git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/btree_update_interior.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "alloc_foreground.h"
   5 #include "bkey_methods.h"
   6 #include "btree_cache.h"
   7 #include "btree_gc.h"
   8 #include "btree_update.h"
   9 #include "btree_update_interior.h"
  10 #include "btree_io.h"
  11 #include "btree_iter.h"
  12 #include "btree_locking.h"
  13 #include "buckets.h"
  14 #include "error.h"
  15 #include "extents.h"
  16 #include "journal.h"
  17 #include "journal_reclaim.h"
  18 #include "keylist.h"
  19 #include "recovery.h"
  20 #include "replicas.h"
  21 #include "super-io.h"
  22
  23 #include <linux/random.h>
  24 #include <trace/events/bcachefs.h>
  25
  26 static void bch2_btree_insert_node(struct btree_update *, struct btree_trans *,
  27                                    struct btree_path *, struct btree *,
  28                                    struct keylist *, unsigned);
  29 static void bch2_btree_update_add_new_node(struct btree_update *, struct btree *);
  30
  31 /* Debug code: */
  32
  33 /*
  34  * Verify that child nodes correctly span parent node's range:
  35  */
  36 static void btree_node_interior_verify(struct bch_fs *c, struct btree *b)
  37 {
  38 #ifdef CONFIG_BCACHEFS_DEBUG
  39         struct bpos next_node = b->data->min_key;
  40         struct btree_node_iter iter;
  41         struct bkey_s_c k;
  42         struct bkey_s_c_btree_ptr_v2 bp;
  43         struct bkey unpacked;
  44         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
  45
  46         BUG_ON(!b->c.level);
  47
  48         if (!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))
  49                 return;
  50
  51         bch2_btree_node_iter_init_from_start(&iter, b);
  52
  53         while (1) {
  54                 k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked);
  55                 if (k.k->type != KEY_TYPE_btree_ptr_v2)
  56                         break;
  57                 bp = bkey_s_c_to_btree_ptr_v2(k);
  58
  59                 if (bpos_cmp(next_node, bp.v->min_key)) {
  60                         bch2_dump_btree_node(c, b);
  61                         bch2_bpos_to_text(&buf1, next_node);
  62                         bch2_bpos_to_text(&buf2, bp.v->min_key);
  63                         panic("expected next min_key %s got %s\n", buf1.buf, buf2.buf);
  64                 }
  65
  66                 bch2_btree_node_iter_advance(&iter, b);
  67
  68                 if (bch2_btree_node_iter_end(&iter)) {
  69                         if (bpos_cmp(k.k->p, b->key.k.p)) {
  70                                 bch2_dump_btree_node(c, b);
  71                                 bch2_bpos_to_text(&buf1, b->key.k.p);
  72                                 bch2_bpos_to_text(&buf2, k.k->p);
  73                                 panic("expected end %s got %s\n", buf1.buf, buf2.buf);
  74                         }
  75                         break;
  76                 }
  77
  78                 next_node = bpos_successor(k.k->p);
  79         }
  80 #endif
  81 }
  82
  83 /* Calculate ideal packed bkey format for new btree nodes: */
  84
  85 void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b)
  86 {
  87         struct bkey_packed *k;
  88         struct bset_tree *t;
  89         struct bkey uk;
  90
  91         for_each_bset(b, t)
  92                 bset_tree_for_each_key(b, t, k)
  93                         if (!bkey_deleted(k)) {
  94                                 uk = bkey_unpack_key(b, k);
  95                                 bch2_bkey_format_add_key(s, &uk);
  96                         }
  97 }
  98
  99 static struct bkey_format bch2_btree_calc_format(struct btree *b)
 100 {
 101         struct bkey_format_state s;
 102
 103         bch2_bkey_format_init(&s);
 104         bch2_bkey_format_add_pos(&s, b->data->min_key);
 105         bch2_bkey_format_add_pos(&s, b->data->max_key);
 106         __bch2_btree_calc_format(&s, b);
 107
 108         return bch2_bkey_format_done(&s);
 109 }
 110
 111 static size_t btree_node_u64s_with_format(struct btree *b,
 112                                           struct bkey_format *new_f)
 113 {
 114         struct bkey_format *old_f = &b->format;
 115
 116         /* stupid integer promotion rules */
 117         ssize_t delta =
 118             (((int) new_f->key_u64s - old_f->key_u64s) *
 119              (int) b->nr.packed_keys) +
 120             (((int) new_f->key_u64s - BKEY_U64s) *
 121              (int) b->nr.unpacked_keys);
 122
 123         BUG_ON(delta + b->nr.live_u64s < 0);
 124
 125         return b->nr.live_u64s + delta;
 126 }
 127
 128 /**
 129  * btree_node_format_fits - check if we could rewrite node with a new format
 130  *
 131  * This assumes all keys can pack with the new format -- it just checks if
 132  * the re-packed keys would fit inside the node itself.
 133  */
 134 bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
 135                                  struct bkey_format *new_f)
 136 {
 137         size_t u64s = btree_node_u64s_with_format(b, new_f);
 138
 139         return __vstruct_bytes(struct btree_node, u64s) < btree_bytes(c);
 140 }
 141
 142 /* Btree node freeing/allocation: */
 143
 144 static void __btree_node_free(struct bch_fs *c, struct btree *b)
 145 {
 146         trace_btree_node_free(c, b);
 147
 148         BUG_ON(btree_node_dirty(b));
 149         BUG_ON(btree_node_need_write(b));
 150         BUG_ON(b == btree_node_root(c, b));
 151         BUG_ON(b->ob.nr);
 152         BUG_ON(!list_empty(&b->write_blocked));
 153         BUG_ON(b->will_make_reachable);
 154
 155         clear_btree_node_noevict(b);
 156
 157         mutex_lock(&c->btree_cache.lock);
 158         list_move(&b->list, &c->btree_cache.freeable);
 159         mutex_unlock(&c->btree_cache.lock);
 160 }
 161
 162 static void bch2_btree_node_free_inmem(struct btree_trans *trans,
 163                                        struct btree *b)
 164 {
 165         struct bch_fs *c = trans->c;
 166         struct btree_path *path;
 167
 168         trans_for_each_path(trans, path)
 169                 BUG_ON(path->l[b->c.level].b == b &&
 170                        path->l[b->c.level].lock_seq == b->c.lock.state.seq);
 171
 172         six_lock_write(&b->c.lock, NULL, NULL);
 173
 174         bch2_btree_node_hash_remove(&c->btree_cache, b);
 175         __btree_node_free(c, b);
 176
 177         six_unlock_write(&b->c.lock);
 178         six_unlock_intent(&b->c.lock);
 179 }
 180
 181 static struct btree *__bch2_btree_node_alloc(struct btree_trans *trans,
 182                                              struct disk_reservation *res,
 183                                              struct closure *cl,
 184                                              bool interior_node,
 185                                              unsigned flags)
 186 {
 187         struct bch_fs *c = trans->c;
 188         struct write_point *wp;
 189         struct btree *b;
 190         __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
 191         struct open_buckets ob = { .nr = 0 };
 192         struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
 193         unsigned nr_reserve;
 194         enum alloc_reserve alloc_reserve;
 195
 196         if (flags & BTREE_INSERT_USE_RESERVE) {
 197                 nr_reserve      = 0;
 198                 alloc_reserve   = RESERVE_btree_movinggc;
 199         } else {
 200                 nr_reserve      = BTREE_NODE_RESERVE;
 201                 alloc_reserve   = RESERVE_btree;
 202         }
 203
 204         mutex_lock(&c->btree_reserve_cache_lock);
 205         if (c->btree_reserve_cache_nr > nr_reserve) {
 206                 struct btree_alloc *a =
 207                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
 208
 209                 ob = a->ob;
 210                 bkey_copy(&tmp.k, &a->k);
 211                 mutex_unlock(&c->btree_reserve_cache_lock);
 212                 goto mem_alloc;
 213         }
 214         mutex_unlock(&c->btree_reserve_cache_lock);
 215
 216 retry:
 217         wp = bch2_alloc_sectors_start_trans(trans,
 218                                       c->opts.metadata_target ?:
 219                                       c->opts.foreground_target,
 220                                       0,
 221                                       writepoint_ptr(&c->btree_write_point),
 222                                       &devs_have,
 223                                       res->nr_replicas,
 224                                       c->opts.metadata_replicas_required,
 225                                       alloc_reserve, 0, cl);
 226         if (IS_ERR(wp))
 227                 return ERR_CAST(wp);
 228
 229         if (wp->sectors_free < btree_sectors(c)) {
 230                 struct open_bucket *ob;
 231                 unsigned i;
 232
 233                 open_bucket_for_each(c, &wp->ptrs, ob, i)
 234                         if (ob->sectors_free < btree_sectors(c))
 235                                 ob->sectors_free = 0;
 236
 237                 bch2_alloc_sectors_done(c, wp);
 238                 goto retry;
 239         }
 240
 241         bkey_btree_ptr_v2_init(&tmp.k);
 242         bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, btree_sectors(c), false);
 243
 244         bch2_open_bucket_get(c, wp, &ob);
 245         bch2_alloc_sectors_done(c, wp);
 246 mem_alloc:
 247         b = bch2_btree_node_mem_alloc(c, interior_node);
 248         six_unlock_write(&b->c.lock);
 249         six_unlock_intent(&b->c.lock);
 250
 251         /* we hold cannibalize_lock: */
 252         BUG_ON(IS_ERR(b));
 253         BUG_ON(b->ob.nr);
 254
 255         bkey_copy(&b->key, &tmp.k);
 256         b->ob = ob;
 257
 258         return b;
 259 }
 260
 261 static struct btree *bch2_btree_node_alloc(struct btree_update *as, unsigned level)
 262 {
 263         struct bch_fs *c = as->c;
 264         struct btree *b;
 265         struct prealloc_nodes *p = &as->prealloc_nodes[!!level];
 266         int ret;
 267
 268         BUG_ON(level >= BTREE_MAX_DEPTH);
 269         BUG_ON(!p->nr);
 270
 271         b = p->b[--p->nr];
 272
 273         six_lock_intent(&b->c.lock, NULL, NULL);
 274         six_lock_write(&b->c.lock, NULL, NULL);
 275
 276         set_btree_node_accessed(b);
 277         set_btree_node_dirty_acct(c, b);
 278         set_btree_node_need_write(b);
 279
 280         bch2_bset_init_first(b, &b->data->keys);
 281         b->c.level      = level;
 282         b->c.btree_id   = as->btree_id;
 283         b->version_ondisk = c->sb.version;
 284
 285         memset(&b->nr, 0, sizeof(b->nr));
 286         b->data->magic = cpu_to_le64(bset_magic(c));
 287         memset(&b->data->_ptr, 0, sizeof(b->data->_ptr));
