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
 591         if (as->reserve)
 592                 bch2_btree_reserve_put(c, as->reserve);
 593
 594         list_del(&as->unwritten_list);
 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 inline bool six_trylock_intentwrite(struct six_lock *lock)
 613 {
 614         if (!six_trylock_intent(lock))
 615                 return false;
 616
 617         if (!six_trylock_write(lock)) {
 618                 six_unlock_intent(lock);
 619                 return false;
 620         }
 621
 622         return true;
 623 }
 624
 625 static void btree_update_nodes_written(struct closure *cl)
 626 {
 627         struct btree_update *as = container_of(cl, struct btree_update, cl);
 628         struct btree *nodes_need_write[BTREE_MAX_DEPTH * 2 + GC_MERGE_NODES + 1];
 629         unsigned nr_nodes_need_write;
 630         struct journal_res res = { 0 };
 631         struct bch_fs *c = as->c;
 632         struct btree_root *r;
 633         struct btree *b;
 634         int ret;
 635
 636         /*
 637          * We did an update to a parent node where the pointers we added pointed
 638          * to child nodes that weren't written yet: now, the child nodes have
 639          * been written so we can write out the update to the interior node.
 640          */
 641         mutex_lock(&c->btree_interior_update_lock);
 642         as->nodes_written = true;
 643 again:
 644         nr_nodes_need_write = 0;
 645         as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
 646                                       struct btree_update, unwritten_list);
 647         if (!as || !as->nodes_written) {
 648                 mutex_unlock(&c->btree_interior_update_lock);
 649                 return;
 650         }
 651
 652         b = as->b;
 653         if (b && !six_trylock_intentwrite(&b->lock)) {
 654                 mutex_unlock(&c->btree_interior_update_lock);
 655
 656                 btree_node_lock_type(c, b, SIX_LOCK_intent);
 657                 six_lock_write(&b->lock);
 658
 659                 six_unlock_write(&b->lock);
 660                 six_unlock_intent(&b->lock);
 661
 662                 mutex_lock(&c->btree_interior_update_lock);
 663                 goto again;
 664         }
 665
 666         ret = bch2_journal_res_get(&c->journal, &res, as->journal_u64s,
 667                                    JOURNAL_RES_GET_NONBLOCK|
 668                                    JOURNAL_RES_GET_RESERVED);
 669         if (ret == -EAGAIN) {
 670                 unsigned u64s = as->journal_u64s;
 671
 672                 if (b) {
 673                         six_unlock_write(&b->lock);
 674                         six_unlock_intent(&b->lock);
 675                 }
 676
 677                 mutex_unlock(&c->btree_interior_update_lock);
 678
 679                 ret = bch2_journal_res_get(&c->journal, &res, u64s,
 680                                            JOURNAL_RES_GET_CHECK|
 681                                            JOURNAL_RES_GET_RESERVED);
 682                 if (!ret) {
 683                         mutex_lock(&c->btree_interior_update_lock);
 684                         goto again;
 685                 }
 686         }
 687
 688         if (!ret) {
 689                 struct journal_buf *buf = &c->journal.buf[res.idx];
 690                 struct jset_entry *entry = vstruct_idx(buf->data, res.offset);
 691
 692                 res.offset      += as->journal_u64s;
 693                 res.u64s        -= as->journal_u64s;
 694                 memcpy_u64s(entry, as->journal_entries, as->journal_u64s);
 695         } else {
 696                 /*
 697                  * On journal error we have to run most of the normal path so
 698                  * that shutdown works - unblocking btree node writes in
 699                  * particular and writing them if needed - except for
 700                  * journalling the update:
 701                  */
 702
 703                 BUG_ON(!bch2_journal_error(&c->journal));
 704         }
 705
 706         switch (as->mode) {
 707         case BTREE_INTERIOR_NO_UPDATE:
 708                 BUG();
 709         case BTREE_INTERIOR_UPDATING_NODE:
 710                 /* @b is the node we did the final insert into: */
 711
 712                 /*
 713                  * On failure to get a journal reservation, we still have to
 714                  * unblock the write and allow most of the write path to happen
 715                  * so that shutdown works, but the i->journal_seq mechanism
 716                  * won't work to prevent the btree write from being visible (we
 717                  * didn't get a journal sequence number) - instead
 718                  * __bch2_btree_node_write() doesn't do the actual write if
 719                  * we're in journal error state:
 720                  */
 721
 722                 list_del(&as->write_blocked_list);
 723
 724                 if (!ret) {
 725                         struct bset *i = btree_bset_last(b);
 726
 727                         i->journal_seq = cpu_to_le64(
 728                                 max(res.seq,
 729                                     le64_to_cpu(i->journal_seq)));
 730
 731                         bch2_btree_add_journal_pin(c, b, res.seq);
 732                 }
 733
 734                 nodes_need_write[nr_nodes_need_write++] = b;
 735
 736                 six_unlock_write(&b->lock);
 737                 six_unlock_intent(&b->lock);
 738                 break;
 739
 740         case BTREE_INTERIOR_UPDATING_AS:
 741                 BUG_ON(b);
 742                 break;
 743
 744         case BTREE_INTERIOR_UPDATING_ROOT:
 745                 r = &c->btree_roots[as->btree_id];
 746
 747                 BUG_ON(b);
 748
 749                 mutex_lock(&c->btree_root_lock);
 750                 bkey_copy(&r->key, as->parent_keys.keys);
 751                 r->level = as->level;
 752                 r->alive = true;
 753                 c->btree_roots_dirty = true;
 754                 mutex_unlock(&c->btree_root_lock);
 755                 break;
 756         }
 757
 758         bch2_journal_pin_drop(&c->journal, &as->journal);
 759
 760         bch2_journal_res_put(&c->journal, &res);
 761         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 762
 763         while (as->nr_new_nodes) {
 764                 b = as->new_nodes[--as->nr_new_nodes];
 765
 766                 BUG_ON(b->will_make_reachable != (unsigned long) as);
 767                 b->will_make_reachable = 0;
 768
 769                 nodes_need_write[nr_nodes_need_write++] = b;
 770         }
 771
 772         while (as->nr_pending)
 773                 bch2_btree_node_free_ondisk(c,
 774                         &as->pending[--as->nr_pending], res.seq);
 775
 776         __bch2_btree_update_free(as);
 777         /*
 778          * for flush_held_btree_writes() waiting on updates to flush or
 779          * nodes to be writeable:
 780          */
 781         closure_wake_up(&c->btree_interior_update_wait);
 782
 783         /*
 784          * Can't take btree node locks while holding btree_interior_update_lock:
 785          * */
 786         mutex_unlock(&c->btree_interior_update_lock);
 787
 788         /* Do btree writes after dropping journal res/locks: */
 789         while (nr_nodes_need_write) {
 790                 b = nodes_need_write[--nr_nodes_need_write];
 791
 792                 btree_node_lock_type(c, b, SIX_LOCK_read);
 793                 bch2_btree_node_write_cond(c, b, btree_node_need_write(b));
 794                 six_unlock_read(&b->lock);
 795         }
 796
 797         mutex_lock(&c->btree_interior_update_lock);
 798         goto again;
 799 }
 800
 801 /*
 802  * We're updating @b with pointers to nodes that haven't finished writing yet:
 803  * block @b from being written until @as completes
 804  */
 805 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
 806 {
 807         struct bch_fs *c = as->c;
 808
 809         mutex_lock(&c->btree_interior_update_lock);
 810         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 811
 812         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 813         BUG_ON(!btree_node_dirty(b));
 814
 815         as->mode        = BTREE_INTERIOR_UPDATING_NODE;
 816         as->b           = b;
 817         as->level       = b->level;
 818         list_add(&as->write_blocked_list, &b->write_blocked);
 819
 820         mutex_unlock(&c->btree_interior_update_lock);
 821 }
 822
 823 static void btree_update_reparent(struct btree_update *as,
 824                                   struct btree_update *child)
 825 {
 826         struct bch_fs *c = as->c;
 827
 828         lockdep_assert_held(&c->btree_interior_update_lock);
 829
 830         child->b = NULL;
 831         child->mode = BTREE_INTERIOR_UPDATING_AS;
 832
 833         /*
 834          * When we write a new btree root, we have to drop our journal pin
 835          * _before_ the new nodes are technically reachable; see
 836          * btree_update_nodes_written().
