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 btree_trans *trans, struct btree *b)
 167 {
 168         struct bch_fs *c = trans->c;
 169
 170         trace_and_count(c, btree_node_free, trans, b);
 171
 172         BUG_ON(btree_node_write_blocked(b));
 173         BUG_ON(btree_node_dirty(b));
 174         BUG_ON(btree_node_need_write(b));
 175         BUG_ON(b == btree_node_root(c, b));
 176         BUG_ON(b->ob.nr);
 177         BUG_ON(!list_empty(&b->write_blocked));
 178         BUG_ON(b->will_make_reachable);
 179
 180         clear_btree_node_noevict(b);
 181
 182         mutex_lock(&c->btree_cache.lock);
 183         list_move(&b->list, &c->btree_cache.freeable);
 184         mutex_unlock(&c->btree_cache.lock);
 185 }
 186
 187 static void bch2_btree_node_free_inmem(struct btree_trans *trans,
 188                                        struct btree_path *path,
 189                                        struct btree *b)
 190 {
 191         struct bch_fs *c = trans->c;
 192         unsigned level = b->c.level;
 193
 194         bch2_btree_node_lock_write_nofail(trans, path, &b->c);
 195         bch2_btree_node_hash_remove(&c->btree_cache, b);
 196         __btree_node_free(trans, b);
 197         six_unlock_write(&b->c.lock);
 198         mark_btree_node_locked_noreset(path, level, BTREE_NODE_INTENT_LOCKED);
 199
 200         trans_for_each_path(trans, path)
 201                 if (path->l[level].b == b) {
 202                         btree_node_unlock(trans, path, level);
 203                         path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
 204                 }
 205 }
 206
 207 static void bch2_btree_node_free_never_used(struct btree_update *as,
 208                                             struct btree_trans *trans,
 209                                             struct btree *b)
 210 {
 211         struct bch_fs *c = as->c;
 212         struct prealloc_nodes *p = &as->prealloc_nodes[b->c.lock.readers != NULL];
 213         struct btree_path *path;
 214         unsigned level = b->c.level;
 215
 216         BUG_ON(!list_empty(&b->write_blocked));
 217         BUG_ON(b->will_make_reachable != (1UL|(unsigned long) as));
 218
 219         b->will_make_reachable = 0;
 220         closure_put(&as->cl);
 221
 222         clear_btree_node_will_make_reachable(b);
 223         clear_btree_node_accessed(b);
 224         clear_btree_node_dirty_acct(c, b);
 225         clear_btree_node_need_write(b);
 226
 227         mutex_lock(&c->btree_cache.lock);
 228         list_del_init(&b->list);
 229         bch2_btree_node_hash_remove(&c->btree_cache, b);
 230         mutex_unlock(&c->btree_cache.lock);
 231
 232         BUG_ON(p->nr >= ARRAY_SIZE(p->b));
 233         p->b[p->nr++] = b;
 234
 235         six_unlock_intent(&b->c.lock);
 236
 237         trans_for_each_path(trans, path)
 238                 if (path->l[level].b == b) {
 239                         btree_node_unlock(trans, path, level);
 240                         path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
 241                 }
 242 }
 243
 244 static struct btree *__bch2_btree_node_alloc(struct btree_trans *trans,
 245                                              struct disk_reservation *res,
 246                                              struct closure *cl,
 247                                              bool interior_node,
 248                                              unsigned flags)
 249 {
 250         struct bch_fs *c = trans->c;
 251         struct write_point *wp;
 252         struct btree *b;
 253         BKEY_PADDED_ONSTACK(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
 254         struct open_buckets obs = { .nr = 0 };
 255         struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
 256         enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
 257         unsigned nr_reserve = watermark > BCH_WATERMARK_reclaim
 258                 ? BTREE_NODE_RESERVE
 259                 : 0;
 260         int ret;
 261
 262         mutex_lock(&c->btree_reserve_cache_lock);
 263         if (c->btree_reserve_cache_nr > nr_reserve) {
 264                 struct btree_alloc *a =
 265                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
 266
 267                 obs = a->ob;
 268                 bkey_copy(&tmp.k, &a->k);
 269                 mutex_unlock(&c->btree_reserve_cache_lock);
 270                 goto mem_alloc;
 271         }
 272         mutex_unlock(&c->btree_reserve_cache_lock);
 273
 274 retry:
 275         ret = bch2_alloc_sectors_start_trans(trans,
 276                                       c->opts.metadata_target ?:
 277                                       c->opts.foreground_target,
 278                                       0,
 279                                       writepoint_ptr(&c->btree_write_point),
 280                                       &devs_have,
 281                                       res->nr_replicas,
 282                                       c->opts.metadata_replicas_required,
 283                                       watermark, 0, cl, &wp);
 284         if (unlikely(ret))
 285                 return ERR_PTR(ret);
 286
 287         if (wp->sectors_free < btree_sectors(c)) {
 288                 struct open_bucket *ob;
 289                 unsigned i;
 290
 291                 open_bucket_for_each(c, &wp->ptrs, ob, i)
 292                         if (ob->sectors_free < btree_sectors(c))
 293                                 ob->sectors_free = 0;
 294
 295                 bch2_alloc_sectors_done(c, wp);
 296                 goto retry;
 297         }
 298
 299         bkey_btree_ptr_v2_init(&tmp.k);
 300         bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, btree_sectors(c), false);
 301
 302         bch2_open_bucket_get(c, wp, &obs);
 303         bch2_alloc_sectors_done(c, wp);
 304 mem_alloc:
 305         b = bch2_btree_node_mem_alloc(trans, interior_node);
 306         six_unlock_write(&b->c.lock);
 307         six_unlock_intent(&b->c.lock);
 308
 309         /* we hold cannibalize_lock: */
 310         BUG_ON(IS_ERR(b));
 311         BUG_ON(b->ob.nr);
 312
 313         bkey_copy(&b->key, &tmp.k);
 314         b->ob = obs;
 315
 316         return b;
 317 }
 318
 319 static struct btree *bch2_btree_node_alloc(struct btree_update *as,
 320                                            struct btree_trans *trans,
 321                                            unsigned level)
 322 {
 323         struct bch_fs *c = as->c;
 324         struct btree *b;
 325         struct prealloc_nodes *p = &as->prealloc_nodes[!!level];
 326         int ret;
 327
 328         BUG_ON(level >= BTREE_MAX_DEPTH);
 329         BUG_ON(!p->nr);
 330
 331         b = p->b[--p->nr];
 332
 333         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
 334         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
 335
 336         set_btree_node_accessed(b);
 337         set_btree_node_dirty_acct(c, b);
 338         set_btree_node_need_write(b);
 339
 340         bch2_bset_init_first(b, &b->data->keys);
 341         b->c.level      = level;
 342         b->c.btree_id   = as->btree_id;
 343         b->version_ondisk = c->sb.version;
 344
 345         memset(&b->nr, 0, sizeof(b->nr));
 346         b->data->magic = cpu_to_le64(bset_magic(c));
 347         memset(&b->data->_ptr, 0, sizeof(b->data->_ptr));
 348         b->data->flags = 0;
 349         SET_BTREE_NODE_ID(b->data, as->btree_id);
 350         SET_BTREE_NODE_LEVEL(b->data, level);
 351
 352         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 353                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
 354
 355                 bp->v.mem_ptr           = 0;
 356                 bp->v.seq               = b->data->keys.seq;
 357                 bp->v.sectors_written   = 0;
 358         }
 359
 360         SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
 361
 362         bch2_btree_build_aux_trees(b);
 363
 364         ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
 365         BUG_ON(ret);
 366
 367         trace_and_count(c, btree_node_alloc, trans, b);
 368         bch2_increment_clock(c, btree_sectors(c), WRITE);
 369         return b;
 370 }
 371
 372 static void btree_set_min(struct btree *b, struct bpos pos)
 373 {
 374         if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
 375                 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
 376         b->data->min_key = pos;
 377 }
 378
 379 static void btree_set_max(struct btree *b, struct bpos pos)
 380 {
 381         b->key.k.p = pos;
 382         b->data->max_key = pos;
 383 }
 384
 385 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
 386                                                        struct btree_trans *trans,
 387                                                        struct btree *b)
 388 {
 389         struct btree *n = bch2_btree_node_alloc(as, trans, b->c.level);
 390         struct bkey_format format = bch2_btree_calc_format(b);
 391
 392         /*
 393          * The keys might expand with the new format - if they wouldn't fit in
 394          * the btree node anymore, use the old format for now:
 395          */
 396         if (!bch2_btree_node_format_fits(as->c, b, &format))
 397                 format = b->format;
 398
 399         SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
 400
 401         btree_set_min(n, b->data->min_key);
 402         btree_set_max(n, b->data->max_key);
 403
 404         n->data->format         = format;
 405         btree_node_set_format(n, format);
 406
 407         bch2_btree_sort_into(as->c, n, b);
 408
 409         btree_node_reset_sib_u64s(n);
 410         return n;
 411 }
 412
 413 static struct btree *__btree_root_alloc(struct btree_update *as,
 414                                 struct btree_trans *trans, unsigned level)
 415 {
 416         struct btree *b = bch2_btree_node_alloc(as, trans, level);
 417
 418         btree_set_min(b, POS_MIN);
 419         btree_set_max(b, SPOS_MAX);
 420         b->data->format = bch2_btree_calc_format(b);
 421
 422         btree_node_set_format(b, b->data->format);
 423         bch2_btree_build_aux_trees(b);
 424
 425         return b;
 426 }
 427
 428 static void bch2_btree_reserve_put(struct btree_update *as, struct btree_trans *trans)
 429 {
 430         struct bch_fs *c = as->c;
 431         struct prealloc_nodes *p;
 432
 433         for (p = as->prealloc_nodes;
 434              p < as->prealloc_nodes + ARRAY_SIZE(as->prealloc_nodes);
 435              p++) {
 436                 while (p->nr) {
 437                         struct btree *b = p->b[--p->nr];
 438
 439                         mutex_lock(&c->btree_reserve_cache_lock);
 440
 441                         if (c->btree_reserve_cache_nr <
 442                             ARRAY_SIZE(c->btree_reserve_cache)) {
 443                                 struct btree_alloc *a =
 444                                         &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
 445
 446                                 a->ob = b->ob;
 447                                 b->ob.nr = 0;
 448                                 bkey_copy(&a->k, &b->key);
 449                         } else {