 288         b->data->flags = 0;
 289         SET_BTREE_NODE_ID(b->data, as->btree_id);
 290         SET_BTREE_NODE_LEVEL(b->data, level);
 291
 292         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 293                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
 294
 295                 bp->v.mem_ptr           = 0;
 296                 bp->v.seq               = b->data->keys.seq;
 297                 bp->v.sectors_written   = 0;
 298         }
 299
 300         SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
 301
 302         bch2_btree_build_aux_trees(b);
 303
 304         ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
 305         BUG_ON(ret);
 306
 307         trace_btree_node_alloc(c, b);
 308         return b;
 309 }
 310
 311 static void btree_set_min(struct btree *b, struct bpos pos)
 312 {
 313         if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
 314                 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
 315         b->data->min_key = pos;
 316 }
 317
 318 static void btree_set_max(struct btree *b, struct bpos pos)
 319 {
 320         b->key.k.p = pos;
 321         b->data->max_key = pos;
 322 }
 323
 324 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *as,
 325                                                   struct btree *b,
 326                                                   struct bkey_format format)
 327 {
 328         struct btree *n;
 329
 330         n = bch2_btree_node_alloc(as, b->c.level);
 331
 332         SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
 333
 334         btree_set_min(n, b->data->min_key);
 335         btree_set_max(n, b->data->max_key);
 336
 337         n->data->format         = format;
 338         btree_node_set_format(n, format);
 339
 340         bch2_btree_sort_into(as->c, n, b);
 341
 342         btree_node_reset_sib_u64s(n);
 343
 344         n->key.k.p = b->key.k.p;
 345         return n;
 346 }
 347
 348 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
 349                                                        struct btree *b)
 350 {
 351         struct bkey_format new_f = bch2_btree_calc_format(b);
 352
 353         /*
 354          * The keys might expand with the new format - if they wouldn't fit in
 355          * the btree node anymore, use the old format for now:
 356          */
 357         if (!bch2_btree_node_format_fits(as->c, b, &new_f))
 358                 new_f = b->format;
 359
 360         return __bch2_btree_node_alloc_replacement(as, b, new_f);
 361 }
 362
 363 static struct btree *__btree_root_alloc(struct btree_update *as, unsigned level)
 364 {
 365         struct btree *b = bch2_btree_node_alloc(as, level);
 366
 367         btree_set_min(b, POS_MIN);
 368         btree_set_max(b, SPOS_MAX);
 369         b->data->format = bch2_btree_calc_format(b);
 370
 371         btree_node_set_format(b, b->data->format);
 372         bch2_btree_build_aux_trees(b);
 373
 374         bch2_btree_update_add_new_node(as, b);
 375         six_unlock_write(&b->c.lock);
 376
 377         return b;
 378 }
 379
 380 static void bch2_btree_reserve_put(struct btree_update *as)
 381 {
 382         struct bch_fs *c = as->c;
 383         struct prealloc_nodes *p;
 384
 385         for (p = as->prealloc_nodes;
 386              p < as->prealloc_nodes + ARRAY_SIZE(as->prealloc_nodes);
 387              p++) {
 388                 while (p->nr) {
 389                         struct btree *b = p->b[--p->nr];
 390
 391                         mutex_lock(&c->btree_reserve_cache_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                         mutex_unlock(&c->btree_reserve_cache_lock);
 406
 407                         six_lock_intent(&b->c.lock, NULL, NULL);
 408                         six_lock_write(&b->c.lock, NULL, NULL);
 409                         __btree_node_free(c, b);
 410                         six_unlock_write(&b->c.lock);
 411                         six_unlock_intent(&b->c.lock);
 412                 }
 413         }
 414 }
 415
 416 static int bch2_btree_reserve_get(struct btree_trans *trans,
 417                                   struct btree_update *as,
 418                                   unsigned nr_nodes[2],
 419                                   unsigned flags,
 420                                   struct closure *cl)
 421 {
 422         struct bch_fs *c = as->c;
 423         struct btree *b;
 424         unsigned interior;
 425         int ret = 0;
 426
 427         BUG_ON(nr_nodes[0] + nr_nodes[1] > BTREE_RESERVE_MAX);
 428
 429         /*
 430          * Protects reaping from the btree node cache and using the btree node
 431          * open bucket reserve:
 432          *
 433          * BTREE_INSERT_NOWAIT only applies to btree node allocation, not
 434          * blocking on this lock:
 435          */
 436         ret = bch2_btree_cache_cannibalize_lock(c, cl);
 437         if (ret)
 438                 return ret;
 439
 440         for (interior = 0; interior < 2; interior++) {
 441                 struct prealloc_nodes *p = as->prealloc_nodes + interior;
 442
 443                 while (p->nr < nr_nodes[interior]) {
 444                         b = __bch2_btree_node_alloc(trans, &as->disk_res,
 445                                         flags & BTREE_INSERT_NOWAIT ? NULL : cl,
 446                                         interior, flags);
 447                         if (IS_ERR(b)) {
 448                                 ret = PTR_ERR(b);
 449                                 goto err;
 450                         }
 451
 452                         p->b[p->nr++] = b;
 453                 }
 454         }
 455 err:
 456         bch2_btree_cache_cannibalize_unlock(c);
 457         return ret;
 458 }
 459
 460 /* Asynchronous interior node update machinery */
 461
 462 static void bch2_btree_update_free(struct btree_update *as)
 463 {
 464         struct bch_fs *c = as->c;
 465
 466         if (as->took_gc_lock)
 467                 up_read(&c->gc_lock);
 468         as->took_gc_lock = false;
 469
 470         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 471
 472         bch2_journal_pin_drop(&c->journal, &as->journal);
 473         bch2_journal_pin_flush(&c->journal, &as->journal);
 474         bch2_disk_reservation_put(c, &as->disk_res);
 475         bch2_btree_reserve_put(as);
 476
 477         bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_total],
 478                                as->start_time);
 479
 480         mutex_lock(&c->btree_interior_update_lock);
 481         list_del(&as->unwritten_list);
 482         list_del(&as->list);
 483
 484         closure_debug_destroy(&as->cl);
 485         mempool_free(as, &c->btree_interior_update_pool);
 486
 487         /*
 488          * Have to do the wakeup with btree_interior_update_lock still held,
 489          * since being on btree_interior_update_list is our ref on @c:
 490          */
 491         closure_wake_up(&c->btree_interior_update_wait);
 492
 493         mutex_unlock(&c->btree_interior_update_lock);
 494 }
 495
 496 static void btree_update_add_key(struct btree_update *as,
 497                                  struct keylist *keys, struct btree *b)
 498 {
 499         struct bkey_i *k = &b->key;
 500
 501         BUG_ON(bch2_keylist_u64s(keys) + k->k.u64s >
 502                ARRAY_SIZE(as->_old_keys));
 503
 504         bkey_copy(keys->top, k);
 505         bkey_i_to_btree_ptr_v2(keys->top)->v.mem_ptr = b->c.level + 1;
 506
 507         bch2_keylist_push(keys);
 508 }
 509
 510 /*
 511  * The transactional part of an interior btree node update, where we journal the
 512  * update we did to the interior node and update alloc info:
 513  */
 514 static int btree_update_nodes_written_trans(struct btree_trans *trans,
 515                                             struct btree_update *as)
 516 {
 517         struct bkey_i *k;
 518         int ret;
 519
 520         ret = darray_make_room(&trans->extra_journal_entries, as->journal_u64s);
 521         if (ret)
 522                 return ret;
 523
 524         memcpy(&darray_top(trans->extra_journal_entries),
 525                as->journal_entries,
 526                as->journal_u64s * sizeof(u64));
 527         trans->extra_journal_entries.nr += as->journal_u64s;
 528
 529         trans->journal_pin = &as->journal;
 530
 531         for_each_keylist_key(&as->old_keys, k) {
 532                 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
 533
 534                 ret = bch2_trans_mark_old(trans, as->btree_id, level, bkey_i_to_s_c(k), 0);
 535                 if (ret)
 536                         return ret;
 537         }
 538
 539         for_each_keylist_key(&as->new_keys, k) {
 540                 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
 541
 542                 ret = bch2_trans_mark_new(trans, as->btree_id, level, k, 0);
 543                 if (ret)
 544                         return ret;
 545         }
 546
 547         return 0;
 548 }
 549
 550 static void btree_update_nodes_written(struct btree_update *as)
 551 {
 552         struct bch_fs *c = as->c;
 553         struct btree *b = as->b;
 554         struct btree_trans trans;
 555         u64 journal_seq = 0;
 556         unsigned i;
 557         int ret;
 558
 559         /*
 560          * If we're already in an error state, it might be because a btree node
 561          * was never written, and we might be trying to free that same btree
 562          * node here, but it won't have been marked as allocated and we'll see
 563          * spurious disk usage inconsistencies in the transactional part below
 564          * if we don't skip it:
 565          */
 566         ret = bch2_journal_error(&c->journal);
 567         if (ret)
 568                 goto err;
 569
 570         /*
 571          * Wait for any in flight writes to finish before we free the old nodes
 572          * on disk:
 573          */
 574         for (i = 0; i < as->nr_old_nodes; i++) {
 575                 struct btree *old = as->old_nodes[i];
 576                 __le64 seq;
 577
 578                 six_lock_read(&old->c.lock, NULL, NULL);
 579                 seq = old->data ? old->data->keys.seq : 0;
 580                 six_unlock_read(&old->c.lock);
 581
 582                 if (seq == as->old_nodes_seq[i])
 583                         wait_on_bit_io(&old->flags, BTREE_NODE_write_in_flight_inner,
 584                                        TASK_UNINTERRUPTIBLE);
 585         }
 586
 587         /*
 588          * We did an update to a parent node where the pointers we added pointed
 589          * to child nodes that weren't written yet: now, the child nodes have
 590          * been written so we can write out the update to the interior node.
