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                 bset_tree_for_each_key(b, t, k)
  83                         if (!bkey_whiteout(k)) {
  84                                 uk = bkey_unpack_key(b, k);
  85                                 bch2_bkey_format_add_key(s, &uk);
  86                         }
  87 }
  88
  89 static struct bkey_format bch2_btree_calc_format(struct btree *b)
  90 {
  91         struct bkey_format_state s;
  92
  93         bch2_bkey_format_init(&s);
  94         __bch2_btree_calc_format(&s, b);
  95
  96         return bch2_bkey_format_done(&s);
  97 }
  98
  99 static size_t btree_node_u64s_with_format(struct btree *b,
 100                                           struct bkey_format *new_f)
 101 {
 102         struct bkey_format *old_f = &b->format;
 103
 104         /* stupid integer promotion rules */
 105         ssize_t delta =
 106             (((int) new_f->key_u64s - old_f->key_u64s) *
 107              (int) b->nr.packed_keys) +
 108             (((int) new_f->key_u64s - BKEY_U64s) *
 109              (int) b->nr.unpacked_keys);
 110
 111         BUG_ON(delta + b->nr.live_u64s < 0);
 112
 113         return b->nr.live_u64s + delta;
 114 }
 115
 116 /**
 117  * btree_node_format_fits - check if we could rewrite node with a new format
 118  *
 119  * This assumes all keys can pack with the new format -- it just checks if
 120  * the re-packed keys would fit inside the node itself.
 121  */
 122 bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
 123                                  struct bkey_format *new_f)
 124 {
 125         size_t u64s = btree_node_u64s_with_format(b, new_f);
 126
 127         return __vstruct_bytes(struct btree_node, u64s) < btree_bytes(c);
 128 }
 129
 130 /* Btree node freeing/allocation: */
 131
 132 static bool btree_key_matches(struct bch_fs *c,
 133                               struct bkey_s_c l,
 134                               struct bkey_s_c r)
 135 {
 136         struct bkey_ptrs_c ptrs1 = bch2_bkey_ptrs_c(l);
 137         struct bkey_ptrs_c ptrs2 = bch2_bkey_ptrs_c(r);
 138         const struct bch_extent_ptr *ptr1, *ptr2;
 139
 140         bkey_for_each_ptr(ptrs1, ptr1)
 141                 bkey_for_each_ptr(ptrs2, ptr2)
 142                         if (ptr1->dev == ptr2->dev &&
 143                             ptr1->gen == ptr2->gen &&
 144                             ptr1->offset == ptr2->offset)
 145                                 return true;
 146
 147         return false;
 148 }
 149
 150 /*
 151  * We're doing the index update that makes @b unreachable, update stuff to
 152  * reflect that:
 153  *
 154  * Must be called _before_ btree_update_updated_root() or
 155  * btree_update_updated_node:
 156  */
 157 static void bch2_btree_node_free_index(struct btree_update *as, struct btree *b,
 158                                        struct bkey_s_c k,
 159                                        struct bch_fs_usage *stats)
 160 {
 161         struct bch_fs *c = as->c;
 162         struct pending_btree_node_free *d;
 163
 164         for (d = as->pending; d < as->pending + as->nr_pending; d++)
 165                 if (!bkey_cmp(k.k->p, d->key.k.p) &&
 166                     btree_key_matches(c, k, bkey_i_to_s_c(&d->key)))
 167                         goto found;
 168         BUG();
 169 found:
 170         BUG_ON(d->index_update_done);
 171         d->index_update_done = true;
 172
 173         /*
 174          * We're dropping @k from the btree, but it's still live until the
 175          * index update is persistent so we need to keep a reference around for
 176          * mark and sweep to find - that's primarily what the
 177          * btree_node_pending_free list is for.
 178          *
 179          * So here (when we set index_update_done = true), we're moving an
 180          * existing reference to a different part of the larger "gc keyspace" -
 181          * and the new position comes after the old position, since GC marks
 182          * the pending free list after it walks the btree.
 183          *
 184          * If we move the reference while mark and sweep is _between_ the old
 185          * and the new position, mark and sweep will see the reference twice
 186          * and it'll get double accounted - so check for that here and subtract
 187          * to cancel out one of mark and sweep's markings if necessary:
 188          */
 189
 190         if (gc_pos_cmp(c->gc_pos, b
 191                        ? gc_pos_btree_node(b)
 192                        : gc_pos_btree_root(as->btree_id)) >= 0 &&
 193             gc_pos_cmp(c->gc_pos, gc_phase(GC_PHASE_PENDING_DELETE)) < 0)
 194                 bch2_mark_key_locked(c, bkey_i_to_s_c(&d->key),
 195                               0, 0, NULL, 0,
 196                               BTREE_TRIGGER_OVERWRITE|
 197                               BTREE_TRIGGER_GC);
 198 }
 199
 200 static void __btree_node_free(struct bch_fs *c, struct btree *b)
 201 {
 202         trace_btree_node_free(c, b);
 203
 204         BUG_ON(btree_node_dirty(b));
 205         BUG_ON(btree_node_need_write(b));
 206         BUG_ON(b == btree_node_root(c, b));
 207         BUG_ON(b->ob.nr);
 208         BUG_ON(!list_empty(&b->write_blocked));
 209         BUG_ON(b->will_make_reachable);
 210
 211         clear_btree_node_noevict(b);
 212
 213         bch2_btree_node_hash_remove(&c->btree_cache, b);
 214
 215         mutex_lock(&c->btree_cache.lock);
 216         list_move(&b->list, &c->btree_cache.freeable);
 217         mutex_unlock(&c->btree_cache.lock);
 218 }
 219
 220 void bch2_btree_node_free_never_inserted(struct bch_fs *c, struct btree *b)
 221 {
 222         struct open_buckets ob = b->ob;
 223
 224         btree_update_drop_new_node(c, b);
 225
 226         b->ob.nr = 0;
 227
 228         clear_btree_node_dirty(b);
 229
 230         btree_node_lock_type(c, b, SIX_LOCK_write);
 231         __btree_node_free(c, b);
 232         six_unlock_write(&b->lock);
 233
 234         bch2_open_buckets_put(c, &ob);
 235 }
 236
 237 void bch2_btree_node_free_inmem(struct bch_fs *c, struct btree *b,
 238                                 struct btree_iter *iter)
 239 {
 240         struct btree_iter *linked;
 241
 242         trans_for_each_iter(iter->trans, linked)
 243                 BUG_ON(linked->l[b->level].b == b);
 244
 245         /*
 246          * Is this a node that isn't reachable on disk yet?
 247          *
 248          * Nodes that aren't reachable yet have writes blocked until they're
 249          * reachable - now that we've cancelled any pending writes and moved
 250          * things waiting on that write to wait on this update, we can drop this
 251          * node from the list of nodes that the other update is making
 252          * reachable, prior to freeing it:
 253          */
 254         btree_update_drop_new_node(c, b);
 255
 256         six_lock_write(&b->lock);
 257         __btree_node_free(c, b);
 258         six_unlock_write(&b->lock);
 259         six_unlock_intent(&b->lock);
 260 }
 261
 262 static void bch2_btree_node_free_ondisk(struct bch_fs *c,
 263                                         struct pending_btree_node_free *pending)
 264 {
 265         BUG_ON(!pending->index_update_done);
 266
 267         bch2_mark_key(c, bkey_i_to_s_c(&pending->key),
 268                       0, 0, NULL, 0, BTREE_TRIGGER_OVERWRITE);
 269
 270         if (gc_visited(c, gc_phase(GC_PHASE_PENDING_DELETE)))
 271                 bch2_mark_key(c, bkey_i_to_s_c(&pending->key),
 272                               0, 0, NULL, 0,
 273                               BTREE_TRIGGER_OVERWRITE|
 274                               BTREE_TRIGGER_GC);
 275 }
 276
 277 static struct btree *__bch2_btree_node_alloc(struct bch_fs *c,
 278                                              struct disk_reservation *res,
 279                                              struct closure *cl,
 280                                              unsigned flags)
 281 {
 282         struct write_point *wp;
 283         struct btree *b;
 284         BKEY_PADDED(k) tmp;
 285         struct open_buckets ob = { .nr = 0 };
 286         struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
 287         unsigned nr_reserve;
 288         enum alloc_reserve alloc_reserve;
 289
 290         if (flags & BTREE_INSERT_USE_ALLOC_RESERVE) {
 291                 nr_reserve      = 0;
 292                 alloc_reserve   = RESERVE_ALLOC;
 293         } else if (flags & BTREE_INSERT_USE_RESERVE) {
 294                 nr_reserve      = BTREE_NODE_RESERVE / 2;
 295                 alloc_reserve   = RESERVE_BTREE;
 296         } else {
 297                 nr_reserve      = BTREE_NODE_RESERVE;
 298                 alloc_reserve   = RESERVE_NONE;
 299         }
 300
 301         mutex_lock(&c->btree_reserve_cache_lock);
 302         if (c->btree_reserve_cache_nr > nr_reserve) {
 303                 struct btree_alloc *a =
 304                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
 305
 306                 ob = a->ob;
 307                 bkey_copy(&tmp.k, &a->k);
 308                 mutex_unlock(&c->btree_reserve_cache_lock);
 309                 goto mem_alloc;
 310         }
 311         mutex_unlock(&c->btree_reserve_cache_lock);
 312
 313 retry:
 314         wp = bch2_alloc_sectors_start(c, c->opts.foreground_target, 0,
 315                                       writepoint_ptr(&c->btree_write_point),
 316                                       &devs_have,
 317                                       res->nr_replicas,
 318                                       c->opts.metadata_replicas_required,
 319                                       alloc_reserve, 0, cl);
 320         if (IS_ERR(wp))
 321                 return ERR_CAST(wp);
 322
 323         if (wp->sectors_free < c->opts.btree_node_size) {
 324                 struct open_bucket *ob;
 325                 unsigned i;
 326
 327                 open_bucket_for_each(c, &wp->ptrs, ob, i)
 328                         if (ob->sectors_free < c->opts.btree_node_size)
 329                                 ob->sectors_free = 0;
 330
 331                 bch2_alloc_sectors_done(c, wp);
 332                 goto retry;
 333         }
 334
 335         bkey_btree_ptr_init(&tmp.k);
 336         bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, c->opts.btree_node_size);
 337
 338         bch2_open_bucket_get(c, wp, &ob);
 339         bch2_alloc_sectors_done(c, wp);
 340 mem_alloc:
 341         b = bch2_btree_node_mem_alloc(c);
 342
 343         /* we hold cannibalize_lock: */
 344         BUG_ON(IS_ERR(b));
 345         BUG_ON(b->ob.nr);
 346
 347         bkey_copy(&b->key, &tmp.k);
 348         b->ob = ob;
 349