 450                                 bch2_open_buckets_put(c, &b->ob);
 451                         }
 452
 453                         mutex_unlock(&c->btree_reserve_cache_lock);
 454
 455                         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
 456                         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
 457                         __btree_node_free(trans, b);
 458                         six_unlock_write(&b->c.lock);
 459                         six_unlock_intent(&b->c.lock);
 460                 }
 461         }
 462 }
 463
 464 static int bch2_btree_reserve_get(struct btree_trans *trans,
 465                                   struct btree_update *as,
 466                                   unsigned nr_nodes[2],
 467                                   unsigned flags,
 468                                   struct closure *cl)
 469 {
 470         struct btree *b;
 471         unsigned interior;
 472         int ret = 0;
 473
 474         BUG_ON(nr_nodes[0] + nr_nodes[1] > BTREE_RESERVE_MAX);
 475
 476         /*
 477          * Protects reaping from the btree node cache and using the btree node
 478          * open bucket reserve:
 479          */
 480         ret = bch2_btree_cache_cannibalize_lock(trans, cl);
 481         if (ret)
 482                 return ret;
 483
 484         for (interior = 0; interior < 2; interior++) {
 485                 struct prealloc_nodes *p = as->prealloc_nodes + interior;
 486
 487                 while (p->nr < nr_nodes[interior]) {
 488                         b = __bch2_btree_node_alloc(trans, &as->disk_res, cl,
 489                                                     interior, flags);
 490                         if (IS_ERR(b)) {
 491                                 ret = PTR_ERR(b);
 492                                 goto err;
 493                         }
 494
 495                         p->b[p->nr++] = b;
 496                 }
 497         }
 498 err:
 499         bch2_btree_cache_cannibalize_unlock(trans);
 500         return ret;
 501 }
 502
 503 /* Asynchronous interior node update machinery */
 504
 505 static void bch2_btree_update_free(struct btree_update *as, struct btree_trans *trans)
 506 {
 507         struct bch_fs *c = as->c;
 508
 509         if (as->took_gc_lock)
 510                 up_read(&c->gc_lock);
 511         as->took_gc_lock = false;
 512
 513         bch2_journal_pin_drop(&c->journal, &as->journal);
 514         bch2_journal_pin_flush(&c->journal, &as->journal);
 515         bch2_disk_reservation_put(c, &as->disk_res);
 516         bch2_btree_reserve_put(as, trans);
 517
 518         bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_total],
 519                                as->start_time);
 520
 521         mutex_lock(&c->btree_interior_update_lock);
 522         list_del(&as->unwritten_list);
 523         list_del(&as->list);
 524
 525         closure_debug_destroy(&as->cl);
 526         mempool_free(as, &c->btree_interior_update_pool);
 527
 528         /*
 529          * Have to do the wakeup with btree_interior_update_lock still held,
 530          * since being on btree_interior_update_list is our ref on @c:
 531          */
 532         closure_wake_up(&c->btree_interior_update_wait);
 533
 534         mutex_unlock(&c->btree_interior_update_lock);
 535 }
 536
 537 static void btree_update_add_key(struct btree_update *as,
 538                                  struct keylist *keys, struct btree *b)
 539 {
 540         struct bkey_i *k = &b->key;
 541
 542         BUG_ON(bch2_keylist_u64s(keys) + k->k.u64s >
 543                ARRAY_SIZE(as->_old_keys));
 544
 545         bkey_copy(keys->top, k);
 546         bkey_i_to_btree_ptr_v2(keys->top)->v.mem_ptr = b->c.level + 1;
 547
 548         bch2_keylist_push(keys);
 549 }
 550
 551 /*
 552  * The transactional part of an interior btree node update, where we journal the
 553  * update we did to the interior node and update alloc info:
 554  */
 555 static int btree_update_nodes_written_trans(struct btree_trans *trans,
 556                                             struct btree_update *as)
 557 {
 558         struct bkey_i *k;
 559         int ret;
 560
 561         ret = darray_make_room(&trans->extra_journal_entries, as->journal_u64s);
 562         if (ret)
 563                 return ret;
 564
 565         memcpy(&darray_top(trans->extra_journal_entries),
 566                as->journal_entries,
 567                as->journal_u64s * sizeof(u64));
 568         trans->extra_journal_entries.nr += as->journal_u64s;
 569
 570         trans->journal_pin = &as->journal;
 571
 572         for_each_keylist_key(&as->old_keys, k) {
 573                 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
 574
 575                 ret = bch2_trans_mark_old(trans, as->btree_id, level, bkey_i_to_s_c(k), 0);
 576                 if (ret)
 577                         return ret;
 578         }
 579
 580         for_each_keylist_key(&as->new_keys, k) {
 581                 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
 582
 583                 ret = bch2_trans_mark_new(trans, as->btree_id, level, k, 0);
 584                 if (ret)
 585                         return ret;
 586         }
 587
 588         return 0;
 589 }
 590
 591 static void btree_update_nodes_written(struct btree_update *as)
 592 {
 593         struct bch_fs *c = as->c;
 594         struct btree *b;
 595         struct btree_trans *trans = bch2_trans_get(c);
 596         u64 journal_seq = 0;
 597         unsigned i;
 598         int ret;
 599
 600         /*
 601          * If we're already in an error state, it might be because a btree node
 602          * was never written, and we might be trying to free that same btree
 603          * node here, but it won't have been marked as allocated and we'll see
 604          * spurious disk usage inconsistencies in the transactional part below
 605          * if we don't skip it:
 606          */
 607         ret = bch2_journal_error(&c->journal);
 608         if (ret)
 609                 goto err;
 610
 611         /*
 612          * Wait for any in flight writes to finish before we free the old nodes
 613          * on disk:
 614          */
 615         for (i = 0; i < as->nr_old_nodes; i++) {
 616                 __le64 seq;
 617
 618                 b = as->old_nodes[i];
 619
 620                 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
 621                 seq = b->data ? b->data->keys.seq : 0;
 622                 six_unlock_read(&b->c.lock);
 623
 624                 if (seq == as->old_nodes_seq[i])
 625                         wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight_inner,
 626                                        TASK_UNINTERRUPTIBLE);
 627         }
 628
 629         /*
 630          * We did an update to a parent node where the pointers we added pointed
 631          * to child nodes that weren't written yet: now, the child nodes have
 632          * been written so we can write out the update to the interior node.
 633          */
 634
 635         /*
 636          * We can't call into journal reclaim here: we'd block on the journal
 637          * reclaim lock, but we may need to release the open buckets we have
 638          * pinned in order for other btree updates to make forward progress, and
 639          * journal reclaim does btree updates when flushing bkey_cached entries,
 640          * which may require allocations as well.
 641          */
 642         ret = commit_do(trans, &as->disk_res, &journal_seq,
 643                         BCH_WATERMARK_reclaim|
 644                         BCH_TRANS_COMMIT_no_enospc|
 645                         BCH_TRANS_COMMIT_no_check_rw|
 646                         BCH_TRANS_COMMIT_journal_reclaim,
 647                         btree_update_nodes_written_trans(trans, as));
 648         bch2_trans_unlock(trans);
 649
 650         bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
 651                              "%s(): error %s", __func__, bch2_err_str(ret));
 652 err:
 653         if (as->b) {
 654                 struct btree_path *path;
 655
 656                 b = as->b;
 657                 path = get_unlocked_mut_path(trans, as->btree_id, b->c.level, b->key.k.p);
 658                 /*
 659                  * @b is the node we did the final insert into:
 660                  *
 661                  * On failure to get a journal reservation, we still have to
 662                  * unblock the write and allow most of the write path to happen
 663                  * so that shutdown works, but the i->journal_seq mechanism
 664                  * won't work to prevent the btree write from being visible (we
 665                  * didn't get a journal sequence number) - instead
 666                  * __bch2_btree_node_write() doesn't do the actual write if
 667                  * we're in journal error state:
 668                  */
 669
 670                 /*
 671                  * Ensure transaction is unlocked before using
 672                  * btree_node_lock_nopath() (the use of which is always suspect,
 673                  * we need to work on removing this in the future)
 674                  *
 675                  * It should be, but get_unlocked_mut_path() -> bch2_path_get()
 676                  * calls bch2_path_upgrade(), before we call path_make_mut(), so
 677                  * we may rarely end up with a locked path besides the one we
 678                  * have here:
 679                  */
 680                 bch2_trans_unlock(trans);
 681                 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
 682                 mark_btree_node_locked(trans, path, b->c.level, BTREE_NODE_INTENT_LOCKED);
 683                 path->l[b->c.level].lock_seq = six_lock_seq(&b->c.lock);
 684                 path->l[b->c.level].b = b;
 685
 686                 bch2_btree_node_lock_write_nofail(trans, path, &b->c);
 687
 688                 mutex_lock(&c->btree_interior_update_lock);
 689
 690                 list_del(&as->write_blocked_list);
 691                 if (list_empty(&b->write_blocked))
 692                         clear_btree_node_write_blocked(b);
 693
 694                 /*
 695                  * Node might have been freed, recheck under
 696                  * btree_interior_update_lock:
 697                  */
 698                 if (as->b == b) {
 699                         BUG_ON(!b->c.level);
 700                         BUG_ON(!btree_node_dirty(b));
 701
 702                         if (!ret) {
 703                                 struct bset *last = btree_bset_last(b);
 704
 705                                 last->journal_seq = cpu_to_le64(
 706                                                              max(journal_seq,
 707                                                                  le64_to_cpu(last->journal_seq)));
 708
 709                                 bch2_btree_add_journal_pin(c, b, journal_seq);
 710                         } else {
 711                                 /*
 712                                  * If we didn't get a journal sequence number we
 713                                  * can't write this btree node, because recovery
