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