 837          *
 838          * This goes for journal pins that are recursively blocked on us - so,
 839          * just transfer the journal pin to the new interior update so
 840          * btree_update_nodes_written() can drop it.
 841          */
 842         bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL);
 843         bch2_journal_pin_drop(&c->journal, &child->journal);
 844 }
 845
 846 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
 847 {
 848         struct bch_fs *c = as->c;
 849
 850         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 851         BUG_ON(!bch2_keylist_empty(&as->parent_keys));
 852
 853         mutex_lock(&c->btree_interior_update_lock);
 854         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 855
 856         as->mode        = BTREE_INTERIOR_UPDATING_ROOT;
 857         as->level       = b->level;
 858         bch2_keylist_add(&as->parent_keys, &b->key);
 859         mutex_unlock(&c->btree_interior_update_lock);
 860 }
 861
 862 static void btree_node_will_make_reachable(struct btree_update *as,
 863                                            struct btree *b)
 864 {
 865         struct bch_fs *c = as->c;
 866
 867         mutex_lock(&c->btree_interior_update_lock);
 868         BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
 869         BUG_ON(b->will_make_reachable);
 870
 871         as->new_nodes[as->nr_new_nodes++] = b;
 872         b->will_make_reachable = 1UL|(unsigned long) as;
 873
 874         closure_get(&as->cl);
 875         mutex_unlock(&c->btree_interior_update_lock);
 876 }
 877
 878 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
 879 {
 880         struct btree_update *as;
 881         unsigned long v;
 882         unsigned i;
 883
 884         mutex_lock(&c->btree_interior_update_lock);
 885         v = xchg(&b->will_make_reachable, 0);
 886         as = (struct btree_update *) (v & ~1UL);
 887
 888         if (!as) {
 889                 mutex_unlock(&c->btree_interior_update_lock);
 890                 return;
 891         }
 892
 893         for (i = 0; i < as->nr_new_nodes; i++)
 894                 if (as->new_nodes[i] == b)
 895                         goto found;
 896
 897         BUG();
 898 found:
 899         array_remove_item(as->new_nodes, as->nr_new_nodes, i);
 900         mutex_unlock(&c->btree_interior_update_lock);
 901
 902         if (v & 1)
 903                 closure_put(&as->cl);
 904 }
 905
 906 static void btree_interior_update_add_node_reference(struct btree_update *as,
 907                                                      struct btree *b)
 908 {
 909         struct bch_fs *c = as->c;
 910         struct pending_btree_node_free *d;
 911
 912         mutex_lock(&c->btree_interior_update_lock);
 913
 914         /* Add this node to the list of nodes being freed: */
 915         BUG_ON(as->nr_pending >= ARRAY_SIZE(as->pending));
 916
 917         d = &as->pending[as->nr_pending++];
 918         d->index_update_done    = false;
 919         d->seq                  = b->data->keys.seq;
 920         d->btree_id             = b->btree_id;
 921         d->level                = b->level;
 922         bkey_copy(&d->key, &b->key);
 923
 924         mutex_unlock(&c->btree_interior_update_lock);
 925 }
 926
 927 /*
 928  * @b is being split/rewritten: it may have pointers to not-yet-written btree
 929  * nodes and thus outstanding btree_updates - redirect @b's
 930  * btree_updates to point to this btree_update:
 931  */
 932 void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 933                                                struct btree *b)
 934 {
 935         struct bch_fs *c = as->c;
 936         struct btree_update *p, *n;
 937         struct btree_write *w;
 938
 939         set_btree_node_dying(b);
 940
 941         if (btree_node_fake(b))
 942                 return;
 943
 944         btree_interior_update_add_node_reference(as, b);
 945
 946         mutex_lock(&c->btree_interior_update_lock);
 947
 948         /*
 949          * Does this node have any btree_update operations preventing
 950          * it from being written?
 951          *
 952          * If so, redirect them to point to this btree_update: we can
 953          * write out our new nodes, but we won't make them visible until those
 954          * operations complete
 955          */
 956         list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
 957                 list_del(&p->write_blocked_list);
 958                 btree_update_reparent(as, p);
 959
 960                 /*
 961                  * for flush_held_btree_writes() waiting on updates to flush or
 962                  * nodes to be writeable:
 963                  */
 964                 closure_wake_up(&c->btree_interior_update_wait);
 965         }
 966
 967         clear_btree_node_dirty(b);
 968         clear_btree_node_need_write(b);
 969
 970         /*
 971          * Does this node have unwritten data that has a pin on the journal?