 591          */
 592
 593         /*
 594          * We can't call into journal reclaim here: we'd block on the journal
 595          * reclaim lock, but we may need to release the open buckets we have
 596          * pinned in order for other btree updates to make forward progress, and
 597          * journal reclaim does btree updates when flushing bkey_cached entries,
 598          * which may require allocations as well.
 599          */
 600         bch2_trans_init(&trans, c, 0, 512);
 601         ret = commit_do(&trans, &as->disk_res, &journal_seq,
 602                               BTREE_INSERT_NOFAIL|
 603                               BTREE_INSERT_NOCHECK_RW|
 604                               BTREE_INSERT_JOURNAL_RECLAIM|
 605                               JOURNAL_WATERMARK_reserved,
 606                               btree_update_nodes_written_trans(&trans, as));
 607         bch2_trans_exit(&trans);
 608
 609         bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
 610                              "error %i in btree_update_nodes_written()", ret);
 611 err:
 612         if (b) {
 613                 /*
 614                  * @b is the node we did the final insert into:
 615                  *
 616                  * On failure to get a journal reservation, we still have to
 617                  * unblock the write and allow most of the write path to happen
 618                  * so that shutdown works, but the i->journal_seq mechanism
 619                  * won't work to prevent the btree write from being visible (we
 620                  * didn't get a journal sequence number) - instead
 621                  * __bch2_btree_node_write() doesn't do the actual write if
 622                  * we're in journal error state:
 623                  */
 624
 625                 six_lock_intent(&b->c.lock, NULL, NULL);
 626                 six_lock_write(&b->c.lock, NULL, NULL);
 627                 mutex_lock(&c->btree_interior_update_lock);
 628
 629                 list_del(&as->write_blocked_list);
 630                 if (list_empty(&b->write_blocked))
 631                         clear_btree_node_write_blocked(b);
 632
 633                 /*
 634                  * Node might have been freed, recheck under
 635                  * btree_interior_update_lock:
 636                  */
 637                 if (as->b == b) {
 638                         struct bset *i = btree_bset_last(b);
 639
 640                         BUG_ON(!b->c.level);
 641                         BUG_ON(!btree_node_dirty(b));
 642
 643                         if (!ret) {
 644                                 i->journal_seq = cpu_to_le64(
 645                                                              max(journal_seq,
 646                                                                  le64_to_cpu(i->journal_seq)));
 647
 648                                 bch2_btree_add_journal_pin(c, b, journal_seq);
 649                         } else {
 650                                 /*
 651                                  * If we didn't get a journal sequence number we
 652                                  * can't write this btree node, because recovery
 653                                  * won't know to ignore this write:
 654                                  */
 655                                 set_btree_node_never_write(b);
 656                         }
 657                 }
 658
 659                 mutex_unlock(&c->btree_interior_update_lock);
 660                 six_unlock_write(&b->c.lock);
 661
 662                 btree_node_write_if_need(c, b, SIX_LOCK_intent);
 663                 six_unlock_intent(&b->c.lock);
 664         }
 665
 666         bch2_journal_pin_drop(&c->journal, &as->journal);
 667
 668         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 669
 670         mutex_lock(&c->btree_interior_update_lock);
 671         for (i = 0; i < as->nr_new_nodes; i++) {
 672                 b = as->new_nodes[i];
 673
 674                 BUG_ON(b->will_make_reachable != (unsigned long) as);
 675                 b->will_make_reachable = 0;
 676                 clear_btree_node_will_make_reachable(b);
 677         }
 678         mutex_unlock(&c->btree_interior_update_lock);
 679
 680         for (i = 0; i < as->nr_new_nodes; i++) {
 681                 b = as->new_nodes[i];
 682
 683                 six_lock_read(&b->c.lock, NULL, NULL);
 684                 btree_node_write_if_need(c, b, SIX_LOCK_read);
 685                 six_unlock_read(&b->c.lock);
 686         }
 687
 688         for (i = 0; i < as->nr_open_buckets; i++)
 689                 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
 690
 691         bch2_btree_update_free(as);
 692 }
 693
 694 static void btree_interior_update_work(struct work_struct *work)
 695 {
 696         struct bch_fs *c =
 697                 container_of(work, struct bch_fs, btree_interior_update_work);
 698         struct btree_update *as;
 699
 700         while (1) {
 701                 mutex_lock(&c->btree_interior_update_lock);
 702                 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
 703                                               struct btree_update, unwritten_list);
 704                 if (as && !as->nodes_written)
 705                         as = NULL;
 706                 mutex_unlock(&c->btree_interior_update_lock);
 707
 708                 if (!as)
 709                         break;
 710
 711                 btree_update_nodes_written(as);
 712         }
 713 }
 714
 715 static void btree_update_set_nodes_written(struct closure *cl)
 716 {
 717         struct btree_update *as = container_of(cl, struct btree_update, cl);
 718         struct bch_fs *c = as->c;
 719
 720         mutex_lock(&c->btree_interior_update_lock);
 721         as->nodes_written = true;
 722         mutex_unlock(&c->btree_interior_update_lock);
 723
 724         queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
 725 }
 726
 727 /*
 728  * We're updating @b with pointers to nodes that haven't finished writing yet:
 729  * block @b from being written until @as completes
 730  */
 731 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
 732 {
 733         struct bch_fs *c = as->c;
 734
 735         mutex_lock(&c->btree_interior_update_lock);
 736         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 737
 738         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 739         BUG_ON(!btree_node_dirty(b));
 740
 741         as->mode        = BTREE_INTERIOR_UPDATING_NODE;
 742         as->b           = b;
 743
 744         set_btree_node_write_blocked(b);
 745         list_add(&as->write_blocked_list, &b->write_blocked);
 746
 747         mutex_unlock(&c->btree_interior_update_lock);
 748 }
 749
 750 static void btree_update_reparent(struct btree_update *as,
 751                                   struct btree_update *child)
 752 {
 753         struct bch_fs *c = as->c;
 754
 755         lockdep_assert_held(&c->btree_interior_update_lock);
 756
 757         child->b = NULL;
 758         child->mode = BTREE_INTERIOR_UPDATING_AS;
 759
 760         bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL);
 761 }
 762
 763 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
 764 {
 765         struct bkey_i *insert = &b->key;
 766         struct bch_fs *c = as->c;
 767
 768         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 769
 770         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
 771                ARRAY_SIZE(as->journal_entries));
 772
 773         as->journal_u64s +=
 774                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
 775                                   BCH_JSET_ENTRY_btree_root,
 776                                   b->c.btree_id, b->c.level,
 777                                   insert, insert->k.u64s);
 778
 779         mutex_lock(&c->btree_interior_update_lock);
 780         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 781
 782         as->mode        = BTREE_INTERIOR_UPDATING_ROOT;
 783         mutex_unlock(&c->btree_interior_update_lock);
 784 }
 785
 786 /*
 787  * bch2_btree_update_add_new_node:
 788  *
 789  * This causes @as to wait on @b to be written, before it gets to
 790  * bch2_btree_update_nodes_written
 791  *
 792  * Additionally, it sets b->will_make_reachable to prevent any additional writes
 793  * to @b from happening besides the first until @b is reachable on disk
 794  *
 795  * And it adds @b to the list of @as's new nodes, so that we can update sector
 796  * counts in bch2_btree_update_nodes_written:
 797  */
 798 static void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
 799 {
 800         struct bch_fs *c = as->c;
 801
 802         closure_get(&as->cl);
 803
 804         mutex_lock(&c->btree_interior_update_lock);
 805         BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
 806         BUG_ON(b->will_make_reachable);
 807
 808         as->new_nodes[as->nr_new_nodes++] = b;
 809         b->will_make_reachable = 1UL|(unsigned long) as;
 810         set_btree_node_will_make_reachable(b);
 811
 812         mutex_unlock(&c->btree_interior_update_lock);
 813
 814         btree_update_add_key(as, &as->new_keys, b);
 815 }
 816
 817 /*
 818  * returns true if @b was a new node
 819  */
 820 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
 821 {
 822         struct btree_update *as;
 823         unsigned long v;
 824         unsigned i;
 825
 826         mutex_lock(&c->btree_interior_update_lock);
 827         /*
 828          * When b->will_make_reachable != 0, it owns a ref on as->cl that's
 829          * dropped when it gets written by bch2_btree_complete_write - the
 830          * xchg() is for synchronization with bch2_btree_complete_write:
 831          */
 832         v = xchg(&b->will_make_reachable, 0);
 833         clear_btree_node_will_make_reachable(b);
 834         as = (struct btree_update *) (v & ~1UL);
 835
 836         if (!as) {
 837                 mutex_unlock(&c->btree_interior_update_lock);
 838                 return;
 839         }
 840
 841         for (i = 0; i < as->nr_new_nodes; i++)
 842                 if (as->new_nodes[i] == b)
 843                         goto found;
 844
 845         BUG();
 846 found:
 847         array_remove_item(as->new_nodes, as->nr_new_nodes, i);
 848         mutex_unlock(&c->btree_interior_update_lock);
 849
 850         if (v & 1)
 851                 closure_put(&as->cl);
 852 }
 853
 854 static void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
 855 {
 856         while (b->ob.nr)
 857                 as->open_buckets[as->nr_open_buckets++] =
 858                         b->ob.v[--b->ob.nr];
 859 }
 860
 861 /*
 862  * @b is being split/rewritten: it may have pointers to not-yet-written btree
 863  * nodes and thus outstanding btree_updates - redirect @b's
 864  * btree_updates to point to this btree_update:
 865  */
 866 static void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 867                                                       struct btree *b)
 868 {
 869         struct bch_fs *c = as->c;
 870         struct btree_update *p, *n;
 871         struct btree_write *w;
 872
 873         set_btree_node_dying(b);
 874
 875         if (btree_node_fake(b))
 876                 return;
 877
 878         mutex_lock(&c->btree_interior_update_lock);
 879
 880         /*
 881          * Does this node have any btree_update operations preventing
 882          * it from being written?
