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_iter *, struct btree *,
  27                                    struct keylist *, unsigned);
  28
  29 /* Debug code: */
  30
  31 /*
  32  * Verify that child nodes correctly span parent node's range:
  33  */
  34 static void btree_node_interior_verify(struct bch_fs *c, struct btree *b)
  35 {
  36 #ifdef CONFIG_BCACHEFS_DEBUG
  37         struct bpos next_node = b->data->min_key;
  38         struct btree_node_iter iter;
  39         struct bkey_s_c k;
  40         struct bkey_s_c_btree_ptr_v2 bp;
  41         struct bkey unpacked;
  42         char buf1[100], buf2[100];
  43
  44         BUG_ON(!b->c.level);
  45
  46         if (!test_bit(BCH_FS_BTREE_INTERIOR_REPLAY_DONE, &c->flags))
  47                 return;
  48
  49         bch2_btree_node_iter_init_from_start(&iter, b);
  50
  51         while (1) {
  52                 k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked);
  53                 if (k.k->type != KEY_TYPE_btree_ptr_v2)
  54                         break;
  55                 bp = bkey_s_c_to_btree_ptr_v2(k);
  56
  57                 if (bpos_cmp(next_node, bp.v->min_key)) {
  58                         bch2_dump_btree_node(c, b);
  59                         panic("expected next min_key %s got %s\n",
  60                               (bch2_bpos_to_text(&PBUF(buf1), next_node), buf1),
  61                               (bch2_bpos_to_text(&PBUF(buf2), bp.v->min_key), buf2));
  62                 }
  63
  64                 bch2_btree_node_iter_advance(&iter, b);
  65
  66                 if (bch2_btree_node_iter_end(&iter)) {
  67                         if (bpos_cmp(k.k->p, b->key.k.p)) {
  68                                 bch2_dump_btree_node(c, b);
  69                                 panic("expected end %s got %s\n",
  70                                       (bch2_bpos_to_text(&PBUF(buf1), b->key.k.p), buf1),
  71                                       (bch2_bpos_to_text(&PBUF(buf2), k.k->p), buf2));
  72                         }
  73                         break;
  74                 }
  75
  76                 next_node = bpos_successor(k.k->p);
  77         }
  78 #endif
  79 }
  80
  81 /* Calculate ideal packed bkey format for new btree nodes: */
  82
  83 void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b)
  84 {
  85         struct bkey_packed *k;
  86         struct bset_tree *t;
  87         struct bkey uk;
  88
  89         for_each_bset(b, t)
  90                 bset_tree_for_each_key(b, t, k)
  91                         if (!bkey_deleted(k)) {
  92                                 uk = bkey_unpack_key(b, k);
  93                                 bch2_bkey_format_add_key(s, &uk);
  94                         }
  95 }
  96
  97 static struct bkey_format bch2_btree_calc_format(struct btree *b)
  98 {
  99         struct bkey_format_state s;
 100
 101         bch2_bkey_format_init(&s);
 102         bch2_bkey_format_add_pos(&s, b->data->min_key);
 103         bch2_bkey_format_add_pos(&s, b->data->max_key);
 104         __bch2_btree_calc_format(&s, b);
 105
 106         return bch2_bkey_format_done(&s);
 107 }
 108
 109 static size_t btree_node_u64s_with_format(struct btree *b,
 110                                           struct bkey_format *new_f)
 111 {
 112         struct bkey_format *old_f = &b->format;
 113
 114         /* stupid integer promotion rules */
 115         ssize_t delta =
 116             (((int) new_f->key_u64s - old_f->key_u64s) *
 117              (int) b->nr.packed_keys) +
 118             (((int) new_f->key_u64s - BKEY_U64s) *
 119              (int) b->nr.unpacked_keys);
 120
 121         BUG_ON(delta + b->nr.live_u64s < 0);
 122
 123         return b->nr.live_u64s + delta;
 124 }
 125
 126 /**
 127  * btree_node_format_fits - check if we could rewrite node with a new format
 128  *
 129  * This assumes all keys can pack with the new format -- it just checks if
 130  * the re-packed keys would fit inside the node itself.
 131  */
 132 bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
 133                                  struct bkey_format *new_f)
 134 {
 135         size_t u64s = btree_node_u64s_with_format(b, new_f);
 136
 137         return __vstruct_bytes(struct btree_node, u64s) < btree_bytes(c);
 138 }
 139
 140 /* Btree node freeing/allocation: */
 141
 142 static void __btree_node_free(struct bch_fs *c, struct btree *b)
 143 {
 144         trace_btree_node_free(c, b);
 145
 146         BUG_ON(btree_node_dirty(b));
 147         BUG_ON(btree_node_need_write(b));
 148         BUG_ON(b == btree_node_root(c, b));
 149         BUG_ON(b->ob.nr);
 150         BUG_ON(!list_empty(&b->write_blocked));
 151         BUG_ON(b->will_make_reachable);
 152
 153         clear_btree_node_noevict(b);
 154
 155         bch2_btree_node_hash_remove(&c->btree_cache, 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 void bch2_btree_node_free_never_inserted(struct bch_fs *c, struct btree *b)
 163 {
 164         struct open_buckets ob = b->ob;
 165
 166         b->ob.nr = 0;
 167
 168         clear_btree_node_dirty(c, b);
 169
 170         btree_node_lock_type(c, b, SIX_LOCK_write);
 171         __btree_node_free(c, b);
 172         six_unlock_write(&b->c.lock);
 173
 174         bch2_open_buckets_put(c, &ob);
 175 }
 176
 177 void bch2_btree_node_free_inmem(struct bch_fs *c, struct btree *b,
 178                                 struct btree_iter *iter)
 179 {
 180         struct btree_iter *linked;
 181
 182         trans_for_each_iter(iter->trans, linked)
 183                 BUG_ON(linked->l[b->c.level].b == b);
 184
 185         six_lock_write(&b->c.lock, NULL, NULL);
 186         __btree_node_free(c, b);
 187         six_unlock_write(&b->c.lock);
 188         six_unlock_intent(&b->c.lock);
 189 }
 190
 191 static struct btree *__bch2_btree_node_alloc(struct bch_fs *c,
 192                                              struct disk_reservation *res,
 193                                              struct closure *cl,
 194                                              unsigned flags)
 195 {
 196         struct write_point *wp;
 197         struct btree *b;
 198         __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
 199         struct open_buckets ob = { .nr = 0 };
 200         struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
 201         unsigned nr_reserve;
 202         enum alloc_reserve alloc_reserve;
 203
 204         if (flags & BTREE_INSERT_USE_RESERVE) {
 205                 nr_reserve      = 0;
 206                 alloc_reserve   = RESERVE_BTREE_MOVINGGC;
 207         } else {
 208                 nr_reserve      = BTREE_NODE_RESERVE;
 209                 alloc_reserve   = RESERVE_BTREE;
 210         }
 211
 212         mutex_lock(&c->btree_reserve_cache_lock);
 213         if (c->btree_reserve_cache_nr > nr_reserve) {
 214                 struct btree_alloc *a =
 215                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
 216
 217                 ob = a->ob;
 218                 bkey_copy(&tmp.k, &a->k);
 219                 mutex_unlock(&c->btree_reserve_cache_lock);
 220                 goto mem_alloc;
 221         }
 222         mutex_unlock(&c->btree_reserve_cache_lock);
 223
 224 retry:
 225         wp = bch2_alloc_sectors_start(c,
 226                                       c->opts.metadata_target ?:
 227                                       c->opts.foreground_target,
 228                                       0,
 229                                       writepoint_ptr(&c->btree_write_point),
 230                                       &devs_have,
 231                                       res->nr_replicas,
 232                                       c->opts.metadata_replicas_required,
 233                                       alloc_reserve, 0, cl);
 234         if (IS_ERR(wp))
 235                 return ERR_CAST(wp);
 236
 237         if (wp->sectors_free < c->opts.btree_node_size) {
 238                 struct open_bucket *ob;
 239                 unsigned i;
 240
 241                 open_bucket_for_each(c, &wp->ptrs, ob, i)
 242                         if (ob->sectors_free < c->opts.btree_node_size)
 243                                 ob->sectors_free = 0;
 244
 245                 bch2_alloc_sectors_done(c, wp);
 246                 goto retry;
 247         }
 248
 249         bkey_btree_ptr_v2_init(&tmp.k);
 250         bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, c->opts.btree_node_size);
 251
 252         bch2_open_bucket_get(c, wp, &ob);
 253         bch2_alloc_sectors_done(c, wp);
 254 mem_alloc:
 255         b = bch2_btree_node_mem_alloc(c);
 256
 257         /* we hold cannibalize_lock: */
 258         BUG_ON(IS_ERR(b));
 259         BUG_ON(b->ob.nr);
 260
 261         bkey_copy(&b->key, &tmp.k);
 262         b->ob = ob;
 263
 264         return b;
 265 }
 266
 267 static struct btree *bch2_btree_node_alloc(struct btree_update *as, unsigned level)
 268 {
 269         struct bch_fs *c = as->c;
 270         struct btree *b;
 271         int ret;
 272
 273         BUG_ON(level >= BTREE_MAX_DEPTH);
 274         BUG_ON(!as->nr_prealloc_nodes);
 275
 276         b = as->prealloc_nodes[--as->nr_prealloc_nodes];
 277
 278         set_btree_node_accessed(b);
 279         set_btree_node_dirty(c, b);
 280         set_btree_node_need_write(b);
 281
 282         bch2_bset_init_first(b, &b->data->keys);
 283         b->c.level      = level;
 284         b->c.btree_id   = as->btree_id;
 285         b->version_ondisk = c->sb.version;
 286
 287         memset(&b->nr, 0, sizeof(b->nr));
 288         b->data->magic = cpu_to_le64(bset_magic(c));
 289         memset(&b->data->_ptr, 0, sizeof(b->data->_ptr));
 290         b->data->flags = 0;
 291         SET_BTREE_NODE_ID(b->data, as->btree_id);
 292         SET_BTREE_NODE_LEVEL(b->data, level);
 293
 294         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 295                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
 296
 297                 bp->v.mem_ptr           = 0;
