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_update.h"
   9 #include "btree_update_interior.h"
  10 #include "btree_io.h"
  11 #include "btree_iter.h"
  12 #include "btree_locking.h"
  13 #include "buckets.h"
  14 #include "error.h"
  15 #include "extents.h"
  16 #include "journal.h"
  17 #include "journal_reclaim.h"
  18 #include "keylist.h"
  19 #include "recovery.h"
  20 #include "replicas.h"
  21 #include "super-io.h"
  22
  23 #include <linux/random.h>
  24 #include <trace/events/bcachefs.h>
  25
  26 static void bch2_btree_insert_node(struct btree_update *, struct btree_trans *,
  27                                    struct btree_path *, struct btree *,
  28                                    struct keylist *, unsigned);
  29 static void bch2_btree_update_add_new_node(struct btree_update *, struct btree *);
  30
  31 /* Debug code: */
  32
  33 /*
  34  * Verify that child nodes correctly span parent node's range:
  35  */
  36 static void btree_node_interior_verify(struct bch_fs *c, struct btree *b)
  37 {
  38 #ifdef CONFIG_BCACHEFS_DEBUG
  39         struct bpos next_node = b->data->min_key;
  40         struct btree_node_iter iter;
  41         struct bkey_s_c k;
  42         struct bkey_s_c_btree_ptr_v2 bp;
  43         struct bkey unpacked;
  44         char buf1[100], buf2[100];
  45
  46         BUG_ON(!b->c.level);
  47
  48         if (!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))
  49                 return;
  50
  51         bch2_btree_node_iter_init_from_start(&iter, b);
  52
  53         while (1) {
  54                 k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked);
  55                 if (k.k->type != KEY_TYPE_btree_ptr_v2)
  56                         break;
  57                 bp = bkey_s_c_to_btree_ptr_v2(k);
  58
  59                 if (bpos_cmp(next_node, bp.v->min_key)) {
  60                         bch2_dump_btree_node(c, b);
  61                         panic("expected next min_key %s got %s\n",
  62                               (bch2_bpos_to_text(&PBUF(buf1), next_node), buf1),
  63                               (bch2_bpos_to_text(&PBUF(buf2), bp.v->min_key), buf2));
  64                 }
  65
  66                 bch2_btree_node_iter_advance(&iter, b);
  67
  68                 if (bch2_btree_node_iter_end(&iter)) {
  69                         if (bpos_cmp(k.k->p, b->key.k.p)) {
  70                                 bch2_dump_btree_node(c, b);
  71                                 panic("expected end %s got %s\n",
  72                                       (bch2_bpos_to_text(&PBUF(buf1), b->key.k.p), buf1),
  73                                       (bch2_bpos_to_text(&PBUF(buf2), k.k->p), buf2));
  74                         }
  75                         break;
  76                 }
  77
  78                 next_node = bpos_successor(k.k->p);
  79         }
  80 #endif
  81 }
  82
  83 /* Calculate ideal packed bkey format for new btree nodes: */
  84
  85 void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b)
  86 {
  87         struct bkey_packed *k;
  88         struct bset_tree *t;
  89         struct bkey uk;
  90
  91         for_each_bset(b, t)
  92                 bset_tree_for_each_key(b, t, k)
  93                         if (!bkey_deleted(k)) {
  94                                 uk = bkey_unpack_key(b, k);
  95                                 bch2_bkey_format_add_key(s, &uk);
  96                         }
  97 }
  98
  99 static struct bkey_format bch2_btree_calc_format(struct btree *b)
 100 {
 101         struct bkey_format_state s;
 102
 103         bch2_bkey_format_init(&s);
 104         bch2_bkey_format_add_pos(&s, b->data->min_key);
 105         bch2_bkey_format_add_pos(&s, b->data->max_key);
 106         __bch2_btree_calc_format(&s, b);
 107
 108         return bch2_bkey_format_done(&s);
 109 }
 110
 111 static size_t btree_node_u64s_with_format(struct btree *b,
 112                                           struct bkey_format *new_f)
 113 {
 114         struct bkey_format *old_f = &b->format;
 115
 116         /* stupid integer promotion rules */
 117         ssize_t delta =
 118             (((int) new_f->key_u64s - old_f->key_u64s) *
 119              (int) b->nr.packed_keys) +
 120             (((int) new_f->key_u64s - BKEY_U64s) *
 121              (int) b->nr.unpacked_keys);
 122
 123         BUG_ON(delta + b->nr.live_u64s < 0);
 124
 125         return b->nr.live_u64s + delta;
 126 }
 127
 128 /**
 129  * btree_node_format_fits - check if we could rewrite node with a new format
 130  *
 131  * This assumes all keys can pack with the new format -- it just checks if
 132  * the re-packed keys would fit inside the node itself.
 133  */
 134 bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
 135                                  struct bkey_format *new_f)
 136 {
 137         size_t u64s = btree_node_u64s_with_format(b, new_f);
 138
 139         return __vstruct_bytes(struct btree_node, u64s) < btree_bytes(c);
 140 }
 141
 142 /* Btree node freeing/allocation: */
 143
 144 static void __btree_node_free(struct bch_fs *c, struct btree *b)
 145 {
 146         trace_btree_node_free(c, b);
 147
 148         BUG_ON(btree_node_dirty(b));
 149         BUG_ON(btree_node_need_write(b));
 150         BUG_ON(b == btree_node_root(c, b));
 151         BUG_ON(b->ob.nr);
 152         BUG_ON(!list_empty(&b->write_blocked));
 153         BUG_ON(b->will_make_reachable);
 154
 155         clear_btree_node_noevict(b);
 156
 157         mutex_lock(&c->btree_cache.lock);
 158         list_move(&b->list, &c->btree_cache.freeable);
 159         mutex_unlock(&c->btree_cache.lock);
 160 }
 161
 162 static void bch2_btree_node_free_inmem(struct btree_trans *trans,
 163                                        struct btree *b)
 164 {
 165         struct bch_fs *c = trans->c;
 166         struct btree_path *path;
 167
 168         trans_for_each_path(trans, path)
 169                 BUG_ON(path->l[b->c.level].b == b &&
 170                        path->l[b->c.level].lock_seq == b->c.lock.state.seq);
 171
 172         six_lock_write(&b->c.lock, NULL, NULL);
 173
 174         bch2_btree_node_hash_remove(&c->btree_cache, b);
 175         __btree_node_free(c, b);
 176
 177         six_unlock_write(&b->c.lock);
 178         six_unlock_intent(&b->c.lock);
 179 }
 180
 181 static struct btree *__bch2_btree_node_alloc(struct bch_fs *c,
 182                                              struct disk_reservation *res,
 183                                              struct closure *cl,
 184                                              unsigned flags)
 185 {
 186         struct write_point *wp;
 187         struct btree *b;
 188         __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
 189         struct open_buckets ob = { .nr = 0 };
 190         struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
 191         unsigned nr_reserve;
 192         enum alloc_reserve alloc_reserve;
 193
 194         if (flags & BTREE_INSERT_USE_RESERVE) {
 195                 nr_reserve      = 0;
 196                 alloc_reserve   = RESERVE_BTREE_MOVINGGC;
 197         } else {
 198                 nr_reserve      = BTREE_NODE_RESERVE;
 199                 alloc_reserve   = RESERVE_BTREE;
 200         }
 201
 202         mutex_lock(&c->btree_reserve_cache_lock);
 203         if (c->btree_reserve_cache_nr > nr_reserve) {
 204                 struct btree_alloc *a =
 205                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
 206
 207                 ob = a->ob;
 208                 bkey_copy(&tmp.k, &a->k);
 209                 mutex_unlock(&c->btree_reserve_cache_lock);
 210                 goto mem_alloc;
 211         }
 212         mutex_unlock(&c->btree_reserve_cache_lock);
 213
 214 retry:
 215         wp = bch2_alloc_sectors_start(c,
 216                                       c->opts.metadata_target ?:
 217                                       c->opts.foreground_target,
 218                                       0,
 219                                       writepoint_ptr(&c->btree_write_point),
 220                                       &devs_have,
 221                                       res->nr_replicas,
 222                                       c->opts.metadata_replicas_required,
 223                                       alloc_reserve, 0, cl);
 224         if (IS_ERR(wp))
 225                 return ERR_CAST(wp);
 226
 227         if (wp->sectors_free < btree_sectors(c)) {
 228                 struct open_bucket *ob;
 229                 unsigned i;
 230
 231                 open_bucket_for_each(c, &wp->ptrs, ob, i)
 232                         if (ob->sectors_free < btree_sectors(c))
 233                                 ob->sectors_free = 0;
 234
 235                 bch2_alloc_sectors_done(c, wp);
 236                 goto retry;
 237         }
 238
 239         bkey_btree_ptr_v2_init(&tmp.k);
 240         bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, btree_sectors(c), false);
 241
 242         bch2_open_bucket_get(c, wp, &ob);
 243         bch2_alloc_sectors_done(c, wp);
 244 mem_alloc:
 245         b = bch2_btree_node_mem_alloc(c);
 246         six_unlock_write(&b->c.lock);
 247         six_unlock_intent(&b->c.lock);
 248
 249         /* we hold cannibalize_lock: */
 250         BUG_ON(IS_ERR(b));
 251         BUG_ON(b->ob.nr);
 252
 253         bkey_copy(&b->key, &tmp.k);
 254         b->ob = ob;
 255
 256         return b;
 257 }
 258
 259 static struct btree *bch2_btree_node_alloc(struct btree_update *as, unsigned level)
 260 {
 261         struct bch_fs *c = as->c;
 262         struct btree *b;
 263         int ret;
 264
 265         BUG_ON(level >= BTREE_MAX_DEPTH);
 266         BUG_ON(!as->nr_prealloc_nodes);
 267
 268         b = as->prealloc_nodes[--as->nr_prealloc_nodes];
 269
 270         six_lock_intent(&b->c.lock, NULL, NULL);
 271         six_lock_write(&b->c.lock, NULL, NULL);
 272
 273         set_btree_node_accessed(b);
 274         set_btree_node_dirty(c, b);
 275         set_btree_node_need_write(b);
 276
 277         bch2_bset_init_first(b, &b->data->keys);
 278         b->c.level      = level;
 279         b->c.btree_id   = as->btree_id;
 280         b->version_ondisk = c->sb.version;
 281
 282         memset(&b->nr, 0, sizeof(b->nr));
 283         b->data->magic = cpu_to_le64(bset_magic(c));
 284         memset(&b->data->_ptr, 0, sizeof(b->data->_ptr));
 285         b->data->flags = 0;
 286         SET_BTREE_NODE_ID(b->data, as->btree_id);
 287         SET_BTREE_NODE_LEVEL(b->data, level);
 288
 289         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 290                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
 291
 292                 bp->v.mem_ptr           = 0;