 883          *
 884          * If so, redirect them to point to this btree_update: we can
 885          * write out our new nodes, but we won't make them visible until those
 886          * operations complete
 887          */
 888         list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
 889                 list_del_init(&p->write_blocked_list);
 890                 btree_update_reparent(as, p);
 891
 892                 /*
 893                  * for flush_held_btree_writes() waiting on updates to flush or
 894                  * nodes to be writeable:
 895                  */
 896                 closure_wake_up(&c->btree_interior_update_wait);
 897         }
 898
 899         clear_btree_node_dirty_acct(c, b);
 900         clear_btree_node_need_write(b);
 901
 902         /*
 903          * Does this node have unwritten data that has a pin on the journal?
 904          *
 905          * If so, transfer that pin to the btree_update operation -
 906          * note that if we're freeing multiple nodes, we only need to keep the
 907          * oldest pin of any of the nodes we're freeing. We'll release the pin
 908          * when the new nodes are persistent and reachable on disk:
 909          */
 910         w = btree_current_write(b);
 911         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 912         bch2_journal_pin_drop(&c->journal, &w->journal);
 913
 914         w = btree_prev_write(b);
 915         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 916         bch2_journal_pin_drop(&c->journal, &w->journal);
 917
 918         mutex_unlock(&c->btree_interior_update_lock);
 919
 920         /*
 921          * Is this a node that isn't reachable on disk yet?
 922          *
 923          * Nodes that aren't reachable yet have writes blocked until they're
 924          * reachable - now that we've cancelled any pending writes and moved
 925          * things waiting on that write to wait on this update, we can drop this
 926          * node from the list of nodes that the other update is making
 927          * reachable, prior to freeing it:
 928          */
 929         btree_update_drop_new_node(c, b);
 930
 931         btree_update_add_key(as, &as->old_keys, b);
 932
 933         as->old_nodes[as->nr_old_nodes] = b;
 934         as->old_nodes_seq[as->nr_old_nodes] = b->data->keys.seq;
 935         as->nr_old_nodes++;
 936 }
 937
 938 static void bch2_btree_update_done(struct btree_update *as)
 939 {
 940         struct bch_fs *c = as->c;
 941         u64 start_time = as->start_time;
 942
 943         BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
 944
 945         if (as->took_gc_lock)
 946                 up_read(&as->c->gc_lock);
 947         as->took_gc_lock = false;
 948
 949         bch2_btree_reserve_put(as);
 950
 951         continue_at(&as->cl, btree_update_set_nodes_written,
 952                     as->c->btree_interior_update_worker);
 953
 954         bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_foreground],
 955                                start_time);
 956 }
 957
 958 static struct btree_update *
 959 bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
 960                         unsigned level, bool split, unsigned flags)
 961 {
 962         struct bch_fs *c = trans->c;
 963         struct btree_update *as;
 964         u64 start_time = local_clock();
 965         int disk_res_flags = (flags & BTREE_INSERT_NOFAIL)
 966                 ? BCH_DISK_RESERVATION_NOFAIL : 0;
 967         unsigned nr_nodes[2] = { 0, 0 };
 968         unsigned update_level = level;
 969         int journal_flags = flags & JOURNAL_WATERMARK_MASK;
 970         int ret = 0;
 971         u32 restart_count = trans->restart_count;
 972
 973         BUG_ON(!path->should_be_locked);
 974
 975         if (flags & BTREE_INSERT_JOURNAL_RECLAIM)
 976                 journal_flags |= JOURNAL_RES_GET_NONBLOCK;
 977
 978         while (1) {
 979                 nr_nodes[!!update_level] += 1 + split;
 980                 update_level++;
 981
 982                 if (!btree_path_node(path, update_level))
 983                         break;
 984
 985                 /*
 986                  * XXX: figure out how far we might need to split,
 987                  * instead of locking/reserving all the way to the root:
 988                  */
 989                 split = update_level + 1 < BTREE_MAX_DEPTH;
 990         }
 991
 992         /* Might have to allocate a new root: */
 993         if (update_level < BTREE_MAX_DEPTH)
 994                 nr_nodes[1] += 1;
 995
 996         if (!bch2_btree_path_upgrade(trans, path, U8_MAX)) {
 997                 trace_trans_restart_iter_upgrade(trans, _RET_IP_, path);
 998                 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_upgrade);
 999                 return ERR_PTR(ret);
1000         }
1001
1002         if (flags & BTREE_INSERT_GC_LOCK_HELD)
1003                 lockdep_assert_held(&c->gc_lock);
1004         else if (!down_read_trylock(&c->gc_lock)) {
1005                 bch2_trans_unlock(trans);
1006                 down_read(&c->gc_lock);
1007                 ret = bch2_trans_relock(trans);
1008                 if (ret) {
1009                         up_read(&c->gc_lock);
1010                         return ERR_PTR(ret);
1011                 }
1012         }
1013
1014         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
1015         memset(as, 0, sizeof(*as));
1016         closure_init(&as->cl, NULL);
1017         as->c           = c;
1018         as->start_time  = start_time;
1019         as->mode        = BTREE_INTERIOR_NO_UPDATE;
1020         as->took_gc_lock = !(flags & BTREE_INSERT_GC_LOCK_HELD);
1021         as->btree_id    = path->btree_id;
1022         INIT_LIST_HEAD(&as->list);
1023         INIT_LIST_HEAD(&as->unwritten_list);
1024         INIT_LIST_HEAD(&as->write_blocked_list);
1025         bch2_keylist_init(&as->old_keys, as->_old_keys);
1026         bch2_keylist_init(&as->new_keys, as->_new_keys);
1027         bch2_keylist_init(&as->parent_keys, as->inline_keys);
1028
1029         mutex_lock(&c->btree_interior_update_lock);
1030         list_add_tail(&as->list, &c->btree_interior_update_list);
1031         mutex_unlock(&c->btree_interior_update_lock);
1032
1033         /*
1034          * We don't want to allocate if we're in an error state, that can cause
1035          * deadlock on emergency shutdown due to open buckets getting stuck in
1036          * the btree_reserve_cache after allocator shutdown has cleared it out.
1037          * This check needs to come after adding us to the btree_interior_update
1038          * list but before calling bch2_btree_reserve_get, to synchronize with
1039          * __bch2_fs_read_only().
1040          */
1041         ret = bch2_journal_error(&c->journal);
1042         if (ret)
1043                 goto err;
1044
1045         ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
1046                                       BTREE_UPDATE_JOURNAL_RES,
1047                                       journal_flags|JOURNAL_RES_GET_NONBLOCK);
1048         if (ret) {
1049                 bch2_trans_unlock(trans);
1050
1051                 ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
1052                                               BTREE_UPDATE_JOURNAL_RES,
1053                                               journal_flags);
1054                 if (ret) {
1055                         trace_trans_restart_journal_preres_get(trans, _RET_IP_);
1056                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_journal_preres_get);
1057                         goto err;
1058                 }
1059
1060                 ret = bch2_trans_relock(trans);
1061                 if (ret)
1062                         goto err;
1063         }
1064
1065         ret = bch2_disk_reservation_get(c, &as->disk_res,
1066                         (nr_nodes[0] + nr_nodes[1]) * btree_sectors(c),
1067                         c->opts.metadata_replicas,
1068                         disk_res_flags);
1069         if (ret)
1070                 goto err;
1071
1072         ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, NULL);
1073         if (ret == -EAGAIN ||
1074             ret == -ENOMEM) {
1075                 struct closure cl;
1076
1077                 closure_init_stack(&cl);
1078
1079                 bch2_trans_unlock(trans);
1080
1081                 do {
1082                         ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, &cl);
1083                         closure_sync(&cl);
1084                 } while (ret == -EAGAIN);
1085         }
1086
1087         if (ret) {
1088                 trace_btree_reserve_get_fail(trans->fn, _RET_IP_,
1089                                              nr_nodes[0] + nr_nodes[1]);
1090                 goto err;
1091         }
1092
1093         ret = bch2_trans_relock(trans);
1094         if (ret)
1095                 goto err;
1096
1097         bch2_trans_verify_not_restarted(trans, restart_count);
1098         return as;
1099 err:
1100         bch2_btree_update_free(as);
1101         return ERR_PTR(ret);
1102 }
1103
1104 /* Btree root updates: */
1105
1106 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1107 {
1108         /* Root nodes cannot be reaped */
1109         mutex_lock(&c->btree_cache.lock);
1110         list_del_init(&b->list);
1111         mutex_unlock(&c->btree_cache.lock);
1112
1113         mutex_lock(&c->btree_root_lock);
1114         BUG_ON(btree_node_root(c, b) &&
1115                (b->c.level < btree_node_root(c, b)->c.level ||
1116                 !btree_node_dying(btree_node_root(c, b))));
1117
1118         btree_node_root(c, b) = b;
1119         mutex_unlock(&c->btree_root_lock);
1120
1121         bch2_recalc_btree_reserve(c);
1122 }
1123
1124 /**
1125  * bch_btree_set_root - update the root in memory and on disk
1126  *
1127  * To ensure forward progress, the current task must not be holding any
1128  * btree node write locks. However, you must hold an intent lock on the
1129  * old root.
