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         int ret = 0;
1058         u32 restart_count = trans->restart_count;
1059
1060         BUG_ON(!path->should_be_locked);
1061
1062         if (watermark == BCH_WATERMARK_copygc)
1063                 watermark = BCH_WATERMARK_btree_copygc;
1064         if (watermark < BCH_WATERMARK_btree)
1065                 watermark = BCH_WATERMARK_btree;
1066
1067         flags &= ~BCH_WATERMARK_MASK;
1068         flags |= watermark;
1069
1070         while (1) {
1071                 nr_nodes[!!update_level] += 1 + split;
1072                 update_level++;
1073
1074                 ret = bch2_btree_path_upgrade(trans, path, update_level + 1);
1075                 if (ret)
1076                         return ERR_PTR(ret);
1077
1078                 if (!btree_path_node(path, update_level)) {
1079                         /* Allocating new root? */
1080                         nr_nodes[1] += split;
1081                         update_level = BTREE_MAX_DEPTH;
1082                         break;
1083                 }
1084
1085                 if (bch2_btree_node_insert_fits(c, path->l[update_level].b,
1086                                         BKEY_BTREE_PTR_U64s_MAX * (1 + split)))
1087                         break;
1088
1089                 split = path->l[update_level].b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c);
1090         }
1091
1092         if (!down_read_trylock(&c->gc_lock)) {
1093                 ret = drop_locks_do(trans, (down_read(&c->gc_lock), 0));
1094                 if (ret) {
1095                         up_read(&c->gc_lock);
1096                         return ERR_PTR(ret);
1097                 }
1098         }
1099
1100         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOFS);
1101         memset(as, 0, sizeof(*as));
1102         closure_init(&as->cl, NULL);
1103         as->c           = c;
1104         as->start_time  = start_time;
1105         as->mode        = BTREE_INTERIOR_NO_UPDATE;
1106         as->took_gc_lock = true;
1107         as->btree_id    = path->btree_id;
1108         as->update_level = update_level;
1109         INIT_LIST_HEAD(&as->list);
1110         INIT_LIST_HEAD(&as->unwritten_list);
1111         INIT_LIST_HEAD(&as->write_blocked_list);
1112         bch2_keylist_init(&as->old_keys, as->_old_keys);
1113         bch2_keylist_init(&as->new_keys, as->_new_keys);
1114         bch2_keylist_init(&as->parent_keys, as->inline_keys);
1115
1116         mutex_lock(&c->btree_interior_update_lock);
1117         list_add_tail(&as->list, &c->btree_interior_update_list);
1118         mutex_unlock(&c->btree_interior_update_lock);
1119
1120         /*
1121          * We don't want to allocate if we're in an error state, that can cause
1122          * deadlock on emergency shutdown due to open buckets getting stuck in
1123          * the btree_reserve_cache after allocator shutdown has cleared it out.
1124          * This check needs to come after adding us to the btree_interior_update
1125          * list but before calling bch2_btree_reserve_get, to synchronize with
1126          * __bch2_fs_read_only().
1127          */
1128         ret = bch2_journal_error(&c->journal);
1129         if (ret)
1130                 goto err;
1131
1132         ret = bch2_disk_reservation_get(c, &as->disk_res,
1133                         (nr_nodes[0] + nr_nodes[1]) * btree_sectors(c),
1134                         c->opts.metadata_replicas,
1135                         disk_res_flags);
1136         if (ret)
1137                 goto err;
1138
1139         ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, NULL);
1140         if (bch2_err_matches(ret, ENOSPC) ||
1141             bch2_err_matches(ret, ENOMEM)) {
1142                 struct closure cl;
1143
1144                 /*
1145                  * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
1146                  * flag
1147                  */
1148                 if (bch2_err_matches(ret, ENOSPC) &&
1149                     (flags & BCH_TRANS_COMMIT_journal_reclaim) &&
1150                     watermark != BCH_WATERMARK_reclaim) {
1151                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
1152                         goto err;
1153                 }
1154
1155                 closure_init_stack(&cl);
1156
1157                 do {
1158                         ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, &cl);
1159
1160                         bch2_trans_unlock(trans);
1161                         closure_sync(&cl);
1162                 } while (bch2_err_matches(ret, BCH_ERR_operation_blocked));
1163         }
1164
1165         if (ret) {
1166                 trace_and_count(c, btree_reserve_get_fail, trans->fn,
1167                                 _RET_IP_, nr_nodes[0] + nr_nodes[1], ret);
1168                 goto err;
1169         }
1170
1171         ret = bch2_trans_relock(trans);
1172         if (ret)
1173                 goto err;
1174
1175         bch2_trans_verify_not_restarted(trans, restart_count);
1176         return as;
1177 err:
1178         bch2_btree_update_free(as, trans);
1179         return ERR_PTR(ret);
1180 }
1181
1182 /* Btree root updates: */
1183
1184 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1185 {
1186         /* Root nodes cannot be reaped */
1187         mutex_lock(&c->btree_cache.lock);
1188         list_del_init(&b->list);
1189         mutex_unlock(&c->btree_cache.lock);
1190
1191         mutex_lock(&c->btree_root_lock);
1192         BUG_ON(btree_node_root(c, b) &&
1193                (b->c.level < btree_node_root(c, b)->c.level ||
1194                 !btree_node_dying(btree_node_root(c, b))));
1195
1196         bch2_btree_id_root(c, b->c.btree_id)->b = b;
1197         mutex_unlock(&c->btree_root_lock);
1198
1199         bch2_recalc_btree_reserve(c);
1200 }
1201
1202 static void bch2_btree_set_root(struct btree_update *as,
1203                                 struct btree_trans *trans,
1204                                 struct btree_path *path,
1205                                 struct btree *b)
1206 {
1207         struct bch_fs *c = as->c;
1208         struct btree *old;
1209
1210         trace_and_count(c, btree_node_set_root, c, b);
1211
1212         old = btree_node_root(c, b);
1213
1214         /*
1215          * Ensure no one is using the old root while we switch to the
1216          * new root:
1217          */
1218         bch2_btree_node_lock_write_nofail(trans, path, &old->c);
1219
1220         bch2_btree_set_root_inmem(c, b);
1221
1222         btree_update_updated_root(as, b);
1223
1224         /*
1225          * Unlock old root after new root is visible:
1226          *
1227          * The new root isn't persistent, but that's ok: we still have
1228          * an intent lock on the new root, and any updates that would
1229          * depend on the new root would have to update the new root.
