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 "error.h"
  15 #include "extents.h"
  16 #include "journal.h"
  17 #include "journal_reclaim.h"
  18 #include "keylist.h"
  19 #include "recovery.h"
  20 #include "replicas.h"
  21 #include "super-io.h"
  22
  23 #include <linux/random.h>
  24 #include <trace/events/bcachefs.h>
  25
  26 static int bch2_btree_insert_node(struct btree_update *, struct btree_trans *,
  27                                   struct btree_path *, struct btree *,
  28                                   struct keylist *, unsigned);
  29 static void bch2_btree_update_add_new_node(struct btree_update *, struct btree *);
  30
  31 static struct btree_path *get_unlocked_mut_path(struct btree_trans *trans,
  32                                                 enum btree_id btree_id,
  33                                                 unsigned level,
  34                                                 struct bpos pos)
  35 {
  36         struct btree_path *path;
  37
  38         path = bch2_path_get(trans, btree_id, pos, level + 1, level,
  39                              BTREE_ITER_NOPRESERVE|
  40                              BTREE_ITER_INTENT, _RET_IP_);
  41         path = bch2_btree_path_make_mut(trans, path, true, _RET_IP_);
  42         bch2_btree_path_downgrade(trans, path);
  43         __bch2_btree_path_unlock(trans, path);
  44         return path;
  45 }
  46
  47 /* Debug code: */
  48
  49 /*
  50  * Verify that child nodes correctly span parent node's range:
  51  */
  52 static void btree_node_interior_verify(struct bch_fs *c, struct btree *b)
  53 {
  54 #ifdef CONFIG_BCACHEFS_DEBUG
  55         struct bpos next_node = b->data->min_key;
  56         struct btree_node_iter iter;
  57         struct bkey_s_c k;
  58         struct bkey_s_c_btree_ptr_v2 bp;
  59         struct bkey unpacked;
  60         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
  61
  62         BUG_ON(!b->c.level);
  63
  64         if (!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))
  65                 return;
  66
  67         bch2_btree_node_iter_init_from_start(&iter, b);
  68
  69         while (1) {
  70                 k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked);
  71                 if (k.k->type != KEY_TYPE_btree_ptr_v2)
  72                         break;
  73                 bp = bkey_s_c_to_btree_ptr_v2(k);
  74
  75                 if (!bpos_eq(next_node, bp.v->min_key)) {
  76                         bch2_dump_btree_node(c, b);
  77                         bch2_bpos_to_text(&buf1, next_node);
  78                         bch2_bpos_to_text(&buf2, bp.v->min_key);
  79                         panic("expected next min_key %s got %s\n", buf1.buf, buf2.buf);
  80                 }
  81
  82                 bch2_btree_node_iter_advance(&iter, b);
  83
  84                 if (bch2_btree_node_iter_end(&iter)) {
  85                         if (!bpos_eq(k.k->p, b->key.k.p)) {
  86                                 bch2_dump_btree_node(c, b);
  87                                 bch2_bpos_to_text(&buf1, b->key.k.p);
  88                                 bch2_bpos_to_text(&buf2, k.k->p);
  89                                 panic("expected end %s got %s\n", buf1.buf, buf2.buf);
  90                         }
  91                         break;
  92                 }
  93
  94                 next_node = bpos_successor(k.k->p);
  95         }
  96 #endif
  97 }
  98
  99 /* Calculate ideal packed bkey format for new btree nodes: */
 100
 101 void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b)
 102 {
 103         struct bkey_packed *k;
 104         struct bset_tree *t;
 105         struct bkey uk;
 106
 107         for_each_bset(b, t)
 108                 bset_tree_for_each_key(b, t, k)
 109                         if (!bkey_deleted(k)) {
 110                                 uk = bkey_unpack_key(b, k);
 111                                 bch2_bkey_format_add_key(s, &uk);
 112                         }
 113 }
 114
 115 static struct bkey_format bch2_btree_calc_format(struct btree *b)
 116 {
 117         struct bkey_format_state s;
 118
 119         bch2_bkey_format_init(&s);
 120         bch2_bkey_format_add_pos(&s, b->data->min_key);
 121         bch2_bkey_format_add_pos(&s, b->data->max_key);
 122         __bch2_btree_calc_format(&s, b);
 123
 124         return bch2_bkey_format_done(&s);
 125 }
 126
 127 static size_t btree_node_u64s_with_format(struct btree *b,
 128                                           struct bkey_format *new_f)
 129 {
 130         struct bkey_format *old_f = &b->format;
 131
 132         /* stupid integer promotion rules */
 133         ssize_t delta =
 134             (((int) new_f->key_u64s - old_f->key_u64s) *
 135              (int) b->nr.packed_keys) +
 136             (((int) new_f->key_u64s - BKEY_U64s) *
 137              (int) b->nr.unpacked_keys);
 138
 139         BUG_ON(delta + b->nr.live_u64s < 0);
 140
 141         return b->nr.live_u64s + delta;
 142 }
 143
 144 /**
 145  * btree_node_format_fits - check if we could rewrite node with a new format
 146  *
 147  * This assumes all keys can pack with the new format -- it just checks if
 148  * the re-packed keys would fit inside the node itself.
 149  */
 150 bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
 151                                  struct bkey_format *new_f)
 152 {
 153         size_t u64s = btree_node_u64s_with_format(b, new_f);
 154
 155         return __vstruct_bytes(struct btree_node, u64s) < btree_bytes(c);
 156 }
 157
 158 /* Btree node freeing/allocation: */
 159
 160 static void __btree_node_free(struct bch_fs *c, struct btree *b)
 161 {
 162         trace_and_count(c, btree_node_free, c, b);
 163
 164         BUG_ON(btree_node_dirty(b));
 165         BUG_ON(btree_node_need_write(b));
 166         BUG_ON(b == btree_node_root(c, b));
 167         BUG_ON(b->ob.nr);
 168         BUG_ON(!list_empty(&b->write_blocked));
 169         BUG_ON(b->will_make_reachable);
 170
 171         clear_btree_node_noevict(b);
 172
 173         mutex_lock(&c->btree_cache.lock);
 174         list_move(&b->list, &c->btree_cache.freeable);
 175         mutex_unlock(&c->btree_cache.lock);
 176 }
 177
 178 static void bch2_btree_node_free_inmem(struct btree_trans *trans,
 179                                        struct btree_path *path,
 180                                        struct btree *b)
 181 {
 182         struct bch_fs *c = trans->c;
 183         unsigned level = b->c.level;
 184
 185         bch2_btree_node_lock_write_nofail(trans, path, &b->c);
 186         bch2_btree_node_hash_remove(&c->btree_cache, b);
 187         __btree_node_free(c, b);
 188         six_unlock_write(&b->c.lock);
 189         mark_btree_node_locked_noreset(path, level, SIX_LOCK_intent);
 190
 191         trans_for_each_path(trans, path)
 192                 if (path->l[level].b == b) {
 193                         btree_node_unlock(trans, path, level);
 194                         path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
 195                 }
 196 }
 197
 198 static void bch2_btree_node_free_never_used(struct btree_update *as,
 199                                             struct btree_trans *trans,
 200                                             struct btree *b)
 201 {
 202         struct bch_fs *c = as->c;
 203         struct prealloc_nodes *p = &as->prealloc_nodes[b->c.lock.readers != NULL];
 204         struct btree_path *path;
 205         unsigned level = b->c.level;
 206
 207         BUG_ON(!list_empty(&b->write_blocked));
 208         BUG_ON(b->will_make_reachable != (1UL|(unsigned long) as));
 209
 210         b->will_make_reachable = 0;
 211         closure_put(&as->cl);
 212
 213         clear_btree_node_will_make_reachable(b);
 214         clear_btree_node_accessed(b);
 215         clear_btree_node_dirty_acct(c, b);
 216         clear_btree_node_need_write(b);
 217
 218         mutex_lock(&c->btree_cache.lock);
 219         list_del_init(&b->list);
 220         bch2_btree_node_hash_remove(&c->btree_cache, b);
 221         mutex_unlock(&c->btree_cache.lock);
 222
 223         BUG_ON(p->nr >= ARRAY_SIZE(p->b));
 224         p->b[p->nr++] = b;
 225
 226         six_unlock_intent(&b->c.lock);
 227
 228         trans_for_each_path(trans, path)
 229                 if (path->l[level].b == b) {
 230                         btree_node_unlock(trans, path, level);
 231                         path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
 232                 }
 233 }
 234
 235 static struct btree *__bch2_btree_node_alloc(struct btree_trans *trans,
 236                                              struct disk_reservation *res,
 237                                              struct closure *cl,
 238                                              bool interior_node,
 239                                              unsigned flags)
 240 {
 241         struct bch_fs *c = trans->c;
 242         struct write_point *wp;
 243         struct btree *b;
 244         __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
 245         struct open_buckets ob = { .nr = 0 };
 246         struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
 247         unsigned nr_reserve;
 248         enum alloc_reserve alloc_reserve;
 249         int ret;
 250
 251         if (flags & BTREE_INSERT_USE_RESERVE) {
 252                 nr_reserve      = 0;
 253                 alloc_reserve   = RESERVE_btree_movinggc;
 254         } else {
 255                 nr_reserve      = BTREE_NODE_RESERVE;
 256                 alloc_reserve   = RESERVE_btree;
 257         }
 258
 259         mutex_lock(&c->btree_reserve_cache_lock);
 260         if (c->btree_reserve_cache_nr > nr_reserve) {
 261                 struct btree_alloc *a =
 262                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
 263
 264                 ob = a->ob;
 265                 bkey_copy(&tmp.k, &a->k);
 266                 mutex_unlock(&c->btree_reserve_cache_lock);
 267                 goto mem_alloc;
 268         }
 269         mutex_unlock(&c->btree_reserve_cache_lock);
 270
 271 retry:
 272         ret = bch2_alloc_sectors_start_trans(trans,
 273                                       c->opts.metadata_target ?:
 274                                       c->opts.foreground_target,
 275                                       0,
 276                                       writepoint_ptr(&c->btree_write_point),
 277                                       &devs_have,
 278                                       res->nr_replicas,
 279                                       c->opts.metadata_replicas_required,
 280                                       alloc_reserve, 0, cl, &wp);
 281         if (unlikely(ret))
 282                 return ERR_PTR(ret);
 283
 284         if (wp->sectors_free < btree_sectors(c)) {
 285                 struct open_bucket *ob;
 286                 unsigned i;
 287
 288                 open_bucket_for_each(c, &wp->ptrs, ob, i)
 289                         if (ob->sectors_free < btree_sectors(c))
 290                                 ob->sectors_free = 0;
 291
 292                 bch2_alloc_sectors_done(c, wp);
 293                 goto retry;
 294         }
 295
 296         bkey_btree_ptr_v2_init(&tmp.k);
 297         bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, btree_sectors(c), false);
 298
 299         bch2_open_bucket_get(c, wp, &ob);
 300         bch2_alloc_sectors_done(c, wp);
 301 mem_alloc:
 302         b = bch2_btree_node_mem_alloc(c, interior_node);
 303         six_unlock_write(&b->c.lock);
 304         six_unlock_intent(&b->c.lock);
 305
 306         /* we hold cannibalize_lock: */
 307         BUG_ON(IS_ERR(b));
 308         BUG_ON(b->ob.nr);
 309
 310         bkey_copy(&b->key, &tmp.k);
 311         b->ob = ob;
 312
 313         return b;
 314 }
 315
 316 static struct btree *bch2_btree_node_alloc(struct btree_update *as,
 317                                            struct btree_trans *trans,
 318                                            unsigned level)