 972          *
 973          * If so, transfer that pin to the btree_update operation -
 974          * note that if we're freeing multiple nodes, we only need to keep the
 975          * oldest pin of any of the nodes we're freeing. We'll release the pin
 976          * when the new nodes are persistent and reachable on disk:
 977          */
 978         w = btree_current_write(b);
 979         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 980         bch2_journal_pin_drop(&c->journal, &w->journal);
 981
 982         w = btree_prev_write(b);
 983         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 984         bch2_journal_pin_drop(&c->journal, &w->journal);
 985
 986         mutex_unlock(&c->btree_interior_update_lock);
 987 }
 988
 989 void bch2_btree_update_done(struct btree_update *as)
 990 {
 991         BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
 992
 993         bch2_btree_reserve_put(as->c, as->reserve);
 994         as->reserve = NULL;
 995
 996         continue_at(&as->cl, btree_update_nodes_written, system_freezable_wq);
 997 }
 998
 999 struct btree_update *
1000 bch2_btree_update_start(struct btree_trans *trans, enum btree_id id,
1001                         unsigned nr_nodes, unsigned flags,
1002                         struct closure *cl)
1003 {
1004         struct bch_fs *c = trans->c;
1005         struct journal_preres journal_preres = { 0 };
1006         struct btree_reserve *reserve;
1007         struct btree_update *as;
1008         int ret;
1009
1010         ret = bch2_journal_preres_get(&c->journal, &journal_preres,
1011                                       BTREE_UPDATE_JOURNAL_RES,
1012                                       JOURNAL_RES_GET_NONBLOCK);
1013         if (ret == -EAGAIN) {
1014                 if (flags & BTREE_INSERT_NOUNLOCK)
1015                         return ERR_PTR(-EINTR);
1016
1017                 bch2_trans_unlock(trans);
1018
1019                 ret = bch2_journal_preres_get(&c->journal, &journal_preres,
1020                                               BTREE_UPDATE_JOURNAL_RES, 0);
1021                 if (ret)
1022                         return ERR_PTR(ret);
1023
1024                 if (!bch2_trans_relock(trans)) {
1025                         bch2_journal_preres_put(&c->journal, &journal_preres);
1026                         return ERR_PTR(-EINTR);
1027                 }
1028         }
1029
1030         reserve = bch2_btree_reserve_get(c, nr_nodes, flags, cl);
1031         if (IS_ERR(reserve)) {
1032                 bch2_journal_preres_put(&c->journal, &journal_preres);
1033                 return ERR_CAST(reserve);
1034         }
1035
1036         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
1037         memset(as, 0, sizeof(*as));
1038         closure_init(&as->cl, NULL);
1039         as->c           = c;
1040         as->mode        = BTREE_INTERIOR_NO_UPDATE;
1041         as->btree_id    = id;
1042         as->reserve     = reserve;
1043         INIT_LIST_HEAD(&as->write_blocked_list);
1044         INIT_LIST_HEAD(&as->unwritten_list);
1045         as->journal_preres = journal_preres;
1046
1047         bch2_keylist_init(&as->parent_keys, as->inline_keys);
1048
1049         mutex_lock(&c->btree_interior_update_lock);
1050         list_add_tail(&as->list, &c->btree_interior_update_list);
1051         mutex_unlock(&c->btree_interior_update_lock);
1052
1053         return as;
1054 }
1055
1056 /* Btree root updates: */
1057
1058 static void __bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1059 {
1060         /* Root nodes cannot be reaped */
1061         mutex_lock(&c->btree_cache.lock);
1062         list_del_init(&b->list);
1063         mutex_unlock(&c->btree_cache.lock);
1064
1065         mutex_lock(&c->btree_root_lock);
1066         BUG_ON(btree_node_root(c, b) &&
1067                (b->level < btree_node_root(c, b)->level ||
1068                 !btree_node_dying(btree_node_root(c, b))));
1069
1070         btree_node_root(c, b) = b;
1071         mutex_unlock(&c->btree_root_lock);
1072
1073         bch2_recalc_btree_reserve(c);
1074 }
1075
1076 static void bch2_btree_set_root_inmem(struct btree_update *as, struct btree *b)
1077 {
1078         struct bch_fs *c = as->c;
1079         struct btree *old = btree_node_root(c, b);
1080         struct bch_fs_usage *fs_usage;
1081
1082         __bch2_btree_set_root_inmem(c, b);
1083
1084         mutex_lock(&c->btree_interior_update_lock);
1085         percpu_down_read(&c->mark_lock);
1086         fs_usage = bch2_fs_usage_scratch_get(c);
1087
1088         bch2_mark_key_locked(c, bkey_i_to_s_c(&b->key),
1089                       0, 0, fs_usage, 0,
1090                       BTREE_TRIGGER_INSERT);
1091         if (gc_visited(c, gc_pos_btree_root(b->btree_id)))
1092                 bch2_mark_key_locked(c, bkey_i_to_s_c(&b->key),
1093                                      0, 0, NULL, 0,
1094                                      BTREE_TRIGGER_INSERT|
1095                                      BTREE_TRIGGER_GC);
1096
1097         if (old && !btree_node_fake(old))
1098                 bch2_btree_node_free_index(as, NULL,
1099                                            bkey_i_to_s_c(&old->key),
1100                                            fs_usage);
1101         bch2_fs_usage_apply(c, fs_usage, &as->reserve->disk_res, 0);
1102
1103         bch2_fs_usage_scratch_put(c, fs_usage);
1104         percpu_up_read(&c->mark_lock);
1105         mutex_unlock(&c->btree_interior_update_lock);
1106 }
1107
1108 /**
1109  * bch_btree_set_root - update the root in memory and on disk
1110  *
1111  * To ensure forward progress, the current task must not be holding any
1112  * btree node write locks. However, you must hold an intent lock on the
1113  * old root.
1114  *
1115  * Note: This allocates a journal entry but doesn't add any keys to
1116  * it.  All the btree roots are part of every journal write, so there
1117  * is nothing new to be done.  This just guarantees that there is a
1118  * journal write.
1119  */
1120 static void bch2_btree_set_root(struct btree_update *as, struct btree *b,
1121                                 struct btree_iter *iter)
1122 {
1123         struct bch_fs *c = as->c;
1124         struct btree *old;
1125
1126         trace_btree_set_root(c, b);
1127         BUG_ON(!b->written &&
1128                !test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags));
1129
1130         old = btree_node_root(c, b);
1131
1132         /*
1133          * Ensure no one is using the old root while we switch to the
1134          * new root:
1135          */
1136         bch2_btree_node_lock_write(old, iter);
1137
1138         bch2_btree_set_root_inmem(as, b);
1139
1140         btree_update_updated_root(as, b);
1141
1142         /*
1143          * Unlock old root after new root is visible:
1144          *
1145          * The new root isn't persistent, but that's ok: we still have
1146          * an intent lock on the new root, and any updates that would
1147          * depend on the new root would have to update the new root.
1148          */
1149         bch2_btree_node_unlock_write(old, iter);
1150 }
1151
1152 /* Interior node updates: */
1153
1154 static void bch2_insert_fixup_btree_ptr(struct btree_update *as, struct btree *b,
1155                                         struct btree_iter *iter,
1156                                         struct bkey_i *insert,
1157                                         struct btree_node_iter *node_iter)
1158 {
1159         struct bch_fs *c = as->c;
1160         struct bch_fs_usage *fs_usage;
1161         struct jset_entry *entry;
1162         struct bkey_packed *k;
1163         struct bkey tmp;
1164
1165         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1166                ARRAY_SIZE(as->journal_entries));
