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 static void bch2_btree_set_root_ondisk(struct bch_fs *, struct btree *, int);
  28
  29 /* Debug code: */
  30
  31 static void btree_node_interior_verify(struct btree *b)
  32 {
  33         struct btree_node_iter iter;
  34         struct bkey_packed *k;
  35
  36         BUG_ON(!b->level);
  37
  38         bch2_btree_node_iter_init(&iter, b, &b->key.k.p);
  39 #if 1
  40         BUG_ON(!(k = bch2_btree_node_iter_peek(&iter, b)) ||
  41                bkey_cmp_left_packed(b, k, &b->key.k.p));
  42
  43         BUG_ON((bch2_btree_node_iter_advance(&iter, b),
  44                 !bch2_btree_node_iter_end(&iter)));
  45 #else
  46         const char *msg;
  47
  48         msg = "not found";
  49         k = bch2_btree_node_iter_peek(&iter, b);
  50         if (!k)
  51                 goto err;
  52
  53         msg = "isn't what it should be";
  54         if (bkey_cmp_left_packed(b, k, &b->key.k.p))
  55                 goto err;
  56
  57         bch2_btree_node_iter_advance(&iter, b);
  58
  59         msg = "isn't last key";
  60         if (!bch2_btree_node_iter_end(&iter))
  61                 goto err;
  62         return;
  63 err:
  64         bch2_dump_btree_node(b);
  65         printk(KERN_ERR "last key %llu:%llu %s\n", b->key.k.p.inode,
  66                b->key.k.p.offset, msg);
  67         BUG();
  68 #endif
  69 }
  70
  71 /* Calculate ideal packed bkey format for new btree nodes: */
  72
  73 void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b)
  74 {
  75         struct bkey_packed *k;
  76         struct bset_tree *t;
  77         struct bkey uk;
  78
  79         bch2_bkey_format_add_pos(s, b->data->min_key);
  80
  81         for_each_bset(b, t)
  82                 for (k = btree_bkey_first(b, t);
  83                      k != btree_bkey_last(b, t);
  84                      k = bkey_next(k))
  85                         if (!bkey_whiteout(k)) {
  86                                 uk = bkey_unpack_key(b, k);
  87                                 bch2_bkey_format_add_key(s, &uk);
  88                         }
  89 }
  90
  91 static struct bkey_format bch2_btree_calc_format(struct btree *b)
  92 {
  93         struct bkey_format_state s;
  94
  95         bch2_bkey_format_init(&s);
  96         __bch2_btree_calc_format(&s, b);
  97
  98         return bch2_bkey_format_done(&s);
  99 }
 100
 101 static size_t btree_node_u64s_with_format(struct btree *b,
 102                                           struct bkey_format *new_f)
 103 {
 104         struct bkey_format *old_f = &b->format;
 105
 106         /* stupid integer promotion rules */
 107         ssize_t delta =
 108             (((int) new_f->key_u64s - old_f->key_u64s) *
 109              (int) b->nr.packed_keys) +
 110             (((int) new_f->key_u64s - BKEY_U64s) *
 111              (int) b->nr.unpacked_keys);
 112
 113         BUG_ON(delta + b->nr.live_u64s < 0);
 114
 115         return b->nr.live_u64s + delta;
 116 }
 117
 118 /**
 119  * btree_node_format_fits - check if we could rewrite node with a new format
 120  *
 121  * This assumes all keys can pack with the new format -- it just checks if
 122  * the re-packed keys would fit inside the node itself.
 123  */
 124 bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
 125                                  struct bkey_format *new_f)
 126 {
 127         size_t u64s = btree_node_u64s_with_format(b, new_f);
 128
 129         return __vstruct_bytes(struct btree_node, u64s) < btree_bytes(c);
 130 }
 131
 132 /* Btree node freeing/allocation: */
 133
 134 static bool btree_key_matches(struct bch_fs *c,
 135                               struct bkey_s_c l,
 136                               struct bkey_s_c r)
 137 {
 138         struct bkey_ptrs_c ptrs1 = bch2_bkey_ptrs_c(l);
 139         struct bkey_ptrs_c ptrs2 = bch2_bkey_ptrs_c(r);
 140         const struct bch_extent_ptr *ptr1, *ptr2;
 141
 142         bkey_for_each_ptr(ptrs1, ptr1)
 143                 bkey_for_each_ptr(ptrs2, ptr2)
 144                         if (ptr1->dev == ptr2->dev &&
 145                             ptr1->gen == ptr2->gen &&
 146                             ptr1->offset == ptr2->offset)
 147                                 return true;
 148
 149         return false;
 150 }
 151
 152 /*
 153  * We're doing the index update that makes @b unreachable, update stuff to
 154  * reflect that:
 155  *
 156  * Must be called _before_ btree_update_updated_root() or
 157  * btree_update_updated_node:
 158  */
 159 static void bch2_btree_node_free_index(struct btree_update *as, struct btree *b,
 160                                        struct bkey_s_c k,
 161                                        struct bch_fs_usage *stats)
 162 {
 163         struct bch_fs *c = as->c;
 164         struct pending_btree_node_free *d;
 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         if (gc_pos_cmp(c->gc_pos, b
 193                        ? gc_pos_btree_node(b)
 194                        : gc_pos_btree_root(as->btree_id)) >= 0 &&
 195             gc_pos_cmp(c->gc_pos, gc_phase(GC_PHASE_PENDING_DELETE)) < 0)
 196                 bch2_mark_key_locked(c, bkey_i_to_s_c(&d->key),
 197                               0, 0, NULL, 0,
 198                               BCH_BUCKET_MARK_OVERWRITE|
 199                               BCH_BUCKET_MARK_GC);
 200 }
 201
 202 static void __btree_node_free(struct bch_fs *c, struct btree *b)
 203 {
 204         trace_btree_node_free(c, b);
 205
 206         BUG_ON(btree_node_dirty(b));
 207         BUG_ON(btree_node_need_write(b));
 208         BUG_ON(b == btree_node_root(c, b));
 209         BUG_ON(b->ob.nr);
 210         BUG_ON(!list_empty(&b->write_blocked));
 211         BUG_ON(b->will_make_reachable);
 212
 213         clear_btree_node_noevict(b);
 214
 215         bch2_btree_node_hash_remove(&c->btree_cache, b);
 216
 217         mutex_lock(&c->btree_cache.lock);
 218         list_move(&b->list, &c->btree_cache.freeable);
 219         mutex_unlock(&c->btree_cache.lock);
 220 }
 221
 222 void bch2_btree_node_free_never_inserted(struct bch_fs *c, struct btree *b)
 223 {
 224         struct open_buckets ob = b->ob;
 225
 226         btree_update_drop_new_node(c, b);
 227
 228         b->ob.nr = 0;
 229
 230         clear_btree_node_dirty(b);
 231
 232         btree_node_lock_type(c, b, SIX_LOCK_write);
 233         __btree_node_free(c, b);
 234         six_unlock_write(&b->lock);
 235
 236         bch2_open_buckets_put(c, &ob);
 237 }
 238
 239 void bch2_btree_node_free_inmem(struct bch_fs *c, struct btree *b,
 240                                 struct btree_iter *iter)
 241 {
 242         struct btree_iter *linked;
 243
 244         trans_for_each_iter(iter->trans, linked)
 245                 BUG_ON(linked->l[b->level].b == b);
 246
 247         /*
 248          * Is this a node that isn't reachable on disk yet?