 350         return b;
 351 }
 352
 353 static struct btree *bch2_btree_node_alloc(struct btree_update *as, unsigned level)
 354 {
 355         struct bch_fs *c = as->c;
 356         struct btree *b;
 357
 358         BUG_ON(level >= BTREE_MAX_DEPTH);
 359         BUG_ON(!as->reserve->nr);
 360
 361         b = as->reserve->b[--as->reserve->nr];
 362
 363         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id));
 364
 365         set_btree_node_accessed(b);
 366         set_btree_node_dirty(b);
 367         set_btree_node_need_write(b);
 368
 369         bch2_bset_init_first(b, &b->data->keys);
 370         memset(&b->nr, 0, sizeof(b->nr));
 371         b->data->magic = cpu_to_le64(bset_magic(c));
 372         b->data->flags = 0;
 373         SET_BTREE_NODE_ID(b->data, as->btree_id);
 374         SET_BTREE_NODE_LEVEL(b->data, level);
 375         b->data->ptr = bkey_i_to_btree_ptr(&b->key)->v.start[0];
 376
 377         if (c->sb.features & (1ULL << BCH_FEATURE_new_extent_overwrite))
 378                 SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
 379
 380         if (btree_node_is_extents(b) &&
 381             !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data))
 382                 set_btree_node_old_extent_overwrite(b);
 383
 384         bch2_btree_build_aux_trees(b);
 385
 386         btree_node_will_make_reachable(as, b);
 387
 388         trace_btree_node_alloc(c, b);
 389         return b;
 390 }
 391
 392 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *as,
 393                                                   struct btree *b,
 394                                                   struct bkey_format format)
 395 {
 396         struct btree *n;
 397
 398         n = bch2_btree_node_alloc(as, b->level);
 399
 400         n->data->min_key        = b->data->min_key;
 401         n->data->max_key        = b->data->max_key;
 402         n->data->format         = format;
 403         SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
 404
 405         btree_node_set_format(n, format);
 406
 407         bch2_btree_sort_into(as->c, n, b);
 408
 409         btree_node_reset_sib_u64s(n);
 410
 411         n->key.k.p = b->key.k.p;
 412         return n;
 413 }
 414
 415 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
 416                                                        struct btree *b)
 417 {
 418         struct bkey_format new_f = bch2_btree_calc_format(b);
 419
 420         /*
 421          * The keys might expand with the new format - if they wouldn't fit in
 422          * the btree node anymore, use the old format for now:
 423          */
 424         if (!bch2_btree_node_format_fits(as->c, b, &new_f))
 425                 new_f = b->format;
 426
 427         return __bch2_btree_node_alloc_replacement(as, b, new_f);
 428 }
 429
 430 static struct btree *__btree_root_alloc(struct btree_update *as, unsigned level)
 431 {
 432         struct btree *b = bch2_btree_node_alloc(as, level);
 433
 434         b->data->min_key = POS_MIN;
 435         b->data->max_key = POS_MAX;
 436         b->data->format = bch2_btree_calc_format(b);
 437         b->key.k.p = POS_MAX;
 438
 439         btree_node_set_format(b, b->data->format);
 440         bch2_btree_build_aux_trees(b);
 441
 442         six_unlock_write(&b->lock);
 443
 444         return b;
 445 }
 446
 447 static void bch2_btree_reserve_put(struct bch_fs *c, struct btree_reserve *reserve)
 448 {
 449         bch2_disk_reservation_put(c, &reserve->disk_res);
 450
 451         mutex_lock(&c->btree_reserve_cache_lock);
 452
 453         while (reserve->nr) {
 454                 struct btree *b = reserve->b[--reserve->nr];
 455
 456                 six_unlock_write(&b->lock);
 457
 458                 if (c->btree_reserve_cache_nr <
 459                     ARRAY_SIZE(c->btree_reserve_cache)) {
 460                         struct btree_alloc *a =
 461                                 &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
 462
 463                         a->ob = b->ob;
 464                         b->ob.nr = 0;
 465                         bkey_copy(&a->k, &b->key);
 466                 } else {
 467                         bch2_open_buckets_put(c, &b->ob);
 468                 }
 469
 470                 btree_node_lock_type(c, b, SIX_LOCK_write);
 471                 __btree_node_free(c, b);
 472                 six_unlock_write(&b->lock);
 473
 474                 six_unlock_intent(&b->lock);
 475         }
 476
 477         mutex_unlock(&c->btree_reserve_cache_lock);
 478
 479         mempool_free(reserve, &c->btree_reserve_pool);
 480 }
 481
 482 static struct btree_reserve *bch2_btree_reserve_get(struct bch_fs *c,
 483                                                     unsigned nr_nodes,
 484                                                     unsigned flags,
 485                                                     struct closure *cl)
 486 {
 487         struct btree_reserve *reserve;
 488         struct btree *b;
 489         struct disk_reservation disk_res = { 0, 0 };
 490         unsigned sectors = nr_nodes * c->opts.btree_node_size;
 491         int ret, disk_res_flags = 0;
 492
 493         if (flags & BTREE_INSERT_NOFAIL)
 494                 disk_res_flags |= BCH_DISK_RESERVATION_NOFAIL;
 495
 496         /*
 497          * This check isn't necessary for correctness - it's just to potentially
 498          * prevent us from doing a lot of work that'll end up being wasted:
 499          */
 500         ret = bch2_journal_error(&c->journal);
 501         if (ret)
 502                 return ERR_PTR(ret);
 503
 504         if (bch2_disk_reservation_get(c, &disk_res, sectors,
 505                                       c->opts.metadata_replicas,
 506                                       disk_res_flags))
 507                 return ERR_PTR(-ENOSPC);
 508
 509         BUG_ON(nr_nodes > BTREE_RESERVE_MAX);
 510
 511         /*
 512          * Protects reaping from the btree node cache and using the btree node
 513          * open bucket reserve:
 514          */
 515         ret = bch2_btree_cache_cannibalize_lock(c, cl);
 516         if (ret) {
 517                 bch2_disk_reservation_put(c, &disk_res);
 518                 return ERR_PTR(ret);
 519         }
 520
 521         reserve = mempool_alloc(&c->btree_reserve_pool, GFP_NOIO);
 522
 523         reserve->disk_res = disk_res;
 524         reserve->nr = 0;
 525
 526         while (reserve->nr < nr_nodes) {
 527                 b = __bch2_btree_node_alloc(c, &disk_res,
 528                                             flags & BTREE_INSERT_NOWAIT
 529                                             ? NULL : cl, flags);
 530                 if (IS_ERR(b)) {
 531                         ret = PTR_ERR(b);
 532                         goto err_free;
 533                 }
 534
 535                 ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(&b->key));
 536                 if (ret)
 537                         goto err_free;
 538
 539                 reserve->b[reserve->nr++] = b;
 540         }
 541
 542         bch2_btree_cache_cannibalize_unlock(c);
 543         return reserve;
 544 err_free:
 545         bch2_btree_reserve_put(c, reserve);
 546         bch2_btree_cache_cannibalize_unlock(c);
 547         trace_btree_reserve_get_fail(c, nr_nodes, cl);
 548         return ERR_PTR(ret);
 549 }
 550
 551 /* Asynchronous interior node update machinery */
 552
 553 static void bch2_btree_update_free(struct btree_update *as)
 554 {
 555         struct bch_fs *c = as->c;
 556
 557         bch2_journal_pin_flush(&c->journal, &as->journal);
 558
 559         BUG_ON(as->nr_new_nodes);
 560         BUG_ON(as->nr_pending);
 561
 562         if (as->reserve)
 563                 bch2_btree_reserve_put(c, as->reserve);
 564
 565         mutex_lock(&c->btree_interior_update_lock);
 566         list_del(&as->list);
 567
 568         closure_debug_destroy(&as->cl);
 569         mempool_free(as, &c->btree_interior_update_pool);
 570
 571         closure_wake_up(&c->btree_interior_update_wait);
 572         mutex_unlock(&c->btree_interior_update_lock);
 573 }
 574
 575 static void btree_update_nodes_reachable(struct closure *cl)
 576 {
 577         struct btree_update *as = container_of(cl, struct btree_update, cl);
 578         struct bch_fs *c = as->c;
 579
 580         bch2_journal_pin_drop(&c->journal, &as->journal);
 581
 582         mutex_lock(&c->btree_interior_update_lock);
 583
 584         while (as->nr_new_nodes) {
 585                 struct btree *b = as->new_nodes[--as->nr_new_nodes];
 586
 587                 BUG_ON(b->will_make_reachable != (unsigned long) as);
 588                 b->will_make_reachable = 0;
 589                 mutex_unlock(&c->btree_interior_update_lock);
 590
 591                 /*
 592                  * b->will_make_reachable prevented it from being written, so
 593                  * write it now if it needs to be written:
 594                  */
 595                 btree_node_lock_type(c, b, SIX_LOCK_read);
 596                 bch2_btree_node_write_cond(c, b, btree_node_need_write(b));
 597                 six_unlock_read(&b->lock);
 598                 mutex_lock(&c->btree_interior_update_lock);
 599         }
 600
 601         while (as->nr_pending)
 602                 bch2_btree_node_free_ondisk(c, &as->pending[--as->nr_pending]);
 603
 604         mutex_unlock(&c->btree_interior_update_lock);
 605
 606         closure_wake_up(&as->wait);
 607
 608         bch2_btree_update_free(as);
 609 }
 610
 611 static void btree_update_wait_on_journal(struct closure *cl)
 612 {
 613         struct btree_update *as = container_of(cl, struct btree_update, cl);
 614         struct bch_fs *c = as->c;
 615         int ret;
 616
 617         ret = bch2_journal_open_seq_async(&c->journal, as->journal_seq, cl);
 618         if (ret == -EAGAIN) {
 619                 continue_at(cl, btree_update_wait_on_journal, system_wq);
 620                 return;
 621         }
 622         if (ret < 0)
 623                 goto err;
 624
 625         bch2_journal_flush_seq_async(&c->journal, as->journal_seq, cl);
 626 err:
 627         continue_at(cl, btree_update_nodes_reachable, system_wq);
 628 }
 629
 630 static void btree_update_nodes_written(struct closure *cl)
 631 {
 632         struct btree_update *as = container_of(cl, struct btree_update, cl);
 633         struct bch_fs *c = as->c;
 634         struct btree *b;
 635
 636         /*
 637          * We did an update to a parent node where the pointers we added pointed
 638          * to child nodes that weren't written yet: now, the child nodes have
 639          * been written so we can write out the update to the interior node.