1230          */
1231         bch2_btree_node_unlock_write(trans, path, old);
1232 }
1233
1234 /* Interior node updates: */
1235
1236 static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
1237                                         struct btree_trans *trans,
1238                                         struct btree_path *path,
1239                                         struct btree *b,
1240                                         struct btree_node_iter *node_iter,
1241                                         struct bkey_i *insert)
1242 {
1243         struct bch_fs *c = as->c;
1244         struct bkey_packed *k;
1245         struct printbuf buf = PRINTBUF;
1246         unsigned long old, new, v;
1247
1248         BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 &&
1249                !btree_ptr_sectors_written(insert));
1250
1251         if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
1252                 bch2_journal_key_overwritten(c, b->c.btree_id, b->c.level, insert->k.p);
1253
1254         if (bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1255                               btree_node_type(b), WRITE, &buf) ?:
1256             bch2_bkey_in_btree_node(c, b, bkey_i_to_s_c(insert), &buf)) {
1257                 printbuf_reset(&buf);
1258                 prt_printf(&buf, "inserting invalid bkey\n  ");
1259                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
1260                 prt_printf(&buf, "\n  ");
1261                 bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1262                                   btree_node_type(b), WRITE, &buf);
1263                 bch2_bkey_in_btree_node(c, b, bkey_i_to_s_c(insert), &buf);
1264
1265                 bch2_fs_inconsistent(c, "%s", buf.buf);
1266                 dump_stack();
1267         }
1268
1269         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1270                ARRAY_SIZE(as->journal_entries));
1271
1272         as->journal_u64s +=
1273                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1274                                   BCH_JSET_ENTRY_btree_keys,
1275                                   b->c.btree_id, b->c.level,
1276                                   insert, insert->k.u64s);
1277
1278         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1279                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1280                 bch2_btree_node_iter_advance(node_iter, b);
1281
1282         bch2_btree_bset_insert_key(trans, path, b, node_iter, insert);
1283         set_btree_node_dirty_acct(c, b);
1284
1285         v = READ_ONCE(b->flags);
1286         do {
1287                 old = new = v;
1288
1289                 new &= ~BTREE_WRITE_TYPE_MASK;
1290                 new |= BTREE_WRITE_interior;
1291                 new |= 1 << BTREE_NODE_need_write;
1292         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1293
1294         printbuf_exit(&buf);
1295 }
1296
1297 static void
1298 __bch2_btree_insert_keys_interior(struct btree_update *as,
1299                                   struct btree_trans *trans,
1300                                   struct btree_path *path,
1301                                   struct btree *b,
1302                                   struct btree_node_iter node_iter,
1303                                   struct keylist *keys)
1304 {
1305         struct bkey_i *insert = bch2_keylist_front(keys);
1306         struct bkey_packed *k;
1307
1308         BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1309
1310         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1311                (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1312                 ;
1313
1314         while (!bch2_keylist_empty(keys)) {
1315                 insert = bch2_keylist_front(keys);
1316
1317                 if (bpos_gt(insert->k.p, b->key.k.p))
1318                         break;
1319
1320                 bch2_insert_fixup_btree_ptr(as, trans, path, b, &node_iter, insert);
1321                 bch2_keylist_pop_front(keys);
1322         }
1323 }
1324
1325 /*
1326  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1327  * node)
1328  */
1329 static void __btree_split_node(struct btree_update *as,
1330                                struct btree_trans *trans,
1331                                struct btree *b,
1332                                struct btree *n[2])
1333 {
1334         struct bkey_packed *k;
1335         struct bpos n1_pos = POS_MIN;
1336         struct btree_node_iter iter;
1337         struct bset *bsets[2];
1338         struct bkey_format_state format[2];
1339         struct bkey_packed *out[2];
1340         struct bkey uk;
1341         unsigned u64s, n1_u64s = (b->nr.live_u64s * 3) / 5;
1342         int i;
1343
1344         for (i = 0; i < 2; i++) {
1345                 BUG_ON(n[i]->nsets != 1);
1346
1347                 bsets[i] = btree_bset_first(n[i]);
1348                 out[i] = bsets[i]->start;
1349
1350                 SET_BTREE_NODE_SEQ(n[i]->data, BTREE_NODE_SEQ(b->data) + 1);
1351                 bch2_bkey_format_init(&format[i]);
1352         }
1353
1354         u64s = 0;
1355         for_each_btree_node_key(b, k, &iter) {
1356                 if (bkey_deleted(k))
1357                         continue;
1358
1359                 i = u64s >= n1_u64s;
1360                 u64s += k->u64s;
1361                 uk = bkey_unpack_key(b, k);
1362                 if (!i)
1363                         n1_pos = uk.p;
1364                 bch2_bkey_format_add_key(&format[i], &uk);
1365         }
1366
1367         btree_set_min(n[0], b->data->min_key);
1368         btree_set_max(n[0], n1_pos);
1369         btree_set_min(n[1], bpos_successor(n1_pos));
1370         btree_set_max(n[1], b->data->max_key);
1371
1372         for (i = 0; i < 2; i++) {
1373                 bch2_bkey_format_add_pos(&format[i], n[i]->data->min_key);
1374                 bch2_bkey_format_add_pos(&format[i], n[i]->data->max_key);
1375
1376                 n[i]->data->format = bch2_bkey_format_done(&format[i]);
1377                 btree_node_set_format(n[i], n[i]->data->format);
1378         }
1379
1380         u64s = 0;
1381         for_each_btree_node_key(b, k, &iter) {
1382                 if (bkey_deleted(k))
1383                         continue;
1384
1385                 i = u64s >= n1_u64s;
1386                 u64s += k->u64s;
1387
1388                 if (bch2_bkey_transform(&n[i]->format, out[i], bkey_packed(k)
1389                                         ? &b->format: &bch2_bkey_format_current, k))
1390                         out[i]->format = KEY_FORMAT_LOCAL_BTREE;
1391                 else
1392                         bch2_bkey_unpack(b, (void *) out[i], k);
1393
1394                 out[i]->needs_whiteout = false;
1395
1396                 btree_keys_account_key_add(&n[i]->nr, 0, out[i]);
1397                 out[i] = bkey_p_next(out[i]);
1398         }
1399
1400         for (i = 0; i < 2; i++) {
1401                 bsets[i]->u64s = cpu_to_le16((u64 *) out[i] - bsets[i]->_data);
1402
1403                 BUG_ON(!bsets[i]->u64s);
1404
1405                 set_btree_bset_end(n[i], n[i]->set);
1406
1407                 btree_node_reset_sib_u64s(n[i]);
1408
1409                 bch2_verify_btree_nr_keys(n[i]);
1410
1411                 if (b->c.level)
1412                         btree_node_interior_verify(as->c, n[i]);
1413         }
1414 }
1415
1416 /*
1417  * For updates to interior nodes, we've got to do the insert before we split
1418  * because the stuff we're inserting has to be inserted atomically. Post split,
1419  * the keys might have to go in different nodes and the split would no longer be
1420  * atomic.