 249          *
 250          * Nodes that aren't reachable yet have writes blocked until they're
 251          * reachable - now that we've cancelled any pending writes and moved
 252          * things waiting on that write to wait on this update, we can drop this
 253          * node from the list of nodes that the other update is making
 254          * reachable, prior to freeing it:
 255          */
 256         btree_update_drop_new_node(c, b);
 257
 258         six_lock_write(&b->lock);
 259         __btree_node_free(c, b);
 260         six_unlock_write(&b->lock);
 261         six_unlock_intent(&b->lock);
 262 }
 263
 264 static void bch2_btree_node_free_ondisk(struct bch_fs *c,
 265                                         struct pending_btree_node_free *pending)
 266 {
 267         BUG_ON(!pending->index_update_done);
 268
 269         bch2_mark_key(c, bkey_i_to_s_c(&pending->key),
 270                       0, 0, NULL, 0, BCH_BUCKET_MARK_OVERWRITE);
 271
 272         if (gc_visited(c, gc_phase(GC_PHASE_PENDING_DELETE)))
 273                 bch2_mark_key(c, bkey_i_to_s_c(&pending->key),
 274                               0, 0, NULL, 0,
 275                               BCH_BUCKET_MARK_OVERWRITE|
 276                               BCH_BUCKET_MARK_GC);
 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 = 0;
 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(&c->mark_lock);
1078         fs_usage = bch2_fs_usage_scratch_get(c);
1079
1080         bch2_mark_key_locked(c, bkey_i_to_s_c(&b->key),
1081                       0, 0, fs_usage, 0,
1082                       BCH_BUCKET_MARK_INSERT);
1083         if (gc_visited(c, gc_pos_btree_root(b->btree_id)))
1084                 bch2_mark_key_locked(c, bkey_i_to_s_c(&b->key),
1085                                      0, 0, NULL, 0,
1086                                      BCH_BUCKET_MARK_INSERT|
1087                                      BCH_BUCKET_MARK_GC);
1088
1089         if (old && !btree_node_fake(old))
1090                 bch2_btree_node_free_index(as, NULL,
1091                                            bkey_i_to_s_c(&old->key),
1092                                            fs_usage);
1093         bch2_fs_usage_apply(c, fs_usage, &as->reserve->disk_res, 0);
1094
1095         bch2_fs_usage_scratch_put(c, fs_usage);
1096         percpu_up_read(&c->mark_lock);
1097         mutex_unlock(&c->btree_interior_update_lock);
1098 }
1099
1100 static void bch2_btree_set_root_ondisk(struct bch_fs *c, struct btree *b, int rw)
1101 {
1102         struct btree_root *r = &c->btree_roots[b->btree_id];
1103
1104         mutex_lock(&c->btree_root_lock);
1105
1106         BUG_ON(b != r->b);
1107         bkey_copy(&r->key, &b->key);
1108         r->level = b->level;
1109         r->alive = true;
1110         if (rw == WRITE)
1111                 c->btree_roots_dirty = true;
1112
1113         mutex_unlock(&c->btree_root_lock);
1114 }
1115
1116 /**
1117  * bch_btree_set_root - update the root in memory and on disk
1118  *
1119  * To ensure forward progress, the current task must not be holding any
1120  * btree node write locks. However, you must hold an intent lock on the
1121  * old root.
1122  *
1123  * Note: This allocates a journal entry but doesn't add any keys to
1124  * it.  All the btree roots are part of every journal write, so there
1125  * is nothing new to be done.  This just guarantees that there is a
1126  * journal write.
1127  */
1128 static void bch2_btree_set_root(struct btree_update *as, struct btree *b,
1129                                 struct btree_iter *iter)
1130 {
1131         struct bch_fs *c = as->c;
1132         struct btree *old;
1133
1134         trace_btree_set_root(c, b);
1135         BUG_ON(!b->written &&
1136                !test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags));
1137
1138         old = btree_node_root(c, b);
1139
1140         /*
1141          * Ensure no one is using the old root while we switch to the
1142          * new root:
1143          */
1144         bch2_btree_node_lock_write(old, iter);
1145
1146         bch2_btree_set_root_inmem(as, b);
1147
1148         btree_update_updated_root(as);
1149
1150         /*
1151          * Unlock old root after new root is visible:
1152          *
1153          * The new root isn't persistent, but that's ok: we still have
1154          * an intent lock on the new root, and any updates that would
1155          * depend on the new root would have to update the new root.
