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         g->g[g->nr++] = (struct trans_waiting_for_lock) {
 117                 .trans          = trans,
 118                 .node_want      = trans->locking,
 119                 .lock_want      = trans->locking_wait.lock_want,
 120         };
 121 }
 122
 123 static void lock_graph_down(struct lock_graph *g, struct btree_trans *trans)
 124 {
 125         closure_get(&trans->ref);
 126         __lock_graph_down(g, trans);
 127 }
 128
 129 static bool lock_graph_remove_non_waiters(struct lock_graph *g)
 130 {
 131         struct trans_waiting_for_lock *i;
 132
 133         for (i = g->g + 1; i < g->g + g->nr; i++)
 134                 if (i->trans->locking != i->node_want ||
 135                     i->trans->locking_wait.start_time != i[-1].lock_start_time) {
 136                         while (g->g + g->nr > i)
 137                                 lock_graph_up(g);
 138                         return true;
 139                 }
 140
 141         return false;
 142 }
 143
 144 static int abort_lock(struct lock_graph *g, struct trans_waiting_for_lock *i)
 145 {
 146         if (i == g->g) {
 147                 trace_and_count(i->trans->c, trans_restart_would_deadlock, i->trans, _RET_IP_);
 148                 return btree_trans_restart(i->trans, BCH_ERR_transaction_restart_would_deadlock);
 149         } else {
 150                 i->trans->lock_must_abort = true;
 151                 wake_up_process(i->trans->locking_wait.task);
 152                 return 0;
 153         }
 154 }
 155
 156 static int btree_trans_abort_preference(struct btree_trans *trans)
 157 {
 158         if (trans->lock_may_not_fail)
 159                 return 0;
 160         if (trans->locking_wait.lock_want == SIX_LOCK_write)
 161                 return 1;
 162         if (!trans->in_traverse_all)
 163                 return 2;
 164         return 3;
 165 }
 166
 167 static noinline int break_cycle(struct lock_graph *g, struct printbuf *cycle)
 168 {
 169         struct trans_waiting_for_lock *i, *abort = NULL;
 170         unsigned best = 0, pref;
 171         int ret;
 172
 173         if (lock_graph_remove_non_waiters(g))
 174                 return 0;
 175
 176         /* Only checking, for debugfs: */
 177         if (cycle) {
 178                 print_cycle(cycle, g);
 179                 ret = -1;
 180                 goto out;
 181         }
 182
 183         for (i = g->g; i < g->g + g->nr; i++) {
 184                 pref = btree_trans_abort_preference(i->trans);
 185                 if (pref > best) {
 186                         abort = i;
 187                         best = pref;
 188                 }
 189         }
 190
 191         if (unlikely(!best)) {
 192                 struct printbuf buf = PRINTBUF;
 193
 194                 prt_printf(&buf, bch2_fmt(g->g->trans->c, "cycle of nofail locks"));
 195
 196                 for (i = g->g; i < g->g + g->nr; i++) {
 197                         struct btree_trans *trans = i->trans;
 198
 199                         bch2_btree_trans_to_text(&buf, trans);
 200
 201                         prt_printf(&buf, "backtrace:");
 202                         prt_newline(&buf);
 203                         printbuf_indent_add(&buf, 2);
 204                         bch2_prt_task_backtrace(&buf, trans->locking_wait.task);
 205                         printbuf_indent_sub(&buf, 2);
 206                         prt_newline(&buf);
 207                 }
 208
 209                 bch2_print_string_as_lines(KERN_ERR, buf.buf);
 210                 printbuf_exit(&buf);
 211                 BUG();
 212         }
 213
 214         ret = abort_lock(g, abort);
 215 out:
 216         if (ret)
 217                 while (g->nr)
 218                         lock_graph_up(g);
 219         return ret;
 220 }
 221
 222 static int lock_graph_descend(struct lock_graph *g, struct btree_trans *trans,