1167
1168         entry = (void *) &as->journal_entries[as->journal_u64s];
1169         memset(entry, 0, sizeof(*entry));
1170         entry->u64s     = cpu_to_le16(insert->k.u64s);
1171         entry->type     = BCH_JSET_ENTRY_btree_keys;
1172         entry->btree_id = b->btree_id;
1173         entry->level    = b->level;
1174         memcpy_u64s_small(entry->_data, insert, insert->k.u64s);
1175         as->journal_u64s += jset_u64s(insert->k.u64s);
1176
1177         mutex_lock(&c->btree_interior_update_lock);
1178         percpu_down_read(&c->mark_lock);
1179         fs_usage = bch2_fs_usage_scratch_get(c);
1180
1181         bch2_mark_key_locked(c, bkey_i_to_s_c(insert),
1182                              0, 0, fs_usage, 0,
1183                              BTREE_TRIGGER_INSERT);
1184
1185         if (gc_visited(c, gc_pos_btree_node(b)))
1186                 bch2_mark_key_locked(c, bkey_i_to_s_c(insert),
1187                                      0, 0, NULL, 0,
1188                                      BTREE_TRIGGER_INSERT|
1189                                      BTREE_TRIGGER_GC);
1190
1191         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1192                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1193                 bch2_btree_node_iter_advance(node_iter, b);
1194
1195         /*
1196          * If we're overwriting, look up pending delete and mark so that gc
1197          * marks it on the pending delete list:
1198          */
1199         if (k && !bkey_cmp_packed(b, k, &insert->k))
1200                 bch2_btree_node_free_index(as, b,
1201                                            bkey_disassemble(b, k, &tmp),
1202                                            fs_usage);
1203
1204         bch2_fs_usage_apply(c, fs_usage, &as->reserve->disk_res, 0);
1205
1206         bch2_fs_usage_scratch_put(c, fs_usage);
1207         percpu_up_read(&c->mark_lock);
1208         mutex_unlock(&c->btree_interior_update_lock);
1209
1210         bch2_btree_bset_insert_key(iter, b, node_iter, insert);
1211         set_btree_node_dirty(b);
1212         set_btree_node_need_write(b);
1213 }
1214
1215 /*
1216  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1217  * node)
1218  */
1219 static struct btree *__btree_split_node(struct btree_update *as,
1220                                         struct btree *n1,
1221                                         struct btree_iter *iter)
1222 {
1223         size_t nr_packed = 0, nr_unpacked = 0;
1224         struct btree *n2;
1225         struct bset *set1, *set2;
1226         struct bkey_packed *k, *prev = NULL;
1227
1228         n2 = bch2_btree_node_alloc(as, n1->level);
1229
1230         n2->data->max_key       = n1->data->max_key;
1231         n2->data->format        = n1->format;
1232         SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data));
1233         n2->key.k.p = n1->key.k.p;
1234
1235         btree_node_set_format(n2, n2->data->format);
1236
1237         set1 = btree_bset_first(n1);
1238         set2 = btree_bset_first(n2);
1239
1240         /*
1241          * Has to be a linear search because we don't have an auxiliary
1242          * search tree yet
1243          */
1244         k = set1->start;
1245         while (1) {
1246                 struct bkey_packed *n = bkey_next_skip_noops(k, vstruct_last(set1));
1247
1248                 if (n == vstruct_last(set1))
1249                         break;
1250                 if (k->_data - set1->_data >= (le16_to_cpu(set1->u64s) * 3) / 5)
1251                         break;
1252
1253                 if (bkey_packed(k))
1254                         nr_packed++;
1255                 else
1256                         nr_unpacked++;
1257
1258                 prev = k;
1259                 k = n;
1260         }
1261
1262         BUG_ON(!prev);
1263
1264         btree_set_max(n1, bkey_unpack_pos(n1, prev));
1265         btree_set_min(n2, bkey_successor(n1->key.k.p));
1266
1267         set2->u64s = cpu_to_le16((u64 *) vstruct_end(set1) - (u64 *) k);
1268         set1->u64s = cpu_to_le16(le16_to_cpu(set1->u64s) - le16_to_cpu(set2->u64s));
1269
1270         set_btree_bset_end(n1, n1->set);
1271         set_btree_bset_end(n2, n2->set);
1272
1273         n2->nr.live_u64s        = le16_to_cpu(set2->u64s);
1274         n2->nr.bset_u64s[0]     = le16_to_cpu(set2->u64s);
1275         n2->nr.packed_keys      = n1->nr.packed_keys - nr_packed;
1276         n2->nr.unpacked_keys    = n1->nr.unpacked_keys - nr_unpacked;
1277
1278         n1->nr.live_u64s        = le16_to_cpu(set1->u64s);
1279         n1->nr.bset_u64s[0]     = le16_to_cpu(set1->u64s);
1280         n1->nr.packed_keys      = nr_packed;
1281         n1->nr.unpacked_keys    = nr_unpacked;
1282
1283         BUG_ON(!set1->u64s);
1284         BUG_ON(!set2->u64s);
1285
1286         memcpy_u64s(set2->start,
1287                     vstruct_end(set1),
1288                     le16_to_cpu(set2->u64s));
1289
1290         btree_node_reset_sib_u64s(n1);
1291         btree_node_reset_sib_u64s(n2);
1292
1293         bch2_verify_btree_nr_keys(n1);
1294         bch2_verify_btree_nr_keys(n2);
1295
1296         if (n1->level) {
1297                 btree_node_interior_verify(n1);
1298                 btree_node_interior_verify(n2);
1299         }
1300
1301         return n2;
1302 }
1303
1304 /*
1305  * For updates to interior nodes, we've got to do the insert before we split
1306  * because the stuff we're inserting has to be inserted atomically. Post split,
1307  * the keys might have to go in different nodes and the split would no longer be
1308  * atomic.
1309  *
1310  * Worse, if the insert is from btree node coalescing, if we do the insert after
1311  * we do the split (and pick the pivot) - the pivot we pick might be between
1312  * nodes that were coalesced, and thus in the middle of a child node post
1313  * coalescing:
1314  */
1315 static void btree_split_insert_keys(struct btree_update *as, struct btree *b,
1316                                     struct btree_iter *iter,
1317                                     struct keylist *keys)
1318 {
1319         struct btree_node_iter node_iter;
1320         struct bkey_i *k = bch2_keylist_front(keys);
1321         struct bkey_packed *src, *dst, *n;
1322         struct bset *i;
1323
1324         /*
1325          * XXX
1326          *
1327          * these updates must be journalled
1328          *
1329          * oops
1330          */
1331
1332         BUG_ON(btree_node_type(b) != BKEY_TYPE_BTREE);
1333
1334         bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
1335
1336         while (!bch2_keylist_empty(keys)) {
1337                 k = bch2_keylist_front(keys);
1338
1339                 bch2_insert_fixup_btree_ptr(as, b, iter, k, &node_iter);
1340                 bch2_keylist_pop_front(keys);
1341         }
1342
1343         /*
1344          * We can't tolerate whiteouts here - with whiteouts there can be
1345          * duplicate keys, and it would be rather bad if we picked a duplicate
1346          * for the pivot:
1347          */
1348         i = btree_bset_first(b);
1349         src = dst = i->start;
1350         while (src != vstruct_last(i)) {
1351                 n = bkey_next_skip_noops(src, vstruct_last(i));
1352                 if (!bkey_deleted(src)) {
1353                         memmove_u64s_down(dst, src, src->u64s);
1354                         dst = bkey_next(dst);
1355                 }
1356                 src = n;
1357         }
1358
1359         i->u64s = cpu_to_le16((u64 *) dst - i->_data);
1360         set_btree_bset_end(b, b->set);
1361
1362         BUG_ON(b->nsets != 1 ||
1363                b->nr.live_u64s != le16_to_cpu(btree_bset_first(b)->u64s));
1364
1365         btree_node_interior_verify(b);
1366 }
1367
