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 "extents.h"
  15 #include "journal.h"
  16 #include "journal_reclaim.h"
  17 #include "keylist.h"
  18 #include "replicas.h"
  19 #include "super-io.h"
  20
  21 #include <linux/random.h>
  22 #include <trace/events/bcachefs.h>
  23
  24 static void btree_node_will_make_reachable(struct btree_update *,
  25                                            struct btree *);
  26 static void btree_update_drop_new_node(struct bch_fs *, struct btree *);
  27
  28 /* Debug code: */
  29
  30 /*
  31  * Verify that child nodes correctly span parent node's range:
  32  */
  33 static void btree_node_interior_verify(struct btree *b)
  34 {
  35 #ifdef CONFIG_BCACHEFS_DEBUG
  36         struct bpos next_node = b->data->min_key;
  37         struct btree_node_iter iter;
  38         struct bkey_s_c k;
  39         struct bkey_s_c_btree_ptr_v2 bp;
  40         struct bkey unpacked;
  41
  42         BUG_ON(!b->level);
  43
  44         bch2_btree_node_iter_init_from_start(&iter, b);
  45
  46         while (1) {
  47                 k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked);
  48                 if (k.k->type != KEY_TYPE_btree_ptr_v2)
  49                         break;
  50                 bp = bkey_s_c_to_btree_ptr_v2(k);
  51
  52                 BUG_ON(bkey_cmp(next_node, bp.v->min_key));
  53
  54                 bch2_btree_node_iter_advance(&iter, b);
  55
  56                 if (bch2_btree_node_iter_end(&iter)) {
  57                         BUG_ON(bkey_cmp(k.k->p, b->key.k.p));
  58                         break;
  59                 }
  60
  61                 next_node = bkey_successor(k.k->p);
  62         }
  63 #endif
  64 }
  65
  66 /* Calculate ideal packed bkey format for new btree nodes: */
  67
  68 void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b)
  69 {
  70         struct bkey_packed *k;
  71         struct bset_tree *t;
  72         struct bkey uk;
  73
  74         bch2_bkey_format_add_pos(s, b->data->min_key);
  75
  76         for_each_bset(b, t)
  77                 bset_tree_for_each_key(b, t, k)
  78                         if (!bkey_whiteout(k)) {
  79                                 uk = bkey_unpack_key(b, k);
  80                                 bch2_bkey_format_add_key(s, &uk);
  81                         }
  82 }
  83
  84 static struct bkey_format bch2_btree_calc_format(struct btree *b)
  85 {
  86         struct bkey_format_state s;
  87
  88         bch2_bkey_format_init(&s);
  89         __bch2_btree_calc_format(&s, b);
  90
  91         return bch2_bkey_format_done(&s);
  92 }
  93
  94 static size_t btree_node_u64s_with_format(struct btree *b,
  95                                           struct bkey_format *new_f)
  96 {
  97         struct bkey_format *old_f = &b->format;
  98
  99         /* stupid integer promotion rules */
 100         ssize_t delta =
 101             (((int) new_f->key_u64s - old_f->key_u64s) *
 102              (int) b->nr.packed_keys) +
 103             (((int) new_f->key_u64s - BKEY_U64s) *
 104              (int) b->nr.unpacked_keys);
 105
 106         BUG_ON(delta + b->nr.live_u64s < 0);
 107
 108         return b->nr.live_u64s + delta;
 109 }
 110
 111 /**
 112  * btree_node_format_fits - check if we could rewrite node with a new format
 113  *
 114  * This assumes all keys can pack with the new format -- it just checks if
 115  * the re-packed keys would fit inside the node itself.
 116  */
 117 bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
 118                                  struct bkey_format *new_f)
 119 {
 120         size_t u64s = btree_node_u64s_with_format(b, new_f);
 121
 122         return __vstruct_bytes(struct btree_node, u64s) < btree_bytes(c);
 123 }
 124
 125 /* Btree node freeing/allocation: */
 126
 127 static bool btree_key_matches(struct bch_fs *c,
 128                               struct bkey_s_c l,
 129                               struct bkey_s_c r)
 130 {
 131         struct bkey_ptrs_c ptrs1 = bch2_bkey_ptrs_c(l);
 132         struct bkey_ptrs_c ptrs2 = bch2_bkey_ptrs_c(r);
 133         const struct bch_extent_ptr *ptr1, *ptr2;
 134
 135         bkey_for_each_ptr(ptrs1, ptr1)
 136                 bkey_for_each_ptr(ptrs2, ptr2)
 137                         if (ptr1->dev == ptr2->dev &&
 138                             ptr1->gen == ptr2->gen &&
 139                             ptr1->offset == ptr2->offset)
 140                                 return true;
 141
 142         return false;
 143 }
 144
 145 /*
 146  * We're doing the index update that makes @b unreachable, update stuff to
 147  * reflect that:
 148  *
 149  * Must be called _before_ btree_update_updated_root() or
 150  * btree_update_updated_node:
 151  */
 152 static void bch2_btree_node_free_index(struct btree_update *as, struct btree *b,
 153                                        struct bkey_s_c k,
 154                                        struct bch_fs_usage *stats)
 155 {
 156         struct bch_fs *c = as->c;
 157         struct pending_btree_node_free *d;
 158
 159         for (d = as->pending; d < as->pending + as->nr_pending; d++)
 160                 if (!bkey_cmp(k.k->p, d->key.k.p) &&
 161                     btree_key_matches(c, k, bkey_i_to_s_c(&d->key)))
 162                         goto found;
 163         BUG();
 164 found:
 165         BUG_ON(d->index_update_done);
 166         d->index_update_done = true;
 167
 168         /*
 169          * We're dropping @k from the btree, but it's still live until the
 170          * index update is persistent so we need to keep a reference around for
 171          * mark and sweep to find - that's primarily what the
 172          * btree_node_pending_free list is for.
 173          *
 174          * So here (when we set index_update_done = true), we're moving an
 175          * existing reference to a different part of the larger "gc keyspace" -
 176          * and the new position comes after the old position, since GC marks
 177          * the pending free list after it walks the btree.
 178          *
 179          * If we move the reference while mark and sweep is _between_ the old
 180          * and the new position, mark and sweep will see the reference twice
 181          * and it'll get double accounted - so check for that here and subtract
 182          * to cancel out one of mark and sweep's markings if necessary:
 183          */
 184
 185         if (gc_pos_cmp(c->gc_pos, b
 186                        ? gc_pos_btree_node(b)
 187                        : gc_pos_btree_root(as->btree_id)) >= 0 &&
 188             gc_pos_cmp(c->gc_pos, gc_phase(GC_PHASE_PENDING_DELETE)) < 0)
 189                 bch2_mark_key_locked(c, bkey_i_to_s_c(&d->key),
 190                               0, 0, NULL, 0,
 191                               BTREE_TRIGGER_OVERWRITE|
 192                               BTREE_TRIGGER_GC);
 193 }
 194
 195 static void __btree_node_free(struct bch_fs *c, struct btree *b)
 196 {
 197         trace_btree_node_free(c, b);
 198
 199         BUG_ON(btree_node_dirty(b));
 200         BUG_ON(btree_node_need_write(b));
 201         BUG_ON(b == btree_node_root(c, b));
 202         BUG_ON(b->ob.nr);
 203         BUG_ON(!list_empty(&b->write_blocked));
 204         BUG_ON(b->will_make_reachable);
 205
 206         clear_btree_node_noevict(b);
 207
 208         bch2_btree_node_hash_remove(&c->btree_cache, b);
 209
 210         mutex_lock(&c->btree_cache.lock);
 211         list_move(&b->list, &c->btree_cache.freeable);
 212         mutex_unlock(&c->btree_cache.lock);
 213 }
 214
 215 void bch2_btree_node_free_never_inserted(struct bch_fs *c, struct btree *b)
 216 {
 217         struct open_buckets ob = b->ob;
 218
 219         btree_update_drop_new_node(c, b);
 220
 221         b->ob.nr = 0;
 222
 223         clear_btree_node_dirty(b);
 224
 225         btree_node_lock_type(c, b, SIX_LOCK_write);
 226         __btree_node_free(c, b);
 227         six_unlock_write(&b->lock);
 228
 229         bch2_open_buckets_put(c, &ob);
 230 }
 231
 232 void bch2_btree_node_free_inmem(struct bch_fs *c, struct btree *b,
 233                                 struct btree_iter *iter)
 234 {
 235         struct btree_iter *linked;
 236
 237         trans_for_each_iter(iter->trans, linked)
 238                 BUG_ON(linked->l[b->level].b == b);
 239
 240         /*
 241          * Is this a node that isn't reachable on disk yet?