 223                               struct printbuf *cycle)
 224 {
 225         struct btree_trans *orig_trans = g->g->trans;
 226         struct trans_waiting_for_lock *i;
 227
 228         for (i = g->g; i < g->g + g->nr; i++)
 229                 if (i->trans == trans) {
 230                         closure_put(&trans->ref);
 231                         return break_cycle(g, cycle);
 232                 }
 233
 234         if (g->nr == ARRAY_SIZE(g->g)) {
 235                 closure_put(&trans->ref);
 236
 237                 if (orig_trans->lock_may_not_fail)
 238                         return 0;
 239
 240                 while (g->nr)
 241                         lock_graph_up(g);
 242
 243                 if (cycle)
 244                         return 0;
 245
 246                 trace_and_count(trans->c, trans_restart_would_deadlock_recursion_limit, trans, _RET_IP_);
 247                 return btree_trans_restart(orig_trans, BCH_ERR_transaction_restart_deadlock_recursion_limit);
 248         }
 249
 250         __lock_graph_down(g, trans);
 251         return 0;
 252 }
 253
 254 static bool lock_type_conflicts(enum six_lock_type t1, enum six_lock_type t2)
 255 {
 256         return t1 + t2 > 1;
 257 }
 258
 259 int bch2_check_for_deadlock(struct btree_trans *trans, struct printbuf *cycle)
 260 {
 261         struct lock_graph g;
 262         struct trans_waiting_for_lock *top;
 263         struct btree_bkey_cached_common *b;
 264         struct btree_path *path;
 265         unsigned path_idx;
 266         int ret;
 267
 268         if (trans->lock_must_abort) {
 269                 if (cycle)
 270                         return -1;
 271
 272                 trace_and_count(trans->c, trans_restart_would_deadlock, trans, _RET_IP_);
 273                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock);
 274         }
 275
 276         g.nr = 0;
 277         lock_graph_down(&g, trans);
 278 next:
 279         if (!g.nr)
 280                 return 0;
 281
 282         top = &g.g[g.nr - 1];
 283
 284         trans_for_each_path_safe_from(top->trans, path, path_idx, top->path_idx) {
 285                 if (!path->nodes_locked)
 286                         continue;
 287
 288                 if (path_idx != top->path_idx) {
 289                         top->path_idx           = path_idx;
 290                         top->level              = 0;
 291                         top->lock_start_time    = 0;
 292                 }
 293
 294                 for (;
 295                      top->level < BTREE_MAX_DEPTH;
 296                      top->level++, top->lock_start_time = 0) {
 297                         int lock_held = btree_node_locked_type(path, top->level);
 298
 299                         if (lock_held == BTREE_NODE_UNLOCKED)
 300                                 continue;
 301
 302                         b = &READ_ONCE(path->l[top->level].b)->c;
 303
 304                         if (IS_ERR_OR_NULL(b)) {
 305                                 /*
 306                                  * If we get here, it means we raced with the
 307                                  * other thread updating its btree_path
 308                                  * structures - which means it can't be blocked
 309                                  * waiting on a lock:
 310                                  */
 311                                 if (!lock_graph_remove_non_waiters(&g)) {
 312                                         /*
 313                                          * If lock_graph_remove_non_waiters()
 314                                          * didn't do anything, it must be
 315                                          * because we're being called by debugfs
 316                                          * checking for lock cycles, which
 317                                          * invokes us on btree_transactions that
 318                                          * aren't actually waiting on anything.