 714                                  * won't know to ignore this write:
 715                                  */
 716                                 set_btree_node_never_write(b);
 717                         }
 718                 }
 719
 720                 mutex_unlock(&c->btree_interior_update_lock);
 721
 722                 mark_btree_node_locked_noreset(path, b->c.level, BTREE_NODE_INTENT_LOCKED);
 723                 six_unlock_write(&b->c.lock);
 724
 725                 btree_node_write_if_need(c, b, SIX_LOCK_intent);
 726                 btree_node_unlock(trans, path, b->c.level);
 727                 bch2_path_put(trans, path, true);
 728         }
 729
 730         bch2_journal_pin_drop(&c->journal, &as->journal);
 731
 732         mutex_lock(&c->btree_interior_update_lock);
 733         for (i = 0; i < as->nr_new_nodes; i++) {
 734                 b = as->new_nodes[i];
 735
 736                 BUG_ON(b->will_make_reachable != (unsigned long) as);
 737                 b->will_make_reachable = 0;
 738                 clear_btree_node_will_make_reachable(b);
 739         }
 740         mutex_unlock(&c->btree_interior_update_lock);
 741
 742         for (i = 0; i < as->nr_new_nodes; i++) {
 743                 b = as->new_nodes[i];
 744
 745                 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
 746                 btree_node_write_if_need(c, b, SIX_LOCK_read);
 747                 six_unlock_read(&b->c.lock);
 748         }
 749
 750         for (i = 0; i < as->nr_open_buckets; i++)
 751                 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
 752
 753         bch2_btree_update_free(as, trans);
 754         bch2_trans_put(trans);
 755 }
 756
 757 static void btree_interior_update_work(struct work_struct *work)
 758 {
 759         struct bch_fs *c =
 760                 container_of(work, struct bch_fs, btree_interior_update_work);
 761         struct btree_update *as;
 762
 763         while (1) {
 764                 mutex_lock(&c->btree_interior_update_lock);
 765                 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
 766                                               struct btree_update, unwritten_list);
 767                 if (as && !as->nodes_written)
 768                         as = NULL;
 769                 mutex_unlock(&c->btree_interior_update_lock);
 770
 771                 if (!as)
 772                         break;
 773
 774                 btree_update_nodes_written(as);
 775         }
 776 }
 777
 778 static CLOSURE_CALLBACK(btree_update_set_nodes_written)
 779 {
 780         closure_type(as, struct btree_update, cl);
 781         struct bch_fs *c = as->c;
 782
 783         mutex_lock(&c->btree_interior_update_lock);
 784         as->nodes_written = true;
 785         mutex_unlock(&c->btree_interior_update_lock);
 786
 787         queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
 788 }
 789
 790 /*
 791  * We're updating @b with pointers to nodes that haven't finished writing yet:
 792  * block @b from being written until @as completes
 793  */
 794 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
 795 {
 796         struct bch_fs *c = as->c;
 797
 798         mutex_lock(&c->btree_interior_update_lock);
 799         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 800
 801         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 802         BUG_ON(!btree_node_dirty(b));
 803         BUG_ON(!b->c.level);
 804
 805         as->mode        = BTREE_INTERIOR_UPDATING_NODE;
 806         as->b           = b;
 807
 808         set_btree_node_write_blocked(b);
 809         list_add(&as->write_blocked_list, &b->write_blocked);
 810
 811         mutex_unlock(&c->btree_interior_update_lock);
 812 }
 813
 814 static int bch2_update_reparent_journal_pin_flush(struct journal *j,
 815                                 struct journal_entry_pin *_pin, u64 seq)
 816 {
 817         return 0;
 818 }
 819
 820 static void btree_update_reparent(struct btree_update *as,
 821                                   struct btree_update *child)
 822 {
 823         struct bch_fs *c = as->c;
 824
 825         lockdep_assert_held(&c->btree_interior_update_lock);
 826
 827         child->b = NULL;
 828         child->mode = BTREE_INTERIOR_UPDATING_AS;
 829
 830         bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal,
 831                               bch2_update_reparent_journal_pin_flush);
 832 }
 833
 834 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
 835 {
 836         struct bkey_i *insert = &b->key;
 837         struct bch_fs *c = as->c;
 838
 839         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 840
 841         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
 842                ARRAY_SIZE(as->journal_entries));
 843
 844         as->journal_u64s +=
 845                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
 846                                   BCH_JSET_ENTRY_btree_root,
 847                                   b->c.btree_id, b->c.level,
 848                                   insert, insert->k.u64s);
 849
 850         mutex_lock(&c->btree_interior_update_lock);
 851         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 852
 853         as->mode        = BTREE_INTERIOR_UPDATING_ROOT;
 854         mutex_unlock(&c->btree_interior_update_lock);
 855 }
 856
 857 /*
 858  * bch2_btree_update_add_new_node:
 859  *
 860  * This causes @as to wait on @b to be written, before it gets to
 861  * bch2_btree_update_nodes_written
 862  *
 863  * Additionally, it sets b->will_make_reachable to prevent any additional writes
 864  * to @b from happening besides the first until @b is reachable on disk
 865  *
 866  * And it adds @b to the list of @as's new nodes, so that we can update sector
 867  * counts in bch2_btree_update_nodes_written:
 868  */
 869 static void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
 870 {
 871         struct bch_fs *c = as->c;
 872
 873         closure_get(&as->cl);
 874
 875         mutex_lock(&c->btree_interior_update_lock);
 876         BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
 877         BUG_ON(b->will_make_reachable);
 878
 879         as->new_nodes[as->nr_new_nodes++] = b;
 880         b->will_make_reachable = 1UL|(unsigned long) as;
 881         set_btree_node_will_make_reachable(b);
 882
 883         mutex_unlock(&c->btree_interior_update_lock);
 884
 885         btree_update_add_key(as, &as->new_keys, b);
 886
 887         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 888                 unsigned bytes = vstruct_end(&b->data->keys) - (void *) b->data;
 889                 unsigned sectors = round_up(bytes, block_bytes(c)) >> 9;
 890
 891                 bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written =
 892                         cpu_to_le16(sectors);
 893         }
 894 }
 895
 896 /*
 897  * returns true if @b was a new node
 898  */
 899 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
 900 {
 901         struct btree_update *as;
 902         unsigned long v;
 903         unsigned i;
 904
 905         mutex_lock(&c->btree_interior_update_lock);
 906         /*
 907          * When b->will_make_reachable != 0, it owns a ref on as->cl that's
 908          * dropped when it gets written by bch2_btree_complete_write - the
 909          * xchg() is for synchronization with bch2_btree_complete_write:
 910          */
 911         v = xchg(&b->will_make_reachable, 0);
 912         clear_btree_node_will_make_reachable(b);
 913         as = (struct btree_update *) (v & ~1UL);
 914
 915         if (!as) {
 916                 mutex_unlock(&c->btree_interior_update_lock);
 917                 return;
 918         }
 919
 920         for (i = 0; i < as->nr_new_nodes; i++)
 921                 if (as->new_nodes[i] == b)
 922                         goto found;
 923
 924         BUG();
 925 found:
 926         array_remove_item(as->new_nodes, as->nr_new_nodes, i);
 927         mutex_unlock(&c->btree_interior_update_lock);
 928
 929         if (v & 1)
 930                 closure_put(&as->cl);
 931 }
 932
 933 static void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
 934 {
 935         while (b->ob.nr)
 936                 as->open_buckets[as->nr_open_buckets++] =
 937                         b->ob.v[--b->ob.nr];
 938 }
 939
 940 static int bch2_btree_update_will_free_node_journal_pin_flush(struct journal *j,
 941                                 struct journal_entry_pin *_pin, u64 seq)
 942 {
 943         return 0;
 944 }
 945
 946 /*
 947  * @b is being split/rewritten: it may have pointers to not-yet-written btree
 948  * nodes and thus outstanding btree_updates - redirect @b's
 949  * btree_updates to point to this btree_update:
 950  */
 951 static void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 952                                                       struct btree *b)
 953 {
 954         struct bch_fs *c = as->c;
 955         struct btree_update *p, *n;
 956         struct btree_write *w;
 957
 958         set_btree_node_dying(b);
 959
 960         if (btree_node_fake(b))
 961                 return;
 962
 963         mutex_lock(&c->btree_interior_update_lock);
 964
 965         /*
 966          * Does this node have any btree_update operations preventing
 967          * it from being written?
 968          *
 969          * If so, redirect them to point to this btree_update: we can
 970          * write out our new nodes, but we won't make them visible until those
 971          * operations complete
 972          */
 973         list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
 974                 list_del_init(&p->write_blocked_list);
 975                 btree_update_reparent(as, p);
 976
 977                 /*
 978                  * for flush_held_btree_writes() waiting on updates to flush or
 979                  * nodes to be writeable:
 980                  */
 981                 closure_wake_up(&c->btree_interior_update_wait);
 982         }
 983
 984         clear_btree_node_dirty_acct(c, b);
 985         clear_btree_node_need_write(b);
 986         clear_btree_node_write_blocked(b);
 987
 988         /*
 989          * Does this node have unwritten data that has a pin on the journal?
 990          *
 991          * If so, transfer that pin to the btree_update operation -
 992          * note that if we're freeing multiple nodes, we only need to keep the
 993          * oldest pin of any of the nodes we're freeing. We'll release the pin
 994          * when the new nodes are persistent and reachable on disk:
 995          */
 996         w = btree_current_write(b);
 997         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal,
 998                               bch2_btree_update_will_free_node_journal_pin_flush);
 999         bch2_journal_pin_drop(&c->journal, &w->journal);
1000
1001         w = btree_prev_write(b);
1002         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal,
1003                               bch2_btree_update_will_free_node_journal_pin_flush);
1004         bch2_journal_pin_drop(&c->journal, &w->journal);
1005
1006         mutex_unlock(&c->btree_interior_update_lock);
1007
1008         /*
1009          * Is this a node that isn't reachable on disk yet?
