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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "btree_locking.h"
   5 #include "btree_types.h"
   6
   7 static struct lock_class_key bch2_btree_node_lock_key;
   8
   9 void bch2_btree_lock_init(struct btree_bkey_cached_common *b,
  10                           enum six_lock_init_flags flags)
  11 {
  12         __six_lock_init(&b->lock, "b->c.lock", &bch2_btree_node_lock_key, flags);
  13 #ifdef CONFIG_DEBUG_LOCK_ALLOC
  14         lockdep_set_no_check_recursion(&b->lock.dep_map);
  15 #endif
  16 }
  17
  18 #ifdef CONFIG_LOCKDEP
  19 void bch2_assert_btree_nodes_not_locked(void)
  20 {
  21         BUG_ON(lock_class_is_held(&bch2_btree_node_lock_key));
  22 }
  23 #endif
  24
  25 /* Btree node locking: */
  26
  27 struct six_lock_count bch2_btree_node_lock_counts(struct btree_trans *trans,
  28                                                   struct btree_path *skip,
  29                                                   struct btree_bkey_cached_common *b,
  30                                                   unsigned level)
  31 {
  32         struct btree_path *path;
  33         struct six_lock_count ret;
  34
  35         memset(&ret, 0, sizeof(ret));
  36
  37         if (IS_ERR_OR_NULL(b))
  38                 return ret;
  39
  40         trans_for_each_path(trans, path)
  41                 if (path != skip && &path->l[level].b->c == b) {
  42                         int t = btree_node_locked_type(path, level);
  43
  44                         if (t != BTREE_NODE_UNLOCKED)
  45                                 ret.n[t]++;
  46                 }
  47
  48         return ret;
  49 }
  50
  51 /* unlock */
  52
  53 void bch2_btree_node_unlock_write(struct btree_trans *trans,
  54                         struct btree_path *path, struct btree *b)
  55 {
  56         bch2_btree_node_unlock_write_inlined(trans, path, b);
  57 }
  58
  59 /* lock */
  60
  61 /*
  62  * @trans wants to lock @b with type @type
  63  */
  64 struct trans_waiting_for_lock {
  65         struct btree_trans              *trans;
  66         struct btree_bkey_cached_common *node_want;
  67         enum six_lock_type              lock_want;
  68
  69         /* for iterating over held locks :*/
  70         u8                              path_idx;
  71         u8                              level;
  72         u64                             lock_start_time;
  73 };
  74
  75 struct lock_graph {
  76         struct trans_waiting_for_lock   g[8];
  77         unsigned                        nr;
  78 };
  79
  80 static noinline void print_cycle(struct printbuf *out, struct lock_graph *g)
  81 {
  82         struct trans_waiting_for_lock *i;
  83
  84         prt_printf(out, "Found lock cycle (%u entries):", g->nr);
  85         prt_newline(out);
  86
  87         for (i = g->g; i < g->g + g->nr; i++)
  88                 bch2_btree_trans_to_text(out, i->trans);
  89 }
  90
  91 static noinline void print_chain(struct printbuf *out, struct lock_graph *g)
  92 {
  93         struct trans_waiting_for_lock *i;
  94
  95         for (i = g->g; i != g->g + g->nr; i++) {
  96                 if (i != g->g)
  97                         prt_str(out, "<- ");
  98                 prt_printf(out, "%u ", i->trans->locking_wait.task->pid);
  99         }
 100         prt_newline(out);
 101 }
 102
 103 static void lock_graph_up(struct lock_graph *g)
 104 {
 105         closure_put(&g->g[--g->nr].trans->ref);
 106 }
 107
 108 static noinline void lock_graph_pop_all(struct lock_graph *g)
 109 {
 110         while (g->nr)
 111                 lock_graph_up(g);
 112 }
 113
 114 static void lock_graph_down(struct lock_graph *g, struct btree_trans *trans)
 115 {
 116         closure_get(&trans->ref);
 117
 118         g->g[g->nr++] = (struct trans_waiting_for_lock) {
 119                 .trans          = trans,
 120                 .node_want      = trans->locking,
 121                 .lock_want      = trans->locking_wait.lock_want,
 122         };
 123 }
 124
 125 static bool lock_graph_remove_non_waiters(struct lock_graph *g)
 126 {
 127         struct trans_waiting_for_lock *i;
 128
 129         for (i = g->g + 1; i < g->g + g->nr; i++)
 130                 if (i->trans->locking != i->node_want ||