 319                                          * Just bail out:
 320                                          */
 321                                         lock_graph_pop_all(&g);
 322                                 }
 323
 324                                 goto next;
 325                         }
 326
 327                         if (list_empty_careful(&b->lock.wait_list))
 328                                 continue;
 329
 330                         raw_spin_lock(&b->lock.wait_lock);
 331                         list_for_each_entry(trans, &b->lock.wait_list, locking_wait.list) {
 332                                 BUG_ON(b != trans->locking);
 333
 334                                 if (top->lock_start_time &&
 335                                     time_after_eq64(top->lock_start_time, trans->locking_wait.start_time))
 336                                         continue;
 337
 338                                 top->lock_start_time = trans->locking_wait.start_time;
 339
 340                                 /* Don't check for self deadlock: */
 341                                 if (trans == top->trans ||
 342                                     !lock_type_conflicts(lock_held, trans->locking_wait.lock_want))
 343                                         continue;
 344
 345                                 closure_get(&trans->ref);
 346                                 raw_spin_unlock(&b->lock.wait_lock);
 347
 348                                 ret = lock_graph_descend(&g, trans, cycle);
 349                                 if (ret)
 350                                         return ret;
 351                                 goto next;
 352
 353                         }
 354                         raw_spin_unlock(&b->lock.wait_lock);
 355                 }
 356         }
 357
 358         if (g.nr > 1 && cycle)
 359                 print_chain(cycle, &g);
 360         lock_graph_up(&g);
 361         goto next;
 362 }
 363
 364 int bch2_six_check_for_deadlock(struct six_lock *lock, void *p)
 365 {
 366         struct btree_trans *trans = p;
 367
 368         return bch2_check_for_deadlock(trans, NULL);
 369 }
 370
 371 int __bch2_btree_node_lock_write(struct btree_trans *trans, struct btree_path *path,
 372                                  struct btree_bkey_cached_common *b,
 373                                  bool lock_may_not_fail)
 374 {
 375         int readers = bch2_btree_node_lock_counts(trans, NULL, b, b->level).n[SIX_LOCK_read];
 376         int ret;
 377
 378         /*
 379          * Must drop our read locks before calling six_lock_write() -
 380          * six_unlock() won't do wakeups until the reader count
 381          * goes to 0, and it's safe because we have the node intent
 382          * locked:
 383          */
 384         six_lock_readers_add(&b->lock, -readers);
 385         ret = __btree_node_lock_nopath(trans, b, SIX_LOCK_write,
 386                                        lock_may_not_fail, _RET_IP_);
 387         six_lock_readers_add(&b->lock, readers);
 388
 389         if (ret)
 390                 mark_btree_node_locked_noreset(path, b->level, BTREE_NODE_INTENT_LOCKED);
 391
 392         return ret;
 393 }
 394
 395 void bch2_btree_node_lock_write_nofail(struct btree_trans *trans,
 396                                        struct btree_path *path,
 397                                        struct btree_bkey_cached_common *b)
 398 {
 399         struct btree_path *linked;
 400         unsigned i;
 401         int ret;
 402
 403         /*
 404          * XXX BIG FAT NOTICE
 405          *
 406          * Drop all read locks before taking a write lock:
 407          *
 408          * This is a hack, because bch2_btree_node_lock_write_nofail() is a
 409          * hack - but by dropping read locks first, this should never fail, and
 410          * we only use this in code paths where whatever read locks we've
 411          * already taken are no longer needed:
 412          */
 413
 414         trans_for_each_path(trans, linked) {
 415                 if (!linked->nodes_locked)
 416                         continue;
 417
 418                 for (i = 0; i < BTREE_MAX_DEPTH; i++)
 419                         if (btree_node_read_locked(linked, i)) {
 420                                 btree_node_unlock(trans, linked, i);
 421                                 btree_path_set_dirty(linked, BTREE_ITER_NEED_RELOCK);
 422                         }
 423         }
 424
 425         ret = __btree_node_lock_write(trans, path, b, true);
 426         BUG_ON(ret);
 427 }
 428
 429 /* relock */
 430