1368 static void btree_split(struct btree_update *as, struct btree *b,
1369                         struct btree_iter *iter, struct keylist *keys,
1370                         unsigned flags)
1371 {
1372         struct bch_fs *c = as->c;
1373         struct btree *parent = btree_node_parent(iter, b);
1374         struct btree *n1, *n2 = NULL, *n3 = NULL;
1375         u64 start_time = local_clock();
1376
1377         BUG_ON(!parent && (b != btree_node_root(c, b)));
1378         BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->level));
1379
1380         bch2_btree_interior_update_will_free_node(as, b);
1381
1382         n1 = bch2_btree_node_alloc_replacement(as, b);
1383
1384         if (keys)
1385                 btree_split_insert_keys(as, n1, iter, keys);
1386
1387         if (bset_u64s(&n1->set[0]) > BTREE_SPLIT_THRESHOLD(c)) {
1388                 trace_btree_split(c, b);
1389
1390                 n2 = __btree_split_node(as, n1, iter);
1391
1392                 bch2_btree_build_aux_trees(n2);
1393                 bch2_btree_build_aux_trees(n1);
1394                 six_unlock_write(&n2->lock);
1395                 six_unlock_write(&n1->lock);
1396
1397                 bch2_btree_node_write(c, n2, SIX_LOCK_intent);
1398
1399                 /*
1400                  * Note that on recursive parent_keys == keys, so we
1401                  * can't start adding new keys to parent_keys before emptying it
1402                  * out (which we did with btree_split_insert_keys() above)
1403                  */
1404                 bch2_keylist_add(&as->parent_keys, &n1->key);
1405                 bch2_keylist_add(&as->parent_keys, &n2->key);
1406
1407                 if (!parent) {
1408                         /* Depth increases, make a new root */
1409                         n3 = __btree_root_alloc(as, b->level + 1);
1410
1411                         n3->sib_u64s[0] = U16_MAX;
1412                         n3->sib_u64s[1] = U16_MAX;
1413
1414                         btree_split_insert_keys(as, n3, iter, &as->parent_keys);
1415
1416                         bch2_btree_node_write(c, n3, SIX_LOCK_intent);
1417                 }
1418         } else {
1419                 trace_btree_compact(c, b);
1420
1421                 bch2_btree_build_aux_trees(n1);
1422                 six_unlock_write(&n1->lock);
1423
1424                 if (parent)
1425                         bch2_keylist_add(&as->parent_keys, &n1->key);
1426         }
1427
1428         bch2_btree_node_write(c, n1, SIX_LOCK_intent);
1429
1430         /* New nodes all written, now make them visible: */
1431
1432         if (parent) {
1433                 /* Split a non root node */
1434                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1435         } else if (n3) {
1436                 bch2_btree_set_root(as, n3, iter);
1437         } else {
1438                 /* Root filled up but didn't need to be split */
1439                 bch2_btree_set_root(as, n1, iter);
1440         }
1441
1442         bch2_open_buckets_put(c, &n1->ob);
1443         if (n2)
1444                 bch2_open_buckets_put(c, &n2->ob);
1445         if (n3)
1446                 bch2_open_buckets_put(c, &n3->ob);
1447
1448         /* Successful split, update the iterator to point to the new nodes: */
1449
1450         six_lock_increment(&b->lock, SIX_LOCK_intent);
1451         bch2_btree_iter_node_drop(iter, b);
1452         if (n3)
1453                 bch2_btree_iter_node_replace(iter, n3);
1454         if (n2)
1455                 bch2_btree_iter_node_replace(iter, n2);
1456         bch2_btree_iter_node_replace(iter, n1);
1457
1458         /*
1459          * The old node must be freed (in memory) _before_ unlocking the new
1460          * nodes - else another thread could re-acquire a read lock on the old
1461          * node after another thread has locked and updated the new node, thus
1462          * seeing stale data:
1463          */
1464         bch2_btree_node_free_inmem(c, b, iter);
1465
1466         if (n3)
1467                 six_unlock_intent(&n3->lock);
1468         if (n2)
1469                 six_unlock_intent(&n2->lock);
1470         six_unlock_intent(&n1->lock);
1471
1472         bch2_btree_trans_verify_locks(iter->trans);
1473
1474         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_split],
1475                                start_time);
1476 }
1477
1478 static void
1479 bch2_btree_insert_keys_interior(struct btree_update *as, struct btree *b,
1480                                 struct btree_iter *iter, struct keylist *keys)
1481 {
1482         struct btree_iter *linked;
1483         struct btree_node_iter node_iter;
1484         struct bkey_i *insert = bch2_keylist_front(keys);
1485         struct bkey_packed *k;
1486
1487         /* Don't screw up @iter's position: */
1488         node_iter = iter->l[b->level].iter;
1489
1490         /*
1491          * btree_split(), btree_gc_coalesce() will insert keys before
1492          * the iterator's current position - they know the keys go in
1493          * the node the iterator points to:
1494          */
1495         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1496                (bkey_cmp_packed(b, k, &insert->k) >= 0))
1497                 ;
1498
1499         for_each_keylist_key(keys, insert)
1500                 bch2_insert_fixup_btree_ptr(as, b, iter, insert, &node_iter);
1501
1502         btree_update_updated_node(as, b);
1503
1504         trans_for_each_iter_with_node(iter->trans, b, linked)
1505                 bch2_btree_node_iter_peek(&linked->l[b->level].iter, b);
1506
1507         bch2_btree_trans_verify_iters(iter->trans, b);
1508 }
1509
1510 /**
1511  * bch_btree_insert_node - insert bkeys into a given btree node
1512  *
1513  * @iter:               btree iterator
1514  * @keys:               list of keys to insert
1515  * @hook:               insert callback
1516  * @persistent:         if not null, @persistent will wait on journal write
1517  *
1518  * Inserts as many keys as it can into a given btree node, splitting it if full.
1519  * If a split occurred, this function will return early. This can only happen
1520  * for leaf nodes -- inserts into interior nodes have to be atomic.
1521  */
1522 void bch2_btree_insert_node(struct btree_update *as, struct btree *b,
1523                             struct btree_iter *iter, struct keylist *keys,
1524                             unsigned flags)
1525 {
1526         struct bch_fs *c = as->c;
1527         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1528         int old_live_u64s = b->nr.live_u64s;
1529         int live_u64s_added, u64s_added;
1530
1531         BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->level));
1532         BUG_ON(!b->level);
1533         BUG_ON(!as || as->b);
1534         bch2_verify_keylist_sorted(keys);
1535
1536         if (as->must_rewrite)
1537                 goto split;
1538
1539         bch2_btree_node_lock_for_insert(c, b, iter);
1540
1541         if (!bch2_btree_node_insert_fits(c, b, bch_keylist_u64s(keys))) {
1542                 bch2_btree_node_unlock_write(b, iter);
1543                 goto split;
1544         }
1545
1546         bch2_btree_insert_keys_interior(as, b, iter, keys);
1547
1548         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1549         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1550
1551         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1552                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1553         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1554                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1555
1556         if (u64s_added > live_u64s_added &&
1557             bch2_maybe_compact_whiteouts(c, b))
1558                 bch2_btree_iter_reinit_node(iter, b);
1559
