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

   1
   2 #include "bcachefs.h"
   3 #include "alloc_foreground.h"
   4 #include "bkey_methods.h"
   5 #include "btree_cache.h"
   6 #include "btree_gc.h"
   7 #include "btree_update.h"
   8 #include "btree_update_interior.h"
   9 #include "btree_io.h"
  10 #include "btree_iter.h"
  11 #include "btree_locking.h"
  12 #include "buckets.h"
  13 #include "extents.h"
  14 #include "journal.h"
  15 #include "journal_reclaim.h"
  16 #include "keylist.h"
  17 #include "replicas.h"
  18 #include "super-io.h"
  19
  20 #include <linux/random.h>
  21 #include <trace/events/bcachefs.h>
  22
  23 static void btree_node_will_make_reachable(struct btree_update *,
  24                                            struct btree *);
  25 static void btree_update_drop_new_node(struct bch_fs *, struct btree *);
  26 static void bch2_btree_set_root_ondisk(struct bch_fs *, struct btree *, int);
  27
  28 /* Debug code: */
  29
  30 static void btree_node_interior_verify(struct btree *b)
  31 {
  32         struct btree_node_iter iter;
  33         struct bkey_packed *k;
  34
  35         BUG_ON(!b->level);
  36
  37         bch2_btree_node_iter_init(&iter, b, &b->key.k.p);
  38 #if 1
  39         BUG_ON(!(k = bch2_btree_node_iter_peek(&iter, b)) ||
  40                bkey_cmp_left_packed(b, k, &b->key.k.p));
  41
  42         BUG_ON((bch2_btree_node_iter_advance(&iter, b),
  43                 !bch2_btree_node_iter_end(&iter)));
  44 #else
  45         const char *msg;
  46
  47         msg = "not found";
  48         k = bch2_btree_node_iter_peek(&iter, b);
  49         if (!k)
  50                 goto err;
  51
  52         msg = "isn't what it should be";
  53         if (bkey_cmp_left_packed(b, k, &b->key.k.p))
  54                 goto err;
  55
  56         bch2_btree_node_iter_advance(&iter, b);
  57
  58         msg = "isn't last key";
  59         if (!bch2_btree_node_iter_end(&iter))
  60                 goto err;
  61         return;
  62 err:
  63         bch2_dump_btree_node(b);
  64         printk(KERN_ERR "last key %llu:%llu %s\n", b->key.k.p.inode,
  65                b->key.k.p.offset, msg);
  66         BUG();
  67 #endif
  68 }
  69
  70 /* Calculate ideal packed bkey format for new btree nodes: */
  71
  72 void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b)
  73 {
  74         struct bkey_packed *k;
  75         struct bset_tree *t;
  76         struct bkey uk;
  77
  78         bch2_bkey_format_add_pos(s, b->data->min_key);
  79
  80         for_each_bset(b, t)
  81                 for (k = btree_bkey_first(b, t);
  82                      k != btree_bkey_last(b, t);
  83                      k = bkey_next(k))
  84                         if (!bkey_whiteout(k)) {
  85                                 uk = bkey_unpack_key(b, k);
  86                                 bch2_bkey_format_add_key(s, &uk);
  87                         }
  88 }
  89
  90 static struct bkey_format bch2_btree_calc_format(struct btree *b)
  91 {
  92         struct bkey_format_state s;
  93
  94         bch2_bkey_format_init(&s);
  95         __bch2_btree_calc_format(&s, b);
  96
  97         return bch2_bkey_format_done(&s);
  98 }
  99
 100 static size_t btree_node_u64s_with_format(struct btree *b,
 101                                           struct bkey_format *new_f)
 102 {
 103         struct bkey_format *old_f = &b->format;
 104
 105         /* stupid integer promotion rules */
 106         ssize_t delta =
 107             (((int) new_f->key_u64s - old_f->key_u64s) *
 108              (int) b->nr.packed_keys) +
 109             (((int) new_f->key_u64s - BKEY_U64s) *
 110              (int) b->nr.unpacked_keys);
 111
 112         BUG_ON(delta + b->nr.live_u64s < 0);
 113
 114         return b->nr.live_u64s + delta;
 115 }
 116
 117 /**
 118  * btree_node_format_fits - check if we could rewrite node with a new format
 119  *
 120  * This assumes all keys can pack with the new format -- it just checks if
 121  * the re-packed keys would fit inside the node itself.
 122  */
 123 bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
 124                                  struct bkey_format *new_f)
 125 {
 126         size_t u64s = btree_node_u64s_with_format(b, new_f);
 127
 128         return __vstruct_bytes(struct btree_node, u64s) < btree_bytes(c);
 129 }
 130
 131 /* Btree node freeing/allocation: */
 132
 133 static bool btree_key_matches(struct bch_fs *c,
 134                               struct bkey_s_c l,
 135                               struct bkey_s_c r)
 136 {
 137         struct bkey_ptrs_c ptrs1 = bch2_bkey_ptrs_c(l);
 138         struct bkey_ptrs_c ptrs2 = bch2_bkey_ptrs_c(r);
 139         const struct bch_extent_ptr *ptr1, *ptr2;
 140
 141         bkey_for_each_ptr(ptrs1, ptr1)
 142                 bkey_for_each_ptr(ptrs2, ptr2)
 143                         if (ptr1->dev == ptr2->dev &&
 144                             ptr1->gen == ptr2->gen &&
 145                             ptr1->offset == ptr2->offset)
 146                                 return true;
 147
 148         return false;
 149 }
 150
 151 /*
 152  * We're doing the index update that makes @b unreachable, update stuff to
 153  * reflect that:
 154  *
 155  * Must be called _before_ btree_update_updated_root() or
 156  * btree_update_updated_node:
 157  */
 158 static void bch2_btree_node_free_index(struct btree_update *as, struct btree *b,
 159                                        struct bkey_s_c k,
 160                                        struct bch_fs_usage *stats)
 161 {
 162         struct bch_fs *c = as->c;
 163         struct pending_btree_node_free *d;
 164         struct gc_pos pos = { 0 };
 165
 166         for (d = as->pending; d < as->pending + as->nr_pending; d++)
 167                 if (!bkey_cmp(k.k->p, d->key.k.p) &&
 168                     btree_key_matches(c, k, bkey_i_to_s_c(&d->key)))
 169                         goto found;
 170         BUG();
 171 found:
 172         BUG_ON(d->index_update_done);
 173         d->index_update_done = true;
 174
 175         /*
 176          * We're dropping @k from the btree, but it's still live until the
 177          * index update is persistent so we need to keep a reference around for
 178          * mark and sweep to find - that's primarily what the
 179          * btree_node_pending_free list is for.
 180          *
 181          * So here (when we set index_update_done = true), we're moving an
 182          * existing reference to a different part of the larger "gc keyspace" -
 183          * and the new position comes after the old position, since GC marks
 184          * the pending free list after it walks the btree.
 185          *
 186          * If we move the reference while mark and sweep is _between_ the old
 187          * and the new position, mark and sweep will see the reference twice
 188          * and it'll get double accounted - so check for that here and subtract
 189          * to cancel out one of mark and sweep's markings if necessary:
 190          */
 191
 192         /*
 193          * bch2_mark_key() compares the current gc pos to the pos we're
 194          * moving this reference from, hence one comparison here:
 195          */
 196         if (gc_pos_cmp(c->gc_pos, b
 197                        ? gc_pos_btree_node(b)
 198                        : gc_pos_btree_root(as->btree_id)) >= 0 &&
 199             gc_pos_cmp(c->gc_pos, gc_phase(GC_PHASE_PENDING_DELETE)) < 0)
 200                 bch2_mark_key_locked(c,
 201                               bkey_i_to_s_c(&d->key),
 202                               false, 0, pos,
 203                               NULL, 0, BCH_BUCKET_MARK_GC);
 204 }
 205
 206 static void __btree_node_free(struct bch_fs *c, struct btree *b)
 207 {
 208         trace_btree_node_free(c, b);
 209
 210         BUG_ON(btree_node_dirty(b));
 211         BUG_ON(btree_node_need_write(b));
 212         BUG_ON(b == btree_node_root(c, b));
 213         BUG_ON(b->ob.nr);
 214         BUG_ON(!list_empty(&b->write_blocked));
 215         BUG_ON(b->will_make_reachable);
 216
 217         clear_btree_node_noevict(b);
 218
 219         bch2_btree_node_hash_remove(&c->btree_cache, b);
 220
 221         mutex_lock(&c->btree_cache.lock);
 222         list_move(&b->list, &c->btree_cache.freeable);
 223         mutex_unlock(&c->btree_cache.lock);
 224 }
 225
 226 void bch2_btree_node_free_never_inserted(struct bch_fs *c, struct btree *b)
 227 {
 228         struct open_buckets ob = b->ob;
 229
 230         btree_update_drop_new_node(c, b);
 231
 232         b->ob.nr = 0;
 233
 234         clear_btree_node_dirty(b);
 235
 236         btree_node_lock_type(c, b, SIX_LOCK_write);
 237         __btree_node_free(c, b);
 238         six_unlock_write(&b->lock);
 239
 240         bch2_open_buckets_put(c, &ob);
 241 }
 242
 243 void bch2_btree_node_free_inmem(struct bch_fs *c, struct btree *b,
 244                                 struct btree_iter *iter)
 245 {
 246         struct btree_iter *linked;
 247
 248         for_each_btree_iter(iter, linked)
 249                 BUG_ON(linked->l[b->level].b == b);
 250
 251         /*
 252          * Is this a node that isn't reachable on disk yet?