 242          *
 243          * Nodes that aren't reachable yet have writes blocked until they're
 244          * reachable - now that we've cancelled any pending writes and moved
 245          * things waiting on that write to wait on this update, we can drop this
 246          * node from the list of nodes that the other update is making
 247          * reachable, prior to freeing it:
 248          */
 249         btree_update_drop_new_node(c, b);
 250
 251         six_lock_write(&b->lock);
 252         __btree_node_free(c, b);
 253         six_unlock_write(&b->lock);
 254         six_unlock_intent(&b->lock);
 255 }
 256
 257 static void bch2_btree_node_free_ondisk(struct bch_fs *c,
 258                         struct pending_btree_node_free *pending,
 259                         u64 journal_seq)
 260 {
 261         BUG_ON(!pending->index_update_done);
 262
 263         bch2_mark_key(c, bkey_i_to_s_c(&pending->key),
 264                       0, 0, NULL, journal_seq, BTREE_TRIGGER_OVERWRITE);
 265
 266         if (gc_visited(c, gc_phase(GC_PHASE_PENDING_DELETE)))
 267                 bch2_mark_key(c, bkey_i_to_s_c(&pending->key),
 268                               0, 0, NULL, journal_seq,
 269                               BTREE_TRIGGER_OVERWRITE|
 270                               BTREE_TRIGGER_GC);
 271 }
 272
 273 static struct btree *__bch2_btree_node_alloc(struct bch_fs *c,
 274                                              struct disk_reservation *res,
 275                                              struct closure *cl,
 276                                              unsigned flags)
 277 {
 278         struct write_point *wp;
 279         struct btree *b;
 280         BKEY_PADDED(k) tmp;
 281         struct open_buckets ob = { .nr = 0 };
 282         struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
 283         unsigned nr_reserve;
 284         enum alloc_reserve alloc_reserve;
 285
 286         if (flags & BTREE_INSERT_USE_ALLOC_RESERVE) {
 287                 nr_reserve      = 0;
 288                 alloc_reserve   = RESERVE_ALLOC;
 289         } else if (flags & BTREE_INSERT_USE_RESERVE) {
 290                 nr_reserve      = BTREE_NODE_RESERVE / 2;
 291                 alloc_reserve   = RESERVE_BTREE;
 292         } else {
 293                 nr_reserve      = BTREE_NODE_RESERVE;
 294                 alloc_reserve   = RESERVE_NONE;
 295         }
 296
 297         mutex_lock(&c->btree_reserve_cache_lock);
 298         if (c->btree_reserve_cache_nr > nr_reserve) {
 299                 struct btree_alloc *a =
 300                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
 301
 302                 ob = a->ob;
 303                 bkey_copy(&tmp.k, &a->k);
 304                 mutex_unlock(&c->btree_reserve_cache_lock);
 305                 goto mem_alloc;
 306         }
 307         mutex_unlock(&c->btree_reserve_cache_lock);
 308
 309 retry:
 310         wp = bch2_alloc_sectors_start(c, c->opts.foreground_target, 0,
 311                                       writepoint_ptr(&c->btree_write_point),
 312                                       &devs_have,
 313                                       res->nr_replicas,
 314                                       c->opts.metadata_replicas_required,
 315                                       alloc_reserve, 0, cl);
 316         if (IS_ERR(wp))
 317                 return ERR_CAST(wp);
 318
 319         if (wp->sectors_free < c->opts.btree_node_size) {
 320                 struct open_bucket *ob;
 321                 unsigned i;
 322
 323                 open_bucket_for_each(c, &wp->ptrs, ob, i)
 324                         if (ob->sectors_free < c->opts.btree_node_size)
 325                                 ob->sectors_free = 0;
 326
 327                 bch2_alloc_sectors_done(c, wp);
 328                 goto retry;
 329         }
 330
 331         if (c->sb.features & (1ULL << BCH_FEATURE_btree_ptr_v2))
 332                 bkey_btree_ptr_v2_init(&tmp.k);
 333         else
 334                 bkey_btree_ptr_init(&tmp.k);
 335
 336         bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, c->opts.btree_node_size);
 337
 338         bch2_open_bucket_get(c, wp, &ob);
 339         bch2_alloc_sectors_done(c, wp);
 340 mem_alloc:
 341         b = bch2_btree_node_mem_alloc(c);
 342
 343         /* we hold cannibalize_lock: */
 344         BUG_ON(IS_ERR(b));
 345         BUG_ON(b->ob.nr);
 346
 347         bkey_copy(&b->key, &tmp.k);
 348         b->ob = ob;
 349
 350         return b;
 351 }
 352
 353 static struct btree *bch2_btree_node_alloc(struct btree_update *as, unsigned level)
 354 {
 355         struct bch_fs *c = as->c;
 356         struct btree *b;
 357         int ret;
 358
 359         BUG_ON(level >= BTREE_MAX_DEPTH);
 360         BUG_ON(!as->reserve->nr);
 361
 362         b = as->reserve->b[--as->reserve->nr];
 363
 364         set_btree_node_accessed(b);
 365         set_btree_node_dirty(b);
 366         set_btree_node_need_write(b);
 367
 368         bch2_bset_init_first(b, &b->data->keys);
 369         b->level        = level;
 370         b->btree_id     = as->btree_id;
 371
 372         memset(&b->nr, 0, sizeof(b->nr));
 373         b->data->magic = cpu_to_le64(bset_magic(c));
 374         b->data->flags = 0;
 375         SET_BTREE_NODE_ID(b->data, as->btree_id);
 376         SET_BTREE_NODE_LEVEL(b->data, level);
 377         b->data->ptr = bch2_bkey_ptrs_c(bkey_i_to_s_c(&b->key)).start->ptr;
 378
 379         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 380                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
 381
 382                 bp->v.mem_ptr           = 0;
 383                 bp->v.seq               = b->data->keys.seq;
 384                 bp->v.sectors_written   = 0;
 385                 bp->v.sectors           = cpu_to_le16(c->opts.btree_node_size);
 386         }
 387
 388         if (c->sb.features & (1ULL << BCH_FEATURE_new_extent_overwrite))
 389                 SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
 390
 391         if (btree_node_is_extents(b) &&
 392             !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data))
 393                 set_btree_node_old_extent_overwrite(b);
 394
 395         bch2_btree_build_aux_trees(b);
 396
 397         btree_node_will_make_reachable(as, b);
 398
 399         ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
 400         BUG_ON(ret);
 401
 402         trace_btree_node_alloc(c, b);
 403         return b;
 404 }
 405
 406 static void btree_set_min(struct btree *b, struct bpos pos)
 407 {
 408         if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
 409                 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
 410         b->data->min_key = pos;
 411 }
 412
 413 static void btree_set_max(struct btree *b, struct bpos pos)
 414 {
 415         b->key.k.p = pos;
 416         b->data->max_key = pos;
 417 }
 418
 419 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *as,
 420                                                   struct btree *b,
 421                                                   struct bkey_format format)
 422 {
 423         struct btree *n;
 424
 425         n = bch2_btree_node_alloc(as, b->level);
 426
 427         SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
 428
 429         btree_set_min(n, b->data->min_key);
 430         btree_set_max(n, b->data->max_key);
 431
 432         n->data->format         = format;
 433         btree_node_set_format(n, format);
 434
 435         bch2_btree_sort_into(as->c, n, b);
 436
 437         btree_node_reset_sib_u64s(n);
 438
 439         n->key.k.p = b->key.k.p;
 440         return n;
 441 }
 442
 443 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
 444                                                        struct btree *b)
 445 {
 446         struct bkey_format new_f = bch2_btree_calc_format(b);
 447
 448         /*
 449          * The keys might expand with the new format - if they wouldn't fit in
 450          * the btree node anymore, use the old format for now:
 451          */
 452         if (!bch2_btree_node_format_fits(as->c, b, &new_f))
 453                 new_f = b->format;
 454
 455         return __bch2_btree_node_alloc_replacement(as, b, new_f);
 456 }
 457
 458 static struct btree *__btree_root_alloc(struct btree_update *as, unsigned level)
 459 {
 460         struct btree *b = bch2_btree_node_alloc(as, level);
 461
 462         btree_set_min(b, POS_MIN);
 463         btree_set_max(b, POS_MAX);
 464         b->data->format = bch2_btree_calc_format(b);
 465
 466         btree_node_set_format(b, b->data->format);
 467         bch2_btree_build_aux_trees(b);
 468
 469         six_unlock_write(&b->lock);
 470
 471         return b;
 472 }
 473
 474 static void bch2_btree_reserve_put(struct bch_fs *c, struct btree_reserve *reserve)
 475 {
 476         bch2_disk_reservation_put(c, &reserve->disk_res);
 477
 478         mutex_lock(&c->btree_reserve_cache_lock);
 479
 480         while (reserve->nr) {
 481                 struct btree *b = reserve->b[--reserve->nr];
 482
 483                 six_unlock_write(&b->lock);
 484
 485                 if (c->btree_reserve_cache_nr <
 486                     ARRAY_SIZE(c->btree_reserve_cache)) {
 487                         struct btree_alloc *a =
 488                                 &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
 489
 490                         a->ob = b->ob;
 491                         b->ob.nr = 0;
 492                         bkey_copy(&a->k, &b->key);
 493                 } else {
 494                         bch2_open_buckets_put(c, &b->ob);
 495                 }
 496
 497                 btree_node_lock_type(c, b, SIX_LOCK_write);
 498                 __btree_node_free(c, b);
 499                 six_unlock_write(&b->lock);
 500
 501                 six_unlock_intent(&b->lock);
 502         }
 503
 504         mutex_unlock(&c->btree_reserve_cache_lock);
 505
 506         mempool_free(reserve, &c->btree_reserve_pool);
 507 }
 508
 509 static struct btree_reserve *bch2_btree_reserve_get(struct bch_fs *c,
 510                                                     unsigned nr_nodes,
 511                                                     unsigned flags,
 512                                                     struct closure *cl)
 513 {
 514         struct btree_reserve *reserve;
 515         struct btree *b;
 516         struct disk_reservation disk_res = { 0, 0 };