 319 {
 320         struct bch_fs *c = as->c;
 321         struct btree *b;
 322         struct prealloc_nodes *p = &as->prealloc_nodes[!!level];
 323         int ret;
 324
 325         BUG_ON(level >= BTREE_MAX_DEPTH);
 326         BUG_ON(!p->nr);
 327
 328         b = p->b[--p->nr];
 329
 330         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
 331         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
 332
 333         set_btree_node_accessed(b);
 334         set_btree_node_dirty_acct(c, b);
 335         set_btree_node_need_write(b);
 336
 337         bch2_bset_init_first(b, &b->data->keys);
 338         b->c.level      = level;
 339         b->c.btree_id   = as->btree_id;
 340         b->version_ondisk = c->sb.version;
 341
 342         memset(&b->nr, 0, sizeof(b->nr));
 343         b->data->magic = cpu_to_le64(bset_magic(c));
 344         memset(&b->data->_ptr, 0, sizeof(b->data->_ptr));
 345         b->data->flags = 0;
 346         SET_BTREE_NODE_ID(b->data, as->btree_id);
 347         SET_BTREE_NODE_LEVEL(b->data, level);
 348
 349         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 350                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
 351
 352                 bp->v.mem_ptr           = 0;
 353                 bp->v.seq               = b->data->keys.seq;
 354                 bp->v.sectors_written   = 0;
 355         }
 356
 357         SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
 358
 359         bch2_btree_build_aux_trees(b);
 360
 361         ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
 362         BUG_ON(ret);
 363
 364         trace_and_count(c, btree_node_alloc, c, b);
 365         return b;
 366 }
 367
 368 static void btree_set_min(struct btree *b, struct bpos pos)
 369 {
 370         if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
 371                 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
 372         b->data->min_key = pos;
 373 }
 374
 375 static void btree_set_max(struct btree *b, struct bpos pos)
 376 {
 377         b->key.k.p = pos;
 378         b->data->max_key = pos;
 379 }
 380
 381 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
 382                                                        struct btree_trans *trans,
 383                                                        struct btree *b)
 384 {
 385         struct btree *n = bch2_btree_node_alloc(as, trans, b->c.level);
 386         struct bkey_format format = bch2_btree_calc_format(b);
 387
 388         /*
 389          * The keys might expand with the new format - if they wouldn't fit in
 390          * the btree node anymore, use the old format for now:
 391          */
 392         if (!bch2_btree_node_format_fits(as->c, b, &format))
 393                 format = b->format;
 394
 395         SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
 396
 397         btree_set_min(n, b->data->min_key);
 398         btree_set_max(n, b->data->max_key);
 399
 400         n->data->format         = format;
 401         btree_node_set_format(n, format);
 402
 403         bch2_btree_sort_into(as->c, n, b);
 404
 405         btree_node_reset_sib_u64s(n);
 406         return n;
 407 }
 408
 409 static struct btree *__btree_root_alloc(struct btree_update *as,
 410                                 struct btree_trans *trans, unsigned level)
 411 {
 412         struct btree *b = bch2_btree_node_alloc(as, trans, level);
 413
 414         btree_set_min(b, POS_MIN);
 415         btree_set_max(b, SPOS_MAX);
 416         b->data->format = bch2_btree_calc_format(b);
 417
 418         btree_node_set_format(b, b->data->format);
 419         bch2_btree_build_aux_trees(b);
 420
 421         return b;
 422 }
 423
 424 static void bch2_btree_reserve_put(struct btree_update *as, struct btree_trans *trans)
 425 {
 426         struct bch_fs *c = as->c;
 427         struct prealloc_nodes *p;
 428
 429         for (p = as->prealloc_nodes;
 430              p < as->prealloc_nodes + ARRAY_SIZE(as->prealloc_nodes);
 431              p++) {
 432                 while (p->nr) {
 433                         struct btree *b = p->b[--p->nr];
 434
 435                         mutex_lock(&c->btree_reserve_cache_lock);
 436
 437                         if (c->btree_reserve_cache_nr <
 438                             ARRAY_SIZE(c->btree_reserve_cache)) {
 439                                 struct btree_alloc *a =
 440                                         &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
 441
 442                                 a->ob = b->ob;
 443                                 b->ob.nr = 0;
 444                                 bkey_copy(&a->k, &b->key);
 445                         } else {
 446                                 bch2_open_buckets_put(c, &b->ob);
 447                         }
 448
 449                         mutex_unlock(&c->btree_reserve_cache_lock);
 450
 451                         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
 452                         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
 453                         __btree_node_free(c, b);
 454                         six_unlock_write(&b->c.lock);
 455                         six_unlock_intent(&b->c.lock);
 456                 }
 457         }
 458 }
 459
 460 static int bch2_btree_reserve_get(struct btree_trans *trans,
 461                                   struct btree_update *as,
 462                                   unsigned nr_nodes[2],
 463                                   unsigned flags,
 464                                   struct closure *cl)
 465 {
 466         struct bch_fs *c = as->c;
 467         struct btree *b;
 468         unsigned interior;
 469         int ret = 0;
 470
 471         BUG_ON(nr_nodes[0] + nr_nodes[1] > BTREE_RESERVE_MAX);
 472
 473         /*
 474          * Protects reaping from the btree node cache and using the btree node
 475          * open bucket reserve:
 476          *
 477          * BTREE_INSERT_NOWAIT only applies to btree node allocation, not
 478          * blocking on this lock:
 479          */
 480         ret = bch2_btree_cache_cannibalize_lock(c, cl);
 481         if (ret)
 482                 return ret;
 483
 484         for (interior = 0; interior < 2; interior++) {
 485                 struct prealloc_nodes *p = as->prealloc_nodes + interior;
 486
 487                 while (p->nr < nr_nodes[interior]) {
 488                         b = __bch2_btree_node_alloc(trans, &as->disk_res,
 489                                         flags & BTREE_INSERT_NOWAIT ? NULL : cl,
 490                                         interior, flags);
 491                         if (IS_ERR(b)) {
 492                                 ret = PTR_ERR(b);
 493                                 goto err;
 494                         }
 495
 496                         p->b[p->nr++] = b;
 497                 }
 498         }
 499 err:
 500         bch2_btree_cache_cannibalize_unlock(c);
 501         return ret;
 502 }
 503
 504 /* Asynchronous interior node update machinery */
 505
 506 static void bch2_btree_update_free(struct btree_update *as, struct btree_trans *trans)
 507 {
 508         struct bch_fs *c = as->c;
 509
 510         if (as->took_gc_lock)
 511                 up_read(&c->gc_lock);
 512         as->took_gc_lock = false;
 513
 514         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 515
 516         bch2_journal_pin_drop(&c->journal, &as->journal);
 517         bch2_journal_pin_flush(&c->journal, &as->journal);
 518         bch2_disk_reservation_put(c, &as->disk_res);
 519         bch2_btree_reserve_put(as, trans);
 520
 521         bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_total],
 522                                as->start_time);
 523
 524         mutex_lock(&c->btree_interior_update_lock);
 525         list_del(&as->unwritten_list);
 526         list_del(&as->list);
 527
 528         closure_debug_destroy(&as->cl);
 529         mempool_free(as, &c->btree_interior_update_pool);
 530
 531         /*
 532          * Have to do the wakeup with btree_interior_update_lock still held,
 533          * since being on btree_interior_update_list is our ref on @c:
 534          */
 535         closure_wake_up(&c->btree_interior_update_wait);
 536
 537         mutex_unlock(&c->btree_interior_update_lock);
 538 }
 539
 540 static void btree_update_add_key(struct btree_update *as,
 541                                  struct keylist *keys, struct btree *b)
 542 {
 543         struct bkey_i *k = &b->key;
 544
 545         BUG_ON(bch2_keylist_u64s(keys) + k->k.u64s >
 546                ARRAY_SIZE(as->_old_keys));
 547
 548         bkey_copy(keys->top, k);
 549         bkey_i_to_btree_ptr_v2(keys->top)->v.mem_ptr = b->c.level + 1;
 550
 551         bch2_keylist_push(keys);
 552 }
 553
 554 /*
 555  * The transactional part of an interior btree node update, where we journal the
 556  * update we did to the interior node and update alloc info:
 557  */
 558 static int btree_update_nodes_written_trans(struct btree_trans *trans,
 559                                             struct btree_update *as)
 560 {
 561         struct bkey_i *k;
 562         int ret;
 563
 564         ret = darray_make_room(&trans->extra_journal_entries, as->journal_u64s);
 565         if (ret)
 566                 return ret;
 567
 568         memcpy(&darray_top(trans->extra_journal_entries),
 569                as->journal_entries,
 570                as->journal_u64s * sizeof(u64));
 571         trans->extra_journal_entries.nr += as->journal_u64s;
 572
 573         trans->journal_pin = &as->journal;
 574
 575         for_each_keylist_key(&as->old_keys, k) {
 576                 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
 577
 578                 ret = bch2_trans_mark_old(trans, as->btree_id, level, bkey_i_to_s_c(k), 0);
 579                 if (ret)
 580                         return ret;
 581         }
 582
 583         for_each_keylist_key(&as->new_keys, k) {
 584                 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
 585
 586                 ret = bch2_trans_mark_new(trans, as->btree_id, level, k, 0);
 587                 if (ret)
 588                         return ret;
 589         }
 590
 591         return 0;
 592 }
 593
 594 static void btree_update_nodes_written(struct btree_update *as)
 595 {
 596         struct bch_fs *c = as->c;
 597         struct btree *b;
 598         struct btree_trans trans;
 599         u64 journal_seq = 0;
 600         unsigned i;
 601         int ret;
 602
 603         bch2_trans_init(&trans, c, 0, 512);
 604         /*
 605          * If we're already in an error state, it might be because a btree node
 606          * was never written, and we might be trying to free that same btree
 607          * node here, but it won't have been marked as allocated and we'll see
 608          * spurious disk usage inconsistencies in the transactional part below
 609          * if we don't skip it:
 610          */
 611         ret = bch2_journal_error(&c->journal);
 612         if (ret)
 613                 goto err;
 614
 615         /*
 616          * Wait for any in flight writes to finish before we free the old nodes
 617          * on disk:
 618          */
 619         for (i = 0; i < as->nr_old_nodes; i++) {
 620                 __le64 seq;
 621
 622                 b = as->old_nodes[i];
 623
 624                 btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
 625                 seq = b->data ? b->data->keys.seq : 0;
 626                 six_unlock_read(&b->c.lock);
 627
 628                 if (seq == as->old_nodes_seq[i])
 629                         wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight_inner,
 630                                        TASK_UNINTERRUPTIBLE);
 631         }
 632
 633         /*
 634          * We did an update to a parent node where the pointers we added pointed
 635          * to child nodes that weren't written yet: now, the child nodes have
 636          * been written so we can write out the update to the interior node.