 253          *
 254          * Nodes that aren't reachable yet have writes blocked until they're
 255          * reachable - now that we've cancelled any pending writes and moved
 256          * things waiting on that write to wait on this update, we can drop this
 257          * node from the list of nodes that the other update is making
 258          * reachable, prior to freeing it:
 259          */
 260         btree_update_drop_new_node(c, b);
 261
 262         six_lock_write(&b->lock);
 263         __btree_node_free(c, b);
 264         six_unlock_write(&b->lock);
 265         six_unlock_intent(&b->lock);
 266 }
 267
 268 static void bch2_btree_node_free_ondisk(struct bch_fs *c,
 269                                         struct pending_btree_node_free *pending)
 270 {
 271         BUG_ON(!pending->index_update_done);
 272
 273         bch2_mark_key(c, bkey_i_to_s_c(&pending->key),
 274                       false, 0,
 275                       gc_phase(GC_PHASE_PENDING_DELETE),
 276                       NULL, 0, 0);
 277 }
 278
 279 static struct btree *__bch2_btree_node_alloc(struct bch_fs *c,
 280                                              struct disk_reservation *res,
 281                                              struct closure *cl,
 282                                              unsigned flags)
 283 {
 284         struct write_point *wp;
 285         struct btree *b;
 286         BKEY_PADDED(k) tmp;
 287         struct open_buckets ob = { .nr = 0 };
 288         struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
 289         unsigned nr_reserve;
 290         enum alloc_reserve alloc_reserve;
 291
 292         if (flags & BTREE_INSERT_USE_ALLOC_RESERVE) {
 293                 nr_reserve      = 0;
 294                 alloc_reserve   = RESERVE_ALLOC;
 295         } else if (flags & BTREE_INSERT_USE_RESERVE) {
 296                 nr_reserve      = BTREE_NODE_RESERVE / 2;
 297                 alloc_reserve   = RESERVE_BTREE;
 298         } else {
 299                 nr_reserve      = BTREE_NODE_RESERVE;
 300                 alloc_reserve   = RESERVE_NONE;
 301         }
 302
 303         mutex_lock(&c->btree_reserve_cache_lock);
 304         if (c->btree_reserve_cache_nr > nr_reserve) {
 305                 struct btree_alloc *a =
 306                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
 307
 308                 ob = a->ob;
 309                 bkey_copy(&tmp.k, &a->k);
 310                 mutex_unlock(&c->btree_reserve_cache_lock);
 311                 goto mem_alloc;
 312         }
 313         mutex_unlock(&c->btree_reserve_cache_lock);
 314
 315 retry:
 316         wp = bch2_alloc_sectors_start(c, c->opts.foreground_target, 0,
 317                                       writepoint_ptr(&c->btree_write_point),
 318                                       &devs_have,
 319                                       res->nr_replicas,
 320                                       c->opts.metadata_replicas_required,
 321                                       alloc_reserve, 0, cl);
 322         if (IS_ERR(wp))
 323                 return ERR_CAST(wp);
 324
 325         if (wp->sectors_free < c->opts.btree_node_size) {
 326                 struct open_bucket *ob;
 327                 unsigned i;
 328
 329                 open_bucket_for_each(c, &wp->ptrs, ob, i)
 330                         if (ob->sectors_free < c->opts.btree_node_size)
 331                                 ob->sectors_free = 0;
 332
 333                 bch2_alloc_sectors_done(c, wp);
 334                 goto retry;
 335         }
 336
 337         bkey_btree_ptr_init(&tmp.k);
 338         bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, c->opts.btree_node_size);
 339
 340         bch2_open_bucket_get(c, wp, &ob);
 341         bch2_alloc_sectors_done(c, wp);
 342 mem_alloc:
 343         b = bch2_btree_node_mem_alloc(c);
 344
 345         /* we hold cannibalize_lock: */
 346         BUG_ON(IS_ERR(b));
 347         BUG_ON(b->ob.nr);
 348
 349         bkey_copy(&b->key, &tmp.k);
 350         b->ob = ob;
 351
 352         return b;
 353 }
 354
 355 static struct btree *bch2_btree_node_alloc(struct btree_update *as, unsigned level)
 356 {
 357         struct bch_fs *c = as->c;
 358         struct btree *b;
 359
 360         BUG_ON(level >= BTREE_MAX_DEPTH);
 361         BUG_ON(!as->reserve->nr);
 362
 363         b = as->reserve->b[--as->reserve->nr];
 364
 365         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id));
 366
 367         set_btree_node_accessed(b);
 368         set_btree_node_dirty(b);
 369         set_btree_node_need_write(b);
 370
 371         bch2_bset_init_first(b, &b->data->keys);
 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 = bkey_i_to_btree_ptr(&b->key)->v.start[0];
 378
 379         bch2_btree_build_aux_trees(b);
 380
 381         btree_node_will_make_reachable(as, b);
 382
 383         trace_btree_node_alloc(c, b);
 384         return b;
 385 }
 386
 387 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *as,
 388                                                   struct btree *b,
 389                                                   struct bkey_format format)
 390 {
 391         struct btree *n;
 392
 393         n = bch2_btree_node_alloc(as, b->level);
 394
 395         n->data->min_key        = b->data->min_key;
 396         n->data->max_key        = b->data->max_key;
 397         n->data->format         = format;
 398         SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
 399
 400         btree_node_set_format(n, format);
 401
 402         bch2_btree_sort_into(as->c, n, b);
 403
 404         btree_node_reset_sib_u64s(n);
 405
 406         n->key.k.p = b->key.k.p;
 407         return n;
 408 }
 409
 410 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
 411                                                        struct btree *b)
 412 {
 413         struct bkey_format new_f = bch2_btree_calc_format(b);
 414
 415         /*
 416          * The keys might expand with the new format - if they wouldn't fit in
 417          * the btree node anymore, use the old format for now:
 418          */
 419         if (!bch2_btree_node_format_fits(as->c, b, &new_f))
 420                 new_f = b->format;
 421
 422         return __bch2_btree_node_alloc_replacement(as, b, new_f);
 423 }
 424
 425 static struct btree *__btree_root_alloc(struct btree_update *as, unsigned level)
 426 {
 427         struct btree *b = bch2_btree_node_alloc(as, level);
 428
 429         b->data->min_key = POS_MIN;
 430         b->data->max_key = POS_MAX;
 431         b->data->format = bch2_btree_calc_format(b);
 432         b->key.k.p = POS_MAX;
 433
 434         btree_node_set_format(b, b->data->format);
 435         bch2_btree_build_aux_trees(b);
 436
 437         six_unlock_write(&b->lock);
 438
 439         return b;
 440 }
 441
 442 static void bch2_btree_reserve_put(struct bch_fs *c, struct btree_reserve *reserve)
 443 {
 444         bch2_disk_reservation_put(c, &reserve->disk_res);
 445
 446         mutex_lock(&c->btree_reserve_cache_lock);
 447
 448         while (reserve->nr) {
 449                 struct btree *b = reserve->b[--reserve->nr];
 450
 451                 six_unlock_write(&b->lock);
 452
 453                 if (c->btree_reserve_cache_nr <
 454                     ARRAY_SIZE(c->btree_reserve_cache)) {
 455                         struct btree_alloc *a =
 456                                 &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
 457
 458                         a->ob = b->ob;
 459                         b->ob.nr = 0;
 460                         bkey_copy(&a->k, &b->key);
 461                 } else {
 462                         bch2_open_buckets_put(c, &b->ob);
 463                 }
 464
 465                 btree_node_lock_type(c, b, SIX_LOCK_write);
 466                 __btree_node_free(c, b);
 467                 six_unlock_write(&b->lock);
 468
 469                 six_unlock_intent(&b->lock);
 470         }
 471
 472         mutex_unlock(&c->btree_reserve_cache_lock);
 473
 474         mempool_free(reserve, &c->btree_reserve_pool);
 475 }
 476
 477 static struct btree_reserve *bch2_btree_reserve_get(struct bch_fs *c,
 478                                                     unsigned nr_nodes,
 479                                                     unsigned flags,
 480                                                     struct closure *cl)
 481 {
 482         struct btree_reserve *reserve;
 483         struct btree *b;
 484         struct disk_reservation disk_res = { 0, 0 };
 485         unsigned sectors = nr_nodes * c->opts.btree_node_size;
 486         int ret, disk_res_flags = BCH_DISK_RESERVATION_GC_LOCK_HELD;
 487
 488         if (flags & BTREE_INSERT_NOFAIL)
 489                 disk_res_flags |= BCH_DISK_RESERVATION_NOFAIL;
 490
 491         /*
 492          * This check isn't necessary for correctness - it's just to potentially
 493          * prevent us from doing a lot of work that'll end up being wasted:
 494          */
 495         ret = bch2_journal_error(&c->journal);
 496         if (ret)
 497                 return ERR_PTR(ret);
 498
 499         if (bch2_disk_reservation_get(c, &disk_res, sectors,
 500                                       c->opts.metadata_replicas,
 501                                       disk_res_flags))
 502                 return ERR_PTR(-ENOSPC);
 503
 504         BUG_ON(nr_nodes > BTREE_RESERVE_MAX);
 505
 506         /*
 507          * Protects reaping from the btree node cache and using the btree node
 508          * open bucket reserve:
 509          */
 510         ret = bch2_btree_cache_cannibalize_lock(c, cl);
 511         if (ret) {
 512                 bch2_disk_reservation_put(c, &disk_res);
 513                 return ERR_PTR(ret);
 514         }
 515
 516         reserve = mempool_alloc(&c->btree_reserve_pool, GFP_NOIO);
 517
 518         reserve->disk_res = disk_res;
 519         reserve->nr = 0;
 520
 521         while (reserve->nr < nr_nodes) {
 522                 b = __bch2_btree_node_alloc(c, &disk_res,
 523                                             flags & BTREE_INSERT_NOWAIT
 524                                             ? NULL : cl, flags);
 525                 if (IS_ERR(b)) {
 526                         ret = PTR_ERR(b);
 527                         goto err_free;
 528                 }
 529
 530                 ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(&b->key));
 531                 if (ret)
 532                         goto err_free;
 533
 534                 reserve->b[reserve->nr++] = b;
 535         }
 536
 537         bch2_btree_cache_cannibalize_unlock(c);
 538         return reserve;
 539 err_free:
 540         bch2_btree_reserve_put(c, reserve);
 541         bch2_btree_cache_cannibalize_unlock(c);
 542         trace_btree_reserve_get_fail(c, nr_nodes, cl);
 543         return ERR_PTR(ret);
 544 }
 545
 546 /* Asynchronous interior node update machinery */
 547