1560         bch2_btree_node_unlock_write(b, iter);
1561
1562         btree_node_interior_verify(b);
1563
1564         /*
1565          * when called from the btree_split path the new nodes aren't added to
1566          * the btree iterator yet, so the merge path's unlock/wait/relock dance
1567          * won't work:
1568          */
1569         bch2_foreground_maybe_merge(c, iter, b->level,
1570                                     flags|BTREE_INSERT_NOUNLOCK);
1571         return;
1572 split:
1573         btree_split(as, b, iter, keys, flags);
1574 }
1575
1576 int bch2_btree_split_leaf(struct bch_fs *c, struct btree_iter *iter,
1577                           unsigned flags)
1578 {
1579         struct btree_trans *trans = iter->trans;
1580         struct btree *b = iter_l(iter)->b;
1581         struct btree_update *as;
1582         struct closure cl;
1583         int ret = 0;
1584         struct btree_iter *linked;
1585
1586         /*
1587          * We already have a disk reservation and open buckets pinned; this
1588          * allocation must not block:
1589          */
1590         trans_for_each_iter(trans, linked)
1591                 if (linked->btree_id == BTREE_ID_EXTENTS)
1592                         flags |= BTREE_INSERT_USE_RESERVE;
1593
1594         closure_init_stack(&cl);
1595
1596         /* Hack, because gc and splitting nodes doesn't mix yet: */
1597         if (!(flags & BTREE_INSERT_GC_LOCK_HELD) &&
1598             !down_read_trylock(&c->gc_lock)) {
1599                 if (flags & BTREE_INSERT_NOUNLOCK) {
1600                         trace_transaction_restart_ip(trans->ip, _THIS_IP_);
1601                         return -EINTR;
1602                 }
1603
1604                 bch2_trans_unlock(trans);
1605                 down_read(&c->gc_lock);
1606
1607                 if (!bch2_trans_relock(trans))
1608                         ret = -EINTR;
1609         }
1610
1611         /*
1612          * XXX: figure out how far we might need to split,
1613          * instead of locking/reserving all the way to the root:
1614          */
1615         if (!bch2_btree_iter_upgrade(iter, U8_MAX)) {
1616                 trace_trans_restart_iter_upgrade(trans->ip);
1617                 ret = -EINTR;
1618                 goto out;
1619         }
1620
1621         as = bch2_btree_update_start(trans, iter->btree_id,
1622                 btree_update_reserve_required(c, b), flags,
1623                 !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1624         if (IS_ERR(as)) {
1625                 ret = PTR_ERR(as);
1626                 if (ret == -EAGAIN) {
1627                         BUG_ON(flags & BTREE_INSERT_NOUNLOCK);
1628                         bch2_trans_unlock(trans);
1629                         ret = -EINTR;
1630
1631                         trace_transaction_restart_ip(trans->ip, _THIS_IP_);
1632                 }
1633                 goto out;
1634         }
1635
1636         btree_split(as, b, iter, NULL, flags);
1637         bch2_btree_update_done(as);
1638
1639         /*
1640          * We haven't successfully inserted yet, so don't downgrade all the way
1641          * back to read locks;
1642          */
1643         __bch2_btree_iter_downgrade(iter, 1);
1644 out:
1645         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1646                 up_read(&c->gc_lock);
1647         closure_sync(&cl);
1648         return ret;
1649 }
1650
1651 void __bch2_foreground_maybe_merge(struct bch_fs *c,
1652                                    struct btree_iter *iter,
1653                                    unsigned level,
1654                                    unsigned flags,
1655                                    enum btree_node_sibling sib)
1656 {
1657         struct btree_trans *trans = iter->trans;
1658         struct btree_update *as;
1659         struct bkey_format_state new_s;
1660         struct bkey_format new_f;
1661         struct bkey_i delete;
1662         struct btree *b, *m, *n, *prev, *next, *parent;
1663         struct closure cl;
1664         size_t sib_u64s;
1665         int ret = 0;
1666
1667         BUG_ON(!btree_node_locked(iter, level));
1668
1669         closure_init_stack(&cl);
1670 retry:
1671         BUG_ON(!btree_node_locked(iter, level));
1672
1673         b = iter->l[level].b;
1674
1675         parent = btree_node_parent(iter, b);
1676         if (!parent)
1677                 goto out;
1678
1679         if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c))
1680                 goto out;
1681
1682         /* XXX: can't be holding read locks */
1683         m = bch2_btree_node_get_sibling(c, iter, b, sib);
1684         if (IS_ERR(m)) {
1685                 ret = PTR_ERR(m);
1686                 goto err;
1687         }
1688
1689         /* NULL means no sibling: */
1690         if (!m) {
1691                 b->sib_u64s[sib] = U16_MAX;
1692                 goto out;
1693         }
1694
1695         if (sib == btree_prev_sib) {
1696                 prev = m;
1697                 next = b;
1698         } else {
1699                 prev = b;
1700                 next = m;
1701         }
1702
1703         bch2_bkey_format_init(&new_s);
1704         __bch2_btree_calc_format(&new_s, b);
1705         __bch2_btree_calc_format(&new_s, m);
1706         new_f = bch2_bkey_format_done(&new_s);
1707
1708         sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1709                 btree_node_u64s_with_format(m, &new_f);
1710
1711         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1712                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1713                 sib_u64s /= 2;
1714                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1715         }
1716
1717         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1718         b->sib_u64s[sib] = sib_u64s;
1719
1720         if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c)) {
1721                 six_unlock_intent(&m->lock);
1722                 goto out;
1723         }
1724
1725         /* We're changing btree topology, doesn't mix with gc: */
1726         if (!(flags & BTREE_INSERT_GC_LOCK_HELD) &&
1727             !down_read_trylock(&c->gc_lock))
1728                 goto err_cycle_gc_lock;
1729
1730         if (!bch2_btree_iter_upgrade(iter, U8_MAX)) {
1731                 ret = -EINTR;
1732                 goto err_unlock;
1733         }
1734
1735         as = bch2_btree_update_start(trans, iter->btree_id,
1736                          btree_update_reserve_required(c, parent) + 1,
1737                          flags|
1738                          BTREE_INSERT_NOFAIL|
1739                          BTREE_INSERT_USE_RESERVE,
1740                          !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1741         if (IS_ERR(as)) {
1742                 ret = PTR_ERR(as);
1743                 goto err_unlock;
1744         }
1745
1746         trace_btree_merge(c, b);
1747
1748         bch2_btree_interior_update_will_free_node(as, b);
1749         bch2_btree_interior_update_will_free_node(as, m);
1750
1751         n = bch2_btree_node_alloc(as, b->level);
1752
1753         btree_set_min(n, prev->data->min_key);
1754         btree_set_max(n, next->data->max_key);
1755         n->data->format         = new_f;
1756
1757         btree_node_set_format(n, new_f);
1758
1759         bch2_btree_sort_into(c, n, prev);
1760         bch2_btree_sort_into(c, n, next);
1761
1762         bch2_btree_build_aux_trees(n);
1763         six_unlock_write(&n->lock);
1764
1765         bkey_init(&delete.k);
1766         delete.k.p = prev->key.k.p;
1767         bch2_keylist_add(&as->parent_keys, &delete);
1768         bch2_keylist_add(&as->parent_keys, &n->key);
1769
1770         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1771