 298                 bp->v.seq               = b->data->keys.seq;
 299                 bp->v.sectors_written   = 0;
 300         }
 301
 302         SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
 303
 304         bch2_btree_build_aux_trees(b);
 305
 306         ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
 307         BUG_ON(ret);
 308
 309         trace_btree_node_alloc(c, b);
 310         return b;
 311 }
 312
 313 static void btree_set_min(struct btree *b, struct bpos pos)
 314 {
 315         if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
 316                 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
 317         b->data->min_key = pos;
 318 }
 319
 320 static void btree_set_max(struct btree *b, struct bpos pos)
 321 {
 322         b->key.k.p = pos;
 323         b->data->max_key = pos;
 324 }
 325
 326 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *as,
 327                                                   struct btree *b,
 328                                                   struct bkey_format format)
 329 {
 330         struct btree *n;
 331
 332         n = bch2_btree_node_alloc(as, b->c.level);
 333
 334         SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
 335
 336         btree_set_min(n, b->data->min_key);
 337         btree_set_max(n, b->data->max_key);
 338
 339         n->data->format         = format;
 340         btree_node_set_format(n, format);
 341
 342         bch2_btree_sort_into(as->c, n, b);
 343
 344         btree_node_reset_sib_u64s(n);
 345
 346         n->key.k.p = b->key.k.p;
 347         return n;
 348 }
 349
 350 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
 351                                                        struct btree *b)
 352 {
 353         struct bkey_format new_f = bch2_btree_calc_format(b);
 354
 355         /*
 356          * The keys might expand with the new format - if they wouldn't fit in
 357          * the btree node anymore, use the old format for now:
 358          */
 359         if (!bch2_btree_node_format_fits(as->c, b, &new_f))
 360                 new_f = b->format;
 361
 362         return __bch2_btree_node_alloc_replacement(as, b, new_f);
 363 }
 364
 365 static struct btree *__btree_root_alloc(struct btree_update *as, unsigned level)
 366 {
 367         struct btree *b = bch2_btree_node_alloc(as, level);
 368
 369         btree_set_min(b, POS_MIN);
 370         btree_set_max(b, SPOS_MAX);
 371         b->data->format = bch2_btree_calc_format(b);
 372
 373         btree_node_set_format(b, b->data->format);
 374         bch2_btree_build_aux_trees(b);
 375
 376         bch2_btree_update_add_new_node(as, b);
 377         six_unlock_write(&b->c.lock);
 378
 379         return b;
 380 }
 381
 382 static void bch2_btree_reserve_put(struct btree_update *as)
 383 {
 384         struct bch_fs *c = as->c;
 385
 386         mutex_lock(&c->btree_reserve_cache_lock);
 387
 388         while (as->nr_prealloc_nodes) {
 389                 struct btree *b = as->prealloc_nodes[--as->nr_prealloc_nodes];
 390
 391                 six_unlock_write(&b->c.lock);
 392
 393                 if (c->btree_reserve_cache_nr <
 394                     ARRAY_SIZE(c->btree_reserve_cache)) {
 395                         struct btree_alloc *a =
 396                                 &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
 397
 398                         a->ob = b->ob;
 399                         b->ob.nr = 0;
 400                         bkey_copy(&a->k, &b->key);
 401                 } else {
 402                         bch2_open_buckets_put(c, &b->ob);
 403                 }
 404
 405                 btree_node_lock_type(c, b, SIX_LOCK_write);
 406                 __btree_node_free(c, b);
 407                 six_unlock_write(&b->c.lock);
 408
 409                 six_unlock_intent(&b->c.lock);
 410         }
 411
 412         mutex_unlock(&c->btree_reserve_cache_lock);
 413 }
 414
 415 static int bch2_btree_reserve_get(struct btree_update *as, unsigned nr_nodes,
 416                                   unsigned flags, struct closure *cl)
 417 {
 418         struct bch_fs *c = as->c;
 419         struct btree *b;
 420         int ret;
 421
 422         BUG_ON(nr_nodes > BTREE_RESERVE_MAX);
 423
 424         /*
 425          * Protects reaping from the btree node cache and using the btree node
 426          * open bucket reserve:
 427          */
 428         ret = bch2_btree_cache_cannibalize_lock(c, cl);
 429         if (ret)
 430                 return ret;
 431
 432         while (as->nr_prealloc_nodes < nr_nodes) {
 433                 b = __bch2_btree_node_alloc(c, &as->disk_res,
 434                                             flags & BTREE_INSERT_NOWAIT
 435                                             ? NULL : cl, flags);
 436                 if (IS_ERR(b)) {
 437                         ret = PTR_ERR(b);
 438                         goto err_free;
 439                 }
 440
 441                 as->prealloc_nodes[as->nr_prealloc_nodes++] = b;
 442         }
 443
 444         bch2_btree_cache_cannibalize_unlock(c);
 445         return 0;
 446 err_free:
 447         bch2_btree_cache_cannibalize_unlock(c);
 448         trace_btree_reserve_get_fail(c, nr_nodes, cl);
 449         return ret;
 450 }
 451
 452 /* Asynchronous interior node update machinery */
 453
 454 static void bch2_btree_update_free(struct btree_update *as)
 455 {
 456         struct bch_fs *c = as->c;
 457
 458         if (as->took_gc_lock)
 459                 up_read(&c->gc_lock);
 460         as->took_gc_lock = false;
 461
 462         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 463
 464         bch2_journal_pin_drop(&c->journal, &as->journal);
 465         bch2_journal_pin_flush(&c->journal, &as->journal);
 466         bch2_disk_reservation_put(c, &as->disk_res);
 467         bch2_btree_reserve_put(as);
 468
 469         mutex_lock(&c->btree_interior_update_lock);
 470         list_del(&as->unwritten_list);
 471         list_del(&as->list);
 472         mutex_unlock(&c->btree_interior_update_lock);
 473
 474         closure_debug_destroy(&as->cl);
 475         mempool_free(as, &c->btree_interior_update_pool);
 476
 477         closure_wake_up(&c->btree_interior_update_wait);
 478 }
 479
 480 static void btree_update_will_delete_key(struct btree_update *as,
 481                                          struct bkey_i *k)
 482 {
 483         BUG_ON(bch2_keylist_u64s(&as->old_keys) + k->k.u64s >
 484                ARRAY_SIZE(as->_old_keys));
 485         bch2_keylist_add(&as->old_keys, k);
 486 }
 487
 488 static void btree_update_will_add_key(struct btree_update *as,
 489                                       struct bkey_i *k)
 490 {
 491         BUG_ON(bch2_keylist_u64s(&as->new_keys) + k->k.u64s >
 492                ARRAY_SIZE(as->_new_keys));
 493         bch2_keylist_add(&as->new_keys, k);
 494 }
 495
 496 /*
 497  * The transactional part of an interior btree node update, where we journal the
 498  * update we did to the interior node and update alloc info:
 499  */
 500 static int btree_update_nodes_written_trans(struct btree_trans *trans,
 501                                             struct btree_update *as)
 502 {
 503         struct bkey_i *k;
 504         int ret;
 505
 506         trans->extra_journal_entries = (void *) &as->journal_entries[0];
 507         trans->extra_journal_entry_u64s = as->journal_u64s;
 508         trans->journal_pin = &as->journal;
 509
 510         for_each_keylist_key(&as->new_keys, k) {
 511                 ret = bch2_trans_mark_key(trans,
 512                                           bkey_s_c_null,
 513                                           bkey_i_to_s_c(k),
 514                                           BTREE_TRIGGER_INSERT);
 515                 if (ret)
 516                         return ret;
 517         }
 518
 519         for_each_keylist_key(&as->old_keys, k) {
 520                 ret = bch2_trans_mark_key(trans,
 521                                           bkey_i_to_s_c(k),
 522                                           bkey_s_c_null,
 523                                           BTREE_TRIGGER_OVERWRITE);
 524                 if (ret)
 525                         return ret;
 526         }
 527
 528         return 0;
 529 }
 530
 531 static void btree_update_nodes_written(struct btree_update *as)
 532 {
 533         struct bch_fs *c = as->c;
 534         struct btree *b = as->b;
 535         struct btree_trans trans;
 536         u64 journal_seq = 0;
 537         unsigned i;
 538         int ret;
 539
 540         /*
 541          * If we're already in an error state, it might be because a btree node
 542          * was never written, and we might be trying to free that same btree
 543          * node here, but it won't have been marked as allocated and we'll see
 544          * spurious disk usage inconsistencies in the transactional part below
 545          * if we don't skip it:
 546          */
 547         ret = bch2_journal_error(&c->journal);
 548         if (ret)
 549                 goto err;
 550
 551         BUG_ON(!journal_pin_active(&as->journal));
 552
 553         /*
 554          * Wait for any in flight writes to finish before we free the old nodes
 555          * on disk:
 556          */
 557         for (i = 0; i < as->nr_old_nodes; i++) {
 558                 struct btree *old = as->old_nodes[i];
 559                 __le64 seq;
 560
 561                 six_lock_read(&old->c.lock, NULL, NULL);
 562                 seq = old->data ? old->data->keys.seq : 0;
 563                 six_unlock_read(&old->c.lock);
 564
 565                 if (seq == as->old_nodes_seq[i])
 566                         wait_on_bit_io(&old->flags, BTREE_NODE_write_in_flight_inner,
 567                                        TASK_UNINTERRUPTIBLE);
 568         }
 569
 570         /*
 571          * We did an update to a parent node where the pointers we added pointed
 572          * to child nodes that weren't written yet: now, the child nodes have
 573          * been written so we can write out the update to the interior node.