 517         unsigned sectors = nr_nodes * c->opts.btree_node_size;
 518         int ret, disk_res_flags = 0;
 519
 520         if (flags & BTREE_INSERT_NOFAIL)
 521                 disk_res_flags |= BCH_DISK_RESERVATION_NOFAIL;
 522
 523         /*
 524          * This check isn't necessary for correctness - it's just to potentially
 525          * prevent us from doing a lot of work that'll end up being wasted:
 526          */
 527         ret = bch2_journal_error(&c->journal);
 528         if (ret)
 529                 return ERR_PTR(ret);
 530
 531         if (bch2_disk_reservation_get(c, &disk_res, sectors,
 532                                       c->opts.metadata_replicas,
 533                                       disk_res_flags))
 534                 return ERR_PTR(-ENOSPC);
 535
 536         BUG_ON(nr_nodes > BTREE_RESERVE_MAX);
 537
 538         /*
 539          * Protects reaping from the btree node cache and using the btree node
 540          * open bucket reserve:
 541          */
 542         ret = bch2_btree_cache_cannibalize_lock(c, cl);
 543         if (ret) {
 544                 bch2_disk_reservation_put(c, &disk_res);
 545                 return ERR_PTR(ret);
 546         }
 547
 548         reserve = mempool_alloc(&c->btree_reserve_pool, GFP_NOIO);
 549
 550         reserve->disk_res = disk_res;
 551         reserve->nr = 0;
 552
 553         while (reserve->nr < nr_nodes) {
 554                 b = __bch2_btree_node_alloc(c, &disk_res,
 555                                             flags & BTREE_INSERT_NOWAIT
 556                                             ? NULL : cl, flags);
 557                 if (IS_ERR(b)) {
 558                         ret = PTR_ERR(b);
 559                         goto err_free;
 560                 }
 561
 562                 ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(&b->key));
 563                 if (ret)
 564                         goto err_free;
 565
 566                 reserve->b[reserve->nr++] = b;
 567         }
 568
 569         bch2_btree_cache_cannibalize_unlock(c);
 570         return reserve;
 571 err_free:
 572         bch2_btree_reserve_put(c, reserve);
 573         bch2_btree_cache_cannibalize_unlock(c);
 574         trace_btree_reserve_get_fail(c, nr_nodes, cl);
 575         return ERR_PTR(ret);
 576 }
 577
 578 /* Asynchronous interior node update machinery */
 579
 580 static void __bch2_btree_update_free(struct btree_update *as)
 581 {
 582         struct bch_fs *c = as->c;
 583
 584         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 585
 586         bch2_journal_pin_drop(&c->journal, &as->journal);
 587         bch2_journal_pin_flush(&c->journal, &as->journal);
 588
 589         BUG_ON((as->nr_new_nodes || as->nr_pending) &&
 590                !bch2_journal_error(&c->journal));;
 591
 592         if (as->reserve)
 593                 bch2_btree_reserve_put(c, as->reserve);
 594
 595         list_del(&as->list);
 596
 597         closure_debug_destroy(&as->cl);
 598         mempool_free(as, &c->btree_interior_update_pool);
 599
 600         closure_wake_up(&c->btree_interior_update_wait);
 601 }
 602
 603 static void bch2_btree_update_free(struct btree_update *as)
 604 {
 605         struct bch_fs *c = as->c;
 606
 607         mutex_lock(&c->btree_interior_update_lock);
 608         __bch2_btree_update_free(as);
 609         mutex_unlock(&c->btree_interior_update_lock);
 610 }
 611
 612 static void btree_update_nodes_reachable(struct btree_update *as, u64 seq)
 613 {
 614         struct bch_fs *c = as->c;
 615
 616         while (as->nr_new_nodes) {
 617                 struct btree *b = as->new_nodes[--as->nr_new_nodes];
 618
 619                 BUG_ON(b->will_make_reachable != (unsigned long) as);
 620                 b->will_make_reachable = 0;
 621
 622                 /*
 623                  * b->will_make_reachable prevented it from being written, so
 624                  * write it now if it needs to be written:
 625                  */
 626                 btree_node_lock_type(c, b, SIX_LOCK_read);
 627                 bch2_btree_node_write_cond(c, b, btree_node_need_write(b));
 628                 six_unlock_read(&b->lock);
 629         }
 630
 631         while (as->nr_pending)
 632                 bch2_btree_node_free_ondisk(c, &as->pending[--as->nr_pending],
 633                                             seq);
 634 }
 635
 636 static void btree_update_nodes_written(struct closure *cl)
 637 {
 638         struct btree_update *as = container_of(cl, struct btree_update, cl);
 639         struct journal_res res = { 0 };
 640         struct bch_fs *c = as->c;
 641         struct btree *b;
 642         struct bset *i;
 643         int ret;
 644
 645         /*
 646          * We did an update to a parent node where the pointers we added pointed
 647          * to child nodes that weren't written yet: now, the child nodes have
 648          * been written so we can write out the update to the interior node.
 649          */
 650         mutex_lock(&c->btree_interior_update_lock);
 651         as->nodes_written = true;
 652 again:
 653         as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
 654                                       struct btree_update, unwritten_list);
 655         if (!as || !as->nodes_written) {
 656                 mutex_unlock(&c->btree_interior_update_lock);
 657                 return;
 658         }
 659
 660         b = as->b;
 661         if (b && !six_trylock_intent(&b->lock)) {
 662                 mutex_unlock(&c->btree_interior_update_lock);
 663                 btree_node_lock_type(c, b, SIX_LOCK_intent);
 664                 six_unlock_intent(&b->lock);
 665                 mutex_lock(&c->btree_interior_update_lock);
 666                 goto again;
 667         }
 668
 669         list_del(&as->unwritten_list);
 670
 671         ret = bch2_journal_res_get(&c->journal, &res, as->journal_u64s,
 672                                    JOURNAL_RES_GET_RESERVED);
 673         if (ret) {
 674                 BUG_ON(!bch2_journal_error(&c->journal));
 675                 /* can't unblock btree writes */
 676                 goto free_update;
 677         }
 678
 679         {
 680                 struct journal_buf *buf = &c->journal.buf[res.idx];
 681                 struct jset_entry *entry = vstruct_idx(buf->data, res.offset);
 682
 683                 res.offset      += as->journal_u64s;
 684                 res.u64s        -= as->journal_u64s;
 685                 memcpy_u64s(entry, as->journal_entries, as->journal_u64s);
 686         }
 687
 688         switch (as->mode) {
 689         case BTREE_INTERIOR_NO_UPDATE:
 690                 BUG();
 691         case BTREE_INTERIOR_UPDATING_NODE:
 692                 /* @b is the node we did the final insert into: */
 693                 BUG_ON(!res.ref);
 694
 695                 six_lock_write(&b->lock);
 696                 list_del(&as->write_blocked_list);
 697
 698                 i = btree_bset_last(b);
 699                 i->journal_seq = cpu_to_le64(
 700                         max(res.seq,
 701                             le64_to_cpu(i->journal_seq)));
 702
 703                 bch2_btree_add_journal_pin(c, b, res.seq);
 704                 six_unlock_write(&b->lock);
 705                 break;
 706
 707         case BTREE_INTERIOR_UPDATING_AS:
 708                 BUG_ON(b);
 709                 break;
 710
 711         case BTREE_INTERIOR_UPDATING_ROOT: {
 712                 struct btree_root *r = &c->btree_roots[as->btree_id];
 713
 714                 BUG_ON(b);
 715
 716                 mutex_lock(&c->btree_root_lock);
 717                 bkey_copy(&r->key, as->parent_keys.keys);
 718                 r->level = as->level;
 719                 r->alive = true;
 720                 c->btree_roots_dirty = true;
 721                 mutex_unlock(&c->btree_root_lock);
 722                 break;
 723         }
 724         }
 725
 726         bch2_journal_pin_drop(&c->journal, &as->journal);
 727
 728         bch2_journal_res_put(&c->journal, &res);
 729         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 730 free_update:
 731         /* Do btree write after dropping journal res: */
 732         if (b) {
 733                 /*
 734                  * b->write_blocked prevented it from being written, so
 735                  * write it now if it needs to be written:
 736                  */
 737                 btree_node_write_if_need(c, b, SIX_LOCK_intent);
 738                 six_unlock_intent(&b->lock);
 739         }
 740
 741         if (!ret)
 742                 btree_update_nodes_reachable(as, res.seq);
 743
 744         __bch2_btree_update_free(as);
 745         /*
 746          * for flush_held_btree_writes() waiting on updates to flush or
 747          * nodes to be writeable:
 748          */
 749         closure_wake_up(&c->btree_interior_update_wait);
 750         goto again;
 751 }
 752
 753 /*
 754  * We're updating @b with pointers to nodes that haven't finished writing yet:
 755  * block @b from being written until @as completes
 756  */
 757 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
 758 {
 759         struct bch_fs *c = as->c;
 760
 761         mutex_lock(&c->btree_interior_update_lock);
 762         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 763
 764         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 765         BUG_ON(!btree_node_dirty(b));
 766
 767         as->mode        = BTREE_INTERIOR_UPDATING_NODE;
 768         as->b           = b;
 769         as->level       = b->level;
 770         list_add(&as->write_blocked_list, &b->write_blocked);
 771
 772         mutex_unlock(&c->btree_interior_update_lock);
 773 }
 774
 775 static void btree_update_reparent(struct btree_update *as,
 776                                   struct btree_update *child)
 777 {
 778         struct bch_fs *c = as->c;
 779
 780         lockdep_assert_held(&c->btree_interior_update_lock);
 781
 782         child->b = NULL;
 783         child->mode = BTREE_INTERIOR_UPDATING_AS;
 784
 785         /*
 786          * When we write a new btree root, we have to drop our journal pin
 787          * _before_ the new nodes are technically reachable; see
 788          * btree_update_nodes_written().
 789          *
 790          * This goes for journal pins that are recursively blocked on us - so,
 791          * just transfer the journal pin to the new interior update so
 792          * btree_update_nodes_written() can drop it.