1156          */
1157         bch2_btree_node_unlock_write(old, iter);
1158 }
1159
1160 /* Interior node updates: */
1161
1162 static void bch2_insert_fixup_btree_ptr(struct btree_update *as, struct btree *b,
1163                                         struct btree_iter *iter,
1164                                         struct bkey_i *insert,
1165                                         struct btree_node_iter *node_iter)
1166 {
1167         struct bch_fs *c = as->c;
1168         struct bch_fs_usage *fs_usage;
1169         struct bkey_packed *k;
1170         struct bkey tmp;
1171
1172         BUG_ON(insert->k.u64s > bch_btree_keys_u64s_remaining(c, b));
1173
1174         mutex_lock(&c->btree_interior_update_lock);
1175         percpu_down_read(&c->mark_lock);
1176         fs_usage = bch2_fs_usage_scratch_get(c);
1177
1178         bch2_mark_key_locked(c, bkey_i_to_s_c(insert),
1179                              0, 0, fs_usage, 0,
1180                              BCH_BUCKET_MARK_INSERT);
1181
1182         if (gc_visited(c, gc_pos_btree_node(b)))
1183                 bch2_mark_key_locked(c, bkey_i_to_s_c(insert),
1184                                      0, 0, NULL, 0,
1185                                      BCH_BUCKET_MARK_INSERT|
1186                                      BCH_BUCKET_MARK_GC);
1187
1188         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1189                bkey_iter_pos_cmp(b, &insert->k.p, k) > 0)
1190                 bch2_btree_node_iter_advance(node_iter, b);
1191
1192         /*
1193          * If we're overwriting, look up pending delete and mark so that gc
1194          * marks it on the pending delete list:
1195          */
1196         if (k && !bkey_cmp_packed(b, k, &insert->k))
1197                 bch2_btree_node_free_index(as, b,
1198                                            bkey_disassemble(b, k, &tmp),
1199                                            fs_usage);
1200
1201         bch2_fs_usage_apply(c, fs_usage, &as->reserve->disk_res, 0);
1202
1203         bch2_fs_usage_scratch_put(c, fs_usage);
1204         percpu_up_read(&c->mark_lock);
1205         mutex_unlock(&c->btree_interior_update_lock);
1206
1207         bch2_btree_bset_insert_key(iter, b, node_iter, insert);
1208         set_btree_node_dirty(b);
1209         set_btree_node_need_write(b);
1210 }
1211
1212 /*
1213  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1214  * node)
1215  */
1216 static struct btree *__btree_split_node(struct btree_update *as,
1217                                         struct btree *n1,
1218                                         struct btree_iter *iter)
1219 {
1220         size_t nr_packed = 0, nr_unpacked = 0;
1221         struct btree *n2;
1222         struct bset *set1, *set2;
1223         struct bkey_packed *k, *prev = NULL;
1224
1225         n2 = bch2_btree_node_alloc(as, n1->level);
1226
1227         n2->data->max_key       = n1->data->max_key;
1228         n2->data->format        = n1->format;
1229         SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data));
1230         n2->key.k.p = n1->key.k.p;
1231
1232         btree_node_set_format(n2, n2->data->format);
1233
1234         set1 = btree_bset_first(n1);
1235         set2 = btree_bset_first(n2);
1236
1237         /*
1238          * Has to be a linear search because we don't have an auxiliary
1239          * search tree yet
1240          */
1241         k = set1->start;
1242         while (1) {
1243                 if (bkey_next(k) == vstruct_last(set1))
1244                         break;
1245                 if (k->_data - set1->_data >= (le16_to_cpu(set1->u64s) * 3) / 5)
1246                         break;
1247
1248                 if (bkey_packed(k))
1249                         nr_packed++;
1250                 else
1251                         nr_unpacked++;
1252
1253                 prev = k;
1254                 k = bkey_next(k);
1255         }
1256
1257         BUG_ON(!prev);
1258
1259         n1->key.k.p = bkey_unpack_pos(n1, prev);
1260         n1->data->max_key = n1->key.k.p;
1261         n2->data->min_key =
1262                 btree_type_successor(n1->btree_id, n1->key.k.p);
1263
1264         set2->u64s = cpu_to_le16((u64 *) vstruct_end(set1) - (u64 *) k);
1265         set1->u64s = cpu_to_le16(le16_to_cpu(set1->u64s) - le16_to_cpu(set2->u64s));
1266
1267         set_btree_bset_end(n1, n1->set);
1268         set_btree_bset_end(n2, n2->set);
1269
1270         n2->nr.live_u64s        = le16_to_cpu(set2->u64s);
1271         n2->nr.bset_u64s[0]     = le16_to_cpu(set2->u64s);
1272         n2->nr.packed_keys      = n1->nr.packed_keys - nr_packed;
1273         n2->nr.unpacked_keys    = n1->nr.unpacked_keys - nr_unpacked;
1274
1275         n1->nr.live_u64s        = le16_to_cpu(set1->u64s);
1276         n1->nr.bset_u64s[0]     = le16_to_cpu(set1->u64s);
1277         n1->nr.packed_keys      = nr_packed;
1278         n1->nr.unpacked_keys    = nr_unpacked;
1279
1280         BUG_ON(!set1->u64s);
1281         BUG_ON(!set2->u64s);
1282
1283         memcpy_u64s(set2->start,
1284                     vstruct_end(set1),
1285                     le16_to_cpu(set2->u64s));
1286
1287         btree_node_reset_sib_u64s(n1);
1288         btree_node_reset_sib_u64s(n2);
1289
1290         bch2_verify_btree_nr_keys(n1);
1291         bch2_verify_btree_nr_keys(n2);
1292
1293         if (n1->level) {
1294                 btree_node_interior_verify(n1);
1295                 btree_node_interior_verify(n2);
1296         }
1297
1298         return n2;
1299 }
1300
1301 /*
1302  * For updates to interior nodes, we've got to do the insert before we split
1303  * because the stuff we're inserting has to be inserted atomically. Post split,
1304  * the keys might have to go in different nodes and the split would no longer be
1305  * atomic.