 131                     i->trans->locking_wait.start_time != i[-1].lock_start_time) {
 132                         while (g->g + g->nr > i)
 133                                 lock_graph_up(g);
 134                         return true;
 135                 }
 136
 137         return false;
 138 }
 139
 140 static int abort_lock(struct lock_graph *g, struct trans_waiting_for_lock *i)
 141 {
 142         if (i == g->g) {
 143                 trace_and_count(i->trans->c, trans_restart_would_deadlock, i->trans, _RET_IP_);
 144                 return btree_trans_restart(i->trans, BCH_ERR_transaction_restart_would_deadlock);
 145         } else {
 146                 i->trans->lock_must_abort = true;
 147                 wake_up_process(i->trans->locking_wait.task);
 148                 return 0;
 149         }
 150 }
 151
 152 static int btree_trans_abort_preference(struct btree_trans *trans)
 153 {
 154         if (trans->lock_may_not_fail)
 155                 return 0;
 156         if (trans->locking_wait.lock_want == SIX_LOCK_write)
 157                 return 1;
 158         if (!trans->in_traverse_all)
 159                 return 2;
 160         return 3;
 161 }
 162
 163 static noinline int break_cycle(struct lock_graph *g, struct printbuf *cycle)
 164 {
 165         struct trans_waiting_for_lock *i, *abort = NULL;
 166         unsigned best = 0, pref;
 167         int ret;
 168
 169         if (lock_graph_remove_non_waiters(g))
 170                 return 0;
 171
 172         /* Only checking, for debugfs: */
 173         if (cycle) {
 174                 print_cycle(cycle, g);
 175                 ret = -1;
 176                 goto out;
 177         }
 178
 179         for (i = g->g; i < g->g + g->nr; i++) {
 180                 pref = btree_trans_abort_preference(i->trans);
 181                 if (pref > best) {
 182                         abort = i;
 183                         best = pref;
 184                 }
 185         }
 186
 187         if (unlikely(!best)) {
 188                 struct printbuf buf = PRINTBUF;
 189
 190                 prt_printf(&buf, bch2_fmt(g->g->trans->c, "cycle of nofail locks"));
 191
 192                 for (i = g->g; i < g->g + g->nr; i++) {
 193                         struct btree_trans *trans = i->trans;
 194
 195                         bch2_btree_trans_to_text(&buf, trans);
 196
 197                         prt_printf(&buf, "backtrace:");
 198                         prt_newline(&buf);
 199                         printbuf_indent_add(&buf, 2);
 200                         bch2_prt_task_backtrace(&buf, trans->locking_wait.task);
 201                         printbuf_indent_sub(&buf, 2);
 202                         prt_newline(&buf);
 203                 }
 204
 205                 bch2_print_string_as_lines(KERN_ERR, buf.buf);
 206                 printbuf_exit(&buf);
 207                 BUG();
 208         }
 209
 210         ret = abort_lock(g, abort);
 211 out:
 212         if (ret)
 213                 while (g->nr)
 214                         lock_graph_up(g);
 215         return ret;
 216 }
 217
 218 static int lock_graph_descend(struct lock_graph *g, struct btree_trans *trans,
 219                               struct printbuf *cycle)
 220 {
 221         struct btree_trans *orig_trans = g->g->trans;
 222         struct trans_waiting_for_lock *i;
 223
 224         for (i = g->g; i < g->g + g->nr; i++)
 225                 if (i->trans == trans)
 226                         return break_cycle(g, cycle);
 227
 228         if (g->nr == ARRAY_SIZE(g->g)) {
 229                 if (orig_trans->lock_may_not_fail)
 230                         return 0;
 231
 232                 while (g->nr)
 233                         lock_graph_up(g);
 234
 235                 if (cycle)
 236                         return 0;
 237
 238                 trace_and_count(trans->c, trans_restart_would_deadlock_recursion_limit, trans, _RET_IP_);
 239                 return btree_trans_restart(orig_trans, BCH_ERR_transaction_restart_deadlock_recursion_limit);
 240         }
 241
 242         lock_graph_down(g, trans);
 243         return 0;
 244 }
 245
 246 static bool lock_type_conflicts(enum six_lock_type t1, enum six_lock_type t2)
 247 {
 248         return t1 + t2 > 1;