1130  *
1131  * Note: This allocates a journal entry but doesn't add any keys to
1132  * it.  All the btree roots are part of every journal write, so there
1133  * is nothing new to be done.  This just guarantees that there is a
1134  * journal write.
1135  */
1136 static void bch2_btree_set_root(struct btree_update *as,
1137                                 struct btree_trans *trans,
1138                                 struct btree_path *path,
1139                                 struct btree *b)
1140 {
1141         struct bch_fs *c = as->c;
1142         struct btree *old;
1143
1144         trace_btree_set_root(c, b);
1145         BUG_ON(!b->written);
1146
1147         old = btree_node_root(c, b);
1148
1149         /*
1150          * Ensure no one is using the old root while we switch to the
1151          * new root:
1152          */
1153         bch2_btree_node_lock_write(trans, path, old);
1154
1155         bch2_btree_set_root_inmem(c, b);
1156
1157         btree_update_updated_root(as, b);
1158
1159         /*
1160          * Unlock old root after new root is visible:
1161          *
1162          * The new root isn't persistent, but that's ok: we still have
1163          * an intent lock on the new root, and any updates that would
1164          * depend on the new root would have to update the new root.
1165          */
1166         bch2_btree_node_unlock_write(trans, path, old);
1167 }
1168
1169 /* Interior node updates: */
1170
1171 static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
1172                                         struct btree_trans *trans,
1173                                         struct btree_path *path,
1174                                         struct btree *b,
1175                                         struct btree_node_iter *node_iter,
1176                                         struct bkey_i *insert)
1177 {
1178         struct bch_fs *c = as->c;
1179         struct bkey_packed *k;
1180         struct printbuf buf = PRINTBUF;
1181
1182         BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 &&
1183                !btree_ptr_sectors_written(insert));
1184
1185         if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
1186                 bch2_journal_key_overwritten(c, b->c.btree_id, b->c.level, insert->k.p);
1187
1188         if (bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1189                               btree_node_type(b), WRITE, &buf) ?:
1190             bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert), &buf)) {
1191                 printbuf_reset(&buf);
1192                 prt_printf(&buf, "inserting invalid bkey\n  ");
1193                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
1194                 prt_printf(&buf, "\n  ");
1195                 bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1196                                   btree_node_type(b), WRITE, &buf);
1197                 bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert), &buf);
1198
1199                 bch2_fs_inconsistent(c, "%s", buf.buf);
1200                 dump_stack();
1201         }
1202
1203         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1204                ARRAY_SIZE(as->journal_entries));
1205
1206         as->journal_u64s +=
1207                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1208                                   BCH_JSET_ENTRY_btree_keys,
1209                                   b->c.btree_id, b->c.level,
1210                                   insert, insert->k.u64s);
1211
1212         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1213                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1214                 bch2_btree_node_iter_advance(node_iter, b);
1215
1216         bch2_btree_bset_insert_key(trans, path, b, node_iter, insert);
1217         set_btree_node_dirty_acct(c, b);
1218         set_btree_node_need_write(b);
1219
1220         printbuf_exit(&buf);
1221 }
1222
1223 static void
1224 __bch2_btree_insert_keys_interior(struct btree_update *as,
1225                                   struct btree_trans *trans,
1226                                   struct btree_path *path,
1227                                   struct btree *b,
1228                                   struct btree_node_iter node_iter,
1229                                   struct keylist *keys)
1230 {
1231         struct bkey_i *insert = bch2_keylist_front(keys);
1232         struct bkey_packed *k;
1233
1234         BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1235
1236         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1237                (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1238                 ;
1239
1240         while (!bch2_keylist_empty(keys)) {
1241                 bch2_insert_fixup_btree_ptr(as, trans, path, b,
1242                                 &node_iter, bch2_keylist_front(keys));
1243                 bch2_keylist_pop_front(keys);
1244         }
1245 }
1246
1247 /*
1248  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1249  * node)
1250  */
1251 static struct btree *__btree_split_node(struct btree_update *as,
1252                                         struct btree *n1)
1253 {
1254         struct bkey_format_state s;
1255         size_t nr_packed = 0, nr_unpacked = 0;
1256         struct btree *n2;
1257         struct bset *set1, *set2;
1258         struct bkey_packed *k, *set2_start, *set2_end, *out, *prev = NULL;
1259         struct bpos n1_pos;
1260
1261         n2 = bch2_btree_node_alloc(as, n1->c.level);
1262
1263         n2->data->max_key       = n1->data->max_key;
1264         n2->data->format        = n1->format;
1265         SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data));
1266         n2->key.k.p = n1->key.k.p;
1267
1268         bch2_btree_update_add_new_node(as, n2);
1269
1270         set1 = btree_bset_first(n1);
1271         set2 = btree_bset_first(n2);
1272
1273         /*
1274          * Has to be a linear search because we don't have an auxiliary
1275          * search tree yet
1276          */
1277         k = set1->start;
1278         while (1) {
1279                 struct bkey_packed *n = bkey_next(k);
1280
1281                 if (n == vstruct_last(set1))
1282                         break;
1283                 if (k->_data - set1->_data >= (le16_to_cpu(set1->u64s) * 3) / 5)
1284                         break;
1285
1286                 if (bkey_packed(k))
1287                         nr_packed++;
1288                 else
1289                         nr_unpacked++;
1290
1291                 prev = k;
1292                 k = n;
1293         }
1294
1295         BUG_ON(!prev);
1296         set2_start      = k;
1297         set2_end        = vstruct_last(set1);
1298
1299         set1->u64s = cpu_to_le16((u64 *) set2_start - set1->_data);
1300         set_btree_bset_end(n1, n1->set);
1301
1302         n1->nr.live_u64s        = le16_to_cpu(set1->u64s);
1303         n1->nr.bset_u64s[0]     = le16_to_cpu(set1->u64s);
1304         n1->nr.packed_keys      = nr_packed;
1305         n1->nr.unpacked_keys    = nr_unpacked;
1306
1307         n1_pos = bkey_unpack_pos(n1, prev);
1308         if (as->c->sb.version < bcachefs_metadata_version_snapshot)
1309                 n1_pos.snapshot = U32_MAX;
1310
1311         btree_set_max(n1, n1_pos);
1312         btree_set_min(n2, bpos_successor(n1->key.k.p));
1313
1314         bch2_bkey_format_init(&s);
1315         bch2_bkey_format_add_pos(&s, n2->data->min_key);
1316         bch2_bkey_format_add_pos(&s, n2->data->max_key);
1317
1318         for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1319                 struct bkey uk = bkey_unpack_key(n1, k);
1320                 bch2_bkey_format_add_key(&s, &uk);
1321         }
1322
1323         n2->data->format = bch2_bkey_format_done(&s);
1324         btree_node_set_format(n2, n2->data->format);
1325
1326         out = set2->start;
1327         memset(&n2->nr, 0, sizeof(n2->nr));
1328
1329         for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1330                 BUG_ON(!bch2_bkey_transform(&n2->format, out, bkey_packed(k)
1331                                        ? &n1->format : &bch2_bkey_format_current, k));
1332                 out->format = KEY_FORMAT_LOCAL_BTREE;
1333                 btree_keys_account_key_add(&n2->nr, 0, out);
1334                 out = bkey_next(out);
1335         }
1336
1337         set2->u64s = cpu_to_le16((u64 *) out - set2->_data);
1338         set_btree_bset_end(n2, n2->set);
1339
1340         BUG_ON(!set1->u64s);
1341         BUG_ON(!set2->u64s);
1342
1343         btree_node_reset_sib_u64s(n1);
1344         btree_node_reset_sib_u64s(n2);
1345
1346         bch2_verify_btree_nr_keys(n1);
1347         bch2_verify_btree_nr_keys(n2);
1348
1349         if (n1->c.level) {
1350                 btree_node_interior_verify(as->c, n1);
1351                 btree_node_interior_verify(as->c, n2);
1352         }
1353
1354         return n2;
1355 }
1356
1357 /*
1358  * For updates to interior nodes, we've got to do the insert before we split
1359  * because the stuff we're inserting has to be inserted atomically. Post split,
1360  * the keys might have to go in different nodes and the split would no longer be
1361  * atomic.