1010          *
1011          * Nodes that aren't reachable yet have writes blocked until they're
1012          * reachable - now that we've cancelled any pending writes and moved
1013          * things waiting on that write to wait on this update, we can drop this
1014          * node from the list of nodes that the other update is making
1015          * reachable, prior to freeing it:
1016          */
1017         btree_update_drop_new_node(c, b);
1018
1019         btree_update_add_key(as, &as->old_keys, b);
1020
1021         as->old_nodes[as->nr_old_nodes] = b;
1022         as->old_nodes_seq[as->nr_old_nodes] = b->data->keys.seq;
1023         as->nr_old_nodes++;
1024 }
1025
1026 static void bch2_btree_update_done(struct btree_update *as, struct btree_trans *trans)
1027 {
1028         struct bch_fs *c = as->c;
1029         u64 start_time = as->start_time;
1030
1031         BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
1032
1033         if (as->took_gc_lock)
1034                 up_read(&as->c->gc_lock);
1035         as->took_gc_lock = false;
1036
1037         bch2_btree_reserve_put(as, trans);
1038
1039         continue_at(&as->cl, btree_update_set_nodes_written,
1040                     as->c->btree_interior_update_worker);
1041
1042         bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_foreground],
1043                                start_time);
1044 }
1045
1046 static struct btree_update *
1047 bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
1048                         unsigned level, bool split, unsigned flags)
1049 {
1050         struct bch_fs *c = trans->c;
1051         struct btree_update *as;
1052         u64 start_time = local_clock();
1053         int disk_res_flags = (flags & BCH_TRANS_COMMIT_no_enospc)
1054                 ? BCH_DISK_RESERVATION_NOFAIL : 0;
1055         unsigned nr_nodes[2] = { 0, 0 };
1056         unsigned update_level = level;
1057         enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
1058         int ret = 0;
1059         u32 restart_count = trans->restart_count;
1060
1061         BUG_ON(!path->should_be_locked);
1062
1063         if (watermark == BCH_WATERMARK_copygc)
1064                 watermark = BCH_WATERMARK_btree_copygc;
1065         if (watermark < BCH_WATERMARK_btree)
1066                 watermark = BCH_WATERMARK_btree;
1067
1068         flags &= ~BCH_WATERMARK_MASK;
1069         flags |= watermark;
1070
1071         if (!(flags & BCH_TRANS_COMMIT_journal_reclaim) &&
1072             watermark < c->journal.watermark) {
1073                 struct journal_res res = { 0 };
1074
1075                 ret = drop_locks_do(trans,
1076                         bch2_journal_res_get(&c->journal, &res, 1,
1077                                              watermark|JOURNAL_RES_GET_CHECK));
1078                 if (ret)
1079                         return ERR_PTR(ret);
1080         }
1081
1082         while (1) {
1083                 nr_nodes[!!update_level] += 1 + split;
1084                 update_level++;
1085
1086                 ret = bch2_btree_path_upgrade(trans, path, update_level + 1);
1087                 if (ret)
1088                         return ERR_PTR(ret);
1089
1090                 if (!btree_path_node(path, update_level)) {
1091                         /* Allocating new root? */
1092                         nr_nodes[1] += split;
1093                         update_level = BTREE_MAX_DEPTH;
1094                         break;
1095                 }
1096
1097                 /*
1098                  * Always check for space for two keys, even if we won't have to
1099                  * split at prior level - it might have been a merge instead:
1100                  */
1101                 if (bch2_btree_node_insert_fits(c, path->l[update_level].b,
1102                                                 BKEY_BTREE_PTR_U64s_MAX * 2))
1103                         break;
1104
1105                 split = path->l[update_level].b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c);
1106         }
1107
1108         if (!down_read_trylock(&c->gc_lock)) {
1109                 ret = drop_locks_do(trans, (down_read(&c->gc_lock), 0));
1110                 if (ret) {
1111                         up_read(&c->gc_lock);
1112                         return ERR_PTR(ret);
1113                 }
1114         }
1115
1116         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOFS);
1117         memset(as, 0, sizeof(*as));
1118         closure_init(&as->cl, NULL);
1119         as->c           = c;
1120         as->start_time  = start_time;
1121         as->mode        = BTREE_INTERIOR_NO_UPDATE;
1122         as->took_gc_lock = true;
1123         as->btree_id    = path->btree_id;
1124         as->update_level = update_level;
1125         INIT_LIST_HEAD(&as->list);
1126         INIT_LIST_HEAD(&as->unwritten_list);
1127         INIT_LIST_HEAD(&as->write_blocked_list);
1128         bch2_keylist_init(&as->old_keys, as->_old_keys);
1129         bch2_keylist_init(&as->new_keys, as->_new_keys);
1130         bch2_keylist_init(&as->parent_keys, as->inline_keys);
1131
1132         mutex_lock(&c->btree_interior_update_lock);
1133         list_add_tail(&as->list, &c->btree_interior_update_list);
1134         mutex_unlock(&c->btree_interior_update_lock);
1135
1136         /*
1137          * We don't want to allocate if we're in an error state, that can cause
1138          * deadlock on emergency shutdown due to open buckets getting stuck in
1139          * the btree_reserve_cache after allocator shutdown has cleared it out.
1140          * This check needs to come after adding us to the btree_interior_update
1141          * list but before calling bch2_btree_reserve_get, to synchronize with
1142          * __bch2_fs_read_only().
1143          */
1144         ret = bch2_journal_error(&c->journal);
1145         if (ret)
1146                 goto err;
1147
1148         ret = bch2_disk_reservation_get(c, &as->disk_res,
1149                         (nr_nodes[0] + nr_nodes[1]) * btree_sectors(c),
1150                         c->opts.metadata_replicas,
1151                         disk_res_flags);
1152         if (ret)
1153                 goto err;
1154
1155         ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, NULL);
1156         if (bch2_err_matches(ret, ENOSPC) ||
1157             bch2_err_matches(ret, ENOMEM)) {
1158                 struct closure cl;
1159
1160                 /*
1161                  * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
1162                  * flag
1163                  */
1164                 if (bch2_err_matches(ret, ENOSPC) &&
1165                     (flags & BCH_TRANS_COMMIT_journal_reclaim) &&
1166                     watermark != BCH_WATERMARK_reclaim) {
1167                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
1168                         goto err;
1169                 }
1170
1171                 closure_init_stack(&cl);
1172
1173                 do {
1174                         ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, &cl);
1175
1176                         bch2_trans_unlock(trans);
1177                         closure_sync(&cl);
1178                 } while (bch2_err_matches(ret, BCH_ERR_operation_blocked));
1179         }
1180
1181         if (ret) {
1182                 trace_and_count(c, btree_reserve_get_fail, trans->fn,
1183                                 _RET_IP_, nr_nodes[0] + nr_nodes[1], ret);
1184                 goto err;
1185         }
1186
1187         ret = bch2_trans_relock(trans);
1188         if (ret)
1189                 goto err;
1190
1191         bch2_trans_verify_not_restarted(trans, restart_count);
1192         return as;
1193 err:
1194         bch2_btree_update_free(as, trans);
1195         return ERR_PTR(ret);
1196 }
1197
1198 /* Btree root updates: */
1199
1200 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1201 {
1202         /* Root nodes cannot be reaped */
1203         mutex_lock(&c->btree_cache.lock);
1204         list_del_init(&b->list);
1205         mutex_unlock(&c->btree_cache.lock);
1206
1207         mutex_lock(&c->btree_root_lock);
1208         BUG_ON(btree_node_root(c, b) &&
1209                (b->c.level < btree_node_root(c, b)->c.level ||
1210                 !btree_node_dying(btree_node_root(c, b))));
1211
1212         bch2_btree_id_root(c, b->c.btree_id)->b = b;
1213         mutex_unlock(&c->btree_root_lock);
1214
1215         bch2_recalc_btree_reserve(c);
1216 }
1217
1218 static void bch2_btree_set_root(struct btree_update *as,
1219                                 struct btree_trans *trans,
1220                                 struct btree_path *path,
1221                                 struct btree *b)
1222 {
1223         struct bch_fs *c = as->c;
1224         struct btree *old;
1225
1226         trace_and_count(c, btree_node_set_root, trans, b);
1227
1228         old = btree_node_root(c, b);
1229
1230         /*
1231          * Ensure no one is using the old root while we switch to the
1232          * new root:
1233          */
1234         bch2_btree_node_lock_write_nofail(trans, path, &old->c);
1235
1236         bch2_btree_set_root_inmem(c, b);
1237
1238         btree_update_updated_root(as, b);
1239
1240         /*
1241          * Unlock old root after new root is visible:
1242          *
1243          * The new root isn't persistent, but that's ok: we still have
1244          * an intent lock on the new root, and any updates that would
1245          * depend on the new root would have to update the new root.