 431 static inline bool btree_path_get_locks(struct btree_trans *trans,
 432                                         struct btree_path *path,
 433                                         bool upgrade)
 434 {
 435         unsigned l = path->level;
 436         int fail_idx = -1;
 437
 438         do {
 439                 if (!btree_path_node(path, l))
 440                         break;
 441
 442                 if (!(upgrade
 443                       ? bch2_btree_node_upgrade(trans, path, l)
 444                       : bch2_btree_node_relock(trans, path, l)))
 445                         fail_idx = l;
 446
 447                 l++;
 448         } while (l < path->locks_want);
 449
 450         /*
 451          * When we fail to get a lock, we have to ensure that any child nodes
 452          * can't be relocked so bch2_btree_path_traverse has to walk back up to
 453          * the node that we failed to relock:
 454          */
 455         if (fail_idx >= 0) {
 456                 __bch2_btree_path_unlock(trans, path);
 457                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 458
 459                 do {
 460                         path->l[fail_idx].b = upgrade
 461                                 ? ERR_PTR(-BCH_ERR_no_btree_node_upgrade)
 462                                 : ERR_PTR(-BCH_ERR_no_btree_node_relock);
 463                         --fail_idx;
 464                 } while (fail_idx >= 0);
 465         }
 466
 467         if (path->uptodate == BTREE_ITER_NEED_RELOCK)
 468                 path->uptodate = BTREE_ITER_UPTODATE;
 469
 470         bch2_trans_verify_locks(trans);
 471
 472         return path->uptodate < BTREE_ITER_NEED_RELOCK;
 473 }
 474
 475 bool __bch2_btree_node_relock(struct btree_trans *trans,
 476                               struct btree_path *path, unsigned level,
 477                               bool trace)
 478 {
 479         struct btree *b = btree_path_node(path, level);
 480         int want = __btree_lock_want(path, level);
 481
 482         if (race_fault())
 483                 goto fail;
 484
 485         if (six_relock_type(&b->c.lock, want, path->l[level].lock_seq) ||
 486             (btree_node_lock_seq_matches(path, b, level) &&
 487              btree_node_lock_increment(trans, &b->c, level, want))) {
 488                 mark_btree_node_locked(trans, path, level, want);
 489                 return true;
 490         }
 491 fail:
 492         if (trace && !trans->notrace_relock_fail)
 493                 trace_and_count(trans->c, btree_path_relock_fail, trans, _RET_IP_, path, level);
 494         return false;
 495 }
 496
 497 /* upgrade */
 498
 499 bool bch2_btree_node_upgrade(struct btree_trans *trans,
 500                              struct btree_path *path, unsigned level)
 501 {
 502         struct btree *b = path->l[level].b;
 503         struct six_lock_count count = bch2_btree_node_lock_counts(trans, path, &b->c, level);
 504
 505         if (!is_btree_node(path, level))
 506                 return false;
 507
 508         switch (btree_lock_want(path, level)) {
 509         case BTREE_NODE_UNLOCKED:
 510                 BUG_ON(btree_node_locked(path, level));
 511                 return true;
 512         case BTREE_NODE_READ_LOCKED:
 513                 BUG_ON(btree_node_intent_locked(path, level));
 514                 return bch2_btree_node_relock(trans, path, level);
 515         case BTREE_NODE_INTENT_LOCKED:
 516                 break;
 517         case BTREE_NODE_WRITE_LOCKED:
 518                 BUG();
 519         }
 520
 521         if (btree_node_intent_locked(path, level))
 522                 return true;
 523
 524         if (race_fault())
 525                 return false;
 526
 527         if (btree_node_locked(path, level)) {
 528                 bool ret;
 529
 530                 six_lock_readers_add(&b->c.lock, -count.n[SIX_LOCK_read]);
 531                 ret = six_lock_tryupgrade(&b->c.lock);
 532                 six_lock_readers_add(&b->c.lock, count.n[SIX_LOCK_read]);
 533
 534                 if (ret)
 535                         goto success;
 536         } else {
 537                 if (six_relock_type(&b->c.lock, SIX_LOCK_intent, path->l[level].lock_seq))
 538                         goto success;
 539         }
 540
 541         /*
 542          * Do we already have an intent lock via another path? If so, just bump
 543          * lock count:
 544          */
 545         if (btree_node_lock_seq_matches(path, b, level) &&
 546             btree_node_lock_increment(trans, &b->c, level, BTREE_NODE_INTENT_LOCKED)) {
 547                 btree_node_unlock(trans, path, level);