 249 }
 250
 251 int bch2_check_for_deadlock(struct btree_trans *trans, struct printbuf *cycle)
 252 {
 253         struct lock_graph g;
 254         struct trans_waiting_for_lock *top;
 255         struct btree_bkey_cached_common *b;
 256         struct btree_path *path;
 257         unsigned path_idx;
 258         int ret;
 259
 260         if (trans->lock_must_abort) {
 261                 if (cycle)
 262                         return -1;
 263
 264                 trace_and_count(trans->c, trans_restart_would_deadlock, trans, _RET_IP_);
 265                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock);
 266         }
 267
 268         g.nr = 0;
 269         lock_graph_down(&g, trans);
 270 next:
 271         if (!g.nr)
 272                 return 0;
 273
 274         top = &g.g[g.nr - 1];
 275
 276         trans_for_each_path_safe_from(top->trans, path, path_idx, top->path_idx) {
 277                 if (!path->nodes_locked)
 278                         continue;
 279
 280                 if (path_idx != top->path_idx) {
 281                         top->path_idx           = path_idx;
 282                         top->level              = 0;
 283                         top->lock_start_time    = 0;
 284                 }
 285
 286                 for (;
 287                      top->level < BTREE_MAX_DEPTH;
 288                      top->level++, top->lock_start_time = 0) {
 289                         int lock_held = btree_node_locked_type(path, top->level);
 290
 291                         if (lock_held == BTREE_NODE_UNLOCKED)
 292                                 continue;
 293
 294                         b = &READ_ONCE(path->l[top->level].b)->c;
 295
 296                         if (IS_ERR_OR_NULL(b)) {
 297                                 /*
 298                                  * If we get here, it means we raced with the
 299                                  * other thread updating its btree_path
 300                                  * structures - which means it can't be blocked
 301                                  * waiting on a lock:
 302                                  */
 303                                 if (!lock_graph_remove_non_waiters(&g)) {
 304                                         /*
 305                                          * If lock_graph_remove_non_waiters()
 306                                          * didn't do anything, it must be
 307                                          * because we're being called by debugfs
 308                                          * checking for lock cycles, which
 309                                          * invokes us on btree_transactions that
 310                                          * aren't actually waiting on anything.
 311                                          * Just bail out:
 312                                          */
 313                                         lock_graph_pop_all(&g);
 314                                 }
 315
 316                                 goto next;
 317                         }
 318
 319                         if (list_empty_careful(&b->lock.wait_list))
 320                                 continue;
 321
 322                         raw_spin_lock(&b->lock.wait_lock);
 323                         list_for_each_entry(trans, &b->lock.wait_list, locking_wait.list) {
 324                                 BUG_ON(b != trans->locking);
 325
 326                                 if (top->lock_start_time &&
 327                                     time_after_eq64(top->lock_start_time, trans->locking_wait.start_time))
 328                                         continue;
 329
 330                                 top->lock_start_time = trans->locking_wait.start_time;
 331
 332                                 /* Don't check for self deadlock: */
 333                                 if (trans == top->trans ||
 334                                     !lock_type_conflicts(lock_held, trans->locking_wait.lock_want))
 335                                         continue;
 336
 337                                 ret = lock_graph_descend(&g, trans, cycle);
 338                                 raw_spin_unlock(&b->lock.wait_lock);
 339
 340                                 if (ret)