 548 static void bch2_btree_update_free(struct btree_update *as)
 549 {
 550         struct bch_fs *c = as->c;
 551
 552         bch2_journal_pin_flush(&c->journal, &as->journal);
 553
 554         BUG_ON(as->nr_new_nodes);
 555         BUG_ON(as->nr_pending);
 556
 557         if (as->reserve)
 558                 bch2_btree_reserve_put(c, as->reserve);
 559
 560         mutex_lock(&c->btree_interior_update_lock);
 561         list_del(&as->list);
 562
 563         closure_debug_destroy(&as->cl);
 564         mempool_free(as, &c->btree_interior_update_pool);
 565
 566         closure_wake_up(&c->btree_interior_update_wait);
 567         mutex_unlock(&c->btree_interior_update_lock);
 568 }
 569
 570 static void btree_update_nodes_reachable(struct closure *cl)
 571 {
 572         struct btree_update *as = container_of(cl, struct btree_update, cl);
 573         struct bch_fs *c = as->c;
 574
 575         bch2_journal_pin_drop(&c->journal, &as->journal);
 576
 577         mutex_lock(&c->btree_interior_update_lock);
 578
 579         while (as->nr_new_nodes) {
 580                 struct btree *b = as->new_nodes[--as->nr_new_nodes];
 581
 582                 BUG_ON(b->will_make_reachable != (unsigned long) as);
 583                 b->will_make_reachable = 0;
 584                 mutex_unlock(&c->btree_interior_update_lock);
 585
 586                 /*
 587                  * b->will_make_reachable prevented it from being written, so
 588                  * write it now if it needs to be written:
 589                  */
 590                 btree_node_lock_type(c, b, SIX_LOCK_read);
 591                 bch2_btree_node_write_cond(c, b, btree_node_need_write(b));
 592                 six_unlock_read(&b->lock);
 593                 mutex_lock(&c->btree_interior_update_lock);
 594         }
 595
 596         while (as->nr_pending)
 597                 bch2_btree_node_free_ondisk(c, &as->pending[--as->nr_pending]);
 598
 599         mutex_unlock(&c->btree_interior_update_lock);
 600
 601         closure_wake_up(&as->wait);
 602
 603         bch2_btree_update_free(as);
 604 }
 605
 606 static void btree_update_wait_on_journal(struct closure *cl)
 607 {
 608         struct btree_update *as = container_of(cl, struct btree_update, cl);
 609         struct bch_fs *c = as->c;
 610         int ret;
 611
 612         ret = bch2_journal_open_seq_async(&c->journal, as->journal_seq, cl);
 613         if (ret == -EAGAIN) {
 614                 continue_at(cl, btree_update_wait_on_journal, system_wq);
 615                 return;
 616         }
 617         if (ret < 0)
 618                 goto err;
 619
 620         bch2_journal_flush_seq_async(&c->journal, as->journal_seq, cl);
 621 err:
 622         continue_at(cl, btree_update_nodes_reachable, system_wq);
 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 bch_fs *c = as->c;
 629         struct btree *b;
 630
 631         /*
 632          * We did an update to a parent node where the pointers we added pointed
 633          * to child nodes that weren't written yet: now, the child nodes have
 634          * been written so we can write out the update to the interior node.
 635          */
 636 retry:
 637         mutex_lock(&c->btree_interior_update_lock);
 638         as->nodes_written = true;
 639
 640         switch (as->mode) {
 641         case BTREE_INTERIOR_NO_UPDATE:
 642                 BUG();
 643         case BTREE_INTERIOR_UPDATING_NODE:
 644                 /* The usual case: */
 645                 b = READ_ONCE(as->b);
 646
 647                 if (!six_trylock_read(&b->lock)) {
 648                         mutex_unlock(&c->btree_interior_update_lock);
 649                         btree_node_lock_type(c, b, SIX_LOCK_read);
 650                         six_unlock_read(&b->lock);
 651                         goto retry;
 652                 }
 653
 654                 BUG_ON(!btree_node_dirty(b));
 655                 closure_wait(&btree_current_write(b)->wait, cl);
 656
 657                 list_del(&as->write_blocked_list);
 658
 659                 /*
 660                  * for flush_held_btree_writes() waiting on updates to flush or
 661                  * nodes to be writeable:
 662                  */
 663                 closure_wake_up(&c->btree_interior_update_wait);
 664                 mutex_unlock(&c->btree_interior_update_lock);
 665
 666                 /*
 667                  * b->write_blocked prevented it from being written, so
 668                  * write it now if it needs to be written:
 669                  */
 670                 bch2_btree_node_write_cond(c, b, true);
 671                 six_unlock_read(&b->lock);
 672                 break;
 673
 674         case BTREE_INTERIOR_UPDATING_AS:
 675                 /*
 676                  * The btree node we originally updated has been freed and is
 677                  * being rewritten - so we need to write anything here, we just
 678                  * need to signal to that btree_update that it's ok to make the
 679                  * new replacement node visible:
 680                  */
 681                 closure_put(&as->parent_as->cl);
 682
 683                 /*
 684                  * and then we have to wait on that btree_update to finish:
 685                  */
 686                 closure_wait(&as->parent_as->wait, cl);
 687                 mutex_unlock(&c->btree_interior_update_lock);
 688                 break;
 689
 690         case BTREE_INTERIOR_UPDATING_ROOT:
 691                 /* b is the new btree root: */
 692                 b = READ_ONCE(as->b);
 693
 694                 if (!six_trylock_read(&b->lock)) {
 695                         mutex_unlock(&c->btree_interior_update_lock);
 696                         btree_node_lock_type(c, b, SIX_LOCK_read);
 697                         six_unlock_read(&b->lock);
 698                         goto retry;
 699                 }
 700
 701                 BUG_ON(c->btree_roots[b->btree_id].as != as);
 702                 c->btree_roots[b->btree_id].as = NULL;
 703
 704                 bch2_btree_set_root_ondisk(c, b, WRITE);
 705
 706                 /*
 707                  * We don't have to wait anything anything here (before
 708                  * btree_update_nodes_reachable frees the old nodes
 709                  * ondisk) - we've ensured that the very next journal write will
 710                  * have the pointer to the new root, and before the allocator
 711                  * can reuse the old nodes it'll have to do a journal commit:
 712                  */
 713                 six_unlock_read(&b->lock);
 714                 mutex_unlock(&c->btree_interior_update_lock);
 715
 716                 /*
 717                  * Bit of funny circularity going on here we have to break:
 718                  *
 719                  * We have to drop our journal pin before writing the journal
 720                  * entry that points to the new btree root: else, we could
 721                  * deadlock if the journal currently happens to be full.
 722                  *
 723                  * This mean we're dropping the journal pin _before_ the new
 724                  * nodes are technically reachable - but this is safe, because
 725                  * after the bch2_btree_set_root_ondisk() call above they will
 726                  * be reachable as of the very next journal write:
 727                  */
 728                 bch2_journal_pin_drop(&c->journal, &as->journal);
 729
 730                 as->journal_seq = bch2_journal_last_unwritten_seq(&c->journal);
 731
 732                 btree_update_wait_on_journal(cl);
 733                 return;
 734         }
 735
 736         continue_at(cl, btree_update_nodes_reachable, system_wq);
 737 }
 738
 739 /*
 740  * We're updating @b with pointers to nodes that haven't finished writing yet:
 741  * block @b from being written until @as completes
 742  */
 743 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
 744 {
 745         struct bch_fs *c = as->c;
 746
 747         mutex_lock(&c->btree_interior_update_lock);
 748
 749         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 750         BUG_ON(!btree_node_dirty(b));
 751
 752         as->mode = BTREE_INTERIOR_UPDATING_NODE;
 753         as->b = b;
 754         list_add(&as->write_blocked_list, &b->write_blocked);
 755
 756         mutex_unlock(&c->btree_interior_update_lock);
 757
 758         /*
 759          * In general, when you're staging things in a journal that will later
 760          * be written elsewhere, and you also want to guarantee ordering: that
 761          * is, if you have updates a, b, c, after a crash you should never see c
 762          * and not a or b - there's a problem:
 763          *
 764          * If the final destination of the update(s) (i.e. btree node) can be
 765          * written/flushed _before_ the relevant journal entry - oops, that
 766          * breaks ordering, since the various leaf nodes can be written in any
 767          * order.
 768          *
 769          * Normally we use bset->journal_seq to deal with this - if during
 770          * recovery we find a btree node write that's newer than the newest
 771          * journal entry, we just ignore it - we don't need it, anything we're
 772          * supposed to have (that we reported as completed via fsync()) will
 773          * still be in the journal, and as far as the state of the journal is
 774          * concerned that btree node write never happened.
 775          *
 776          * That breaks when we're rewriting/splitting/merging nodes, since we're
 777          * mixing btree node writes that haven't happened yet with previously
 778          * written data that has been reported as completed to the journal.
 779          *
 780          * Thus, before making the new nodes reachable, we have to wait the
 781          * newest journal sequence number we have data for to be written (if it
 782          * hasn't been yet).
 783          */
 784         bch2_journal_wait_on_seq(&c->journal, as->journal_seq, &as->cl);
 785 }
 786
 787 static void interior_update_flush(struct journal *j,
 788                         struct journal_entry_pin *pin, u64 seq)
 789 {
 790         struct btree_update *as =
 791                 container_of(pin, struct btree_update, journal);
 792
 793         bch2_journal_flush_seq_async(j, as->journal_seq, NULL);
 794 }
 795
 796 static void btree_update_reparent(struct btree_update *as,
 797                                   struct btree_update *child)
 798 {
 799         struct bch_fs *c = as->c;
 800
 801         child->b = NULL;
 802         child->mode = BTREE_INTERIOR_UPDATING_AS;
 803         child->parent_as = as;
 804         closure_get(&as->cl);
 805
 806         /*
 807          * When we write a new btree root, we have to drop our journal pin
 808          * _before_ the new nodes are technically reachable; see
 809          * btree_update_nodes_written().