1246          */
1247         bch2_btree_node_unlock_write(trans, path, old);
1248 }
1249
1250 /* Interior node updates: */
1251
1252 static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
1253                                         struct btree_trans *trans,
1254                                         struct btree_path *path,
1255                                         struct btree *b,
1256                                         struct btree_node_iter *node_iter,
1257                                         struct bkey_i *insert)
1258 {
1259         struct bch_fs *c = as->c;
1260         struct bkey_packed *k;
1261         struct printbuf buf = PRINTBUF;
1262         unsigned long old, new, v;
1263
1264         BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 &&
1265                !btree_ptr_sectors_written(insert));
1266
1267         if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
1268                 bch2_journal_key_overwritten(c, b->c.btree_id, b->c.level, insert->k.p);
1269
1270         if (bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1271                               btree_node_type(b), WRITE, &buf) ?:
1272             bch2_bkey_in_btree_node(c, b, bkey_i_to_s_c(insert), &buf)) {
1273                 printbuf_reset(&buf);
1274                 prt_printf(&buf, "inserting invalid bkey\n  ");
1275                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
1276                 prt_printf(&buf, "\n  ");
1277                 bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1278                                   btree_node_type(b), WRITE, &buf);
1279                 bch2_bkey_in_btree_node(c, b, bkey_i_to_s_c(insert), &buf);
1280
1281                 bch2_fs_inconsistent(c, "%s", buf.buf);
1282                 dump_stack();
1283         }
1284
1285         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1286                ARRAY_SIZE(as->journal_entries));
1287
1288         as->journal_u64s +=
1289                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1290                                   BCH_JSET_ENTRY_btree_keys,
1291                                   b->c.btree_id, b->c.level,
1292                                   insert, insert->k.u64s);
1293
1294         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1295                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1296                 bch2_btree_node_iter_advance(node_iter, b);
1297
1298         bch2_btree_bset_insert_key(trans, path, b, node_iter, insert);
1299         set_btree_node_dirty_acct(c, b);
1300
1301         v = READ_ONCE(b->flags);
1302         do {
1303                 old = new = v;
1304
1305                 new &= ~BTREE_WRITE_TYPE_MASK;
1306                 new |= BTREE_WRITE_interior;
1307                 new |= 1 << BTREE_NODE_need_write;
1308         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1309
1310         printbuf_exit(&buf);
1311 }
1312
1313 static void
1314 __bch2_btree_insert_keys_interior(struct btree_update *as,
1315                                   struct btree_trans *trans,
1316                                   struct btree_path *path,
1317                                   struct btree *b,
1318                                   struct btree_node_iter node_iter,
1319                                   struct keylist *keys)
1320 {
1321         struct bkey_i *insert = bch2_keylist_front(keys);
1322         struct bkey_packed *k;
1323
1324         BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1325
1326         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1327                (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1328                 ;
1329
1330         while (!bch2_keylist_empty(keys)) {
1331                 insert = bch2_keylist_front(keys);
1332
1333                 if (bpos_gt(insert->k.p, b->key.k.p))
1334                         break;
1335
1336                 bch2_insert_fixup_btree_ptr(as, trans, path, b, &node_iter, insert);
1337                 bch2_keylist_pop_front(keys);
1338         }
1339 }
1340
1341 /*
1342  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1343  * node)
1344  */
1345 static void __btree_split_node(struct btree_update *as,
1346                                struct btree_trans *trans,
1347                                struct btree *b,
1348                                struct btree *n[2])
1349 {
1350         struct bkey_packed *k;
1351         struct bpos n1_pos = POS_MIN;
1352         struct btree_node_iter iter;
1353         struct bset *bsets[2];
1354         struct bkey_format_state format[2];
1355         struct bkey_packed *out[2];
1356         struct bkey uk;
1357         unsigned u64s, n1_u64s = (b->nr.live_u64s * 3) / 5;
1358         int i;
1359
1360         for (i = 0; i < 2; i++) {
1361                 BUG_ON(n[i]->nsets != 1);
1362
1363                 bsets[i] = btree_bset_first(n[i]);
1364                 out[i] = bsets[i]->start;
1365
1366                 SET_BTREE_NODE_SEQ(n[i]->data, BTREE_NODE_SEQ(b->data) + 1);
1367                 bch2_bkey_format_init(&format[i]);
1368         }
1369
1370         u64s = 0;
1371         for_each_btree_node_key(b, k, &iter) {
1372                 if (bkey_deleted(k))
1373                         continue;
1374
1375                 i = u64s >= n1_u64s;
1376                 u64s += k->u64s;
1377                 uk = bkey_unpack_key(b, k);
1378                 if (!i)
1379                         n1_pos = uk.p;
1380                 bch2_bkey_format_add_key(&format[i], &uk);
1381         }
1382
1383         btree_set_min(n[0], b->data->min_key);
1384         btree_set_max(n[0], n1_pos);
1385         btree_set_min(n[1], bpos_successor(n1_pos));
1386         btree_set_max(n[1], b->data->max_key);
1387
1388         for (i = 0; i < 2; i++) {
1389                 bch2_bkey_format_add_pos(&format[i], n[i]->data->min_key);
1390                 bch2_bkey_format_add_pos(&format[i], n[i]->data->max_key);
1391
1392                 n[i]->data->format = bch2_bkey_format_done(&format[i]);
1393                 btree_node_set_format(n[i], n[i]->data->format);
1394         }
1395
1396         u64s = 0;
1397         for_each_btree_node_key(b, k, &iter) {
1398                 if (bkey_deleted(k))
1399                         continue;
1400
1401                 i = u64s >= n1_u64s;
1402                 u64s += k->u64s;
1403
1404                 if (bch2_bkey_transform(&n[i]->format, out[i], bkey_packed(k)
1405                                         ? &b->format: &bch2_bkey_format_current, k))
1406                         out[i]->format = KEY_FORMAT_LOCAL_BTREE;
1407                 else
1408                         bch2_bkey_unpack(b, (void *) out[i], k);
1409
1410                 out[i]->needs_whiteout = false;
1411
1412                 btree_keys_account_key_add(&n[i]->nr, 0, out[i]);
1413                 out[i] = bkey_p_next(out[i]);
1414         }
1415
1416         for (i = 0; i < 2; i++) {
1417                 bsets[i]->u64s = cpu_to_le16((u64 *) out[i] - bsets[i]->_data);
1418
1419                 BUG_ON(!bsets[i]->u64s);
1420
1421                 set_btree_bset_end(n[i], n[i]->set);
1422
1423                 btree_node_reset_sib_u64s(n[i]);
1424
1425                 bch2_verify_btree_nr_keys(n[i]);
1426
1427                 if (b->c.level)
1428                         btree_node_interior_verify(as->c, n[i]);
1429         }
1430 }
1431
1432 /*
1433  * For updates to interior nodes, we've got to do the insert before we split
1434  * because the stuff we're inserting has to be inserted atomically. Post split,
1435  * the keys might have to go in different nodes and the split would no longer be
1436  * atomic.
1437  *
1438  * Worse, if the insert is from btree node coalescing, if we do the insert after
1439  * we do the split (and pick the pivot) - the pivot we pick might be between
1440  * nodes that were coalesced, and thus in the middle of a child node post
1441  * coalescing:
1442  */
1443 static void btree_split_insert_keys(struct btree_update *as,
1444                                     struct btree_trans *trans,
1445                                     struct btree_path *path,
1446                                     struct btree *b,
1447                                     struct keylist *keys)
1448 {
1449         if (!bch2_keylist_empty(keys) &&
1450             bpos_le(bch2_keylist_front(keys)->k.p, b->data->max_key)) {
1451                 struct btree_node_iter node_iter;
1452
1453                 bch2_btree_node_iter_init(&node_iter, b, &bch2_keylist_front(keys)->k.p);
1454
1455                 __bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
1456
1457                 btree_node_interior_verify(as->c, b);
1458         }
1459 }
1460
1461 static int btree_split(struct btree_update *as, struct btree_trans *trans,
1462                        struct btree_path *path, struct btree *b,
1463                        struct keylist *keys, unsigned flags)
1464 {
1465         struct bch_fs *c = as->c;
1466         struct btree *parent = btree_node_parent(path, b);
1467         struct btree *n1, *n2 = NULL, *n3 = NULL;
1468         struct btree_path *path1 = NULL, *path2 = NULL;
1469         u64 start_time = local_clock();
1470         int ret = 0;
1471
1472         BUG_ON(!parent && (b != btree_node_root(c, b)));
1473         BUG_ON(parent && !btree_node_intent_locked(path, b->c.level + 1));
1474
1475         bch2_btree_interior_update_will_free_node(as, b);
1476
1477         if (b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c)) {
1478                 struct btree *n[2];
1479
1480                 trace_and_count(c, btree_node_split, trans, b);
1481
1482                 n[0] = n1 = bch2_btree_node_alloc(as, trans, b->c.level);
1483                 n[1] = n2 = bch2_btree_node_alloc(as, trans, b->c.level);
1484
1485                 __btree_split_node(as, trans, b, n);
1486
1487                 if (keys) {
1488                         btree_split_insert_keys(as, trans, path, n1, keys);
1489                         btree_split_insert_keys(as, trans, path, n2, keys);
1490                         BUG_ON(!bch2_keylist_empty(keys));
1491                 }
1492
1493                 bch2_btree_build_aux_trees(n2);
1494                 bch2_btree_build_aux_trees(n1);
1495
1496                 bch2_btree_update_add_new_node(as, n1);
1497                 bch2_btree_update_add_new_node(as, n2);
1498                 six_unlock_write(&n2->c.lock);
1499                 six_unlock_write(&n1->c.lock);
1500
1501                 path1 = get_unlocked_mut_path(trans, path->btree_id, n1->c.level, n1->key.k.p);
1502                 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1503                 mark_btree_node_locked(trans, path1, n1->c.level, BTREE_NODE_INTENT_LOCKED);
1504                 bch2_btree_path_level_init(trans, path1, n1);
1505
1506                 path2 = get_unlocked_mut_path(trans, path->btree_id, n2->c.level, n2->key.k.p);
1507                 six_lock_increment(&n2->c.lock, SIX_LOCK_intent);
1508                 mark_btree_node_locked(trans, path2, n2->c.level, BTREE_NODE_INTENT_LOCKED);
1509                 bch2_btree_path_level_init(trans, path2, n2);
1510
1511                 /*
1512                  * Note that on recursive parent_keys == keys, so we
1513                  * can't start adding new keys to parent_keys before emptying it
1514                  * out (which we did with btree_split_insert_keys() above)
1515                  */
1516                 bch2_keylist_add(&as->parent_keys, &n1->key);
1517                 bch2_keylist_add(&as->parent_keys, &n2->key);
1518
1519                 if (!parent) {
1520                         /* Depth increases, make a new root */
1521                         n3 = __btree_root_alloc(as, trans, b->c.level + 1);
1522
1523                         bch2_btree_update_add_new_node(as, n3);
1524                         six_unlock_write(&n3->c.lock);
1525
1526                         path2->locks_want++;
1527                         BUG_ON(btree_node_locked(path2, n3->c.level));
1528                         six_lock_increment(&n3->c.lock, SIX_LOCK_intent);
1529                         mark_btree_node_locked(trans, path2, n3->c.level, BTREE_NODE_INTENT_LOCKED);
1530                         bch2_btree_path_level_init(trans, path2, n3);
1531
1532                         n3->sib_u64s[0] = U16_MAX;
1533                         n3->sib_u64s[1] = U16_MAX;
1534
1535                         btree_split_insert_keys(as, trans, path, n3, &as->parent_keys);
1536                 }
1537         } else {
1538                 trace_and_count(c, btree_node_compact, trans, b);
1539
1540                 n1 = bch2_btree_node_alloc_replacement(as, trans, b);
1541
1542                 if (keys) {
1543                         btree_split_insert_keys(as, trans, path, n1, keys);
1544                         BUG_ON(!bch2_keylist_empty(keys));
1545                 }
1546
1547                 bch2_btree_build_aux_trees(n1);
1548                 bch2_btree_update_add_new_node(as, n1);
1549                 six_unlock_write(&n1->c.lock);
1550
1551                 path1 = get_unlocked_mut_path(trans, path->btree_id, n1->c.level, n1->key.k.p);
1552                 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1553                 mark_btree_node_locked(trans, path1, n1->c.level, BTREE_NODE_INTENT_LOCKED);
1554                 bch2_btree_path_level_init(trans, path1, n1);
1555
1556                 if (parent)
1557                         bch2_keylist_add(&as->parent_keys, &n1->key);
1558         }
1559
1560         /* New nodes all written, now make them visible: */
1561
1562         if (parent) {
1563                 /* Split a non root node */
1564                 ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1565                 if (ret)
1566                         goto err;
1567         } else if (n3) {
1568                 bch2_btree_set_root(as, trans, path, n3);
1569         } else {
1570                 /* Root filled up but didn't need to be split */
1571                 bch2_btree_set_root(as, trans, path, n1);
1572         }
1573
1574         if (n3) {
1575                 bch2_btree_update_get_open_buckets(as, n3);
1576                 bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0);
1577         }
1578         if (n2) {
1579                 bch2_btree_update_get_open_buckets(as, n2);
1580                 bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0);
1581         }
1582         bch2_btree_update_get_open_buckets(as, n1);
1583         bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1584
1585         /*
1586          * The old node must be freed (in memory) _before_ unlocking the new
1587          * nodes - else another thread could re-acquire a read lock on the old
1588          * node after another thread has locked and updated the new node, thus
1589          * seeing stale data:
1590          */
1591         bch2_btree_node_free_inmem(trans, path, b);
1592
1593         if (n3)
1594                 bch2_trans_node_add(trans, n3);
1595         if (n2)
1596                 bch2_trans_node_add(trans, n2);
1597         bch2_trans_node_add(trans, n1);
1598
1599         if (n3)
1600                 six_unlock_intent(&n3->c.lock);
1601         if (n2)
1602                 six_unlock_intent(&n2->c.lock);
1603         six_unlock_intent(&n1->c.lock);
1604 out:
1605         if (path2) {
1606                 __bch2_btree_path_unlock(trans, path2);
1607                 bch2_path_put(trans, path2, true);
1608         }
1609         if (path1) {
1610                 __bch2_btree_path_unlock(trans, path1);
1611                 bch2_path_put(trans, path1, true);
1612         }
1613
1614         bch2_trans_verify_locks(trans);
1615
1616         bch2_time_stats_update(&c->times[n2
1617                                ? BCH_TIME_btree_node_split
1618                                : BCH_TIME_btree_node_compact],
1619                                start_time);
1620         return ret;
1621 err:
1622         if (n3)
1623                 bch2_btree_node_free_never_used(as, trans, n3);
1624         if (n2)
1625                 bch2_btree_node_free_never_used(as, trans, n2);
1626         bch2_btree_node_free_never_used(as, trans, n1);
1627         goto out;
1628 }
1629
1630 static void
1631 bch2_btree_insert_keys_interior(struct btree_update *as,
1632                                 struct btree_trans *trans,
1633                                 struct btree_path *path,
1634                                 struct btree *b,
1635                                 struct keylist *keys)
1636 {
1637         struct btree_path *linked;
1638
1639         __bch2_btree_insert_keys_interior(as, trans, path, b,
1640                                           path->l[b->c.level].iter, keys);
1641
1642         btree_update_updated_node(as, b);
1643
1644         trans_for_each_path_with_node(trans, b, linked)
1645                 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1646
1647         bch2_trans_verify_paths(trans);
1648 }
1649
1650 /**
1651  * bch2_btree_insert_node - insert bkeys into a given btree node
1652  *
1653  * @as:                 btree_update object
1654  * @trans:              btree_trans object
1655  * @path:               path that points to current node
1656  * @b:                  node to insert keys into
1657  * @keys:               list of keys to insert
1658  * @flags:              transaction commit flags
1659  *
1660  * Returns: 0 on success, typically transaction restart error on failure
1661  *
1662  * Inserts as many keys as it can into a given btree node, splitting it if full.