 574          */
 575
 576         /*
 577          * We can't call into journal reclaim here: we'd block on the journal
 578          * reclaim lock, but we may need to release the open buckets we have
 579          * pinned in order for other btree updates to make forward progress, and
 580          * journal reclaim does btree updates when flushing bkey_cached entries,
 581          * which may require allocations as well.
 582          */
 583         bch2_trans_init(&trans, c, 0, 512);
 584         ret = __bch2_trans_do(&trans, &as->disk_res, &journal_seq,
 585                               BTREE_INSERT_NOFAIL|
 586                               BTREE_INSERT_NOCHECK_RW|
 587                               BTREE_INSERT_JOURNAL_RECLAIM|
 588                               BTREE_INSERT_JOURNAL_RESERVED,
 589                               btree_update_nodes_written_trans(&trans, as));
 590         bch2_trans_exit(&trans);
 591
 592         bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
 593                              "error %i in btree_update_nodes_written()", ret);
 594 err:
 595         if (b) {
 596                 /*
 597                  * @b is the node we did the final insert into:
 598                  *
 599                  * On failure to get a journal reservation, we still have to
 600                  * unblock the write and allow most of the write path to happen
 601                  * so that shutdown works, but the i->journal_seq mechanism
 602                  * won't work to prevent the btree write from being visible (we
 603                  * didn't get a journal sequence number) - instead
 604                  * __bch2_btree_node_write() doesn't do the actual write if
 605                  * we're in journal error state:
 606                  */
 607
 608                 btree_node_lock_type(c, b, SIX_LOCK_intent);
 609                 btree_node_lock_type(c, b, SIX_LOCK_write);
 610                 mutex_lock(&c->btree_interior_update_lock);
 611
 612                 list_del(&as->write_blocked_list);
 613
 614                 /*
 615                  * Node might have been freed, recheck under
 616                  * btree_interior_update_lock:
 617                  */
 618                 if (as->b == b) {
 619                         struct bset *i = btree_bset_last(b);
 620
 621                         BUG_ON(!b->c.level);
 622                         BUG_ON(!btree_node_dirty(b));
 623
 624                         if (!ret) {
 625                                 i->journal_seq = cpu_to_le64(
 626                                         max(journal_seq,
 627                                             le64_to_cpu(i->journal_seq)));
 628
 629                                 bch2_btree_add_journal_pin(c, b, journal_seq);
 630                         } else {
 631                                 /*
 632                                  * If we didn't get a journal sequence number we
 633                                  * can't write this btree node, because recovery
 634                                  * won't know to ignore this write:
 635                                  */
 636                                 set_btree_node_never_write(b);
 637                         }
 638                 }
 639
 640                 mutex_unlock(&c->btree_interior_update_lock);
 641                 six_unlock_write(&b->c.lock);
 642
 643                 btree_node_write_if_need(c, b, SIX_LOCK_intent);
 644                 six_unlock_intent(&b->c.lock);
 645         }
 646
 647         bch2_journal_pin_drop(&c->journal, &as->journal);
 648
 649         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 650
 651         mutex_lock(&c->btree_interior_update_lock);
 652         for (i = 0; i < as->nr_new_nodes; i++) {
 653                 b = as->new_nodes[i];
 654
 655                 BUG_ON(b->will_make_reachable != (unsigned long) as);
 656                 b->will_make_reachable = 0;
 657         }
 658         mutex_unlock(&c->btree_interior_update_lock);
 659
 660         for (i = 0; i < as->nr_new_nodes; i++) {
 661                 b = as->new_nodes[i];
 662
 663                 btree_node_lock_type(c, b, SIX_LOCK_read);
 664                 btree_node_write_if_need(c, b, SIX_LOCK_read);
 665                 six_unlock_read(&b->c.lock);
 666         }
 667
 668         for (i = 0; i < as->nr_open_buckets; i++)
 669                 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
 670
 671         bch2_btree_update_free(as);
 672 }
 673
 674 static void btree_interior_update_work(struct work_struct *work)
 675 {
 676         struct bch_fs *c =
 677                 container_of(work, struct bch_fs, btree_interior_update_work);
 678         struct btree_update *as;
 679
 680         while (1) {
 681                 mutex_lock(&c->btree_interior_update_lock);
 682                 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
 683                                               struct btree_update, unwritten_list);
 684                 if (as && !as->nodes_written)
 685                         as = NULL;
 686                 mutex_unlock(&c->btree_interior_update_lock);
 687
 688                 if (!as)
 689                         break;
 690
 691                 btree_update_nodes_written(as);
 692         }
 693 }
 694
 695 static void btree_update_set_nodes_written(struct closure *cl)
 696 {
 697         struct btree_update *as = container_of(cl, struct btree_update, cl);
 698         struct bch_fs *c = as->c;
 699
 700         mutex_lock(&c->btree_interior_update_lock);
 701         as->nodes_written = true;
 702         mutex_unlock(&c->btree_interior_update_lock);
 703
 704         queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
 705 }
 706
 707 /*
 708  * We're updating @b with pointers to nodes that haven't finished writing yet:
 709  * block @b from being written until @as completes
 710  */
 711 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
 712 {
 713         struct bch_fs *c = as->c;
 714
 715         mutex_lock(&c->btree_interior_update_lock);
 716         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 717
 718         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 719         BUG_ON(!btree_node_dirty(b));
 720
 721         as->mode        = BTREE_INTERIOR_UPDATING_NODE;
 722         as->b           = b;
 723         list_add(&as->write_blocked_list, &b->write_blocked);
 724
 725         mutex_unlock(&c->btree_interior_update_lock);
 726 }
 727
 728 static void btree_update_reparent(struct btree_update *as,
 729                                   struct btree_update *child)
 730 {
 731         struct bch_fs *c = as->c;
 732
 733         lockdep_assert_held(&c->btree_interior_update_lock);
 734
 735         child->b = NULL;
 736         child->mode = BTREE_INTERIOR_UPDATING_AS;
 737
 738         bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL);
 739 }
 740
 741 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
 742 {
 743         struct bkey_i *insert = &b->key;
 744         struct bch_fs *c = as->c;
 745
 746         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 747
 748         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
 749                ARRAY_SIZE(as->journal_entries));
 750
 751         as->journal_u64s +=
 752                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
 753                                   BCH_JSET_ENTRY_btree_root,
 754                                   b->c.btree_id, b->c.level,
 755                                   insert, insert->k.u64s);
 756
 757         mutex_lock(&c->btree_interior_update_lock);
 758         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 759
 760         as->mode        = BTREE_INTERIOR_UPDATING_ROOT;
 761         mutex_unlock(&c->btree_interior_update_lock);
 762 }
 763
 764 /*
 765  * bch2_btree_update_add_new_node:
 766  *
 767  * This causes @as to wait on @b to be written, before it gets to
 768  * bch2_btree_update_nodes_written
 769  *
 770  * Additionally, it sets b->will_make_reachable to prevent any additional writes
 771  * to @b from happening besides the first until @b is reachable on disk
 772  *
 773  * And it adds @b to the list of @as's new nodes, so that we can update sector
 774  * counts in bch2_btree_update_nodes_written:
 775  */
 776 void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
 777 {
 778         struct bch_fs *c = as->c;
 779
 780         closure_get(&as->cl);
 781
 782         mutex_lock(&c->btree_interior_update_lock);
 783         BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
 784         BUG_ON(b->will_make_reachable);
 785
 786         as->new_nodes[as->nr_new_nodes++] = b;
 787         b->will_make_reachable = 1UL|(unsigned long) as;
 788
 789         mutex_unlock(&c->btree_interior_update_lock);
 790
 791         btree_update_will_add_key(as, &b->key);
 792 }
 793
 794 /*
 795  * returns true if @b was a new node
 796  */
 797 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
 798 {
 799         struct btree_update *as;
 800         unsigned long v;
 801         unsigned i;
 802
 803         mutex_lock(&c->btree_interior_update_lock);
 804         /*
 805          * When b->will_make_reachable != 0, it owns a ref on as->cl that's
 806          * dropped when it gets written by bch2_btree_complete_write - the
 807          * xchg() is for synchronization with bch2_btree_complete_write:
 808          */
 809         v = xchg(&b->will_make_reachable, 0);
 810         as = (struct btree_update *) (v & ~1UL);
 811
 812         if (!as) {
 813                 mutex_unlock(&c->btree_interior_update_lock);
 814                 return;
 815         }
 816
 817         for (i = 0; i < as->nr_new_nodes; i++)
 818                 if (as->new_nodes[i] == b)
 819                         goto found;
 820
 821         BUG();
 822 found:
 823         array_remove_item(as->new_nodes, as->nr_new_nodes, i);
 824         mutex_unlock(&c->btree_interior_update_lock);
 825
 826         if (v & 1)
 827                 closure_put(&as->cl);
 828 }
 829
 830 void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
 831 {
 832         while (b->ob.nr)
 833                 as->open_buckets[as->nr_open_buckets++] =
 834                         b->ob.v[--b->ob.nr];
 835 }
 836
 837 /*
 838  * @b is being split/rewritten: it may have pointers to not-yet-written btree
 839  * nodes and thus outstanding btree_updates - redirect @b's
 840  * btree_updates to point to this btree_update:
 841  */
 842 void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 843                                                struct btree *b)
 844 {
 845         struct bch_fs *c = as->c;
 846         struct btree_update *p, *n;
 847         struct btree_write *w;
 848
 849         set_btree_node_dying(b);
 850
 851         if (btree_node_fake(b))
 852                 return;
 853
 854         mutex_lock(&c->btree_interior_update_lock);
 855
 856         /*
 857          * Does this node have any btree_update operations preventing
 858          * it from being written?