1772         bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1773
1774         bch2_open_buckets_put(c, &n->ob);
1775
1776         six_lock_increment(&b->lock, SIX_LOCK_intent);
1777         bch2_btree_iter_node_drop(iter, b);
1778         bch2_btree_iter_node_drop(iter, m);
1779
1780         bch2_btree_iter_node_replace(iter, n);
1781
1782         bch2_btree_trans_verify_iters(trans, n);
1783
1784         bch2_btree_node_free_inmem(c, b, iter);
1785         bch2_btree_node_free_inmem(c, m, iter);
1786
1787         six_unlock_intent(&n->lock);
1788
1789         bch2_btree_update_done(as);
1790
1791         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1792                 up_read(&c->gc_lock);
1793 out:
1794         bch2_btree_trans_verify_locks(trans);
1795
1796         /*
1797          * Don't downgrade locks here: we're called after successful insert,
1798          * and the caller will downgrade locks after a successful insert
1799          * anyways (in case e.g. a split was required first)
1800          *
1801          * And we're also called when inserting into interior nodes in the
1802          * split path, and downgrading to read locks in there is potentially
1803          * confusing:
1804          */
1805         closure_sync(&cl);
1806         return;
1807
1808 err_cycle_gc_lock:
1809         six_unlock_intent(&m->lock);
1810
1811         if (flags & BTREE_INSERT_NOUNLOCK)
1812                 goto out;
1813
1814         bch2_trans_unlock(trans);
1815
1816         down_read(&c->gc_lock);
1817         up_read(&c->gc_lock);
1818         ret = -EINTR;
1819         goto err;
1820
1821 err_unlock:
1822         six_unlock_intent(&m->lock);
1823         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1824                 up_read(&c->gc_lock);
1825 err:
1826         BUG_ON(ret == -EAGAIN && (flags & BTREE_INSERT_NOUNLOCK));
1827
1828         if ((ret == -EAGAIN || ret == -EINTR) &&
1829             !(flags & BTREE_INSERT_NOUNLOCK)) {
1830                 bch2_trans_unlock(trans);
1831                 closure_sync(&cl);
1832                 ret = bch2_btree_iter_traverse(iter);
1833                 if (ret)
1834                         goto out;
1835
1836                 goto retry;
1837         }
1838
1839         goto out;
1840 }
1841
1842 static int __btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
1843                                 struct btree *b, unsigned flags,
1844                                 struct closure *cl)
1845 {
1846         struct btree *n, *parent = btree_node_parent(iter, b);
1847         struct btree_update *as;
1848
1849         as = bch2_btree_update_start(iter->trans, iter->btree_id,
1850                 (parent
1851                  ? btree_update_reserve_required(c, parent)
1852                  : 0) + 1,
1853                 flags, cl);
1854         if (IS_ERR(as)) {
1855                 trace_btree_gc_rewrite_node_fail(c, b);
1856                 return PTR_ERR(as);
1857         }
1858
1859         bch2_btree_interior_update_will_free_node(as, b);
1860
1861         n = bch2_btree_node_alloc_replacement(as, b);
1862
1863         bch2_btree_build_aux_trees(n);
1864         six_unlock_write(&n->lock);
1865
1866         trace_btree_gc_rewrite_node(c, b);
1867
1868         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1869
1870         if (parent) {
1871                 bch2_keylist_add(&as->parent_keys, &n->key);
1872                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1873         } else {
1874                 bch2_btree_set_root(as, n, iter);
1875         }
1876
1877         bch2_open_buckets_put(c, &n->ob);
1878
1879         six_lock_increment(&b->lock, SIX_LOCK_intent);
1880         bch2_btree_iter_node_drop(iter, b);
1881         bch2_btree_iter_node_replace(iter, n);
1882         bch2_btree_node_free_inmem(c, b, iter);
1883         six_unlock_intent(&n->lock);
1884
1885         bch2_btree_update_done(as);
1886         return 0;
1887 }
1888
1889 /**
1890  * bch_btree_node_rewrite - Rewrite/move a btree node
1891  *
1892  * Returns 0 on success, -EINTR or -EAGAIN on failure (i.e.
1893  * btree_check_reserve() has to wait)
1894  */
1895 int bch2_btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
1896                             __le64 seq, unsigned flags)
1897 {
1898         struct btree_trans *trans = iter->trans;
1899         struct closure cl;
1900         struct btree *b;
1901         int ret;
1902
1903         flags |= BTREE_INSERT_NOFAIL;
1904
1905         closure_init_stack(&cl);
1906
1907         bch2_btree_iter_upgrade(iter, U8_MAX);
1908
1909         if (!(flags & BTREE_INSERT_GC_LOCK_HELD)) {
1910                 if (!down_read_trylock(&c->gc_lock)) {
1911                         bch2_trans_unlock(trans);
1912                         down_read(&c->gc_lock);
1913                 }
1914         }
1915
1916         while (1) {
1917                 ret = bch2_btree_iter_traverse(iter);
1918                 if (ret)
1919                         break;
1920
1921                 b = bch2_btree_iter_peek_node(iter);
1922                 if (!b || b->data->keys.seq != seq)
1923                         break;
1924
1925                 ret = __btree_node_rewrite(c, iter, b, flags, &cl);
1926                 if (ret != -EAGAIN &&
1927                     ret != -EINTR)
1928                         break;
1929
1930                 bch2_trans_unlock(trans);
1931                 closure_sync(&cl);
1932         }
1933
1934         bch2_btree_iter_downgrade(iter);
1935
1936         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1937                 up_read(&c->gc_lock);
1938
1939         closure_sync(&cl);
1940         return ret;
1941 }
1942
1943 static void __bch2_btree_node_update_key(struct bch_fs *c,
1944                                          struct btree_update *as,
1945                                          struct btree_iter *iter,
1946                                          struct btree *b, struct btree *new_hash,
1947                                          struct bkey_i *new_key)
1948 {
1949         struct btree *parent;
1950         int ret;
1951
1952         /*
1953          * Two corner cases that need to be thought about here:
1954          *
1955          * @b may not be reachable yet - there might be another interior update
1956          * operation waiting on @b to be written, and we're gonna deliver the
1957          * write completion to that interior update operation _before_
1958          * persisting the new_key update
1959          *
1960          * That ends up working without us having to do anything special here:
1961          * the reason is, we do kick off (and do the in memory updates) for the
1962          * update for @new_key before we return, creating a new interior_update
1963          * operation here.
1964          *
1965          * The new interior update operation here will in effect override the
1966          * previous one. The previous one was going to terminate - make @b
1967          * reachable - in one of two ways:
1968          * - updating the btree root pointer
1969          *   In that case,
1970          *   no, this doesn't work. argh.
1971          */
1972
1973         if (b->will_make_reachable)
1974                 as->must_rewrite = true;
1975
1976         btree_interior_update_add_node_reference(as, b);
1977
1978         /*
1979          * XXX: the rest of the update path treats this like we're actually
1980          * inserting a new node and deleting the existing node, so the
1981          * reservation needs to include enough space for @b
1982          *
1983          * that is actually sketch as fuck though and I am surprised the code
1984          * seems to work like that, definitely need to go back and rework it
1985          * into something saner.