1663  * If a split occurred, this function will return early. This can only happen
1664  * for leaf nodes -- inserts into interior nodes have to be atomic.
1665  */
1666 static int bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans,
1667                                   struct btree_path *path, struct btree *b,
1668                                   struct keylist *keys, unsigned flags)
1669 {
1670         struct bch_fs *c = as->c;
1671         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1672         int old_live_u64s = b->nr.live_u64s;
1673         int live_u64s_added, u64s_added;
1674         int ret;
1675
1676         lockdep_assert_held(&c->gc_lock);
1677         BUG_ON(!btree_node_intent_locked(path, b->c.level));
1678         BUG_ON(!b->c.level);
1679         BUG_ON(!as || as->b);
1680         bch2_verify_keylist_sorted(keys);
1681
1682         ret = bch2_btree_node_lock_write(trans, path, &b->c);
1683         if (ret)
1684                 return ret;
1685
1686         bch2_btree_node_prep_for_write(trans, path, b);
1687
1688         if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1689                 bch2_btree_node_unlock_write(trans, path, b);
1690                 goto split;
1691         }
1692
1693         btree_node_interior_verify(c, b);
1694
1695         bch2_btree_insert_keys_interior(as, trans, path, b, keys);
1696
1697         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1698         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1699
1700         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1701                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1702         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1703                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1704
1705         if (u64s_added > live_u64s_added &&
1706             bch2_maybe_compact_whiteouts(c, b))
1707                 bch2_trans_node_reinit_iter(trans, b);
1708
1709         bch2_btree_node_unlock_write(trans, path, b);
1710
1711         btree_node_interior_verify(c, b);
1712         return 0;
1713 split:
1714         /*
1715          * We could attempt to avoid the transaction restart, by calling
1716          * bch2_btree_path_upgrade() and allocating more nodes:
1717          */
1718         if (b->c.level >= as->update_level) {
1719                 trace_and_count(c, trans_restart_split_race, trans, _THIS_IP_, b);
1720                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_split_race);
1721         }
1722
1723         return btree_split(as, trans, path, b, keys, flags);
1724 }
1725
1726 int bch2_btree_split_leaf(struct btree_trans *trans,
1727                           struct btree_path *path,
1728                           unsigned flags)
1729 {
1730         struct btree *b = path_l(path)->b;
1731         struct btree_update *as;
1732         unsigned l;
1733         int ret = 0;
1734
1735         as = bch2_btree_update_start(trans, path, path->level,
1736                                      true, flags);
1737         if (IS_ERR(as))
1738                 return PTR_ERR(as);
1739
1740         ret = btree_split(as, trans, path, b, NULL, flags);
1741         if (ret) {
1742                 bch2_btree_update_free(as, trans);
1743                 return ret;
1744         }
1745
1746         bch2_btree_update_done(as, trans);
1747
1748         for (l = path->level + 1; btree_node_intent_locked(path, l) && !ret; l++)
1749                 ret = bch2_foreground_maybe_merge(trans, path, l, flags);
1750
1751         return ret;
1752 }
1753
1754 int __bch2_foreground_maybe_merge(struct btree_trans *trans,
1755                                   struct btree_path *path,
1756                                   unsigned level,
1757                                   unsigned flags,
1758                                   enum btree_node_sibling sib)
1759 {
1760         struct bch_fs *c = trans->c;
1761         struct btree_path *sib_path = NULL, *new_path = NULL;
1762         struct btree_update *as;
1763         struct bkey_format_state new_s;
1764         struct bkey_format new_f;
1765         struct bkey_i delete;
1766         struct btree *b, *m, *n, *prev, *next, *parent;
1767         struct bpos sib_pos;
1768         size_t sib_u64s;
1769         u64 start_time = local_clock();
1770         int ret = 0;
1771
1772         BUG_ON(!path->should_be_locked);
1773         BUG_ON(!btree_node_locked(path, level));
1774
1775         b = path->l[level].b;
1776
1777         if ((sib == btree_prev_sib && bpos_eq(b->data->min_key, POS_MIN)) ||
1778             (sib == btree_next_sib && bpos_eq(b->data->max_key, SPOS_MAX))) {
1779                 b->sib_u64s[sib] = U16_MAX;
1780                 return 0;
1781         }
1782
1783         sib_pos = sib == btree_prev_sib
1784                 ? bpos_predecessor(b->data->min_key)
1785                 : bpos_successor(b->data->max_key);
1786
1787         sib_path = bch2_path_get(trans, path->btree_id, sib_pos,
1788                                  U8_MAX, level, BTREE_ITER_INTENT, _THIS_IP_);
1789         ret = bch2_btree_path_traverse(trans, sib_path, false);
1790         if (ret)
1791                 goto err;
1792
1793         btree_path_set_should_be_locked(sib_path);
1794
1795         m = sib_path->l[level].b;
1796
1797         if (btree_node_parent(path, b) !=
1798             btree_node_parent(sib_path, m)) {
1799                 b->sib_u64s[sib] = U16_MAX;
1800                 goto out;
1801         }
1802
1803         if (sib == btree_prev_sib) {
1804                 prev = m;
1805                 next = b;
1806         } else {
1807                 prev = b;
1808                 next = m;
1809         }
1810
1811         if (!bpos_eq(bpos_successor(prev->data->max_key), next->data->min_key)) {
1812                 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1813
1814                 bch2_bpos_to_text(&buf1, prev->data->max_key);
1815                 bch2_bpos_to_text(&buf2, next->data->min_key);
1816                 bch_err(c,
1817                         "%s(): btree topology error:\n"
1818                         "  prev ends at   %s\n"
1819                         "  next starts at %s",
1820                         __func__, buf1.buf, buf2.buf);
1821                 printbuf_exit(&buf1);
1822                 printbuf_exit(&buf2);
1823                 bch2_topology_error(c);
1824                 ret = -EIO;
1825                 goto err;
1826         }
1827
1828         bch2_bkey_format_init(&new_s);
1829         bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
1830         __bch2_btree_calc_format(&new_s, prev);
1831         __bch2_btree_calc_format(&new_s, next);
1832         bch2_bkey_format_add_pos(&new_s, next->data->max_key);
1833         new_f = bch2_bkey_format_done(&new_s);
1834
1835         sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1836                 btree_node_u64s_with_format(m, &new_f);
1837
1838         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1839                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1840                 sib_u64s /= 2;
1841                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1842         }
1843
1844         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1845         sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
1846         b->sib_u64s[sib] = sib_u64s;
1847
1848         if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
1849                 goto out;
1850
1851         parent = btree_node_parent(path, b);
1852         as = bch2_btree_update_start(trans, path, level, false,
1853                                      BCH_TRANS_COMMIT_no_enospc|flags);
1854         ret = PTR_ERR_OR_ZERO(as);
1855         if (ret)
1856                 goto err;
1857
1858         trace_and_count(c, btree_node_merge, trans, b);
1859
1860         bch2_btree_interior_update_will_free_node(as, b);
1861         bch2_btree_interior_update_will_free_node(as, m);
1862
1863         n = bch2_btree_node_alloc(as, trans, b->c.level);
1864
1865         SET_BTREE_NODE_SEQ(n->data,
1866                            max(BTREE_NODE_SEQ(b->data),
1867                                BTREE_NODE_SEQ(m->data)) + 1);
1868
1869         btree_set_min(n, prev->data->min_key);
1870         btree_set_max(n, next->data->max_key);
1871
1872         n->data->format  = new_f;
1873         btree_node_set_format(n, new_f);
1874
1875         bch2_btree_sort_into(c, n, prev);
1876         bch2_btree_sort_into(c, n, next);
1877
1878         bch2_btree_build_aux_trees(n);
1879         bch2_btree_update_add_new_node(as, n);
1880         six_unlock_write(&n->c.lock);
1881
1882         new_path = get_unlocked_mut_path(trans, path->btree_id, n->c.level, n->key.k.p);
1883         six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1884         mark_btree_node_locked(trans, new_path, n->c.level, BTREE_NODE_INTENT_LOCKED);
1885         bch2_btree_path_level_init(trans, new_path, n);
1886
1887         bkey_init(&delete.k);
1888         delete.k.p = prev->key.k.p;
1889         bch2_keylist_add(&as->parent_keys, &delete);
1890         bch2_keylist_add(&as->parent_keys, &n->key);
1891
1892         bch2_trans_verify_paths(trans);
1893
1894         ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1895         if (ret)
1896                 goto err_free_update;
1897
1898         bch2_trans_verify_paths(trans);
1899
1900         bch2_btree_update_get_open_buckets(as, n);
1901         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1902
1903         bch2_btree_node_free_inmem(trans, path, b);
1904         bch2_btree_node_free_inmem(trans, sib_path, m);
1905
1906         bch2_trans_node_add(trans, n);
1907
1908         bch2_trans_verify_paths(trans);
1909
1910         six_unlock_intent(&n->c.lock);
1911
1912         bch2_btree_update_done(as, trans);
1913
1914         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time);
1915 out:
1916 err:
1917         if (new_path)
1918                 bch2_path_put(trans, new_path, true);
1919         bch2_path_put(trans, sib_path, true);
1920         bch2_trans_verify_locks(trans);
1921         return ret;
1922 err_free_update:
1923         bch2_btree_node_free_never_used(as, trans, n);
1924         bch2_btree_update_free(as, trans);