 859          *
 860          * If so, redirect them to point to this btree_update: we can
 861          * write out our new nodes, but we won't make them visible until those
 862          * operations complete
 863          */
 864         list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
 865                 list_del_init(&p->write_blocked_list);
 866                 btree_update_reparent(as, p);
 867
 868                 /*
 869                  * for flush_held_btree_writes() waiting on updates to flush or
 870                  * nodes to be writeable:
 871                  */
 872                 closure_wake_up(&c->btree_interior_update_wait);
 873         }
 874
 875         clear_btree_node_dirty(c, b);
 876         clear_btree_node_need_write(b);
 877
 878         /*
 879          * Does this node have unwritten data that has a pin on the journal?
 880          *
 881          * If so, transfer that pin to the btree_update operation -
 882          * note that if we're freeing multiple nodes, we only need to keep the
 883          * oldest pin of any of the nodes we're freeing. We'll release the pin
 884          * when the new nodes are persistent and reachable on disk:
 885          */
 886         w = btree_current_write(b);
 887         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 888         bch2_journal_pin_drop(&c->journal, &w->journal);
 889
 890         w = btree_prev_write(b);
 891         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 892         bch2_journal_pin_drop(&c->journal, &w->journal);
 893
 894         mutex_unlock(&c->btree_interior_update_lock);
 895
 896         /*
 897          * Is this a node that isn't reachable on disk yet?
 898          *
 899          * Nodes that aren't reachable yet have writes blocked until they're
 900          * reachable - now that we've cancelled any pending writes and moved
 901          * things waiting on that write to wait on this update, we can drop this
 902          * node from the list of nodes that the other update is making
 903          * reachable, prior to freeing it:
 904          */
 905         btree_update_drop_new_node(c, b);
 906
 907         btree_update_will_delete_key(as, &b->key);
 908
 909         as->old_nodes[as->nr_old_nodes] = b;
 910         as->old_nodes_seq[as->nr_old_nodes] = b->data->keys.seq;
 911         as->nr_old_nodes++;
 912 }
 913
 914 void bch2_btree_update_done(struct btree_update *as)
 915 {
 916         BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
 917
 918         if (as->took_gc_lock)
 919                 up_read(&as->c->gc_lock);
 920         as->took_gc_lock = false;
 921
 922         bch2_btree_reserve_put(as);
 923
 924         continue_at(&as->cl, btree_update_set_nodes_written,
 925                     as->c->btree_interior_update_worker);
 926 }
 927
 928 struct btree_update *
 929 bch2_btree_update_start(struct btree_iter *iter, unsigned level,
 930                         unsigned nr_nodes, unsigned flags)
 931 {
 932         struct btree_trans *trans = iter->trans;
 933         struct bch_fs *c = trans->c;
 934         struct btree_update *as;
 935         struct closure cl;
 936         int disk_res_flags = (flags & BTREE_INSERT_NOFAIL)
 937                 ? BCH_DISK_RESERVATION_NOFAIL : 0;
 938         int journal_flags = 0;
 939         int ret = 0;
 940
 941         BUG_ON(!iter->should_be_locked);
 942
 943         if (flags & BTREE_INSERT_JOURNAL_RESERVED)
 944                 journal_flags |= JOURNAL_RES_GET_RESERVED;
 945
 946         closure_init_stack(&cl);
 947 retry:
 948
 949         /*
 950          * XXX: figure out how far we might need to split,
 951          * instead of locking/reserving all the way to the root:
 952          */
 953         if (!bch2_btree_iter_upgrade(iter, U8_MAX)) {
 954                 trace_trans_restart_iter_upgrade(trans->ip, _RET_IP_,
 955                                                  iter->btree_id,
 956                                                  &iter->real_pos);
 957                 return ERR_PTR(-EINTR);
 958         }
 959
 960         if (flags & BTREE_INSERT_GC_LOCK_HELD)
 961                 lockdep_assert_held(&c->gc_lock);
 962         else if (!down_read_trylock(&c->gc_lock)) {
 963                 if (flags & BTREE_INSERT_NOUNLOCK)
 964                         return ERR_PTR(-EINTR);
 965
 966                 bch2_trans_unlock(trans);
 967                 down_read(&c->gc_lock);
 968                 if (!bch2_trans_relock(trans)) {
 969                         up_read(&c->gc_lock);
 970                         return ERR_PTR(-EINTR);
 971                 }
 972         }
 973
 974         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
 975         memset(as, 0, sizeof(*as));
 976         closure_init(&as->cl, NULL);
 977         as->c           = c;
 978         as->mode        = BTREE_INTERIOR_NO_UPDATE;
 979         as->took_gc_lock = !(flags & BTREE_INSERT_GC_LOCK_HELD);
 980         as->btree_id    = iter->btree_id;
 981         INIT_LIST_HEAD(&as->list);
 982         INIT_LIST_HEAD(&as->unwritten_list);
 983         INIT_LIST_HEAD(&as->write_blocked_list);
 984         bch2_keylist_init(&as->old_keys, as->_old_keys);
 985         bch2_keylist_init(&as->new_keys, as->_new_keys);
 986         bch2_keylist_init(&as->parent_keys, as->inline_keys);
 987
 988         mutex_lock(&c->btree_interior_update_lock);
 989         list_add_tail(&as->list, &c->btree_interior_update_list);
 990         mutex_unlock(&c->btree_interior_update_lock);
 991
 992         /*
 993          * We don't want to allocate if we're in an error state, that can cause
 994          * deadlock on emergency shutdown due to open buckets getting stuck in
 995          * the btree_reserve_cache after allocator shutdown has cleared it out.
 996          * This check needs to come after adding us to the btree_interior_update
 997          * list but before calling bch2_btree_reserve_get, to synchronize with
 998          * __bch2_fs_read_only().
 999          */
1000         ret = bch2_journal_error(&c->journal);
1001         if (ret)
1002                 goto err;
1003
1004         ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
1005                                       BTREE_UPDATE_JOURNAL_RES,
1006                                       journal_flags|JOURNAL_RES_GET_NONBLOCK);
1007         if (ret == -EAGAIN) {
1008                 /*
1009                  * this would be cleaner if bch2_journal_preres_get() took a
1010                  * closure argument
1011                  */
1012                 if (flags & BTREE_INSERT_NOUNLOCK) {
1013                         trace_trans_restart_journal_preres_get(trans->ip, _RET_IP_);
1014                         ret = -EINTR;
1015                         goto err;
1016                 }
1017
1018                 bch2_trans_unlock(trans);
1019
1020                 if (flags & BTREE_INSERT_JOURNAL_RECLAIM) {
1021                         bch2_btree_update_free(as);
1022                         return ERR_PTR(ret);
1023                 }
1024
1025                 ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
1026                                 BTREE_UPDATE_JOURNAL_RES,
1027                                 journal_flags);
1028                 if (ret) {
1029                         trace_trans_restart_journal_preres_get(trans->ip, _RET_IP_);
1030                         goto err;
1031                 }
1032
1033                 if (!bch2_trans_relock(trans)) {
1034                         ret = -EINTR;
1035                         goto err;
1036                 }
1037         }
1038
1039         ret = bch2_disk_reservation_get(c, &as->disk_res,
1040                         nr_nodes * c->opts.btree_node_size,
1041                         c->opts.metadata_replicas,
1042                         disk_res_flags);
1043         if (ret)
1044                 goto err;
1045
1046         ret = bch2_btree_reserve_get(as, nr_nodes, flags,
1047                 !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1048         if (ret)
1049                 goto err;
1050
1051         bch2_journal_pin_add(&c->journal,
1052                              atomic64_read(&c->journal.seq),
1053                              &as->journal, NULL);
1054
1055         return as;
1056 err:
1057         bch2_btree_update_free(as);
1058
1059         if (ret == -EAGAIN) {
1060                 BUG_ON(flags & BTREE_INSERT_NOUNLOCK);
1061
1062                 bch2_trans_unlock(trans);
1063                 closure_sync(&cl);
1064                 ret = -EINTR;
1065         }
1066
1067         if (ret == -EINTR && bch2_trans_relock(trans))
1068                 goto retry;
1069
1070         return ERR_PTR(ret);
1071 }
1072
1073 /* Btree root updates: */
1074
1075 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1076 {
1077         /* Root nodes cannot be reaped */
1078         mutex_lock(&c->btree_cache.lock);
1079         list_del_init(&b->list);
1080         mutex_unlock(&c->btree_cache.lock);
1081
1082         if (b->c.level)
1083                 six_lock_pcpu_alloc(&b->c.lock);
1084         else
1085                 six_lock_pcpu_free(&b->c.lock);
1086
1087         mutex_lock(&c->btree_root_lock);
1088         BUG_ON(btree_node_root(c, b) &&
1089                (b->c.level < btree_node_root(c, b)->c.level ||
1090                 !btree_node_dying(btree_node_root(c, b))));
1091
1092         btree_node_root(c, b) = b;
1093         mutex_unlock(&c->btree_root_lock);
1094
1095         bch2_recalc_btree_reserve(c);
1096 }
1097
1098 /**
1099  * bch_btree_set_root - update the root in memory and on disk
1100  *
1101  * To ensure forward progress, the current task must not be holding any
1102  * btree node write locks. However, you must hold an intent lock on the
1103  * old root.