 793          */
 794         bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL);
 795         bch2_journal_pin_drop(&c->journal, &child->journal);
 796 }
 797
 798 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
 799 {
 800         struct bch_fs *c = as->c;
 801
 802         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 803         BUG_ON(!bch2_keylist_empty(&as->parent_keys));
 804
 805         mutex_lock(&c->btree_interior_update_lock);
 806         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 807
 808         as->mode        = BTREE_INTERIOR_UPDATING_ROOT;
 809         as->level       = b->level;
 810         bch2_keylist_add(&as->parent_keys, &b->key);
 811         mutex_unlock(&c->btree_interior_update_lock);
 812 }
 813
 814 static void btree_node_will_make_reachable(struct btree_update *as,
 815                                            struct btree *b)
 816 {
 817         struct bch_fs *c = as->c;
 818
 819         mutex_lock(&c->btree_interior_update_lock);
 820         BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
 821         BUG_ON(b->will_make_reachable);
 822
 823         as->new_nodes[as->nr_new_nodes++] = b;
 824         b->will_make_reachable = 1UL|(unsigned long) as;
 825
 826         closure_get(&as->cl);
 827         mutex_unlock(&c->btree_interior_update_lock);
 828 }
 829
 830 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
 831 {
 832         struct btree_update *as;
 833         unsigned long v;
 834         unsigned i;
 835
 836         mutex_lock(&c->btree_interior_update_lock);
 837         v = xchg(&b->will_make_reachable, 0);
 838         as = (struct btree_update *) (v & ~1UL);
 839
 840         if (!as) {
 841                 mutex_unlock(&c->btree_interior_update_lock);
 842                 return;
 843         }
 844
 845         for (i = 0; i < as->nr_new_nodes; i++)
 846                 if (as->new_nodes[i] == b)
 847                         goto found;
 848
 849         BUG();
 850 found:
 851         array_remove_item(as->new_nodes, as->nr_new_nodes, i);
 852         mutex_unlock(&c->btree_interior_update_lock);
 853
 854         if (v & 1)
 855                 closure_put(&as->cl);
 856 }
 857
 858 static void btree_interior_update_add_node_reference(struct btree_update *as,
 859                                                      struct btree *b)
 860 {
 861         struct bch_fs *c = as->c;
 862         struct pending_btree_node_free *d;
 863
 864         mutex_lock(&c->btree_interior_update_lock);
 865
 866         /* Add this node to the list of nodes being freed: */
 867         BUG_ON(as->nr_pending >= ARRAY_SIZE(as->pending));
 868
 869         d = &as->pending[as->nr_pending++];
 870         d->index_update_done    = false;
 871         d->seq                  = b->data->keys.seq;
 872         d->btree_id             = b->btree_id;
 873         d->level                = b->level;
 874         bkey_copy(&d->key, &b->key);
 875
 876         mutex_unlock(&c->btree_interior_update_lock);
 877 }
 878
 879 /*
 880  * @b is being split/rewritten: it may have pointers to not-yet-written btree
 881  * nodes and thus outstanding btree_updates - redirect @b's
 882  * btree_updates to point to this btree_update:
 883  */
 884 void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 885                                                struct btree *b)
 886 {
 887         struct bch_fs *c = as->c;
 888         struct btree_update *p, *n;
 889         struct btree_write *w;
 890
 891         set_btree_node_dying(b);
 892
 893         if (btree_node_fake(b))
 894                 return;
 895
 896         btree_interior_update_add_node_reference(as, b);
 897
 898         mutex_lock(&c->btree_interior_update_lock);
 899
 900         /*
 901          * Does this node have any btree_update operations preventing
 902          * it from being written?
 903          *
 904          * If so, redirect them to point to this btree_update: we can
 905          * write out our new nodes, but we won't make them visible until those
 906          * operations complete
 907          */
 908         list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
 909                 list_del(&p->write_blocked_list);
 910                 btree_update_reparent(as, p);
 911
 912                 /*
 913                  * for flush_held_btree_writes() waiting on updates to flush or
 914                  * nodes to be writeable:
 915                  */
 916                 closure_wake_up(&c->btree_interior_update_wait);
 917         }
 918
 919         clear_btree_node_dirty(b);
 920         clear_btree_node_need_write(b);
 921
 922         /*
 923          * Does this node have unwritten data that has a pin on the journal?
 924          *
 925          * If so, transfer that pin to the btree_update operation -
 926          * note that if we're freeing multiple nodes, we only need to keep the
 927          * oldest pin of any of the nodes we're freeing. We'll release the pin
 928          * when the new nodes are persistent and reachable on disk:
 929          */
 930         w = btree_current_write(b);
 931         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 932         bch2_journal_pin_drop(&c->journal, &w->journal);
 933
 934         w = btree_prev_write(b);
 935         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 936         bch2_journal_pin_drop(&c->journal, &w->journal);
 937
 938         mutex_unlock(&c->btree_interior_update_lock);
 939 }
 940
 941 void bch2_btree_update_done(struct btree_update *as)
 942 {
 943         BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
 944
 945         bch2_btree_reserve_put(as->c, as->reserve);
 946         as->reserve = NULL;
 947
 948         continue_at(&as->cl, btree_update_nodes_written, system_freezable_wq);
 949 }
 950
 951 struct btree_update *
 952 bch2_btree_update_start(struct bch_fs *c, enum btree_id id,
 953                         unsigned nr_nodes, unsigned flags,
 954                         struct closure *cl)
 955 {
 956         struct btree_reserve *reserve;
 957         struct btree_update *as;
 958         int ret;
 959
 960         reserve = bch2_btree_reserve_get(c, nr_nodes, flags, cl);
 961         if (IS_ERR(reserve))
 962                 return ERR_CAST(reserve);
 963
 964         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
 965         memset(as, 0, sizeof(*as));
 966         closure_init(&as->cl, NULL);
 967         as->c           = c;
 968         as->mode        = BTREE_INTERIOR_NO_UPDATE;
 969         as->btree_id    = id;
 970         as->reserve     = reserve;
 971         INIT_LIST_HEAD(&as->write_blocked_list);
 972
 973         bch2_keylist_init(&as->parent_keys, as->inline_keys);
 974
 975         ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
 976                                       ARRAY_SIZE(as->journal_entries), 0);
 977         if (ret) {
 978                 bch2_btree_reserve_put(c, reserve);
 979                 closure_debug_destroy(&as->cl);
 980                 mempool_free(as, &c->btree_interior_update_pool);
 981                 return ERR_PTR(ret);
 982         }
 983
 984         mutex_lock(&c->btree_interior_update_lock);
 985         list_add_tail(&as->list, &c->btree_interior_update_list);
 986         mutex_unlock(&c->btree_interior_update_lock);
 987
 988         return as;
 989 }
 990
 991 /* Btree root updates: */
 992
 993 static void __bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
 994 {
 995         /* Root nodes cannot be reaped */
 996         mutex_lock(&c->btree_cache.lock);
 997         list_del_init(&b->list);
 998         mutex_unlock(&c->btree_cache.lock);
 999
1000         mutex_lock(&c->btree_root_lock);
1001         BUG_ON(btree_node_root(c, b) &&
1002                (b->level < btree_node_root(c, b)->level ||
1003                 !btree_node_dying(btree_node_root(c, b))));
1004
1005         btree_node_root(c, b) = b;
1006         mutex_unlock(&c->btree_root_lock);
1007
1008         bch2_recalc_btree_reserve(c);
1009 }
1010
1011 static void bch2_btree_set_root_inmem(struct btree_update *as, struct btree *b)
1012 {
1013         struct bch_fs *c = as->c;
1014         struct btree *old = btree_node_root(c, b);
1015         struct bch_fs_usage *fs_usage;
1016
1017         __bch2_btree_set_root_inmem(c, b);
1018
1019         mutex_lock(&c->btree_interior_update_lock);
1020         percpu_down_read(&c->mark_lock);
1021         fs_usage = bch2_fs_usage_scratch_get(c);
1022
1023         bch2_mark_key_locked(c, bkey_i_to_s_c(&b->key),
1024                       0, 0, fs_usage, 0,
1025                       BTREE_TRIGGER_INSERT);
1026         if (gc_visited(c, gc_pos_btree_root(b->btree_id)))
1027                 bch2_mark_key_locked(c, bkey_i_to_s_c(&b->key),
1028                                      0, 0, NULL, 0,
1029                                      BTREE_TRIGGER_INSERT|
1030                                      BTREE_TRIGGER_GC);
1031
1032         if (old && !btree_node_fake(old))
1033                 bch2_btree_node_free_index(as, NULL,
1034                                            bkey_i_to_s_c(&old->key),
1035                                            fs_usage);
1036         bch2_fs_usage_apply(c, fs_usage, &as->reserve->disk_res, 0);
1037
1038         bch2_fs_usage_scratch_put(c, fs_usage);
1039         percpu_up_read(&c->mark_lock);
1040         mutex_unlock(&c->btree_interior_update_lock);
1041 }
1042
1043 /**
1044  * bch_btree_set_root - update the root in memory and on disk
1045  *
1046  * To ensure forward progress, the current task must not be holding any
1047  * btree node write locks. However, you must hold an intent lock on the
1048  * old root.
1049  *
1050  * Note: This allocates a journal entry but doesn't add any keys to
1051  * it.  All the btree roots are part of every journal write, so there
1052  * is nothing new to be done.  This just guarantees that there is a
1053  * journal write.
1054  */
1055 static void bch2_btree_set_root(struct btree_update *as, struct btree *b,
1056                                 struct btree_iter *iter)
1057 {
1058         struct bch_fs *c = as->c;
1059         struct btree *old;
1060
1061         trace_btree_set_root(c, b);
1062         BUG_ON(!b->written &&
1063                !test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags));
1064
1065         old = btree_node_root(c, b);
1066
1067         /*
1068          * Ensure no one is using the old root while we switch to the
1069          * new root:
1070          */
1071         bch2_btree_node_lock_write(old, iter);
1072
1073         bch2_btree_set_root_inmem(as, b);
1074
1075         btree_update_updated_root(as, b);
1076
1077         /*
1078          * Unlock old root after new root is visible:
1079          *
1080          * The new root isn't persistent, but that's ok: we still have
1081          * an intent lock on the new root, and any updates that would
1082          * depend on the new root would have to update the new root.