 637          */
 638
 639         /*
 640          * We can't call into journal reclaim here: we'd block on the journal
 641          * reclaim lock, but we may need to release the open buckets we have
 642          * pinned in order for other btree updates to make forward progress, and
 643          * journal reclaim does btree updates when flushing bkey_cached entries,
 644          * which may require allocations as well.
 645          */
 646         ret = commit_do(&trans, &as->disk_res, &journal_seq,
 647                         BTREE_INSERT_NOFAIL|
 648                         BTREE_INSERT_NOCHECK_RW|
 649                         BTREE_INSERT_USE_RESERVE|
 650                         BTREE_INSERT_JOURNAL_RECLAIM|
 651                         JOURNAL_WATERMARK_reserved,
 652                         btree_update_nodes_written_trans(&trans, as));
 653         bch2_trans_unlock(&trans);
 654
 655         bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
 656                              "%s(): error %s", __func__, bch2_err_str(ret));
 657 err:
 658         if (as->b) {
 659                 struct btree_path *path;
 660
 661                 b = as->b;
 662                 path = get_unlocked_mut_path(&trans, as->btree_id, b->c.level, b->key.k.p);
 663                 /*
 664                  * @b is the node we did the final insert into:
 665                  *
 666                  * On failure to get a journal reservation, we still have to
 667                  * unblock the write and allow most of the write path to happen
 668                  * so that shutdown works, but the i->journal_seq mechanism
 669                  * won't work to prevent the btree write from being visible (we
 670                  * didn't get a journal sequence number) - instead
 671                  * __bch2_btree_node_write() doesn't do the actual write if
 672                  * we're in journal error state:
 673                  */
 674
 675                 /*
 676                  * Ensure transaction is unlocked before using
 677                  * btree_node_lock_nopath() (the use of which is always suspect,
 678                  * we need to work on removing this in the future)
 679                  *
 680                  * It should be, but get_unlocked_mut_path() -> bch2_path_get()
 681                  * calls bch2_path_upgrade(), before we call path_make_mut(), so
 682                  * we may rarely end up with a locked path besides the one we
 683                  * have here:
 684                  */
 685                 bch2_trans_unlock(&trans);
 686                 btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_intent);
 687                 mark_btree_node_locked(&trans, path, b->c.level, SIX_LOCK_intent);
 688                 bch2_btree_path_level_init(&trans, path, b);
 689
 690                 bch2_btree_node_lock_write_nofail(&trans, path, &b->c);
 691
 692                 mutex_lock(&c->btree_interior_update_lock);
 693
 694                 list_del(&as->write_blocked_list);
 695                 if (list_empty(&b->write_blocked))
 696                         clear_btree_node_write_blocked(b);
 697
 698                 /*
 699                  * Node might have been freed, recheck under
 700                  * btree_interior_update_lock:
 701                  */
 702                 if (as->b == b) {
 703                         struct bset *i = btree_bset_last(b);
 704
 705                         BUG_ON(!b->c.level);
 706                         BUG_ON(!btree_node_dirty(b));
 707
 708                         if (!ret) {
 709                                 i->journal_seq = cpu_to_le64(
 710                                                              max(journal_seq,
 711                                                                  le64_to_cpu(i->journal_seq)));
 712
 713                                 bch2_btree_add_journal_pin(c, b, journal_seq);
 714                         } else {
 715                                 /*
 716                                  * If we didn't get a journal sequence number we
 717                                  * can't write this btree node, because recovery
 718                                  * won't know to ignore this write:
 719                                  */
 720                                 set_btree_node_never_write(b);
 721                         }
 722                 }
 723
 724                 mutex_unlock(&c->btree_interior_update_lock);
 725
 726                 mark_btree_node_locked_noreset(path, b->c.level, SIX_LOCK_intent);
 727                 six_unlock_write(&b->c.lock);
 728
 729                 btree_node_write_if_need(c, b, SIX_LOCK_intent);
 730                 btree_node_unlock(&trans, path, b->c.level);
 731                 bch2_path_put(&trans, path, true);
 732         }
 733
 734         bch2_journal_pin_drop(&c->journal, &as->journal);
 735
 736         bch2_journal_preres_put(&c->journal, &as->journal_preres);
 737
 738         mutex_lock(&c->btree_interior_update_lock);
 739         for (i = 0; i < as->nr_new_nodes; i++) {
 740                 b = as->new_nodes[i];
 741
 742                 BUG_ON(b->will_make_reachable != (unsigned long) as);
 743                 b->will_make_reachable = 0;
 744                 clear_btree_node_will_make_reachable(b);
 745         }
 746         mutex_unlock(&c->btree_interior_update_lock);
 747
 748         for (i = 0; i < as->nr_new_nodes; i++) {
 749                 b = as->new_nodes[i];
 750
 751                 btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
 752                 btree_node_write_if_need(c, b, SIX_LOCK_read);
 753                 six_unlock_read(&b->c.lock);
 754         }
 755
 756         for (i = 0; i < as->nr_open_buckets; i++)
 757                 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
 758
 759         bch2_btree_update_free(as, &trans);
 760         bch2_trans_exit(&trans);
 761 }
 762
 763 static void btree_interior_update_work(struct work_struct *work)
 764 {
 765         struct bch_fs *c =
 766                 container_of(work, struct bch_fs, btree_interior_update_work);
 767         struct btree_update *as;
 768
 769         while (1) {
 770                 mutex_lock(&c->btree_interior_update_lock);
 771                 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
 772                                               struct btree_update, unwritten_list);
 773                 if (as && !as->nodes_written)
 774                         as = NULL;
 775                 mutex_unlock(&c->btree_interior_update_lock);
 776
 777                 if (!as)
 778                         break;
 779
 780                 btree_update_nodes_written(as);
 781         }
 782 }
 783
 784 static void btree_update_set_nodes_written(struct closure *cl)
 785 {
 786         struct btree_update *as = container_of(cl, struct btree_update, cl);
 787         struct bch_fs *c = as->c;
 788
 789         mutex_lock(&c->btree_interior_update_lock);
 790         as->nodes_written = true;
 791         mutex_unlock(&c->btree_interior_update_lock);
 792
 793         queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
 794 }
 795
 796 /*
 797  * We're updating @b with pointers to nodes that haven't finished writing yet:
 798  * block @b from being written until @as completes
 799  */
 800 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
 801 {
 802         struct bch_fs *c = as->c;
 803
 804         mutex_lock(&c->btree_interior_update_lock);
 805         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 806
 807         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 808         BUG_ON(!btree_node_dirty(b));
 809
 810         as->mode        = BTREE_INTERIOR_UPDATING_NODE;
 811         as->b           = b;
 812
 813         set_btree_node_write_blocked(b);
 814         list_add(&as->write_blocked_list, &b->write_blocked);
 815
 816         mutex_unlock(&c->btree_interior_update_lock);
 817 }
 818
 819 static void btree_update_reparent(struct btree_update *as,
 820                                   struct btree_update *child)
 821 {
 822         struct bch_fs *c = as->c;
 823
 824         lockdep_assert_held(&c->btree_interior_update_lock);
 825
 826         child->b = NULL;
 827         child->mode = BTREE_INTERIOR_UPDATING_AS;
 828
 829         bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL);
 830 }
 831
 832 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
 833 {
 834         struct bkey_i *insert = &b->key;
 835         struct bch_fs *c = as->c;
 836
 837         BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 838
 839         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
 840                ARRAY_SIZE(as->journal_entries));
 841
 842         as->journal_u64s +=
 843                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
 844                                   BCH_JSET_ENTRY_btree_root,
 845                                   b->c.btree_id, b->c.level,
 846                                   insert, insert->k.u64s);
 847
 848         mutex_lock(&c->btree_interior_update_lock);
 849         list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 850
 851         as->mode        = BTREE_INTERIOR_UPDATING_ROOT;
 852         mutex_unlock(&c->btree_interior_update_lock);
 853 }
 854
 855 /*
 856  * bch2_btree_update_add_new_node:
 857  *
 858  * This causes @as to wait on @b to be written, before it gets to
 859  * bch2_btree_update_nodes_written
 860  *
 861  * Additionally, it sets b->will_make_reachable to prevent any additional writes
 862  * to @b from happening besides the first until @b is reachable on disk
 863  *
 864  * And it adds @b to the list of @as's new nodes, so that we can update sector
 865  * counts in bch2_btree_update_nodes_written:
 866  */
 867 static void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
 868 {
 869         struct bch_fs *c = as->c;
 870
 871         closure_get(&as->cl);
 872
 873         mutex_lock(&c->btree_interior_update_lock);
 874         BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
 875         BUG_ON(b->will_make_reachable);
 876
 877         as->new_nodes[as->nr_new_nodes++] = b;
 878         b->will_make_reachable = 1UL|(unsigned long) as;
 879         set_btree_node_will_make_reachable(b);
 880
 881         mutex_unlock(&c->btree_interior_update_lock);
 882
 883         btree_update_add_key(as, &as->new_keys, b);
 884
 885         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 886                 unsigned bytes = vstruct_end(&b->data->keys) - (void *) b->data;
 887                 unsigned sectors = round_up(bytes, block_bytes(c)) >> 9;
 888
 889                 bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written =
 890                         cpu_to_le16(sectors);
 891         }
 892 }
 893
 894 /*
 895  * returns true if @b was a new node
 896  */
 897 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
 898 {
 899         struct btree_update *as;
 900         unsigned long v;
 901         unsigned i;
 902
 903         mutex_lock(&c->btree_interior_update_lock);
 904         /*
 905          * When b->will_make_reachable != 0, it owns a ref on as->cl that's
 906          * dropped when it gets written by bch2_btree_complete_write - the
 907          * xchg() is for synchronization with bch2_btree_complete_write:
 908          */
 909         v = xchg(&b->will_make_reachable, 0);
 910         clear_btree_node_will_make_reachable(b);
 911         as = (struct btree_update *) (v & ~1UL);
 912
 913         if (!as) {
 914                 mutex_unlock(&c->btree_interior_update_lock);
 915                 return;
 916         }
 917
 918         for (i = 0; i < as->nr_new_nodes; i++)
 919                 if (as->new_nodes[i] == b)
 920                         goto found;
 921
 922         BUG();
 923 found:
 924         array_remove_item(as->new_nodes, as->nr_new_nodes, i);
 925         mutex_unlock(&c->btree_interior_update_lock);
 926
 927         if (v & 1)
 928                 closure_put(&as->cl);
 929 }
 930
 931 static void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
 932 {
 933         while (b->ob.nr)
 934                 as->open_buckets[as->nr_open_buckets++] =
 935                         b->ob.v[--b->ob.nr];
 936 }
 937
 938 /*
 939  * @b is being split/rewritten: it may have pointers to not-yet-written btree
 940  * nodes and thus outstanding btree_updates - redirect @b's
 941  * btree_updates to point to this btree_update:
 942  */
 943 static void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 944                                                       struct btree *b)
 945 {
 946         struct bch_fs *c = as->c;
 947         struct btree_update *p, *n;
 948         struct btree_write *w;
 949
 950         set_btree_node_dying(b);
 951
 952         if (btree_node_fake(b))
 953                 return;
 954
 955         mutex_lock(&c->btree_interior_update_lock);
 956
 957         /*
 958          * Does this node have any btree_update operations preventing
 959          * it from being written?