1104  *
1105  * Note: This allocates a journal entry but doesn't add any keys to
1106  * it.  All the btree roots are part of every journal write, so there
1107  * is nothing new to be done.  This just guarantees that there is a
1108  * journal write.
1109  */
1110 static void bch2_btree_set_root(struct btree_update *as, struct btree *b,
1111                                 struct btree_iter *iter)
1112 {
1113         struct bch_fs *c = as->c;
1114         struct btree *old;
1115
1116         trace_btree_set_root(c, b);
1117         BUG_ON(!b->written &&
1118                !test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags));
1119
1120         old = btree_node_root(c, b);
1121
1122         /*
1123          * Ensure no one is using the old root while we switch to the
1124          * new root:
1125          */
1126         bch2_btree_node_lock_write(old, iter);
1127
1128         bch2_btree_set_root_inmem(c, b);
1129
1130         btree_update_updated_root(as, b);
1131
1132         /*
1133          * Unlock old root after new root is visible:
1134          *
1135          * The new root isn't persistent, but that's ok: we still have
1136          * an intent lock on the new root, and any updates that would
1137          * depend on the new root would have to update the new root.
1138          */
1139         bch2_btree_node_unlock_write(old, iter);
1140 }
1141
1142 /* Interior node updates: */
1143
1144 static void bch2_insert_fixup_btree_ptr(struct btree_update *as, struct btree *b,
1145                                         struct btree_iter *iter,
1146                                         struct bkey_i *insert,
1147                                         struct btree_node_iter *node_iter)
1148 {
1149         struct bch_fs *c = as->c;
1150         struct bkey_packed *k;
1151         const char *invalid;
1152
1153         BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 &&
1154                !btree_ptr_sectors_written(insert));
1155
1156         invalid = bch2_bkey_invalid(c, bkey_i_to_s_c(insert), btree_node_type(b)) ?:
1157                 bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert));
1158         if (invalid) {
1159                 char buf[160];
1160
1161                 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(insert));
1162                 bch2_fs_inconsistent(c, "inserting invalid bkey %s: %s", buf, invalid);
1163                 dump_stack();
1164         }
1165
1166         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1167                ARRAY_SIZE(as->journal_entries));
1168
1169         as->journal_u64s +=
1170                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1171                                   BCH_JSET_ENTRY_btree_keys,
1172                                   b->c.btree_id, b->c.level,
1173                                   insert, insert->k.u64s);
1174
1175         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1176                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1177                 bch2_btree_node_iter_advance(node_iter, b);
1178
1179         bch2_btree_bset_insert_key(iter, b, node_iter, insert);
1180         set_btree_node_dirty(c, b);
1181         set_btree_node_need_write(b);
1182 }
1183
1184 static void
1185 __bch2_btree_insert_keys_interior(struct btree_update *as, struct btree *b,
1186                                   struct btree_iter *iter, struct keylist *keys,
1187                                   struct btree_node_iter node_iter)
1188 {
1189         struct bkey_i *insert = bch2_keylist_front(keys);
1190         struct bkey_packed *k;
1191
1192         BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1193
1194         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1195                (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1196                 ;
1197
1198         while (!bch2_keylist_empty(keys)) {
1199                 bch2_insert_fixup_btree_ptr(as, b, iter,
1200                                 bch2_keylist_front(keys), &node_iter);
1201                 bch2_keylist_pop_front(keys);
1202         }
1203 }
1204
1205 /*
1206  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1207  * node)
1208  */
1209 static struct btree *__btree_split_node(struct btree_update *as,
1210                                         struct btree *n1,
1211                                         struct btree_iter *iter)
1212 {
1213         struct bkey_format_state s;
1214         size_t nr_packed = 0, nr_unpacked = 0;
1215         struct btree *n2;
1216         struct bset *set1, *set2;
1217         struct bkey_packed *k, *set2_start, *set2_end, *out, *prev = NULL;
1218         struct bpos n1_pos;
1219
1220         n2 = bch2_btree_node_alloc(as, n1->c.level);
1221         bch2_btree_update_add_new_node(as, n2);
1222
1223         n2->data->max_key       = n1->data->max_key;
1224         n2->data->format        = n1->format;
1225         SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data));
1226         n2->key.k.p = n1->key.k.p;
1227
1228         set1 = btree_bset_first(n1);
1229         set2 = btree_bset_first(n2);
1230
1231         /*
1232          * Has to be a linear search because we don't have an auxiliary
1233          * search tree yet
1234          */
1235         k = set1->start;
1236         while (1) {
1237                 struct bkey_packed *n = bkey_next(k);
1238
1239                 if (n == vstruct_last(set1))
1240                         break;
1241                 if (k->_data - set1->_data >= (le16_to_cpu(set1->u64s) * 3) / 5)
1242                         break;
1243
1244                 if (bkey_packed(k))
1245                         nr_packed++;
1246                 else
1247                         nr_unpacked++;
1248
1249                 prev = k;
1250                 k = n;
1251         }
1252
1253         BUG_ON(!prev);
1254         set2_start      = k;
1255         set2_end        = vstruct_last(set1);
1256
1257         set1->u64s = cpu_to_le16((u64 *) set2_start - set1->_data);
1258         set_btree_bset_end(n1, n1->set);
1259
1260         n1->nr.live_u64s        = le16_to_cpu(set1->u64s);
1261         n1->nr.bset_u64s[0]     = le16_to_cpu(set1->u64s);
1262         n1->nr.packed_keys      = nr_packed;
1263         n1->nr.unpacked_keys    = nr_unpacked;
1264
1265         n1_pos = bkey_unpack_pos(n1, prev);
1266         if (as->c->sb.version < bcachefs_metadata_version_snapshot)
1267                 n1_pos.snapshot = U32_MAX;
1268
1269         btree_set_max(n1, n1_pos);
1270         btree_set_min(n2, bpos_successor(n1->key.k.p));
1271
1272         bch2_bkey_format_init(&s);
1273         bch2_bkey_format_add_pos(&s, n2->data->min_key);
1274         bch2_bkey_format_add_pos(&s, n2->data->max_key);
1275
1276         for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1277                 struct bkey uk = bkey_unpack_key(n1, k);
1278                 bch2_bkey_format_add_key(&s, &uk);
1279         }
1280
1281         n2->data->format = bch2_bkey_format_done(&s);
1282         btree_node_set_format(n2, n2->data->format);
1283
1284         out = set2->start;
1285         memset(&n2->nr, 0, sizeof(n2->nr));
1286
1287         for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1288                 BUG_ON(!bch2_bkey_transform(&n2->format, out, bkey_packed(k)
1289                                        ? &n1->format : &bch2_bkey_format_current, k));
1290                 out->format = KEY_FORMAT_LOCAL_BTREE;
1291                 btree_keys_account_key_add(&n2->nr, 0, out);
1292                 out = bkey_next(out);
1293         }
1294
1295         set2->u64s = cpu_to_le16((u64 *) out - set2->_data);
1296         set_btree_bset_end(n2, n2->set);
1297
1298         BUG_ON(!set1->u64s);
1299         BUG_ON(!set2->u64s);
1300
1301         btree_node_reset_sib_u64s(n1);
1302         btree_node_reset_sib_u64s(n2);
1303
1304         bch2_verify_btree_nr_keys(n1);
1305         bch2_verify_btree_nr_keys(n2);
1306
1307         if (n1->c.level) {
1308                 btree_node_interior_verify(as->c, n1);
1309                 btree_node_interior_verify(as->c, n2);
1310         }
1311
1312         return n2;
1313 }
1314
1315 /*
1316  * For updates to interior nodes, we've got to do the insert before we split
1317  * because the stuff we're inserting has to be inserted atomically. Post split,
1318  * the keys might have to go in different nodes and the split would no longer be
1319  * atomic.