 640          */
 641 retry:
 642         mutex_lock(&c->btree_interior_update_lock);
 643         as->nodes_written = true;
 644
 645         switch (as->mode) {
 646         case BTREE_INTERIOR_NO_UPDATE:
 647                 BUG();
 648         case BTREE_INTERIOR_UPDATING_NODE:
 649                 /* The usual case: */
 650                 b = READ_ONCE(as->b);
 651
 652                 if (!six_trylock_read(&b->lock)) {
 653                         mutex_unlock(&c->btree_interior_update_lock);
 654                         btree_node_lock_type(c, b, SIX_LOCK_read);
 655                         six_unlock_read(&b->lock);
 656                         goto retry;
 657                 }
 658
 659                 BUG_ON(!btree_node_dirty(b));
 660                 closure_wait(&btree_current_write(b)->wait, cl);
 661
 662                 list_del(&as->write_blocked_list);
 663
 664                 /*
 665                  * for flush_held_btree_writes() waiting on updates to flush or
 666                  * nodes to be writeable:
 667                  */
 668                 closure_wake_up(&c->btree_interior_update_wait);
 669                 mutex_unlock(&c->btree_interior_update_lock);
 670
 671                 /*
 672                  * b->write_blocked prevented it from being written, so
 673                  * write it now if it needs to be written:
 674                  */
 675                 bch2_btree_node_write_cond(c, b, true);
 676                 six_unlock_read(&b->lock);
 677                 break;
 678
 679         case BTREE_INTERIOR_UPDATING_AS:
 680                 /*
 681                  * The btree node we originally updated has been freed and is
 682                  * being rewritten - so we need to write anything here, we just
 683                  * need to signal to that btree_update that it's ok to make the
 684                  * new replacement node visible:
 685                  */
 686                 closure_put(&as->parent_as->cl);
 687
 688                 /*
 689                  * and then we have to wait on that btree_update to finish:
 690                  */
 691                 closure_wait(&as->parent_as->wait, cl);
 692                 mutex_unlock(&c->btree_interior_update_lock);
 693                 break;
 694
 695         case BTREE_INTERIOR_UPDATING_ROOT:
 696                 /* b is the new btree root: */
 697                 b = READ_ONCE(as->b);
 698
 699                 if (!six_trylock_read(&b->lock)) {
 700                         mutex_unlock(&c->btree_interior_update_lock);
 701                         btree_node_lock_type(c, b, SIX_LOCK_read);
 702                         six_unlock_read(&b->lock);
 703                         goto retry;
 704                 }
 705
 706                 BUG_ON(c->btree_roots[b->btree_id].as != as);
 707                 c->btree_roots[b->btree_id].as = NULL;
 708
 709                 bch2_btree_set_root_ondisk(c, b, WRITE);
 710
 711                 /*
 712                  * We don't have to wait anything anything here (before
 713                  * btree_update_nodes_reachable frees the old nodes
 714                  * ondisk) - we've ensured that the very next journal write will
 715                  * have the pointer to the new root, and before the allocator
 716                  * can reuse the old nodes it'll have to do a journal commit:
 717                  */
 718                 six_unlock_read(&b->lock);
 719                 mutex_unlock(&c->btree_interior_update_lock);
 720
 721                 /*
 722                  * Bit of funny circularity going on here we have to break:
 723                  *
 724                  * We have to drop our journal pin before writing the journal
 725                  * entry that points to the new btree root: else, we could
 726                  * deadlock if the journal currently happens to be full.
 727                  *
 728                  * This mean we're dropping the journal pin _before_ the new
 729                  * nodes are technically reachable - but this is safe, because
 730                  * after the bch2_btree_set_root_ondisk() call above they will
 731                  * be reachable as of the very next journal write:
 732                  */
 733                 bch2_journal_pin_drop(&c->journal, &as->journal);
 734
 735                 as->journal_seq = bch2_journal_last_unwritten_seq(&c->journal);
 736
 737                 btree_update_wait_on_journal(cl);
 738                 return;
 739         }
 740
 741         continue_at(cl, btree_update_nodes_reachable, system_wq);
 742 }
 743
 744 /*
 745  * We're updating @b with pointers to nodes that haven't finished writing yet:
 746  * block @b from being written until @as completes
 747  */
 748 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
 749 {
 750         struct bch_fs *c = as->c;
 751
 752         mutex_lock(&c->btree_interior_update_lock);
 753
 754         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 755         BUG_ON(!btree_node_dirty(b));
 756
 757         as->mode = BTREE_INTERIOR_UPDATING_NODE;
 758         as->b = b;
 759         list_add(&as->write_blocked_list, &b->write_blocked);
 760
 761         mutex_unlock(&c->btree_interior_update_lock);
 762
 763         /*
 764          * In general, when you're staging things in a journal that will later
 765          * be written elsewhere, and you also want to guarantee ordering: that
 766          * is, if you have updates a, b, c, after a crash you should never see c
 767          * and not a or b - there's a problem:
 768          *
 769          * If the final destination of the update(s) (i.e. btree node) can be
 770          * written/flushed _before_ the relevant journal entry - oops, that
 771          * breaks ordering, since the various leaf nodes can be written in any
 772          * order.
 773          *
 774          * Normally we use bset->journal_seq to deal with this - if during
 775          * recovery we find a btree node write that's newer than the newest
 776          * journal entry, we just ignore it - we don't need it, anything we're
 777          * supposed to have (that we reported as completed via fsync()) will
 778          * still be in the journal, and as far as the state of the journal is
 779          * concerned that btree node write never happened.
 780          *
 781          * That breaks when we're rewriting/splitting/merging nodes, since we're
 782          * mixing btree node writes that haven't happened yet with previously
 783          * written data that has been reported as completed to the journal.
 784          *
 785          * Thus, before making the new nodes reachable, we have to wait the
 786          * newest journal sequence number we have data for to be written (if it
 787          * hasn't been yet).
 788          */
 789         bch2_journal_wait_on_seq(&c->journal, as->journal_seq, &as->cl);
 790 }
 791
 792 static void interior_update_flush(struct journal *j,
 793                         struct journal_entry_pin *pin, u64 seq)
 794 {
 795         struct btree_update *as =
 796                 container_of(pin, struct btree_update, journal);
 797
 798         bch2_journal_flush_seq_async(j, as->journal_seq, NULL);
 799 }
 800
 801 static void btree_update_reparent(struct btree_update *as,
 802                                   struct btree_update *child)
 803 {
 804         struct bch_fs *c = as->c;
 805
 806         child->b = NULL;
 807         child->mode = BTREE_INTERIOR_UPDATING_AS;
 808         child->parent_as = as;
 809         closure_get(&as->cl);
 810
 811         /*
 812          * When we write a new btree root, we have to drop our journal pin
 813          * _before_ the new nodes are technically reachable; see
 814          * btree_update_nodes_written().
 815          *
 816          * This goes for journal pins that are recursively blocked on us - so,
 817          * just transfer the journal pin to the new interior update so
 818          * btree_update_nodes_written() can drop it.
 819          */
 820         bch2_journal_pin_add_if_older(&c->journal, &child->journal,
 821                                       &as->journal, interior_update_flush);
 822         bch2_journal_pin_drop(&c->journal, &child->journal);
 823
 824         as->journal_seq = max(as->journal_seq, child->journal_seq);
 825 }
 826
 827 static void btree_update_updated_root(struct btree_update *as)
 828 {
 829         struct bch_fs *c = as->c;
 830         struct btree_root *r = &c->btree_roots[as->btree_id];
 831
 832         mutex_lock(&c->btree_interior_update_lock);
 833
 834         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 835
 836         /*
 837          * Old root might not be persistent yet - if so, redirect its
 838          * btree_update operation to point to us:
 839          */
 840         if (r->as)
 841                 btree_update_reparent(as, r->as);
 842
 843         as->mode = BTREE_INTERIOR_UPDATING_ROOT;
 844         as->b = r->b;
 845         r->as = as;
 846
 847         mutex_unlock(&c->btree_interior_update_lock);
 848
 849         /*
 850          * When we're rewriting nodes and updating interior nodes, there's an
 851          * issue with updates that haven't been written in the journal getting
 852          * mixed together with older data - see btree_update_updated_node()
 853          * for the explanation.
 854          *
 855          * However, this doesn't affect us when we're writing a new btree root -
 856          * because to make that new root reachable we have to write out a new
 857          * journal entry, which must necessarily be newer than as->journal_seq.