1362  *
1363  * Worse, if the insert is from btree node coalescing, if we do the insert after
1364  * we do the split (and pick the pivot) - the pivot we pick might be between
1365  * nodes that were coalesced, and thus in the middle of a child node post
1366  * coalescing:
1367  */
1368 static void btree_split_insert_keys(struct btree_update *as,
1369                                     struct btree_trans *trans,
1370                                     struct btree_path *path,
1371                                     struct btree *b,
1372                                     struct keylist *keys)
1373 {
1374         struct btree_node_iter node_iter;
1375         struct bkey_i *k = bch2_keylist_front(keys);
1376         struct bkey_packed *src, *dst, *n;
1377         struct bset *i;
1378
1379         bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
1380
1381         __bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
1382
1383         /*
1384          * We can't tolerate whiteouts here - with whiteouts there can be
1385          * duplicate keys, and it would be rather bad if we picked a duplicate
1386          * for the pivot:
1387          */
1388         i = btree_bset_first(b);
1389         src = dst = i->start;
1390         while (src != vstruct_last(i)) {
1391                 n = bkey_next(src);
1392                 if (!bkey_deleted(src)) {
1393                         memmove_u64s_down(dst, src, src->u64s);
1394                         dst = bkey_next(dst);
1395                 }
1396                 src = n;
1397         }
1398
1399         /* Also clear out the unwritten whiteouts area: */
1400         b->whiteout_u64s = 0;
1401
1402         i->u64s = cpu_to_le16((u64 *) dst - i->_data);
1403         set_btree_bset_end(b, b->set);
1404
1405         BUG_ON(b->nsets != 1 ||
1406                b->nr.live_u64s != le16_to_cpu(btree_bset_first(b)->u64s));
1407
1408         btree_node_interior_verify(as->c, b);
1409 }
1410
1411 static void btree_split(struct btree_update *as, struct btree_trans *trans,
1412                         struct btree_path *path, struct btree *b,
1413                         struct keylist *keys, unsigned flags)
1414 {
1415         struct bch_fs *c = as->c;
1416         struct btree *parent = btree_node_parent(path, b);
1417         struct btree *n1, *n2 = NULL, *n3 = NULL;
1418         u64 start_time = local_clock();
1419
1420         BUG_ON(!parent && (b != btree_node_root(c, b)));
1421         BUG_ON(!btree_node_intent_locked(path, btree_node_root(c, b)->c.level));
1422
1423         bch2_btree_interior_update_will_free_node(as, b);
1424
1425         n1 = bch2_btree_node_alloc_replacement(as, b);
1426
1427         if (keys)
1428                 btree_split_insert_keys(as, trans, path, n1, keys);
1429
1430         if (bset_u64s(&n1->set[0]) > BTREE_SPLIT_THRESHOLD(c)) {
1431                 trace_btree_split(c, b);
1432
1433                 n2 = __btree_split_node(as, n1);
1434
1435                 bch2_btree_build_aux_trees(n2);
1436                 bch2_btree_build_aux_trees(n1);
1437                 six_unlock_write(&n2->c.lock);
1438                 six_unlock_write(&n1->c.lock);
1439
1440                 bch2_btree_update_add_new_node(as, n1);
1441
1442                 bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1443                 bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0);
1444
1445                 /*
1446                  * Note that on recursive parent_keys == keys, so we
1447                  * can't start adding new keys to parent_keys before emptying it
1448                  * out (which we did with btree_split_insert_keys() above)
1449                  */
1450                 bch2_keylist_add(&as->parent_keys, &n1->key);
1451                 bch2_keylist_add(&as->parent_keys, &n2->key);
1452
1453                 if (!parent) {
1454                         /* Depth increases, make a new root */
1455                         n3 = __btree_root_alloc(as, b->c.level + 1);
1456
1457                         n3->sib_u64s[0] = U16_MAX;
1458                         n3->sib_u64s[1] = U16_MAX;
1459
1460                         btree_split_insert_keys(as, trans, path, n3, &as->parent_keys);
1461
1462                         bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0);
1463                 }
1464         } else {
1465                 trace_btree_compact(c, b);
1466
1467                 bch2_btree_build_aux_trees(n1);
1468                 six_unlock_write(&n1->c.lock);
1469
1470                 bch2_btree_update_add_new_node(as, n1);
1471
1472                 bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1473
1474                 if (parent)
1475                         bch2_keylist_add(&as->parent_keys, &n1->key);
1476         }
1477
1478         /* New nodes all written, now make them visible: */
1479
1480         if (parent) {
1481                 /* Split a non root node */
1482                 bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1483         } else if (n3) {
1484                 bch2_btree_set_root(as, trans, path, n3);
1485         } else {
1486                 /* Root filled up but didn't need to be split */
1487                 bch2_btree_set_root(as, trans, path, n1);
1488         }
1489
1490         bch2_btree_update_get_open_buckets(as, n1);
1491         if (n2)
1492                 bch2_btree_update_get_open_buckets(as, n2);
1493         if (n3)
1494                 bch2_btree_update_get_open_buckets(as, n3);
1495
1496         /* Successful split, update the path to point to the new nodes: */
1497
1498         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1499         if (n3)
1500                 bch2_trans_node_add(trans, n3);
1501         if (n2)
1502                 bch2_trans_node_add(trans, n2);
1503         bch2_trans_node_add(trans, n1);
1504
1505         /*
1506          * The old node must be freed (in memory) _before_ unlocking the new
1507          * nodes - else another thread could re-acquire a read lock on the old
1508          * node after another thread has locked and updated the new node, thus
1509          * seeing stale data:
1510          */
1511         bch2_btree_node_free_inmem(trans, b);
1512
1513         if (n3)
1514                 six_unlock_intent(&n3->c.lock);
1515         if (n2)
1516                 six_unlock_intent(&n2->c.lock);
1517         six_unlock_intent(&n1->c.lock);
1518
1519         bch2_trans_verify_locks(trans);
1520
1521         bch2_time_stats_update(&c->times[n2
1522                                ? BCH_TIME_btree_node_split
1523                                : BCH_TIME_btree_node_compact],
1524                                start_time);
1525 }
1526
1527 static void
1528 bch2_btree_insert_keys_interior(struct btree_update *as,
1529                                 struct btree_trans *trans,
1530                                 struct btree_path *path,
1531                                 struct btree *b,
1532                                 struct keylist *keys)
1533 {
1534         struct btree_path *linked;
1535
1536         __bch2_btree_insert_keys_interior(as, trans, path, b,
1537                                           path->l[b->c.level].iter, keys);
1538
1539         btree_update_updated_node(as, b);
1540
1541         trans_for_each_path_with_node(trans, b, linked)
1542                 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1543
1544         bch2_trans_verify_paths(trans);
1545 }
1546
1547 /**
1548  * bch_btree_insert_node - insert bkeys into a given btree node
1549  *
1550  * @iter:               btree iterator
1551  * @keys:               list of keys to insert
1552  * @hook:               insert callback
1553  * @persistent:         if not null, @persistent will wait on journal write
1554  *
1555  * Inserts as many keys as it can into a given btree node, splitting it if full.
1556  * If a split occurred, this function will return early. This can only happen
1557  * for leaf nodes -- inserts into interior nodes have to be atomic.
1558  */
1559 static void bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans,
1560                                    struct btree_path *path, struct btree *b,
1561                                    struct keylist *keys, unsigned flags)
1562 {
1563         struct bch_fs *c = as->c;
1564         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1565         int old_live_u64s = b->nr.live_u64s;
1566         int live_u64s_added, u64s_added;
1567
1568         lockdep_assert_held(&c->gc_lock);
1569         BUG_ON(!btree_node_intent_locked(path, btree_node_root(c, b)->c.level));
1570         BUG_ON(!b->c.level);
1571         BUG_ON(!as || as->b);
1572         bch2_verify_keylist_sorted(keys);
1573
1574         bch2_btree_node_lock_for_insert(trans, path, b);
1575
1576         if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1577                 bch2_btree_node_unlock_write(trans, path, b);
1578                 goto split;
1579         }
1580
1581         btree_node_interior_verify(c, b);
1582
1583         bch2_btree_insert_keys_interior(as, trans, path, b, keys);
1584
1585         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1586         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1587
1588         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1589                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1590         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1591                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1592
1593         if (u64s_added > live_u64s_added &&
1594             bch2_maybe_compact_whiteouts(c, b))
1595                 bch2_trans_node_reinit_iter(trans, b);
1596
1597         bch2_btree_node_unlock_write(trans, path, b);
1598
1599         btree_node_interior_verify(c, b);
1600         return;
1601 split:
1602         btree_split(as, trans, path, b, keys, flags);
1603 }
1604
1605 int bch2_btree_split_leaf(struct btree_trans *trans,
1606                           struct btree_path *path,
1607                           unsigned flags)
1608 {
1609         struct btree *b = path_l(path)->b;
1610         struct btree_update *as;
1611         unsigned l;
1612         int ret = 0;
1613
1614         as = bch2_btree_update_start(trans, path, path->level,
1615                                      true, flags);
1616         if (IS_ERR(as))
1617                 return PTR_ERR(as);
1618
1619         btree_split(as, trans, path, b, NULL, flags);
1620         bch2_btree_update_done(as);
1621
1622         for (l = path->level + 1; btree_path_node(path, l) && !ret; l++)
1623                 ret = bch2_foreground_maybe_merge(trans, path, l, flags);
1624
1625         return ret;
1626 }
1627
1628 int __bch2_foreground_maybe_merge(struct btree_trans *trans,
1629                                   struct btree_path *path,
1630                                   unsigned level,
1631                                   unsigned flags,
1632                                   enum btree_node_sibling sib)
1633 {
1634         struct bch_fs *c = trans->c;
1635         struct btree_path *sib_path = NULL;
1636         struct btree_update *as;
1637         struct bkey_format_state new_s;
1638         struct bkey_format new_f;
1639         struct bkey_i delete;
1640         struct btree *b, *m, *n, *prev, *next, *parent;
1641         struct bpos sib_pos;
1642         size_t sib_u64s;
1643         u64 start_time = local_clock();
1644         int ret = 0;
1645
1646         BUG_ON(!path->should_be_locked);