1320  *
1321  * Worse, if the insert is from btree node coalescing, if we do the insert after
1322  * we do the split (and pick the pivot) - the pivot we pick might be between
1323  * nodes that were coalesced, and thus in the middle of a child node post
1324  * coalescing:
1325  */
1326 static void btree_split_insert_keys(struct btree_update *as, struct btree *b,
1327                                     struct btree_iter *iter,
1328                                     struct keylist *keys)
1329 {
1330         struct btree_node_iter node_iter;
1331         struct bkey_i *k = bch2_keylist_front(keys);
1332         struct bkey_packed *src, *dst, *n;
1333         struct bset *i;
1334
1335         bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
1336
1337         __bch2_btree_insert_keys_interior(as, b, iter, keys, node_iter);
1338
1339         /*
1340          * We can't tolerate whiteouts here - with whiteouts there can be
1341          * duplicate keys, and it would be rather bad if we picked a duplicate
1342          * for the pivot:
1343          */
1344         i = btree_bset_first(b);
1345         src = dst = i->start;
1346         while (src != vstruct_last(i)) {
1347                 n = bkey_next(src);
1348                 if (!bkey_deleted(src)) {
1349                         memmove_u64s_down(dst, src, src->u64s);
1350                         dst = bkey_next(dst);
1351                 }
1352                 src = n;
1353         }
1354
1355         /* Also clear out the unwritten whiteouts area: */
1356         b->whiteout_u64s = 0;
1357
1358         i->u64s = cpu_to_le16((u64 *) dst - i->_data);
1359         set_btree_bset_end(b, b->set);
1360
1361         BUG_ON(b->nsets != 1 ||
1362                b->nr.live_u64s != le16_to_cpu(btree_bset_first(b)->u64s));
1363
1364         btree_node_interior_verify(as->c, b);
1365 }
1366
1367 static void btree_split(struct btree_update *as,
1368                         struct btree_trans *trans, struct btree_iter *iter,
1369                         struct btree *b, struct keylist *keys,
1370                         unsigned flags)
1371 {
1372         struct bch_fs *c = as->c;
1373         struct btree *parent = btree_node_parent(iter, b);
1374         struct btree *n1, *n2 = NULL, *n3 = NULL;
1375         u64 start_time = local_clock();
1376
1377         BUG_ON(!parent && (b != btree_node_root(c, b)));
1378         BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->c.level));
1379
1380         bch2_btree_interior_update_will_free_node(as, b);
1381
1382         n1 = bch2_btree_node_alloc_replacement(as, b);
1383         bch2_btree_update_add_new_node(as, n1);
1384
1385         if (keys)
1386                 btree_split_insert_keys(as, n1, iter, keys);
1387
1388         if (bset_u64s(&n1->set[0]) > BTREE_SPLIT_THRESHOLD(c)) {
1389                 trace_btree_split(c, b);
1390
1391                 n2 = __btree_split_node(as, n1, iter);
1392
1393                 bch2_btree_build_aux_trees(n2);
1394                 bch2_btree_build_aux_trees(n1);
1395                 six_unlock_write(&n2->c.lock);
1396                 six_unlock_write(&n1->c.lock);
1397
1398                 bch2_btree_node_write(c, n1, SIX_LOCK_intent);
1399                 bch2_btree_node_write(c, n2, SIX_LOCK_intent);
1400
1401                 /*
1402                  * Note that on recursive parent_keys == keys, so we
1403                  * can't start adding new keys to parent_keys before emptying it
1404                  * out (which we did with btree_split_insert_keys() above)
1405                  */
1406                 bch2_keylist_add(&as->parent_keys, &n1->key);
1407                 bch2_keylist_add(&as->parent_keys, &n2->key);
1408
1409                 if (!parent) {
1410                         /* Depth increases, make a new root */
1411                         n3 = __btree_root_alloc(as, b->c.level + 1);
1412
1413                         n3->sib_u64s[0] = U16_MAX;
1414                         n3->sib_u64s[1] = U16_MAX;
1415
1416                         btree_split_insert_keys(as, n3, iter, &as->parent_keys);
1417
1418                         bch2_btree_node_write(c, n3, SIX_LOCK_intent);
1419                 }
1420         } else {
1421                 trace_btree_compact(c, b);
1422
1423                 bch2_btree_build_aux_trees(n1);
1424                 six_unlock_write(&n1->c.lock);
1425
1426                 bch2_btree_node_write(c, n1, SIX_LOCK_intent);
1427
1428                 if (parent)
1429                         bch2_keylist_add(&as->parent_keys, &n1->key);
1430         }
1431
1432         /* New nodes all written, now make them visible: */
1433
1434         if (parent) {
1435                 /* Split a non root node */
1436                 bch2_btree_insert_node(as, trans, iter, parent, &as->parent_keys, flags);
1437         } else if (n3) {
1438                 bch2_btree_set_root(as, n3, iter);
1439         } else {
1440                 /* Root filled up but didn't need to be split */
1441                 bch2_btree_set_root(as, n1, iter);
1442         }
1443
1444         bch2_btree_update_get_open_buckets(as, n1);
1445         if (n2)
1446                 bch2_btree_update_get_open_buckets(as, n2);
1447         if (n3)
1448                 bch2_btree_update_get_open_buckets(as, n3);
1449
1450         /* Successful split, update the iterator to point to the new nodes: */
1451
1452         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1453         bch2_btree_iter_node_drop(iter, b);
1454         if (n3)
1455                 bch2_btree_iter_node_replace(iter, n3);
1456         if (n2)
1457                 bch2_btree_iter_node_replace(iter, n2);
1458         bch2_btree_iter_node_replace(iter, n1);
1459
1460         /*
1461          * The old node must be freed (in memory) _before_ unlocking the new
1462          * nodes - else another thread could re-acquire a read lock on the old
1463          * node after another thread has locked and updated the new node, thus
1464          * seeing stale data:
1465          */
1466         bch2_btree_node_free_inmem(c, b, iter);
1467
1468         if (n3)
1469                 six_unlock_intent(&n3->c.lock);
1470         if (n2)
1471                 six_unlock_intent(&n2->c.lock);
1472         six_unlock_intent(&n1->c.lock);
1473
1474         bch2_btree_trans_verify_locks(trans);
1475
1476         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_split],
1477                                start_time);
1478 }
1479
1480 static void
1481 bch2_btree_insert_keys_interior(struct btree_update *as, struct btree *b,
1482                                 struct btree_iter *iter, struct keylist *keys)
1483 {
1484         struct btree_iter *linked;
1485
1486         __bch2_btree_insert_keys_interior(as, b, iter, keys, iter->l[b->c.level].iter);
1487
1488         btree_update_updated_node(as, b);
1489
1490         trans_for_each_iter_with_node(iter->trans, b, linked)
1491                 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1492
1493         bch2_btree_trans_verify_iters(iter->trans, b);
1494 }
1495
1496 /**
1497  * bch_btree_insert_node - insert bkeys into a given btree node
1498  *
1499  * @iter:               btree iterator
1500  * @keys:               list of keys to insert
1501  * @hook:               insert callback
1502  * @persistent:         if not null, @persistent will wait on journal write
1503  *
1504  * Inserts as many keys as it can into a given btree node, splitting it if full.
1505  * If a split occurred, this function will return early. This can only happen
1506  * for leaf nodes -- inserts into interior nodes have to be atomic.
1507  */
1508 static void bch2_btree_insert_node(struct btree_update *as,
1509                                    struct btree_trans *trans, struct btree_iter *iter,
1510                                    struct btree *b, struct keylist *keys,
1511                                    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(iter, 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, iter, b);
1525
1526         if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1527                 bch2_btree_node_unlock_write(b, iter);
1528                 goto split;
1529         }
1530
1531         btree_node_interior_verify(c, b);
1532
1533         bch2_btree_insert_keys_interior(as, b, iter, 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_btree_iter_reinit_node(iter, b);
1546
1547         bch2_btree_node_unlock_write(b, iter);
1548
1549         btree_node_interior_verify(c, b);
1550         return;
1551 split:
1552         btree_split(as, trans, iter, b, keys, flags);
1553 }
1554
1555 int bch2_btree_split_leaf(struct btree_trans *trans,
1556                           struct btree_iter *iter,
1557                           unsigned flags)
1558 {
1559         struct bch_fs *c = trans->c;
1560         struct btree *b = iter_l(iter)->b;
1561         struct btree_update *as;
1562         unsigned l;
1563         int ret = 0;
1564
1565         as = bch2_btree_update_start(iter, iter->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, iter, b, NULL, flags);
1571         bch2_btree_update_done(as);
1572
1573         for (l = iter->level + 1; btree_iter_node(iter, l) && !ret; l++)
1574                 ret = bch2_foreground_maybe_merge(trans, iter, l, flags);
1575
1576         return ret;
1577 }
1578
1579 int __bch2_foreground_maybe_merge(struct btree_trans *trans,
1580                                   struct btree_iter *iter,
1581                                   unsigned level,
1582                                   unsigned flags,
1583                                   enum btree_node_sibling sib)
1584 {
1585         struct bch_fs *c = trans->c;
1586         struct btree_iter *sib_iter = 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         int ret = 0, ret2 = 0;
1595
1596         BUG_ON(!btree_node_locked(iter, level));
1597 retry:
1598         ret = bch2_btree_iter_traverse(iter);
1599         if (ret)
1600                 goto err;
1601
1602         BUG_ON(!btree_node_locked(iter, level));
1603
1604         b = iter->l[level].b;
1605
1606         if ((sib == btree_prev_sib && !bpos_cmp(b->data->min_key, POS_MIN)) ||
1607             (sib == btree_next_sib && !bpos_cmp(b->data->max_key, SPOS_MAX))) {
1608                 b->sib_u64s[sib] = U16_MAX;
1609                 goto out;
1610         }
1611
1612         sib_pos = sib == btree_prev_sib
1613                 ? bpos_predecessor(b->data->min_key)
1614                 : bpos_successor(b->data->max_key);
1615
1616         sib_iter = bch2_trans_get_node_iter(trans, iter->btree_id,
1617                                             sib_pos, U8_MAX, level,
1618                                             BTREE_ITER_INTENT);
1619         ret = bch2_btree_iter_traverse(sib_iter);
1620         if (ret)
1621                 goto err;
1622
1623         m = sib_iter->l[level].b;
1624
1625         if (btree_node_parent(iter, b) !=
1626             btree_node_parent(sib_iter, m)) {