1083          */
1084         bch2_btree_node_unlock_write(old, iter);
1085 }
1086
1087 /* Interior node updates: */
1088
1089 static void bch2_insert_fixup_btree_ptr(struct btree_update *as, struct btree *b,
1090                                         struct btree_iter *iter,
1091                                         struct bkey_i *insert,
1092                                         struct btree_node_iter *node_iter)
1093 {
1094         struct bch_fs *c = as->c;
1095         struct bch_fs_usage *fs_usage;
1096         struct jset_entry *entry;
1097         struct bkey_packed *k;
1098         struct bkey tmp;
1099
1100         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1101                ARRAY_SIZE(as->journal_entries));
1102
1103         entry = (void *) &as->journal_entries[as->journal_u64s];
1104         memset(entry, 0, sizeof(*entry));
1105         entry->u64s     = cpu_to_le16(insert->k.u64s);
1106         entry->type     = BCH_JSET_ENTRY_btree_keys;
1107         entry->btree_id = b->btree_id;
1108         entry->level    = b->level;
1109         memcpy_u64s_small(entry->_data, insert, insert->k.u64s);
1110         as->journal_u64s += jset_u64s(insert->k.u64s);
1111
1112         mutex_lock(&c->btree_interior_update_lock);
1113         percpu_down_read(&c->mark_lock);
1114         fs_usage = bch2_fs_usage_scratch_get(c);
1115
1116         bch2_mark_key_locked(c, bkey_i_to_s_c(insert),
1117                              0, 0, fs_usage, 0,
1118                              BTREE_TRIGGER_INSERT);
1119
1120         if (gc_visited(c, gc_pos_btree_node(b)))
1121                 bch2_mark_key_locked(c, bkey_i_to_s_c(insert),
1122                                      0, 0, NULL, 0,
1123                                      BTREE_TRIGGER_INSERT|
1124                                      BTREE_TRIGGER_GC);
1125
1126         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1127                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1128                 bch2_btree_node_iter_advance(node_iter, b);
1129
1130         /*
1131          * If we're overwriting, look up pending delete and mark so that gc
1132          * marks it on the pending delete list:
1133          */
1134         if (k && !bkey_cmp_packed(b, k, &insert->k))
1135                 bch2_btree_node_free_index(as, b,
1136                                            bkey_disassemble(b, k, &tmp),
1137                                            fs_usage);
1138
1139         bch2_fs_usage_apply(c, fs_usage, &as->reserve->disk_res, 0);
1140
1141         bch2_fs_usage_scratch_put(c, fs_usage);
1142         percpu_up_read(&c->mark_lock);
1143         mutex_unlock(&c->btree_interior_update_lock);
1144
1145         bch2_btree_bset_insert_key(iter, b, node_iter, insert);
1146         set_btree_node_dirty(b);
1147         set_btree_node_need_write(b);
1148 }
1149
1150 /*
1151  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1152  * node)
1153  */
1154 static struct btree *__btree_split_node(struct btree_update *as,
1155                                         struct btree *n1,
1156                                         struct btree_iter *iter)
1157 {
1158         size_t nr_packed = 0, nr_unpacked = 0;
1159         struct btree *n2;
1160         struct bset *set1, *set2;
1161         struct bkey_packed *k, *prev = NULL;
1162
1163         n2 = bch2_btree_node_alloc(as, n1->level);
1164
1165         n2->data->max_key       = n1->data->max_key;
1166         n2->data->format        = n1->format;
1167         SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data));
1168         n2->key.k.p = n1->key.k.p;
1169
1170         btree_node_set_format(n2, n2->data->format);
1171
1172         set1 = btree_bset_first(n1);
1173         set2 = btree_bset_first(n2);
1174
1175         /*
1176          * Has to be a linear search because we don't have an auxiliary
1177          * search tree yet
1178          */
1179         k = set1->start;
1180         while (1) {
1181                 struct bkey_packed *n = bkey_next_skip_noops(k, vstruct_last(set1));
1182
1183                 if (n == vstruct_last(set1))
1184                         break;
1185                 if (k->_data - set1->_data >= (le16_to_cpu(set1->u64s) * 3) / 5)
1186                         break;
1187
1188                 if (bkey_packed(k))
1189                         nr_packed++;
1190                 else
1191                         nr_unpacked++;
1192
1193                 prev = k;
1194                 k = n;
1195         }
1196
1197         BUG_ON(!prev);
1198
1199         btree_set_max(n1, bkey_unpack_pos(n1, prev));
1200         btree_set_min(n2, bkey_successor(n1->key.k.p));
1201
1202         set2->u64s = cpu_to_le16((u64 *) vstruct_end(set1) - (u64 *) k);
1203         set1->u64s = cpu_to_le16(le16_to_cpu(set1->u64s) - le16_to_cpu(set2->u64s));
1204
1205         set_btree_bset_end(n1, n1->set);
1206         set_btree_bset_end(n2, n2->set);
1207
1208         n2->nr.live_u64s        = le16_to_cpu(set2->u64s);
1209         n2->nr.bset_u64s[0]     = le16_to_cpu(set2->u64s);
1210         n2->nr.packed_keys      = n1->nr.packed_keys - nr_packed;
1211         n2->nr.unpacked_keys    = n1->nr.unpacked_keys - nr_unpacked;
1212
1213         n1->nr.live_u64s        = le16_to_cpu(set1->u64s);
1214         n1->nr.bset_u64s[0]     = le16_to_cpu(set1->u64s);
1215         n1->nr.packed_keys      = nr_packed;
1216         n1->nr.unpacked_keys    = nr_unpacked;
1217
1218         BUG_ON(!set1->u64s);
1219         BUG_ON(!set2->u64s);
1220
1221         memcpy_u64s(set2->start,
1222                     vstruct_end(set1),
1223                     le16_to_cpu(set2->u64s));
1224
1225         btree_node_reset_sib_u64s(n1);
1226         btree_node_reset_sib_u64s(n2);
1227
1228         bch2_verify_btree_nr_keys(n1);
1229         bch2_verify_btree_nr_keys(n2);
1230
1231         if (n1->level) {
1232                 btree_node_interior_verify(n1);
1233                 btree_node_interior_verify(n2);
1234         }
1235
1236         return n2;
1237 }
1238
1239 /*
1240  * For updates to interior nodes, we've got to do the insert before we split
1241  * because the stuff we're inserting has to be inserted atomically. Post split,
1242  * the keys might have to go in different nodes and the split would no longer be
1243  * atomic.
1244  *
1245  * Worse, if the insert is from btree node coalescing, if we do the insert after
1246  * we do the split (and pick the pivot) - the pivot we pick might be between
1247  * nodes that were coalesced, and thus in the middle of a child node post
1248  * coalescing:
1249  */
1250 static void btree_split_insert_keys(struct btree_update *as, struct btree *b,
1251                                     struct btree_iter *iter,
1252                                     struct keylist *keys)
1253 {
1254         struct btree_node_iter node_iter;
1255         struct bkey_i *k = bch2_keylist_front(keys);
1256         struct bkey_packed *src, *dst, *n;
1257         struct bset *i;
1258
1259         /*
1260          * XXX
1261          *
1262          * these updates must be journalled
1263          *
1264          * oops
1265          */
1266
1267         BUG_ON(btree_node_type(b) != BKEY_TYPE_BTREE);
1268
1269         bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
1270
1271         while (!bch2_keylist_empty(keys)) {
1272                 k = bch2_keylist_front(keys);
1273
1274                 bch2_insert_fixup_btree_ptr(as, b, iter, k, &node_iter);
1275                 bch2_keylist_pop_front(keys);
1276         }
1277
1278         /*
1279          * We can't tolerate whiteouts here - with whiteouts there can be
1280          * duplicate keys, and it would be rather bad if we picked a duplicate
1281          * for the pivot:
1282          */
1283         i = btree_bset_first(b);
1284         src = dst = i->start;
1285         while (src != vstruct_last(i)) {
1286                 n = bkey_next_skip_noops(src, vstruct_last(i));
1287                 if (!bkey_deleted(src)) {
1288                         memmove_u64s_down(dst, src, src->u64s);
1289                         dst = bkey_next(dst);
1290                 }
1291                 src = n;
1292         }
1293
1294         i->u64s = cpu_to_le16((u64 *) dst - i->_data);
1295         set_btree_bset_end(b, b->set);
1296
1297         BUG_ON(b->nsets != 1 ||
1298                b->nr.live_u64s != le16_to_cpu(btree_bset_first(b)->u64s));
1299
1300         btree_node_interior_verify(b);
1301 }
1302
1303 static void btree_split(struct btree_update *as, struct btree *b,
1304                         struct btree_iter *iter, struct keylist *keys,
1305                         unsigned flags)
1306 {
1307         struct bch_fs *c = as->c;
1308         struct btree *parent = btree_node_parent(iter, b);
1309         struct btree *n1, *n2 = NULL, *n3 = NULL;
1310         u64 start_time = local_clock();
1311
1312         BUG_ON(!parent && (b != btree_node_root(c, b)));
1313         BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->level));
1314
1315         bch2_btree_interior_update_will_free_node(as, b);
1316
1317         n1 = bch2_btree_node_alloc_replacement(as, b);
1318
1319         if (keys)
1320                 btree_split_insert_keys(as, n1, iter, keys);
1321
1322         if (bset_u64s(&n1->set[0]) > BTREE_SPLIT_THRESHOLD(c)) {
1323                 trace_btree_split(c, b);
1324
1325                 n2 = __btree_split_node(as, n1, iter);
1326
1327                 bch2_btree_build_aux_trees(n2);
1328                 bch2_btree_build_aux_trees(n1);
1329                 six_unlock_write(&n2->lock);
1330                 six_unlock_write(&n1->lock);
1331
1332                 bch2_btree_node_write(c, n2, SIX_LOCK_intent);
1333
1334                 /*
1335                  * Note that on recursive parent_keys == keys, so we
1336                  * can't start adding new keys to parent_keys before emptying it
1337                  * out (which we did with btree_split_insert_keys() above)
1338                  */
1339                 bch2_keylist_add(&as->parent_keys, &n1->key);
1340                 bch2_keylist_add(&as->parent_keys, &n2->key);
1341
1342                 if (!parent) {
1343                         /* Depth increases, make a new root */
1344                         n3 = __btree_root_alloc(as, b->level + 1);