1647         BUG_ON(!btree_node_locked(path, level));
1648
1649         b = path->l[level].b;
1650
1651         if ((sib == btree_prev_sib && !bpos_cmp(b->data->min_key, POS_MIN)) ||
1652             (sib == btree_next_sib && !bpos_cmp(b->data->max_key, SPOS_MAX))) {
1653                 b->sib_u64s[sib] = U16_MAX;
1654                 return 0;
1655         }
1656
1657         sib_pos = sib == btree_prev_sib
1658                 ? bpos_predecessor(b->data->min_key)
1659                 : bpos_successor(b->data->max_key);
1660
1661         sib_path = bch2_path_get(trans, path->btree_id, sib_pos,
1662                                  U8_MAX, level, BTREE_ITER_INTENT, _THIS_IP_);
1663         ret = bch2_btree_path_traverse(trans, sib_path, false);
1664         if (ret)
1665                 goto err;
1666
1667         btree_path_set_should_be_locked(sib_path);
1668
1669         m = sib_path->l[level].b;
1670
1671         if (btree_node_parent(path, b) !=
1672             btree_node_parent(sib_path, m)) {
1673                 b->sib_u64s[sib] = U16_MAX;
1674                 goto out;
1675         }
1676
1677         if (sib == btree_prev_sib) {
1678                 prev = m;
1679                 next = b;
1680         } else {
1681                 prev = b;
1682                 next = m;
1683         }
1684
1685         if (bkey_cmp(bpos_successor(prev->data->max_key), next->data->min_key)) {
1686                 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1687
1688                 bch2_bpos_to_text(&buf1, prev->data->max_key);
1689                 bch2_bpos_to_text(&buf2, next->data->min_key);
1690                 bch_err(c,
1691                         "btree topology error in btree merge:\n"
1692                         "  prev ends at   %s\n"
1693                         "  next starts at %s",
1694                         buf1.buf, buf2.buf);
1695                 printbuf_exit(&buf1);
1696                 printbuf_exit(&buf2);
1697                 bch2_topology_error(c);
1698                 ret = -EIO;
1699                 goto err;
1700         }
1701
1702         bch2_bkey_format_init(&new_s);
1703         bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
1704         __bch2_btree_calc_format(&new_s, prev);
1705         __bch2_btree_calc_format(&new_s, next);
1706         bch2_bkey_format_add_pos(&new_s, next->data->max_key);
1707         new_f = bch2_bkey_format_done(&new_s);
1708
1709         sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1710                 btree_node_u64s_with_format(m, &new_f);
1711
1712         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1713                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1714                 sib_u64s /= 2;
1715                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1716         }
1717
1718         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1719         sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
1720         b->sib_u64s[sib] = sib_u64s;
1721
1722         if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
1723                 goto out;
1724
1725         parent = btree_node_parent(path, b);
1726         as = bch2_btree_update_start(trans, path, level, false,
1727                          BTREE_INSERT_NOFAIL|
1728                          BTREE_INSERT_USE_RESERVE|
1729                          flags);
1730         ret = PTR_ERR_OR_ZERO(as);
1731         if (ret)
1732                 goto err;
1733
1734         trace_btree_merge(c, b);
1735
1736         bch2_btree_interior_update_will_free_node(as, b);
1737         bch2_btree_interior_update_will_free_node(as, m);
1738
1739         n = bch2_btree_node_alloc(as, b->c.level);
1740
1741         SET_BTREE_NODE_SEQ(n->data,
1742                            max(BTREE_NODE_SEQ(b->data),
1743                                BTREE_NODE_SEQ(m->data)) + 1);
1744
1745         btree_set_min(n, prev->data->min_key);
1746         btree_set_max(n, next->data->max_key);
1747
1748         bch2_btree_update_add_new_node(as, n);
1749
1750         n->data->format  = new_f;
1751         btree_node_set_format(n, new_f);
1752
1753         bch2_btree_sort_into(c, n, prev);
1754         bch2_btree_sort_into(c, n, next);
1755
1756         bch2_btree_build_aux_trees(n);
1757         six_unlock_write(&n->c.lock);
1758
1759         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1760
1761         bkey_init(&delete.k);
1762         delete.k.p = prev->key.k.p;
1763         bch2_keylist_add(&as->parent_keys, &delete);
1764         bch2_keylist_add(&as->parent_keys, &n->key);
1765
1766         bch2_trans_verify_paths(trans);
1767
1768         bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1769
1770         bch2_trans_verify_paths(trans);
1771
1772         bch2_btree_update_get_open_buckets(as, n);
1773
1774         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1775         six_lock_increment(&m->c.lock, SIX_LOCK_intent);
1776
1777         bch2_trans_node_add(trans, n);
1778
1779         bch2_trans_verify_paths(trans);
1780
1781         bch2_btree_node_free_inmem(trans, b);
1782         bch2_btree_node_free_inmem(trans, m);
1783
1784         six_unlock_intent(&n->c.lock);
1785
1786         bch2_btree_update_done(as);
1787
1788         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time);
1789 out:
1790 err:
1791         bch2_path_put(trans, sib_path, true);
1792         bch2_trans_verify_locks(trans);
1793         return ret;
1794 }
1795
1796 /**
1797  * bch_btree_node_rewrite - Rewrite/move a btree node
1798  */
1799 int bch2_btree_node_rewrite(struct btree_trans *trans,
1800                             struct btree_iter *iter,
1801                             struct btree *b,
1802                             unsigned flags)
1803 {
1804         struct bch_fs *c = trans->c;
1805         struct btree *n, *parent;
1806         struct btree_update *as;
1807         int ret;
1808
1809         flags |= BTREE_INSERT_NOFAIL;
1810
1811         parent = btree_node_parent(iter->path, b);
1812         as = bch2_btree_update_start(trans, iter->path, b->c.level,
1813                                      false, flags);
1814         ret = PTR_ERR_OR_ZERO(as);
1815         if (ret)
1816                 goto out;
1817
1818         bch2_btree_interior_update_will_free_node(as, b);
1819
1820         n = bch2_btree_node_alloc_replacement(as, b);
1821         bch2_btree_update_add_new_node(as, n);
1822
1823         bch2_btree_build_aux_trees(n);
1824         six_unlock_write(&n->c.lock);
1825
1826         trace_btree_rewrite(c, b);
1827
1828         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1829
1830         if (parent) {
1831                 bch2_keylist_add(&as->parent_keys, &n->key);
1832                 bch2_btree_insert_node(as, trans, iter->path, parent,
1833                                        &as->parent_keys, flags);
1834         } else {
1835                 bch2_btree_set_root(as, trans, iter->path, n);
1836         }
1837
1838         bch2_btree_update_get_open_buckets(as, n);
1839
1840         six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1841         bch2_trans_node_add(trans, n);
1842         bch2_btree_node_free_inmem(trans, b);
1843         six_unlock_intent(&n->c.lock);
1844
1845         bch2_btree_update_done(as);
1846 out:
1847         bch2_btree_path_downgrade(trans, iter->path);
1848         return ret;
1849 }
1850
1851 struct async_btree_rewrite {
1852         struct bch_fs           *c;
1853         struct work_struct      work;
1854         enum btree_id           btree_id;
1855         unsigned                level;
1856         struct bpos             pos;
1857         __le64                  seq;
1858 };
1859
1860 static int async_btree_node_rewrite_trans(struct btree_trans *trans,
1861                                           struct async_btree_rewrite *a)
1862 {
1863         struct btree_iter iter;
1864         struct btree *b;
1865         int ret;
1866
1867         bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos,
1868                                   BTREE_MAX_DEPTH, a->level, 0);
1869         b = bch2_btree_iter_peek_node(&iter);
1870         ret = PTR_ERR_OR_ZERO(b);
1871         if (ret)
1872                 goto out;
1873
1874         if (!b || b->data->keys.seq != a->seq)
1875                 goto out;
1876
1877         ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
1878 out :
1879         bch2_trans_iter_exit(trans, &iter);
1880
1881         return ret;
1882 }
1883
1884 void async_btree_node_rewrite_work(struct work_struct *work)
1885 {
1886         struct async_btree_rewrite *a =
1887                 container_of(work, struct async_btree_rewrite, work);
1888         struct bch_fs *c = a->c;
1889
1890         bch2_trans_do(c, NULL, NULL, 0,
1891                       async_btree_node_rewrite_trans(&trans, a));
1892         percpu_ref_put(&c->writes);
1893         kfree(a);
1894 }
1895
1896 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
1897 {
1898         struct async_btree_rewrite *a;
1899
1900         if (!percpu_ref_tryget_live(&c->writes))
1901                 return;
1902
1903         a = kmalloc(sizeof(*a), GFP_NOFS);
1904         if (!a) {
1905                 percpu_ref_put(&c->writes);
1906                 return;
1907         }
1908
1909         a->c            = c;
1910         a->btree_id     = b->c.btree_id;
1911         a->level        = b->c.level;
1912         a->pos          = b->key.k.p;
1913         a->seq          = b->data->keys.seq;
1914
1915         INIT_WORK(&a->work, async_btree_node_rewrite_work);
1916         queue_work(c->btree_interior_update_worker, &a->work);
1917 }
1918
1919 static int __bch2_btree_node_update_key(struct btree_trans *trans,
1920                                         struct btree_iter *iter,
1921                                         struct btree *b, struct btree *new_hash,
1922                                         struct bkey_i *new_key,
1923                                         bool skip_triggers)
1924 {
1925         struct bch_fs *c = trans->c;
1926         struct btree_iter iter2 = { NULL };
1927         struct btree *parent;
1928         int ret;
1929
1930         if (!skip_triggers) {
1931                 ret = bch2_trans_mark_old(trans, b->c.btree_id, b->c.level + 1,
1932                                           bkey_i_to_s_c(&b->key), 0);
1933                 if (ret)
1934                         return ret;
1935
1936                 ret = bch2_trans_mark_new(trans, b->c.btree_id, b->c.level + 1,
1937                                           new_key, 0);
1938                 if (ret)
1939                         return ret;
1940         }
1941
1942         if (new_hash) {
1943                 bkey_copy(&new_hash->key, new_key);
1944                 ret = bch2_btree_node_hash_insert(&c->btree_cache,
1945                                 new_hash, b->c.level, b->c.btree_id);
1946                 BUG_ON(ret);
1947         }
1948
1949         parent = btree_node_parent(iter->path, b);
1950         if (parent) {
1951                 bch2_trans_copy_iter(&iter2, iter);
1952
1953                 iter2.path = bch2_btree_path_make_mut(trans, iter2.path,
1954                                 iter2.flags & BTREE_ITER_INTENT,
1955                                 _THIS_IP_);
1956
1957                 BUG_ON(iter2.path->level != b->c.level);
1958                 BUG_ON(bpos_cmp(iter2.path->pos, new_key->k.p));
1959
1960                 btree_path_set_level_up(trans, iter2.path);