 341                                         return ret;
 342                                 goto next;
 343
 344                         }
 345                         raw_spin_unlock(&b->lock.wait_lock);
 346                 }
 347         }
 348
 349         if (g.nr > 1 && cycle)
 350                 print_chain(cycle, &g);
 351         lock_graph_up(&g);
 352         goto next;
 353 }
 354
 355 int bch2_six_check_for_deadlock(struct six_lock *lock, void *p)
 356 {
 357         struct btree_trans *trans = p;
 358
 359         return bch2_check_for_deadlock(trans, NULL);
 360 }
 361
 362 int __bch2_btree_node_lock_write(struct btree_trans *trans, struct btree_path *path,
 363                                  struct btree_bkey_cached_common *b,
 364                                  bool lock_may_not_fail)
 365 {
 366         int readers = bch2_btree_node_lock_counts(trans, NULL, b, b->level).n[SIX_LOCK_read];
 367         int ret;
 368
 369         /*
 370          * Must drop our read locks before calling six_lock_write() -
 371          * six_unlock() won't do wakeups until the reader count
 372          * goes to 0, and it's safe because we have the node intent
 373          * locked:
 374          */
 375         six_lock_readers_add(&b->lock, -readers);
 376         ret = __btree_node_lock_nopath(trans, b, SIX_LOCK_write,
 377                                        lock_may_not_fail, _RET_IP_);
 378         six_lock_readers_add(&b->lock, readers);
 379
 380         if (ret)
 381                 mark_btree_node_locked_noreset(path, b->level, SIX_LOCK_intent);
 382
 383         return ret;
 384 }
 385
 386 void bch2_btree_node_lock_write_nofail(struct btree_trans *trans,
 387                                        struct btree_path *path,
 388                                        struct btree_bkey_cached_common *b)
 389 {
 390         struct btree_path *linked;
 391         unsigned i;
 392         int ret;
 393
 394         /*
 395          * XXX BIG FAT NOTICE
 396          *
 397          * Drop all read locks before taking a write lock:
 398          *
 399          * This is a hack, because bch2_btree_node_lock_write_nofail() is a
 400          * hack - but by dropping read locks first, this should never fail, and
 401          * we only use this in code paths where whatever read locks we've
 402          * already taken are no longer needed:
 403          */
 404
 405         trans_for_each_path(trans, linked) {
 406                 if (!linked->nodes_locked)
 407                         continue;
 408
 409                 for (i = 0; i < BTREE_MAX_DEPTH; i++)
 410                         if (btree_node_read_locked(linked, i)) {
 411                                 btree_node_unlock(trans, linked, i);
 412                                 btree_path_set_dirty(linked, BTREE_ITER_NEED_RELOCK);
 413                         }
 414         }
 415
 416         ret = __btree_node_lock_write(trans, path, b, true);
 417         BUG_ON(ret);
 418 }
 419
 420 /* relock */
 421
 422 static inline bool btree_path_get_locks(struct btree_trans *trans,
 423                                         struct btree_path *path,
 424                                         bool upgrade)
 425 {
 426         unsigned l = path->level;
 427         int fail_idx = -1;
 428
 429         do {
 430                 if (!btree_path_node(path, l))
 431                         break;
 432
 433                 if (!(upgrade
 434                       ? bch2_btree_node_upgrade(trans, path, l)
 435                       : bch2_btree_node_relock(trans, path, l)))
 436                         fail_idx = l;
 437
 438                 l++;
 439         } while (l < path->locks_want);
 440
 441         /*
 442          * When we fail to get a lock, we have to ensure that any child nodes
 443          * can't be relocked so bch2_btree_path_traverse has to walk back up to
 444          * the node that we failed to relock:
 445          */
 446         if (fail_idx >= 0) {
 447                 __bch2_btree_path_unlock(trans, path);
 448                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 449
 450                 do {
 451                         path->l[fail_idx].b = upgrade
 452                                 ? ERR_PTR(-BCH_ERR_no_btree_node_upgrade)