1421  *
1422  * Worse, if the insert is from btree node coalescing, if we do the insert after
1423  * we do the split (and pick the pivot) - the pivot we pick might be between
1424  * nodes that were coalesced, and thus in the middle of a child node post
1425  * coalescing:
1426  */
1427 static void btree_split_insert_keys(struct btree_update *as,
1428                                     struct btree_trans *trans,
1429                                     struct btree_path *path,
1430                                     struct btree *b,
1431                                     struct keylist *keys)
1432 {
1433         if (!bch2_keylist_empty(keys) &&
1434             bpos_le(bch2_keylist_front(keys)->k.p, b->data->max_key)) {
1435                 struct btree_node_iter node_iter;
1436
1437                 bch2_btree_node_iter_init(&node_iter, b, &bch2_keylist_front(keys)->k.p);
1438
1439                 __bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
1440
1441                 btree_node_interior_verify(as->c, b);
1442         }
1443 }
1444
1445 static int btree_split(struct btree_update *as, struct btree_trans *trans,
1446                        struct btree_path *path, struct btree *b,
1447                        struct keylist *keys, unsigned flags)
1448 {
1449         struct bch_fs *c = as->c;
1450         struct btree *parent = btree_node_parent(path, b);
1451         struct btree *n1, *n2 = NULL, *n3 = NULL;
1452         struct btree_path *path1 = NULL, *path2 = NULL;
1453         u64 start_time = local_clock();
1454         int ret = 0;
1455
1456         BUG_ON(!parent && (b != btree_node_root(c, b)));
1457         BUG_ON(parent && !btree_node_intent_locked(path, b->c.level + 1));
1458
1459         bch2_btree_interior_update_will_free_node(as, b);
1460
1461         if (b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c)) {
1462                 struct btree *n[2];
1463
1464                 trace_and_count(c, btree_node_split, c, b);
1465
1466                 n[0] = n1 = bch2_btree_node_alloc(as, trans, b->c.level);
1467                 n[1] = n2 = bch2_btree_node_alloc(as, trans, b->c.level);
1468
1469                 __btree_split_node(as, trans, b, n);
1470
1471                 if (keys) {
1472                         btree_split_insert_keys(as, trans, path, n1, keys);
1473                         btree_split_insert_keys(as, trans, path, n2, keys);
1474                         BUG_ON(!bch2_keylist_empty(keys));
1475                 }
1476
1477                 bch2_btree_build_aux_trees(n2);
1478                 bch2_btree_build_aux_trees(n1);
1479
1480                 bch2_btree_update_add_new_node(as, n1);
1481                 bch2_btree_update_add_new_node(as, n2);
1482                 six_unlock_write(&n2->c.lock);
1483                 six_unlock_write(&n1->c.lock);
1484
1485                 path1 = get_unlocked_mut_path(trans, path->btree_id, n1->c.level, n1->key.k.p);
1486                 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1487                 mark_btree_node_locked(trans, path1, n1->c.level, BTREE_NODE_INTENT_LOCKED);
1488                 bch2_btree_path_level_init(trans, path1, n1);
1489
1490                 path2 = get_unlocked_mut_path(trans, path->btree_id, n2->c.level, n2->key.k.p);
1491                 six_lock_increment(&n2->c.lock, SIX_LOCK_intent);
1492                 mark_btree_node_locked(trans, path2, n2->c.level, BTREE_NODE_INTENT_LOCKED);
1493                 bch2_btree_path_level_init(trans, path2, n2);
1494
1495                 /*
1496                  * Note that on recursive parent_keys == keys, so we
1497                  * can't start adding new keys to parent_keys before emptying it
1498                  * out (which we did with btree_split_insert_keys() above)
1499                  */
1500                 bch2_keylist_add(&as->parent_keys, &n1->key);
1501                 bch2_keylist_add(&as->parent_keys, &n2->key);
1502
1503                 if (!parent) {
1504                         /* Depth increases, make a new root */
1505                         n3 = __btree_root_alloc(as, trans, b->c.level + 1);
1506
1507                         bch2_btree_update_add_new_node(as, n3);
1508                         six_unlock_write(&n3->c.lock);
1509
1510                         path2->locks_want++;
1511                         BUG_ON(btree_node_locked(path2, n3->c.level));
1512                         six_lock_increment(&n3->c.lock, SIX_LOCK_intent);
1513                         mark_btree_node_locked(trans, path2, n3->c.level, BTREE_NODE_INTENT_LOCKED);
1514                         bch2_btree_path_level_init(trans, path2, n3);
1515
1516                         n3->sib_u64s[0] = U16_MAX;
1517                         n3->sib_u64s[1] = U16_MAX;
1518
1519                         btree_split_insert_keys(as, trans, path, n3, &as->parent_keys);
1520                 }
1521         } else {
1522                 trace_and_count(c, btree_node_compact, c, b);
1523
1524                 n1 = bch2_btree_node_alloc_replacement(as, trans, b);
1525
1526                 if (keys) {
1527                         btree_split_insert_keys(as, trans, path, n1, keys);
1528                         BUG_ON(!bch2_keylist_empty(keys));
1529                 }
1530
1531                 bch2_btree_build_aux_trees(n1);
1532                 bch2_btree_update_add_new_node(as, n1);
1533                 six_unlock_write(&n1->c.lock);
1534
1535                 path1 = get_unlocked_mut_path(trans, path->btree_id, n1->c.level, n1->key.k.p);
1536                 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1537                 mark_btree_node_locked(trans, path1, n1->c.level, BTREE_NODE_INTENT_LOCKED);
1538                 bch2_btree_path_level_init(trans, path1, n1);
1539
1540                 if (parent)
1541                         bch2_keylist_add(&as->parent_keys, &n1->key);
1542         }
1543
1544         /* New nodes all written, now make them visible: */
1545
1546         if (parent) {
1547                 /* Split a non root node */
1548                 ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1549                 if (ret)
1550                         goto err;
1551         } else if (n3) {
1552                 bch2_btree_set_root(as, trans, path, n3);
1553         } else {
1554                 /* Root filled up but didn't need to be split */
1555                 bch2_btree_set_root(as, trans, path, n1);
1556         }
1557
1558         if (n3) {
1559                 bch2_btree_update_get_open_buckets(as, n3);
1560                 bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0);
1561         }
1562         if (n2) {
1563                 bch2_btree_update_get_open_buckets(as, n2);
1564                 bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0);