1306  *
1307  * Worse, if the insert is from btree node coalescing, if we do the insert after
1308  * we do the split (and pick the pivot) - the pivot we pick might be between
1309  * nodes that were coalesced, and thus in the middle of a child node post
1310  * coalescing:
1311  */
1312 static void btree_split_insert_keys(struct btree_update *as, struct btree *b,
1313                                     struct btree_iter *iter,
1314                                     struct keylist *keys)
1315 {
1316         struct btree_node_iter node_iter;
1317         struct bkey_i *k = bch2_keylist_front(keys);
1318         struct bkey_packed *p;
1319         struct bset *i;
1320
1321         BUG_ON(btree_node_type(b) != BKEY_TYPE_BTREE);
1322
1323         bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
1324
1325         while (!bch2_keylist_empty(keys)) {
1326                 k = bch2_keylist_front(keys);
1327
1328                 BUG_ON(bch_keylist_u64s(keys) >
1329                        bch_btree_keys_u64s_remaining(as->c, b));
1330                 BUG_ON(bkey_cmp(k->k.p, b->data->min_key) < 0);
1331                 BUG_ON(bkey_cmp(k->k.p, b->data->max_key) > 0);
1332
1333                 bch2_insert_fixup_btree_ptr(as, b, iter, k, &node_iter);
1334                 bch2_keylist_pop_front(keys);
1335         }
1336
1337         /*
1338          * We can't tolerate whiteouts here - with whiteouts there can be
1339          * duplicate keys, and it would be rather bad if we picked a duplicate
1340          * for the pivot:
1341          */
1342         i = btree_bset_first(b);
1343         p = i->start;
1344         while (p != vstruct_last(i))
1345                 if (bkey_deleted(p)) {
1346                         le16_add_cpu(&i->u64s, -p->u64s);
1347                         set_btree_bset_end(b, b->set);
1348                         memmove_u64s_down(p, bkey_next(p),
1349                                           (u64 *) vstruct_last(i) -
1350                                           (u64 *) p);
1351                 } else
1352                         p = bkey_next(p);
1353
1354         BUG_ON(b->nsets != 1 ||
1355                b->nr.live_u64s != le16_to_cpu(btree_bset_first(b)->u64s));
1356
1357         btree_node_interior_verify(b);
1358 }
1359
1360 static void btree_split(struct btree_update *as, struct btree *b,
1361                         struct btree_iter *iter, struct keylist *keys,
1362                         unsigned flags)
1363 {
1364         struct bch_fs *c = as->c;
1365         struct btree *parent = btree_node_parent(iter, b);
1366         struct btree *n1, *n2 = NULL, *n3 = NULL;
1367         u64 start_time = local_clock();
1368
1369         BUG_ON(!parent && (b != btree_node_root(c, b)));
1370         BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->level));
1371
1372         bch2_btree_interior_update_will_free_node(as, b);
1373
1374         n1 = bch2_btree_node_alloc_replacement(as, b);
1375
1376         if (keys)
1377                 btree_split_insert_keys(as, n1, iter, keys);
1378
1379         if (vstruct_blocks(n1->data, c->block_bits) > BTREE_SPLIT_THRESHOLD(c)) {
1380                 trace_btree_split(c, b);
1381
1382                 n2 = __btree_split_node(as, n1, iter);
1383
1384                 bch2_btree_build_aux_trees(n2);
1385                 bch2_btree_build_aux_trees(n1);
1386                 six_unlock_write(&n2->lock);
1387                 six_unlock_write(&n1->lock);
1388
1389                 bch2_btree_node_write(c, n2, SIX_LOCK_intent);
1390
1391                 /*
1392                  * Note that on recursive parent_keys == keys, so we
1393                  * can't start adding new keys to parent_keys before emptying it
1394                  * out (which we did with btree_split_insert_keys() above)
1395                  */
1396                 bch2_keylist_add(&as->parent_keys, &n1->key);
1397                 bch2_keylist_add(&as->parent_keys, &n2->key);
1398
1399                 if (!parent) {
1400                         /* Depth increases, make a new root */
1401                         n3 = __btree_root_alloc(as, b->level + 1);
1402
1403                         n3->sib_u64s[0] = U16_MAX;
1404                         n3->sib_u64s[1] = U16_MAX;
1405
1406                         btree_split_insert_keys(as, n3, iter, &as->parent_keys);
1407
1408                         bch2_btree_node_write(c, n3, SIX_LOCK_intent);
1409                 }
1410         } else {
1411                 trace_btree_compact(c, b);
1412
1413                 bch2_btree_build_aux_trees(n1);
1414                 six_unlock_write(&n1->lock);
1415
1416                 bch2_keylist_add(&as->parent_keys, &n1->key);
1417         }
1418
1419         bch2_btree_node_write(c, n1, SIX_LOCK_intent);
1420
1421         /* New nodes all written, now make them visible: */
1422
1423         if (parent) {
1424                 /* Split a non root node */
1425                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1426         } else if (n3) {
1427                 bch2_btree_set_root(as, n3, iter);
1428         } else {
1429                 /* Root filled up but didn't need to be split */
1430                 bch2_btree_set_root(as, n1, iter);
1431         }
1432
1433         bch2_open_buckets_put(c, &n1->ob);
1434         if (n2)
1435                 bch2_open_buckets_put(c, &n2->ob);
1436         if (n3)
1437                 bch2_open_buckets_put(c, &n3->ob);
1438
1439         /* Successful split, update the iterator to point to the new nodes: */
1440
1441         six_lock_increment(&b->lock, SIX_LOCK_intent);
1442         bch2_btree_iter_node_drop(iter, b);
1443         if (n3)
1444                 bch2_btree_iter_node_replace(iter, n3);
1445         if (n2)
1446                 bch2_btree_iter_node_replace(iter, n2);
1447         bch2_btree_iter_node_replace(iter, n1);
1448
1449         /*
1450          * The old node must be freed (in memory) _before_ unlocking the new
1451          * nodes - else another thread could re-acquire a read lock on the old
1452          * node after another thread has locked and updated the new node, thus
1453          * seeing stale data:
1454          */
1455         bch2_btree_node_free_inmem(c, b, iter);
1456
1457         if (n3)
1458                 six_unlock_intent(&n3->lock);
1459         if (n2)
1460                 six_unlock_intent(&n2->lock);
1461         six_unlock_intent(&n1->lock);
1462
1463         bch2_btree_trans_verify_locks(iter->trans);
1464
1465         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_split],
1466                                start_time);
1467 }
1468
1469 static void
1470 bch2_btree_insert_keys_interior(struct btree_update *as, struct btree *b,
1471                                 struct btree_iter *iter, struct keylist *keys)
1472 {
1473         struct btree_iter *linked;
1474         struct btree_node_iter node_iter;
1475         struct bkey_i *insert = bch2_keylist_front(keys);
1476         struct bkey_packed *k;
1477
1478         /* Don't screw up @iter's position: */
1479         node_iter = iter->l[b->level].iter;
1480
1481         /*
1482          * btree_split(), btree_gc_coalesce() will insert keys before
1483          * the iterator's current position - they know the keys go in
1484          * the node the iterator points to:
1485          */
1486         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1487                (bkey_cmp_packed(b, k, &insert->k) >= 0))
1488                 ;
1489
1490         while (!bch2_keylist_empty(keys)) {
1491                 insert = bch2_keylist_front(keys);
1492
1493                 bch2_insert_fixup_btree_ptr(as, b, iter, insert, &node_iter);
1494                 bch2_keylist_pop_front(keys);
1495         }
1496
1497         btree_update_updated_node(as, b);
1498
1499         trans_for_each_iter_with_node(iter->trans, b, linked)
1500                 bch2_btree_node_iter_peek(&linked->l[b->level].iter, b);
1501
1502         bch2_btree_iter_verify(iter, b);
1503 }
1504
1505 /**
1506  * bch_btree_insert_node - insert bkeys into a given btree node
1507  *
1508  * @iter:               btree iterator
1509  * @keys:               list of keys to insert
1510  * @hook:               insert callback
1511  * @persistent:         if not null, @persistent will wait on journal write
1512  *
1513  * Inserts as many keys as it can into a given btree node, splitting it if full.