 810          *
 811          * This goes for journal pins that are recursively blocked on us - so,
 812          * just transfer the journal pin to the new interior update so
 813          * btree_update_nodes_written() can drop it.
 814          */
 815         bch2_journal_pin_add_if_older(&c->journal, &child->journal,
 816                                       &as->journal, interior_update_flush);
 817         bch2_journal_pin_drop(&c->journal, &child->journal);
 818
 819         as->journal_seq = max(as->journal_seq, child->journal_seq);
 820 }
 821
 822 static void btree_update_updated_root(struct btree_update *as)
 823 {
 824         struct bch_fs *c = as->c;
 825         struct btree_root *r = &c->btree_roots[as->btree_id];
 826
 827         mutex_lock(&c->btree_interior_update_lock);
 828
 829         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 830
 831         /*
 832          * Old root might not be persistent yet - if so, redirect its
 833          * btree_update operation to point to us:
 834          */
 835         if (r->as)
 836                 btree_update_reparent(as, r->as);
 837
 838         as->mode = BTREE_INTERIOR_UPDATING_ROOT;
 839         as->b = r->b;
 840         r->as = as;
 841
 842         mutex_unlock(&c->btree_interior_update_lock);
 843
 844         /*
 845          * When we're rewriting nodes and updating interior nodes, there's an
 846          * issue with updates that haven't been written in the journal getting
 847          * mixed together with older data - see btree_update_updated_node()
 848          * for the explanation.
 849          *
 850          * However, this doesn't affect us when we're writing a new btree root -
 851          * because to make that new root reachable we have to write out a new
 852          * journal entry, which must necessarily be newer than as->journal_seq.
 853          */
 854 }
 855
 856 static void btree_node_will_make_reachable(struct btree_update *as,
 857                                            struct btree *b)
 858 {
 859         struct bch_fs *c = as->c;
 860
 861         mutex_lock(&c->btree_interior_update_lock);
 862         BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
 863         BUG_ON(b->will_make_reachable);
 864
 865         as->new_nodes[as->nr_new_nodes++] = b;
 866         b->will_make_reachable = 1UL|(unsigned long) as;
 867
 868         closure_get(&as->cl);
 869         mutex_unlock(&c->btree_interior_update_lock);
 870 }
 871
 872 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
 873 {
 874         struct btree_update *as;
 875         unsigned long v;
 876         unsigned i;
 877
 878         mutex_lock(&c->btree_interior_update_lock);
 879         v = xchg(&b->will_make_reachable, 0);
 880         as = (struct btree_update *) (v & ~1UL);
 881
 882         if (!as) {
 883                 mutex_unlock(&c->btree_interior_update_lock);
 884                 return;
 885         }
 886
 887         for (i = 0; i < as->nr_new_nodes; i++)
 888                 if (as->new_nodes[i] == b)
 889                         goto found;
 890
 891         BUG();
 892 found:
 893         array_remove_item(as->new_nodes, as->nr_new_nodes, i);
 894         mutex_unlock(&c->btree_interior_update_lock);
 895
 896         if (v & 1)
 897                 closure_put(&as->cl);
 898 }
 899
 900 static void btree_interior_update_add_node_reference(struct btree_update *as,
 901                                                      struct btree *b)
 902 {
 903         struct bch_fs *c = as->c;
 904         struct pending_btree_node_free *d;
 905
 906         mutex_lock(&c->btree_interior_update_lock);
 907
 908         /* Add this node to the list of nodes being freed: */
 909         BUG_ON(as->nr_pending >= ARRAY_SIZE(as->pending));
 910
 911         d = &as->pending[as->nr_pending++];
 912         d->index_update_done    = false;
 913         d->seq                  = b->data->keys.seq;
 914         d->btree_id             = b->btree_id;
 915         d->level                = b->level;
 916         bkey_copy(&d->key, &b->key);
 917
 918         mutex_unlock(&c->btree_interior_update_lock);
 919 }
 920
 921 /*
 922  * @b is being split/rewritten: it may have pointers to not-yet-written btree
 923  * nodes and thus outstanding btree_updates - redirect @b's
 924  * btree_updates to point to this btree_update:
 925  */
 926 void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 927                                                struct btree *b)
 928 {
 929         struct bch_fs *c = as->c;
 930         struct closure *cl, *cl_n;
 931         struct btree_update *p, *n;
 932         struct btree_write *w;
 933         struct bset_tree *t;
 934
 935         set_btree_node_dying(b);
 936
 937         if (btree_node_fake(b))
 938                 return;
 939
 940         btree_interior_update_add_node_reference(as, b);
 941
 942         /*
 943          * Does this node have data that hasn't been written in the journal?
 944          *
 945          * If so, we have to wait for the corresponding journal entry to be
 946          * written before making the new nodes reachable - we can't just carry
 947          * over the bset->journal_seq tracking, since we'll be mixing those keys
 948          * in with keys that aren't in the journal anymore:
 949          */
 950         for_each_bset(b, t)
 951                 as->journal_seq = max(as->journal_seq,
 952                                       le64_to_cpu(bset(b, t)->journal_seq));
 953
 954         mutex_lock(&c->btree_interior_update_lock);
 955
 956         /*
 957          * Does this node have any btree_update operations preventing
 958          * it from being written?
 959          *
 960          * If so, redirect them to point to this btree_update: we can
 961          * write out our new nodes, but we won't make them visible until those
 962          * operations complete
 963          */
 964         list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
 965                 list_del(&p->write_blocked_list);
 966                 btree_update_reparent(as, p);
 967
 968                 /*
 969                  * for flush_held_btree_writes() waiting on updates to flush or
 970                  * nodes to be writeable:
 971                  */
 972                 closure_wake_up(&c->btree_interior_update_wait);
 973         }
 974
 975         clear_btree_node_dirty(b);
 976         clear_btree_node_need_write(b);
 977         w = btree_current_write(b);
 978
 979         /*
 980          * Does this node have any btree_update operations waiting on this node
 981          * to be written?
 982          *
 983          * If so, wake them up when this btree_update operation is reachable:
 984          */
 985         llist_for_each_entry_safe(cl, cl_n, llist_del_all(&w->wait.list), list)
 986                 llist_add(&cl->list, &as->wait.list);
 987
 988         /*
 989          * Does this node have unwritten data that has a pin on the journal?
 990          *
 991          * If so, transfer that pin to the btree_update operation -
 992          * note that if we're freeing multiple nodes, we only need to keep the
 993          * oldest pin of any of the nodes we're freeing. We'll release the pin
 994          * when the new nodes are persistent and reachable on disk:
 995          */
 996         bch2_journal_pin_add_if_older(&c->journal, &w->journal,
 997                                       &as->journal, interior_update_flush);
 998         bch2_journal_pin_drop(&c->journal, &w->journal);
 999
1000         w = btree_prev_write(b);
1001         bch2_journal_pin_add_if_older(&c->journal, &w->journal,
1002                                       &as->journal, interior_update_flush);
1003         bch2_journal_pin_drop(&c->journal, &w->journal);
1004
1005         mutex_unlock(&c->btree_interior_update_lock);
1006 }
1007
1008 void bch2_btree_update_done(struct btree_update *as)
1009 {
1010         BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
1011
1012         bch2_btree_reserve_put(as->c, as->reserve);
1013         as->reserve = NULL;
1014
1015         continue_at(&as->cl, btree_update_nodes_written, system_freezable_wq);
1016 }
1017
1018 struct btree_update *
1019 bch2_btree_update_start(struct bch_fs *c, enum btree_id id,
1020                         unsigned nr_nodes, unsigned flags,
1021                         struct closure *cl)
1022 {
1023         struct btree_reserve *reserve;
1024         struct btree_update *as;
1025
1026         reserve = bch2_btree_reserve_get(c, nr_nodes, flags, cl);
1027         if (IS_ERR(reserve))
1028                 return ERR_CAST(reserve);
1029
1030         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
1031         memset(as, 0, sizeof(*as));
1032         closure_init(&as->cl, NULL);
1033         as->c           = c;
1034         as->mode        = BTREE_INTERIOR_NO_UPDATE;
1035         as->btree_id    = id;
1036         as->reserve     = reserve;
1037         INIT_LIST_HEAD(&as->write_blocked_list);
1038
1039         bch2_keylist_init(&as->parent_keys, as->inline_keys);
1040
1041         mutex_lock(&c->btree_interior_update_lock);
1042         list_add_tail(&as->list, &c->btree_interior_update_list);
1043         mutex_unlock(&c->btree_interior_update_lock);
1044
1045         return as;
1046 }
1047
1048 /* Btree root updates: */
1049
1050 static void __bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1051 {
1052         /* Root nodes cannot be reaped */
1053         mutex_lock(&c->btree_cache.lock);
1054         list_del_init(&b->list);
1055         mutex_unlock(&c->btree_cache.lock);
1056
1057         mutex_lock(&c->btree_root_lock);
1058         BUG_ON(btree_node_root(c, b) &&
1059                (b->level < btree_node_root(c, b)->level ||
1060                 !btree_node_dying(btree_node_root(c, b))));
1061
1062         btree_node_root(c, b) = b;
1063         mutex_unlock(&c->btree_root_lock);
1064
1065         bch2_recalc_btree_reserve(c);
1066 }
1067
1068 static void bch2_btree_set_root_inmem(struct btree_update *as, struct btree *b)
1069 {
1070         struct bch_fs *c = as->c;
1071         struct btree *old = btree_node_root(c, b);
1072         struct bch_fs_usage *fs_usage;
1073
1074         __bch2_btree_set_root_inmem(c, b);
1075
1076         mutex_lock(&c->btree_interior_update_lock);
1077         percpu_down_read_preempt_disable(&c->mark_lock);
1078         fs_usage = bch2_fs_usage_get_scratch(c);
1079
1080         bch2_mark_key_locked(c, bkey_i_to_s_c(&b->key),
1081                       true, 0,
1082                       gc_pos_btree_root(b->btree_id),
1083                       fs_usage, 0, 0);
1084
1085         if (old && !btree_node_fake(old))
1086                 bch2_btree_node_free_index(as, NULL,
1087                                            bkey_i_to_s_c(&old->key),
1088                                            fs_usage);
1089         bch2_fs_usage_apply(c, fs_usage, &as->reserve->disk_res,
1090                             gc_pos_btree_root(b->btree_id));
1091
1092         percpu_up_read_preempt_enable(&c->mark_lock);
1093         mutex_unlock(&c->btree_interior_update_lock);
1094 }
1095
1096 static void bch2_btree_set_root_ondisk(struct bch_fs *c, struct btree *b, int rw)
1097 {
1098         struct btree_root *r = &c->btree_roots[b->btree_id];
1099
1100         mutex_lock(&c->btree_root_lock);
1101
1102         BUG_ON(b != r->b);
1103         bkey_copy(&r->key, &b->key);
1104         r->level = b->level;
1105         r->alive = true;
1106         if (rw == WRITE)
1107                 c->btree_roots_dirty = true;
1108
1109         mutex_unlock(&c->btree_root_lock);
1110 }
1111
1112 /**
1113  * bch_btree_set_root - update the root in memory and on disk
1114  *
1115  * To ensure forward progress, the current task must not be holding any
1116  * btree node write locks. However, you must hold an intent lock on the
1117  * old root.