1565         }
1566         bch2_btree_update_get_open_buckets(as, n1);
1567         bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1568
1569         /*
1570          * The old node must be freed (in memory) _before_ unlocking the new
1571          * nodes - else another thread could re-acquire a read lock on the old
1572          * node after another thread has locked and updated the new node, thus
1573          * seeing stale data:
1574          */
1575         bch2_btree_node_free_inmem(trans, path, b);
1576
1577         if (n3)
1578                 bch2_trans_node_add(trans, n3);
1579         if (n2)
1580                 bch2_trans_node_add(trans, n2);
1581         bch2_trans_node_add(trans, n1);
1582
1583         if (n3)
1584                 six_unlock_intent(&n3->c.lock);
1585         if (n2)
1586                 six_unlock_intent(&n2->c.lock);
1587         six_unlock_intent(&n1->c.lock);
1588 out:
1589         if (path2) {
1590                 __bch2_btree_path_unlock(trans, path2);
1591                 bch2_path_put(trans, path2, true);
1592         }
1593         if (path1) {
1594                 __bch2_btree_path_unlock(trans, path1);
1595                 bch2_path_put(trans, path1, true);
1596         }
1597
1598         bch2_trans_verify_locks(trans);
1599
1600         bch2_time_stats_update(&c->times[n2
1601                                ? BCH_TIME_btree_node_split
1602                                : BCH_TIME_btree_node_compact],
1603                                start_time);
1604         return ret;
1605 err:
1606         if (n3)
1607                 bch2_btree_node_free_never_used(as, trans, n3);
1608         if (n2)
1609                 bch2_btree_node_free_never_used(as, trans, n2);
1610         bch2_btree_node_free_never_used(as, trans, n1);
1611         goto out;
1612 }
1613
1614 static void
1615 bch2_btree_insert_keys_interior(struct btree_update *as,
1616                                 struct btree_trans *trans,
1617                                 struct btree_path *path,
1618                                 struct btree *b,
1619                                 struct keylist *keys)
1620 {
1621         struct btree_path *linked;
1622
1623         __bch2_btree_insert_keys_interior(as, trans, path, b,
1624                                           path->l[b->c.level].iter, keys);
1625
1626         btree_update_updated_node(as, b);
1627
1628         trans_for_each_path_with_node(trans, b, linked)
1629                 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1630
1631         bch2_trans_verify_paths(trans);
1632 }
1633
1634 /**
1635  * bch2_btree_insert_node - insert bkeys into a given btree node
1636  *
1637  * @as:                 btree_update object
1638  * @trans:              btree_trans object
1639  * @path:               path that points to current node
1640  * @b:                  node to insert keys into
1641  * @keys:               list of keys to insert
1642  * @flags:              transaction commit flags
1643  *
1644  * Returns: 0 on success, typically transaction restart error on failure
1645  *
1646  * Inserts as many keys as it can into a given btree node, splitting it if full.
1647  * If a split occurred, this function will return early. This can only happen
1648  * for leaf nodes -- inserts into interior nodes have to be atomic.
1649  */
1650 static int bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans,
1651                                   struct btree_path *path, struct btree *b,
1652                                   struct keylist *keys, unsigned flags)
1653 {
1654         struct bch_fs *c = as->c;
1655         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1656         int old_live_u64s = b->nr.live_u64s;
1657         int live_u64s_added, u64s_added;
1658         int ret;
1659
1660         lockdep_assert_held(&c->gc_lock);
1661         BUG_ON(!btree_node_intent_locked(path, b->c.level));
1662         BUG_ON(!b->c.level);
1663         BUG_ON(!as || as->b);
1664         bch2_verify_keylist_sorted(keys);
1665
1666         ret = bch2_btree_node_lock_write(trans, path, &b->c);
1667         if (ret)
1668                 return ret;
1669
1670         bch2_btree_node_prep_for_write(trans, path, b);
1671
1672         if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1673                 bch2_btree_node_unlock_write(trans, path, b);
1674                 goto split;
1675         }
1676
1677         btree_node_interior_verify(c, b);
1678
1679         bch2_btree_insert_keys_interior(as, trans, path, b, keys);
1680
1681         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1682         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1683
1684         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1685                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1686         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1687                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1688
1689         if (u64s_added > live_u64s_added &&
1690             bch2_maybe_compact_whiteouts(c, b))
1691                 bch2_trans_node_reinit_iter(trans, b);
1692
1693         bch2_btree_node_unlock_write(trans, path, b);
1694
1695         btree_node_interior_verify(c, b);
1696         return 0;
1697 split:
1698         /*
1699          * We could attempt to avoid the transaction restart, by calling
1700          * bch2_btree_path_upgrade() and allocating more nodes:
1701          */
1702         if (b->c.level >= as->update_level) {
1703                 trace_and_count(c, trans_restart_split_race, trans, _THIS_IP_, b);
1704                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_split_race);
1705         }
1706
1707         return btree_split(as, trans, path, b, keys, flags);
1708 }
1709
1710 int bch2_btree_split_leaf(struct btree_trans *trans,
1711                           struct btree_path *path,
1712                           unsigned flags)
1713 {
1714         struct btree *b = path_l(path)->b;
1715         struct btree_update *as;
1716         unsigned l;
1717         int ret = 0;
1718
1719         as = bch2_btree_update_start(trans, path, path->level,
1720                                      true, flags);
1721         if (IS_ERR(as))
1722                 return PTR_ERR(as);
1723
1724         ret = btree_split(as, trans, path, b, NULL, flags);
1725         if (ret) {
1726                 bch2_btree_update_free(as, trans);
1727                 return ret;
1728         }
1729
1730         bch2_btree_update_done(as, trans);
1731
1732         for (l = path->level + 1; btree_node_intent_locked(path, l) && !ret; l++)
1733                 ret = bch2_foreground_maybe_merge(trans, path, l, flags);
1734
1735         return ret;
1736 }