1627                 b->sib_u64s[sib] = U16_MAX;
1628                 goto out;
1629         }
1630
1631         if (sib == btree_prev_sib) {
1632                 prev = m;
1633                 next = b;
1634         } else {
1635                 prev = b;
1636                 next = m;
1637         }
1638
1639         if (bkey_cmp(bpos_successor(prev->data->max_key), next->data->min_key)) {
1640                 char buf1[100], buf2[100];
1641
1642                 bch2_bpos_to_text(&PBUF(buf1), prev->data->max_key);
1643                 bch2_bpos_to_text(&PBUF(buf2), next->data->min_key);
1644                 bch_err(c,
1645                         "btree topology error in btree merge:\n"
1646                         "  prev ends at   %s\n"
1647                         "  next starts at %s",
1648                         buf1, buf2);
1649                 bch2_topology_error(c);
1650                 ret = -EIO;
1651                 goto err;
1652         }
1653
1654         bch2_bkey_format_init(&new_s);
1655         bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
1656         __bch2_btree_calc_format(&new_s, prev);
1657         __bch2_btree_calc_format(&new_s, next);
1658         bch2_bkey_format_add_pos(&new_s, next->data->max_key);
1659         new_f = bch2_bkey_format_done(&new_s);
1660
1661         sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1662                 btree_node_u64s_with_format(m, &new_f);
1663
1664         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1665                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1666                 sib_u64s /= 2;
1667                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1668         }
1669
1670         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1671         sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
1672         b->sib_u64s[sib] = sib_u64s;
1673
1674         if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
1675                 goto out;
1676
1677         parent = btree_node_parent(iter, b);
1678         as = bch2_btree_update_start(iter, level,
1679                          btree_update_reserve_required(c, parent) + 1,
1680                          flags|
1681                          BTREE_INSERT_NOFAIL|
1682                          BTREE_INSERT_USE_RESERVE);
1683         ret = PTR_ERR_OR_ZERO(as);
1684         if (ret)
1685                 goto err;
1686
1687         trace_btree_merge(c, b);
1688
1689         bch2_btree_interior_update_will_free_node(as, b);
1690         bch2_btree_interior_update_will_free_node(as, m);
1691
1692         n = bch2_btree_node_alloc(as, b->c.level);
1693         bch2_btree_update_add_new_node(as, n);
1694
1695         btree_set_min(n, prev->data->min_key);
1696         btree_set_max(n, next->data->max_key);
1697         n->data->format         = new_f;
1698
1699         btree_node_set_format(n, new_f);
1700
1701         bch2_btree_sort_into(c, n, prev);
1702         bch2_btree_sort_into(c, n, next);
1703
1704         bch2_btree_build_aux_trees(n);
1705         six_unlock_write(&n->c.lock);
1706
1707         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1708
1709         bkey_init(&delete.k);
1710         delete.k.p = prev->key.k.p;
1711         bch2_keylist_add(&as->parent_keys, &delete);
1712         bch2_keylist_add(&as->parent_keys, &n->key);
1713
1714         bch2_btree_insert_node(as, trans, iter, parent, &as->parent_keys, flags);
1715
1716         bch2_btree_update_get_open_buckets(as, n);
1717
1718         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1719         six_lock_increment(&m->c.lock, SIX_LOCK_intent);
1720         bch2_btree_iter_node_drop(iter, b);
1721         bch2_btree_iter_node_drop(iter, m);
1722
1723         bch2_btree_iter_node_replace(iter, n);
1724
1725         bch2_btree_trans_verify_iters(trans, n);
1726
1727         bch2_btree_node_free_inmem(c, b, iter);
1728         bch2_btree_node_free_inmem(c, m, iter);
1729
1730         six_unlock_intent(&n->c.lock);
1731
1732         bch2_btree_update_done(as);
1733 out:
1734         bch2_btree_trans_verify_locks(trans);
1735         bch2_trans_iter_free(trans, sib_iter);
1736
1737         /*
1738          * Don't downgrade locks here: we're called after successful insert,
1739          * and the caller will downgrade locks after a successful insert
1740          * anyways (in case e.g. a split was required first)
1741          *
1742          * And we're also called when inserting into interior nodes in the
1743          * split path, and downgrading to read locks in there is potentially
1744          * confusing:
1745          */
1746         return ret ?: ret2;
1747 err:
1748         bch2_trans_iter_put(trans, sib_iter);
1749         sib_iter = NULL;
1750
1751         if (ret == -EINTR && bch2_trans_relock(trans))
1752                 goto retry;
1753
1754         if (ret == -EINTR && !(flags & BTREE_INSERT_NOUNLOCK)) {
1755                 ret2 = ret;
1756                 ret = bch2_btree_iter_traverse_all(trans);
1757                 if (!ret)
1758                         goto retry;
1759         }
1760
1761         goto out;
1762 }
1763
1764 /**
1765  * bch_btree_node_rewrite - Rewrite/move a btree node
1766  */
1767 int bch2_btree_node_rewrite(struct btree_trans *trans,
1768                             struct btree_iter *iter,
1769                             __le64 seq, unsigned flags)
1770 {
1771         struct bch_fs *c = trans->c;
1772         struct btree *b, *n, *parent;
1773         struct btree_update *as;
1774         int ret;
1775
1776         flags |= BTREE_INSERT_NOFAIL;
1777 retry:
1778         ret = bch2_btree_iter_traverse(iter);
1779         if (ret)
1780                 goto out;
1781
1782         b = bch2_btree_iter_peek_node(iter);
1783         if (!b || b->data->keys.seq != seq)
1784                 goto out;
1785
1786         parent = btree_node_parent(iter, b);
1787         as = bch2_btree_update_start(iter, b->c.level,
1788                 (parent
1789                  ? btree_update_reserve_required(c, parent)
1790                  : 0) + 1,
1791                 flags);
1792         ret = PTR_ERR_OR_ZERO(as);
1793         if (ret == -EINTR)
1794                 goto retry;
1795         if (ret) {
1796                 trace_btree_gc_rewrite_node_fail(c, b);
1797                 goto out;
1798         }
1799
1800         bch2_btree_interior_update_will_free_node(as, b);
1801
1802         n = bch2_btree_node_alloc_replacement(as, b);
1803         bch2_btree_update_add_new_node(as, n);
1804
1805         bch2_btree_build_aux_trees(n);
1806         six_unlock_write(&n->c.lock);
1807
1808         trace_btree_gc_rewrite_node(c, b);
1809
1810         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1811
1812         if (parent) {
1813                 bch2_keylist_add(&as->parent_keys, &n->key);
1814                 bch2_btree_insert_node(as, trans, iter, parent,
1815                                        &as->parent_keys, flags);
1816         } else {
1817                 bch2_btree_set_root(as, n, iter);
1818         }
1819
1820         bch2_btree_update_get_open_buckets(as, n);
1821
1822         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1823         bch2_btree_iter_node_drop(iter, b);
1824         bch2_btree_iter_node_replace(iter, n);
1825         bch2_btree_node_free_inmem(c, b, iter);
1826         six_unlock_intent(&n->c.lock);
1827
1828         bch2_btree_update_done(as);
1829 out:
1830         bch2_btree_iter_downgrade(iter);
1831         return ret;
1832 }
1833
1834 struct async_btree_rewrite {
1835         struct bch_fs           *c;
1836         struct work_struct      work;
1837         enum btree_id           btree_id;
1838         unsigned                level;
1839         struct bpos             pos;
1840         __le64                  seq;
1841 };
1842
1843 void async_btree_node_rewrite_work(struct work_struct *work)
1844 {
1845         struct async_btree_rewrite *a =
1846                 container_of(work, struct async_btree_rewrite, work);
1847         struct bch_fs *c = a->c;
1848         struct btree_trans trans;
1849         struct btree_iter *iter;
1850
1851         bch2_trans_init(&trans, c, 0, 0);
1852         iter = bch2_trans_get_node_iter(&trans, a->btree_id, a->pos,
1853                                         BTREE_MAX_DEPTH, a->level, 0);
1854         bch2_btree_node_rewrite(&trans, iter, a->seq, 0);
1855         bch2_trans_iter_put(&trans, iter);
1856         bch2_trans_exit(&trans);
1857         percpu_ref_put(&c->writes);
1858         kfree(a);
1859 }
1860
1861 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
1862 {
1863         struct async_btree_rewrite *a;
1864
1865         if (!test_bit(BCH_FS_BTREE_INTERIOR_REPLAY_DONE, &c->flags))
1866                 return;
1867
1868         if (!percpu_ref_tryget(&c->writes))
1869                 return;
1870
1871         a = kmalloc(sizeof(*a), GFP_NOFS);
1872         if (!a) {
1873                 percpu_ref_put(&c->writes);
1874                 return;
1875         }
1876
1877         a->c            = c;
1878         a->btree_id     = b->c.btree_id;
1879         a->level        = b->c.level;
1880         a->pos          = b->key.k.p;
1881         a->seq          = b->data->keys.seq;
1882
1883         INIT_WORK(&a->work, async_btree_node_rewrite_work);
1884         queue_work(c->btree_interior_update_worker, &a->work);
1885 }
1886
1887 static int __bch2_btree_node_update_key(struct btree_trans *trans,
1888                                         struct btree_iter *iter,
1889                                         struct btree *b, struct btree *new_hash,
1890                                         struct bkey_i *new_key,
1891                                         bool skip_triggers)
1892 {
1893         struct bch_fs *c = trans->c;
1894         struct btree_iter *iter2 = NULL;
1895         struct btree *parent;
1896         u64 journal_entries[BKEY_BTREE_PTR_U64s_MAX];
1897         int ret;
1898
1899         if (!skip_triggers) {
1900                 ret = bch2_trans_mark_key(trans,
1901                                           bkey_s_c_null,
1902                                           bkey_i_to_s_c(new_key),
1903                                           BTREE_TRIGGER_INSERT);
1904                 if (ret)
1905                         return ret;
1906
1907                 ret = bch2_trans_mark_key(trans,
1908                                           bkey_i_to_s_c(&b->key),
1909                                           bkey_s_c_null,
1910                                           BTREE_TRIGGER_OVERWRITE);
1911                 if (ret)
1912                         return ret;
1913         }
1914
1915         if (new_hash) {
1916                 bkey_copy(&new_hash->key, new_key);
1917                 ret = bch2_btree_node_hash_insert(&c->btree_cache,
1918                                 new_hash, b->c.level, b->c.btree_id);
1919                 BUG_ON(ret);
1920         }