 293                 bp->v.seq               = b->data->keys.seq;
 294                 bp->v.sectors_written   = 0;
 295         }
 296
 297         SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
 298
 299         bch2_btree_build_aux_trees(b);
 300
 301         ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
 302         BUG_ON(ret);
 303
 304         trace_btree_node_alloc(c, b);
 305         return b;
 306 }
 307
 308 static void btree_set_min(struct btree *b, struct bpos pos)
 309 {
 310         if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
 311                 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
 312         b->data->min_key = pos;
 313 }
 314
 315 static void btree_set_max(struct btree *b, struct bpos pos)
 316 {
 317         b->key.k.p = pos;
 318         b->data->max_key = pos;
 319 }
 320
 321 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *as,
 322                                                   struct btree *b,
 323                                                   struct bkey_format format)
 324 {
 325         struct btree *n;
 326
 327         n = bch2_btree_node_alloc(as, b->c.level);
 328
 329         SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
 330
 331         btree_set_min(n, b->data->min_key);
 332         btree_set_max(n, b->data->max_key);
 333
 334         n->data->format         = format;
 335         btree_node_set_format(n, format);
 336
 337         bch2_btree_sort_into(as->c, n, b);
 338
 339         btree_node_reset_sib_u64s(n);
 340
 341         n->key.k.p = b->key.k.p;
 342         return n;
 343 }
 344
 345 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
 346                                                        struct btree *b)
 347 {
 348         struct bkey_format new_f = bch2_btree_calc_format(b);
 349
 350         /*
 351          * The keys might expand with the new format - if they wouldn't fit in
 352          * the btree node anymore, use the old format for now:
 353          */
 354         if (!bch2_btree_node_format_fits(as->c, b, &new_f))
 355                 new_f = b->format;
 356
 357         return __bch2_btree_node_alloc_replacement(as, b, new_f);
 358 }
 359
 360 static struct btree *__btree_root_alloc(struct btree_update *as, unsigned level)
 361 {
 362         struct btree *b = bch2_btree_node_alloc(as, level);
 363
 364         btree_set_min(b, POS_MIN);
 365         btree_set_max(b, SPOS_MAX);
 366         b->data->format = bch2_btree_calc_format(b);
 367
 368         btree_node_set_format(b, b->data->format);
 369         bch2_btree_build_aux_trees(b);
 370
 371         bch2_btree_update_add_new_node(as, b);
 372         six_unlock_write(&b->c.lock);
 373
 374         return b;
 375 }
 376
 377 static void bch2_btree_reserve_put(struct btree_update *as)
 378 {
 379         struct bch_fs *c = as->c;
 380
 381         mutex_lock(&c->btree_reserve_cache_lock);
 382
 383         while (as->nr_prealloc_nodes) {
 384                 struct btree *b = as->prealloc_nodes[--as->nr_prealloc_nodes];
 385
 386                 six_lock_intent(&b->c.lock, NULL, NULL);
 387                 six_lock_write(&b->c.lock, NULL, NULL);
 388
 389                 if (c->btree_reserve_cache_nr <
 390                     ARRAY_SIZE(c->btree_reserve_cache)) {
 391                         struct btree_alloc *a =
 392                                 &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
 393
 394                         a->ob = b->ob;
 395                         b->ob.nr = 0;
 396                         bkey_copy(&a->k, &b->key);
 397                 } else {
 398                         bch2_open_buckets_put(c, &b->ob);
 399                 }
 400
 401                 __btree_node_free(c, b);
 402                 six_unlock_write(&b->c.lock);
 403                 six_unlock_intent(&b->c.lock);
 404         }
 405
 406         mutex_unlock(&c->btree_reserve_cache_lock);
 407 }
 408
 409 static int bch2_btree_reserve_get(struct btree_update *as, unsigned nr_nodes,
 410                                   unsigned flags)
 411 {
 412         struct bch_fs *c = as->c;
 413         struct closure cl;
 414         struct btree *b;
 415         int ret;
 416
 417         closure_init_stack(&cl);
 418 retry:
 419
 420         BUG_ON(nr_nodes > BTREE_RESERVE_MAX);
 421
 422         /*
 423          * Protects reaping from the btree node cache and using the btree node
 424          * open bucket reserve:
 425          *
 426          * BTREE_INSERT_NOWAIT only applies to btree node allocation, not
 427          * blocking on this lock:
 428          */
 429         ret = bch2_btree_cache_cannibalize_lock(c, &cl);
 430         if (ret)
 431                 goto err;
 432
 433         while (as->nr_prealloc_nodes < nr_nodes) {
 434                 b = __bch2_btree_node_alloc(c, &as->disk_res,
 435                                             flags & BTREE_INSERT_NOWAIT
 436                                             ? NULL : &cl, flags);
 437                 if (IS_ERR(b)) {
 438                         ret = PTR_ERR(b);
 439                         goto err;
 440                 }
 441
 442                 as->prealloc_nodes[as->nr_prealloc_nodes++] = b;
 443         }
 444
 445         bch2_btree_cache_cannibalize_unlock(c);
 446         closure_sync(&cl);
 447         return 0;
 448 err:
 449         bch2_btree_cache_cannibalize_unlock(c);
 450         closure_sync(&cl);
 451
 452         if (ret == -EAGAIN)
 453                 goto retry;
 454
 455         trace_btree_reserve_get_fail(c, nr_nodes, &cl);
 456         return ret;
 457 }
 458
 459 /* Asynchronous interior node update machinery */
 460
 461 static void bch2_btree_update_free(struct btree_update *as)
 462 {
 463         struct bch_fs *c = as->c;
 464
 465         if (as->took_gc_lock)
 466                 up_read(&c->gc_lock);
 467         as->took_gc_lock = false;
 468
 469         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 470
 471         bch2_journal_pin_drop(&c->journal, &as->journal);
 472         bch2_journal_pin_flush(&c->journal, &as->journal);
 473         bch2_disk_reservation_put(c, &as->disk_res);
 474         bch2_btree_reserve_put(as);
 475
 476         bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_total],
 477                                as->start_time);
 478
 479         mutex_lock(&c->btree_interior_update_lock);
 480         list_del(&as->unwritten_list);
 481         list_del(&as->list);
 482
 483         closure_debug_destroy(&as->cl);
 484         mempool_free(as, &c->btree_interior_update_pool);
 485
 486         /*
 487          * Have to do the wakeup with btree_interior_update_lock still held,
 488          * since being on btree_interior_update_list is our ref on @c:
 489          */
 490         closure_wake_up(&c->btree_interior_update_wait);
 491
 492         mutex_unlock(&c->btree_interior_update_lock);
 493 }
 494
 495 static void btree_update_will_delete_key(struct btree_update *as,
 496                                          struct bkey_i *k)
 497 {
 498         BUG_ON(bch2_keylist_u64s(&as->old_keys) + k->k.u64s >
 499                ARRAY_SIZE(as->_old_keys));
 500         bch2_keylist_add(&as->old_keys, k);
 501 }
 502
 503 static void btree_update_will_add_key(struct btree_update *as,
 504                                       struct bkey_i *k)
 505 {
 506         BUG_ON(bch2_keylist_u64s(&as->new_keys) + k->k.u64s >
 507                ARRAY_SIZE(as->_new_keys));
 508         bch2_keylist_add(&as->new_keys, k);
 509 }
 510
 511 /*
 512  * The transactional part of an interior btree node update, where we journal the
 513  * update we did to the interior node and update alloc info:
 514  */
 515 static int btree_update_nodes_written_trans(struct btree_trans *trans,
 516                                             struct btree_update *as)
 517 {
 518         struct bkey_i *k;
 519         int ret;
 520
 521         trans->extra_journal_entries = (void *) &as->journal_entries[0];
 522         trans->extra_journal_entry_u64s = as->journal_u64s;
 523         trans->journal_pin = &as->journal;
 524
 525         for_each_keylist_key(&as->new_keys, k) {
 526                 ret = bch2_trans_mark_key(trans,
 527                                           bkey_s_c_null,
 528                                           bkey_i_to_s_c(k),
 529                                           BTREE_TRIGGER_INSERT);
 530                 if (ret)
 531                         return ret;
 532         }
 533
 534         for_each_keylist_key(&as->old_keys, k) {
 535                 ret = bch2_trans_mark_key(trans,
 536                                           bkey_i_to_s_c(k),
 537                                           bkey_s_c_null,
 538                                           BTREE_TRIGGER_OVERWRITE);
 539                 if (ret)
 540                         return ret;
 541         }
 542
 543         return 0;
 544 }
 545
 546 static void btree_update_nodes_written(struct btree_update *as)
 547 {
 548         struct bch_fs *c = as->c;
 549         struct btree *b = as->b;
 550         struct btree_trans trans;
 551         u64 journal_seq = 0;
 552         unsigned i;
 553         int ret;
 554
 555         /*
 556          * If we're already in an error state, it might be because a btree node
 557          * was never written, and we might be trying to free that same btree
 558          * node here, but it won't have been marked as allocated and we'll see
 559          * spurious disk usage inconsistencies in the transactional part below
 560          * if we don't skip it:
 561          */
 562         ret = bch2_journal_error(&c->journal);
 563         if (ret)
 564                 goto err;
 565
 566         BUG_ON(!journal_pin_active(&as->journal));
 567
 568         /*
 569          * Wait for any in flight writes to finish before we free the old nodes
 570          * on disk:
 571          */
 572         for (i = 0; i < as->nr_old_nodes; i++) {
 573                 struct btree *old = as->old_nodes[i];
 574                 __le64 seq;
 575
 576                 six_lock_read(&old->c.lock, NULL, NULL);
 577                 seq = old->data ? old->data->keys.seq : 0;
 578                 six_unlock_read(&old->c.lock);
 579
 580                 if (seq == as->old_nodes_seq[i])
 581                         wait_on_bit_io(&old->flags, BTREE_NODE_write_in_flight_inner,
 582                                        TASK_UNINTERRUPTIBLE);
 583         }
 584
 585         /*
 586          * We did an update to a parent node where the pointers we added pointed
 587          * to child nodes that weren't written yet: now, the child nodes have
 588          * been written so we can write out the update to the interior node.