 453                                 : ERR_PTR(-BCH_ERR_no_btree_node_relock);
 454                         --fail_idx;
 455                 } while (fail_idx >= 0);
 456         }
 457
 458         if (path->uptodate == BTREE_ITER_NEED_RELOCK)
 459                 path->uptodate = BTREE_ITER_UPTODATE;
 460
 461         bch2_trans_verify_locks(trans);
 462
 463         return path->uptodate < BTREE_ITER_NEED_RELOCK;
 464 }
 465
 466 bool __bch2_btree_node_relock(struct btree_trans *trans,
 467                               struct btree_path *path, unsigned level,
 468                               bool trace)
 469 {
 470         struct btree *b = btree_path_node(path, level);
 471         int want = __btree_lock_want(path, level);
 472
 473         if (race_fault())
 474                 goto fail;
 475
 476         if (six_relock_type(&b->c.lock, want, path->l[level].lock_seq) ||
 477             (btree_node_lock_seq_matches(path, b, level) &&
 478              btree_node_lock_increment(trans, &b->c, level, want))) {
 479                 mark_btree_node_locked(trans, path, level, want);
 480                 return true;
 481         }
 482 fail:
 483         if (trace && !trans->notrace_relock_fail)
 484                 trace_and_count(trans->c, btree_path_relock_fail, trans, _RET_IP_, path, level);
 485         return false;
 486 }
 487
 488 /* upgrade */
 489
 490 bool bch2_btree_node_upgrade(struct btree_trans *trans,
 491                              struct btree_path *path, unsigned level)
 492 {
 493         struct btree *b = path->l[level].b;
 494         struct six_lock_count count = bch2_btree_node_lock_counts(trans, path, &b->c, level);
 495
 496         if (!is_btree_node(path, level))
 497                 return false;
 498
 499         switch (btree_lock_want(path, level)) {
 500         case BTREE_NODE_UNLOCKED:
 501                 BUG_ON(btree_node_locked(path, level));
 502                 return true;
 503         case BTREE_NODE_READ_LOCKED:
 504                 BUG_ON(btree_node_intent_locked(path, level));
 505                 return bch2_btree_node_relock(trans, path, level);
 506         case BTREE_NODE_INTENT_LOCKED:
 507                 break;
 508         case BTREE_NODE_WRITE_LOCKED:
 509                 BUG();
 510         }
 511
 512         if (btree_node_intent_locked(path, level))
 513                 return true;
 514
 515         if (race_fault())
 516                 return false;
 517
 518         if (btree_node_locked(path, level)) {
 519                 bool ret;
 520
 521                 six_lock_readers_add(&b->c.lock, -count.n[SIX_LOCK_read]);
 522                 ret = six_lock_tryupgrade(&b->c.lock);
 523                 six_lock_readers_add(&b->c.lock, count.n[SIX_LOCK_read]);
 524
 525                 if (ret)
 526                         goto success;
 527         } else {
 528                 if (six_relock_type(&b->c.lock, SIX_LOCK_intent, path->l[level].lock_seq))
 529                         goto success;
 530         }
 531
 532         /*
 533          * Do we already have an intent lock via another path? If so, just bump
 534          * lock count:
 535          */
 536         if (btree_node_lock_seq_matches(path, b, level) &&
 537             btree_node_lock_increment(trans, &b->c, level, BTREE_NODE_INTENT_LOCKED)) {
 538                 btree_node_unlock(trans, path, level);
 539                 goto success;
 540         }
 541
 542         trace_and_count(trans->c, btree_path_upgrade_fail, trans, _RET_IP_, path, level);
 543         return false;
 544 success:
 545         mark_btree_node_locked_noreset(path, level, SIX_LOCK_intent);
 546         return true;
 547 }
 548
 549 /* Btree path locking: */
 550
 551 /*
 552  * Only for btree_cache.c - only relocks intent locks
 553  */
 554 int bch2_btree_path_relock_intent(struct btree_trans *trans,
 555                                   struct btree_path *path)
 556 {
 557         unsigned l;
 558
 559         for (l = path->level;
 560              l < path->locks_want && btree_path_node(path, l);
 561              l++) {
 562                 if (!bch2_btree_node_relock(trans, path, l)) {
 563                         __bch2_btree_path_unlock(trans, path);