1737
1738 int __bch2_foreground_maybe_merge(struct btree_trans *trans,
1739                                   struct btree_path *path,
1740                                   unsigned level,
1741                                   unsigned flags,
1742                                   enum btree_node_sibling sib)
1743 {
1744         struct bch_fs *c = trans->c;
1745         struct btree_path *sib_path = NULL, *new_path = NULL;
1746         struct btree_update *as;
1747         struct bkey_format_state new_s;
1748         struct bkey_format new_f;
1749         struct bkey_i delete;
1750         struct btree *b, *m, *n, *prev, *next, *parent;
1751         struct bpos sib_pos;
1752         size_t sib_u64s;
1753         u64 start_time = local_clock();
1754         int ret = 0;
1755
1756         BUG_ON(!path->should_be_locked);
1757         BUG_ON(!btree_node_locked(path, level));
1758
1759         b = path->l[level].b;
1760
1761         if ((sib == btree_prev_sib && bpos_eq(b->data->min_key, POS_MIN)) ||
1762             (sib == btree_next_sib && bpos_eq(b->data->max_key, SPOS_MAX))) {
1763                 b->sib_u64s[sib] = U16_MAX;
1764                 return 0;
1765         }
1766
1767         sib_pos = sib == btree_prev_sib
1768                 ? bpos_predecessor(b->data->min_key)
1769                 : bpos_successor(b->data->max_key);
1770
1771         sib_path = bch2_path_get(trans, path->btree_id, sib_pos,
1772                                  U8_MAX, level, BTREE_ITER_INTENT, _THIS_IP_);
1773         ret = bch2_btree_path_traverse(trans, sib_path, false);
1774         if (ret)
1775                 goto err;
1776
1777         btree_path_set_should_be_locked(sib_path);
1778
1779         m = sib_path->l[level].b;
1780
1781         if (btree_node_parent(path, b) !=
1782             btree_node_parent(sib_path, m)) {
1783                 b->sib_u64s[sib] = U16_MAX;
1784                 goto out;
1785         }
1786
1787         if (sib == btree_prev_sib) {
1788                 prev = m;
1789                 next = b;
1790         } else {
1791                 prev = b;
1792                 next = m;
1793         }
1794
1795         if (!bpos_eq(bpos_successor(prev->data->max_key), next->data->min_key)) {
1796                 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1797
1798                 bch2_bpos_to_text(&buf1, prev->data->max_key);
1799                 bch2_bpos_to_text(&buf2, next->data->min_key);
1800                 bch_err(c,
1801                         "%s(): btree topology error:\n"
1802                         "  prev ends at   %s\n"
1803                         "  next starts at %s",
1804                         __func__, buf1.buf, buf2.buf);
1805                 printbuf_exit(&buf1);
1806                 printbuf_exit(&buf2);
1807                 bch2_topology_error(c);
1808                 ret = -EIO;
1809                 goto err;
1810         }
1811
1812         bch2_bkey_format_init(&new_s);
1813         bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
1814         __bch2_btree_calc_format(&new_s, prev);
1815         __bch2_btree_calc_format(&new_s, next);
1816         bch2_bkey_format_add_pos(&new_s, next->data->max_key);
1817         new_f = bch2_bkey_format_done(&new_s);
1818
1819         sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1820                 btree_node_u64s_with_format(m, &new_f);
1821
1822         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1823                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1824                 sib_u64s /= 2;
1825                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1826         }
1827
1828         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1829         sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
1830         b->sib_u64s[sib] = sib_u64s;
1831
1832         if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
1833                 goto out;
1834
1835         parent = btree_node_parent(path, b);
1836         as = bch2_btree_update_start(trans, path, level, false,
1837                                      BCH_TRANS_COMMIT_no_enospc|flags);
1838         ret = PTR_ERR_OR_ZERO(as);
1839         if (ret)
1840                 goto err;
1841
1842         trace_and_count(c, btree_node_merge, c, b);
1843
1844         bch2_btree_interior_update_will_free_node(as, b);
1845         bch2_btree_interior_update_will_free_node(as, m);
1846
1847         n = bch2_btree_node_alloc(as, trans, b->c.level);
1848
1849         SET_BTREE_NODE_SEQ(n->data,
1850                            max(BTREE_NODE_SEQ(b->data),
1851                                BTREE_NODE_SEQ(m->data)) + 1);
1852
1853         btree_set_min(n, prev->data->min_key);
1854         btree_set_max(n, next->data->max_key);
1855
1856         n->data->format  = new_f;
1857         btree_node_set_format(n, new_f);
1858
1859         bch2_btree_sort_into(c, n, prev);
1860         bch2_btree_sort_into(c, n, next);
1861
1862         bch2_btree_build_aux_trees(n);
1863         bch2_btree_update_add_new_node(as, n);
1864         six_unlock_write(&n->c.lock);
1865
1866         new_path = get_unlocked_mut_path(trans, path->btree_id, n->c.level, n->key.k.p);
1867         six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1868         mark_btree_node_locked(trans, new_path, n->c.level, BTREE_NODE_INTENT_LOCKED);
1869         bch2_btree_path_level_init(trans, new_path, n);
1870
1871         bkey_init(&delete.k);
1872         delete.k.p = prev->key.k.p;
1873         bch2_keylist_add(&as->parent_keys, &delete);
1874         bch2_keylist_add(&as->parent_keys, &n->key);
1875
1876         bch2_trans_verify_paths(trans);
1877
1878         ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1879         if (ret)
1880                 goto err_free_update;
1881
1882         bch2_trans_verify_paths(trans);
1883
1884         bch2_btree_update_get_open_buckets(as, n);
1885         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1886
1887         bch2_btree_node_free_inmem(trans, path, b);
1888         bch2_btree_node_free_inmem(trans, sib_path, m);
1889
1890         bch2_trans_node_add(trans, n);
1891
1892         bch2_trans_verify_paths(trans);
1893
1894         six_unlock_intent(&n->c.lock);
1895
1896         bch2_btree_update_done(as, trans);
1897
1898         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time);
1899 out:
1900 err:
1901         if (new_path)
1902                 bch2_path_put(trans, new_path, true);
1903         bch2_path_put(trans, sib_path, true);
1904         bch2_trans_verify_locks(trans);
1905         return ret;
1906 err_free_update:
1907         bch2_btree_node_free_never_used(as, trans, n);
1908         bch2_btree_update_free(as, trans);
1909         goto out;
1910 }
1911