1118  *
1119  * Note: This allocates a journal entry but doesn't add any keys to
1120  * it.  All the btree roots are part of every journal write, so there
1121  * is nothing new to be done.  This just guarantees that there is a
1122  * journal write.
1123  */
1124 static void bch2_btree_set_root(struct btree_update *as, struct btree *b,
1125                                 struct btree_iter *iter)
1126 {
1127         struct bch_fs *c = as->c;
1128         struct btree *old;
1129
1130         trace_btree_set_root(c, b);
1131         BUG_ON(!b->written &&
1132                !test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags));
1133
1134         old = btree_node_root(c, b);
1135
1136         /*
1137          * Ensure no one is using the old root while we switch to the
1138          * new root:
1139          */
1140         bch2_btree_node_lock_write(old, iter);
1141
1142         bch2_btree_set_root_inmem(as, b);
1143
1144         btree_update_updated_root(as);
1145
1146         /*
1147          * Unlock old root after new root is visible:
1148          *
1149          * The new root isn't persistent, but that's ok: we still have
1150          * an intent lock on the new root, and any updates that would
1151          * depend on the new root would have to update the new root.
1152          */
1153         bch2_btree_node_unlock_write(old, iter);
1154 }
1155
1156 /* Interior node updates: */
1157
1158 static void bch2_insert_fixup_btree_ptr(struct btree_update *as, struct btree *b,
1159                                         struct btree_iter *iter,
1160                                         struct bkey_i *insert,
1161                                         struct btree_node_iter *node_iter)
1162 {
1163         struct bch_fs *c = as->c;
1164         struct bch_fs_usage *fs_usage;
1165         struct bkey_packed *k;
1166         struct bkey tmp;
1167
1168         BUG_ON(insert->k.u64s > bch_btree_keys_u64s_remaining(c, b));
1169
1170         mutex_lock(&c->btree_interior_update_lock);
1171         percpu_down_read_preempt_disable(&c->mark_lock);
1172         fs_usage = bch2_fs_usage_get_scratch(c);
1173
1174         bch2_mark_key_locked(c, bkey_i_to_s_c(insert),
1175                              true, 0,
1176                              gc_pos_btree_node(b), fs_usage, 0, 0);
1177
1178         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1179                bkey_iter_pos_cmp(b, &insert->k.p, k) > 0)
1180                 bch2_btree_node_iter_advance(node_iter, b);
1181
1182         /*
1183          * If we're overwriting, look up pending delete and mark so that gc
1184          * marks it on the pending delete list:
1185          */
1186         if (k && !bkey_cmp_packed(b, k, &insert->k))
1187                 bch2_btree_node_free_index(as, b,
1188                                            bkey_disassemble(b, k, &tmp),
1189                                            fs_usage);
1190
1191         bch2_fs_usage_apply(c, fs_usage, &as->reserve->disk_res,
1192                             gc_pos_btree_node(b));
1193
1194         percpu_up_read_preempt_enable(&c->mark_lock);
1195         mutex_unlock(&c->btree_interior_update_lock);
1196
1197         bch2_btree_bset_insert_key(iter, b, node_iter, insert);
1198         set_btree_node_dirty(b);
1199         set_btree_node_need_write(b);
1200 }
1201
1202 /*
1203  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1204  * node)
1205  */
1206 static struct btree *__btree_split_node(struct btree_update *as,
1207                                         struct btree *n1,
1208                                         struct btree_iter *iter)
1209 {
1210         size_t nr_packed = 0, nr_unpacked = 0;
1211         struct btree *n2;
1212         struct bset *set1, *set2;
1213         struct bkey_packed *k, *prev = NULL;
1214
1215         n2 = bch2_btree_node_alloc(as, n1->level);
1216
1217         n2->data->max_key       = n1->data->max_key;
1218         n2->data->format        = n1->format;
1219         SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data));
1220         n2->key.k.p = n1->key.k.p;
1221
1222         btree_node_set_format(n2, n2->data->format);
1223
1224         set1 = btree_bset_first(n1);
1225         set2 = btree_bset_first(n2);
1226
1227         /*
1228          * Has to be a linear search because we don't have an auxiliary
1229          * search tree yet
1230          */
1231         k = set1->start;
1232         while (1) {
1233                 if (bkey_next(k) == vstruct_last(set1))
1234                         break;
1235                 if (k->_data - set1->_data >= (le16_to_cpu(set1->u64s) * 3) / 5)
1236                         break;
1237
1238                 if (bkey_packed(k))
1239                         nr_packed++;
1240                 else
1241                         nr_unpacked++;
1242
1243                 prev = k;
1244                 k = bkey_next(k);
1245         }
1246
1247         BUG_ON(!prev);
1248
1249         n1->key.k.p = bkey_unpack_pos(n1, prev);
1250         n1->data->max_key = n1->key.k.p;
1251         n2->data->min_key =
1252                 btree_type_successor(n1->btree_id, n1->key.k.p);
1253
1254         set2->u64s = cpu_to_le16((u64 *) vstruct_end(set1) - (u64 *) k);
1255         set1->u64s = cpu_to_le16(le16_to_cpu(set1->u64s) - le16_to_cpu(set2->u64s));
1256
1257         set_btree_bset_end(n1, n1->set);
1258         set_btree_bset_end(n2, n2->set);
1259
1260         n2->nr.live_u64s        = le16_to_cpu(set2->u64s);
1261         n2->nr.bset_u64s[0]     = le16_to_cpu(set2->u64s);
1262         n2->nr.packed_keys      = n1->nr.packed_keys - nr_packed;
1263         n2->nr.unpacked_keys    = n1->nr.unpacked_keys - nr_unpacked;
1264
1265         n1->nr.live_u64s        = le16_to_cpu(set1->u64s);
1266         n1->nr.bset_u64s[0]     = le16_to_cpu(set1->u64s);
1267         n1->nr.packed_keys      = nr_packed;
1268         n1->nr.unpacked_keys    = nr_unpacked;
1269
1270         BUG_ON(!set1->u64s);
1271         BUG_ON(!set2->u64s);
1272
1273         memcpy_u64s(set2->start,
1274                     vstruct_end(set1),
1275                     le16_to_cpu(set2->u64s));
1276
1277         btree_node_reset_sib_u64s(n1);
1278         btree_node_reset_sib_u64s(n2);
1279
1280         bch2_verify_btree_nr_keys(n1);
1281         bch2_verify_btree_nr_keys(n2);
1282
1283         if (n1->level) {
1284                 btree_node_interior_verify(n1);
1285                 btree_node_interior_verify(n2);
1286         }
1287
1288         return n2;
1289 }
1290
1291 /*
1292  * For updates to interior nodes, we've got to do the insert before we split
1293  * because the stuff we're inserting has to be inserted atomically. Post split,
1294  * the keys might have to go in different nodes and the split would no longer be
1295  * atomic.