 589          */
 590
 591         /*
 592          * We can't call into journal reclaim here: we'd block on the journal
 593          * reclaim lock, but we may need to release the open buckets we have
 594          * pinned in order for other btree updates to make forward progress, and
 595          * journal reclaim does btree updates when flushing bkey_cached entries,
 596          * which may require allocations as well.
 597          */
 598         bch2_trans_init(&trans, c, 0, 512);
 599         ret = __bch2_trans_do(&trans, &as->disk_res, &journal_seq,
 600                               BTREE_INSERT_NOFAIL|
 601                               BTREE_INSERT_NOCHECK_RW|
 602                               BTREE_INSERT_JOURNAL_RECLAIM|
 603                               BTREE_INSERT_JOURNAL_RESERVED,
 604                               btree_update_nodes_written_trans(&trans, as));
 605         bch2_trans_exit(&trans);
 606
 607         bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
 608                              "error %i in btree_update_nodes_written()", ret);
 609 err:
 610         if (b) {
 611                 /*
 612                  * @b is the node we did the final insert into:
 613                  *
 614                  * On failure to get a journal reservation, we still have to
 615                  * unblock the write and allow most of the write path to happen
 616                  * so that shutdown works, but the i->journal_seq mechanism
 617                  * won't work to prevent the btree write from being visible (we
 618                  * didn't get a journal sequence number) - instead
 619                  * __bch2_btree_node_write() doesn't do the actual write if
 620                  * we're in journal error state:
 621                  */
 622
 623                 six_lock_intent(&b->c.lock, NULL, NULL);
 624                 six_lock_write(&b->c.lock, NULL, NULL);
 625                 mutex_lock(&c->btree_interior_update_lock);
 626
 627                 list_del(&as->write_blocked_list);
 628
 629                 /*
 630                  * Node might have been freed, recheck under
 631                  * btree_interior_update_lock:
 632                  */
 633                 if (as->b == b) {
 634                         struct bset *i = btree_bset_last(b);
 635
 636                         BUG_ON(!b->c.level);
 637                         BUG_ON(!btree_node_dirty(b));
 638
 639                         if (!ret) {
 640                                 i->journal_seq = cpu_to_le64(
 641                                         max(journal_seq,
 642                                             le64_to_cpu(i->journal_seq)));
 643
 644                                 bch2_btree_add_journal_pin(c, b, journal_seq);
 645                         } else {
 646                                 /*
 647                                  * If we didn't get a journal sequence number we
 648                                  * can't write this btree node, because recovery
 649                                  * won't know to ignore this write:
 650                                  */
 651                                 set_btree_node_never_write(b);
 652                         }
 653                 }
 654
 655                 mutex_unlock(&c->btree_interior_update_lock);
 656                 six_unlock_write(&b->c.lock);
 657
 658                 btree_node_write_if_need(c, b, SIX_LOCK_intent);
 659                 six_unlock_intent(&b->c.lock);
 660         }
 661
 662         bch2_journal_pin_drop(&c->journal, &as->journal);
 663
 664         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 665
 666         mutex_lock(&c->btree_interior_update_lock);
 667         for (i = 0; i < as->nr_new_nodes; i++) {
 668                 b = as->new_nodes[i];
 669
 670                 BUG_ON(b->will_make_reachable != (unsigned long) as);
 671                 b->will_make_reachable = 0;
 672         }
 673         mutex_unlock(&c->btree_interior_update_lock);
 674
 675         for (i = 0; i < as->nr_new_nodes; i++) {
 676                 b = as->new_nodes[i];
 677
 678                 six_lock_read(&b->c.lock, NULL, NULL);
 679                 btree_node_write_if_need(c, b, SIX_LOCK_read);
 680                 six_unlock_read(&b->c.lock);
 681         }
 682
 683         for (i = 0; i < as->nr_open_buckets; i++)
 684                 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
 685
 686         bch2_btree_update_free(as);
 687 }
 688
 689 static void btree_interior_update_work(struct work_struct *work)
 690 {
 691         struct bch_fs *c =
 692                 container_of(work, struct bch_fs, btree_interior_update_work);
 693         struct btree_update *as;
 694
 695         while (1) {
 696                 mutex_lock(&c->btree_interior_update_lock);
 697                 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
 698                                               struct btree_update, unwritten_list);
 699                 if (as && !as->nodes_written)
 700                         as = NULL;
 701                 mutex_unlock(&c->btree_interior_update_lock);
 702
 703                 if (!as)
 704                         break;
 705
 706                 btree_update_nodes_written(as);
 707         }
 708 }
 709
 710 static void btree_update_set_nodes_written(struct closure *cl)
 711 {
 712         struct btree_update *as = container_of(cl, struct btree_update, cl);
 713         struct bch_fs *c = as->c;
 714
 715         mutex_lock(&c->btree_interior_update_lock);
 716         as->nodes_written = true;
 717         mutex_unlock(&c->btree_interior_update_lock);
 718
 719         queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
 720 }
 721
 722 /*
 723  * We're updating @b with pointers to nodes that haven't finished writing yet:
 724  * block @b from being written until @as completes
 725  */
 726 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
 727 {
 728         struct bch_fs *c = as->c;
 729
 730         mutex_lock(&c->btree_interior_update_lock);
 731         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 732
 733         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 734         BUG_ON(!btree_node_dirty(b));
 735
 736         as->mode        = BTREE_INTERIOR_UPDATING_NODE;
 737         as->b           = b;
 738         list_add(&as->write_blocked_list, &b->write_blocked);
 739
 740         mutex_unlock(&c->btree_interior_update_lock);
 741 }
 742
 743 static void btree_update_reparent(struct btree_update *as,
 744                                   struct btree_update *child)
 745 {
 746         struct bch_fs *c = as->c;
 747
 748         lockdep_assert_held(&c->btree_interior_update_lock);
 749
 750         child->b = NULL;
 751         child->mode = BTREE_INTERIOR_UPDATING_AS;
 752
 753         bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL);
 754 }
 755
 756 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
 757 {
 758         struct bkey_i *insert = &b->key;
 759         struct bch_fs *c = as->c;
 760
 761         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 762
 763         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
 764                ARRAY_SIZE(as->journal_entries));
 765
 766         as->journal_u64s +=
 767                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
 768                                   BCH_JSET_ENTRY_btree_root,
 769                                   b->c.btree_id, b->c.level,
 770                                   insert, insert->k.u64s);
 771
 772         mutex_lock(&c->btree_interior_update_lock);
 773         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 774
 775         as->mode        = BTREE_INTERIOR_UPDATING_ROOT;
 776         mutex_unlock(&c->btree_interior_update_lock);
 777 }
 778
 779 /*
 780  * bch2_btree_update_add_new_node:
 781  *
 782  * This causes @as to wait on @b to be written, before it gets to
 783  * bch2_btree_update_nodes_written
 784  *
 785  * Additionally, it sets b->will_make_reachable to prevent any additional writes
 786  * to @b from happening besides the first until @b is reachable on disk
 787  *
 788  * And it adds @b to the list of @as's new nodes, so that we can update sector
 789  * counts in bch2_btree_update_nodes_written:
 790  */
 791 static void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
 792 {
 793         struct bch_fs *c = as->c;
 794
 795         closure_get(&as->cl);
 796
 797         mutex_lock(&c->btree_interior_update_lock);
 798         BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
 799         BUG_ON(b->will_make_reachable);
 800
 801         as->new_nodes[as->nr_new_nodes++] = b;
 802         b->will_make_reachable = 1UL|(unsigned long) as;
 803
 804         mutex_unlock(&c->btree_interior_update_lock);
 805
 806         btree_update_will_add_key(as, &b->key);
 807 }
 808
 809 /*
 810  * returns true if @b was a new node
 811  */
 812 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
 813 {
 814         struct btree_update *as;
 815         unsigned long v;
 816         unsigned i;
 817
 818         mutex_lock(&c->btree_interior_update_lock);
 819         /*
 820          * When b->will_make_reachable != 0, it owns a ref on as->cl that's
 821          * dropped when it gets written by bch2_btree_complete_write - the
 822          * xchg() is for synchronization with bch2_btree_complete_write:
 823          */
 824         v = xchg(&b->will_make_reachable, 0);
 825         as = (struct btree_update *) (v & ~1UL);
 826
 827         if (!as) {
 828                 mutex_unlock(&c->btree_interior_update_lock);
 829                 return;
 830         }
 831
 832         for (i = 0; i < as->nr_new_nodes; i++)
 833                 if (as->new_nodes[i] == b)
 834                         goto found;
 835
 836         BUG();
 837 found:
 838         array_remove_item(as->new_nodes, as->nr_new_nodes, i);
 839         mutex_unlock(&c->btree_interior_update_lock);
 840
 841         if (v & 1)
 842                 closure_put(&as->cl);
 843 }
 844
 845 static void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
 846 {
 847         while (b->ob.nr)
 848                 as->open_buckets[as->nr_open_buckets++] =
 849                         b->ob.v[--b->ob.nr];
 850 }
 851
 852 /*
 853  * @b is being split/rewritten: it may have pointers to not-yet-written btree
 854  * nodes and thus outstanding btree_updates - redirect @b's
 855  * btree_updates to point to this btree_update:
 856  */
 857 static void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 858                                                struct btree *b)
 859 {
 860         struct bch_fs *c = as->c;
 861         struct btree_update *p, *n;
 862         struct btree_write *w;
 863
 864         set_btree_node_dying(b);
 865
 866         if (btree_node_fake(b))
 867                 return;
 868
 869         mutex_lock(&c->btree_interior_update_lock);
 870
 871         /*
 872          * Does this node have any btree_update operations preventing
 873          * it from being written?