 960          *
 961          * If so, redirect them to point to this btree_update: we can
 962          * write out our new nodes, but we won't make them visible until those
 963          * operations complete
 964          */
 965         list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
 966                 list_del_init(&p->write_blocked_list);
 967                 btree_update_reparent(as, p);
 968
 969                 /*
 970                  * for flush_held_btree_writes() waiting on updates to flush or
 971                  * nodes to be writeable:
 972                  */
 973                 closure_wake_up(&c->btree_interior_update_wait);
 974         }
 975
 976         clear_btree_node_dirty_acct(c, b);
 977         clear_btree_node_need_write(b);
 978
 979         /*
 980          * Does this node have unwritten data that has a pin on the journal?
 981          *
 982          * If so, transfer that pin to the btree_update operation -
 983          * note that if we're freeing multiple nodes, we only need to keep the
 984          * oldest pin of any of the nodes we're freeing. We'll release the pin
 985          * when the new nodes are persistent and reachable on disk:
 986          */
 987         w = btree_current_write(b);
 988         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 989         bch2_journal_pin_drop(&c->journal, &w->journal);
 990
 991         w = btree_prev_write(b);
 992         bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 993         bch2_journal_pin_drop(&c->journal, &w->journal);
 994
 995         mutex_unlock(&c->btree_interior_update_lock);
 996
 997         /*
 998          * Is this a node that isn't reachable on disk yet?
 999          *
1000          * Nodes that aren't reachable yet have writes blocked until they're
1001          * reachable - now that we've cancelled any pending writes and moved
1002          * things waiting on that write to wait on this update, we can drop this
1003          * node from the list of nodes that the other update is making
1004          * reachable, prior to freeing it:
1005          */
1006         btree_update_drop_new_node(c, b);
1007
1008         btree_update_add_key(as, &as->old_keys, b);
1009
1010         as->old_nodes[as->nr_old_nodes] = b;
1011         as->old_nodes_seq[as->nr_old_nodes] = b->data->keys.seq;
1012         as->nr_old_nodes++;
1013 }
1014
1015 static void bch2_btree_update_done(struct btree_update *as, struct btree_trans *trans)
1016 {
1017         struct bch_fs *c = as->c;
1018         u64 start_time = as->start_time;
1019
1020         BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
1021
1022         if (as->took_gc_lock)
1023                 up_read(&as->c->gc_lock);
1024         as->took_gc_lock = false;
1025
1026         bch2_btree_reserve_put(as, trans);
1027
1028         continue_at(&as->cl, btree_update_set_nodes_written,
1029                     as->c->btree_interior_update_worker);
1030
1031         bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_foreground],
1032                                start_time);
1033 }
1034
1035 static struct btree_update *
1036 bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
1037                         unsigned level, bool split, unsigned flags)
1038 {
1039         struct bch_fs *c = trans->c;
1040         struct btree_update *as;
1041         u64 start_time = local_clock();
1042         int disk_res_flags = (flags & BTREE_INSERT_NOFAIL)
1043                 ? BCH_DISK_RESERVATION_NOFAIL : 0;
1044         unsigned nr_nodes[2] = { 0, 0 };
1045         unsigned update_level = level;
1046         int journal_flags = flags & JOURNAL_WATERMARK_MASK;
1047         int ret = 0;
1048         u32 restart_count = trans->restart_count;
1049
1050         BUG_ON(!path->should_be_locked);
1051
1052         if (flags & BTREE_INSERT_JOURNAL_RECLAIM)
1053                 journal_flags |= JOURNAL_RES_GET_NONBLOCK;
1054
1055         while (1) {
1056                 nr_nodes[!!update_level] += 1 + split;
1057                 update_level++;
1058
1059                 ret = bch2_btree_path_upgrade(trans, path, update_level + 1);
1060                 if (ret)
1061                         return ERR_PTR(ret);
1062
1063                 if (!btree_path_node(path, update_level)) {
1064                         /* Allocating new root? */
1065                         nr_nodes[1] += split;
1066                         update_level = BTREE_MAX_DEPTH;
1067                         break;
1068                 }
1069
1070                 if (bch2_btree_node_insert_fits(c, path->l[update_level].b,
1071                                         BKEY_BTREE_PTR_U64s_MAX * (1 + split)))
1072                         break;
1073
1074                 split = true;
1075         }
1076
1077         if (flags & BTREE_INSERT_GC_LOCK_HELD)
1078                 lockdep_assert_held(&c->gc_lock);
1079         else if (!down_read_trylock(&c->gc_lock)) {
1080                 bch2_trans_unlock(trans);
1081                 down_read(&c->gc_lock);
1082                 ret = bch2_trans_relock(trans);
1083                 if (ret) {
1084                         up_read(&c->gc_lock);
1085                         return ERR_PTR(ret);
1086                 }
1087         }
1088
1089         as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
1090         memset(as, 0, sizeof(*as));
1091         closure_init(&as->cl, NULL);
1092         as->c           = c;
1093         as->start_time  = start_time;
1094         as->mode        = BTREE_INTERIOR_NO_UPDATE;
1095         as->took_gc_lock = !(flags & BTREE_INSERT_GC_LOCK_HELD);
1096         as->btree_id    = path->btree_id;
1097         as->update_level = update_level;
1098         INIT_LIST_HEAD(&as->list);
1099         INIT_LIST_HEAD(&as->unwritten_list);
1100         INIT_LIST_HEAD(&as->write_blocked_list);
1101         bch2_keylist_init(&as->old_keys, as->_old_keys);
1102         bch2_keylist_init(&as->new_keys, as->_new_keys);
1103         bch2_keylist_init(&as->parent_keys, as->inline_keys);
1104
1105         mutex_lock(&c->btree_interior_update_lock);
1106         list_add_tail(&as->list, &c->btree_interior_update_list);
1107         mutex_unlock(&c->btree_interior_update_lock);
1108
1109         /*
1110          * We don't want to allocate if we're in an error state, that can cause
1111          * deadlock on emergency shutdown due to open buckets getting stuck in
1112          * the btree_reserve_cache after allocator shutdown has cleared it out.
1113          * This check needs to come after adding us to the btree_interior_update
1114          * list but before calling bch2_btree_reserve_get, to synchronize with
1115          * __bch2_fs_read_only().
1116          */
1117         ret = bch2_journal_error(&c->journal);
1118         if (ret)
1119                 goto err;
1120
1121         ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
1122                                       BTREE_UPDATE_JOURNAL_RES,
1123                                       journal_flags|JOURNAL_RES_GET_NONBLOCK);
1124         if (ret) {
1125                 bch2_trans_unlock(trans);
1126
1127                 if (flags & BTREE_INSERT_JOURNAL_RECLAIM) {
1128                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
1129                         goto err;
1130                 }
1131
1132                 ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
1133                                               BTREE_UPDATE_JOURNAL_RES,
1134                                               journal_flags);
1135                 if (ret) {
1136                         trace_and_count(c, trans_restart_journal_preres_get, trans, _RET_IP_, journal_flags);
1137                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_journal_preres_get);
1138                         goto err;
1139                 }
1140
1141                 ret = bch2_trans_relock(trans);
1142                 if (ret)
1143                         goto err;
1144         }
1145
1146         ret = bch2_disk_reservation_get(c, &as->disk_res,
1147                         (nr_nodes[0] + nr_nodes[1]) * btree_sectors(c),
1148                         c->opts.metadata_replicas,
1149                         disk_res_flags);
1150         if (ret)
1151                 goto err;
1152
1153         ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, NULL);
1154         if (bch2_err_matches(ret, ENOSPC) ||
1155             bch2_err_matches(ret, ENOMEM)) {
1156                 struct closure cl;
1157
1158                 closure_init_stack(&cl);
1159
1160                 do {
1161                         ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, &cl);
1162
1163                         bch2_trans_unlock(trans);
1164                         closure_sync(&cl);
1165                 } while (bch2_err_matches(ret, BCH_ERR_operation_blocked));
1166         }
1167
1168         if (ret) {
1169                 trace_and_count(c, btree_reserve_get_fail, trans->fn,
1170                                 _RET_IP_, nr_nodes[0] + nr_nodes[1], ret);
1171                 goto err;
1172         }
1173
1174         ret = bch2_trans_relock(trans);
1175         if (ret)
1176                 goto err;
1177
1178         bch2_trans_verify_not_restarted(trans, restart_count);
1179         return as;
1180 err:
1181         bch2_btree_update_free(as, trans);
1182         return ERR_PTR(ret);
1183 }
1184
1185 /* Btree root updates: */
1186
1187 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1188 {
1189         /* Root nodes cannot be reaped */
1190         mutex_lock(&c->btree_cache.lock);
1191         list_del_init(&b->list);
1192         mutex_unlock(&c->btree_cache.lock);
1193
1194         mutex_lock(&c->btree_root_lock);
1195         BUG_ON(btree_node_root(c, b) &&
1196                (b->c.level < btree_node_root(c, b)->c.level ||
1197                 !btree_node_dying(btree_node_root(c, b))));
1198
1199         btree_node_root(c, b) = b;
1200         mutex_unlock(&c->btree_root_lock);
1201
1202         bch2_recalc_btree_reserve(c);
1203 }
1204
1205 /**
1206  * bch_btree_set_root - update the root in memory and on disk
1207  *
1208  * To ensure forward progress, the current task must not be holding any
1209  * btree node write locks. However, you must hold an intent lock on the
1210  * old root.