1925         goto out;
1926 }
1927
1928 int bch2_btree_node_rewrite(struct btree_trans *trans,
1929                             struct btree_iter *iter,
1930                             struct btree *b,
1931                             unsigned flags)
1932 {
1933         struct bch_fs *c = trans->c;
1934         struct btree_path *new_path = NULL;
1935         struct btree *n, *parent;
1936         struct btree_update *as;
1937         int ret;
1938
1939         flags |= BCH_TRANS_COMMIT_no_enospc;
1940
1941         parent = btree_node_parent(iter->path, b);
1942         as = bch2_btree_update_start(trans, iter->path, b->c.level,
1943                                      false, flags);
1944         ret = PTR_ERR_OR_ZERO(as);
1945         if (ret)
1946                 goto out;
1947
1948         bch2_btree_interior_update_will_free_node(as, b);
1949
1950         n = bch2_btree_node_alloc_replacement(as, trans, b);
1951
1952         bch2_btree_build_aux_trees(n);
1953         bch2_btree_update_add_new_node(as, n);
1954         six_unlock_write(&n->c.lock);
1955
1956         new_path = get_unlocked_mut_path(trans, iter->btree_id, n->c.level, n->key.k.p);
1957         six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1958         mark_btree_node_locked(trans, new_path, n->c.level, BTREE_NODE_INTENT_LOCKED);
1959         bch2_btree_path_level_init(trans, new_path, n);
1960
1961         trace_and_count(c, btree_node_rewrite, trans, b);
1962
1963         if (parent) {
1964                 bch2_keylist_add(&as->parent_keys, &n->key);
1965                 ret = bch2_btree_insert_node(as, trans, iter->path, parent,
1966                                              &as->parent_keys, flags);
1967                 if (ret)
1968                         goto err;
1969         } else {
1970                 bch2_btree_set_root(as, trans, iter->path, n);
1971         }
1972
1973         bch2_btree_update_get_open_buckets(as, n);
1974         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1975
1976         bch2_btree_node_free_inmem(trans, iter->path, b);
1977
1978         bch2_trans_node_add(trans, n);
1979         six_unlock_intent(&n->c.lock);
1980
1981         bch2_btree_update_done(as, trans);
1982 out:
1983         if (new_path)
1984                 bch2_path_put(trans, new_path, true);
1985         bch2_trans_downgrade(trans);
1986         return ret;
1987 err:
1988         bch2_btree_node_free_never_used(as, trans, n);
1989         bch2_btree_update_free(as, trans);
1990         goto out;
1991 }
1992
1993 struct async_btree_rewrite {
1994         struct bch_fs           *c;
1995         struct work_struct      work;
1996         struct list_head        list;
1997         enum btree_id           btree_id;
1998         unsigned                level;
1999         struct bpos             pos;
2000         __le64                  seq;
2001 };
2002
2003 static int async_btree_node_rewrite_trans(struct btree_trans *trans,
2004                                           struct async_btree_rewrite *a)
2005 {
2006         struct bch_fs *c = trans->c;
2007         struct btree_iter iter;
2008         struct btree *b;
2009         int ret;
2010
2011         bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos,
2012                                   BTREE_MAX_DEPTH, a->level, 0);
2013         b = bch2_btree_iter_peek_node(&iter);
2014         ret = PTR_ERR_OR_ZERO(b);
2015         if (ret)
2016                 goto out;
2017
2018         if (!b || b->data->keys.seq != a->seq) {
2019                 struct printbuf buf = PRINTBUF;
2020
2021                 if (b)
2022                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
2023                 else
2024                         prt_str(&buf, "(null");
2025                 bch_info(c, "%s: node to rewrite not found:, searching for seq %llu, got\n%s",
2026                          __func__, a->seq, buf.buf);
2027                 printbuf_exit(&buf);
2028                 goto out;
2029         }
2030
2031         ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
2032 out:
2033         bch2_trans_iter_exit(trans, &iter);
2034
2035         return ret;
2036 }
2037
2038 static void async_btree_node_rewrite_work(struct work_struct *work)
2039 {
2040         struct async_btree_rewrite *a =
2041                 container_of(work, struct async_btree_rewrite, work);
2042         struct bch_fs *c = a->c;
2043         int ret;
2044
2045         ret = bch2_trans_do(c, NULL, NULL, 0,
2046                       async_btree_node_rewrite_trans(trans, a));
2047         if (ret)
2048                 bch_err_fn(c, ret);
2049         bch2_write_ref_put(c, BCH_WRITE_REF_node_rewrite);
2050         kfree(a);
2051 }
2052
2053 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
2054 {
2055         struct async_btree_rewrite *a;
2056         int ret;
2057
2058         a = kmalloc(sizeof(*a), GFP_NOFS);
2059         if (!a) {
2060                 bch_err(c, "%s: error allocating memory", __func__);
2061                 return;
2062         }
2063
2064         a->c            = c;
2065         a->btree_id     = b->c.btree_id;
2066         a->level        = b->c.level;
2067         a->pos          = b->key.k.p;
2068         a->seq          = b->data->keys.seq;
2069         INIT_WORK(&a->work, async_btree_node_rewrite_work);
2070
2071         if (unlikely(!test_bit(BCH_FS_may_go_rw, &c->flags))) {
2072                 mutex_lock(&c->pending_node_rewrites_lock);
2073                 list_add(&a->list, &c->pending_node_rewrites);
2074                 mutex_unlock(&c->pending_node_rewrites_lock);
2075                 return;
2076         }
2077
2078         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_node_rewrite)) {
2079                 if (test_bit(BCH_FS_started, &c->flags)) {
2080                         bch_err(c, "%s: error getting c->writes ref", __func__);
2081                         kfree(a);
2082                         return;
2083                 }
2084
2085                 ret = bch2_fs_read_write_early(c);
2086                 if (ret) {
2087                         bch_err_msg(c, ret, "going read-write");
2088                         kfree(a);
2089                         return;
2090                 }
2091
2092                 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2093         }
2094
2095         queue_work(c->btree_interior_update_worker, &a->work);
2096 }
2097
2098 void bch2_do_pending_node_rewrites(struct bch_fs *c)
2099 {
2100         struct async_btree_rewrite *a, *n;
2101
2102         mutex_lock(&c->pending_node_rewrites_lock);
2103         list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2104                 list_del(&a->list);
2105
2106                 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2107                 queue_work(c->btree_interior_update_worker, &a->work);
2108         }
2109         mutex_unlock(&c->pending_node_rewrites_lock);
2110 }
2111
2112 void bch2_free_pending_node_rewrites(struct bch_fs *c)
2113 {
2114         struct async_btree_rewrite *a, *n;
2115
2116         mutex_lock(&c->pending_node_rewrites_lock);
2117         list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2118                 list_del(&a->list);
2119
2120                 kfree(a);
2121         }
2122         mutex_unlock(&c->pending_node_rewrites_lock);
2123 }
2124
2125 static int __bch2_btree_node_update_key(struct btree_trans *trans,
2126                                         struct btree_iter *iter,
2127                                         struct btree *b, struct btree *new_hash,
2128                                         struct bkey_i *new_key,
2129                                         unsigned commit_flags,
2130                                         bool skip_triggers)
2131 {
2132         struct bch_fs *c = trans->c;
2133         struct btree_iter iter2 = { NULL };
2134         struct btree *parent;
2135         int ret;
2136
2137         if (!skip_triggers) {
2138                 ret = bch2_trans_mark_old(trans, b->c.btree_id, b->c.level + 1,
2139                                           bkey_i_to_s_c(&b->key), 0);
2140                 if (ret)
2141                         return ret;
2142
2143                 ret = bch2_trans_mark_new(trans, b->c.btree_id, b->c.level + 1,
2144                                           new_key, 0);
2145                 if (ret)
2146                         return ret;
2147         }
2148
2149         if (new_hash) {
2150                 bkey_copy(&new_hash->key, new_key);
2151                 ret = bch2_btree_node_hash_insert(&c->btree_cache,
2152                                 new_hash, b->c.level, b->c.btree_id);
2153                 BUG_ON(ret);
2154         }
2155
2156         parent = btree_node_parent(iter->path, b);
2157         if (parent) {
2158                 bch2_trans_copy_iter(&iter2, iter);
2159
2160                 iter2.path = bch2_btree_path_make_mut(trans, iter2.path,
2161                                 iter2.flags & BTREE_ITER_INTENT,
2162                                 _THIS_IP_);
2163
2164                 BUG_ON(iter2.path->level != b->c.level);
2165                 BUG_ON(!bpos_eq(iter2.path->pos, new_key->k.p));
2166
2167                 btree_path_set_level_up(trans, iter2.path);
2168
2169                 trans->paths_sorted = false;
2170
2171                 ret   = bch2_btree_iter_traverse(&iter2) ?:
2172                         bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_NORUN);
2173                 if (ret)
2174                         goto err;
2175         } else {
2176                 BUG_ON(btree_node_root(c, b) != b);
2177
2178                 ret = darray_make_room(&trans->extra_journal_entries,
2179                                        jset_u64s(new_key->k.u64s));
2180                 if (ret)
2181                         return ret;
2182
2183                 journal_entry_set((void *) &darray_top(trans->extra_journal_entries),
2184                                   BCH_JSET_ENTRY_btree_root,
2185                                   b->c.btree_id, b->c.level,
2186                                   new_key, new_key->k.u64s);
2187                 trans->extra_journal_entries.nr += jset_u64s(new_key->k.u64s);
2188         }
2189