 858          */
 859 }
 860
 861 static void btree_node_will_make_reachable(struct btree_update *as,
 862                                            struct btree *b)
 863 {
 864         struct bch_fs *c = as->c;
 865
 866         mutex_lock(&c->btree_interior_update_lock);
 867         BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
 868         BUG_ON(b->will_make_reachable);
 869
 870         as->new_nodes[as->nr_new_nodes++] = b;
 871         b->will_make_reachable = 1UL|(unsigned long) as;
 872
 873         closure_get(&as->cl);
 874         mutex_unlock(&c->btree_interior_update_lock);
 875 }
 876
 877 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
 878 {
 879         struct btree_update *as;
 880         unsigned long v;
 881         unsigned i;
 882
 883         mutex_lock(&c->btree_interior_update_lock);
 884         v = xchg(&b->will_make_reachable, 0);
 885         as = (struct btree_update *) (v & ~1UL);
 886
 887         if (!as) {
 888                 mutex_unlock(&c->btree_interior_update_lock);
 889                 return;
 890         }
 891
 892         for (i = 0; i < as->nr_new_nodes; i++)
 893                 if (as->new_nodes[i] == b)
 894                         goto found;
 895
 896         BUG();
 897 found:
 898         array_remove_item(as->new_nodes, as->nr_new_nodes, i);
 899         mutex_unlock(&c->btree_interior_update_lock);
 900
 901         if (v & 1)
 902                 closure_put(&as->cl);
 903 }
 904
 905 static void btree_interior_update_add_node_reference(struct btree_update *as,
 906                                                      struct btree *b)
 907 {
 908         struct bch_fs *c = as->c;
 909         struct pending_btree_node_free *d;
 910
 911         mutex_lock(&c->btree_interior_update_lock);
 912
 913         /* Add this node to the list of nodes being freed: */
 914         BUG_ON(as->nr_pending >= ARRAY_SIZE(as->pending));
 915
 916         d = &as->pending[as->nr_pending++];
 917         d->index_update_done    = false;
 918         d->seq                  = b->data->keys.seq;
 919         d->btree_id             = b->btree_id;
 920         d->level                = b->level;
 921         bkey_copy(&d->key, &b->key);
 922
 923         mutex_unlock(&c->btree_interior_update_lock);
 924 }
 925
 926 /*
 927  * @b is being split/rewritten: it may have pointers to not-yet-written btree
 928  * nodes and thus outstanding btree_updates - redirect @b's
 929  * btree_updates to point to this btree_update:
 930  */
 931 void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 932                                                struct btree *b)
 933 {
 934         struct bch_fs *c = as->c;
 935         struct closure *cl, *cl_n;
 936         struct btree_update *p, *n;
 937         struct btree_write *w;
 938         struct bset_tree *t;
 939
 940         set_btree_node_dying(b);
 941
 942         if (btree_node_fake(b))
 943                 return;
 944
 945         btree_interior_update_add_node_reference(as, b);
 946
 947         /*
 948          * Does this node have data that hasn't been written in the journal?
 949          *
 950          * If so, we have to wait for the corresponding journal entry to be
 951          * written before making the new nodes reachable - we can't just carry
 952          * over the bset->journal_seq tracking, since we'll be mixing those keys
 953          * in with keys that aren't in the journal anymore:
 954          */
 955         for_each_bset(b, t)
 956                 as->journal_seq = max(as->journal_seq,
 957                                       le64_to_cpu(bset(b, t)->journal_seq));
 958
 959         mutex_lock(&c->btree_interior_update_lock);
 960
 961         /*
 962          * Does this node have any btree_update operations preventing
 963          * it from being written?
 964          *
 965          * If so, redirect them to point to this btree_update: we can
 966          * write out our new nodes, but we won't make them visible until those
 967          * operations complete
 968          */
 969         list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
 970                 list_del(&p->write_blocked_list);
 971                 btree_update_reparent(as, p);
 972
 973                 /*
 974                  * for flush_held_btree_writes() waiting on updates to flush or
 975                  * nodes to be writeable:
 976                  */
 977                 closure_wake_up(&c->btree_interior_update_wait);
 978         }
 979
 980         clear_btree_node_dirty(b);
 981         clear_btree_node_need_write(b);
 982         w = btree_current_write(b);
 983
 984         /*
 985          * Does this node have any btree_update operations waiting on this node
 986          * to be written?
 987          *
 988          * If so, wake them up when this btree_update operation is reachable:
 989          */
 990         llist_for_each_entry_safe(cl, cl_n, llist_del_all(&w->wait.list), list)
 991                 llist_add(&cl->list, &as->wait.list);
 992
 993         /*
 994          * Does this node have unwritten data that has a pin on the journal?
 995          *
 996          * If so, transfer that pin to the btree_update operation -
 997          * note that if we're freeing multiple nodes, we only need to keep the
 998          * oldest pin of any of the nodes we're freeing. We'll release the pin
 999          * when the new nodes are persistent and reachable on disk:
1000          */
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         w = btree_prev_write(b);
1006         bch2_journal_pin_add_if_older(&c->journal, &w->journal,
1007                                       &as->journal, interior_update_flush);
1008         bch2_journal_pin_drop(&c->journal, &w->journal);
1009
1010         mutex_unlock(&c->btree_interior_update_lock);
1011 }
1012
1013 void bch2_btree_update_done(struct btree_update *as)
1014 {
1015         BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
1016
1017         bch2_btree_reserve_put(as->c, as->reserve);
1018         as->reserve = NULL;
1019
1020         continue_at(&as->cl, btree_update_nodes_written, system_freezable_wq);
1021 }
1022
1023 struct btree_update *
1024 bch2_btree_update_start(struct bch_fs *c, enum btree_id id,
1025                         unsigned nr_nodes, unsigned flags,
1026                         struct closure *cl)
1027 {
1028         struct btree_reserve *reserve;
1029         struct btree_update *as;
1030
1031         reserve = bch2_btree_reserve_get(c, nr_nodes, flags, cl);
1032         if (IS_ERR(reserve))
1033                 return ERR_CAST(reserve);
1034
1035         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
1036         memset(as, 0, sizeof(*as));
1037         closure_init(&as->cl, NULL);
1038         as->c           = c;
1039         as->mode        = BTREE_INTERIOR_NO_UPDATE;
1040         as->btree_id    = id;
1041         as->reserve     = reserve;
1042         INIT_LIST_HEAD(&as->write_blocked_list);
1043
1044         bch2_keylist_init(&as->parent_keys, as->inline_keys);
1045
1046         mutex_lock(&c->btree_interior_update_lock);
1047         list_add_tail(&as->list, &c->btree_interior_update_list);
1048         mutex_unlock(&c->btree_interior_update_lock);
1049
1050         return as;
1051 }
1052
1053 /* Btree root updates: */
1054
1055 static void __bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1056 {
1057         /* Root nodes cannot be reaped */
1058         mutex_lock(&c->btree_cache.lock);
1059         list_del_init(&b->list);
1060         mutex_unlock(&c->btree_cache.lock);
1061
1062         mutex_lock(&c->btree_root_lock);
1063         BUG_ON(btree_node_root(c, b) &&
1064                (b->level < btree_node_root(c, b)->level ||
1065                 !btree_node_dying(btree_node_root(c, b))));
1066
1067         btree_node_root(c, b) = b;
1068         mutex_unlock(&c->btree_root_lock);
1069
1070         bch2_recalc_btree_reserve(c);
1071 }
1072
1073 static void bch2_btree_set_root_inmem(struct btree_update *as, struct btree *b)
1074 {
1075         struct bch_fs *c = as->c;
1076         struct btree *old = btree_node_root(c, b);
1077         struct bch_fs_usage *fs_usage;
1078
1079         __bch2_btree_set_root_inmem(c, b);
1080
1081         mutex_lock(&c->btree_interior_update_lock);
1082         percpu_down_read(&c->mark_lock);
1083         fs_usage = bch2_fs_usage_scratch_get(c);
1084
1085         bch2_mark_key_locked(c, bkey_i_to_s_c(&b->key),
1086                       0, 0, fs_usage, 0,
1087                       BTREE_TRIGGER_INSERT);
1088         if (gc_visited(c, gc_pos_btree_root(b->btree_id)))
1089                 bch2_mark_key_locked(c, bkey_i_to_s_c(&b->key),
1090                                      0, 0, NULL, 0,
1091                                      BTREE_TRIGGER_INSERT|
1092                                      BTREE_TRIGGER_GC);
1093
1094         if (old && !btree_node_fake(old))
1095                 bch2_btree_node_free_index(as, NULL,
1096                                            bkey_i_to_s_c(&old->key),
1097                                            fs_usage);
1098         bch2_fs_usage_apply(c, fs_usage, &as->reserve->disk_res, 0);
1099
1100         bch2_fs_usage_scratch_put(c, fs_usage);
1101         percpu_up_read(&c->mark_lock);
1102         mutex_unlock(&c->btree_interior_update_lock);
1103 }
1104
1105 static void bch2_btree_set_root_ondisk(struct bch_fs *c, struct btree *b, int rw)
1106 {
1107         struct btree_root *r = &c->btree_roots[b->btree_id];
1108
1109         mutex_lock(&c->btree_root_lock);
1110
1111         BUG_ON(b != r->b);
1112         bkey_copy(&r->key, &b->key);
1113         r->level = b->level;
1114         r->alive = true;
1115         if (rw == WRITE)
1116                 c->btree_roots_dirty = true;
1117
1118         mutex_unlock(&c->btree_root_lock);
1119 }
1120
1121 /**
1122  * bch_btree_set_root - update the root in memory and on disk
1123  *
1124  * To ensure forward progress, the current task must not be holding any
1125  * btree node write locks. However, you must hold an intent lock on the
1126  * old root.
1127  *
1128  * Note: This allocates a journal entry but doesn't add any keys to
1129  * it.  All the btree roots are part of every journal write, so there
1130  * is nothing new to be done.  This just guarantees that there is a
1131  * journal write.