 874          *
 875          * If so, redirect them to point to this btree_update: we can
 876          * write out our new nodes, but we won't make them visible until those
 877          * operations complete
 878          */
 879         list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
 880                 list_del_init(&p->write_blocked_list);
 881                 btree_update_reparent(as, p);
 882
 883                 /*
 884                  * for flush_held_btree_writes() waiting on updates to flush or
 885                  * nodes to be writeable:
 886                  */
 887                 closure_wake_up(&c->btree_interior_update_wait);
 888         }
 889
 890         clear_btree_node_dirty(c, b);
 891         clear_btree_node_need_write(b);
 892
 893         /*
 894          * Does this node have unwritten data that has a pin on the journal?
 895          *
 896          * If so, transfer that pin to the btree_update operation -
 897          * note that if we're freeing multiple nodes, we only need to keep the
 898          * oldest pin of any of the nodes we're freeing. We'll release the pin
 899          * when the new nodes are persistent and reachable on disk:
 900          */
 901         w = btree_current_write(b);
 902         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 903         bch2_journal_pin_drop(&c->journal, &w->journal);
 904
 905         w = btree_prev_write(b);
 906         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 907         bch2_journal_pin_drop(&c->journal, &w->journal);
 908
 909         mutex_unlock(&c->btree_interior_update_lock);
 910
 911         /*
 912          * Is this a node that isn't reachable on disk yet?
 913          *
 914          * Nodes that aren't reachable yet have writes blocked until they're
 915          * reachable - now that we've cancelled any pending writes and moved
 916          * things waiting on that write to wait on this update, we can drop this
 917          * node from the list of nodes that the other update is making
 918          * reachable, prior to freeing it:
 919          */
 920         btree_update_drop_new_node(c, b);
 921
 922         btree_update_will_delete_key(as, &b->key);
 923
 924         as->old_nodes[as->nr_old_nodes] = b;
 925         as->old_nodes_seq[as->nr_old_nodes] = b->data->keys.seq;
 926         as->nr_old_nodes++;
 927 }
 928
 929 static void bch2_btree_update_done(struct btree_update *as)
 930 {
 931         struct bch_fs *c = as->c;
 932         u64 start_time = as->start_time;
 933
 934         BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
 935
 936         if (as->took_gc_lock)
 937                 up_read(&as->c->gc_lock);
 938         as->took_gc_lock = false;
 939
 940         bch2_btree_reserve_put(as);
 941
 942         continue_at(&as->cl, btree_update_set_nodes_written,
 943                     as->c->btree_interior_update_worker);
 944
 945         bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_foreground],
 946                                start_time);
 947 }
 948
 949 static struct btree_update *
 950 bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
 951                         unsigned level, unsigned nr_nodes, unsigned flags)
 952 {
 953         struct bch_fs *c = trans->c;
 954         struct btree_update *as;
 955         u64 start_time = local_clock();
 956         int disk_res_flags = (flags & BTREE_INSERT_NOFAIL)
 957                 ? BCH_DISK_RESERVATION_NOFAIL : 0;
 958         int journal_flags = 0;
 959         int ret = 0;
 960
 961         BUG_ON(!path->should_be_locked);
 962
 963         if (flags & BTREE_INSERT_JOURNAL_RESERVED)
 964                 journal_flags |= JOURNAL_RES_GET_RESERVED;
 965         if (flags & BTREE_INSERT_JOURNAL_RECLAIM)
 966                 journal_flags |= JOURNAL_RES_GET_NONBLOCK;
 967
 968         /*
 969          * XXX: figure out how far we might need to split,
 970          * instead of locking/reserving all the way to the root:
 971          */
 972         if (!bch2_btree_path_upgrade(trans, path, U8_MAX)) {
 973                 trace_trans_restart_iter_upgrade(trans->fn, _RET_IP_,
 974                                                  path->btree_id, &path->pos);
 975                 ret = btree_trans_restart(trans);
 976                 return ERR_PTR(ret);
 977         }
 978
 979         if (flags & BTREE_INSERT_GC_LOCK_HELD)
 980                 lockdep_assert_held(&c->gc_lock);
 981         else if (!down_read_trylock(&c->gc_lock)) {
 982                 bch2_trans_unlock(trans);
 983                 down_read(&c->gc_lock);
 984                 if (!bch2_trans_relock(trans)) {
 985                         up_read(&c->gc_lock);
 986                         return ERR_PTR(-EINTR);
 987                 }
 988         }
 989
 990         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
 991         memset(as, 0, sizeof(*as));
 992         closure_init(&as->cl, NULL);
 993         as->c           = c;
 994         as->start_time  = start_time;
 995         as->mode        = BTREE_INTERIOR_NO_UPDATE;
 996         as->took_gc_lock = !(flags & BTREE_INSERT_GC_LOCK_HELD);
 997         as->btree_id    = path->btree_id;
 998         INIT_LIST_HEAD(&as->list);
 999         INIT_LIST_HEAD(&as->unwritten_list);
1000         INIT_LIST_HEAD(&as->write_blocked_list);
1001         bch2_keylist_init(&as->old_keys, as->_old_keys);
1002         bch2_keylist_init(&as->new_keys, as->_new_keys);
1003         bch2_keylist_init(&as->parent_keys, as->inline_keys);
1004
1005         mutex_lock(&c->btree_interior_update_lock);
1006         list_add_tail(&as->list, &c->btree_interior_update_list);
1007         mutex_unlock(&c->btree_interior_update_lock);
1008
1009         /*
1010          * We don't want to allocate if we're in an error state, that can cause
1011          * deadlock on emergency shutdown due to open buckets getting stuck in
1012          * the btree_reserve_cache after allocator shutdown has cleared it out.
1013          * This check needs to come after adding us to the btree_interior_update
1014          * list but before calling bch2_btree_reserve_get, to synchronize with
1015          * __bch2_fs_read_only().
1016          */
1017         ret = bch2_journal_error(&c->journal);
1018         if (ret)
1019                 goto err;
1020
1021         bch2_trans_unlock(trans);
1022
1023         ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
1024                                       BTREE_UPDATE_JOURNAL_RES,
1025                                       journal_flags);
1026         if (ret) {
1027                 bch2_btree_update_free(as);
1028                 trace_trans_restart_journal_preres_get(trans->fn, _RET_IP_);
1029                 btree_trans_restart(trans);
1030                 return ERR_PTR(ret);
1031         }
1032
1033         ret = bch2_disk_reservation_get(c, &as->disk_res,
1034                         nr_nodes * btree_sectors(c),
1035                         c->opts.metadata_replicas,
1036                         disk_res_flags);
1037         if (ret)
1038                 goto err;
1039
1040         ret = bch2_btree_reserve_get(as, nr_nodes, flags);
1041         if (ret)
1042                 goto err;
1043
1044         if (!bch2_trans_relock(trans)) {
1045                 ret = -EINTR;
1046                 goto err;
1047         }
1048
1049         bch2_journal_pin_add(&c->journal,
1050                              atomic64_read(&c->journal.seq),
1051                              &as->journal, NULL);
1052
1053         return as;
1054 err:
1055         bch2_btree_update_free(as);
1056         return ERR_PTR(ret);
1057 }
1058
1059 /* Btree root updates: */
1060
1061 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1062 {
1063         /* Root nodes cannot be reaped */
1064         mutex_lock(&c->btree_cache.lock);
1065         list_del_init(&b->list);
1066         mutex_unlock(&c->btree_cache.lock);
1067
1068         if (b->c.level)
1069                 six_lock_pcpu_alloc(&b->c.lock);
1070         else
1071                 six_lock_pcpu_free(&b->c.lock);
1072
1073         mutex_lock(&c->btree_root_lock);
1074         BUG_ON(btree_node_root(c, b) &&
1075                (b->c.level < btree_node_root(c, b)->c.level ||
1076                 !btree_node_dying(btree_node_root(c, b))));
1077
1078         btree_node_root(c, b) = b;
1079         mutex_unlock(&c->btree_root_lock);
1080
1081         bch2_recalc_btree_reserve(c);
1082 }
1083
1084 /**
1085  * bch_btree_set_root - update the root in memory and on disk
1086  *
1087  * To ensure forward progress, the current task must not be holding any
1088  * btree node write locks. However, you must hold an intent lock on the
1089  * old root.
1090  *
1091  * Note: This allocates a journal entry but doesn't add any keys to
1092  * it.  All the btree roots are part of every journal write, so there
1093  * is nothing new to be done.  This just guarantees that there is a
1094  * journal write.