 564                         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 565                         trace_and_count(trans->c, trans_restart_relock_path_intent, trans, _RET_IP_, path);
 566                         return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_path_intent);
 567                 }
 568         }
 569
 570         return 0;
 571 }
 572
 573 __flatten
 574 bool bch2_btree_path_relock_norestart(struct btree_trans *trans,
 575                         struct btree_path *path, unsigned long trace_ip)
 576 {
 577         return btree_path_get_locks(trans, path, false);
 578 }
 579
 580 int __bch2_btree_path_relock(struct btree_trans *trans,
 581                         struct btree_path *path, unsigned long trace_ip)
 582 {
 583         if (!bch2_btree_path_relock_norestart(trans, path, trace_ip)) {
 584                 trace_and_count(trans->c, trans_restart_relock_path, trans, trace_ip, path);
 585                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_path);
 586         }
 587
 588         return 0;
 589 }
 590
 591 __flatten
 592 bool bch2_btree_path_upgrade_norestart(struct btree_trans *trans,
 593                         struct btree_path *path, unsigned long trace_ip)
 594 {
 595         return btree_path_get_locks(trans, path, true);
 596 }
 597
 598 bool bch2_btree_path_upgrade_noupgrade_sibs(struct btree_trans *trans,
 599                                struct btree_path *path,
 600                                unsigned new_locks_want)
 601 {
 602         EBUG_ON(path->locks_want >= new_locks_want);
 603
 604         path->locks_want = new_locks_want;
 605
 606         return btree_path_get_locks(trans, path, true);
 607 }
 608
 609 bool __bch2_btree_path_upgrade(struct btree_trans *trans,
 610                                struct btree_path *path,
 611                                unsigned new_locks_want)
 612 {
 613         struct btree_path *linked;
 614
 615         if (bch2_btree_path_upgrade_noupgrade_sibs(trans, path, new_locks_want))
 616                 return true;
 617
 618         /*
 619          * XXX: this is ugly - we'd prefer to not be mucking with other
 620          * iterators in the btree_trans here.
 621          *
 622          * On failure to upgrade the iterator, setting iter->locks_want and
 623          * calling get_locks() is sufficient to make bch2_btree_path_traverse()
 624          * get the locks we want on transaction restart.
 625          *
 626          * But if this iterator was a clone, on transaction restart what we did
 627          * to this iterator isn't going to be preserved.
 628          *
 629          * Possibly we could add an iterator field for the parent iterator when
 630          * an iterator is a copy - for now, we'll just upgrade any other
 631          * iterators with the same btree id.
 632          *
 633          * The code below used to be needed to ensure ancestor nodes get locked
 634          * before interior nodes - now that's handled by
 635          * bch2_btree_path_traverse_all().
 636          */
 637         if (!path->cached && !trans->in_traverse_all)
 638                 trans_for_each_path(trans, linked)
 639                         if (linked != path &&
 640                             linked->cached == path->cached &&
 641                             linked->btree_id == path->btree_id &&
 642                             linked->locks_want < new_locks_want) {
 643                                 linked->locks_want = new_locks_want;
 644                                 btree_path_get_locks(trans, linked, true);
 645                         }
 646
 647         return false;
 648 }
 649
 650 void __bch2_btree_path_downgrade(struct btree_trans *trans,
 651                                  struct btree_path *path,
 652                                  unsigned new_locks_want)
 653 {
 654         unsigned l;
 655
 656         EBUG_ON(path->locks_want < new_locks_want);
 657
 658         path->locks_want = new_locks_want;
 659
 660         while (path->nodes_locked &&
 661                (l = btree_path_highest_level_locked(path)) >= path->locks_want) {
 662                 if (l > path->level) {
 663                         btree_node_unlock(trans, path, l);
 664                 } else {