1912 int bch2_btree_node_rewrite(struct btree_trans *trans,
1913                             struct btree_iter *iter,
1914                             struct btree *b,
1915                             unsigned flags)
1916 {
1917         struct bch_fs *c = trans->c;
1918         struct btree_path *new_path = NULL;
1919         struct btree *n, *parent;
1920         struct btree_update *as;
1921         int ret;
1922
1923         flags |= BCH_TRANS_COMMIT_no_enospc;
1924
1925         parent = btree_node_parent(iter->path, b);
1926         as = bch2_btree_update_start(trans, iter->path, b->c.level,
1927                                      false, flags);
1928         ret = PTR_ERR_OR_ZERO(as);
1929         if (ret)
1930                 goto out;
1931
1932         bch2_btree_interior_update_will_free_node(as, b);
1933
1934         n = bch2_btree_node_alloc_replacement(as, trans, b);
1935
1936         bch2_btree_build_aux_trees(n);
1937         bch2_btree_update_add_new_node(as, n);
1938         six_unlock_write(&n->c.lock);
1939
1940         new_path = get_unlocked_mut_path(trans, iter->btree_id, n->c.level, n->key.k.p);
1941         six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1942         mark_btree_node_locked(trans, new_path, n->c.level, BTREE_NODE_INTENT_LOCKED);
1943         bch2_btree_path_level_init(trans, new_path, n);
1944
1945         trace_and_count(c, btree_node_rewrite, c, b);
1946
1947         if (parent) {
1948                 bch2_keylist_add(&as->parent_keys, &n->key);
1949                 ret = bch2_btree_insert_node(as, trans, iter->path, parent,
1950                                              &as->parent_keys, flags);
1951                 if (ret)
1952                         goto err;
1953         } else {
1954                 bch2_btree_set_root(as, trans, iter->path, n);
1955         }
1956
1957         bch2_btree_update_get_open_buckets(as, n);
1958         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1959
1960         bch2_btree_node_free_inmem(trans, iter->path, b);
1961
1962         bch2_trans_node_add(trans, n);
1963         six_unlock_intent(&n->c.lock);
1964
1965         bch2_btree_update_done(as, trans);
1966 out:
1967         if (new_path)
1968                 bch2_path_put(trans, new_path, true);
1969         bch2_trans_downgrade(trans);
1970         return ret;
1971 err:
1972         bch2_btree_node_free_never_used(as, trans, n);
1973         bch2_btree_update_free(as, trans);
1974         goto out;
1975 }
1976
1977 struct async_btree_rewrite {
1978         struct bch_fs           *c;
1979         struct work_struct      work;
1980         struct list_head        list;
1981         enum btree_id           btree_id;
1982         unsigned                level;
1983         struct bpos             pos;
1984         __le64                  seq;
1985 };
1986
1987 static int async_btree_node_rewrite_trans(struct btree_trans *trans,
1988                                           struct async_btree_rewrite *a)
1989 {
1990         struct bch_fs *c = trans->c;
1991         struct btree_iter iter;
1992         struct btree *b;
1993         int ret;
1994
1995         bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos,
1996                                   BTREE_MAX_DEPTH, a->level, 0);
1997         b = bch2_btree_iter_peek_node(&iter);
1998         ret = PTR_ERR_OR_ZERO(b);
1999         if (ret)
2000                 goto out;
2001
2002         if (!b || b->data->keys.seq != a->seq) {
2003                 struct printbuf buf = PRINTBUF;
2004
2005                 if (b)
2006                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
2007                 else
2008                         prt_str(&buf, "(null");
2009                 bch_info(c, "%s: node to rewrite not found:, searching for seq %llu, got\n%s",
2010                          __func__, a->seq, buf.buf);
2011                 printbuf_exit(&buf);
2012                 goto out;
2013         }
2014
2015         ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
2016 out:
2017         bch2_trans_iter_exit(trans, &iter);
2018
2019         return ret;
2020 }
2021
2022 static void async_btree_node_rewrite_work(struct work_struct *work)
2023 {
2024         struct async_btree_rewrite *a =
2025                 container_of(work, struct async_btree_rewrite, work);
2026         struct bch_fs *c = a->c;
2027         int ret;
2028
2029         ret = bch2_trans_do(c, NULL, NULL, 0,
2030                       async_btree_node_rewrite_trans(trans, a));
2031         if (ret)
2032                 bch_err_fn(c, ret);
2033         bch2_write_ref_put(c, BCH_WRITE_REF_node_rewrite);
2034         kfree(a);
2035 }
2036
2037 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
2038 {
2039         struct async_btree_rewrite *a;
2040         int ret;
2041
2042         a = kmalloc(sizeof(*a), GFP_NOFS);
2043         if (!a) {
2044                 bch_err(c, "%s: error allocating memory", __func__);
2045                 return;
2046         }
2047
2048         a->c            = c;
2049         a->btree_id     = b->c.btree_id;
2050         a->level        = b->c.level;
2051         a->pos          = b->key.k.p;
2052         a->seq          = b->data->keys.seq;
2053         INIT_WORK(&a->work, async_btree_node_rewrite_work);
2054
2055         if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
2056                 mutex_lock(&c->pending_node_rewrites_lock);
2057                 list_add(&a->list, &c->pending_node_rewrites);
2058                 mutex_unlock(&c->pending_node_rewrites_lock);
2059                 return;
2060         }
2061
2062         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_node_rewrite)) {
2063                 if (test_bit(BCH_FS_STARTED, &c->flags)) {
2064                         bch_err(c, "%s: error getting c->writes ref", __func__);
2065                         kfree(a);
2066                         return;
2067                 }
2068
2069                 ret = bch2_fs_read_write_early(c);
2070                 if (ret) {
2071                         bch_err_msg(c, ret, "going read-write");
2072                         kfree(a);
2073                         return;
2074                 }
2075
2076                 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2077         }
2078
2079         queue_work(c->btree_interior_update_worker, &a->work);
2080 }
2081
2082 void bch2_do_pending_node_rewrites(struct bch_fs *c)
2083 {
2084         struct async_btree_rewrite *a, *n;
2085
2086         mutex_lock(&c->pending_node_rewrites_lock);
2087         list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2088                 list_del(&a->list);
2089
2090                 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2091                 queue_work(c->btree_interior_update_worker, &a->work);