1095  */
1096 static void bch2_btree_set_root(struct btree_update *as,
1097                                 struct btree_trans *trans,
1098                                 struct btree_path *path,
1099                                 struct btree *b)
1100 {
1101         struct bch_fs *c = as->c;
1102         struct btree *old;
1103
1104         trace_btree_set_root(c, b);
1105         BUG_ON(!b->written &&
1106                !test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags));
1107
1108         old = btree_node_root(c, b);
1109
1110         /*
1111          * Ensure no one is using the old root while we switch to the
1112          * new root:
1113          */
1114         bch2_btree_node_lock_write(trans, path, old);
1115
1116         bch2_btree_set_root_inmem(c, b);
1117
1118         btree_update_updated_root(as, b);
1119
1120         /*
1121          * Unlock old root after new root is visible:
1122          *
1123          * The new root isn't persistent, but that's ok: we still have
1124          * an intent lock on the new root, and any updates that would
1125          * depend on the new root would have to update the new root.
1126          */
1127         bch2_btree_node_unlock_write(trans, path, old);
1128 }
1129
1130 /* Interior node updates: */
1131
1132 static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
1133                                         struct btree_trans *trans,
1134                                         struct btree_path *path,
1135                                         struct btree *b,
1136                                         struct btree_node_iter *node_iter,
1137                                         struct bkey_i *insert)
1138 {
1139         struct bch_fs *c = as->c;
1140         struct bkey_packed *k;
1141         const char *invalid;
1142
1143         BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 &&
1144                !btree_ptr_sectors_written(insert));
1145
1146         if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
1147                 bch2_journal_key_overwritten(c, b->c.btree_id, b->c.level, insert->k.p);
1148
1149         invalid = bch2_bkey_invalid(c, bkey_i_to_s_c(insert), btree_node_type(b)) ?:
1150                 bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert));
1151         if (invalid) {
1152                 char buf[160];
1153
1154                 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(insert));
1155                 bch2_fs_inconsistent(c, "inserting invalid bkey %s: %s", buf, invalid);
1156                 dump_stack();
1157         }
1158
1159         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1160                ARRAY_SIZE(as->journal_entries));
1161
1162         as->journal_u64s +=
1163                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1164                                   BCH_JSET_ENTRY_btree_keys,
1165                                   b->c.btree_id, b->c.level,
1166                                   insert, insert->k.u64s);
1167
1168         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1169                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1170                 bch2_btree_node_iter_advance(node_iter, b);
1171
1172         bch2_btree_bset_insert_key(trans, path, b, node_iter, insert);
1173         set_btree_node_dirty(c, b);
1174         set_btree_node_need_write(b);
1175 }
1176
1177 static void
1178 __bch2_btree_insert_keys_interior(struct btree_update *as,
1179                                   struct btree_trans *trans,
1180                                   struct btree_path *path,
1181                                   struct btree *b,
1182                                   struct btree_node_iter node_iter,
1183                                   struct keylist *keys)
1184 {
1185         struct bkey_i *insert = bch2_keylist_front(keys);
1186         struct bkey_packed *k;
1187
1188         BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1189
1190         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1191                (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1192                 ;
1193
1194         while (!bch2_keylist_empty(keys)) {
1195                 bch2_insert_fixup_btree_ptr(as, trans, path, b,
1196                                 &node_iter, bch2_keylist_front(keys));
1197                 bch2_keylist_pop_front(keys);
1198         }
1199 }
1200
1201 /*
1202  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1203  * node)
1204  */
1205 static struct btree *__btree_split_node(struct btree_update *as,
1206                                         struct btree *n1)
1207 {
1208         struct bkey_format_state s;
1209         size_t nr_packed = 0, nr_unpacked = 0;
1210         struct btree *n2;
1211         struct bset *set1, *set2;
1212         struct bkey_packed *k, *set2_start, *set2_end, *out, *prev = NULL;
1213         struct bpos n1_pos;
1214
1215         n2 = bch2_btree_node_alloc(as, n1->c.level);
1216         bch2_btree_update_add_new_node(as, n2);
1217
1218         n2->data->max_key       = n1->data->max_key;
1219         n2->data->format        = n1->format;
1220         SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data));
1221         n2->key.k.p = n1->key.k.p;
1222
1223         set1 = btree_bset_first(n1);
1224         set2 = btree_bset_first(n2);
1225
1226         /*
1227          * Has to be a linear search because we don't have an auxiliary
1228          * search tree yet
1229          */
1230         k = set1->start;
1231         while (1) {
1232                 struct bkey_packed *n = bkey_next(k);
1233
1234                 if (n == vstruct_last(set1))
1235                         break;
1236                 if (k->_data - set1->_data >= (le16_to_cpu(set1->u64s) * 3) / 5)
1237                         break;
1238
1239                 if (bkey_packed(k))
1240                         nr_packed++;
1241                 else
1242                         nr_unpacked++;
1243
1244                 prev = k;
1245                 k = n;
1246         }
1247
1248         BUG_ON(!prev);
1249         set2_start      = k;
1250         set2_end        = vstruct_last(set1);
1251
1252         set1->u64s = cpu_to_le16((u64 *) set2_start - set1->_data);
1253         set_btree_bset_end(n1, n1->set);
1254
1255         n1->nr.live_u64s        = le16_to_cpu(set1->u64s);
1256         n1->nr.bset_u64s[0]     = le16_to_cpu(set1->u64s);
1257         n1->nr.packed_keys      = nr_packed;
1258         n1->nr.unpacked_keys    = nr_unpacked;
1259
1260         n1_pos = bkey_unpack_pos(n1, prev);
1261         if (as->c->sb.version < bcachefs_metadata_version_snapshot)
1262                 n1_pos.snapshot = U32_MAX;
1263
1264         btree_set_max(n1, n1_pos);
1265         btree_set_min(n2, bpos_successor(n1->key.k.p));
1266
1267         bch2_bkey_format_init(&s);
1268         bch2_bkey_format_add_pos(&s, n2->data->min_key);
1269         bch2_bkey_format_add_pos(&s, n2->data->max_key);
1270
1271         for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1272                 struct bkey uk = bkey_unpack_key(n1, k);
1273                 bch2_bkey_format_add_key(&s, &uk);
1274         }
1275
1276         n2->data->format = bch2_bkey_format_done(&s);
1277         btree_node_set_format(n2, n2->data->format);
1278
1279         out = set2->start;
1280         memset(&n2->nr, 0, sizeof(n2->nr));
1281
1282         for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1283                 BUG_ON(!bch2_bkey_transform(&n2->format, out, bkey_packed(k)
1284                                        ? &n1->format : &bch2_bkey_format_current, k));
1285                 out->format = KEY_FORMAT_LOCAL_BTREE;
1286                 btree_keys_account_key_add(&n2->nr, 0, out);
1287                 out = bkey_next(out);
1288         }
1289
1290         set2->u64s = cpu_to_le16((u64 *) out - set2->_data);
1291         set_btree_bset_end(n2, n2->set);
1292
1293         BUG_ON(!set1->u64s);
1294         BUG_ON(!set2->u64s);
1295
1296         btree_node_reset_sib_u64s(n1);
1297         btree_node_reset_sib_u64s(n2);
1298
1299         bch2_verify_btree_nr_keys(n1);
1300         bch2_verify_btree_nr_keys(n2);
1301
1302         if (n1->c.level) {
1303                 btree_node_interior_verify(as->c, n1);
1304                 btree_node_interior_verify(as->c, n2);
1305         }
1306
1307         return n2;
1308 }
1309
1310 /*
1311  * For updates to interior nodes, we've got to do the insert before we split
1312  * because the stuff we're inserting has to be inserted atomically. Post split,
1313  * the keys might have to go in different nodes and the split would no longer be
1314  * atomic.