1211  *
1212  * Note: This allocates a journal entry but doesn't add any keys to
1213  * it.  All the btree roots are part of every journal write, so there
1214  * is nothing new to be done.  This just guarantees that there is a
1215  * journal write.
1216  */
1217 static void bch2_btree_set_root(struct btree_update *as,
1218                                 struct btree_trans *trans,
1219                                 struct btree_path *path,
1220                                 struct btree *b)
1221 {
1222         struct bch_fs *c = as->c;
1223         struct btree *old;
1224
1225         trace_and_count(c, btree_node_set_root, c, b);
1226
1227         old = btree_node_root(c, b);
1228
1229         /*
1230          * Ensure no one is using the old root while we switch to the
1231          * new root:
1232          */
1233         bch2_btree_node_lock_write_nofail(trans, path, &old->c);
1234
1235         bch2_btree_set_root_inmem(c, b);
1236
1237         btree_update_updated_root(as, b);
1238
1239         /*
1240          * Unlock old root after new root is visible:
1241          *
1242          * The new root isn't persistent, but that's ok: we still have
1243          * an intent lock on the new root, and any updates that would
1244          * depend on the new root would have to update the new root.
1245          */
1246         bch2_btree_node_unlock_write(trans, path, old);
1247 }
1248
1249 /* Interior node updates: */
1250
1251 static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
1252                                         struct btree_trans *trans,
1253                                         struct btree_path *path,
1254                                         struct btree *b,
1255                                         struct btree_node_iter *node_iter,
1256                                         struct bkey_i *insert)
1257 {
1258         struct bch_fs *c = as->c;
1259         struct bkey_packed *k;
1260         struct printbuf buf = PRINTBUF;
1261         unsigned long old, new, v;
1262
1263         BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 &&
1264                !btree_ptr_sectors_written(insert));
1265
1266         if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
1267                 bch2_journal_key_overwritten(c, b->c.btree_id, b->c.level, insert->k.p);
1268
1269         if (bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1270                               btree_node_type(b), WRITE, &buf) ?:
1271             bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert), &buf)) {
1272                 printbuf_reset(&buf);
1273                 prt_printf(&buf, "inserting invalid bkey\n  ");
1274                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
1275                 prt_printf(&buf, "\n  ");
1276                 bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1277                                   btree_node_type(b), WRITE, &buf);
1278                 bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert), &buf);
1279
1280                 bch2_fs_inconsistent(c, "%s", buf.buf);
1281                 dump_stack();
1282         }
1283
1284         BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1285                ARRAY_SIZE(as->journal_entries));
1286
1287         as->journal_u64s +=
1288                 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1289                                   BCH_JSET_ENTRY_btree_keys,
1290                                   b->c.btree_id, b->c.level,
1291                                   insert, insert->k.u64s);
1292
1293         while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1294                bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1295                 bch2_btree_node_iter_advance(node_iter, b);
1296
1297         bch2_btree_bset_insert_key(trans, path, b, node_iter, insert);
1298         set_btree_node_dirty_acct(c, b);
1299
1300         v = READ_ONCE(b->flags);
1301         do {
1302                 old = new = v;
1303
1304                 new &= ~BTREE_WRITE_TYPE_MASK;
1305                 new |= BTREE_WRITE_interior;
1306                 new |= 1 << BTREE_NODE_need_write;
1307         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1308
1309         printbuf_exit(&buf);
1310 }
1311
1312 static void
1313 __bch2_btree_insert_keys_interior(struct btree_update *as,
1314                                   struct btree_trans *trans,
1315                                   struct btree_path *path,
1316                                   struct btree *b,
1317                                   struct btree_node_iter node_iter,
1318                                   struct keylist *keys)
1319 {
1320         struct bkey_i *insert = bch2_keylist_front(keys);
1321         struct bkey_packed *k;
1322
1323         BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1324
1325         while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1326                (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1327                 ;
1328
1329         while (!bch2_keylist_empty(keys)) {
1330                 struct bkey_i *k = bch2_keylist_front(keys);
1331
1332                 if (bpos_gt(k->k.p, b->key.k.p))
1333                         break;
1334
1335                 bch2_insert_fixup_btree_ptr(as, trans, path, b, &node_iter, k);
1336                 bch2_keylist_pop_front(keys);
1337         }
1338 }
1339
1340 /*
1341  * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1342  * node)
1343  */
1344 static void __btree_split_node(struct btree_update *as,
1345                                struct btree_trans *trans,
1346                                struct btree *b,
1347                                struct btree *n[2])
1348 {
1349         struct bkey_packed *k;
1350         struct bpos n1_pos = POS_MIN;
1351         struct btree_node_iter iter;
1352         struct bset *bsets[2];
1353         struct bkey_format_state format[2];
1354         struct bkey_packed *out[2];
1355         struct bkey uk;
1356         unsigned u64s, n1_u64s = (b->nr.live_u64s * 3) / 5;
1357         int i;
1358
1359         for (i = 0; i < 2; i++) {
1360                 BUG_ON(n[i]->nsets != 1);
1361
1362                 bsets[i] = btree_bset_first(n[i]);
1363                 out[i] = bsets[i]->start;
1364
1365                 SET_BTREE_NODE_SEQ(n[i]->data, BTREE_NODE_SEQ(b->data) + 1);
1366                 bch2_bkey_format_init(&format[i]);
1367         }
1368
1369         u64s = 0;
1370         for_each_btree_node_key(b, k, &iter) {
1371                 if (bkey_deleted(k))
1372                         continue;
1373
1374                 i = u64s >= n1_u64s;
1375                 u64s += k->u64s;
1376                 uk = bkey_unpack_key(b, k);
1377                 if (!i)
1378                         n1_pos = uk.p;
1379                 bch2_bkey_format_add_key(&format[i], &uk);
1380         }
1381
1382         btree_set_min(n[0], b->data->min_key);
1383         btree_set_max(n[0], n1_pos);
1384         btree_set_min(n[1], bpos_successor(n1_pos));
1385         btree_set_max(n[1], b->data->max_key);
1386
1387         for (i = 0; i < 2; i++) {
1388                 bch2_bkey_format_add_pos(&format[i], n[i]->data->min_key);
1389                 bch2_bkey_format_add_pos(&format[i], n[i]->data->max_key);
1390
1391                 n[i]->data->format = bch2_bkey_format_done(&format[i]);
1392                 btree_node_set_format(n[i], n[i]->data->format);
1393         }
1394
1395         u64s = 0;
1396         for_each_btree_node_key(b, k, &iter) {
1397                 if (bkey_deleted(k))
1398                         continue;
1399
1400                 i = u64s >= n1_u64s;
1401                 u64s += k->u64s;
1402
1403                 if (bch2_bkey_transform(&n[i]->format, out[i], bkey_packed(k)
1404                                         ? &b->format: &bch2_bkey_format_current, k))
1405                         out[i]->format = KEY_FORMAT_LOCAL_BTREE;
1406                 else
1407                         bch2_bkey_unpack(b, (void *) out[i], k);
1408
1409                 out[i]->needs_whiteout = false;
1410
1411                 btree_keys_account_key_add(&n[i]->nr, 0, out[i]);
1412                 out[i] = bkey_next(out[i]);
1413         }
1414
1415         for (i = 0; i < 2; i++) {
1416                 bsets[i]->u64s = cpu_to_le16((u64 *) out[i] - bsets[i]->_data);
1417
1418                 BUG_ON(!bsets[i]->u64s);
1419
1420                 set_btree_bset_end(n[i], n[i]->set);
1421
1422                 btree_node_reset_sib_u64s(n[i]);
1423
1424                 bch2_verify_btree_nr_keys(n[i]);
1425
1426                 if (b->c.level)
1427                         btree_node_interior_verify(as->c, n[i]);
1428         }
1429 }
1430
1431 /*
1432  * For updates to interior nodes, we've got to do the insert before we split
1433  * because the stuff we're inserting has to be inserted atomically. Post split,
1434  * the keys might have to go in different nodes and the split would no longer be
1435  * atomic.