 548                 goto success;
 549         }
 550
 551         trace_and_count(trans->c, btree_path_upgrade_fail, trans, _RET_IP_, path, level);
 552         return false;
 553 success:
 554         mark_btree_node_locked_noreset(path, level, BTREE_NODE_INTENT_LOCKED);
 555         return true;
 556 }
 557
 558 /* Btree path locking: */
 559
 560 /*
 561  * Only for btree_cache.c - only relocks intent locks
 562  */
 563 int bch2_btree_path_relock_intent(struct btree_trans *trans,
 564                                   struct btree_path *path)
 565 {
 566         unsigned l;
 567
 568         for (l = path->level;
 569              l < path->locks_want && btree_path_node(path, l);
 570              l++) {
 571                 if (!bch2_btree_node_relock(trans, path, l)) {
 572                         __bch2_btree_path_unlock(trans, path);
 573                         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 574                         trace_and_count(trans->c, trans_restart_relock_path_intent, trans, _RET_IP_, path);
 575                         return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_path_intent);
 576                 }
 577         }
 578
 579         return 0;
 580 }
 581
 582 __flatten
 583 bool bch2_btree_path_relock_norestart(struct btree_trans *trans,
 584                         struct btree_path *path, unsigned long trace_ip)
 585 {
 586         return btree_path_get_locks(trans, path, false);
 587 }
 588
 589 int __bch2_btree_path_relock(struct btree_trans *trans,
 590                         struct btree_path *path, unsigned long trace_ip)
 591 {
 592         if (!bch2_btree_path_relock_norestart(trans, path, trace_ip)) {
 593                 trace_and_count(trans->c, trans_restart_relock_path, trans, trace_ip, path);
 594                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_path);
 595         }
 596
 597         return 0;
 598 }
 599
 600 bool bch2_btree_path_upgrade_noupgrade_sibs(struct btree_trans *trans,
 601                                struct btree_path *path,
 602                                unsigned new_locks_want)
 603 {
 604         EBUG_ON(path->locks_want >= new_locks_want);
 605
 606         path->locks_want = new_locks_want;
 607
 608         return btree_path_get_locks(trans, path, true);
 609 }
 610
 611 bool __bch2_btree_path_upgrade(struct btree_trans *trans,
 612                                struct btree_path *path,
 613                                unsigned new_locks_want)
 614 {
 615         struct btree_path *linked;
 616
 617         if (bch2_btree_path_upgrade_noupgrade_sibs(trans, path, new_locks_want))
 618                 return true;
 619
 620         /*
 621          * XXX: this is ugly - we'd prefer to not be mucking with other
 622          * iterators in the btree_trans here.
 623          *
 624          * On failure to upgrade the iterator, setting iter->locks_want and
 625          * calling get_locks() is sufficient to make bch2_btree_path_traverse()
 626          * get the locks we want on transaction restart.
 627          *
 628          * But if this iterator was a clone, on transaction restart what we did
 629          * to this iterator isn't going to be preserved.
 630          *
 631          * Possibly we could add an iterator field for the parent iterator when
 632          * an iterator is a copy - for now, we'll just upgrade any other
 633          * iterators with the same btree id.
 634          *
 635          * The code below used to be needed to ensure ancestor nodes get locked
 636          * before interior nodes - now that's handled by
 637          * bch2_btree_path_traverse_all().
 638          */
 639         if (!path->cached && !trans->in_traverse_all)
 640                 trans_for_each_path(trans, linked)
 641                         if (linked != path &&
 642                             linked->cached == path->cached &&
 643                             linked->btree_id == path->btree_id &&
 644                             linked->locks_want < new_locks_want) {
 645                                 linked->locks_want = new_locks_want;
 646                                 btree_path_get_locks(trans, linked, true);
 647                         }
 648
 649         return false;
 650 }
 651
 652 void __bch2_btree_path_downgrade(struct btree_trans *trans,
 653                                  struct btree_path *path,
 654                                  unsigned new_locks_want)
 655 {
 656         unsigned l;
 657
 658         EBUG_ON(path->locks_want < new_locks_want);
 659