 665                         if (btree_node_intent_locked(path, l)) {
 666                                 six_lock_downgrade(&path->l[l].b->c.lock);
 667                                 mark_btree_node_locked_noreset(path, l, SIX_LOCK_read);
 668                         }
 669                         break;
 670                 }
 671         }
 672
 673         bch2_btree_path_verify_locks(path);
 674 }
 675
 676 /* Btree transaction locking: */
 677
 678 void bch2_trans_downgrade(struct btree_trans *trans)
 679 {
 680         struct btree_path *path;
 681
 682         trans_for_each_path(trans, path)
 683                 bch2_btree_path_downgrade(trans, path);
 684 }
 685
 686 int bch2_trans_relock(struct btree_trans *trans)
 687 {
 688         struct btree_path *path;
 689
 690         if (unlikely(trans->restarted))
 691                 return -((int) trans->restarted);
 692
 693         trans_for_each_path(trans, path)
 694                 if (path->should_be_locked &&
 695                     !bch2_btree_path_relock_norestart(trans, path, _RET_IP_)) {
 696                         trace_and_count(trans->c, trans_restart_relock, trans, _RET_IP_, path);
 697                         return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
 698                 }
 699         return 0;
 700 }
 701
 702 int bch2_trans_relock_notrace(struct btree_trans *trans)
 703 {
 704         struct btree_path *path;
 705
 706         if (unlikely(trans->restarted))
 707                 return -((int) trans->restarted);
 708
 709         trans_for_each_path(trans, path)
 710                 if (path->should_be_locked &&
 711                     !bch2_btree_path_relock_norestart(trans, path, _RET_IP_)) {
 712                         return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
 713                 }
 714         return 0;
 715 }
 716
 717 void bch2_trans_unlock_noassert(struct btree_trans *trans)
 718 {
 719         struct btree_path *path;
 720
 721         trans_for_each_path(trans, path)
 722                 __bch2_btree_path_unlock(trans, path);
 723 }
 724
 725 void bch2_trans_unlock(struct btree_trans *trans)
 726 {
 727         struct btree_path *path;
 728
 729         trans_for_each_path(trans, path)
 730                 __bch2_btree_path_unlock(trans, path);
 731
 732         /*
 733          * bch2_gc_btree_init_recurse() doesn't use btree iterators for walking
 734          * btree nodes, it implements its own walking:
 735          */
 736         if (!trans->is_initial_gc)
 737                 bch2_assert_btree_nodes_not_locked();
 738 }
 739
 740 bool bch2_trans_locked(struct btree_trans *trans)
 741 {
 742         struct btree_path *path;
 743
 744         trans_for_each_path(trans, path)
 745                 if (path->nodes_locked)
 746                         return true;
 747         return false;
 748 }
 749
 750 int __bch2_trans_mutex_lock(struct btree_trans *trans,
 751                             struct mutex *lock)
 752 {
 753         int ret = drop_locks_do(trans, (mutex_lock(lock), 0));
 754
 755         if (ret)
 756                 mutex_unlock(lock);
 757         return ret;
 758 }
 759
 760 /* Debug */
 761
 762 #ifdef CONFIG_BCACHEFS_DEBUG
 763
 764 void bch2_btree_path_verify_locks(struct btree_path *path)
 765 {
 766         unsigned l;
 767
 768         if (!path->nodes_locked) {
 769                 BUG_ON(path->uptodate == BTREE_ITER_UPTODATE &&
 770                        btree_path_node(path, path->level));
 771                 return;
 772         }
 773
 774         for (l = 0; l < BTREE_MAX_DEPTH; l++) {
 775                 int want = btree_lock_want(path, l);
 776                 int have = btree_node_locked_type(path, l);
 777
 778                 BUG_ON(!is_btree_node(path, l) && have != BTREE_NODE_UNLOCKED);
 779
 780                 BUG_ON(is_btree_node(path, l) &&
 781                        (want == BTREE_NODE_UNLOCKED ||
 782                         have != BTREE_NODE_WRITE_LOCKED) &&
 783                        want != have);
 784         }
 785 }
 786
 787 void bch2_trans_verify_locks(struct btree_trans *trans)
 788 {
 789         struct btree_path *path;
 790
 791         trans_for_each_path(trans, path)
 792                 bch2_btree_path_verify_locks(path);
 793 }
 794
 795 #endif