1514  * If a split occurred, this function will return early. This can only happen
1515  * for leaf nodes -- inserts into interior nodes have to be atomic.
1516  */
1517 void bch2_btree_insert_node(struct btree_update *as, struct btree *b,
1518                             struct btree_iter *iter, struct keylist *keys,
1519                             unsigned flags)
1520 {
1521         struct bch_fs *c = as->c;
1522         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1523         int old_live_u64s = b->nr.live_u64s;
1524         int live_u64s_added, u64s_added;
1525
1526         BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->level));
1527         BUG_ON(!b->level);
1528         BUG_ON(!as || as->b);
1529         bch2_verify_keylist_sorted(keys);
1530
1531         if (as->must_rewrite)
1532                 goto split;
1533
1534         bch2_btree_node_lock_for_insert(c, b, iter);
1535
1536         if (!bch2_btree_node_insert_fits(c, b, bch_keylist_u64s(keys))) {
1537                 bch2_btree_node_unlock_write(b, iter);
1538                 goto split;
1539         }
1540
1541         bch2_btree_insert_keys_interior(as, b, iter, keys);
1542
1543         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1544         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1545
1546         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1547                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1548         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1549                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1550
1551         if (u64s_added > live_u64s_added &&
1552             bch2_maybe_compact_whiteouts(c, b))
1553                 bch2_btree_iter_reinit_node(iter, b);
1554
1555         bch2_btree_node_unlock_write(b, iter);
1556
1557         btree_node_interior_verify(b);
1558
1559         /*
1560          * when called from the btree_split path the new nodes aren't added to
1561          * the btree iterator yet, so the merge path's unlock/wait/relock dance
1562          * won't work:
1563          */
1564         bch2_foreground_maybe_merge(c, iter, b->level,
1565                                     flags|BTREE_INSERT_NOUNLOCK);
1566         return;
1567 split:
1568         btree_split(as, b, iter, keys, flags);
1569 }
1570
1571 int bch2_btree_split_leaf(struct bch_fs *c, struct btree_iter *iter,
1572                           unsigned flags)
1573 {
1574         struct btree_trans *trans = iter->trans;
1575         struct btree *b = iter->l[0].b;
1576         struct btree_update *as;
1577         struct closure cl;
1578         int ret = 0;
1579         struct btree_iter *linked;
1580
1581         /*
1582          * We already have a disk reservation and open buckets pinned; this
1583          * allocation must not block:
1584          */
1585         trans_for_each_iter(trans, linked)
1586                 if (linked->btree_id == BTREE_ID_EXTENTS)
1587                         flags |= BTREE_INSERT_USE_RESERVE;
1588
1589         closure_init_stack(&cl);
1590
1591         /* Hack, because gc and splitting nodes doesn't mix yet: */
1592         if (!(flags & BTREE_INSERT_GC_LOCK_HELD) &&
1593             !down_read_trylock(&c->gc_lock)) {
1594                 if (flags & BTREE_INSERT_NOUNLOCK)
1595                         return -EINTR;
1596
1597                 bch2_trans_unlock(trans);
1598                 down_read(&c->gc_lock);
1599
1600                 if (!bch2_trans_relock(trans))
1601                         ret = -EINTR;
1602         }
1603
1604         /*
1605          * XXX: figure out how far we might need to split,
1606          * instead of locking/reserving all the way to the root:
1607          */
1608         if (!bch2_btree_iter_upgrade(iter, U8_MAX)) {
1609                 trace_trans_restart_iter_upgrade(trans->ip);
1610                 ret = -EINTR;
1611                 goto out;
1612         }
1613
1614         as = bch2_btree_update_start(c, iter->btree_id,
1615                 btree_update_reserve_required(c, b), flags,
1616                 !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1617         if (IS_ERR(as)) {
1618                 ret = PTR_ERR(as);
1619                 if (ret == -EAGAIN) {
1620                         BUG_ON(flags & BTREE_INSERT_NOUNLOCK);
1621                         bch2_trans_unlock(trans);
1622                         ret = -EINTR;
1623                 }
1624                 goto out;
1625         }
1626
1627         btree_split(as, b, iter, NULL, flags);
1628         bch2_btree_update_done(as);
1629
1630         /*
1631          * We haven't successfully inserted yet, so don't downgrade all the way
1632          * back to read locks;
1633          */
1634         __bch2_btree_iter_downgrade(iter, 1);
1635 out:
1636         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1637                 up_read(&c->gc_lock);
1638         closure_sync(&cl);
1639         return ret;
1640 }
1641
1642 void __bch2_foreground_maybe_merge(struct bch_fs *c,
1643                                    struct btree_iter *iter,
1644                                    unsigned level,
1645                                    unsigned flags,
1646                                    enum btree_node_sibling sib)
1647 {
1648         struct btree_trans *trans = iter->trans;
1649         struct btree_update *as;
1650         struct bkey_format_state new_s;
1651         struct bkey_format new_f;
1652         struct bkey_i delete;
1653         struct btree *b, *m, *n, *prev, *next, *parent;
1654         struct closure cl;
1655         size_t sib_u64s;
1656         int ret = 0;
1657
1658         closure_init_stack(&cl);
1659 retry:
1660         BUG_ON(!btree_node_locked(iter, level));
1661
1662         b = iter->l[level].b;
1663
1664         parent = btree_node_parent(iter, b);
1665         if (!parent)
1666                 goto out;
1667
1668         if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c))
1669                 goto out;
1670
1671         /* XXX: can't be holding read locks */
1672         m = bch2_btree_node_get_sibling(c, iter, b, sib);
1673         if (IS_ERR(m)) {
1674                 ret = PTR_ERR(m);
1675                 goto err;
1676         }
1677
1678         /* NULL means no sibling: */
1679         if (!m) {
1680                 b->sib_u64s[sib] = U16_MAX;
1681                 goto out;
1682         }
1683
1684         if (sib == btree_prev_sib) {
1685                 prev = m;
1686                 next = b;
1687         } else {
1688                 prev = b;
1689                 next = m;
1690         }
1691
1692         bch2_bkey_format_init(&new_s);
1693         __bch2_btree_calc_format(&new_s, b);