1315  *
1316  * Worse, if the insert is from btree node coalescing, if we do the insert after
1317  * we do the split (and pick the pivot) - the pivot we pick might be between
1318  * nodes that were coalesced, and thus in the middle of a child node post
1319  * coalescing:
1320  */
1321 static void btree_split_insert_keys(struct btree_update *as,
1322                                     struct btree_trans *trans,
1323                                     struct btree_path *path,
1324                                     struct btree *b,
1325                                     struct keylist *keys)
1326 {
1327         struct btree_node_iter node_iter;
1328         struct bkey_i *k = bch2_keylist_front(keys);
1329         struct bkey_packed *src, *dst, *n;
1330         struct bset *i;
1331
1332         bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
1333
1334         __bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
1335
1336         /*
1337          * We can't tolerate whiteouts here - with whiteouts there can be
1338          * duplicate keys, and it would be rather bad if we picked a duplicate
1339          * for the pivot:
1340          */
1341         i = btree_bset_first(b);
1342         src = dst = i->start;
1343         while (src != vstruct_last(i)) {
1344                 n = bkey_next(src);
1345                 if (!bkey_deleted(src)) {
1346                         memmove_u64s_down(dst, src, src->u64s);
1347                         dst = bkey_next(dst);
1348                 }
1349                 src = n;
1350         }
1351
1352         /* Also clear out the unwritten whiteouts area: */
1353         b->whiteout_u64s = 0;
1354
1355         i->u64s = cpu_to_le16((u64 *) dst - i->_data);
1356         set_btree_bset_end(b, b->set);
1357
1358         BUG_ON(b->nsets != 1 ||
1359                b->nr.live_u64s != le16_to_cpu(btree_bset_first(b)->u64s));
1360
1361         btree_node_interior_verify(as->c, b);
1362 }
1363
1364 static void btree_split(struct btree_update *as, struct btree_trans *trans,
1365                         struct btree_path *path, struct btree *b,
1366                         struct keylist *keys, unsigned flags)
1367 {
1368         struct bch_fs *c = as->c;
1369         struct btree *parent = btree_node_parent(path, b);
1370         struct btree *n1, *n2 = NULL, *n3 = NULL;
1371         u64 start_time = local_clock();
1372
1373         BUG_ON(!parent && (b != btree_node_root(c, b)));
1374         BUG_ON(!btree_node_intent_locked(path, btree_node_root(c, b)->c.level));
1375
1376         bch2_btree_interior_update_will_free_node(as, b);
1377
1378         n1 = bch2_btree_node_alloc_replacement(as, b);
1379         bch2_btree_update_add_new_node(as, n1);
1380
1381         if (keys)
1382                 btree_split_insert_keys(as, trans, path, n1, keys);
1383
1384         if (bset_u64s(&n1->set[0]) > BTREE_SPLIT_THRESHOLD(c)) {
1385                 trace_btree_split(c, b);
1386
1387                 n2 = __btree_split_node(as, n1);
1388
1389                 bch2_btree_build_aux_trees(n2);
1390                 bch2_btree_build_aux_trees(n1);
1391                 six_unlock_write(&n2->c.lock);
1392                 six_unlock_write(&n1->c.lock);
1393
1394                 bch2_btree_node_write(c, n1, SIX_LOCK_intent);
1395                 bch2_btree_node_write(c, n2, SIX_LOCK_intent);
1396
1397                 /*
1398                  * Note that on recursive parent_keys == keys, so we
1399                  * can't start adding new keys to parent_keys before emptying it
1400                  * out (which we did with btree_split_insert_keys() above)
1401                  */
1402                 bch2_keylist_add(&as->parent_keys, &n1->key);
1403                 bch2_keylist_add(&as->parent_keys, &n2->key);
1404
1405                 if (!parent) {
1406                         /* Depth increases, make a new root */
1407                         n3 = __btree_root_alloc(as, b->c.level + 1);
1408
1409                         n3->sib_u64s[0] = U16_MAX;
1410                         n3->sib_u64s[1] = U16_MAX;
1411
1412                         btree_split_insert_keys(as, trans, path, n3, &as->parent_keys);
1413
1414                         bch2_btree_node_write(c, n3, SIX_LOCK_intent);
1415                 }
1416         } else {
1417                 trace_btree_compact(c, b);
1418
1419                 bch2_btree_build_aux_trees(n1);
1420                 six_unlock_write(&n1->c.lock);
1421
1422                 bch2_btree_node_write(c, n1, SIX_LOCK_intent);
1423
1424                 if (parent)
1425                         bch2_keylist_add(&as->parent_keys, &n1->key);
1426         }
1427
1428         /* New nodes all written, now make them visible: */
1429
1430         if (parent) {
1431                 /* Split a non root node */
1432                 bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1433         } else if (n3) {
1434                 bch2_btree_set_root(as, trans, path, n3);
1435         } else {
1436                 /* Root filled up but didn't need to be split */
1437                 bch2_btree_set_root(as, trans, path, n1);
1438         }
1439
1440         bch2_btree_update_get_open_buckets(as, n1);
1441         if (n2)
1442                 bch2_btree_update_get_open_buckets(as, n2);
1443         if (n3)
1444                 bch2_btree_update_get_open_buckets(as, n3);
1445
1446         /* Successful split, update the path to point to the new nodes: */
1447
1448         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1449         if (n3)
1450                 bch2_trans_node_add(trans, n3);
1451         if (n2)
1452                 bch2_trans_node_add(trans, n2);
1453         bch2_trans_node_add(trans, n1);
1454
1455         /*
1456          * The old node must be freed (in memory) _before_ unlocking the new
1457          * nodes - else another thread could re-acquire a read lock on the old
1458          * node after another thread has locked and updated the new node, thus
1459          * seeing stale data:
1460          */
1461         bch2_btree_node_free_inmem(trans, b);
1462
1463         if (n3)
1464                 six_unlock_intent(&n3->c.lock);
1465         if (n2)
1466                 six_unlock_intent(&n2->c.lock);
1467         six_unlock_intent(&n1->c.lock);
1468
1469         bch2_trans_verify_locks(trans);
1470
1471         bch2_time_stats_update(&c->times[n2
1472                                ? BCH_TIME_btree_node_split
1473                                : BCH_TIME_btree_node_compact],
1474                                start_time);
1475 }
1476
1477 static void
1478 bch2_btree_insert_keys_interior(struct btree_update *as,
1479                                 struct btree_trans *trans,
1480                                 struct btree_path *path,
1481                                 struct btree *b,
1482                                 struct keylist *keys)
1483 {
1484         struct btree_path *linked;
1485
1486         __bch2_btree_insert_keys_interior(as, trans, path, b,
1487                                           path->l[b->c.level].iter, keys);
1488
1489         btree_update_updated_node(as, b);
1490
1491         trans_for_each_path_with_node(trans, b, linked)
1492                 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1493
1494         bch2_trans_verify_paths(trans);
1495 }
1496
1497 /**
1498  * bch_btree_insert_node - insert bkeys into a given btree node
1499  *
1500  * @iter:               btree iterator
1501  * @keys:               list of keys to insert
1502  * @hook:               insert callback
1503  * @persistent:         if not null, @persistent will wait on journal write
1504  *
1505  * Inserts as many keys as it can into a given btree node, splitting it if full.
1506  * If a split occurred, this function will return early. This can only happen
1507  * for leaf nodes -- inserts into interior nodes have to be atomic.
1508  */
1509 static void bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans,
1510                                    struct btree_path *path, struct btree *b,
1511                                    struct keylist *keys, unsigned flags)
1512 {
1513         struct bch_fs *c = as->c;
1514         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1515         int old_live_u64s = b->nr.live_u64s;
1516         int live_u64s_added, u64s_added;
1517
1518         lockdep_assert_held(&c->gc_lock);
1519         BUG_ON(!btree_node_intent_locked(path, btree_node_root(c, b)->c.level));
1520         BUG_ON(!b->c.level);
1521         BUG_ON(!as || as->b);
1522         bch2_verify_keylist_sorted(keys);
1523
1524         bch2_btree_node_lock_for_insert(trans, path, b);
1525
1526         if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1527                 bch2_btree_node_unlock_write(trans, path, b);
1528                 goto split;
1529         }
1530
1531         btree_node_interior_verify(c, b);
1532
1533         bch2_btree_insert_keys_interior(as, trans, path, b, keys);
1534
1535         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1536         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1537
1538         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1539                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1540         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1541                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1542
1543         if (u64s_added > live_u64s_added &&
1544             bch2_maybe_compact_whiteouts(c, b))
1545                 bch2_trans_node_reinit_iter(trans, b);
1546
1547         bch2_btree_node_unlock_write(trans, path, b);
1548
1549         btree_node_interior_verify(c, b);
1550         return;
1551 split:
1552         btree_split(as, trans, path, b, keys, flags);
1553 }
1554
1555 int bch2_btree_split_leaf(struct btree_trans *trans,
1556                           struct btree_path *path,
1557                           unsigned flags)
1558 {
1559         struct bch_fs *c = trans->c;
1560         struct btree *b = path_l(path)->b;
1561         struct btree_update *as;
1562         unsigned l;
1563         int ret = 0;
1564
1565         as = bch2_btree_update_start(trans, path, path->level,
1566                 btree_update_reserve_required(c, b), flags);
1567         if (IS_ERR(as))
1568                 return PTR_ERR(as);
1569
1570         btree_split(as, trans, path, b, NULL, flags);
1571         bch2_btree_update_done(as);
1572
1573         for (l = path->level + 1; btree_path_node(path, l) && !ret; l++)
1574                 ret = bch2_foreground_maybe_merge(trans, path, l, flags);
1575
1576         return ret;
1577 }
1578
1579 int __bch2_foreground_maybe_merge(struct btree_trans *trans,
1580                                   struct btree_path *path,
1581                                   unsigned level,
1582                                   unsigned flags,
1583                                   enum btree_node_sibling sib)
1584 {
1585         struct bch_fs *c = trans->c;
1586         struct btree_path *sib_path = NULL;
1587         struct btree_update *as;
1588         struct bkey_format_state new_s;
1589         struct bkey_format new_f;
1590         struct bkey_i delete;
1591         struct btree *b, *m, *n, *prev, *next, *parent;
1592         struct bpos sib_pos;
1593         size_t sib_u64s;
1594         u64 start_time = local_clock();
1595         int ret = 0;
1596