1296  *
1297  * Worse, if the insert is from btree node coalescing, if we do the insert after
1298  * we do the split (and pick the pivot) - the pivot we pick might be between
1299  * nodes that were coalesced, and thus in the middle of a child node post
1300  * coalescing:
1301  */
1302 static void btree_split_insert_keys(struct btree_update *as, struct btree *b,
1303                                     struct btree_iter *iter,
1304                                     struct keylist *keys)
1305 {
1306         struct btree_node_iter node_iter;
1307         struct bkey_i *k = bch2_keylist_front(keys);
1308         struct bkey_packed *p;
1309         struct bset *i;
1310
1311         BUG_ON(btree_node_type(b) != BKEY_TYPE_BTREE);
1312
1313         bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
1314
1315         while (!bch2_keylist_empty(keys)) {
1316                 k = bch2_keylist_front(keys);
1317
1318                 BUG_ON(bch_keylist_u64s(keys) >
1319                        bch_btree_keys_u64s_remaining(as->c, b));
1320                 BUG_ON(bkey_cmp(k->k.p, b->data->min_key) < 0);
1321                 BUG_ON(bkey_cmp(k->k.p, b->data->max_key) > 0);
1322
1323                 bch2_insert_fixup_btree_ptr(as, b, iter, k, &node_iter);
1324                 bch2_keylist_pop_front(keys);
1325         }
1326
1327         /*
1328          * We can't tolerate whiteouts here - with whiteouts there can be
1329          * duplicate keys, and it would be rather bad if we picked a duplicate
1330          * for the pivot:
1331          */
1332         i = btree_bset_first(b);
1333         p = i->start;
1334         while (p != vstruct_last(i))
1335                 if (bkey_deleted(p)) {
1336                         le16_add_cpu(&i->u64s, -p->u64s);
1337                         set_btree_bset_end(b, b->set);
1338                         memmove_u64s_down(p, bkey_next(p),
1339                                           (u64 *) vstruct_last(i) -
1340                                           (u64 *) p);
1341                 } else
1342                         p = bkey_next(p);
1343
1344         BUG_ON(b->nsets != 1 ||
1345                b->nr.live_u64s != le16_to_cpu(btree_bset_first(b)->u64s));
1346
1347         btree_node_interior_verify(b);
1348 }
1349
1350 static void btree_split(struct btree_update *as, struct btree *b,
1351                         struct btree_iter *iter, struct keylist *keys,
1352                         unsigned flags)
1353 {
1354         struct bch_fs *c = as->c;
1355         struct btree *parent = btree_node_parent(iter, b);
1356         struct btree *n1, *n2 = NULL, *n3 = NULL;
1357         u64 start_time = local_clock();
1358
1359         BUG_ON(!parent && (b != btree_node_root(c, b)));
1360         BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->level));
1361
1362         bch2_btree_interior_update_will_free_node(as, b);
1363
1364         n1 = bch2_btree_node_alloc_replacement(as, b);
1365
1366         if (keys)
1367                 btree_split_insert_keys(as, n1, iter, keys);
1368
1369         if (vstruct_blocks(n1->data, c->block_bits) > BTREE_SPLIT_THRESHOLD(c)) {
1370                 trace_btree_split(c, b);
1371
1372                 n2 = __btree_split_node(as, n1, iter);
1373
1374                 bch2_btree_build_aux_trees(n2);
1375                 bch2_btree_build_aux_trees(n1);
1376                 six_unlock_write(&n2->lock);
1377                 six_unlock_write(&n1->lock);
1378
1379                 bch2_btree_node_write(c, n2, SIX_LOCK_intent);
1380
1381                 /*
1382                  * Note that on recursive parent_keys == keys, so we
1383                  * can't start adding new keys to parent_keys before emptying it
1384                  * out (which we did with btree_split_insert_keys() above)
1385                  */
1386                 bch2_keylist_add(&as->parent_keys, &n1->key);
1387                 bch2_keylist_add(&as->parent_keys, &n2->key);
1388
1389                 if (!parent) {
1390                         /* Depth increases, make a new root */
1391                         n3 = __btree_root_alloc(as, b->level + 1);
1392
1393                         n3->sib_u64s[0] = U16_MAX;
1394                         n3->sib_u64s[1] = U16_MAX;
1395
1396                         btree_split_insert_keys(as, n3, iter, &as->parent_keys);
1397
1398                         bch2_btree_node_write(c, n3, SIX_LOCK_intent);
1399                 }
1400         } else {
1401                 trace_btree_compact(c, b);
1402
1403                 bch2_btree_build_aux_trees(n1);
1404                 six_unlock_write(&n1->lock);
1405
1406                 bch2_keylist_add(&as->parent_keys, &n1->key);
1407         }
1408
1409         bch2_btree_node_write(c, n1, SIX_LOCK_intent);
1410
1411         /* New nodes all written, now make them visible: */
1412
1413         if (parent) {
1414                 /* Split a non root node */
1415                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1416         } else if (n3) {
1417                 bch2_btree_set_root(as, n3, iter);
1418         } else {
1419                 /* Root filled up but didn't need to be split */
1420                 bch2_btree_set_root(as, n1, iter);
1421         }
1422
1423         bch2_open_buckets_put(c, &n1->ob);
1424         if (n2)
1425                 bch2_open_buckets_put(c, &n2->ob);
1426         if (n3)
1427                 bch2_open_buckets_put(c, &n3->ob);
1428
1429         /* Successful split, update the iterator to point to the new nodes: */
1430
1431         bch2_btree_iter_node_drop(iter, b);
1432         if (n3)
1433                 bch2_btree_iter_node_replace(iter, n3);
1434         if (n2)
1435                 bch2_btree_iter_node_replace(iter, n2);
1436         bch2_btree_iter_node_replace(iter, n1);
1437
1438         bch2_btree_node_free_inmem(c, b, iter);
1439
1440         bch2_btree_iter_verify_locks(iter);
1441
1442         bch2_time_stats_update(&c->times[BCH_TIME_btree_split], start_time);
1443 }
1444
1445 static void
1446 bch2_btree_insert_keys_interior(struct btree_update *as, struct btree *b,
1447                                 struct btree_iter *iter, struct keylist *keys)
1448 {
1449         struct btree_iter *linked;
1450         struct btree_node_iter node_iter;
1451         struct bkey_i *insert = bch2_keylist_front(keys);
1452         struct bkey_packed *k;
1453
1454         /* Don't screw up @iter's position: */
1455         node_iter = iter->l[b->level].iter;
1456
1457         /*
1458          * btree_split(), btree_gc_coalesce() will insert keys before
1459          * the iterator's current position - they know the keys go in
1460          * the node the iterator points to:
1461          */
1462         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1463                (bkey_cmp_packed(b, k, &insert->k) >= 0))
1464                 ;
1465
1466         while (!bch2_keylist_empty(keys)) {
1467                 insert = bch2_keylist_front(keys);
1468
1469                 bch2_insert_fixup_btree_ptr(as, b, iter, insert, &node_iter);
1470                 bch2_keylist_pop_front(keys);
1471         }
1472
1473         btree_update_updated_node(as, b);
1474
1475         for_each_btree_iter_with_node(iter, b, linked)
1476                 bch2_btree_node_iter_peek(&linked->l[b->level].iter, b);
1477
1478         bch2_btree_iter_verify(iter, b);
1479 }
1480
1481 /**
1482  * bch_btree_insert_node - insert bkeys into a given btree node
1483  *
1484  * @iter:               btree iterator
1485  * @keys:               list of keys to insert
1486  * @hook:               insert callback
1487  * @persistent:         if not null, @persistent will wait on journal write
1488  *
1489  * Inserts as many keys as it can into a given btree node, splitting it if full.
1490  * If a split occurred, this function will return early. This can only happen
1491  * for leaf nodes -- inserts into interior nodes have to be atomic.
1492  */
1493 void bch2_btree_insert_node(struct btree_update *as, struct btree *b,
1494                             struct btree_iter *iter, struct keylist *keys,
1495                             unsigned flags)
1496 {
1497         struct bch_fs *c = as->c;
1498         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1499         int old_live_u64s = b->nr.live_u64s;
1500         int live_u64s_added, u64s_added;
1501
1502         BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->level));
1503         BUG_ON(!b->level);
1504         BUG_ON(!as || as->b);
1505         bch2_verify_keylist_sorted(keys);
1506
1507         if (as->must_rewrite)
1508                 goto split;
1509
1510         bch2_btree_node_lock_for_insert(c, b, iter);
1511
1512         if (!bch2_btree_node_insert_fits(c, b, bch_keylist_u64s(keys))) {
1513                 bch2_btree_node_unlock_write(b, iter);
1514                 goto split;
1515         }
1516
1517         bch2_btree_insert_keys_interior(as, b, iter, keys);
1518
1519         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1520         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1521
1522         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1523                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1524         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1525                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1526
1527         if (u64s_added > live_u64s_added &&
1528             bch2_maybe_compact_whiteouts(c, b))
1529                 bch2_btree_iter_reinit_node(iter, b);
1530
1531         bch2_btree_node_unlock_write(b, iter);
1532
1533         btree_node_interior_verify(b);
1534
1535         /*
1536          * when called from the btree_split path the new nodes aren't added to
1537          * the btree iterator yet, so the merge path's unlock/wait/relock dance
1538          * won't work:
1539          */
1540         bch2_foreground_maybe_merge(c, iter, b->level,
1541                                     flags|BTREE_INSERT_NOUNLOCK);
1542         return;
1543 split:
1544         btree_split(as, b, iter, keys, flags);
1545 }
1546
1547 int bch2_btree_split_leaf(struct bch_fs *c, struct btree_iter *iter,
1548                           unsigned flags)
1549 {
1550         struct btree *b = iter->l[0].b;
1551         struct btree_update *as;
1552         struct closure cl;
1553         int ret = 0;
1554         struct btree_iter *linked;
1555
1556         /*
1557          * We already have a disk reservation and open buckets pinned; this
1558          * allocation must not block:
1559          */
1560         for_each_btree_iter(iter, linked)
1561                 if (linked->btree_id == BTREE_ID_EXTENTS)
1562                         flags |= BTREE_INSERT_USE_RESERVE;
1563
1564         closure_init_stack(&cl);
1565
1566         /* Hack, because gc and splitting nodes doesn't mix yet: */
1567         if (!down_read_trylock(&c->gc_lock)) {
1568                 if (flags & BTREE_INSERT_NOUNLOCK)
1569                         return -EINTR;
1570
1571                 bch2_btree_iter_unlock(iter);
1572                 down_read(&c->gc_lock);
1573
1574                 if (btree_iter_linked(iter))
1575                         ret = -EINTR;
1576         }
1577
1578         /*
1579          * XXX: figure out how far we might need to split,
1580          * instead of locking/reserving all the way to the root:
1581          */
1582         if (!bch2_btree_iter_upgrade(iter, U8_MAX,
1583                         !(flags & BTREE_INSERT_NOUNLOCK))) {
1584                 ret = -EINTR;
1585                 goto out;
1586         }
1587
1588         as = bch2_btree_update_start(c, iter->btree_id,