1694         __bch2_btree_calc_format(&new_s, m);
1695         new_f = bch2_bkey_format_done(&new_s);
1696
1697         sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1698                 btree_node_u64s_with_format(m, &new_f);
1699
1700         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1701                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1702                 sib_u64s /= 2;
1703                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1704         }
1705
1706         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1707         b->sib_u64s[sib] = sib_u64s;
1708
1709         if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c)) {
1710                 six_unlock_intent(&m->lock);
1711                 goto out;
1712         }
1713
1714         /* We're changing btree topology, doesn't mix with gc: */
1715         if (!(flags & BTREE_INSERT_GC_LOCK_HELD) &&
1716             !down_read_trylock(&c->gc_lock))
1717                 goto err_cycle_gc_lock;
1718
1719         if (!bch2_btree_iter_upgrade(iter, U8_MAX)) {
1720                 ret = -EINTR;
1721                 goto err_unlock;
1722         }
1723
1724         as = bch2_btree_update_start(c, iter->btree_id,
1725                          btree_update_reserve_required(c, parent) + 1,
1726                          BTREE_INSERT_NOFAIL|
1727                          BTREE_INSERT_USE_RESERVE,
1728                          !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1729         if (IS_ERR(as)) {
1730                 ret = PTR_ERR(as);
1731                 goto err_unlock;
1732         }
1733
1734         trace_btree_merge(c, b);
1735
1736         bch2_btree_interior_update_will_free_node(as, b);
1737         bch2_btree_interior_update_will_free_node(as, m);
1738
1739         n = bch2_btree_node_alloc(as, b->level);
1740
1741         n->data->min_key        = prev->data->min_key;
1742         n->data->max_key        = next->data->max_key;
1743         n->data->format         = new_f;
1744         n->key.k.p              = next->key.k.p;
1745
1746         btree_node_set_format(n, new_f);
1747
1748         bch2_btree_sort_into(c, n, prev);
1749         bch2_btree_sort_into(c, n, next);
1750
1751         bch2_btree_build_aux_trees(n);
1752         six_unlock_write(&n->lock);
1753
1754         bkey_init(&delete.k);
1755         delete.k.p = prev->key.k.p;
1756         bch2_keylist_add(&as->parent_keys, &delete);
1757         bch2_keylist_add(&as->parent_keys, &n->key);
1758
1759         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1760
1761         bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1762
1763         bch2_open_buckets_put(c, &n->ob);
1764
1765         six_lock_increment(&b->lock, SIX_LOCK_intent);
1766         bch2_btree_iter_node_drop(iter, b);
1767         bch2_btree_iter_node_drop(iter, m);
1768
1769         bch2_btree_iter_node_replace(iter, n);
1770
1771         bch2_btree_iter_verify(iter, n);
1772
1773         bch2_btree_node_free_inmem(c, b, iter);
1774         bch2_btree_node_free_inmem(c, m, iter);
1775
1776         six_unlock_intent(&n->lock);
1777
1778         bch2_btree_update_done(as);
1779
1780         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1781                 up_read(&c->gc_lock);
1782 out:
1783         bch2_btree_trans_verify_locks(trans);
1784
1785         /*
1786          * Don't downgrade locks here: we're called after successful insert,
1787          * and the caller will downgrade locks after a successful insert
1788          * anyways (in case e.g. a split was required first)
1789          *
1790          * And we're also called when inserting into interior nodes in the
1791          * split path, and downgrading to read locks in there is potentially
1792          * confusing:
1793          */
1794         closure_sync(&cl);
1795         return;
1796
1797 err_cycle_gc_lock:
1798         six_unlock_intent(&m->lock);
1799
1800         if (flags & BTREE_INSERT_NOUNLOCK)
1801                 goto out;
1802
1803         bch2_trans_unlock(trans);
1804
1805         down_read(&c->gc_lock);
1806         up_read(&c->gc_lock);
1807         ret = -EINTR;
1808         goto err;
1809
1810 err_unlock:
1811         six_unlock_intent(&m->lock);
1812         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1813                 up_read(&c->gc_lock);
1814 err:
1815         BUG_ON(ret == -EAGAIN && (flags & BTREE_INSERT_NOUNLOCK));
1816
1817         if ((ret == -EAGAIN || ret == -EINTR) &&
1818             !(flags & BTREE_INSERT_NOUNLOCK)) {
1819                 bch2_trans_unlock(trans);
1820                 closure_sync(&cl);
1821                 ret = bch2_btree_iter_traverse(iter);
1822                 if (ret)
1823                         goto out;
1824
1825                 goto retry;
1826         }
1827
1828         goto out;
1829 }
1830
1831 static int __btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
1832                                 struct btree *b, unsigned flags,
1833                                 struct closure *cl)
1834 {
1835         struct btree *n, *parent = btree_node_parent(iter, b);
1836         struct btree_update *as;
1837
1838         as = bch2_btree_update_start(c, iter->btree_id,
1839                 (parent
1840                  ? btree_update_reserve_required(c, parent)
1841                  : 0) + 1,
1842                 flags, cl);
1843         if (IS_ERR(as)) {
1844                 trace_btree_gc_rewrite_node_fail(c, b);
1845                 return PTR_ERR(as);
1846         }
1847
1848         bch2_btree_interior_update_will_free_node(as, b);
1849
1850         n = bch2_btree_node_alloc_replacement(as, b);
1851
1852         bch2_btree_build_aux_trees(n);
1853         six_unlock_write(&n->lock);
1854
1855         trace_btree_gc_rewrite_node(c, b);
1856
1857         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1858
1859         if (parent) {
1860                 bch2_keylist_add(&as->parent_keys, &n->key);
1861                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1862         } else {
1863                 bch2_btree_set_root(as, n, iter);
1864         }
1865
1866         bch2_open_buckets_put(c, &n->ob);
1867
1868         six_lock_increment(&b->lock, SIX_LOCK_intent);
1869         bch2_btree_iter_node_drop(iter, b);
1870         bch2_btree_iter_node_replace(iter, n);
1871         bch2_btree_node_free_inmem(c, b, iter);
1872         six_unlock_intent(&n->lock);
1873
1874         bch2_btree_update_done(as);
1875         return 0;
1876 }
1877
1878 /**
1879  * bch_btree_node_rewrite - Rewrite/move a btree node
1880  *
1881  * Returns 0 on success, -EINTR or -EAGAIN on failure (i.e.