2190         ret = bch2_trans_commit(trans, NULL, NULL, commit_flags);
2191         if (ret)
2192                 goto err;
2193
2194         bch2_btree_node_lock_write_nofail(trans, iter->path, &b->c);
2195
2196         if (new_hash) {
2197                 mutex_lock(&c->btree_cache.lock);
2198                 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
2199                 bch2_btree_node_hash_remove(&c->btree_cache, b);
2200
2201                 bkey_copy(&b->key, new_key);
2202                 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
2203                 BUG_ON(ret);
2204                 mutex_unlock(&c->btree_cache.lock);
2205         } else {
2206                 bkey_copy(&b->key, new_key);
2207         }
2208
2209         bch2_btree_node_unlock_write(trans, iter->path, b);
2210 out:
2211         bch2_trans_iter_exit(trans, &iter2);
2212         return ret;
2213 err:
2214         if (new_hash) {
2215                 mutex_lock(&c->btree_cache.lock);
2216                 bch2_btree_node_hash_remove(&c->btree_cache, b);
2217                 mutex_unlock(&c->btree_cache.lock);
2218         }
2219         goto out;
2220 }
2221
2222 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
2223                                struct btree *b, struct bkey_i *new_key,
2224                                unsigned commit_flags, bool skip_triggers)
2225 {
2226         struct bch_fs *c = trans->c;
2227         struct btree *new_hash = NULL;
2228         struct btree_path *path = iter->path;
2229         struct closure cl;
2230         int ret = 0;
2231
2232         ret = bch2_btree_path_upgrade(trans, path, b->c.level + 1);
2233         if (ret)
2234                 return ret;
2235
2236         closure_init_stack(&cl);
2237
2238         /*
2239          * check btree_ptr_hash_val() after @b is locked by
2240          * btree_iter_traverse():
2241          */
2242         if (btree_ptr_hash_val(new_key) != b->hash_val) {
2243                 ret = bch2_btree_cache_cannibalize_lock(trans, &cl);
2244                 if (ret) {
2245                         ret = drop_locks_do(trans, (closure_sync(&cl), 0));
2246                         if (ret)
2247                                 return ret;
2248                 }
2249
2250                 new_hash = bch2_btree_node_mem_alloc(trans, false);
2251         }
2252
2253         path->intent_ref++;
2254         ret = __bch2_btree_node_update_key(trans, iter, b, new_hash, new_key,
2255                                            commit_flags, skip_triggers);
2256         --path->intent_ref;
2257
2258         if (new_hash) {
2259                 mutex_lock(&c->btree_cache.lock);
2260                 list_move(&new_hash->list, &c->btree_cache.freeable);
2261                 mutex_unlock(&c->btree_cache.lock);
2262
2263                 six_unlock_write(&new_hash->c.lock);
2264                 six_unlock_intent(&new_hash->c.lock);
2265         }
2266         closure_sync(&cl);
2267         bch2_btree_cache_cannibalize_unlock(trans);
2268         return ret;
2269 }
2270
2271 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
2272                                         struct btree *b, struct bkey_i *new_key,
2273                                         unsigned commit_flags, bool skip_triggers)
2274 {
2275         struct btree_iter iter;
2276         int ret;
2277
2278         bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p,
2279                                   BTREE_MAX_DEPTH, b->c.level,
2280                                   BTREE_ITER_INTENT);
2281         ret = bch2_btree_iter_traverse(&iter);
2282         if (ret)
2283                 goto out;
2284
2285         /* has node been freed? */
2286         if (iter.path->l[b->c.level].b != b) {
2287                 /* node has been freed: */
2288                 BUG_ON(!btree_node_dying(b));
2289                 goto out;
2290         }
2291
2292         BUG_ON(!btree_node_hashed(b));
2293
2294         struct bch_extent_ptr *ptr;
2295         bch2_bkey_drop_ptrs(bkey_i_to_s(new_key), ptr,
2296                             !bch2_bkey_has_device(bkey_i_to_s(&b->key), ptr->dev));
2297
2298         ret = bch2_btree_node_update_key(trans, &iter, b, new_key,
2299                                          commit_flags, skip_triggers);
2300 out:
2301         bch2_trans_iter_exit(trans, &iter);
2302         return ret;
2303 }
2304
2305 /* Init code: */
2306
2307 /*
2308  * Only for filesystem bringup, when first reading the btree roots or allocating
2309  * btree roots when initializing a new filesystem:
2310  */
2311 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2312 {
2313         BUG_ON(btree_node_root(c, b));
2314
2315         bch2_btree_set_root_inmem(c, b);
2316 }
2317
2318 static int __bch2_btree_root_alloc(struct btree_trans *trans, enum btree_id id)
2319 {
2320         struct bch_fs *c = trans->c;
2321         struct closure cl;
2322         struct btree *b;
2323         int ret;
2324
2325         closure_init_stack(&cl);
2326
2327         do {
2328                 ret = bch2_btree_cache_cannibalize_lock(trans, &cl);
2329                 closure_sync(&cl);
2330         } while (ret);
2331
2332         b = bch2_btree_node_mem_alloc(trans, false);
2333         bch2_btree_cache_cannibalize_unlock(trans);
2334
2335         set_btree_node_fake(b);
2336         set_btree_node_need_rewrite(b);
2337         b->c.level      = 0;
2338         b->c.btree_id   = id;
2339
2340         bkey_btree_ptr_init(&b->key);
2341         b->key.k.p = SPOS_MAX;
2342         *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2343
2344         bch2_bset_init_first(b, &b->data->keys);
2345         bch2_btree_build_aux_trees(b);
2346
2347         b->data->flags = 0;
2348         btree_set_min(b, POS_MIN);
2349         btree_set_max(b, SPOS_MAX);
2350         b->data->format = bch2_btree_calc_format(b);
2351         btree_node_set_format(b, b->data->format);
2352
2353         ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2354                                           b->c.level, b->c.btree_id);
2355         BUG_ON(ret);
2356
2357         bch2_btree_set_root_inmem(c, b);
2358
2359         six_unlock_write(&b->c.lock);
2360         six_unlock_intent(&b->c.lock);
2361         return 0;
2362 }
2363
2364 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2365 {
2366         bch2_trans_run(c, __bch2_btree_root_alloc(trans, id));
2367 }
2368
2369 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2370 {
2371         struct btree_update *as;
2372
2373         mutex_lock(&c->btree_interior_update_lock);
2374         list_for_each_entry(as, &c->btree_interior_update_list, list)
2375                 prt_printf(out, "%p m %u w %u r %u j %llu\n",
2376                        as,
2377                        as->mode,
2378                        as->nodes_written,
2379                        closure_nr_remaining(&as->cl),
2380                        as->journal.seq);
2381         mutex_unlock(&c->btree_interior_update_lock);
2382 }
2383
2384 static bool bch2_btree_interior_updates_pending(struct bch_fs *c)
2385 {
2386         bool ret;
2387
2388         mutex_lock(&c->btree_interior_update_lock);
2389         ret = !list_empty(&c->btree_interior_update_list);
2390         mutex_unlock(&c->btree_interior_update_lock);
2391
2392         return ret;
2393 }
2394
2395 bool bch2_btree_interior_updates_flush(struct bch_fs *c)
2396 {
2397         bool ret = bch2_btree_interior_updates_pending(c);
2398
2399         if (ret)
2400                 closure_wait_event(&c->btree_interior_update_wait,
2401                                    !bch2_btree_interior_updates_pending(c));
2402         return ret;
2403 }
2404
2405 void bch2_journal_entry_to_btree_root(struct bch_fs *c, struct jset_entry *entry)
2406 {
2407         struct btree_root *r = bch2_btree_id_root(c, entry->btree_id);
2408
2409         mutex_lock(&c->btree_root_lock);
2410
2411         r->level = entry->level;
2412         r->alive = true;
2413         bkey_copy(&r->key, (struct bkey_i *) entry->start);
2414
2415         mutex_unlock(&c->btree_root_lock);
2416 }
2417
2418 struct jset_entry *
2419 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2420                                     struct jset_entry *end,
2421                                     unsigned long skip)
2422 {
2423         unsigned i;
2424
2425         mutex_lock(&c->btree_root_lock);
2426
2427         for (i = 0; i < btree_id_nr_alive(c); i++) {
2428                 struct btree_root *r = bch2_btree_id_root(c, i);
2429
2430                 if (r->alive && !test_bit(i, &skip)) {
2431                         journal_entry_set(end, BCH_JSET_ENTRY_btree_root,
2432                                           i, r->level, &r->key, r->key.k.u64s);
2433                         end = vstruct_next(end);
2434                 }
2435         }
2436
2437         mutex_unlock(&c->btree_root_lock);
2438
2439         return end;
2440 }
2441
2442 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2443 {
2444         if (c->btree_interior_update_worker)
2445                 destroy_workqueue(c->btree_interior_update_worker);
2446         mempool_exit(&c->btree_interior_update_pool);
2447 }
2448
2449 void bch2_fs_btree_interior_update_init_early(struct bch_fs *c)
2450 {
2451         mutex_init(&c->btree_reserve_cache_lock);
2452         INIT_LIST_HEAD(&c->btree_interior_update_list);
2453         INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2454         mutex_init(&c->btree_interior_update_lock);
2455         INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2456
2457         INIT_LIST_HEAD(&c->pending_node_rewrites);
2458         mutex_init(&c->pending_node_rewrites_lock);
2459 }
2460
2461 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2462 {
2463         c->btree_interior_update_worker =
2464                 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2465         if (!c->btree_interior_update_worker)
2466                 return -BCH_ERR_ENOMEM_btree_interior_update_worker_init;
2467
2468         if (mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2469                                       sizeof(struct btree_update)))
2470                 return -BCH_ERR_ENOMEM_btree_interior_update_pool_init;
2471
2472         return 0;
2473 }