1589                 btree_update_reserve_required(c, b), flags,
1590                 !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1591         if (IS_ERR(as)) {
1592                 ret = PTR_ERR(as);
1593                 if (ret == -EAGAIN) {
1594                         BUG_ON(flags & BTREE_INSERT_NOUNLOCK);
1595                         bch2_btree_iter_unlock(iter);
1596                         ret = -EINTR;
1597                 }
1598                 goto out;
1599         }
1600
1601         btree_split(as, b, iter, NULL, flags);
1602         bch2_btree_update_done(as);
1603
1604         /*
1605          * We haven't successfully inserted yet, so don't downgrade all the way
1606          * back to read locks;
1607          */
1608         __bch2_btree_iter_downgrade(iter, 1);
1609 out:
1610         up_read(&c->gc_lock);
1611         closure_sync(&cl);
1612         return ret;
1613 }
1614
1615 void __bch2_foreground_maybe_merge(struct bch_fs *c,
1616                                    struct btree_iter *iter,
1617                                    unsigned level,
1618                                    unsigned flags,
1619                                    enum btree_node_sibling sib)
1620 {
1621         struct btree_update *as;
1622         struct bkey_format_state new_s;
1623         struct bkey_format new_f;
1624         struct bkey_i delete;
1625         struct btree *b, *m, *n, *prev, *next, *parent;
1626         struct closure cl;
1627         size_t sib_u64s;
1628         int ret = 0;
1629
1630         closure_init_stack(&cl);
1631 retry:
1632         BUG_ON(!btree_node_locked(iter, level));
1633
1634         b = iter->l[level].b;
1635
1636         parent = btree_node_parent(iter, b);
1637         if (!parent)
1638                 goto out;
1639
1640         if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c))
1641                 goto out;
1642
1643         /* XXX: can't be holding read locks */
1644         m = bch2_btree_node_get_sibling(c, iter, b,
1645                         !(flags & BTREE_INSERT_NOUNLOCK), sib);
1646         if (IS_ERR(m)) {
1647                 ret = PTR_ERR(m);
1648                 goto err;
1649         }
1650
1651         /* NULL means no sibling: */
1652         if (!m) {
1653                 b->sib_u64s[sib] = U16_MAX;
1654                 goto out;
1655         }
1656
1657         if (sib == btree_prev_sib) {
1658                 prev = m;
1659                 next = b;
1660         } else {
1661                 prev = b;
1662                 next = m;
1663         }
1664
1665         bch2_bkey_format_init(&new_s);
1666         __bch2_btree_calc_format(&new_s, b);
1667         __bch2_btree_calc_format(&new_s, m);
1668         new_f = bch2_bkey_format_done(&new_s);
1669
1670         sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1671                 btree_node_u64s_with_format(m, &new_f);
1672
1673         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1674                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1675                 sib_u64s /= 2;
1676                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1677         }
1678
1679         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1680         b->sib_u64s[sib] = sib_u64s;
1681
1682         if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c)) {
1683                 six_unlock_intent(&m->lock);
1684                 goto out;
1685         }
1686
1687         /* We're changing btree topology, doesn't mix with gc: */
1688         if (!down_read_trylock(&c->gc_lock))
1689                 goto err_cycle_gc_lock;
1690
1691         if (!bch2_btree_iter_upgrade(iter, U8_MAX,
1692                         !(flags & BTREE_INSERT_NOUNLOCK))) {
1693                 ret = -EINTR;
1694                 goto err_unlock;
1695         }
1696
1697         as = bch2_btree_update_start(c, iter->btree_id,
1698                          btree_update_reserve_required(c, parent) + 1,
1699                          BTREE_INSERT_NOFAIL|
1700                          BTREE_INSERT_USE_RESERVE,
1701                          !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1702         if (IS_ERR(as)) {
1703                 ret = PTR_ERR(as);
1704                 goto err_unlock;
1705         }
1706
1707         trace_btree_merge(c, b);
1708
1709         bch2_btree_interior_update_will_free_node(as, b);
1710         bch2_btree_interior_update_will_free_node(as, m);
1711
1712         n = bch2_btree_node_alloc(as, b->level);
1713
1714         n->data->min_key        = prev->data->min_key;
1715         n->data->max_key        = next->data->max_key;
1716         n->data->format         = new_f;
1717         n->key.k.p              = next->key.k.p;
1718
1719         btree_node_set_format(n, new_f);
1720
1721         bch2_btree_sort_into(c, n, prev);
1722         bch2_btree_sort_into(c, n, next);
1723
1724         bch2_btree_build_aux_trees(n);
1725         six_unlock_write(&n->lock);
1726
1727         bkey_init(&delete.k);
1728         delete.k.p = prev->key.k.p;
1729         bch2_keylist_add(&as->parent_keys, &delete);
1730         bch2_keylist_add(&as->parent_keys, &n->key);
1731
1732         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1733
1734         bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1735
1736         bch2_open_buckets_put(c, &n->ob);
1737
1738         bch2_btree_iter_node_drop(iter, b);
1739         bch2_btree_iter_node_replace(iter, n);
1740
1741         bch2_btree_iter_verify(iter, n);
1742
1743         bch2_btree_node_free_inmem(c, b, iter);
1744         bch2_btree_node_free_inmem(c, m, iter);
1745
1746         bch2_btree_update_done(as);
1747
1748         up_read(&c->gc_lock);
1749 out:
1750         bch2_btree_iter_verify_locks(iter);
1751
1752         /*
1753          * Don't downgrade locks here: we're called after successful insert,
1754          * and the caller will downgrade locks after a successful insert
1755          * anyways (in case e.g. a split was required first)
1756          *
1757          * And we're also called when inserting into interior nodes in the
1758          * split path, and downgrading to read locks in there is potentially
1759          * confusing:
1760          */
1761         closure_sync(&cl);
1762         return;
1763
1764 err_cycle_gc_lock:
1765         six_unlock_intent(&m->lock);
1766
1767         if (flags & BTREE_INSERT_NOUNLOCK)
1768                 goto out;
1769
1770         bch2_btree_iter_unlock(iter);
1771
1772         down_read(&c->gc_lock);
1773         up_read(&c->gc_lock);
1774         ret = -EINTR;
1775         goto err;
1776
1777 err_unlock:
1778         six_unlock_intent(&m->lock);
1779         up_read(&c->gc_lock);
1780 err:
1781         BUG_ON(ret == -EAGAIN && (flags & BTREE_INSERT_NOUNLOCK));
1782
1783         if ((ret == -EAGAIN || ret == -EINTR) &&
1784             !(flags & BTREE_INSERT_NOUNLOCK)) {
1785                 bch2_btree_iter_unlock(iter);
1786                 closure_sync(&cl);
1787                 ret = bch2_btree_iter_traverse(iter);
1788                 if (ret)
1789                         goto out;
1790
1791                 goto retry;
1792         }
1793
1794         goto out;
1795 }
1796
1797 static int __btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
1798                                 struct btree *b, unsigned flags,
1799                                 struct closure *cl)
1800 {
1801         struct btree *n, *parent = btree_node_parent(iter, b);
1802         struct btree_update *as;
1803
1804         as = bch2_btree_update_start(c, iter->btree_id,
1805                 (parent
1806                  ? btree_update_reserve_required(c, parent)
1807                  : 0) + 1,
1808                 flags, cl);
1809         if (IS_ERR(as)) {
1810                 trace_btree_gc_rewrite_node_fail(c, b);
1811                 return PTR_ERR(as);
1812         }
1813
1814         bch2_btree_interior_update_will_free_node(as, b);
1815
1816         n = bch2_btree_node_alloc_replacement(as, b);
1817
1818         bch2_btree_build_aux_trees(n);
1819         six_unlock_write(&n->lock);
1820
1821         trace_btree_gc_rewrite_node(c, b);
1822
1823         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1824
1825         if (parent) {
1826                 bch2_keylist_add(&as->parent_keys, &n->key);
1827                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1828         } else {
1829                 bch2_btree_set_root(as, n, iter);
1830         }
1831
1832         bch2_open_buckets_put(c, &n->ob);
1833
1834         bch2_btree_iter_node_drop(iter, b);
1835         bch2_btree_iter_node_replace(iter, n);
1836         bch2_btree_node_free_inmem(c, b, iter);
1837
1838         bch2_btree_update_done(as);
1839         return 0;
1840 }
1841
1842 /**
1843  * bch_btree_node_rewrite - Rewrite/move a btree node
1844  *
1845  * Returns 0 on success, -EINTR or -EAGAIN on failure (i.e.
1846  * btree_check_reserve() has to wait)
1847  */
1848 int bch2_btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
1849                             __le64 seq, unsigned flags)
1850 {
1851         struct closure cl;
1852         struct btree *b;
1853         int ret;
1854
1855         flags |= BTREE_INSERT_NOFAIL;
1856
1857         closure_init_stack(&cl);
1858
1859         bch2_btree_iter_upgrade(iter, U8_MAX, true);
1860
1861         if (!(flags & BTREE_INSERT_GC_LOCK_HELD)) {
1862                 if (!down_read_trylock(&c->gc_lock)) {
1863                         bch2_btree_iter_unlock(iter);
1864                         down_read(&c->gc_lock);
1865                 }
1866         }
1867
1868         while (1) {
1869                 ret = bch2_btree_iter_traverse(iter);
1870                 if (ret)
1871                         break;
1872
1873                 b = bch2_btree_iter_peek_node(iter);
1874                 if (!b || b->data->keys.seq != seq)
1875                         break;
1876
1877                 ret = __btree_node_rewrite(c, iter, b, flags, &cl);
1878                 if (ret != -EAGAIN &&
1879                     ret != -EINTR)
1880                         break;
1881
1882                 bch2_btree_iter_unlock(iter);
1883                 closure_sync(&cl);
1884         }
1885
1886         bch2_btree_iter_downgrade(iter);
1887
1888         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1889                 up_read(&c->gc_lock);
1890
1891         closure_sync(&cl);
1892         return ret;
1893 }
1894
1895 static void __bch2_btree_node_update_key(struct bch_fs *c,
1896                                          struct btree_update *as,
1897                                          struct btree_iter *iter,
1898                                          struct btree *b, struct btree *new_hash,
1899                                          struct bkey_i_btree_ptr *new_key)
1900 {
1901         struct btree *parent;
1902         int ret;
1903
1904         /*
1905          * Two corner cases that need to be thought about here:
1906          *
1907          * @b may not be reachable yet - there might be another interior update
1908          * operation waiting on @b to be written, and we're gonna deliver the
1909          * write completion to that interior update operation _before_
1910          * persisting the new_key update
1911          *
1912          * That ends up working without us having to do anything special here:
1913          * the reason is, we do kick off (and do the in memory updates) for the
1914          * update for @new_key before we return, creating a new interior_update
1915          * operation here.