1921
1922         parent = btree_node_parent(iter, b);
1923         if (parent) {
1924                 iter2 = bch2_trans_copy_iter(trans, iter);
1925
1926                 BUG_ON(iter2->level != b->c.level);
1927                 BUG_ON(bpos_cmp(iter2->pos, new_key->k.p));
1928
1929                 btree_node_unlock(iter2, iter2->level);
1930                 iter2->l[iter2->level].b = BTREE_ITER_NO_NODE_UP;
1931                 iter2->level++;
1932
1933                 ret   = bch2_btree_iter_traverse(iter2) ?:
1934                         bch2_trans_update(trans, iter2, new_key, BTREE_TRIGGER_NORUN);
1935                 if (ret)
1936                         goto err;
1937         } else {
1938                 BUG_ON(btree_node_root(c, b) != b);
1939
1940                 trans->extra_journal_entries = (void *) &journal_entries[0];
1941                 trans->extra_journal_entry_u64s =
1942                         journal_entry_set((void *) &journal_entries[0],
1943                                           BCH_JSET_ENTRY_btree_root,
1944                                           b->c.btree_id, b->c.level,
1945                                           new_key, new_key->k.u64s);
1946         }
1947
1948         ret = bch2_trans_commit(trans, NULL, NULL,
1949                                 BTREE_INSERT_NOFAIL|
1950                                 BTREE_INSERT_NOCHECK_RW|
1951                                 BTREE_INSERT_JOURNAL_RECLAIM|
1952                                 BTREE_INSERT_JOURNAL_RESERVED|
1953                                 BTREE_INSERT_NOUNLOCK);
1954         if (ret)
1955                 goto err;
1956
1957         bch2_btree_node_lock_write(b, iter);
1958
1959         if (new_hash) {
1960                 mutex_lock(&c->btree_cache.lock);
1961                 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
1962                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1963
1964                 bkey_copy(&b->key, new_key);
1965                 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1966                 BUG_ON(ret);
1967                 mutex_unlock(&c->btree_cache.lock);
1968         } else {
1969                 bkey_copy(&b->key, new_key);
1970         }
1971
1972         bch2_btree_node_unlock_write(b, iter);
1973 out:
1974         bch2_trans_iter_put(trans, iter2);
1975         return ret;
1976 err:
1977         if (new_hash) {
1978                 mutex_lock(&c->btree_cache.lock);
1979                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1980                 mutex_unlock(&c->btree_cache.lock);
1981         }
1982         goto out;
1983 }
1984
1985 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
1986                                struct btree *b, struct bkey_i *new_key,
1987                                bool skip_triggers)
1988 {
1989         struct bch_fs *c = trans->c;
1990         struct btree *new_hash = NULL;
1991         struct closure cl;
1992         int ret = 0;
1993
1994         closure_init_stack(&cl);
1995
1996         /*
1997          * check btree_ptr_hash_val() after @b is locked by
1998          * btree_iter_traverse():
1999          */
2000         if (btree_ptr_hash_val(new_key) != b->hash_val) {
2001                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2002                 if (ret) {
2003                         bch2_trans_unlock(trans);
2004                         closure_sync(&cl);
2005                         if (!bch2_trans_relock(trans))
2006                                 return -EINTR;
2007                 }
2008
2009                 new_hash = bch2_btree_node_mem_alloc(c);
2010         }
2011
2012         ret = __bch2_btree_node_update_key(trans, iter, b, new_hash,
2013                                            new_key, skip_triggers);
2014
2015         if (new_hash) {
2016                 mutex_lock(&c->btree_cache.lock);
2017                 list_move(&new_hash->list, &c->btree_cache.freeable);
2018                 mutex_unlock(&c->btree_cache.lock);
2019
2020                 six_unlock_write(&new_hash->c.lock);
2021                 six_unlock_intent(&new_hash->c.lock);
2022         }
2023         closure_sync(&cl);
2024         bch2_btree_cache_cannibalize_unlock(c);
2025         return ret;
2026 }
2027
2028 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
2029                                         struct btree *b, struct bkey_i *new_key,
2030                                         bool skip_triggers)
2031 {
2032         struct btree_iter *iter;
2033         int ret;
2034
2035         iter = bch2_trans_get_node_iter(trans, b->c.btree_id, b->key.k.p,
2036                                         BTREE_MAX_DEPTH, b->c.level,
2037                                         BTREE_ITER_INTENT);
2038         ret = bch2_btree_iter_traverse(iter);
2039         if (ret)
2040                 goto out;
2041
2042         /* has node been freed? */
2043         if (iter->l[b->c.level].b != b) {
2044                 /* node has been freed: */
2045                 BUG_ON(!btree_node_dying(b));
2046                 goto out;
2047         }
2048
2049         BUG_ON(!btree_node_hashed(b));
2050
2051         ret = bch2_btree_node_update_key(trans, iter, b, new_key, skip_triggers);
2052 out:
2053         bch2_trans_iter_put(trans, iter);
2054         return ret;
2055 }
2056
2057 /* Init code: */
2058
2059 /*
2060  * Only for filesystem bringup, when first reading the btree roots or allocating
2061  * btree roots when initializing a new filesystem:
2062  */
2063 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2064 {
2065         BUG_ON(btree_node_root(c, b));
2066
2067         bch2_btree_set_root_inmem(c, b);
2068 }
2069
2070 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2071 {
2072         struct closure cl;
2073         struct btree *b;
2074         int ret;
2075
2076         closure_init_stack(&cl);
2077
2078         do {
2079                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2080                 closure_sync(&cl);
2081         } while (ret);
2082
2083         b = bch2_btree_node_mem_alloc(c);
2084         bch2_btree_cache_cannibalize_unlock(c);
2085
2086         set_btree_node_fake(b);
2087         set_btree_node_need_rewrite(b);
2088         b->c.level      = 0;
2089         b->c.btree_id   = id;
2090
2091         bkey_btree_ptr_init(&b->key);
2092         b->key.k.p = SPOS_MAX;
2093         *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2094
2095         bch2_bset_init_first(b, &b->data->keys);
2096         bch2_btree_build_aux_trees(b);
2097
2098         b->data->flags = 0;
2099         btree_set_min(b, POS_MIN);
2100         btree_set_max(b, SPOS_MAX);
2101         b->data->format = bch2_btree_calc_format(b);
2102         btree_node_set_format(b, b->data->format);
2103
2104         ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2105                                           b->c.level, b->c.btree_id);
2106         BUG_ON(ret);
2107
2108         bch2_btree_set_root_inmem(c, b);
2109
2110         six_unlock_write(&b->c.lock);
2111         six_unlock_intent(&b->c.lock);
2112 }
2113
2114 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2115 {
2116         struct btree_update *as;
2117
2118         mutex_lock(&c->btree_interior_update_lock);
2119         list_for_each_entry(as, &c->btree_interior_update_list, list)
2120                 pr_buf(out, "%p m %u w %u r %u j %llu\n",
2121                        as,
2122                        as->mode,
2123                        as->nodes_written,
2124                        atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2125                        as->journal.seq);
2126         mutex_unlock(&c->btree_interior_update_lock);
2127 }
2128
2129 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *c)
2130 {
2131         size_t ret = 0;
2132         struct list_head *i;
2133
2134         mutex_lock(&c->btree_interior_update_lock);
2135         list_for_each(i, &c->btree_interior_update_list)
2136                 ret++;
2137         mutex_unlock(&c->btree_interior_update_lock);
2138
2139         return ret;
2140 }
2141
2142 void bch2_journal_entries_to_btree_roots(struct bch_fs *c, struct jset *jset)
2143 {
2144         struct btree_root *r;
2145         struct jset_entry *entry;
2146
2147         mutex_lock(&c->btree_root_lock);
2148
2149         vstruct_for_each(jset, entry)
2150                 if (entry->type == BCH_JSET_ENTRY_btree_root) {
2151                         r = &c->btree_roots[entry->btree_id];
2152                         r->level = entry->level;
2153                         r->alive = true;
2154                         bkey_copy(&r->key, &entry->start[0]);
2155                 }
2156
2157         mutex_unlock(&c->btree_root_lock);
2158 }
2159
2160 struct jset_entry *
2161 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2162                                     struct jset_entry *start,
2163                                     struct jset_entry *end)
2164 {
2165         struct jset_entry *entry;
2166         unsigned long have = 0;
2167         unsigned i;
2168
2169         for (entry = start; entry < end; entry = vstruct_next(entry))
2170                 if (entry->type == BCH_JSET_ENTRY_btree_root)
2171                         __set_bit(entry->btree_id, &have);
2172
2173         mutex_lock(&c->btree_root_lock);
2174
2175         for (i = 0; i < BTREE_ID_NR; i++)
2176                 if (c->btree_roots[i].alive && !test_bit(i, &have)) {
2177                         journal_entry_set(end,
2178                                           BCH_JSET_ENTRY_btree_root,
2179                                           i, c->btree_roots[i].level,
2180                                           &c->btree_roots[i].key,
2181                                           c->btree_roots[i].key.u64s);
2182                         end = vstruct_next(end);
2183                 }
2184
2185         mutex_unlock(&c->btree_root_lock);
2186
2187         return end;
2188 }
2189
2190 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2191 {
2192         if (c->btree_interior_update_worker)
2193                 destroy_workqueue(c->btree_interior_update_worker);
2194         mempool_exit(&c->btree_interior_update_pool);
2195 }
2196
2197 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2198 {
2199         mutex_init(&c->btree_reserve_cache_lock);
2200         INIT_LIST_HEAD(&c->btree_interior_update_list);
2201         INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2202         mutex_init(&c->btree_interior_update_lock);
2203         INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2204
2205         c->btree_interior_update_worker =
2206                 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2207         if (!c->btree_interior_update_worker)
2208                 return -ENOMEM;
2209
2210         return mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2211                                          sizeof(struct btree_update));
2212 }