1132  */
1133 static void bch2_btree_set_root(struct btree_update *as, struct btree *b,
1134                                 struct btree_iter *iter)
1135 {
1136         struct bch_fs *c = as->c;
1137         struct btree *old;
1138
1139         trace_btree_set_root(c, b);
1140         BUG_ON(!b->written &&
1141                !test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags));
1142
1143         old = btree_node_root(c, b);
1144
1145         /*
1146          * Ensure no one is using the old root while we switch to the
1147          * new root:
1148          */
1149         bch2_btree_node_lock_write(old, iter);
1150
1151         bch2_btree_set_root_inmem(as, b);
1152
1153         btree_update_updated_root(as);
1154
1155         /*
1156          * Unlock old root after new root is visible:
1157          *
1158          * The new root isn't persistent, but that's ok: we still have
1159          * an intent lock on the new root, and any updates that would
1160          * depend on the new root would have to update the new root.
1161          */
1162         bch2_btree_node_unlock_write(old, iter);
1163 }
1164
1165 /* Interior node updates: */
1166
1167 static void bch2_insert_fixup_btree_ptr(struct btree_update *as, struct btree *b,
1168                                         struct btree_iter *iter,
1169                                         struct bkey_i *insert,
1170                                         struct btree_node_iter *node_iter)
1171 {
1172         struct bch_fs *c = as->c;
1173         struct bch_fs_usage *fs_usage;
1174         struct bkey_packed *k;
1175         struct bkey tmp;
1176
1177         BUG_ON(insert->k.u64s > bch_btree_keys_u64s_remaining(c, b));
1178
1179         mutex_lock(&c->btree_interior_update_lock);
1180         percpu_down_read(&c->mark_lock);
1181         fs_usage = bch2_fs_usage_scratch_get(c);
1182
1183         bch2_mark_key_locked(c, bkey_i_to_s_c(insert),
1184                              0, 0, fs_usage, 0,
1185                              BTREE_TRIGGER_INSERT);
1186
1187         if (gc_visited(c, gc_pos_btree_node(b)))
1188                 bch2_mark_key_locked(c, bkey_i_to_s_c(insert),
1189                                      0, 0, NULL, 0,
1190                                      BTREE_TRIGGER_INSERT|
1191                                      BTREE_TRIGGER_GC);
1192
1193         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1194                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1195                 bch2_btree_node_iter_advance(node_iter, b);
1196
1197         /*
1198          * If we're overwriting, look up pending delete and mark so that gc
1199          * marks it on the pending delete list:
1200          */
1201         if (k && !bkey_cmp_packed(b, k, &insert->k))
1202                 bch2_btree_node_free_index(as, b,
1203                                            bkey_disassemble(b, k, &tmp),
1204                                            fs_usage);
1205
1206         bch2_fs_usage_apply(c, fs_usage, &as->reserve->disk_res, 0);
1207
1208         bch2_fs_usage_scratch_put(c, fs_usage);
1209         percpu_up_read(&c->mark_lock);
1210         mutex_unlock(&c->btree_interior_update_lock);
1211
1212         bch2_btree_bset_insert_key(iter, b, node_iter, insert);
1213         set_btree_node_dirty(b);
1214         set_btree_node_need_write(b);
1215 }
1216
1217 /*
1218  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1219  * node)
1220  */
1221 static struct btree *__btree_split_node(struct btree_update *as,
1222                                         struct btree *n1,
1223                                         struct btree_iter *iter)
1224 {
1225         size_t nr_packed = 0, nr_unpacked = 0;
1226         struct btree *n2;
1227         struct bset *set1, *set2;
1228         struct bkey_packed *k, *prev = NULL;
1229
1230         n2 = bch2_btree_node_alloc(as, n1->level);
1231
1232         n2->data->max_key       = n1->data->max_key;
1233         n2->data->format        = n1->format;
1234         SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data));
1235         n2->key.k.p = n1->key.k.p;
1236
1237         btree_node_set_format(n2, n2->data->format);
1238
1239         set1 = btree_bset_first(n1);
1240         set2 = btree_bset_first(n2);
1241
1242         /*
1243          * Has to be a linear search because we don't have an auxiliary
1244          * search tree yet
1245          */
1246         k = set1->start;
1247         while (1) {
1248                 struct bkey_packed *n = bkey_next_skip_noops(k, vstruct_last(set1));
1249
1250                 if (n == vstruct_last(set1))
1251                         break;
1252                 if (k->_data - set1->_data >= (le16_to_cpu(set1->u64s) * 3) / 5)
1253                         break;
1254
1255                 if (bkey_packed(k))
1256                         nr_packed++;
1257                 else
1258                         nr_unpacked++;
1259
1260                 prev = k;
1261                 k = n;
1262         }
1263
1264         BUG_ON(!prev);
1265
1266         n1->key.k.p = bkey_unpack_pos(n1, prev);
1267         n1->data->max_key = n1->key.k.p;
1268         n2->data->min_key =
1269                 btree_type_successor(n1->btree_id, n1->key.k.p);
1270
1271         set2->u64s = cpu_to_le16((u64 *) vstruct_end(set1) - (u64 *) k);
1272         set1->u64s = cpu_to_le16(le16_to_cpu(set1->u64s) - le16_to_cpu(set2->u64s));
1273
1274         set_btree_bset_end(n1, n1->set);
1275         set_btree_bset_end(n2, n2->set);
1276
1277         n2->nr.live_u64s        = le16_to_cpu(set2->u64s);
1278         n2->nr.bset_u64s[0]     = le16_to_cpu(set2->u64s);
1279         n2->nr.packed_keys      = n1->nr.packed_keys - nr_packed;
1280         n2->nr.unpacked_keys    = n1->nr.unpacked_keys - nr_unpacked;
1281
1282         n1->nr.live_u64s        = le16_to_cpu(set1->u64s);
1283         n1->nr.bset_u64s[0]     = le16_to_cpu(set1->u64s);
1284         n1->nr.packed_keys      = nr_packed;
1285         n1->nr.unpacked_keys    = nr_unpacked;
1286
1287         BUG_ON(!set1->u64s);
1288         BUG_ON(!set2->u64s);
1289
1290         memcpy_u64s(set2->start,
1291                     vstruct_end(set1),
1292                     le16_to_cpu(set2->u64s));
1293
1294         btree_node_reset_sib_u64s(n1);
1295         btree_node_reset_sib_u64s(n2);
1296
1297         bch2_verify_btree_nr_keys(n1);
1298         bch2_verify_btree_nr_keys(n2);
1299
1300         if (n1->level) {
1301                 btree_node_interior_verify(n1);
1302                 btree_node_interior_verify(n2);
1303         }
1304
1305         return n2;
1306 }
1307
1308 /*
1309  * For updates to interior nodes, we've got to do the insert before we split
1310  * because the stuff we're inserting has to be inserted atomically. Post split,
1311  * the keys might have to go in different nodes and the split would no longer be
1312  * atomic.
1313  *
1314  * Worse, if the insert is from btree node coalescing, if we do the insert after
1315  * we do the split (and pick the pivot) - the pivot we pick might be between
1316  * nodes that were coalesced, and thus in the middle of a child node post
1317  * coalescing:
1318  */
1319 static void btree_split_insert_keys(struct btree_update *as, struct btree *b,
1320                                     struct btree_iter *iter,
1321                                     struct keylist *keys)
1322 {
1323         struct btree_node_iter node_iter;
1324         struct bkey_i *k = bch2_keylist_front(keys);
1325         struct bkey_packed *src, *dst, *n;
1326         struct bset *i;
1327
1328         BUG_ON(btree_node_type(b) != BKEY_TYPE_BTREE);
1329
1330         bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
1331
1332         while (!bch2_keylist_empty(keys)) {
1333                 k = bch2_keylist_front(keys);
1334
1335                 BUG_ON(bch_keylist_u64s(keys) >
1336                        bch_btree_keys_u64s_remaining(as->c, b));
1337                 BUG_ON(bkey_cmp(k->k.p, b->data->min_key) < 0);
1338                 BUG_ON(bkey_cmp(k->k.p, b->data->max_key) > 0);
1339
1340                 bch2_insert_fixup_btree_ptr(as, b, iter, k, &node_iter);
1341                 bch2_keylist_pop_front(keys);
1342         }
1343
1344         /*
1345          * We can't tolerate whiteouts here - with whiteouts there can be
1346          * duplicate keys, and it would be rather bad if we picked a duplicate
1347          * for the pivot:
1348          */
1349         i = btree_bset_first(b);
1350         src = dst = i->start;
1351         while (src != vstruct_last(i)) {
1352                 n = bkey_next_skip_noops(src, vstruct_last(i));
1353                 if (!bkey_deleted(src)) {
1354                         memmove_u64s_down(dst, src, src->u64s);
1355                         dst = bkey_next(dst);
1356                 }
1357                 src = n;
1358         }
1359
1360         i->u64s = cpu_to_le16((u64 *) dst - i->_data);
1361         set_btree_bset_end(b, b->set);
1362
1363         BUG_ON(b->nsets != 1 ||
1364                b->nr.live_u64s != le16_to_cpu(btree_bset_first(b)->u64s));
1365
1366         btree_node_interior_verify(b);
1367 }
1368
1369 static void btree_split(struct btree_update *as, struct btree *b,
1370                         struct btree_iter *iter, struct keylist *keys,
1371                         unsigned flags)
1372 {
1373         struct bch_fs *c = as->c;
1374         struct btree *parent = btree_node_parent(iter, b);
1375         struct btree *n1, *n2 = NULL, *n3 = NULL;
1376         u64 start_time = local_clock();
1377
1378         BUG_ON(!parent && (b != btree_node_root(c, b)));
1379         BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->level));
1380
1381         bch2_btree_interior_update_will_free_node(as, b);
1382
1383         n1 = bch2_btree_node_alloc_replacement(as, b);
1384
1385         if (keys)
1386                 btree_split_insert_keys(as, n1, iter, keys);
1387
1388         if (bset_u64s(&n1->set[0]) > BTREE_SPLIT_THRESHOLD(c)) {
1389                 trace_btree_split(c, b);
1390
1391                 n2 = __btree_split_node(as, n1, iter);
1392
1393                 bch2_btree_build_aux_trees(n2);
1394                 bch2_btree_build_aux_trees(n1);
1395                 six_unlock_write(&n2->lock);
1396                 six_unlock_write(&n1->lock);
1397
1398                 bch2_btree_node_write(c, n2, SIX_LOCK_intent);
1399
1400                 /*
1401                  * Note that on recursive parent_keys == keys, so we
1402                  * can't start adding new keys to parent_keys before emptying it