1345
1346                         n3->sib_u64s[0] = U16_MAX;
1347                         n3->sib_u64s[1] = U16_MAX;
1348
1349                         btree_split_insert_keys(as, n3, iter, &as->parent_keys);
1350
1351                         bch2_btree_node_write(c, n3, SIX_LOCK_intent);
1352                 }
1353         } else {
1354                 trace_btree_compact(c, b);
1355
1356                 bch2_btree_build_aux_trees(n1);
1357                 six_unlock_write(&n1->lock);
1358
1359                 if (parent)
1360                         bch2_keylist_add(&as->parent_keys, &n1->key);
1361         }
1362
1363         bch2_btree_node_write(c, n1, SIX_LOCK_intent);
1364
1365         /* New nodes all written, now make them visible: */
1366
1367         if (parent) {
1368                 /* Split a non root node */
1369                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1370         } else if (n3) {
1371                 bch2_btree_set_root(as, n3, iter);
1372         } else {
1373                 /* Root filled up but didn't need to be split */
1374                 bch2_btree_set_root(as, n1, iter);
1375         }
1376
1377         bch2_open_buckets_put(c, &n1->ob);
1378         if (n2)
1379                 bch2_open_buckets_put(c, &n2->ob);
1380         if (n3)
1381                 bch2_open_buckets_put(c, &n3->ob);
1382
1383         /* Successful split, update the iterator to point to the new nodes: */
1384
1385         six_lock_increment(&b->lock, SIX_LOCK_intent);
1386         bch2_btree_iter_node_drop(iter, b);
1387         if (n3)
1388                 bch2_btree_iter_node_replace(iter, n3);
1389         if (n2)
1390                 bch2_btree_iter_node_replace(iter, n2);
1391         bch2_btree_iter_node_replace(iter, n1);
1392
1393         /*
1394          * The old node must be freed (in memory) _before_ unlocking the new
1395          * nodes - else another thread could re-acquire a read lock on the old
1396          * node after another thread has locked and updated the new node, thus
1397          * seeing stale data:
1398          */
1399         bch2_btree_node_free_inmem(c, b, iter);
1400
1401         if (n3)
1402                 six_unlock_intent(&n3->lock);
1403         if (n2)
1404                 six_unlock_intent(&n2->lock);
1405         six_unlock_intent(&n1->lock);
1406
1407         bch2_btree_trans_verify_locks(iter->trans);
1408
1409         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_split],
1410                                start_time);
1411 }
1412
1413 static void
1414 bch2_btree_insert_keys_interior(struct btree_update *as, struct btree *b,
1415                                 struct btree_iter *iter, struct keylist *keys)
1416 {
1417         struct btree_iter *linked;
1418         struct btree_node_iter node_iter;
1419         struct bkey_i *insert = bch2_keylist_front(keys);
1420         struct bkey_packed *k;
1421
1422         /* Don't screw up @iter's position: */
1423         node_iter = iter->l[b->level].iter;
1424
1425         /*
1426          * btree_split(), btree_gc_coalesce() will insert keys before
1427          * the iterator's current position - they know the keys go in
1428          * the node the iterator points to:
1429          */
1430         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1431                (bkey_cmp_packed(b, k, &insert->k) >= 0))
1432                 ;
1433
1434         for_each_keylist_key(keys, insert)
1435                 bch2_insert_fixup_btree_ptr(as, b, iter, insert, &node_iter);
1436
1437         btree_update_updated_node(as, b);
1438
1439         trans_for_each_iter_with_node(iter->trans, b, linked)
1440                 bch2_btree_node_iter_peek(&linked->l[b->level].iter, b);
1441
1442         bch2_btree_trans_verify_iters(iter->trans, b);
1443 }
1444
1445 /**
1446  * bch_btree_insert_node - insert bkeys into a given btree node
1447  *
1448  * @iter:               btree iterator
1449  * @keys:               list of keys to insert
1450  * @hook:               insert callback
1451  * @persistent:         if not null, @persistent will wait on journal write
1452  *
1453  * Inserts as many keys as it can into a given btree node, splitting it if full.
1454  * If a split occurred, this function will return early. This can only happen
1455  * for leaf nodes -- inserts into interior nodes have to be atomic.
1456  */
1457 void bch2_btree_insert_node(struct btree_update *as, struct btree *b,
1458                             struct btree_iter *iter, struct keylist *keys,
1459                             unsigned flags)
1460 {
1461         struct bch_fs *c = as->c;
1462         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1463         int old_live_u64s = b->nr.live_u64s;
1464         int live_u64s_added, u64s_added;
1465
1466         BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->level));
1467         BUG_ON(!b->level);
1468         BUG_ON(!as || as->b);
1469         bch2_verify_keylist_sorted(keys);
1470
1471         if (as->must_rewrite)
1472                 goto split;
1473
1474         bch2_btree_node_lock_for_insert(c, b, iter);
1475
1476         if (!bch2_btree_node_insert_fits(c, b, bch_keylist_u64s(keys))) {
1477                 bch2_btree_node_unlock_write(b, iter);
1478                 goto split;
1479         }
1480
1481         bch2_btree_insert_keys_interior(as, b, iter, keys);
1482
1483         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1484         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1485
1486         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1487                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1488         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1489                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1490
1491         if (u64s_added > live_u64s_added &&
1492             bch2_maybe_compact_whiteouts(c, b))
1493                 bch2_btree_iter_reinit_node(iter, b);
1494
1495         bch2_btree_node_unlock_write(b, iter);
1496
1497         btree_node_interior_verify(b);
1498
1499         /*
1500          * when called from the btree_split path the new nodes aren't added to
1501          * the btree iterator yet, so the merge path's unlock/wait/relock dance
1502          * won't work:
1503          */
1504         bch2_foreground_maybe_merge(c, iter, b->level,
1505                                     flags|BTREE_INSERT_NOUNLOCK);
1506         return;
1507 split:
1508         btree_split(as, b, iter, keys, flags);
1509 }
1510
1511 int bch2_btree_split_leaf(struct bch_fs *c, struct btree_iter *iter,
1512                           unsigned flags)
1513 {
1514         struct btree_trans *trans = iter->trans;
1515         struct btree *b = iter_l(iter)->b;
1516         struct btree_update *as;
1517         struct closure cl;
1518         int ret = 0;
1519         struct btree_iter *linked;
1520
1521         /*
1522          * We already have a disk reservation and open buckets pinned; this
1523          * allocation must not block:
1524          */
1525         trans_for_each_iter(trans, linked)
1526                 if (linked->btree_id == BTREE_ID_EXTENTS)
1527                         flags |= BTREE_INSERT_USE_RESERVE;
1528
1529         closure_init_stack(&cl);
1530
1531         /* Hack, because gc and splitting nodes doesn't mix yet: */
1532         if (!(flags & BTREE_INSERT_GC_LOCK_HELD) &&
1533             !down_read_trylock(&c->gc_lock)) {
1534                 if (flags & BTREE_INSERT_NOUNLOCK)
1535                         return -EINTR;
1536
1537                 bch2_trans_unlock(trans);
1538                 down_read(&c->gc_lock);
1539
1540                 if (!bch2_trans_relock(trans))
1541                         ret = -EINTR;
1542         }
1543
1544         /*
1545          * XXX: figure out how far we might need to split,
1546          * instead of locking/reserving all the way to the root:
1547          */
1548         if (!bch2_btree_iter_upgrade(iter, U8_MAX)) {
1549                 trace_trans_restart_iter_upgrade(trans->ip);
1550                 ret = -EINTR;
1551                 goto out;
1552         }
1553
1554         as = bch2_btree_update_start(c, iter->btree_id,
1555                 btree_update_reserve_required(c, b), flags,
1556                 !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1557         if (IS_ERR(as)) {
1558                 ret = PTR_ERR(as);
1559                 if (ret == -EAGAIN) {
1560                         BUG_ON(flags & BTREE_INSERT_NOUNLOCK);
1561                         bch2_trans_unlock(trans);
1562                         ret = -EINTR;
1563                 }
1564                 goto out;
1565         }
1566
1567         btree_split(as, b, iter, NULL, flags);
1568         bch2_btree_update_done(as);
1569
1570         /*
1571          * We haven't successfully inserted yet, so don't downgrade all the way
1572          * back to read locks;
1573          */
1574         __bch2_btree_iter_downgrade(iter, 1);
1575 out:
1576         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1577                 up_read(&c->gc_lock);
1578         closure_sync(&cl);
1579         return ret;
1580 }
1581
1582 void __bch2_foreground_maybe_merge(struct bch_fs *c,
1583                                    struct btree_iter *iter,
1584                                    unsigned level,
1585                                    unsigned flags,
1586                                    enum btree_node_sibling sib)
1587 {
1588         struct btree_trans *trans = iter->trans;
1589         struct btree_update *as;
1590         struct bkey_format_state new_s;
1591         struct bkey_format new_f;
1592         struct bkey_i delete;
1593         struct btree *b, *m, *n, *prev, *next, *parent;
1594         struct closure cl;
1595         size_t sib_u64s;
1596         int ret = 0;
1597
1598         BUG_ON(!btree_node_locked(iter, level));
1599
1600         closure_init_stack(&cl);
1601 retry:
1602         BUG_ON(!btree_node_locked(iter, level));
1603
1604         b = iter->l[level].b;
1605
1606         parent = btree_node_parent(iter, b);
1607         if (!parent)
1608                 goto out;
1609
1610         if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c))
1611                 goto out;
1612
1613         /* XXX: can't be holding read locks */