1436  *
1437  * Worse, if the insert is from btree node coalescing, if we do the insert after
1438  * we do the split (and pick the pivot) - the pivot we pick might be between
1439  * nodes that were coalesced, and thus in the middle of a child node post
1440  * coalescing:
1441  */
1442 static void btree_split_insert_keys(struct btree_update *as,
1443                                     struct btree_trans *trans,
1444                                     struct btree_path *path,
1445                                     struct btree *b,
1446                                     struct keylist *keys)
1447 {
1448         if (!bch2_keylist_empty(keys) &&
1449             bpos_le(bch2_keylist_front(keys)->k.p, b->data->max_key)) {
1450                 struct btree_node_iter node_iter;
1451
1452                 bch2_btree_node_iter_init(&node_iter, b, &bch2_keylist_front(keys)->k.p);
1453
1454                 __bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
1455
1456                 btree_node_interior_verify(as->c, b);
1457         }
1458 }
1459
1460 static int btree_split(struct btree_update *as, struct btree_trans *trans,
1461                        struct btree_path *path, struct btree *b,
1462                        struct keylist *keys, unsigned flags)
1463 {
1464         struct bch_fs *c = as->c;
1465         struct btree *parent = btree_node_parent(path, b);
1466         struct btree *n1, *n2 = NULL, *n3 = NULL;
1467         struct btree_path *path1 = NULL, *path2 = NULL;
1468         u64 start_time = local_clock();
1469         int ret = 0;
1470
1471         BUG_ON(!parent && (b != btree_node_root(c, b)));
1472         BUG_ON(parent && !btree_node_intent_locked(path, b->c.level + 1));
1473
1474         bch2_btree_interior_update_will_free_node(as, b);
1475
1476         if (b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c)) {
1477                 struct btree *n[2];
1478
1479                 trace_and_count(c, btree_node_split, c, b);
1480
1481                 n[0] = n1 = bch2_btree_node_alloc(as, trans, b->c.level);
1482                 n[1] = n2 = bch2_btree_node_alloc(as, trans, b->c.level);
1483
1484                 __btree_split_node(as, trans, b, n);
1485
1486                 if (keys) {
1487                         btree_split_insert_keys(as, trans, path, n1, keys);
1488                         btree_split_insert_keys(as, trans, path, n2, keys);
1489                         BUG_ON(!bch2_keylist_empty(keys));
1490                 }
1491
1492                 bch2_btree_build_aux_trees(n2);
1493                 bch2_btree_build_aux_trees(n1);
1494
1495                 bch2_btree_update_add_new_node(as, n1);
1496                 bch2_btree_update_add_new_node(as, n2);
1497                 six_unlock_write(&n2->c.lock);
1498                 six_unlock_write(&n1->c.lock);
1499
1500                 path1 = get_unlocked_mut_path(trans, path->btree_id, n1->c.level, n1->key.k.p);
1501                 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1502                 mark_btree_node_locked(trans, path1, n1->c.level, SIX_LOCK_intent);
1503                 bch2_btree_path_level_init(trans, path1, n1);
1504
1505                 path2 = get_unlocked_mut_path(trans, path->btree_id, n2->c.level, n2->key.k.p);
1506                 six_lock_increment(&n2->c.lock, SIX_LOCK_intent);
1507                 mark_btree_node_locked(trans, path2, n2->c.level, SIX_LOCK_intent);
1508                 bch2_btree_path_level_init(trans, path2, n2);
1509
1510                 /*
1511                  * Note that on recursive parent_keys == keys, so we
1512                  * can't start adding new keys to parent_keys before emptying it
1513                  * out (which we did with btree_split_insert_keys() above)
1514                  */
1515                 bch2_keylist_add(&as->parent_keys, &n1->key);
1516                 bch2_keylist_add(&as->parent_keys, &n2->key);
1517
1518                 if (!parent) {
1519                         /* Depth increases, make a new root */
1520                         n3 = __btree_root_alloc(as, trans, b->c.level + 1);
1521
1522                         bch2_btree_update_add_new_node(as, n3);
1523                         six_unlock_write(&n3->c.lock);
1524
1525                         path2->locks_want++;
1526                         BUG_ON(btree_node_locked(path2, n3->c.level));
1527                         six_lock_increment(&n3->c.lock, SIX_LOCK_intent);
1528                         mark_btree_node_locked(trans, path2, n3->c.level, SIX_LOCK_intent);
1529                         bch2_btree_path_level_init(trans, path2, n3);
1530
1531                         n3->sib_u64s[0] = U16_MAX;
1532                         n3->sib_u64s[1] = U16_MAX;
1533
1534                         btree_split_insert_keys(as, trans, path, n3, &as->parent_keys);
1535                 }
1536         } else {
1537                 trace_and_count(c, btree_node_compact, c, b);
1538
1539                 n1 = bch2_btree_node_alloc_replacement(as, trans, b);
1540
1541                 if (keys) {
1542                         btree_split_insert_keys(as, trans, path, n1, keys);
1543                         BUG_ON(!bch2_keylist_empty(keys));
1544                 }
1545
1546                 bch2_btree_build_aux_trees(n1);
1547                 bch2_btree_update_add_new_node(as, n1);
1548                 six_unlock_write(&n1->c.lock);
1549
1550                 path1 = get_unlocked_mut_path(trans, path->btree_id, n1->c.level, n1->key.k.p);
1551                 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1552                 mark_btree_node_locked(trans, path1, n1->c.level, SIX_LOCK_intent);
1553                 bch2_btree_path_level_init(trans, path1, n1);
1554
1555                 if (parent)
1556                         bch2_keylist_add(&as->parent_keys, &n1->key);
1557         }
1558
1559         /* New nodes all written, now make them visible: */
1560
1561         if (parent) {
1562                 /* Split a non root node */
1563                 ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1564                 if (ret)
1565                         goto err;
1566         } else if (n3) {
1567                 bch2_btree_set_root(as, trans, path, n3);
1568         } else {
1569                 /* Root filled up but didn't need to be split */
1570                 bch2_btree_set_root(as, trans, path, n1);
1571         }
1572
1573         if (n3) {
1574                 bch2_btree_update_get_open_buckets(as, n3);
1575                 bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0);
1576         }
1577         if (n2) {
1578                 bch2_btree_update_get_open_buckets(as, n2);
1579                 bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0);
1580         }
1581         bch2_btree_update_get_open_buckets(as, n1);
1582         bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1583
1584         /*
1585          * The old node must be freed (in memory) _before_ unlocking the new
1586          * nodes - else another thread could re-acquire a read lock on the old
1587          * node after another thread has locked and updated the new node, thus
1588          * seeing stale data:
1589          */
1590         bch2_btree_node_free_inmem(trans, path, b);
1591
1592         if (n3)
1593                 bch2_trans_node_add(trans, n3);
1594         if (n2)
1595                 bch2_trans_node_add(trans, n2);
1596         bch2_trans_node_add(trans, n1);
1597
1598         if (n3)
1599                 six_unlock_intent(&n3->c.lock);
1600         if (n2)
1601                 six_unlock_intent(&n2->c.lock);
1602         six_unlock_intent(&n1->c.lock);
1603 out:
1604         if (path2) {
1605                 __bch2_btree_path_unlock(trans, path2);
1606                 bch2_path_put(trans, path2, true);
1607         }
1608         if (path1) {
1609                 __bch2_btree_path_unlock(trans, path1);
1610                 bch2_path_put(trans, path1, true);
1611         }
1612
1613         bch2_trans_verify_locks(trans);
1614
1615         bch2_time_stats_update(&c->times[n2
1616                                ? BCH_TIME_btree_node_split
1617                                : BCH_TIME_btree_node_compact],
1618                                start_time);
1619         return ret;
1620 err:
1621         if (n3)
1622                 bch2_btree_node_free_never_used(as, trans, n3);
1623         if (n2)
1624                 bch2_btree_node_free_never_used(as, trans, n2);
1625         bch2_btree_node_free_never_used(as, trans, n1);
1626         goto out;
1627 }
1628
1629 static void
1630 bch2_btree_insert_keys_interior(struct btree_update *as,
1631                                 struct btree_trans *trans,
1632                                 struct btree_path *path,
1633                                 struct btree *b,
1634                                 struct keylist *keys)
1635 {
1636         struct btree_path *linked;
1637
1638         __bch2_btree_insert_keys_interior(as, trans, path, b,
1639                                           path->l[b->c.level].iter, keys);
1640
1641         btree_update_updated_node(as, b);
1642
1643         trans_for_each_path_with_node(trans, b, linked)
1644                 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1645
1646         bch2_trans_verify_paths(trans);
1647 }
1648
1649 /**
1650  * bch_btree_insert_node - insert bkeys into a given btree node
1651  *
1652  * @iter:               btree iterator
1653  * @keys:               list of keys to insert
1654  * @hook:               insert callback
1655  * @persistent:         if not null, @persistent will wait on journal write
1656  *
1657  * Inserts as many keys as it can into a given btree node, splitting it if full.
1658  * If a split occurred, this function will return early. This can only happen
1659  * for leaf nodes -- inserts into interior nodes have to be atomic.
1660  */
1661 static int bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans,
1662                                   struct btree_path *path, struct btree *b,
1663                                   struct keylist *keys, unsigned flags)
1664 {
1665         struct bch_fs *c = as->c;
1666         int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1667         int old_live_u64s = b->nr.live_u64s;
1668         int live_u64s_added, u64s_added;
1669         int ret;
1670
1671         lockdep_assert_held(&c->gc_lock);
1672         BUG_ON(!btree_node_intent_locked(path, b->c.level));
1673         BUG_ON(!b->c.level);
1674         BUG_ON(!as || as->b);
1675         bch2_verify_keylist_sorted(keys);
1676
1677         if (!(local_clock() & 63))
1678                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_split_race);
1679
1680         ret = bch2_btree_node_lock_write(trans, path, &b->c);
1681         if (ret)
1682                 return ret;
1683
1684         bch2_btree_node_prep_for_write(trans, path, b);
1685
1686         if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1687                 bch2_btree_node_unlock_write(trans, path, b);
1688                 goto split;
1689         }
1690
1691         btree_node_interior_verify(c, b);
1692
1693         bch2_btree_insert_keys_interior(as, trans, path, b, keys);
1694
1695         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1696         u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1697
1698         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1699                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1700         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1701                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1702
1703         if (u64s_added > live_u64s_added &&
1704             bch2_maybe_compact_whiteouts(c, b))
1705                 bch2_trans_node_reinit_iter(trans, b);
1706
1707         bch2_btree_node_unlock_write(trans, path, b);
1708
1709         btree_node_interior_verify(c, b);
1710         return 0;
1711 split:
1712         /*
1713          * We could attempt to avoid the transaction restart, by calling
1714          * bch2_btree_path_upgrade() and allocating more nodes:
1715          */
1716         if (b->c.level >= as->update_level)
1717                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_split_race);
1718
1719         return btree_split(as, trans, path, b, keys, flags);
1720 }
1721
1722 int bch2_btree_split_leaf(struct btree_trans *trans,
1723                           struct btree_path *path,
1724                           unsigned flags)
1725 {
1726         struct btree *b = path_l(path)->b;
1727         struct btree_update *as;
1728         unsigned l;
1729         int ret = 0;
1730
1731         as = bch2_btree_update_start(trans, path, path->level,