1916          *
1917          * The new interior update operation here will in effect override the
1918          * previous one. The previous one was going to terminate - make @b
1919          * reachable - in one of two ways:
1920          * - updating the btree root pointer
1921          *   In that case,
1922          *   no, this doesn't work. argh.
1923          */
1924
1925         if (b->will_make_reachable)
1926                 as->must_rewrite = true;
1927
1928         btree_interior_update_add_node_reference(as, b);
1929
1930         /*
1931          * XXX: the rest of the update path treats this like we're actually
1932          * inserting a new node and deleting the existing node, so the
1933          * reservation needs to include enough space for @b
1934          *
1935          * that is actually sketch as fuck though and I am surprised the code
1936          * seems to work like that, definitely need to go back and rework it
1937          * into something saner.
1938          *
1939          * (I think @b is just getting double counted until the btree update
1940          * finishes and "deletes" @b on disk)
1941          */
1942         ret = bch2_disk_reservation_add(c, &as->reserve->disk_res,
1943                         c->opts.btree_node_size *
1944                         bch2_bkey_nr_ptrs(bkey_i_to_s_c(&new_key->k_i)),
1945                         BCH_DISK_RESERVATION_NOFAIL|
1946                         BCH_DISK_RESERVATION_GC_LOCK_HELD);
1947         BUG_ON(ret);
1948
1949         parent = btree_node_parent(iter, b);
1950         if (parent) {
1951                 if (new_hash) {
1952                         bkey_copy(&new_hash->key, &new_key->k_i);
1953                         ret = bch2_btree_node_hash_insert(&c->btree_cache,
1954                                         new_hash, b->level, b->btree_id);
1955                         BUG_ON(ret);
1956                 }
1957
1958                 bch2_keylist_add(&as->parent_keys, &new_key->k_i);
1959                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, 0);
1960
1961                 if (new_hash) {
1962                         mutex_lock(&c->btree_cache.lock);
1963                         bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
1964
1965                         bch2_btree_node_hash_remove(&c->btree_cache, b);
1966
1967                         bkey_copy(&b->key, &new_key->k_i);
1968                         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1969                         BUG_ON(ret);
1970                         mutex_unlock(&c->btree_cache.lock);
1971                 } else {
1972                         bkey_copy(&b->key, &new_key->k_i);
1973                 }
1974         } else {
1975                 struct bch_fs_usage *fs_usage;
1976
1977                 BUG_ON(btree_node_root(c, b) != b);
1978
1979                 bch2_btree_node_lock_write(b, iter);
1980
1981                 mutex_lock(&c->btree_interior_update_lock);
1982                 percpu_down_read_preempt_disable(&c->mark_lock);
1983                 fs_usage = bch2_fs_usage_get_scratch(c);
1984
1985                 bch2_mark_key_locked(c, bkey_i_to_s_c(&new_key->k_i),
1986                               true, 0,
1987                               gc_pos_btree_root(b->btree_id),
1988                               fs_usage, 0, 0);
1989                 bch2_btree_node_free_index(as, NULL,
1990                                            bkey_i_to_s_c(&b->key),
1991                                            fs_usage);
1992                 bch2_fs_usage_apply(c, fs_usage, &as->reserve->disk_res,
1993                                     gc_pos_btree_root(b->btree_id));
1994
1995                 percpu_up_read_preempt_enable(&c->mark_lock);
1996                 mutex_unlock(&c->btree_interior_update_lock);
1997
1998                 if (PTR_HASH(&new_key->k_i) != PTR_HASH(&b->key)) {
1999                         mutex_lock(&c->btree_cache.lock);
2000                         bch2_btree_node_hash_remove(&c->btree_cache, b);
2001
2002                         bkey_copy(&b->key, &new_key->k_i);
2003                         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
2004                         BUG_ON(ret);
2005                         mutex_unlock(&c->btree_cache.lock);
2006                 } else {
2007                         bkey_copy(&b->key, &new_key->k_i);
2008                 }
2009
2010                 btree_update_updated_root(as);
2011                 bch2_btree_node_unlock_write(b, iter);
2012         }
2013
2014         bch2_btree_update_done(as);
2015 }
2016
2017 int bch2_btree_node_update_key(struct bch_fs *c, struct btree_iter *iter,
2018                                struct btree *b,
2019                                struct bkey_i_btree_ptr *new_key)
2020 {
2021         struct btree *parent = btree_node_parent(iter, b);
2022         struct btree_update *as = NULL;
2023         struct btree *new_hash = NULL;
2024         struct closure cl;
2025         int ret;
2026
2027         closure_init_stack(&cl);
2028
2029         if (!bch2_btree_iter_upgrade(iter, U8_MAX, true))
2030                 return -EINTR;
2031
2032         if (!down_read_trylock(&c->gc_lock)) {
2033                 bch2_btree_iter_unlock(iter);
2034                 down_read(&c->gc_lock);
2035
2036                 if (!bch2_btree_iter_relock(iter)) {
2037                         ret = -EINTR;
2038                         goto err;
2039                 }
2040         }
2041
2042         /* check PTR_HASH() after @b is locked by btree_iter_traverse(): */
2043         if (PTR_HASH(&new_key->k_i) != PTR_HASH(&b->key)) {
2044                 /* bch2_btree_reserve_get will unlock */
2045                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2046                 if (ret) {
2047                         ret = -EINTR;
2048
2049                         bch2_btree_iter_unlock(iter);
2050                         up_read(&c->gc_lock);
2051                         closure_sync(&cl);
2052                         down_read(&c->gc_lock);
2053
2054                         if (!bch2_btree_iter_relock(iter))
2055                                 goto err;
2056                 }
2057
2058                 new_hash = bch2_btree_node_mem_alloc(c);
2059         }
2060
2061         as = bch2_btree_update_start(c, iter->btree_id,
2062                 parent ? btree_update_reserve_required(c, parent) : 0,
2063                 BTREE_INSERT_NOFAIL|
2064                 BTREE_INSERT_USE_RESERVE|
2065                 BTREE_INSERT_USE_ALLOC_RESERVE,
2066                 &cl);
2067
2068         if (IS_ERR(as)) {
2069                 ret = PTR_ERR(as);
2070                 if (ret == -EAGAIN)
2071                         ret = -EINTR;
2072
2073                 if (ret != -EINTR)
2074                         goto err;
2075
2076                 bch2_btree_iter_unlock(iter);
2077                 up_read(&c->gc_lock);
2078                 closure_sync(&cl);
2079                 down_read(&c->gc_lock);
2080
2081                 if (!bch2_btree_iter_relock(iter))
2082                         goto err;
2083         }
2084
2085         ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(&new_key->k_i));
2086         if (ret)
2087                 goto err_free_update;
2088
2089         __bch2_btree_node_update_key(c, as, iter, b, new_hash, new_key);
2090
2091         bch2_btree_iter_downgrade(iter);
2092 err:
2093         if (new_hash) {
2094                 mutex_lock(&c->btree_cache.lock);
2095                 list_move(&new_hash->list, &c->btree_cache.freeable);
2096                 mutex_unlock(&c->btree_cache.lock);
2097
2098                 six_unlock_write(&new_hash->lock);
2099                 six_unlock_intent(&new_hash->lock);
2100         }
2101         up_read(&c->gc_lock);
2102         closure_sync(&cl);
2103         return ret;
2104 err_free_update:
2105         bch2_btree_update_free(as);
2106         goto err;
2107 }
2108
2109 /* Init code: */
2110
2111 /*
2112  * Only for filesystem bringup, when first reading the btree roots or allocating
2113  * btree roots when initializing a new filesystem:
2114  */
2115 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2116 {
2117         BUG_ON(btree_node_root(c, b));
2118
2119         __bch2_btree_set_root_inmem(c, b);
2120 }
2121
2122 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2123 {
2124         struct closure cl;
2125         struct btree *b;
2126         int ret;
2127
2128         closure_init_stack(&cl);
2129
2130         do {
2131                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2132                 closure_sync(&cl);
2133         } while (ret);
2134
2135         b = bch2_btree_node_mem_alloc(c);
2136         bch2_btree_cache_cannibalize_unlock(c);
2137
2138         set_btree_node_fake(b);
2139         b->level        = 0;
2140         b->btree_id     = id;
2141
2142         bkey_btree_ptr_init(&b->key);
2143         b->key.k.p = POS_MAX;
2144         PTR_HASH(&b->key) = U64_MAX - id;
2145
2146         bch2_bset_init_first(b, &b->data->keys);
2147         bch2_btree_build_aux_trees(b);
2148
2149         b->data->min_key = POS_MIN;
2150         b->data->max_key = POS_MAX;
2151         b->data->format = bch2_btree_calc_format(b);
2152         btree_node_set_format(b, b->data->format);
2153
2154         ret = bch2_btree_node_hash_insert(&c->btree_cache, b, b->level, b->btree_id);
2155         BUG_ON(ret);
2156
2157         __bch2_btree_set_root_inmem(c, b);
2158
2159         six_unlock_write(&b->lock);
2160         six_unlock_intent(&b->lock);
2161 }
2162
2163 ssize_t bch2_btree_updates_print(struct bch_fs *c, char *buf)
2164 {
2165         struct printbuf out = _PBUF(buf, PAGE_SIZE);
2166         struct btree_update *as;
2167
2168         mutex_lock(&c->btree_interior_update_lock);
2169         list_for_each_entry(as, &c->btree_interior_update_list, list)
2170                 pr_buf(&out, "%p m %u w %u r %u j %llu\n",
2171                        as,
2172                        as->mode,
2173                        as->nodes_written,
2174                        atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2175                        as->journal.seq);
2176         mutex_unlock(&c->btree_interior_update_lock);
2177
2178         return out.pos - buf;
2179 }
2180
2181 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *c)
2182 {
2183         size_t ret = 0;
2184         struct list_head *i;
2185
2186         mutex_lock(&c->btree_interior_update_lock);
2187         list_for_each(i, &c->btree_interior_update_list)
2188                 ret++;
2189         mutex_unlock(&c->btree_interior_update_lock);
2190
2191         return ret;
2192 }