1614         m = bch2_btree_node_get_sibling(c, iter, b, sib);
1615         if (IS_ERR(m)) {
1616                 ret = PTR_ERR(m);
1617                 goto err;
1618         }
1619
1620         /* NULL means no sibling: */
1621         if (!m) {
1622                 b->sib_u64s[sib] = U16_MAX;
1623                 goto out;
1624         }
1625
1626         if (sib == btree_prev_sib) {
1627                 prev = m;
1628                 next = b;
1629         } else {
1630                 prev = b;
1631                 next = m;
1632         }
1633
1634         bch2_bkey_format_init(&new_s);
1635         __bch2_btree_calc_format(&new_s, b);
1636         __bch2_btree_calc_format(&new_s, m);
1637         new_f = bch2_bkey_format_done(&new_s);
1638
1639         sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1640                 btree_node_u64s_with_format(m, &new_f);
1641
1642         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1643                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1644                 sib_u64s /= 2;
1645                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1646         }
1647
1648         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1649         b->sib_u64s[sib] = sib_u64s;
1650
1651         if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c)) {
1652                 six_unlock_intent(&m->lock);
1653                 goto out;
1654         }
1655
1656         /* We're changing btree topology, doesn't mix with gc: */
1657         if (!(flags & BTREE_INSERT_GC_LOCK_HELD) &&
1658             !down_read_trylock(&c->gc_lock))
1659                 goto err_cycle_gc_lock;
1660
1661         if (!bch2_btree_iter_upgrade(iter, U8_MAX)) {
1662                 ret = -EINTR;
1663                 goto err_unlock;
1664         }
1665
1666         as = bch2_btree_update_start(c, iter->btree_id,
1667                          btree_update_reserve_required(c, parent) + 1,
1668                          BTREE_INSERT_NOFAIL|
1669                          BTREE_INSERT_USE_RESERVE,
1670                          !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1671         if (IS_ERR(as)) {
1672                 ret = PTR_ERR(as);
1673                 goto err_unlock;
1674         }
1675
1676         trace_btree_merge(c, b);
1677
1678         bch2_btree_interior_update_will_free_node(as, b);
1679         bch2_btree_interior_update_will_free_node(as, m);
1680
1681         n = bch2_btree_node_alloc(as, b->level);
1682
1683         btree_set_min(n, prev->data->min_key);
1684         btree_set_max(n, next->data->max_key);
1685         n->data->format         = new_f;
1686
1687         btree_node_set_format(n, new_f);
1688
1689         bch2_btree_sort_into(c, n, prev);
1690         bch2_btree_sort_into(c, n, next);
1691
1692         bch2_btree_build_aux_trees(n);
1693         six_unlock_write(&n->lock);
1694
1695         bkey_init(&delete.k);
1696         delete.k.p = prev->key.k.p;
1697         bch2_keylist_add(&as->parent_keys, &delete);
1698         bch2_keylist_add(&as->parent_keys, &n->key);
1699
1700         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1701
1702         bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1703
1704         bch2_open_buckets_put(c, &n->ob);
1705
1706         six_lock_increment(&b->lock, SIX_LOCK_intent);
1707         bch2_btree_iter_node_drop(iter, b);
1708         bch2_btree_iter_node_drop(iter, m);
1709
1710         bch2_btree_iter_node_replace(iter, n);
1711
1712         bch2_btree_trans_verify_iters(trans, n);
1713
1714         bch2_btree_node_free_inmem(c, b, iter);
1715         bch2_btree_node_free_inmem(c, m, iter);
1716
1717         six_unlock_intent(&n->lock);
1718
1719         bch2_btree_update_done(as);
1720
1721         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1722                 up_read(&c->gc_lock);
1723 out:
1724         bch2_btree_trans_verify_locks(trans);
1725
1726         /*
1727          * Don't downgrade locks here: we're called after successful insert,
1728          * and the caller will downgrade locks after a successful insert
1729          * anyways (in case e.g. a split was required first)
1730          *
1731          * And we're also called when inserting into interior nodes in the
1732          * split path, and downgrading to read locks in there is potentially
1733          * confusing:
1734          */
1735         closure_sync(&cl);
1736         return;
1737
1738 err_cycle_gc_lock:
1739         six_unlock_intent(&m->lock);
1740
1741         if (flags & BTREE_INSERT_NOUNLOCK)
1742                 goto out;
1743
1744         bch2_trans_unlock(trans);
1745
1746         down_read(&c->gc_lock);
1747         up_read(&c->gc_lock);
1748         ret = -EINTR;
1749         goto err;
1750
1751 err_unlock:
1752         six_unlock_intent(&m->lock);
1753         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1754                 up_read(&c->gc_lock);
1755 err:
1756         BUG_ON(ret == -EAGAIN && (flags & BTREE_INSERT_NOUNLOCK));
1757
1758         if ((ret == -EAGAIN || ret == -EINTR) &&
1759             !(flags & BTREE_INSERT_NOUNLOCK)) {
1760                 bch2_trans_unlock(trans);
1761                 closure_sync(&cl);
1762                 ret = bch2_btree_iter_traverse(iter);
1763                 if (ret)
1764                         goto out;
1765
1766                 goto retry;
1767         }
1768
1769         goto out;
1770 }
1771
1772 static int __btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
1773                                 struct btree *b, unsigned flags,
1774                                 struct closure *cl)
1775 {
1776         struct btree *n, *parent = btree_node_parent(iter, b);
1777         struct btree_update *as;
1778
1779         as = bch2_btree_update_start(c, iter->btree_id,
1780                 (parent
1781                  ? btree_update_reserve_required(c, parent)
1782                  : 0) + 1,
1783                 flags, cl);
1784         if (IS_ERR(as)) {
1785                 trace_btree_gc_rewrite_node_fail(c, b);
1786                 return PTR_ERR(as);
1787         }
1788
1789         bch2_btree_interior_update_will_free_node(as, b);
1790
1791         n = bch2_btree_node_alloc_replacement(as, b);
1792
1793         bch2_btree_build_aux_trees(n);
1794         six_unlock_write(&n->lock);
1795
1796         trace_btree_gc_rewrite_node(c, b);
1797
1798         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1799
1800         if (parent) {
1801                 bch2_keylist_add(&as->parent_keys, &n->key);
1802                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1803         } else {
1804                 bch2_btree_set_root(as, n, iter);
1805         }
1806
1807         bch2_open_buckets_put(c, &n->ob);
1808
1809         six_lock_increment(&b->lock, SIX_LOCK_intent);
1810         bch2_btree_iter_node_drop(iter, b);
1811         bch2_btree_iter_node_replace(iter, n);
1812         bch2_btree_node_free_inmem(c, b, iter);
1813         six_unlock_intent(&n->lock);
1814
1815         bch2_btree_update_done(as);
1816         return 0;
1817 }
1818
1819 /**
1820  * bch_btree_node_rewrite - Rewrite/move a btree node
1821  *
1822  * Returns 0 on success, -EINTR or -EAGAIN on failure (i.e.
1823  * btree_check_reserve() has to wait)
1824  */
1825 int bch2_btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
1826                             __le64 seq, unsigned flags)
1827 {
1828         struct btree_trans *trans = iter->trans;
1829         struct closure cl;
1830         struct btree *b;
1831         int ret;
1832
1833         flags |= BTREE_INSERT_NOFAIL;
1834
1835         closure_init_stack(&cl);
1836
1837         bch2_btree_iter_upgrade(iter, U8_MAX);
1838
1839         if (!(flags & BTREE_INSERT_GC_LOCK_HELD)) {
1840                 if (!down_read_trylock(&c->gc_lock)) {
1841                         bch2_trans_unlock(trans);
1842                         down_read(&c->gc_lock);
1843                 }
1844         }
1845
1846         while (1) {
1847                 ret = bch2_btree_iter_traverse(iter);
1848                 if (ret)
1849                         break;
1850
1851                 b = bch2_btree_iter_peek_node(iter);
1852                 if (!b || b->data->keys.seq != seq)
1853                         break;
1854
1855                 ret = __btree_node_rewrite(c, iter, b, flags, &cl);
1856                 if (ret != -EAGAIN &&
1857                     ret != -EINTR)
1858                         break;
1859
1860                 bch2_trans_unlock(trans);
1861                 closure_sync(&cl);
1862         }
1863
1864         bch2_btree_iter_downgrade(iter);
1865
1866         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1867                 up_read(&c->gc_lock);
1868
1869         closure_sync(&cl);
1870         return ret;
1871 }
1872
1873 static void __bch2_btree_node_update_key(struct bch_fs *c,
1874                                          struct btree_update *as,
1875                                          struct btree_iter *iter,
1876                                          struct btree *b, struct btree *new_hash,
1877                                          struct bkey_i *new_key)
1878 {
1879         struct btree *parent;
1880         int ret;
1881
1882         /*
1883          * Two corner cases that need to be thought about here:
1884          *
1885          * @b may not be reachable yet - there might be another interior update
1886          * operation waiting on @b to be written, and we're gonna deliver the
1887          * write completion to that interior update operation _before_
1888          * persisting the new_key update
1889          *
1890          * That ends up working without us having to do anything special here:
1891          * the reason is, we do kick off (and do the in memory updates) for the
1892          * update for @new_key before we return, creating a new interior_update
1893          * operation here.
1894          *
1895          * The new interior update operation here will in effect override the
1896          * previous one. The previous one was going to terminate - make @b
1897          * reachable - in one of two ways:
1898          * - updating the btree root pointer
1899          *   In that case,
1900          *   no, this doesn't work. argh.