1961
1962                 bch2_btree_path_check_sort(trans, iter2.path, 0);
1963
1964                 ret   = bch2_btree_iter_traverse(&iter2) ?:
1965                         bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_NORUN);
1966                 if (ret)
1967                         goto err;
1968         } else {
1969                 BUG_ON(btree_node_root(c, b) != b);
1970
1971                 ret = darray_make_room(&trans->extra_journal_entries,
1972                                        jset_u64s(new_key->k.u64s));
1973                 if (ret)
1974                         return ret;
1975
1976                 journal_entry_set((void *) &darray_top(trans->extra_journal_entries),
1977                                   BCH_JSET_ENTRY_btree_root,
1978                                   b->c.btree_id, b->c.level,
1979                                   new_key, new_key->k.u64s);
1980                 trans->extra_journal_entries.nr += jset_u64s(new_key->k.u64s);
1981         }
1982
1983         ret = bch2_trans_commit(trans, NULL, NULL,
1984                                 BTREE_INSERT_NOFAIL|
1985                                 BTREE_INSERT_NOCHECK_RW|
1986                                 BTREE_INSERT_USE_RESERVE|
1987                                 BTREE_INSERT_JOURNAL_RECLAIM|
1988                                 JOURNAL_WATERMARK_reserved);
1989         if (ret)
1990                 goto err;
1991
1992         bch2_btree_node_lock_write(trans, iter->path, b);
1993
1994         if (new_hash) {
1995                 mutex_lock(&c->btree_cache.lock);
1996                 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
1997                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1998
1999                 bkey_copy(&b->key, new_key);
2000                 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
2001                 BUG_ON(ret);
2002                 mutex_unlock(&c->btree_cache.lock);
2003         } else {
2004                 bkey_copy(&b->key, new_key);
2005         }
2006
2007         bch2_btree_node_unlock_write(trans, iter->path, b);
2008 out:
2009         bch2_trans_iter_exit(trans, &iter2);
2010         return ret;
2011 err:
2012         if (new_hash) {
2013                 mutex_lock(&c->btree_cache.lock);
2014                 bch2_btree_node_hash_remove(&c->btree_cache, b);
2015                 mutex_unlock(&c->btree_cache.lock);
2016         }
2017         goto out;
2018 }
2019
2020 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
2021                                struct btree *b, struct bkey_i *new_key,
2022                                bool skip_triggers)
2023 {
2024         struct bch_fs *c = trans->c;
2025         struct btree *new_hash = NULL;
2026         struct btree_path *path = iter->path;
2027         struct closure cl;
2028         int ret = 0;
2029
2030         if (!btree_node_intent_locked(path, b->c.level) &&
2031             !bch2_btree_path_upgrade(trans, path, b->c.level + 1))
2032                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_upgrade);
2033
2034         closure_init_stack(&cl);
2035
2036         /*
2037          * check btree_ptr_hash_val() after @b is locked by
2038          * btree_iter_traverse():
2039          */
2040         if (btree_ptr_hash_val(new_key) != b->hash_val) {
2041                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2042                 if (ret) {
2043                         bch2_trans_unlock(trans);
2044                         closure_sync(&cl);
2045                         ret = bch2_trans_relock(trans);
2046                         if (ret)
2047                                 return ret;
2048                 }
2049
2050                 new_hash = bch2_btree_node_mem_alloc(c, false);
2051         }
2052
2053         path->intent_ref++;
2054         ret = __bch2_btree_node_update_key(trans, iter, b, new_hash,
2055                                            new_key, skip_triggers);
2056         --path->intent_ref;
2057
2058         if (new_hash) {
2059                 mutex_lock(&c->btree_cache.lock);
2060                 list_move(&new_hash->list, &c->btree_cache.freeable);
2061                 mutex_unlock(&c->btree_cache.lock);
2062
2063                 six_unlock_write(&new_hash->c.lock);
2064                 six_unlock_intent(&new_hash->c.lock);
2065         }
2066         closure_sync(&cl);
2067         bch2_btree_cache_cannibalize_unlock(c);
2068         return ret;
2069 }
2070
2071 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
2072                                         struct btree *b, struct bkey_i *new_key,
2073                                         bool skip_triggers)
2074 {
2075         struct btree_iter iter;
2076         int ret;
2077
2078         bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p,
2079                                   BTREE_MAX_DEPTH, b->c.level,
2080                                   BTREE_ITER_INTENT);
2081         ret = bch2_btree_iter_traverse(&iter);
2082         if (ret)
2083                 goto out;
2084
2085         /* has node been freed? */
2086         if (iter.path->l[b->c.level].b != b) {
2087                 /* node has been freed: */
2088                 BUG_ON(!btree_node_dying(b));
2089                 goto out;
2090         }
2091
2092         BUG_ON(!btree_node_hashed(b));
2093
2094         ret = bch2_btree_node_update_key(trans, &iter, b, new_key, skip_triggers);
2095 out:
2096         bch2_trans_iter_exit(trans, &iter);
2097         return ret;
2098 }
2099
2100 /* Init code: */
2101
2102 /*
2103  * Only for filesystem bringup, when first reading the btree roots or allocating
2104  * btree roots when initializing a new filesystem:
2105  */
2106 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2107 {
2108         BUG_ON(btree_node_root(c, b));
2109
2110         bch2_btree_set_root_inmem(c, b);
2111 }
2112
2113 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2114 {
2115         struct closure cl;
2116         struct btree *b;
2117         int ret;
2118
2119         closure_init_stack(&cl);
2120
2121         do {
2122                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2123                 closure_sync(&cl);
2124         } while (ret);
2125
2126         b = bch2_btree_node_mem_alloc(c, false);
2127         bch2_btree_cache_cannibalize_unlock(c);
2128
2129         set_btree_node_fake(b);
2130         set_btree_node_need_rewrite(b);
2131         b->c.level      = 0;
2132         b->c.btree_id   = id;
2133
2134         bkey_btree_ptr_init(&b->key);
2135         b->key.k.p = SPOS_MAX;
2136         *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2137
2138         bch2_bset_init_first(b, &b->data->keys);
2139         bch2_btree_build_aux_trees(b);
2140
2141         b->data->flags = 0;
2142         btree_set_min(b, POS_MIN);
2143         btree_set_max(b, SPOS_MAX);
2144         b->data->format = bch2_btree_calc_format(b);
2145         btree_node_set_format(b, b->data->format);
2146
2147         ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2148                                           b->c.level, b->c.btree_id);
2149         BUG_ON(ret);
2150
2151         bch2_btree_set_root_inmem(c, b);
2152
2153         six_unlock_write(&b->c.lock);
2154         six_unlock_intent(&b->c.lock);
2155 }
2156
2157 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2158 {
2159         struct btree_update *as;
2160
2161         mutex_lock(&c->btree_interior_update_lock);
2162         list_for_each_entry(as, &c->btree_interior_update_list, list)
2163                 prt_printf(out, "%p m %u w %u r %u j %llu\n",
2164                        as,
2165                        as->mode,
2166                        as->nodes_written,
2167                        atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2168                        as->journal.seq);
2169         mutex_unlock(&c->btree_interior_update_lock);
2170 }
2171
2172 static bool bch2_btree_interior_updates_pending(struct bch_fs *c)
2173 {
2174         bool ret;
2175
2176         mutex_lock(&c->btree_interior_update_lock);
2177         ret = !list_empty(&c->btree_interior_update_list);
2178         mutex_unlock(&c->btree_interior_update_lock);
2179
2180         return ret;
2181 }
2182
2183 bool bch2_btree_interior_updates_flush(struct bch_fs *c)
2184 {
2185         bool ret = bch2_btree_interior_updates_pending(c);
2186
2187         if (ret)
2188                 closure_wait_event(&c->btree_interior_update_wait,
2189                                    !bch2_btree_interior_updates_pending(c));
2190         return ret;
2191 }
2192
2193 void bch2_journal_entries_to_btree_roots(struct bch_fs *c, struct jset *jset)
2194 {
2195         struct btree_root *r;
2196         struct jset_entry *entry;
2197
2198         mutex_lock(&c->btree_root_lock);
2199
2200         vstruct_for_each(jset, entry)
2201                 if (entry->type == BCH_JSET_ENTRY_btree_root) {
2202                         r = &c->btree_roots[entry->btree_id];
2203                         r->level = entry->level;
2204                         r->alive = true;
2205                         bkey_copy(&r->key, &entry->start[0]);
2206                 }
2207
2208         mutex_unlock(&c->btree_root_lock);
2209 }
2210
2211 struct jset_entry *
2212 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2213                                     struct jset_entry *start,
2214                                     struct jset_entry *end)
2215 {
2216         struct jset_entry *entry;
2217         unsigned long have = 0;
2218         unsigned i;
2219
2220         for (entry = start; entry < end; entry = vstruct_next(entry))
2221                 if (entry->type == BCH_JSET_ENTRY_btree_root)
2222                         __set_bit(entry->btree_id, &have);
2223
2224         mutex_lock(&c->btree_root_lock);
2225
2226         for (i = 0; i < BTREE_ID_NR; i++)
2227                 if (c->btree_roots[i].alive && !test_bit(i, &have)) {
2228                         journal_entry_set(end,
2229                                           BCH_JSET_ENTRY_btree_root,
2230                                           i, c->btree_roots[i].level,
2231                                           &c->btree_roots[i].key,
2232                                           c->btree_roots[i].key.u64s);
2233                         end = vstruct_next(end);
2234                 }
2235
2236         mutex_unlock(&c->btree_root_lock);
2237
2238         return end;
2239 }
2240
2241 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2242 {
2243         if (c->btree_interior_update_worker)
2244                 destroy_workqueue(c->btree_interior_update_worker);
2245         mempool_exit(&c->btree_interior_update_pool);
2246 }
2247
2248 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2249 {
2250         mutex_init(&c->btree_reserve_cache_lock);
2251         INIT_LIST_HEAD(&c->btree_interior_update_list);
2252         INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2253         mutex_init(&c->btree_interior_update_lock);
2254         INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2255
2256         c->btree_interior_update_worker =
2257                 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2258         if (!c->btree_interior_update_worker)
2259                 return -ENOMEM;
2260
2261         return mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2262                                          sizeof(struct btree_update));
2263 }