1732                                      true, flags);
1733         if (IS_ERR(as))
1734                 return PTR_ERR(as);
1735
1736         ret = btree_split(as, trans, path, b, NULL, flags);
1737         if (ret) {
1738                 bch2_btree_update_free(as, trans);
1739                 return ret;
1740         }
1741
1742         bch2_btree_update_done(as, trans);
1743
1744         for (l = path->level + 1; btree_node_intent_locked(path, l) && !ret; l++)
1745                 ret = bch2_foreground_maybe_merge(trans, path, l, flags);
1746
1747         return ret;
1748 }
1749
1750 int __bch2_foreground_maybe_merge(struct btree_trans *trans,
1751                                   struct btree_path *path,
1752                                   unsigned level,
1753                                   unsigned flags,
1754                                   enum btree_node_sibling sib)
1755 {
1756         struct bch_fs *c = trans->c;
1757         struct btree_path *sib_path = NULL, *new_path = NULL;
1758         struct btree_update *as;
1759         struct bkey_format_state new_s;
1760         struct bkey_format new_f;
1761         struct bkey_i delete;
1762         struct btree *b, *m, *n, *prev, *next, *parent;
1763         struct bpos sib_pos;
1764         size_t sib_u64s;
1765         u64 start_time = local_clock();
1766         int ret = 0;
1767
1768         BUG_ON(!path->should_be_locked);
1769         BUG_ON(!btree_node_locked(path, level));
1770
1771         b = path->l[level].b;
1772
1773         if ((sib == btree_prev_sib && bpos_eq(b->data->min_key, POS_MIN)) ||
1774             (sib == btree_next_sib && bpos_eq(b->data->max_key, SPOS_MAX))) {
1775                 b->sib_u64s[sib] = U16_MAX;
1776                 return 0;
1777         }
1778
1779         sib_pos = sib == btree_prev_sib
1780                 ? bpos_predecessor(b->data->min_key)
1781                 : bpos_successor(b->data->max_key);
1782
1783         sib_path = bch2_path_get(trans, path->btree_id, sib_pos,
1784                                  U8_MAX, level, BTREE_ITER_INTENT, _THIS_IP_);
1785         ret = bch2_btree_path_traverse(trans, sib_path, false);
1786         if (ret)
1787                 goto err;
1788
1789         btree_path_set_should_be_locked(sib_path);
1790
1791         m = sib_path->l[level].b;
1792
1793         if (btree_node_parent(path, b) !=
1794             btree_node_parent(sib_path, m)) {
1795                 b->sib_u64s[sib] = U16_MAX;
1796                 goto out;
1797         }
1798
1799         if (sib == btree_prev_sib) {
1800                 prev = m;
1801                 next = b;
1802         } else {
1803                 prev = b;
1804                 next = m;
1805         }
1806
1807         if (!bpos_eq(bpos_successor(prev->data->max_key), next->data->min_key)) {
1808                 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1809
1810                 bch2_bpos_to_text(&buf1, prev->data->max_key);
1811                 bch2_bpos_to_text(&buf2, next->data->min_key);
1812                 bch_err(c,
1813                         "%s(): btree topology error:\n"
1814                         "  prev ends at   %s\n"
1815                         "  next starts at %s",
1816                         __func__, buf1.buf, buf2.buf);
1817                 printbuf_exit(&buf1);
1818                 printbuf_exit(&buf2);
1819                 bch2_topology_error(c);
1820                 ret = -EIO;
1821                 goto err;
1822         }
1823
1824         bch2_bkey_format_init(&new_s);
1825         bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
1826         __bch2_btree_calc_format(&new_s, prev);
1827         __bch2_btree_calc_format(&new_s, next);
1828         bch2_bkey_format_add_pos(&new_s, next->data->max_key);
1829         new_f = bch2_bkey_format_done(&new_s);
1830
1831         sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1832                 btree_node_u64s_with_format(m, &new_f);
1833
1834         if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1835                 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1836                 sib_u64s /= 2;
1837                 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1838         }
1839
1840         sib_u64s = min(sib_u64s, btree_max_u64s(c));
1841         sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
1842         b->sib_u64s[sib] = sib_u64s;
1843
1844         if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
1845                 goto out;
1846
1847         parent = btree_node_parent(path, b);
1848         as = bch2_btree_update_start(trans, path, level, false,
1849                          BTREE_INSERT_NOFAIL|
1850                          BTREE_INSERT_USE_RESERVE|
1851                          flags);
1852         ret = PTR_ERR_OR_ZERO(as);
1853         if (ret)
1854                 goto err;
1855
1856         trace_and_count(c, btree_node_merge, c, b);
1857
1858         bch2_btree_interior_update_will_free_node(as, b);
1859         bch2_btree_interior_update_will_free_node(as, m);
1860
1861         n = bch2_btree_node_alloc(as, trans, b->c.level);
1862
1863         SET_BTREE_NODE_SEQ(n->data,
1864                            max(BTREE_NODE_SEQ(b->data),
1865                                BTREE_NODE_SEQ(m->data)) + 1);
1866
1867         btree_set_min(n, prev->data->min_key);
1868         btree_set_max(n, next->data->max_key);
1869
1870         n->data->format  = new_f;
1871         btree_node_set_format(n, new_f);
1872
1873         bch2_btree_sort_into(c, n, prev);
1874         bch2_btree_sort_into(c, n, next);
1875
1876         bch2_btree_build_aux_trees(n);
1877         bch2_btree_update_add_new_node(as, n);
1878         six_unlock_write(&n->c.lock);
1879
1880         new_path = get_unlocked_mut_path(trans, path->btree_id, n->c.level, n->key.k.p);
1881         six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1882         mark_btree_node_locked(trans, new_path, n->c.level, SIX_LOCK_intent);
1883         bch2_btree_path_level_init(trans, new_path, n);
1884
1885         bkey_init(&delete.k);
1886         delete.k.p = prev->key.k.p;
1887         bch2_keylist_add(&as->parent_keys, &delete);
1888         bch2_keylist_add(&as->parent_keys, &n->key);
1889
1890         bch2_trans_verify_paths(trans);
1891
1892         ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1893         if (ret)
1894                 goto err_free_update;
1895
1896         bch2_trans_verify_paths(trans);
1897
1898         bch2_btree_update_get_open_buckets(as, n);
1899         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1900
1901         bch2_btree_node_free_inmem(trans, path, b);
1902         bch2_btree_node_free_inmem(trans, sib_path, m);
1903
1904         bch2_trans_node_add(trans, n);
1905
1906         bch2_trans_verify_paths(trans);
1907
1908         six_unlock_intent(&n->c.lock);
1909
1910         bch2_btree_update_done(as, trans);
1911
1912         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time);
1913 out:
1914 err:
1915         if (new_path)
1916                 bch2_path_put(trans, new_path, true);
1917         bch2_path_put(trans, sib_path, true);
1918         bch2_trans_verify_locks(trans);
1919         return ret;
1920 err_free_update:
1921         bch2_btree_node_free_never_used(as, trans, n);
1922         bch2_btree_update_free(as, trans);
1923         goto out;
1924 }
1925
1926 /**
1927  * bch_btree_node_rewrite - Rewrite/move a btree node
1928  */
1929 int bch2_btree_node_rewrite(struct btree_trans *trans,
1930                             struct btree_iter *iter,
1931                             struct btree *b,
1932                             unsigned flags)
1933 {
1934         struct bch_fs *c = trans->c;
1935         struct btree_path *new_path = NULL;
1936         struct btree *n, *parent;
1937         struct btree_update *as;
1938         int ret;
1939
1940         flags |= BTREE_INSERT_NOFAIL;
1941
1942         parent = btree_node_parent(iter->path, b);
1943         as = bch2_btree_update_start(trans, iter->path, b->c.level,
1944                                      false, flags);
1945         ret = PTR_ERR_OR_ZERO(as);
1946         if (ret)
1947                 goto out;
1948
1949         bch2_btree_interior_update_will_free_node(as, b);
1950
1951         n = bch2_btree_node_alloc_replacement(as, trans, b);
1952
1953         bch2_btree_build_aux_trees(n);
1954         bch2_btree_update_add_new_node(as, n);
1955         six_unlock_write(&n->c.lock);
1956
1957         new_path = get_unlocked_mut_path(trans, iter->btree_id, n->c.level, n->key.k.p);
1958         six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1959         mark_btree_node_locked(trans, new_path, n->c.level, SIX_LOCK_intent);
1960         bch2_btree_path_level_init(trans, new_path, n);
1961
1962         trace_and_count(c, btree_node_rewrite, c, b);
1963
1964         if (parent) {
1965                 bch2_keylist_add(&as->parent_keys, &n->key);
1966                 ret = bch2_btree_insert_node(as, trans, iter->path, parent,
1967                                              &as->parent_keys, flags);
1968                 if (ret)
1969                         goto err;
1970         } else {
1971                 bch2_btree_set_root(as, trans, iter->path, n);
1972         }
1973
1974         bch2_btree_update_get_open_buckets(as, n);
1975         bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1976
1977         bch2_btree_node_free_inmem(trans, iter->path, b);
1978
1979         bch2_trans_node_add(trans, n);
1980         six_unlock_intent(&n->c.lock);
1981
1982         bch2_btree_update_done(as, trans);
1983 out:
1984         if (new_path)
1985                 bch2_path_put(trans, new_path, true);
1986         bch2_btree_path_downgrade(trans, iter->path);
1987         return ret;
1988 err:
1989         bch2_btree_node_free_never_used(as, trans, n);
1990         bch2_btree_update_free(as, trans);
1991         goto out;
1992 }
1993
1994 struct async_btree_rewrite {
1995         struct bch_fs           *c;
1996         struct work_struct      work;
1997         enum btree_id           btree_id;
1998         unsigned                level;
1999         struct bpos             pos;
2000         __le64                  seq;
2001 };
2002
2003 static int async_btree_node_rewrite_trans(struct btree_trans *trans,
2004                                           struct async_btree_rewrite *a)
2005 {
2006         struct btree_iter iter;
2007         struct btree *b;
2008         int ret;
2009
2010         bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos,
2011                                   BTREE_MAX_DEPTH, a->level, 0);
2012         b = bch2_btree_iter_peek_node(&iter);
2013         ret = PTR_ERR_OR_ZERO(b);
2014         if (ret)
2015                 goto out;
2016
2017         if (!b || b->data->keys.seq != a->seq)
2018                 goto out;
2019
2020         ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
2021 out:
2022         bch2_trans_iter_exit(trans, &iter);
2023
2024         return ret;
2025 }
2026
2027 void async_btree_node_rewrite_work(struct work_struct *work)
2028 {
2029         struct async_btree_rewrite *a =
2030                 container_of(work, struct async_btree_rewrite, work);
2031         struct bch_fs *c = a->c;
2032
2033         bch2_trans_do(c, NULL, NULL, 0,
2034                       async_btree_node_rewrite_trans(&trans, a));
2035         bch2_write_ref_put(c, BCH_WRITE_REF_node_rewrite);
2036         kfree(a);
2037 }
2038
2039 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
2040 {
2041         struct async_btree_rewrite *a;
2042
2043         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_node_rewrite))
2044                 return;
2045
2046         a = kmalloc(sizeof(*a), GFP_NOFS);
2047         if (!a) {
2048                 bch2_write_ref_put(c, BCH_WRITE_REF_node_rewrite);
2049                 return;
2050         }
2051
2052         a->c            = c;
2053         a->btree_id     = b->c.btree_id;
2054         a->level        = b->c.level;
2055         a->pos          = b->key.k.p;
2056         a->seq          = b->data->keys.seq;
2057
2058         INIT_WORK(&a->work, async_btree_node_rewrite_work);
2059         queue_work(c->btree_interior_update_worker, &a->work);
2060 }
2061
2062 static int __bch2_btree_node_update_key(struct btree_trans *trans,
2063                                         struct btree_iter *iter,
2064                                         struct btree *b, struct btree *new_hash,
2065                                         struct bkey_i *new_key,
2066                                         bool skip_triggers)
2067 {
2068         struct bch_fs *c = trans->c;
2069         struct btree_iter iter2 = { NULL };
2070         struct btree *parent;