1882  * btree_check_reserve() has to wait)
1883  */
1884 int bch2_btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
1885                             __le64 seq, unsigned flags)
1886 {
1887         struct btree_trans *trans = iter->trans;
1888         struct closure cl;
1889         struct btree *b;
1890         int ret;
1891
1892         flags |= BTREE_INSERT_NOFAIL;
1893
1894         closure_init_stack(&cl);
1895
1896         bch2_btree_iter_upgrade(iter, U8_MAX);
1897
1898         if (!(flags & BTREE_INSERT_GC_LOCK_HELD)) {
1899                 if (!down_read_trylock(&c->gc_lock)) {
1900                         bch2_trans_unlock(trans);
1901                         down_read(&c->gc_lock);
1902                 }
1903         }
1904
1905         while (1) {
1906                 ret = bch2_btree_iter_traverse(iter);
1907                 if (ret)
1908                         break;
1909
1910                 b = bch2_btree_iter_peek_node(iter);
1911                 if (!b || b->data->keys.seq != seq)
1912                         break;
1913
1914                 ret = __btree_node_rewrite(c, iter, b, flags, &cl);
1915                 if (ret != -EAGAIN &&
1916                     ret != -EINTR)
1917                         break;
1918
1919                 bch2_trans_unlock(trans);
1920                 closure_sync(&cl);
1921         }
1922
1923         bch2_btree_iter_downgrade(iter);
1924
1925         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1926                 up_read(&c->gc_lock);
1927
1928         closure_sync(&cl);
1929         return ret;
1930 }
1931
1932 static void __bch2_btree_node_update_key(struct bch_fs *c,
1933                                          struct btree_update *as,
1934                                          struct btree_iter *iter,
1935                                          struct btree *b, struct btree *new_hash,
1936                                          struct bkey_i_btree_ptr *new_key)
1937 {
1938         struct btree *parent;
1939         int ret;
1940
1941         /*
1942          * Two corner cases that need to be thought about here:
1943          *
1944          * @b may not be reachable yet - there might be another interior update
1945          * operation waiting on @b to be written, and we're gonna deliver the
1946          * write completion to that interior update operation _before_
1947          * persisting the new_key update
1948          *
1949          * That ends up working without us having to do anything special here:
1950          * the reason is, we do kick off (and do the in memory updates) for the
1951          * update for @new_key before we return, creating a new interior_update
1952          * operation here.
1953          *
1954          * The new interior update operation here will in effect override the
1955          * previous one. The previous one was going to terminate - make @b
1956          * reachable - in one of two ways:
1957          * - updating the btree root pointer
1958          *   In that case,
1959          *   no, this doesn't work. argh.
1960          */
1961
1962         if (b->will_make_reachable)
1963                 as->must_rewrite = true;
1964
1965         btree_interior_update_add_node_reference(as, b);
1966
1967         /*
1968          * XXX: the rest of the update path treats this like we're actually
1969          * inserting a new node and deleting the existing node, so the
1970          * reservation needs to include enough space for @b
1971          *
1972          * that is actually sketch as fuck though and I am surprised the code
1973          * seems to work like that, definitely need to go back and rework it
1974          * into something saner.
1975          *
1976          * (I think @b is just getting double counted until the btree update
1977          * finishes and "deletes" @b on disk)
1978          */
1979         ret = bch2_disk_reservation_add(c, &as->reserve->disk_res,
1980                         c->opts.btree_node_size *
1981                         bch2_bkey_nr_ptrs(bkey_i_to_s_c(&new_key->k_i)),
1982                         BCH_DISK_RESERVATION_NOFAIL);
1983         BUG_ON(ret);
1984
1985         parent = btree_node_parent(iter, b);
1986         if (parent) {
1987                 if (new_hash) {
1988                         bkey_copy(&new_hash->key, &new_key->k_i);
1989                         ret = bch2_btree_node_hash_insert(&c->btree_cache,
1990                                         new_hash, b->level, b->btree_id);
1991                         BUG_ON(ret);
1992                 }
1993
1994                 bch2_keylist_add(&as->parent_keys, &new_key->k_i);
1995                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, 0);
1996
1997                 if (new_hash) {
1998                         mutex_lock(&c->btree_cache.lock);
1999                         bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
2000
2001                         bch2_btree_node_hash_remove(&c->btree_cache, b);
2002
2003                         bkey_copy(&b->key, &new_key->k_i);
2004                         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
2005                         BUG_ON(ret);
2006                         mutex_unlock(&c->btree_cache.lock);
2007                 } else {
2008                         bkey_copy(&b->key, &new_key->k_i);
2009                 }
2010         } else {
2011                 struct bch_fs_usage *fs_usage;
2012
2013                 BUG_ON(btree_node_root(c, b) != b);
2014
2015                 bch2_btree_node_lock_write(b, iter);
2016
2017                 mutex_lock(&c->btree_interior_update_lock);
2018                 percpu_down_read(&c->mark_lock);
2019                 fs_usage = bch2_fs_usage_scratch_get(c);
2020
2021                 bch2_mark_key_locked(c, bkey_i_to_s_c(&new_key->k_i),
2022                               0, 0, fs_usage, 0,
2023                               BCH_BUCKET_MARK_INSERT);
2024                 if (gc_visited(c, gc_pos_btree_root(b->btree_id)))
2025                         bch2_mark_key_locked(c, bkey_i_to_s_c(&new_key->k_i),
2026                                              0, 0, NULL, 0,
2027                                              BCH_BUCKET_MARK_INSERT||
2028                                              BCH_BUCKET_MARK_GC);
2029
2030                 bch2_btree_node_free_index(as, NULL,
2031                                            bkey_i_to_s_c(&b->key),
2032                                            fs_usage);
2033                 bch2_fs_usage_apply(c, fs_usage, &as->reserve->disk_res, 0);
2034
2035                 bch2_fs_usage_scratch_put(c, fs_usage);
2036                 percpu_up_read(&c->mark_lock);
2037                 mutex_unlock(&c->btree_interior_update_lock);
2038
2039                 if (PTR_HASH(&new_key->k_i) != PTR_HASH(&b->key)) {