1403                  * out (which we did with btree_split_insert_keys() above)
1404                  */
1405                 bch2_keylist_add(&as->parent_keys, &n1->key);
1406                 bch2_keylist_add(&as->parent_keys, &n2->key);
1407
1408                 if (!parent) {
1409                         /* Depth increases, make a new root */
1410                         n3 = __btree_root_alloc(as, b->level + 1);
1411
1412                         n3->sib_u64s[0] = U16_MAX;
1413                         n3->sib_u64s[1] = U16_MAX;
1414
1415                         btree_split_insert_keys(as, n3, iter, &as->parent_keys);
1416
1417                         bch2_btree_node_write(c, n3, SIX_LOCK_intent);
1418                 }
1419         } else {
1420                 trace_btree_compact(c, b);
1421
1422                 bch2_btree_build_aux_trees(n1);
1423                 six_unlock_write(&n1->lock);
1424
1425                 bch2_keylist_add(&as->parent_keys, &n1->key);
1426         }
1427
1428         bch2_btree_node_write(c, n1, SIX_LOCK_intent);
1429
1430         /* New nodes all written, now make them visible: */
1431
1432         if (parent) {
1433                 /* Split a non root node */
1434                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1435         } else if (n3) {
1436                 bch2_btree_set_root(as, n3, iter);
1437         } else {
1438                 /* Root filled up but didn't need to be split */
1439                 bch2_btree_set_root(as, n1, iter);
1440         }
1441
1442         bch2_open_buckets_put(c, &n1->ob);
1443         if (n2)
1444                 bch2_open_buckets_put(c, &n2->ob);
1445         if (n3)
1446                 bch2_open_buckets_put(c, &n3->ob);
1447
1448         /* Successful split, update the iterator to point to the new nodes: */
1449
1450         six_lock_increment(&b->lock, SIX_LOCK_intent);
1451         bch2_btree_iter_node_drop(iter, b);
1452         if (n3)
1453                 bch2_btree_iter_node_replace(iter, n3);
1454         if (n2)
1455                 bch2_btree_iter_node_replace(iter, n2);
1456         bch2_btree_iter_node_replace(iter, n1);
1457
1458         /*
1459          * The old node must be freed (in memory) _before_ unlocking the new
1460          * nodes - else another thread could re-acquire a read lock on the old
1461          * node after another thread has locked and updated the new node, thus
1462          * seeing stale data:
1463          */
1464         bch2_btree_node_free_inmem(c, b, iter);
1465
1466         if (n3)
1467                 six_unlock_intent(&n3->lock);
1468         if (n2)
1469                 six_unlock_intent(&n2->lock);
1470         six_unlock_intent(&n1->lock);
1471
1472         bch2_btree_trans_verify_locks(iter->trans);
1473
1474         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_split],
1475                                start_time);
1476 }
1477
1478 static void
1479 bch2_btree_insert_keys_interior(struct btree_update *as, struct btree *b,
1480                                 struct btree_iter *iter, struct keylist *keys)
1481 {
1482         struct btree_iter *linked;
1483         struct btree_node_iter node_iter;
1484         struct bkey_i *insert = bch2_keylist_front(keys);
1485         struct bkey_packed *k;
1486
1487         /* Don't screw up @iter's position: */
1488         node_iter = iter->l[b->level].iter;
1489
1490         /*
1491          * btree_split(), btree_gc_coalesce() will insert keys before
1492          * the iterator's current position - they know the keys go in
1493          * the node the iterator points to:
1494          */
1495         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1496                (bkey_cmp_packed(b, k, &insert->k) >= 0))
1497                 ;
1498
1499         while (!bch2_keylist_empty(keys)) {
1500                 insert = bch2_keylist_front(keys);
1501
1502                 bch2_insert_fixup_btree_ptr(as, b, iter, insert, &node_iter);
1503                 bch2_keylist_pop_front(keys);
1504         }
1505
1506         btree_update_updated_node(as, b);
1507
1508         trans_for_each_iter_with_node(iter->trans, b, linked)
1509                 bch2_btree_node_iter_peek(&linked->l[b->level].iter, b);
1510
1511         bch2_btree_iter_verify(iter, b);
1512 }
1513
1514 /**
1515  * bch_btree_insert_node - insert bkeys into a given btree node
1516  *
1517  * @iter:               btree iterator
1518  * @keys:               list of keys to insert
1519  * @hook:               insert callback
1520  * @persistent:         if not null, @persistent will wait on journal write
1521  *
1522  * Inserts as many keys as it can into a given btree node, splitting it if full.
1523  * If a split occurred, this function will return early. This can only happen
1524  * for leaf nodes -- inserts into interior nodes have to be atomic.
1525  */
1526 void bch2_btree_insert_node(struct btree_update *as, struct btree *b,
1527                             struct btree_iter *iter, struct keylist *keys,
1528                             unsigned flags)
1529 {
1530         struct bch_fs *c = as->c;
1531         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1532         int old_live_u64s = b->nr.live_u64s;
1533         int live_u64s_added, u64s_added;
1534
1535         BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->level));
1536         BUG_ON(!b->level);
1537         BUG_ON(!as || as->b);
1538         bch2_verify_keylist_sorted(keys);
1539
1540         if (as->must_rewrite)
1541                 goto split;
1542
1543         bch2_btree_node_lock_for_insert(c, b, iter);
1544
1545         if (!bch2_btree_node_insert_fits(c, b, bch_keylist_u64s(keys))) {
1546                 bch2_btree_node_unlock_write(b, iter);
1547                 goto split;
1548         }
1549
1550         bch2_btree_insert_keys_interior(as, b, iter, keys);
1551
1552         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1553         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1554
1555         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1556                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1557         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1558                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1559
1560         if (u64s_added > live_u64s_added &&
1561             bch2_maybe_compact_whiteouts(c, b))
1562                 bch2_btree_iter_reinit_node(iter, b);
1563
1564         bch2_btree_node_unlock_write(b, iter);
1565
1566         btree_node_interior_verify(b);
1567
1568         /*
1569          * when called from the btree_split path the new nodes aren't added to
1570          * the btree iterator yet, so the merge path's unlock/wait/relock dance
1571          * won't work:
1572          */
1573         bch2_foreground_maybe_merge(c, iter, b->level,
1574                                     flags|BTREE_INSERT_NOUNLOCK);
1575         return;
1576 split:
1577         btree_split(as, b, iter, keys, flags);
1578 }
1579
1580 int bch2_btree_split_leaf(struct bch_fs *c, struct btree_iter *iter,
1581                           unsigned flags)
1582 {
1583         struct btree_trans *trans = iter->trans;
1584         struct btree *b = iter->l[0].b;
1585         struct btree_update *as;
1586         struct closure cl;
1587         int ret = 0;
1588         struct btree_iter *linked;
1589
1590         /*
1591          * We already have a disk reservation and open buckets pinned; this
1592          * allocation must not block:
1593          */
1594         trans_for_each_iter(trans, linked)
1595                 if (linked->btree_id == BTREE_ID_EXTENTS)
1596                         flags |= BTREE_INSERT_USE_RESERVE;
1597
1598         closure_init_stack(&cl);
1599
1600         /* Hack, because gc and splitting nodes doesn't mix yet: */
1601         if (!(flags & BTREE_INSERT_GC_LOCK_HELD) &&
1602             !down_read_trylock(&c->gc_lock)) {
1603                 if (flags & BTREE_INSERT_NOUNLOCK)
1604                         return -EINTR;
1605
1606                 bch2_trans_unlock(trans);
1607                 down_read(&c->gc_lock);
1608
1609                 if (!bch2_trans_relock(trans))
1610                         ret = -EINTR;
1611         }
1612
1613         /*
1614          * XXX: figure out how far we might need to split,
1615          * instead of locking/reserving all the way to the root:
1616          */
1617         if (!bch2_btree_iter_upgrade(iter, U8_MAX)) {
1618                 trace_trans_restart_iter_upgrade(trans->ip);
1619                 ret = -EINTR;
1620                 goto out;
1621         }
1622
1623         as = bch2_btree_update_start(c, iter->btree_id,
1624                 btree_update_reserve_required(c, b), flags,
1625                 !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1626         if (IS_ERR(as)) {
1627                 ret = PTR_ERR(as);
1628                 if (ret == -EAGAIN) {
1629                         BUG_ON(flags & BTREE_INSERT_NOUNLOCK);
1630                         bch2_trans_unlock(trans);
1631                         ret = -EINTR;
1632                 }
1633                 goto out;
1634         }
1635
1636         btree_split(as, b, iter, NULL, flags);
1637         bch2_btree_update_done(as);
1638
1639         /*
1640          * We haven't successfully inserted yet, so don't downgrade all the way
1641          * back to read locks;
1642          */
1643         __bch2_btree_iter_downgrade(iter, 1);
1644 out:
1645         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1646                 up_read(&c->gc_lock);
1647         closure_sync(&cl);
1648         return ret;
1649 }
1650
1651 void __bch2_foreground_maybe_merge(struct bch_fs *c,
1652                                    struct btree_iter *iter,
1653                                    unsigned level,
1654                                    unsigned flags,
1655                                    enum btree_node_sibling sib)
1656 {
1657         struct btree_trans *trans = iter->trans;
1658         struct btree_update *as;
1659         struct bkey_format_state new_s;
1660         struct bkey_format new_f;
1661         struct bkey_i delete;
1662         struct btree *b, *m, *n, *prev, *next, *parent;
1663         struct closure cl;
1664         size_t sib_u64s;
1665         int ret = 0;
1666
1667         BUG_ON(!btree_node_locked(iter, level));
1668
1669         closure_init_stack(&cl);
1670 retry:
1671         BUG_ON(!btree_node_locked(iter, level));
1672
1673         b = iter->l[level].b;
1674
1675         parent = btree_node_parent(iter, b);