2092         }
2093         mutex_unlock(&c->pending_node_rewrites_lock);
2094 }
2095
2096 void bch2_free_pending_node_rewrites(struct bch_fs *c)
2097 {
2098         struct async_btree_rewrite *a, *n;
2099
2100         mutex_lock(&c->pending_node_rewrites_lock);
2101         list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2102                 list_del(&a->list);
2103
2104                 kfree(a);
2105         }
2106         mutex_unlock(&c->pending_node_rewrites_lock);
2107 }
2108
2109 static int __bch2_btree_node_update_key(struct btree_trans *trans,
2110                                         struct btree_iter *iter,
2111                                         struct btree *b, struct btree *new_hash,
2112                                         struct bkey_i *new_key,
2113                                         unsigned commit_flags,
2114                                         bool skip_triggers)
2115 {
2116         struct bch_fs *c = trans->c;
2117         struct btree_iter iter2 = { NULL };
2118         struct btree *parent;
2119         int ret;
2120
2121         if (!skip_triggers) {
2122                 ret = bch2_trans_mark_old(trans, b->c.btree_id, b->c.level + 1,
2123                                           bkey_i_to_s_c(&b->key), 0);
2124                 if (ret)
2125                         return ret;
2126
2127                 ret = bch2_trans_mark_new(trans, b->c.btree_id, b->c.level + 1,
2128                                           new_key, 0);
2129                 if (ret)
2130                         return ret;
2131         }
2132
2133         if (new_hash) {
2134                 bkey_copy(&new_hash->key, new_key);
2135                 ret = bch2_btree_node_hash_insert(&c->btree_cache,
2136                                 new_hash, b->c.level, b->c.btree_id);
2137                 BUG_ON(ret);
2138         }
2139
2140         parent = btree_node_parent(iter->path, b);
2141         if (parent) {
2142                 bch2_trans_copy_iter(&iter2, iter);
2143
2144                 iter2.path = bch2_btree_path_make_mut(trans, iter2.path,
2145                                 iter2.flags & BTREE_ITER_INTENT,
2146                                 _THIS_IP_);
2147
2148                 BUG_ON(iter2.path->level != b->c.level);
2149                 BUG_ON(!bpos_eq(iter2.path->pos, new_key->k.p));
2150
2151                 btree_path_set_level_up(trans, iter2.path);
2152
2153                 trans->paths_sorted = false;
2154
2155                 ret   = bch2_btree_iter_traverse(&iter2) ?:
2156                         bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_NORUN);
2157                 if (ret)
2158                         goto err;
2159         } else {
2160                 BUG_ON(btree_node_root(c, b) != b);
2161
2162                 ret = darray_make_room(&trans->extra_journal_entries,
2163                                        jset_u64s(new_key->k.u64s));
2164                 if (ret)
2165                         return ret;
2166
2167                 journal_entry_set((void *) &darray_top(trans->extra_journal_entries),
2168                                   BCH_JSET_ENTRY_btree_root,
2169                                   b->c.btree_id, b->c.level,
2170                                   new_key, new_key->k.u64s);
2171                 trans->extra_journal_entries.nr += jset_u64s(new_key->k.u64s);
2172         }
2173
2174         ret = bch2_trans_commit(trans, NULL, NULL, commit_flags);
2175         if (ret)
2176                 goto err;
2177
2178         bch2_btree_node_lock_write_nofail(trans, iter->path, &b->c);
2179
2180         if (new_hash) {
2181                 mutex_lock(&c->btree_cache.lock);
2182                 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
2183                 bch2_btree_node_hash_remove(&c->btree_cache, b);
2184
2185                 bkey_copy(&b->key, new_key);
2186                 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
2187                 BUG_ON(ret);
2188                 mutex_unlock(&c->btree_cache.lock);
2189         } else {
2190                 bkey_copy(&b->key, new_key);
2191         }
2192
2193         bch2_btree_node_unlock_write(trans, iter->path, b);
2194 out:
2195         bch2_trans_iter_exit(trans, &iter2);
2196         return ret;
2197 err:
2198         if (new_hash) {
2199                 mutex_lock(&c->btree_cache.lock);
2200                 bch2_btree_node_hash_remove(&c->btree_cache, b);
2201                 mutex_unlock(&c->btree_cache.lock);
2202         }
2203         goto out;
2204 }
2205
2206 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
2207                                struct btree *b, struct bkey_i *new_key,
2208                                unsigned commit_flags, bool skip_triggers)
2209 {
2210         struct bch_fs *c = trans->c;
2211         struct btree *new_hash = NULL;
2212         struct btree_path *path = iter->path;
2213         struct closure cl;
2214         int ret = 0;
2215
2216         ret = bch2_btree_path_upgrade(trans, path, b->c.level + 1);
2217         if (ret)
2218                 return ret;
2219
2220         closure_init_stack(&cl);
2221
2222         /*
2223          * check btree_ptr_hash_val() after @b is locked by
2224          * btree_iter_traverse():
2225          */
2226         if (btree_ptr_hash_val(new_key) != b->hash_val) {
2227                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2228                 if (ret) {
2229                         ret = drop_locks_do(trans, (closure_sync(&cl), 0));
2230                         if (ret)
2231                                 return ret;
2232                 }
2233
2234                 new_hash = bch2_btree_node_mem_alloc(trans, false);
2235         }
2236
2237         path->intent_ref++;
2238         ret = __bch2_btree_node_update_key(trans, iter, b, new_hash, new_key,
2239                                            commit_flags, skip_triggers);
2240         --path->intent_ref;
2241
2242         if (new_hash) {
2243                 mutex_lock(&c->btree_cache.lock);
2244                 list_move(&new_hash->list, &c->btree_cache.freeable);
2245                 mutex_unlock(&c->btree_cache.lock);
2246
2247                 six_unlock_write(&new_hash->c.lock);
2248                 six_unlock_intent(&new_hash->c.lock);
2249         }
2250         closure_sync(&cl);
2251         bch2_btree_cache_cannibalize_unlock(c);
2252         return ret;
2253 }
2254
2255 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
2256                                         struct btree *b, struct bkey_i *new_key,
2257                                         unsigned commit_flags, bool skip_triggers)
2258 {
2259         struct btree_iter iter;