1901          */
1902
1903         if (b->will_make_reachable)
1904                 as->must_rewrite = true;
1905
1906         btree_interior_update_add_node_reference(as, b);
1907
1908         /*
1909          * XXX: the rest of the update path treats this like we're actually
1910          * inserting a new node and deleting the existing node, so the
1911          * reservation needs to include enough space for @b
1912          *
1913          * that is actually sketch as fuck though and I am surprised the code
1914          * seems to work like that, definitely need to go back and rework it
1915          * into something saner.
1916          *
1917          * (I think @b is just getting double counted until the btree update
1918          * finishes and "deletes" @b on disk)
1919          */
1920         ret = bch2_disk_reservation_add(c, &as->reserve->disk_res,
1921                         c->opts.btree_node_size *
1922                         bch2_bkey_nr_ptrs(bkey_i_to_s_c(new_key)),
1923                         BCH_DISK_RESERVATION_NOFAIL);
1924         BUG_ON(ret);
1925
1926         parent = btree_node_parent(iter, b);
1927         if (parent) {
1928                 if (new_hash) {
1929                         bkey_copy(&new_hash->key, new_key);
1930                         ret = bch2_btree_node_hash_insert(&c->btree_cache,
1931                                         new_hash, b->level, b->btree_id);
1932                         BUG_ON(ret);
1933                 }
1934
1935                 bch2_keylist_add(&as->parent_keys, new_key);
1936                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, 0);
1937
1938                 if (new_hash) {
1939                         mutex_lock(&c->btree_cache.lock);
1940                         bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
1941
1942                         bch2_btree_node_hash_remove(&c->btree_cache, b);
1943
1944                         bkey_copy(&b->key, new_key);
1945                         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1946                         BUG_ON(ret);
1947                         mutex_unlock(&c->btree_cache.lock);
1948                 } else {
1949                         bkey_copy(&b->key, new_key);
1950                 }
1951         } else {
1952                 struct bch_fs_usage *fs_usage;
1953
1954                 BUG_ON(btree_node_root(c, b) != b);
1955
1956                 bch2_btree_node_lock_write(b, iter);
1957
1958                 mutex_lock(&c->btree_interior_update_lock);
1959                 percpu_down_read(&c->mark_lock);
1960                 fs_usage = bch2_fs_usage_scratch_get(c);
1961
1962                 bch2_mark_key_locked(c, bkey_i_to_s_c(new_key),
1963                               0, 0, fs_usage, 0,
1964                               BTREE_TRIGGER_INSERT);
1965                 if (gc_visited(c, gc_pos_btree_root(b->btree_id)))
1966                         bch2_mark_key_locked(c, bkey_i_to_s_c(new_key),
1967                                              0, 0, NULL, 0,
1968                                              BTREE_TRIGGER_INSERT||
1969                                              BTREE_TRIGGER_GC);
1970
1971                 bch2_btree_node_free_index(as, NULL,
1972                                            bkey_i_to_s_c(&b->key),
1973                                            fs_usage);
1974                 bch2_fs_usage_apply(c, fs_usage, &as->reserve->disk_res, 0);
1975
1976                 bch2_fs_usage_scratch_put(c, fs_usage);
1977                 percpu_up_read(&c->mark_lock);
1978                 mutex_unlock(&c->btree_interior_update_lock);
1979
1980                 if (btree_ptr_hash_val(new_key) != b->hash_val) {
1981                         mutex_lock(&c->btree_cache.lock);
1982                         bch2_btree_node_hash_remove(&c->btree_cache, b);
1983
1984                         bkey_copy(&b->key, new_key);
1985                         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1986                         BUG_ON(ret);
1987                         mutex_unlock(&c->btree_cache.lock);
1988                 } else {
1989                         bkey_copy(&b->key, new_key);
1990                 }
1991
1992                 btree_update_updated_root(as, b);
1993                 bch2_btree_node_unlock_write(b, iter);
1994         }
1995
1996         bch2_btree_update_done(as);
1997 }
1998
1999 int bch2_btree_node_update_key(struct bch_fs *c, struct btree_iter *iter,
2000                                struct btree *b,
2001                                struct bkey_i *new_key)
2002 {
2003         struct btree *parent = btree_node_parent(iter, b);
2004         struct btree_update *as = NULL;
2005         struct btree *new_hash = NULL;
2006         struct closure cl;
2007         int ret;
2008
2009         closure_init_stack(&cl);
2010
2011         if (!bch2_btree_iter_upgrade(iter, U8_MAX))
2012                 return -EINTR;
2013
2014         if (!down_read_trylock(&c->gc_lock)) {
2015                 bch2_trans_unlock(iter->trans);
2016                 down_read(&c->gc_lock);
2017
2018                 if (!bch2_trans_relock(iter->trans)) {
2019                         ret = -EINTR;
2020                         goto err;
2021                 }
2022         }
2023
2024         /*
2025          * check btree_ptr_hash_val() after @b is locked by
2026          * btree_iter_traverse():
2027          */
2028         if (btree_ptr_hash_val(new_key) != b->hash_val) {
2029                 /* bch2_btree_reserve_get will unlock */
2030                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2031                 if (ret) {
2032                         bch2_trans_unlock(iter->trans);
2033                         up_read(&c->gc_lock);
2034                         closure_sync(&cl);
2035                         down_read(&c->gc_lock);
2036
2037                         if (!bch2_trans_relock(iter->trans)) {
2038                                 ret = -EINTR;
2039                                 goto err;
2040                         }
2041                 }
2042
2043                 new_hash = bch2_btree_node_mem_alloc(c);
2044         }
2045
2046         as = bch2_btree_update_start(c, iter->btree_id,
2047                 parent ? btree_update_reserve_required(c, parent) : 0,
2048                 BTREE_INSERT_NOFAIL|
2049                 BTREE_INSERT_USE_RESERVE|
2050                 BTREE_INSERT_USE_ALLOC_RESERVE,
2051                 &cl);
2052
2053         if (IS_ERR(as)) {
2054                 ret = PTR_ERR(as);
2055                 if (ret == -EAGAIN)
2056                         ret = -EINTR;
2057
2058                 if (ret != -EINTR)
2059                         goto err;
2060
2061                 bch2_trans_unlock(iter->trans);
2062                 up_read(&c->gc_lock);
2063                 closure_sync(&cl);
2064                 down_read(&c->gc_lock);
2065
2066                 if (!bch2_trans_relock(iter->trans))
2067                         goto err;
2068         }
2069
2070         ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(new_key));
2071         if (ret)
2072                 goto err_free_update;
2073
2074         __bch2_btree_node_update_key(c, as, iter, b, new_hash, new_key);
2075
2076         bch2_btree_iter_downgrade(iter);
2077 err:
2078         if (new_hash) {
2079                 mutex_lock(&c->btree_cache.lock);
2080                 list_move(&new_hash->list, &c->btree_cache.freeable);
2081                 mutex_unlock(&c->btree_cache.lock);
2082
2083                 six_unlock_write(&new_hash->lock);
2084                 six_unlock_intent(&new_hash->lock);
2085         }
2086         up_read(&c->gc_lock);
2087         closure_sync(&cl);
2088         return ret;
2089 err_free_update:
2090         bch2_btree_update_free(as);
2091         goto err;
2092 }
2093
2094 /* Init code: */
2095
2096 /*
2097  * Only for filesystem bringup, when first reading the btree roots or allocating
2098  * btree roots when initializing a new filesystem:
2099  */
2100 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2101 {
2102         BUG_ON(btree_node_root(c, b));
2103
2104         __bch2_btree_set_root_inmem(c, b);
2105 }
2106
2107 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2108 {
2109         struct closure cl;
2110         struct btree *b;
2111         int ret;
2112
2113         closure_init_stack(&cl);
2114
2115         do {
2116                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2117                 closure_sync(&cl);
2118         } while (ret);
2119
2120         b = bch2_btree_node_mem_alloc(c);
2121         bch2_btree_cache_cannibalize_unlock(c);
2122
2123         set_btree_node_fake(b);
2124         b->level        = 0;
2125         b->btree_id     = id;
2126
2127         bkey_btree_ptr_init(&b->key);
2128         b->key.k.p = POS_MAX;
2129         *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2130
2131         bch2_bset_init_first(b, &b->data->keys);
2132         bch2_btree_build_aux_trees(b);
2133
2134         b->data->flags = 0;
2135         btree_set_min(b, POS_MIN);
2136         btree_set_max(b, POS_MAX);
2137         b->data->format = bch2_btree_calc_format(b);
2138         btree_node_set_format(b, b->data->format);
2139
2140         ret = bch2_btree_node_hash_insert(&c->btree_cache, b, b->level, b->btree_id);
2141         BUG_ON(ret);
2142
2143         __bch2_btree_set_root_inmem(c, b);
2144
2145         six_unlock_write(&b->lock);
2146         six_unlock_intent(&b->lock);
2147 }
2148
2149 ssize_t bch2_btree_updates_print(struct bch_fs *c, char *buf)
2150 {
2151         struct printbuf out = _PBUF(buf, PAGE_SIZE);
2152         struct btree_update *as;
2153
2154         mutex_lock(&c->btree_interior_update_lock);
2155         list_for_each_entry(as, &c->btree_interior_update_list, list)
2156                 pr_buf(&out, "%p m %u w %u r %u j %llu\n",
2157                        as,
2158                        as->mode,
2159                        as->nodes_written,
2160                        atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2161                        as->journal.seq);
2162         mutex_unlock(&c->btree_interior_update_lock);
2163
2164         return out.pos - buf;
2165 }
2166
2167 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *c)
2168 {
2169         size_t ret = 0;
2170         struct list_head *i;
2171
2172         mutex_lock(&c->btree_interior_update_lock);
2173         list_for_each(i, &c->btree_interior_update_list)
2174                 ret++;
2175         mutex_unlock(&c->btree_interior_update_lock);
2176
2177         return ret;
2178 }