_root *r = bch2_btree_id_root(c, entry->btree_id);
2393
2394         mutex_lock(&c->btree_root_lock);
2395
2396         r->level = entry->level;
2397         r->alive = true;
2398         bkey_copy(&r->key, (struct bkey_i *) entry->start);
2399
2400         mutex_unlock(&c->btree_root_lock);
2401 }
2402
2403 struct jset_entry *
2404 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2405                                     struct jset_entry *end,
2406                                     unsigned long skip)
2407 {
2408         unsigned i;
2409
2410         mutex_lock(&c->btree_root_lock);
2411
2412         for (i = 0; i < btree_id_nr_alive(c); i++) {
2413                 struct btree_root *r = bch2_btree_id_root(c, i);
2414
2415                 if (r->alive && !test_bit(i, &skip)) {
2416                         journal_entry_set(end, BCH_JSET_ENTRY_btree_root,
2417                                           i, r->level, &r->key, r->key.k.u64s);
2418                         end = vstruct_next(end);
2419                 }
2420         }
2421
2422         mutex_unlock(&c->btree_root_lock);
2423
2424         return end;
2425 }
2426
2427 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2428 {
2429         if (c->btree_interior_update_worker)
2430                 destroy_workqueue(c->btree_interior_update_worker);
2431         mempool_exit(&c->btree_interior_update_pool);
2432 }
2433
2434 void bch2_fs_btree_interior_update_init_early(struct bch_fs *c)
2435 {
2436         mutex_init(&c->btree_reserve_cache_lock);
2437         INIT_LIST_HEAD(&c->btree_interior_update_list);
2438         INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2439         mutex_init(&c->btree_interior_update_lock);
2440         INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2441
2442         INIT_LIST_HEAD(&c->pending_node_rewrites);
2443         mutex_init(&c->pending_node_rewrites_lock);
2444 }
2445
2446 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2447 {
2448         c->btree_interior_update_worker =
2449                 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2450         if (!c->btree_interior_update_worker)
2451                 return -BCH_ERR_ENOMEM_btree_interior_update_worker_init;
2452
2453         if (mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2454                                       sizeof(struct btree_update)))
2455                 return -BCH_ERR_ENOMEM_btree_interior_update_pool_init;
2456
2457         return 0;
2458 }