1676         if (!parent)
1677                 goto out;
1678
1679         if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c))
1680                 goto out;
1681
1682         /* XXX: can't be holding read locks */
1683         m = bch2_btree_node_get_sibling(c, iter, b, sib);
1684         if (IS_ERR(m)) {
1685                 ret = PTR_ERR(m);
1686                 goto err;
1687         }
1688
1689         /* NULL means no sibling: */
1690         if (!m) {
1691                 b->sib_u64s[sib] = U16_MAX;
1692                 goto out;
1693         }
1694
1695         if (sib == btree_prev_sib) {
1696                 prev = m;
1697                 next = b;
1698         } else {
1699                 prev = b;
1700                 next = m;
1701         }
1702
1703         bch2_bkey_format_init(&new_s);
1704         __bch2_btree_calc_format(&new_s, b);
1705         __bch2_btree_calc_format(&new_s, m);
1706         new_f = bch2_bkey_format_done(&new_s);
1707
1708         sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1709                 btree_node_u64s_with_format(m, &new_f);
1710
1711         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1712                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1713                 sib_u64s /= 2;
1714                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1715         }
1716
1717         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1718         b->sib_u64s[sib] = sib_u64s;
1719
1720         if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c)) {
1721                 six_unlock_intent(&m->lock);
1722                 goto out;
1723         }
1724
1725         /* We're changing btree topology, doesn't mix with gc: */
1726         if (!(flags & BTREE_INSERT_GC_LOCK_HELD) &&
1727             !down_read_trylock(&c->gc_lock))
1728                 goto err_cycle_gc_lock;
1729
1730         if (!bch2_btree_iter_upgrade(iter, U8_MAX)) {
1731                 ret = -EINTR;
1732                 goto err_unlock;
1733         }
1734
1735         as = bch2_btree_update_start(c, iter->btree_id,
1736                          btree_update_reserve_required(c, parent) + 1,
1737                          BTREE_INSERT_NOFAIL|
1738                          BTREE_INSERT_USE_RESERVE,
1739                          !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1740         if (IS_ERR(as)) {
1741                 ret = PTR_ERR(as);
1742                 goto err_unlock;
1743         }
1744
1745         trace_btree_merge(c, b);
1746
1747         bch2_btree_interior_update_will_free_node(as, b);
1748         bch2_btree_interior_update_will_free_node(as, m);
1749
1750         n = bch2_btree_node_alloc(as, b->level);
1751
1752         n->data->min_key        = prev->data->min_key;
1753         n->data->max_key        = next->data->max_key;
1754         n->data->format         = new_f;
1755         n->key.k.p              = next->key.k.p;
1756
1757         btree_node_set_format(n, new_f);
1758
1759         bch2_btree_sort_into(c, n, prev);
1760         bch2_btree_sort_into(c, n, next);
1761
1762         bch2_btree_build_aux_trees(n);
1763         six_unlock_write(&n->lock);
1764
1765         bkey_init(&delete.k);
1766         delete.k.p = prev->key.k.p;
1767         bch2_keylist_add(&as->parent_keys, &delete);
1768         bch2_keylist_add(&as->parent_keys, &n->key);
1769
1770         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1771
1772         bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1773
1774         bch2_open_buckets_put(c, &n->ob);
1775
1776         six_lock_increment(&b->lock, SIX_LOCK_intent);
1777         bch2_btree_iter_node_drop(iter, b);
1778         bch2_btree_iter_node_drop(iter, m);
1779
1780         bch2_btree_iter_node_replace(iter, n);
1781
1782         bch2_btree_iter_verify(iter, n);
1783
1784         bch2_btree_node_free_inmem(c, b, iter);
1785         bch2_btree_node_free_inmem(c, m, iter);
1786
1787         six_unlock_intent(&n->lock);
1788
1789         bch2_btree_update_done(as);
1790
1791         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1792                 up_read(&c->gc_lock);
1793 out:
1794         bch2_btree_trans_verify_locks(trans);
1795
1796         /*
1797          * Don't downgrade locks here: we're called after successful insert,
1798          * and the caller will downgrade locks after a successful insert
1799          * anyways (in case e.g. a split was required first)
1800          *
1801          * And we're also called when inserting into interior nodes in the
1802          * split path, and downgrading to read locks in there is potentially
1803          * confusing:
1804          */
1805         closure_sync(&cl);
1806         return;
1807
1808 err_cycle_gc_lock:
1809         six_unlock_intent(&m->lock);
1810
1811         if (flags & BTREE_INSERT_NOUNLOCK)
1812                 goto out;
1813
1814         bch2_trans_unlock(trans);
1815
1816         down_read(&c->gc_lock);
1817         up_read(&c->gc_lock);
1818         ret = -EINTR;
1819         goto err;
1820
1821 err_unlock:
1822         six_unlock_intent(&m->lock);
1823         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1824                 up_read(&c->gc_lock);
1825 err:
1826         BUG_ON(ret == -EAGAIN && (flags & BTREE_INSERT_NOUNLOCK));
1827
1828         if ((ret == -EAGAIN || ret == -EINTR) &&
1829             !(flags & BTREE_INSERT_NOUNLOCK)) {
1830                 bch2_trans_unlock(trans);
1831                 closure_sync(&cl);
1832                 ret = bch2_btree_iter_traverse(iter);
1833                 if (ret)
1834                         goto out;
1835
1836                 goto retry;
1837         }
1838
1839         goto out;
1840 }
1841
1842 static int __btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
1843                                 struct btree *b, unsigned flags,
1844                                 struct closure *cl)
1845 {
1846         struct btree *n, *parent = btree_node_parent(iter, b);
1847         struct btree_update *as;
1848
1849         as = bch2_btree_update_start(c, iter->btree_id,
1850                 (parent
1851                  ? btree_update_reserve_required(c, parent)
1852                  : 0) + 1,
1853                 flags, cl);
1854         if (IS_ERR(as)) {
1855                 trace_btree_gc_rewrite_node_fail(c, b);
1856                 return PTR_ERR(as);
1857         }
1858
1859         bch2_btree_interior_update_will_free_node(as, b);
1860
1861         n = bch2_btree_node_alloc_replacement(as, b);
1862
1863         bch2_btree_build_aux_trees(n);
1864         six_unlock_write(&n->lock);
1865
1866         trace_btree_gc_rewrite_node(c, b);
1867
1868         bch2_btree_node_write(c, n, SIX_LOCK_intent);
1869
1870         if (parent) {
1871                 bch2_keylist_add(&as->parent_keys, &n->key);
1872                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1873         } else {
1874                 bch2_btree_set_root(as, n, iter);
1875         }
1876
1877         bch2_open_buckets_put(c, &n->ob);
1878
1879         six_lock_increment(&b->lock, SIX_LOCK_intent);
1880         bch2_btree_iter_node_drop(iter, b);
1881         bch2_btree_iter_node_replace(iter, n);
1882         bch2_btree_node_free_inmem(c, b, iter);
1883         six_unlock_intent(&n->lock);
1884
1885         bch2_btree_update_done(as);
1886         return 0;
1887 }
1888
1889 /**
1890  * bch_btree_node_rewrite - Rewrite/move a btree node
1891  *
1892  * Returns 0 on success, -EINTR or -EAGAIN on failure (i.e.
1893  * btree_check_reserve() has to wait)
1894  */
1895 int bch2_btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
1896                             __le64 seq, unsigned flags)
1897 {
1898         struct btree_trans *trans = iter->trans;
1899         struct closure cl;
1900         struct btree *b;
1901         int ret;
1902
1903         flags |= BTREE_INSERT_NOFAIL;
1904
1905         closure_init_stack(&cl);
1906
1907         bch2_btree_iter_upgrade(iter, U8_MAX);
1908
1909         if (!(flags & BTREE_INSERT_GC_LOCK_HELD)) {
1910                 if (!down_read_trylock(&c->gc_lock)) {
1911                         bch2_trans_unlock(trans);
1912                         down_read(&c->gc_lock);
1913                 }
1914         }
1915
1916         while (1) {
1917                 ret = bch2_btree_iter_traverse(iter);
1918                 if (ret)
1919                         break;
1920
1921                 b = bch2_btree_iter_peek_node(iter);
1922                 if (!b || b->data->keys.seq != seq)
1923                         break;
1924
1925                 ret = __btree_node_rewrite(c, iter, b, flags, &cl);
1926                 if (ret != -EAGAIN &&
1927                     ret != -EINTR)
1928                         break;
1929
1930                 bch2_trans_unlock(trans);
1931                 closure_sync(&cl);
1932         }
1933
1934         bch2_btree_iter_downgrade(iter);
1935
1936         if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1937                 up_read(&c->gc_lock);
1938
1939         closure_sync(&cl);
1940         return ret;
1941 }
1942
1943 static void __bch2_btree_node_update_key(struct bch_fs *c,
1944                                          struct btree_update *as,
1945                                          struct btree_iter *iter,
1946                                          struct btree *b, struct btree *new_hash,
1947                                          struct bkey_i_btree_ptr *new_key)
1948 {
1949         struct btree *parent;
1950         int ret;
1951
1952         /*
1953          * Two corner cases that need to be thought about here:
1954          *
1955          * @b may not be reachable yet - there might be another interior update
1956          * operation waiting on @b to be written, and we're gonna deliver the
1957          * write completion to that interior update operation _before_
1958          * persisting the new_key update
1959          *
1960          * That ends up working without us having to do anything special here:
1961          * the reason is, we do kick off (and do the in memory updates) for the
1962          * update for @new_key before we return, creating a new interior_update
1963          * operation here.
1964          *
1965          * The new interior update operation here will in effect override the
1966          * previous one. The previous one was going to terminate - make @b
1967          * reachable - in one of two ways:
1968          * - updating the btree root pointer
1969          *   In that case,
1970          *   no, this doesn't work. argh.