1597         BUG_ON(!path->should_be_locked);
1598         BUG_ON(!btree_node_locked(path, level));
1599
1600         b = path->l[level].b;
1601
1602         if ((sib == btree_prev_sib && !bpos_cmp(b->data->min_key, POS_MIN)) ||
1603             (sib == btree_next_sib && !bpos_cmp(b->data->max_key, SPOS_MAX))) {
1604                 b->sib_u64s[sib] = U16_MAX;
1605                 return 0;
1606         }
1607
1608         sib_pos = sib == btree_prev_sib
1609                 ? bpos_predecessor(b->data->min_key)
1610                 : bpos_successor(b->data->max_key);
1611
1612         sib_path = bch2_path_get(trans, path->btree_id, sib_pos,
1613                                  U8_MAX, level, BTREE_ITER_INTENT, _THIS_IP_);
1614         ret = bch2_btree_path_traverse(trans, sib_path, false);
1615         if (ret)
1616                 goto err;
1617
1618         sib_path->should_be_locked = true;
1619
1620         m = sib_path->l[level].b;
1621
1622         if (btree_node_parent(path, b) !=
1623             btree_node_parent(sib_path, m)) {
1624                 b->sib_u64s[sib] = U16_MAX;
1625                 goto out;
1626         }
1627
1628         if (sib == btree_prev_sib) {
1629                 prev = m;
1630                 next = b;
1631         } else {
1632                 prev = b;
1633                 next = m;
1634         }
1635
1636         if (bkey_cmp(bpos_successor(prev->data->max_key), next->data->min_key)) {
1637                 char buf1[100], buf2[100];
1638
1639                 bch2_bpos_to_text(&PBUF(buf1), prev->data->max_key);
1640                 bch2_bpos_to_text(&PBUF(buf2), next->data->min_key);
1641                 bch_err(c,
1642                         "btree topology error in btree merge:\n"
1643                         "  prev ends at   %s\n"
1644                         "  next starts at %s",
1645                         buf1, buf2);
1646                 bch2_topology_error(c);
1647                 ret = -EIO;
1648                 goto err;
1649         }
1650
1651         bch2_bkey_format_init(&new_s);
1652         bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
1653         __bch2_btree_calc_format(&new_s, prev);
1654         __bch2_btree_calc_format(&new_s, next);
1655         bch2_bkey_format_add_pos(&new_s, next->data->max_key);
1656         new_f = bch2_bkey_format_done(&new_s);
1657
1658         sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1659                 btree_node_u64s_with_format(m, &new_f);
1660
1661         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1662                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1663                 sib_u64s /= 2;
1664                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1665         }
1666
1667         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1668         sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
1669         b->sib_u64s[sib] = sib_u64s;
1670
1671         if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
1672                 goto out;
1673
1674         parent = btree_node_parent(path, b);
1675         as = bch2_btree_update_start(trans, path, level,
1676                          btree_update_reserve_required(c, parent) + 1,
1677                          flags|
1678                          BTREE_INSERT_NOFAIL|
1679                          BTREE_INSERT_USE_RESERVE);
1680         ret = PTR_ERR_OR_ZERO(as);
1681         if (ret)
1682                 goto err;
1683
1684         trace_btree_merge(c, b);
1685
1686         bch2_btree_interior_update_will_free_node(as, b);
1687         bch2_btree_interior_update_will_free_node(as, m);
1688
1689         n = bch2_btree_node_alloc(as, b->c.level);
1690         bch2_btree_update_add_new_node(as, n);
1691
1692         btree_set_min(n, prev->data->min_key);
1693         btree_set_max(n, next->data->max_key);
1694         n->data->format         = new_f;
1695
1696         btree_node_set_format(n, new_f);
1697
1698         bch2_btree_sort_into(c, n, prev);
1699         bch2_btree_sort_into(c, n, next);
1700
1701         bch2_btree_build_aux_trees(n);
1702         six_unlock_write(&n->c.lock);
1703
1704         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1705
1706         bkey_init(&delete.k);
1707         delete.k.p = prev->key.k.p;
1708         bch2_keylist_add(&as->parent_keys, &delete);
1709         bch2_keylist_add(&as->parent_keys, &n->key);
1710
1711         bch2_trans_verify_paths(trans);
1712
1713         bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1714
1715         bch2_trans_verify_paths(trans);
1716
1717         bch2_btree_update_get_open_buckets(as, n);
1718
1719         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1720         six_lock_increment(&m->c.lock, SIX_LOCK_intent);
1721
1722         bch2_trans_node_add(trans, n);
1723
1724         bch2_trans_verify_paths(trans);
1725
1726         bch2_btree_node_free_inmem(trans, b);
1727         bch2_btree_node_free_inmem(trans, m);
1728
1729         six_unlock_intent(&n->c.lock);
1730
1731         bch2_btree_update_done(as);
1732
1733         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time);
1734 out:
1735 err:
1736         bch2_path_put(trans, sib_path, true);
1737         bch2_trans_verify_locks(trans);
1738         return ret;
1739 }
1740
1741 /**
1742  * bch_btree_node_rewrite - Rewrite/move a btree node
1743  */
1744 int bch2_btree_node_rewrite(struct btree_trans *trans,
1745                             struct btree_iter *iter,
1746                             struct btree *b,
1747                             unsigned flags)
1748 {
1749         struct bch_fs *c = trans->c;
1750         struct btree *n, *parent;
1751         struct btree_update *as;
1752         int ret;
1753
1754         flags |= BTREE_INSERT_NOFAIL;
1755
1756         parent = btree_node_parent(iter->path, b);
1757         as = bch2_btree_update_start(trans, iter->path, b->c.level,
1758                 (parent
1759                  ? btree_update_reserve_required(c, parent)
1760                  : 0) + 1,
1761                 flags);
1762         ret = PTR_ERR_OR_ZERO(as);
1763         if (ret) {
1764                 trace_btree_gc_rewrite_node_fail(c, b);
1765                 goto out;
1766         }
1767
1768         bch2_btree_interior_update_will_free_node(as, b);
1769
1770         n = bch2_btree_node_alloc_replacement(as, b);
1771         bch2_btree_update_add_new_node(as, n);
1772
1773         bch2_btree_build_aux_trees(n);
1774         six_unlock_write(&n->c.lock);
1775
1776         trace_btree_gc_rewrite_node(c, b);
1777
1778         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1779
1780         if (parent) {
1781                 bch2_keylist_add(&as->parent_keys, &n->key);
1782                 bch2_btree_insert_node(as, trans, iter->path, parent,
1783                                        &as->parent_keys, flags);
1784         } else {
1785                 bch2_btree_set_root(as, trans, iter->path, n);
1786         }
1787
1788         bch2_btree_update_get_open_buckets(as, n);
1789
1790         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1791         bch2_trans_node_add(trans, n);
1792         bch2_btree_node_free_inmem(trans, b);
1793         six_unlock_intent(&n->c.lock);
1794
1795         bch2_btree_update_done(as);
1796 out:
1797         bch2_btree_path_downgrade(iter->path);
1798         return ret;
1799 }
1800
1801 struct async_btree_rewrite {
1802         struct bch_fs           *c;
1803         struct work_struct      work;
1804         enum btree_id           btree_id;
1805         unsigned                level;
1806         struct bpos             pos;
1807         __le64                  seq;
1808 };
1809
1810 static int async_btree_node_rewrite_trans(struct btree_trans *trans,
1811                                           struct async_btree_rewrite *a)
1812 {
1813         struct btree_iter iter;
1814         struct btree *b;
1815         int ret;
1816
1817         bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos,
1818                                   BTREE_MAX_DEPTH, a->level, 0);
1819         b = bch2_btree_iter_peek_node(&iter);
1820         ret = PTR_ERR_OR_ZERO(b);
1821         if (ret)
1822                 goto out;
1823
1824         if (!b || b->data->keys.seq != a->seq)
1825                 goto out;
1826
1827         ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
1828 out :
1829         bch2_trans_iter_exit(trans, &iter);
1830
1831         return ret;
1832 }
1833
1834 void async_btree_node_rewrite_work(struct work_struct *work)
1835 {
1836         struct async_btree_rewrite *a =
1837                 container_of(work, struct async_btree_rewrite, work);
1838         struct bch_fs *c = a->c;
1839
1840         bch2_trans_do(c, NULL, NULL, 0,
1841                       async_btree_node_rewrite_trans(&trans, a));
1842         percpu_ref_put(&c->writes);
1843         kfree(a);
1844 }
1845
1846 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
1847 {
1848         struct async_btree_rewrite *a;
1849
1850         if (!percpu_ref_tryget(&c->writes))
1851                 return;
1852
1853         a = kmalloc(sizeof(*a), GFP_NOFS);
1854         if (!a) {
1855                 percpu_ref_put(&c->writes);
1856                 return;
1857         }
1858
1859         a->c            = c;
1860         a->btree_id     = b->c.btree_id;
1861         a->level        = b->c.level;
1862         a->pos          = b->key.k.p;
1863         a->seq          = b->data->keys.seq;
1864
1865         INIT_WORK(&a->work, async_btree_node_rewrite_work);
1866         queue_work(c->btree_interior_update_worker, &a->work);
1867 }
1868
1869 static int __bch2_btree_node_update_key(struct btree_trans *trans,
1870                                         struct btree_iter *iter,
1871                                         struct btree *b, struct btree *new_hash,
1872                                         struct bkey_i *new_key,
1873                                         bool skip_triggers)
1874 {
1875         struct bch_fs *c = trans->c;
1876         struct btree_iter iter2 = { NULL };
1877         struct btree *parent;
1878         u64 journal_entries[BKEY_BTREE_PTR_U64s_MAX];
1879         int ret;
1880
1881         if (!skip_triggers) {
1882                 ret = bch2_trans_mark_key(trans,
1883                                           bkey_s_c_null,
1884                                           bkey_i_to_s_c(new_key),
1885                                           BTREE_TRIGGER_INSERT);
1886                 if (ret)
1887                         return ret;
1888
1889                 ret = bch2_trans_mark_key(trans,
1890                                           bkey_i_to_s_c(&b->key),
1891                                           bkey_s_c_null,
1892                                           BTREE_TRIGGER_OVERWRITE);
1893                 if (ret)
1894                         return ret;
1895         }
1896
1897         if (new_hash) {
1898                 bkey_copy(&new_hash->key, new_key);
1899                 ret = bch2_btree_node_hash_insert(&c->btree_cache,
1900                                 new_hash, b->c.level, b->c.btree_id);
1901                 BUG_ON(ret);
1902         }
1903
1904         parent = btree_node_parent(iter->path, b);
1905         if (parent) {
1906                 bch2_trans_copy_iter(&iter2, iter);
1907