2260         int ret;
2261
2262         bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p,
2263                                   BTREE_MAX_DEPTH, b->c.level,
2264                                   BTREE_ITER_INTENT);
2265         ret = bch2_btree_iter_traverse(&iter);
2266         if (ret)
2267                 goto out;
2268
2269         /* has node been freed? */
2270         if (iter.path->l[b->c.level].b != b) {
2271                 /* node has been freed: */
2272                 BUG_ON(!btree_node_dying(b));
2273                 goto out;
2274         }
2275
2276         BUG_ON(!btree_node_hashed(b));
2277
2278         ret = bch2_btree_node_update_key(trans, &iter, b, new_key,
2279                                          commit_flags, skip_triggers);
2280 out:
2281         bch2_trans_iter_exit(trans, &iter);
2282         return ret;
2283 }
2284
2285 /* Init code: */
2286
2287 /*
2288  * Only for filesystem bringup, when first reading the btree roots or allocating
2289  * btree roots when initializing a new filesystem:
2290  */
2291 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2292 {
2293         BUG_ON(btree_node_root(c, b));
2294
2295         bch2_btree_set_root_inmem(c, b);
2296 }
2297
2298 static int __bch2_btree_root_alloc(struct btree_trans *trans, enum btree_id id)
2299 {
2300         struct bch_fs *c = trans->c;
2301         struct closure cl;
2302         struct btree *b;
2303         int ret;
2304
2305         closure_init_stack(&cl);
2306
2307         do {
2308                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2309                 closure_sync(&cl);
2310         } while (ret);
2311
2312         b = bch2_btree_node_mem_alloc(trans, false);
2313         bch2_btree_cache_cannibalize_unlock(c);
2314
2315         set_btree_node_fake(b);
2316         set_btree_node_need_rewrite(b);
2317         b->c.level      = 0;
2318         b->c.btree_id   = id;
2319
2320         bkey_btree_ptr_init(&b->key);
2321         b->key.k.p = SPOS_MAX;
2322         *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2323
2324         bch2_bset_init_first(b, &b->data->keys);
2325         bch2_btree_build_aux_trees(b);
2326
2327         b->data->flags = 0;
2328         btree_set_min(b, POS_MIN);
2329         btree_set_max(b, SPOS_MAX);
2330         b->data->format = bch2_btree_calc_format(b);
2331         btree_node_set_format(b, b->data->format);
2332
2333         ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2334                                           b->c.level, b->c.btree_id);
2335         BUG_ON(ret);
2336
2337         bch2_btree_set_root_inmem(c, b);
2338
2339         six_unlock_write(&b->c.lock);
2340         six_unlock_intent(&b->c.lock);
2341         return 0;
2342 }
2343
2344 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2345 {
2346         bch2_trans_run(c, __bch2_btree_root_alloc(trans, id));
2347 }
2348
2349 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2350 {
2351         struct btree_update *as;
2352
2353         mutex_lock(&c->btree_interior_update_lock);
2354         list_for_each_entry(as, &c->btree_interior_update_list, list)
2355                 prt_printf(out, "%p m %u w %u r %u j %llu\n",
2356                        as,
2357                        as->mode,
2358                        as->nodes_written,
2359                        closure_nr_remaining(&as->cl),
2360                        as->journal.seq);
2361         mutex_unlock(&c->btree_interior_update_lock);
2362 }
2363
2364 static bool bch2_btree_interior_updates_pending(struct bch_fs *c)
2365 {
2366         bool ret;
2367
2368         mutex_lock(&c->btree_interior_update_lock);
2369         ret = !list_empty(&c->btree_interior_update_list);
2370         mutex_unlock(&c->btree_interior_update_lock);
2371
2372         return ret;
2373 }
2374
2375 bool bch2_btree_interior_updates_flush(struct bch_fs *c)
2376 {
2377         bool ret = bch2_btree_interior_updates_pending(c);
2378
2379         if (ret)
2380                 closure_wait_event(&c->btree_interior_update_wait,
2381                                    !bch2_btree_interior_updates_pending(c));
2382         return ret;
2383 }
2384
2385 void bch2_journal_entry_to_btree_root(struct bch_fs *c, struct jset_entry *entry)
2386 {
2387         struct btree_root *r = bch2_btree_id_root(c, entry->btree_id);
2388
2389         mutex_lock(&c->btree_root_lock);
2390
2391         r->level = entry->level;
2392         r->alive = true;
2393         bkey_copy(&r->key, (struct bkey_i *) entry->start);
2394
2395         mutex_unlock(&c->btree_root_lock);
2396 }
2397
2398 struct jset_entry *
2399 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2400                                     struct jset_entry *end,
2401                                     unsigned long skip)
2402 {
2403         unsigned i;
2404
2405         mutex_lock(&c->btree_root_lock);
2406
2407         for (i = 0; i < btree_id_nr_alive(c); i++) {
2408                 struct btree_root *r = bch2_btree_id_root(c, i);
2409
2410                 if (r->alive && !test_bit(i, &skip)) {
2411                         journal_entry_set(end, BCH_JSET_ENTRY_btree_root,
2412                                           i, r->level, &r->key, r->key.k.u64s);
2413                         end = vstruct_next(end);
2414                 }
2415         }
2416
2417         mutex_unlock(&c->btree_root_lock);
2418
2419         return end;
2420 }
2421
2422 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2423 {
2424         if (c->btree_interior_update_worker)
2425                 destroy_workqueue(c->btree_interior_update_worker);
2426         mempool_exit(&c->btree_interior_update_pool);
2427 }
2428
2429 void bch2_fs_btree_interior_update_init_early(struct bch_fs *c)
2430 {
2431         mutex_init(&c->btree_reserve_cache_lock);
2432         INIT_LIST_HEAD(&c->btree_interior_update_list);
2433         INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2434         mutex_init(&c->btree_interior_update_lock);
2435         INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2436
2437         INIT_LIST_HEAD(&c->pending_node_rewrites);
2438         mutex_init(&c->pending_node_rewrites_lock);
2439 }
2440
2441 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2442 {
2443         c->btree_interior_update_worker =
2444                 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2445         if (!c->btree_interior_update_worker)
2446                 return -BCH_ERR_ENOMEM_btree_interior_update_worker_init;
2447
2448         if (mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2449                                       sizeof(struct btree_update)))
2450                 return -BCH_ERR_ENOMEM_btree_interior_update_pool_init;
2451
2452         return 0;
2453 }