1986          *
1987          * (I think @b is just getting double counted until the btree update
1988          * finishes and "deletes" @b on disk)
1989          */
1990         ret = bch2_disk_reservation_add(c, &as->reserve->disk_res,
1991                         c->opts.btree_node_size *
1992                         bch2_bkey_nr_ptrs(bkey_i_to_s_c(new_key)),
1993                         BCH_DISK_RESERVATION_NOFAIL);
1994         BUG_ON(ret);
1995
1996         parent = btree_node_parent(iter, b);
1997         if (parent) {
1998                 if (new_hash) {
1999                         bkey_copy(&new_hash->key, new_key);
2000                         ret = bch2_btree_node_hash_insert(&c->btree_cache,
2001                                         new_hash, b->level, b->btree_id);
2002                         BUG_ON(ret);
2003                 }
2004
2005                 bch2_keylist_add(&as->parent_keys, new_key);
2006                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, 0);
2007
2008                 if (new_hash) {
2009                         mutex_lock(&c->btree_cache.lock);
2010                         bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
2011
2012                         bch2_btree_node_hash_remove(&c->btree_cache, b);
2013
2014                         bkey_copy(&b->key, new_key);
2015                         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
2016                         BUG_ON(ret);
2017                         mutex_unlock(&c->btree_cache.lock);
2018                 } else {
2019                         bkey_copy(&b->key, new_key);
2020                 }
2021         } else {
2022                 struct bch_fs_usage *fs_usage;
2023
2024                 BUG_ON(btree_node_root(c, b) != b);
2025
2026                 bch2_btree_node_lock_write(b, iter);
2027
2028                 mutex_lock(&c->btree_interior_update_lock);
2029                 percpu_down_read(&c->mark_lock);
2030                 fs_usage = bch2_fs_usage_scratch_get(c);
2031
2032                 bch2_mark_key_locked(c, bkey_i_to_s_c(new_key),
2033                               0, 0, fs_usage, 0,
2034                               BTREE_TRIGGER_INSERT);
2035                 if (gc_visited(c, gc_pos_btree_root(b->btree_id)))
2036                         bch2_mark_key_locked(c, bkey_i_to_s_c(new_key),
2037                                              0, 0, NULL, 0,
2038                                              BTREE_TRIGGER_INSERT||
2039                                              BTREE_TRIGGER_GC);
2040
2041                 bch2_btree_node_free_index(as, NULL,
2042                                            bkey_i_to_s_c(&b->key),
2043                                            fs_usage);
2044                 bch2_fs_usage_apply(c, fs_usage, &as->reserve->disk_res, 0);
2045
2046                 bch2_fs_usage_scratch_put(c, fs_usage);
2047                 percpu_up_read(&c->mark_lock);
2048                 mutex_unlock(&c->btree_interior_update_lock);
2049
2050                 if (btree_ptr_hash_val(new_key) != b->hash_val) {
2051                         mutex_lock(&c->btree_cache.lock);
2052                         bch2_btree_node_hash_remove(&c->btree_cache, b);
2053
2054                         bkey_copy(&b->key, new_key);
2055                         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
2056                         BUG_ON(ret);
2057                         mutex_unlock(&c->btree_cache.lock);
2058                 } else {
2059                         bkey_copy(&b->key, new_key);
2060                 }
2061
2062                 btree_update_updated_root(as, b);
2063                 bch2_btree_node_unlock_write(b, iter);
2064         }
2065
2066         bch2_btree_update_done(as);
2067 }
2068
2069 int bch2_btree_node_update_key(struct bch_fs *c, struct btree_iter *iter,
2070                                struct btree *b,
2071                                struct bkey_i *new_key)
2072 {
2073         struct btree *parent = btree_node_parent(iter, b);
2074         struct btree_update *as = NULL;
2075         struct btree *new_hash = NULL;
2076         struct closure cl;
2077         int ret;
2078
2079         closure_init_stack(&cl);
2080
2081         if (!bch2_btree_iter_upgrade(iter, U8_MAX))
2082                 return -EINTR;
2083
2084         if (!down_read_trylock(&c->gc_lock)) {
2085                 bch2_trans_unlock(iter->trans);
2086                 down_read(&c->gc_lock);
2087
2088                 if (!bch2_trans_relock(iter->trans)) {
2089                         ret = -EINTR;
2090                         goto err;
2091                 }
2092         }
2093
2094         /*
2095          * check btree_ptr_hash_val() after @b is locked by
2096          * btree_iter_traverse():
2097          */
2098         if (btree_ptr_hash_val(new_key) != b->hash_val) {
2099                 /* bch2_btree_reserve_get will unlock */
2100                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2101                 if (ret) {
2102                         bch2_trans_unlock(iter->trans);
2103                         up_read(&c->gc_lock);
2104                         closure_sync(&cl);
2105                         down_read(&c->gc_lock);
2106
2107                         if (!bch2_trans_relock(iter->trans)) {
2108                                 ret = -EINTR;
2109                                 goto err;
2110                         }
2111                 }
2112
2113                 new_hash = bch2_btree_node_mem_alloc(c);
2114         }
2115
2116         as = bch2_btree_update_start(iter->trans, iter->btree_id,
2117                 parent ? btree_update_reserve_required(c, parent) : 0,
2118                 BTREE_INSERT_NOFAIL|
2119                 BTREE_INSERT_USE_RESERVE|
2120                 BTREE_INSERT_USE_ALLOC_RESERVE,
2121                 &cl);
2122
2123         if (IS_ERR(as)) {
2124                 ret = PTR_ERR(as);
2125                 if (ret == -EAGAIN)
2126                         ret = -EINTR;
2127
2128                 if (ret != -EINTR)
2129                         goto err;
2130
2131                 bch2_trans_unlock(iter->trans);
2132                 up_read(&c->gc_lock);
2133                 closure_sync(&cl);
2134                 down_read(&c->gc_lock);
2135
2136                 if (!bch2_trans_relock(iter->trans))
2137                         goto err;
2138         }
2139
2140         ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(new_key));
2141         if (ret)
2142                 goto err_free_update;
2143
2144         __bch2_btree_node_update_key(c, as, iter, b, new_hash, new_key);
2145
2146         bch2_btree_iter_downgrade(iter);
2147 err:
2148         if (new_hash) {
2149                 mutex_lock(&c->btree_cache.lock);
2150                 list_move(&new_hash->list, &c->btree_cache.freeable);
2151                 mutex_unlock(&c->btree_cache.lock);
2152
2153                 six_unlock_write(&new_hash->lock);
2154                 six_unlock_intent(&new_hash->lock);
2155         }
2156         up_read(&c->gc_lock);
2157         closure_sync(&cl);
2158         return ret;
2159 err_free_update:
2160         bch2_btree_update_free(as);
2161         goto err;
2162 }
2163
2164 /* Init code: */
2165
2166 /*
2167  * Only for filesystem bringup, when first reading the btree roots or allocating
2168  * btree roots when initializing a new filesystem:
2169  */
2170 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2171 {
2172         BUG_ON(btree_node_root(c, b));
2173
2174         __bch2_btree_set_root_inmem(c, b);
2175 }
2176
2177 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2178 {
2179         struct closure cl;
2180         struct btree *b;
2181         int ret;
2182
2183         closure_init_stack(&cl);
2184
2185         do {
2186                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2187                 closure_sync(&cl);
2188         } while (ret);
2189
2190         b = bch2_btree_node_mem_alloc(c);
2191         bch2_btree_cache_cannibalize_unlock(c);
2192
2193         set_btree_node_fake(b);
2194         b->level        = 0;
2195         b->btree_id     = id;
2196
2197         bkey_btree_ptr_init(&b->key);
2198         b->key.k.p = POS_MAX;
2199         *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2200
2201         bch2_bset_init_first(b, &b->data->keys);
2202         bch2_btree_build_aux_trees(b);
2203
2204         b->data->flags = 0;
2205         btree_set_min(b, POS_MIN);
2206         btree_set_max(b, POS_MAX);
2207         b->data->format = bch2_btree_calc_format(b);
2208         btree_node_set_format(b, b->data->format);
2209
2210         ret = bch2_btree_node_hash_insert(&c->btree_cache, b, b->level, b->btree_id);
2211         BUG_ON(ret);
2212
2213         __bch2_btree_set_root_inmem(c, b);
2214
2215         six_unlock_write(&b->lock);
2216         six_unlock_intent(&b->lock);
2217 }
2218
2219 ssize_t bch2_btree_updates_print(struct bch_fs *c, char *buf)
2220 {
2221         struct printbuf out = _PBUF(buf, PAGE_SIZE);
2222         struct btree_update *as;
2223
2224         mutex_lock(&c->btree_interior_update_lock);
2225         list_for_each_entry(as, &c->btree_interior_update_list, list)
2226                 pr_buf(&out, "%p m %u w %u r %u j %llu\n",
2227                        as,
2228                        as->mode,
2229                        as->nodes_written,
2230                        atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2231                        as->journal.seq);
2232         mutex_unlock(&c->btree_interior_update_lock);
2233
2234         return out.pos - buf;
2235 }
2236
2237 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *c)
2238 {
2239         size_t ret = 0;
2240         struct list_head *i;
2241
2242         mutex_lock(&c->btree_interior_update_lock);
2243         list_for_each(i, &c->btree_interior_update_list)
2244                 ret++;
2245         mutex_unlock(&c->btree_interior_update_lock);
2246
2247         return ret;
2248 }