1908                 iter2.path = bch2_btree_path_make_mut(trans, iter2.path,
1909                                 iter2.flags & BTREE_ITER_INTENT,
1910                                 _THIS_IP_);
1911
1912                 BUG_ON(iter2.path->level != b->c.level);
1913                 BUG_ON(bpos_cmp(iter2.path->pos, new_key->k.p));
1914
1915                 btree_node_unlock(iter2.path, iter2.path->level);
1916                 path_l(iter2.path)->b = BTREE_ITER_NO_NODE_UP;
1917                 iter2.path->level++;
1918
1919                 ret   = bch2_btree_iter_traverse(&iter2) ?:
1920                         bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_NORUN);
1921                 if (ret)
1922                         goto err;
1923         } else {
1924                 BUG_ON(btree_node_root(c, b) != b);
1925
1926                 trans->extra_journal_entries = (void *) &journal_entries[0];
1927                 trans->extra_journal_entry_u64s =
1928                         journal_entry_set((void *) &journal_entries[0],
1929                                           BCH_JSET_ENTRY_btree_root,
1930                                           b->c.btree_id, b->c.level,
1931                                           new_key, new_key->k.u64s);
1932         }
1933
1934         ret = bch2_trans_commit(trans, NULL, NULL,
1935                                 BTREE_INSERT_NOFAIL|
1936                                 BTREE_INSERT_NOCHECK_RW|
1937                                 BTREE_INSERT_USE_RESERVE|
1938                                 BTREE_INSERT_JOURNAL_RECLAIM|
1939                                 BTREE_INSERT_JOURNAL_RESERVED);
1940         if (ret)
1941                 goto err;
1942
1943         bch2_btree_node_lock_write(trans, iter->path, b);
1944
1945         if (new_hash) {
1946                 mutex_lock(&c->btree_cache.lock);
1947                 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
1948                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1949
1950                 bkey_copy(&b->key, new_key);
1951                 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1952                 BUG_ON(ret);
1953                 mutex_unlock(&c->btree_cache.lock);
1954         } else {
1955                 bkey_copy(&b->key, new_key);
1956         }
1957
1958         bch2_btree_node_unlock_write(trans, iter->path, b);
1959 out:
1960         bch2_trans_iter_exit(trans, &iter2);
1961         return ret;
1962 err:
1963         if (new_hash) {
1964                 mutex_lock(&c->btree_cache.lock);
1965                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1966                 mutex_unlock(&c->btree_cache.lock);
1967         }
1968         goto out;
1969 }
1970
1971 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
1972                                struct btree *b, struct bkey_i *new_key,
1973                                bool skip_triggers)
1974 {
1975         struct bch_fs *c = trans->c;
1976         struct btree *new_hash = NULL;
1977         struct btree_path *path = iter->path;
1978         struct closure cl;
1979         int ret = 0;
1980
1981         if (!btree_node_intent_locked(path, b->c.level) &&
1982             !bch2_btree_path_upgrade(trans, path, b->c.level + 1)) {
1983                 btree_trans_restart(trans);
1984                 return -EINTR;
1985         }
1986
1987         closure_init_stack(&cl);
1988
1989         /*
1990          * check btree_ptr_hash_val() after @b is locked by
1991          * btree_iter_traverse():
1992          */
1993         if (btree_ptr_hash_val(new_key) != b->hash_val) {
1994                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1995                 if (ret) {
1996                         bch2_trans_unlock(trans);
1997                         closure_sync(&cl);
1998                         if (!bch2_trans_relock(trans))
1999                                 return -EINTR;
2000                 }
2001
2002                 new_hash = bch2_btree_node_mem_alloc(c);
2003         }
2004
2005         path->intent_ref++;
2006         ret = __bch2_btree_node_update_key(trans, iter, b, new_hash,
2007                                            new_key, skip_triggers);
2008         --path->intent_ref;
2009
2010         if (new_hash) {
2011                 mutex_lock(&c->btree_cache.lock);
2012                 list_move(&new_hash->list, &c->btree_cache.freeable);
2013                 mutex_unlock(&c->btree_cache.lock);
2014
2015                 six_unlock_write(&new_hash->c.lock);
2016                 six_unlock_intent(&new_hash->c.lock);
2017         }
2018         closure_sync(&cl);
2019         bch2_btree_cache_cannibalize_unlock(c);
2020         return ret;
2021 }
2022
2023 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
2024                                         struct btree *b, struct bkey_i *new_key,
2025                                         bool skip_triggers)
2026 {
2027         struct btree_iter iter;
2028         int ret;
2029
2030         bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p,
2031                                   BTREE_MAX_DEPTH, b->c.level,
2032                                   BTREE_ITER_INTENT);
2033         ret = bch2_btree_iter_traverse(&iter);
2034         if (ret)
2035                 goto out;
2036
2037         /* has node been freed? */
2038         if (iter.path->l[b->c.level].b != b) {
2039                 /* node has been freed: */
2040                 BUG_ON(!btree_node_dying(b));
2041                 goto out;
2042         }
2043
2044         BUG_ON(!btree_node_hashed(b));
2045
2046         ret = bch2_btree_node_update_key(trans, &iter, b, new_key, skip_triggers);
2047 out:
2048         bch2_trans_iter_exit(trans, &iter);
2049         return ret;
2050 }
2051
2052 /* Init code: */
2053
2054 /*
2055  * Only for filesystem bringup, when first reading the btree roots or allocating
2056  * btree roots when initializing a new filesystem:
2057  */
2058 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2059 {
2060         BUG_ON(btree_node_root(c, b));
2061
2062         bch2_btree_set_root_inmem(c, b);
2063 }
2064
2065 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2066 {
2067         struct closure cl;
2068         struct btree *b;
2069         int ret;
2070
2071         closure_init_stack(&cl);
2072
2073         do {
2074                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2075                 closure_sync(&cl);
2076         } while (ret);
2077
2078         b = bch2_btree_node_mem_alloc(c);
2079         bch2_btree_cache_cannibalize_unlock(c);
2080
2081         set_btree_node_fake(b);
2082         set_btree_node_need_rewrite(b);
2083         b->c.level      = 0;
2084         b->c.btree_id   = id;
2085
2086         bkey_btree_ptr_init(&b->key);
2087         b->key.k.p = SPOS_MAX;
2088         *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2089
2090         bch2_bset_init_first(b, &b->data->keys);
2091         bch2_btree_build_aux_trees(b);
2092
2093         b->data->flags = 0;
2094         btree_set_min(b, POS_MIN);
2095         btree_set_max(b, SPOS_MAX);
2096         b->data->format = bch2_btree_calc_format(b);
2097         btree_node_set_format(b, b->data->format);
2098
2099         ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2100                                           b->c.level, b->c.btree_id);
2101         BUG_ON(ret);
2102
2103         bch2_btree_set_root_inmem(c, b);
2104
2105         six_unlock_write(&b->c.lock);
2106         six_unlock_intent(&b->c.lock);
2107 }
2108
2109 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2110 {
2111         struct btree_update *as;
2112
2113         mutex_lock(&c->btree_interior_update_lock);
2114         list_for_each_entry(as, &c->btree_interior_update_list, list)
2115                 pr_buf(out, "%p m %u w %u r %u j %llu\n",
2116                        as,
2117                        as->mode,
2118                        as->nodes_written,
2119                        atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2120                        as->journal.seq);
2121         mutex_unlock(&c->btree_interior_update_lock);
2122 }
2123
2124 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *c)
2125 {
2126         size_t ret = 0;
2127         struct list_head *i;
2128
2129         mutex_lock(&c->btree_interior_update_lock);
2130         list_for_each(i, &c->btree_interior_update_list)
2131                 ret++;
2132         mutex_unlock(&c->btree_interior_update_lock);
2133
2134         return ret;
2135 }
2136
2137 void bch2_journal_entries_to_btree_roots(struct bch_fs *c, struct jset *jset)
2138 {
2139         struct btree_root *r;
2140         struct jset_entry *entry;
2141
2142         mutex_lock(&c->btree_root_lock);
2143
2144         vstruct_for_each(jset, entry)
2145                 if (entry->type == BCH_JSET_ENTRY_btree_root) {
2146                         r = &c->btree_roots[entry->btree_id];
2147                         r->level = entry->level;
2148                         r->alive = true;
2149                         bkey_copy(&r->key, &entry->start[0]);
2150                 }
2151
2152         mutex_unlock(&c->btree_root_lock);
2153 }
2154
2155 struct jset_entry *
2156 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2157                                     struct jset_entry *start,
2158                                     struct jset_entry *end)
2159 {
2160         struct jset_entry *entry;
2161         unsigned long have = 0;
2162         unsigned i;
2163
2164         for (entry = start; entry < end; entry = vstruct_next(entry))
2165                 if (entry->type == BCH_JSET_ENTRY_btree_root)
2166                         __set_bit(entry->btree_id, &have);
2167
2168         mutex_lock(&c->btree_root_lock);
2169
2170         for (i = 0; i < BTREE_ID_NR; i++)
2171                 if (c->btree_roots[i].alive && !test_bit(i, &have)) {
2172                         journal_entry_set(end,
2173                                           BCH_JSET_ENTRY_btree_root,
2174                                           i, c->btree_roots[i].level,
2175                                           &c->btree_roots[i].key,
2176                                           c->btree_roots[i].key.u64s);
2177                         end = vstruct_next(end);
2178                 }
2179
2180         mutex_unlock(&c->btree_root_lock);
2181
2182         return end;
2183 }
2184
2185 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2186 {
2187         if (c->btree_interior_update_worker)
2188                 destroy_workqueue(c->btree_interior_update_worker);
2189         mempool_exit(&c->btree_interior_update_pool);
2190 }
2191
2192 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2193 {
2194         mutex_init(&c->btree_reserve_cache_lock);
2195         INIT_LIST_HEAD(&c->btree_interior_update_list);
2196         INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2197         mutex_init(&c->btree_interior_update_lock);
2198         INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2199
2200         c->btree_interior_update_worker =
2201                 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2202         if (!c->btree_interior_update_worker)
2203                 return -ENOMEM;
2204
2205         return mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2206                                          sizeof(struct btree_update));
2207 }