2040                         mutex_lock(&c->btree_cache.lock);
2041                         bch2_btree_node_hash_remove(&c->btree_cache, b);
2042
2043                         bkey_copy(&b->key, &new_key->k_i);
2044                         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
2045                         BUG_ON(ret);
2046                         mutex_unlock(&c->btree_cache.lock);
2047                 } else {
2048                         bkey_copy(&b->key, &new_key->k_i);
2049                 }
2050
2051                 btree_update_updated_root(as);
2052                 bch2_btree_node_unlock_write(b, iter);
2053         }
2054
2055         bch2_btree_update_done(as);
2056 }
2057
2058 int bch2_btree_node_update_key(struct bch_fs *c, struct btree_iter *iter,
2059                                struct btree *b,
2060                                struct bkey_i_btree_ptr *new_key)
2061 {
2062         struct btree *parent = btree_node_parent(iter, b);
2063         struct btree_update *as = NULL;
2064         struct btree *new_hash = NULL;
2065         struct closure cl;
2066         int ret;
2067
2068         closure_init_stack(&cl);
2069
2070         if (!bch2_btree_iter_upgrade(iter, U8_MAX))
2071                 return -EINTR;
2072
2073         if (!down_read_trylock(&c->gc_lock)) {
2074                 bch2_trans_unlock(iter->trans);
2075                 down_read(&c->gc_lock);
2076
2077                 if (!bch2_trans_relock(iter->trans)) {
2078                         ret = -EINTR;
2079                         goto err;
2080                 }
2081         }
2082
2083         /* check PTR_HASH() after @b is locked by btree_iter_traverse(): */
2084         if (PTR_HASH(&new_key->k_i) != PTR_HASH(&b->key)) {
2085                 /* bch2_btree_reserve_get will unlock */
2086                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2087                 if (ret) {
2088                         bch2_trans_unlock(iter->trans);
2089                         up_read(&c->gc_lock);
2090                         closure_sync(&cl);
2091                         down_read(&c->gc_lock);
2092
2093                         if (!bch2_trans_relock(iter->trans)) {
2094                                 ret = -EINTR;
2095                                 goto err;
2096                         }
2097                 }
2098
2099                 new_hash = bch2_btree_node_mem_alloc(c);
2100         }
2101
2102         as = bch2_btree_update_start(c, iter->btree_id,
2103                 parent ? btree_update_reserve_required(c, parent) : 0,
2104                 BTREE_INSERT_NOFAIL|
2105                 BTREE_INSERT_USE_RESERVE|
2106                 BTREE_INSERT_USE_ALLOC_RESERVE,
2107                 &cl);
2108
2109         if (IS_ERR(as)) {
2110                 ret = PTR_ERR(as);
2111                 if (ret == -EAGAIN)
2112                         ret = -EINTR;
2113
2114                 if (ret != -EINTR)
2115                         goto err;
2116
2117                 bch2_trans_unlock(iter->trans);
2118                 up_read(&c->gc_lock);
2119                 closure_sync(&cl);
2120                 down_read(&c->gc_lock);
2121
2122                 if (!bch2_trans_relock(iter->trans))
2123                         goto err;
2124         }
2125
2126         ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(&new_key->k_i));
2127         if (ret)
2128                 goto err_free_update;
2129
2130         __bch2_btree_node_update_key(c, as, iter, b, new_hash, new_key);
2131
2132         bch2_btree_iter_downgrade(iter);
2133 err:
2134         if (new_hash) {
2135                 mutex_lock(&c->btree_cache.lock);
2136                 list_move(&new_hash->list, &c->btree_cache.freeable);
2137                 mutex_unlock(&c->btree_cache.lock);
2138
2139                 six_unlock_write(&new_hash->lock);
2140                 six_unlock_intent(&new_hash->lock);
2141         }
2142         up_read(&c->gc_lock);
2143         closure_sync(&cl);
2144         return ret;
2145 err_free_update:
2146         bch2_btree_update_free(as);
2147         goto err;
2148 }
2149
2150 /* Init code: */
2151
2152 /*
2153  * Only for filesystem bringup, when first reading the btree roots or allocating
2154  * btree roots when initializing a new filesystem:
2155  */
2156 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2157 {
2158         BUG_ON(btree_node_root(c, b));
2159
2160         __bch2_btree_set_root_inmem(c, b);
2161 }
2162
2163 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2164 {
2165         struct closure cl;
2166         struct btree *b;
2167         int ret;
2168
2169         closure_init_stack(&cl);
2170
2171         do {
2172                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2173                 closure_sync(&cl);
2174         } while (ret);
2175
2176         b = bch2_btree_node_mem_alloc(c);
2177         bch2_btree_cache_cannibalize_unlock(c);
2178
2179         set_btree_node_fake(b);
2180         b->level        = 0;
2181         b->btree_id     = id;
2182
2183         bkey_btree_ptr_init(&b->key);
2184         b->key.k.p = POS_MAX;
2185         PTR_HASH(&b->key) = U64_MAX - id;
2186
2187         bch2_bset_init_first(b, &b->data->keys);
2188         bch2_btree_build_aux_trees(b);
2189
2190         b->data->flags = 0;
2191         b->data->min_key = POS_MIN;
2192         b->data->max_key = POS_MAX;
2193         b->data->format = bch2_btree_calc_format(b);
2194         btree_node_set_format(b, b->data->format);
2195
2196         ret = bch2_btree_node_hash_insert(&c->btree_cache, b, b->level, b->btree_id);
2197         BUG_ON(ret);
2198
2199         __bch2_btree_set_root_inmem(c, b);
2200
2201         six_unlock_write(&b->lock);
2202         six_unlock_intent(&b->lock);
2203 }
2204
2205 ssize_t bch2_btree_updates_print(struct bch_fs *c, char *buf)
2206 {
2207         struct printbuf out = _PBUF(buf, PAGE_SIZE);
2208         struct btree_update *as;
2209
2210         mutex_lock(&c->btree_interior_update_lock);
2211         list_for_each_entry(as, &c->btree_interior_update_list, list)
2212                 pr_buf(&out, "%p m %u w %u r %u j %llu\n",
2213                        as,
2214                        as->mode,
2215                        as->nodes_written,
2216                        atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2217                        as->journal.seq);
2218         mutex_unlock(&c->btree_interior_update_lock);
2219
2220         return out.pos - buf;
2221 }
2222
2223 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *c)
2224 {
2225         size_t ret = 0;
2226         struct list_head *i;
2227
2228         mutex_lock(&c->btree_interior_update_lock);
2229         list_for_each(i, &c->btree_interior_update_list)
2230                 ret++;
2231         mutex_unlock(&c->btree_interior_update_lock);
2232
2233         return ret;
2234 }