1971          */
1972
1973         if (b->will_make_reachable)
1974                 as->must_rewrite = true;
1975
1976         btree_interior_update_add_node_reference(as, b);
1977
1978         /*
1979          * XXX: the rest of the update path treats this like we're actually
1980          * inserting a new node and deleting the existing node, so the
1981          * reservation needs to include enough space for @b
1982          *
1983          * that is actually sketch as fuck though and I am surprised the code
1984          * seems to work like that, definitely need to go back and rework it
1985          * into something saner.
1986          *
1987          * (I think @b is just getting double counted until the btree update
1988          * finishes and "deletes" @b on disk)
1989          */
1990         ret = bch2_disk_reservation_add(c, &as->reserve->disk_res,
1991                         c->opts.btree_node_size *
1992                         bch2_bkey_nr_ptrs(bkey_i_to_s_c(&new_key->k_i)),
1993                         BCH_DISK_RESERVATION_NOFAIL);
1994         BUG_ON(ret);
1995
1996         parent = btree_node_parent(iter, b);
1997         if (parent) {
1998                 if (new_hash) {
1999                         bkey_copy(&new_hash->key, &new_key->k_i);
2000                         ret = bch2_btree_node_hash_insert(&c->btree_cache,
2001                                         new_hash, b->level, b->btree_id);
2002                         BUG_ON(ret);
2003                 }
2004
2005                 bch2_keylist_add(&as->parent_keys, &new_key->k_i);
2006                 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, 0);
2007
2008                 if (new_hash) {
2009                         mutex_lock(&c->btree_cache.lock);
2010                         bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
2011
2012                         bch2_btree_node_hash_remove(&c->btree_cache, b);
2013
2014                         bkey_copy(&b->key, &new_key->k_i);
2015                         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
2016                         BUG_ON(ret);
2017                         mutex_unlock(&c->btree_cache.lock);
2018                 } else {
2019                         bkey_copy(&b->key, &new_key->k_i);
2020                 }
2021         } else {
2022                 struct bch_fs_usage *fs_usage;
2023
2024                 BUG_ON(btree_node_root(c, b) != b);
2025
2026                 bch2_btree_node_lock_write(b, iter);
2027
2028                 mutex_lock(&c->btree_interior_update_lock);
2029                 percpu_down_read(&c->mark_lock);
2030                 fs_usage = bch2_fs_usage_scratch_get(c);
2031
2032                 bch2_mark_key_locked(c, bkey_i_to_s_c(&new_key->k_i),
2033                               0, 0, fs_usage, 0,
2034                               BTREE_TRIGGER_INSERT);
2035                 if (gc_visited(c, gc_pos_btree_root(b->btree_id)))
2036                         bch2_mark_key_locked(c, bkey_i_to_s_c(&new_key->k_i),
2037                                              0, 0, NULL, 0,
2038                                              BTREE_TRIGGER_INSERT||
2039                                              BTREE_TRIGGER_GC);
2040
2041                 bch2_btree_node_free_index(as, NULL,
2042                                            bkey_i_to_s_c(&b->key),
2043                                            fs_usage);
2044                 bch2_fs_usage_apply(c, fs_usage, &as->reserve->disk_res, 0);
2045
2046                 bch2_fs_usage_scratch_put(c, fs_usage);
2047                 percpu_up_read(&c->mark_lock);
2048                 mutex_unlock(&c->btree_interior_update_lock);
2049
2050                 if (PTR_HASH(&new_key->k_i) != PTR_HASH(&b->key)) {
2051                         mutex_lock(&c->btree_cache.lock);
2052                         bch2_btree_node_hash_remove(&c->btree_cache, b);
2053
2054                         bkey_copy(&b->key, &new_key->k_i);
2055                         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
2056                         BUG_ON(ret);
2057                         mutex_unlock(&c->btree_cache.lock);
2058                 } else {
2059                         bkey_copy(&b->key, &new_key->k_i);
2060                 }
2061
2062                 btree_update_updated_root(as);
2063                 bch2_btree_node_unlock_write(b, iter);
2064         }
2065
2066         bch2_btree_update_done(as);
2067 }
2068
2069 int bch2_btree_node_update_key(struct bch_fs *c, struct btree_iter *iter,
2070                                struct btree *b,
2071                                struct bkey_i_btree_ptr *new_key)
2072 {
2073         struct btree *parent = btree_node_parent(iter, b);
2074         struct btree_update *as = NULL;
2075         struct btree *new_hash = NULL;
2076         struct closure cl;
2077         int ret;
2078
2079         closure_init_stack(&cl);
2080
2081         if (!bch2_btree_iter_upgrade(iter, U8_MAX))
2082                 return -EINTR;
2083
2084         if (!down_read_trylock(&c->gc_lock)) {
2085                 bch2_trans_unlock(iter->trans);
2086                 down_read(&c->gc_lock);
2087
2088                 if (!bch2_trans_relock(iter->trans)) {
2089                         ret = -EINTR;
2090                         goto err;
2091                 }
2092         }
2093
2094         /* check PTR_HASH() after @b is locked by btree_iter_traverse(): */
2095         if (PTR_HASH(&new_key->k_i) != PTR_HASH(&b->key)) {
2096                 /* bch2_btree_reserve_get will unlock */
2097                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2098                 if (ret) {
2099                         bch2_trans_unlock(iter->trans);
2100                         up_read(&c->gc_lock);
2101                         closure_sync(&cl);
2102                         down_read(&c->gc_lock);
2103
2104                         if (!bch2_trans_relock(iter->trans)) {
2105                                 ret = -EINTR;
2106                                 goto err;
2107                         }
2108                 }
2109
2110                 new_hash = bch2_btree_node_mem_alloc(c);
2111         }
2112
2113         as = bch2_btree_update_start(c, iter->btree_id,
2114                 parent ? btree_update_reserve_required(c, parent) : 0,
2115                 BTREE_INSERT_NOFAIL|
2116                 BTREE_INSERT_USE_RESERVE|
2117                 BTREE_INSERT_USE_ALLOC_RESERVE,
2118                 &cl);
2119
2120         if (IS_ERR(as)) {
2121                 ret = PTR_ERR(as);
2122                 if (ret == -EAGAIN)
2123                         ret = -EINTR;
2124
2125                 if (ret != -EINTR)
2126                         goto err;
2127
2128                 bch2_trans_unlock(iter->trans);
2129                 up_read(&c->gc_lock);
2130                 closure_sync(&cl);
2131                 down_read(&c->gc_lock);
2132
2133                 if (!bch2_trans_relock(iter->trans))
2134                         goto err;
2135         }
2136
2137         ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(&new_key->k_i));
2138         if (ret)
2139                 goto err_free_update;
2140
2141         __bch2_btree_node_update_key(c, as, iter, b, new_hash, new_key);
2142
2143         bch2_btree_iter_downgrade(iter);
2144 err:
2145         if (new_hash) {
2146                 mutex_lock(&c->btree_cache.lock);
2147                 list_move(&new_hash->list, &c->btree_cache.freeable);
2148                 mutex_unlock(&c->btree_cache.lock);
2149
2150                 six_unlock_write(&new_hash->lock);
2151                 six_unlock_intent(&new_hash->lock);
2152         }
2153         up_read(&c->gc_lock);
2154         closure_sync(&cl);
2155         return ret;
2156 err_free_update:
2157         bch2_btree_update_free(as);
2158         goto err;
2159 }
2160
2161 /* Init code: */
2162
2163 /*
2164  * Only for filesystem bringup, when first reading the btree roots or allocating
2165  * btree roots when initializing a new filesystem:
2166  */
2167 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2168 {
2169         BUG_ON(btree_node_root(c, b));
2170
2171         __bch2_btree_set_root_inmem(c, b);
2172 }
2173
2174 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2175 {
2176         struct closure cl;
2177         struct btree *b;
2178         int ret;
2179
2180         closure_init_stack(&cl);
2181
2182         do {
2183                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2184                 closure_sync(&cl);
2185         } while (ret);
2186
2187         b = bch2_btree_node_mem_alloc(c);
2188         bch2_btree_cache_cannibalize_unlock(c);
2189
2190         set_btree_node_fake(b);
2191         b->level        = 0;
2192         b->btree_id     = id;
2193
2194         bkey_btree_ptr_init(&b->key);
2195         b->key.k.p = POS_MAX;
2196         PTR_HASH(&b->key) = U64_MAX - id;
2197
2198         bch2_bset_init_first(b, &b->data->keys);
2199         bch2_btree_build_aux_trees(b);
2200
2201         b->data->flags = 0;
2202         b->data->min_key = POS_MIN;
2203         b->data->max_key = POS_MAX;
2204         b->data->format = bch2_btree_calc_format(b);
2205         btree_node_set_format(b, b->data->format);
2206
2207         ret = bch2_btree_node_hash_insert(&c->btree_cache, b, b->level, b->btree_id);
2208         BUG_ON(ret);
2209
2210         __bch2_btree_set_root_inmem(c, b);
2211
2212         six_unlock_write(&b->lock);
2213         six_unlock_intent(&b->lock);
2214 }
2215
2216 ssize_t bch2_btree_updates_print(struct bch_fs *c, char *buf)
2217 {
2218         struct printbuf out = _PBUF(buf, PAGE_SIZE);
2219         struct btree_update *as;
2220
2221         mutex_lock(&c->btree_interior_update_lock);
2222         list_for_each_entry(as, &c->btree_interior_update_list, list)
2223                 pr_buf(&out, "%p m %u w %u r %u j %llu\n",
2224                        as,
2225                        as->mode,
2226                        as->nodes_written,
2227                        atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2228                        as->journal.seq);
2229         mutex_unlock(&c->btree_interior_update_lock);
2230
2231         return out.pos - buf;
2232 }
2233
2234 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *c)
2235 {
2236         size_t ret = 0;
2237         struct list_head *i;
2238
2239         mutex_lock(&c->btree_interior_update_lock);
2240         list_for_each(i, &c->btree_interior_update_list)
2241                 ret++;
2242         mutex_unlock(&c->btree_interior_update_lock);
2243
2244         return ret;
2245 }