 660         path->locks_want = new_locks_want;
 661
 662         while (path->nodes_locked &&
 663                (l = btree_path_highest_level_locked(path)) >= path->locks_want) {
 664                 if (l > path->level) {
 665                         btree_node_unlock(trans, path, l);
 666                 } else {
 667                         if (btree_node_intent_locked(path, l)) {
 668                                 six_lock_downgrade(&path->l[l].b->c.lock);
 669                                 mark_btree_node_locked_noreset(path, l, BTREE_NODE_READ_LOCKED);
 670                         }
 671                         break;
 672                 }
 673         }
 674
 675         bch2_btree_path_verify_locks(path);
 676 }
 677
 678 /* Btree transaction locking: */
 679
 680 void bch2_trans_downgrade(struct btree_trans *trans)
 681 {
 682         struct btree_path *path;
 683
 684         trans_for_each_path(trans, path)
 685                 bch2_btree_path_downgrade(trans, path);
 686 }
 687
 688 int bch2_trans_relock(struct btree_trans *trans)
 689 {
 690         struct btree_path *path;
 691
 692         if (unlikely(trans->restarted))
 693                 return -((int) trans->restarted);
 694
 695         trans_for_each_path(trans, path)
 696                 if (path->should_be_locked &&
 697                     !bch2_btree_path_relock_norestart(trans, path, _RET_IP_)) {
 698                         trace_and_count(trans->c, trans_restart_relock, trans, _RET_IP_, path);
 699                         return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
 700                 }
 701         return 0;
 702 }
 703
 704 int bch2_trans_relock_notrace(struct btree_trans *trans)
 705 {
 706         struct btree_path *path;
 707
 708         if (unlikely(trans->restarted))
 709                 return -((int) trans->restarted);
 710
 711         trans_for_each_path(trans, path)
 712                 if (path->should_be_locked &&
 713                     !bch2_btree_path_relock_norestart(trans, path, _RET_IP_)) {
 714                         return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
 715                 }
 716         return 0;
 717 }
 718
 719 void bch2_trans_unlock_noassert(struct btree_trans *trans)
 720 {
 721         struct btree_path *path;
 722
 723         trans_for_each_path(trans, path)
 724                 __bch2_btree_path_unlock(trans, path);
 725 }
 726
 727 void bch2_trans_unlock(struct btree_trans *trans)
 728 {
 729         struct btree_path *path;
 730
 731         trans_for_each_path(trans, path)
 732                 __bch2_btree_path_unlock(trans, path);
 733
 734         /*
 735          * bch2_gc_btree_init_recurse() doesn't use btree iterators for walking
 736          * btree nodes, it implements its own walking:
 737          */
 738         if (!trans->is_initial_gc)
 739                 bch2_assert_btree_nodes_not_locked();
 740 }
 741
 742 bool bch2_trans_locked(struct btree_trans *trans)
 743 {
 744         struct btree_path *path;
 745
 746         trans_for_each_path(trans, path)
 747                 if (path->nodes_locked)
 748                         return true;
 749         return false;
 750 }
 751
 752 int __bch2_trans_mutex_lock(struct btree_trans *trans,
 753                             struct mutex *lock)
 754 {
 755         int ret = drop_locks_do(trans, (mutex_lock(lock), 0));
 756
 757         if (ret)
 758                 mutex_unlock(lock);
 759         return ret;
 760 }
 761
 762 /* Debug */
 763
 764 #ifdef CONFIG_BCACHEFS_DEBUG
 765
 766 void bch2_btree_path_verify_locks(struct btree_path *path)
 767 {
 768         unsigned l;
 769
 770         if (!path->nodes_locked) {
 771                 BUG_ON(path->uptodate == BTREE_ITER_UPTODATE &&
 772                        btree_path_node(path, path->level));
 773                 return;
 774         }
 775
 776         for (l = 0; l < BTREE_MAX_DEPTH; l++) {
 777                 int want = btree_lock_want(path, l);
 778                 int have = btree_node_locked_type(path, l);
 779
 780                 BUG_ON(!is_btree_node(path, l) && have != BTREE_NODE_UNLOCKED);
 781
 782                 BUG_ON(is_btree_node(path, l) &&
 783                        (want == BTREE_NODE_UNLOCKED ||
 784                         have != BTREE_NODE_WRITE_LOCKED) &&
 785                        want != have);
 786         }
 787 }
 788
 789 void bch2_trans_verify_locks(struct btree_trans *trans)
 790 {
 791         struct btree_path *path;
 792
 793         trans_for_each_path(trans, path)
 794                 bch2_btree_path_verify_locks(path);
 795 }
 796
 797 #endif