2071         int ret;
2072
2073         if (!skip_triggers) {
2074                 ret = bch2_trans_mark_old(trans, b->c.btree_id, b->c.level + 1,
2075                                           bkey_i_to_s_c(&b->key), 0);
2076                 if (ret)
2077                         return ret;
2078
2079                 ret = bch2_trans_mark_new(trans, b->c.btree_id, b->c.level + 1,
2080                                           new_key, 0);
2081                 if (ret)
2082                         return ret;
2083         }
2084
2085         if (new_hash) {
2086                 bkey_copy(&new_hash->key, new_key);
2087                 ret = bch2_btree_node_hash_insert(&c->btree_cache,
2088                                 new_hash, b->c.level, b->c.btree_id);
2089                 BUG_ON(ret);
2090         }
2091
2092         parent = btree_node_parent(iter->path, b);
2093         if (parent) {
2094                 bch2_trans_copy_iter(&iter2, iter);
2095
2096                 iter2.path = bch2_btree_path_make_mut(trans, iter2.path,
2097                                 iter2.flags & BTREE_ITER_INTENT,
2098                                 _THIS_IP_);
2099
2100                 BUG_ON(iter2.path->level != b->c.level);
2101                 BUG_ON(!bpos_eq(iter2.path->pos, new_key->k.p));
2102
2103                 btree_path_set_level_up(trans, iter2.path);
2104
2105                 trans->paths_sorted = false;
2106
2107                 ret   = bch2_btree_iter_traverse(&iter2) ?:
2108                         bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_NORUN);
2109                 if (ret)
2110                         goto err;
2111         } else {
2112                 BUG_ON(btree_node_root(c, b) != b);
2113
2114                 ret = darray_make_room(&trans->extra_journal_entries,
2115                                        jset_u64s(new_key->k.u64s));
2116                 if (ret)
2117                         return ret;
2118
2119                 journal_entry_set((void *) &darray_top(trans->extra_journal_entries),
2120                                   BCH_JSET_ENTRY_btree_root,
2121                                   b->c.btree_id, b->c.level,
2122                                   new_key, new_key->k.u64s);
2123                 trans->extra_journal_entries.nr += jset_u64s(new_key->k.u64s);
2124         }
2125
2126         ret = bch2_trans_commit(trans, NULL, NULL,
2127                                 BTREE_INSERT_NOFAIL|
2128                                 BTREE_INSERT_NOCHECK_RW|
2129                                 BTREE_INSERT_USE_RESERVE|
2130                                 BTREE_INSERT_JOURNAL_RECLAIM|
2131                                 JOURNAL_WATERMARK_reserved);
2132         if (ret)
2133                 goto err;
2134
2135         bch2_btree_node_lock_write_nofail(trans, iter->path, &b->c);
2136
2137         if (new_hash) {
2138                 mutex_lock(&c->btree_cache.lock);
2139                 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
2140                 bch2_btree_node_hash_remove(&c->btree_cache, b);
2141
2142                 bkey_copy(&b->key, new_key);
2143                 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
2144                 BUG_ON(ret);
2145                 mutex_unlock(&c->btree_cache.lock);
2146         } else {
2147                 bkey_copy(&b->key, new_key);
2148         }
2149
2150         bch2_btree_node_unlock_write(trans, iter->path, b);
2151 out:
2152         bch2_trans_iter_exit(trans, &iter2);
2153         return ret;
2154 err:
2155         if (new_hash) {
2156                 mutex_lock(&c->btree_cache.lock);
2157                 bch2_btree_node_hash_remove(&c->btree_cache, b);
2158                 mutex_unlock(&c->btree_cache.lock);
2159         }
2160         goto out;
2161 }
2162
2163 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
2164                                struct btree *b, struct bkey_i *new_key,
2165                                bool skip_triggers)
2166 {
2167         struct bch_fs *c = trans->c;
2168         struct btree *new_hash = NULL;
2169         struct btree_path *path = iter->path;
2170         struct closure cl;
2171         int ret = 0;
2172
2173         ret = bch2_btree_path_upgrade(trans, path, b->c.level + 1);
2174         if (ret)
2175                 return ret;
2176
2177         closure_init_stack(&cl);
2178
2179         /*
2180          * check btree_ptr_hash_val() after @b is locked by
2181          * btree_iter_traverse():
2182          */
2183         if (btree_ptr_hash_val(new_key) != b->hash_val) {
2184                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2185                 if (ret) {
2186                         bch2_trans_unlock(trans);
2187                         closure_sync(&cl);
2188                         ret = bch2_trans_relock(trans);
2189                         if (ret)
2190                                 return ret;
2191                 }
2192
2193                 new_hash = bch2_btree_node_mem_alloc(c, false);
2194         }
2195
2196         path->intent_ref++;
2197         ret = __bch2_btree_node_update_key(trans, iter, b, new_hash,
2198                                            new_key, skip_triggers);
2199         --path->intent_ref;
2200
2201         if (new_hash) {
2202                 mutex_lock(&c->btree_cache.lock);
2203                 list_move(&new_hash->list, &c->btree_cache.freeable);
2204                 mutex_unlock(&c->btree_cache.lock);
2205
2206                 six_unlock_write(&new_hash->c.lock);
2207                 six_unlock_intent(&new_hash->c.lock);
2208         }
2209         closure_sync(&cl);
2210         bch2_btree_cache_cannibalize_unlock(c);
2211         return ret;
2212 }
2213
2214 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
2215                                         struct btree *b, struct bkey_i *new_key,
2216                                         bool skip_triggers)
2217 {
2218         struct btree_iter iter;
2219         int ret;
2220
2221         bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p,
2222                                   BTREE_MAX_DEPTH, b->c.level,
2223                                   BTREE_ITER_INTENT);
2224         ret = bch2_btree_iter_traverse(&iter);
2225         if (ret)
2226                 goto out;
2227
2228         /* has node been freed? */
2229         if (iter.path->l[b->c.level].b != b) {
2230                 /* node has been freed: */
2231                 BUG_ON(!btree_node_dying(b));
2232                 goto out;
2233         }
2234
2235         BUG_ON(!btree_node_hashed(b));
2236
2237         ret = bch2_btree_node_update_key(trans, &iter, b, new_key, skip_triggers);
2238 out:
2239         bch2_trans_iter_exit(trans, &iter);
2240         return ret;
2241 }
2242
2243 /* Init code: */
2244
2245 /*
2246  * Only for filesystem bringup, when first reading the btree roots or allocating
2247  * btree roots when initializing a new filesystem:
2248  */
2249 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2250 {
2251         BUG_ON(btree_node_root(c, b));
2252
2253         bch2_btree_set_root_inmem(c, b);
2254 }
2255
2256 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2257 {
2258         struct closure cl;
2259         struct btree *b;
2260         int ret;
2261
2262         closure_init_stack(&cl);
2263
2264         do {
2265                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2266                 closure_sync(&cl);
2267         } while (ret);
2268
2269         b = bch2_btree_node_mem_alloc(c, false);
2270         bch2_btree_cache_cannibalize_unlock(c);
2271
2272         set_btree_node_fake(b);
2273         set_btree_node_need_rewrite(b);
2274         b->c.level      = 0;
2275         b->c.btree_id   = id;
2276
2277         bkey_btree_ptr_init(&b->key);
2278         b->key.k.p = SPOS_MAX;
2279         *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2280
2281         bch2_bset_init_first(b, &b->data->keys);
2282         bch2_btree_build_aux_trees(b);
2283
2284         b->data->flags = 0;
2285         btree_set_min(b, POS_MIN);
2286         btree_set_max(b, SPOS_MAX);
2287         b->data->format = bch2_btree_calc_format(b);
2288         btree_node_set_format(b, b->data->format);
2289
2290         ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2291                                           b->c.level, b->c.btree_id);
2292         BUG_ON(ret);
2293
2294         bch2_btree_set_root_inmem(c, b);
2295
2296         six_unlock_write(&b->c.lock);
2297         six_unlock_intent(&b->c.lock);
2298 }
2299
2300 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2301 {
2302         struct btree_update *as;
2303
2304         mutex_lock(&c->btree_interior_update_lock);
2305         list_for_each_entry(as, &c->btree_interior_update_list, list)
2306                 prt_printf(out, "%p m %u w %u r %u j %llu\n",
2307                        as,
2308                        as->mode,
2309                        as->nodes_written,
2310                        atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2311                        as->journal.seq);
2312         mutex_unlock(&c->btree_interior_update_lock);
2313 }
2314
2315 static bool bch2_btree_interior_updates_pending(struct bch_fs *c)
2316 {
2317         bool ret;
2318
2319         mutex_lock(&c->btree_interior_update_lock);
2320         ret = !list_empty(&c->btree_interior_update_list);
2321         mutex_unlock(&c->btree_interior_update_lock);
2322
2323         return ret;
2324 }
2325
2326 bool bch2_btree_interior_updates_flush(struct bch_fs *c)
2327 {
2328         bool ret = bch2_btree_interior_updates_pending(c);
2329
2330         if (ret)
2331                 closure_wait_event(&c->btree_interior_update_wait,
2332                                    !bch2_btree_interior_updates_pending(c));
2333         return ret;
2334 }
2335
2336 void bch2_journal_entries_to_btree_roots(struct bch_fs *c, struct jset *jset)
2337 {
2338         struct btree_root *r;
2339         struct jset_entry *entry;
2340
2341         mutex_lock(&c->btree_root_lock);
2342
2343         vstruct_for_each(jset, entry)
2344                 if (entry->type == BCH_JSET_ENTRY_btree_root) {
2345                         r = &c->btree_roots[entry->btree_id];
2346                         r->level = entry->level;
2347                         r->alive = true;
2348                         bkey_copy(&r->key, &entry->start[0]);
2349                 }
2350
2351         mutex_unlock(&c->btree_root_lock);
2352 }
2353
2354 struct jset_entry *
2355 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2356                                     struct jset_entry *start,
2357                                     struct jset_entry *end)
2358 {
2359         struct jset_entry *entry;
2360         unsigned long have = 0;
2361         unsigned i;
2362
2363         for (entry = start; entry < end; entry = vstruct_next(entry))
2364                 if (entry->type == BCH_JSET_ENTRY_btree_root)
2365                         __set_bit(entry->btree_id, &have);
2366
2367         mutex_lock(&c->btree_root_lock);
2368
2369         for (i = 0; i < BTREE_ID_NR; i++)
2370                 if (c->btree_roots[i].alive && !test_bit(i, &have)) {
2371                         journal_entry_set(end,
2372                                           BCH_JSET_ENTRY_btree_root,
2373                                           i, c->btree_roots[i].level,
2374                                           &c->btree_roots[i].key,
2375                                           c->btree_roots[i].key.u64s);
2376                         end = vstruct_next(end);
2377                 }
2378
2379         mutex_unlock(&c->btree_root_lock);
2380
2381         return end;
2382 }
2383
2384 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2385 {
2386         if (c->btree_interior_update_worker)
2387                 destroy_workqueue(c->btree_interior_update_worker);
2388         mempool_exit(&c->btree_interior_update_pool);
2389 }
2390
2391 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2392 {
2393         mutex_init(&c->btree_reserve_cache_lock);
2394         INIT_LIST_HEAD(&c->btree_interior_update_list);
2395         INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2396         mutex_init(&c->btree_interior_update_lock);
2397         INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2398
2399         c->btree_interior_update_worker =
2400                 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2401         if (!c->btree_interior_update_worker)
2402                 return -ENOMEM;
2403
2404         return mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2405                                          sizeof(struct btree_update));
2406 }