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 {
  11         __six_lock_init(&b->lock, "b->c.lock", &bch2_btree_node_lock_key);
  12 }
  13
  14 #ifdef CONFIG_LOCKDEP
  15 void bch2_assert_btree_nodes_not_locked(void)
  16 {
  17         BUG_ON(lock_class_is_held(&bch2_btree_node_lock_key));
  18 }
  19 #endif
  20
  21 /* Btree node locking: */
  22
  23 static inline void six_lock_readers_add(struct six_lock *lock, int nr)
  24 {
  25         if (lock->readers)
  26                 this_cpu_add(*lock->readers, nr);
  27         else if (nr > 0)
  28                 atomic64_add(__SIX_VAL(read_lock, nr), &lock->state.counter);
  29         else
  30                 atomic64_sub(__SIX_VAL(read_lock, -nr), &lock->state.counter);
  31 }
  32
  33 struct six_lock_count bch2_btree_node_lock_counts(struct btree_trans *trans,
  34                                                   struct btree_path *skip,
  35                                                   struct btree_bkey_cached_common *b,
  36                                                   unsigned level)
  37 {
  38         struct btree_path *path;
  39         struct six_lock_count ret;
  40
  41         memset(&ret, 0, sizeof(ret));
  42
  43         if (IS_ERR_OR_NULL(b))
  44                 return ret;
  45
  46         trans_for_each_path(trans, path)
  47                 if (path != skip && &path->l[level].b->c == b) {
  48                         int t = btree_node_locked_type(path, level);
  49
  50                         if (t != BTREE_NODE_UNLOCKED)
  51                                 ret.n[t]++;
  52                 }
  53
  54         return ret;
  55 }
  56
  57 /* unlock */
  58
  59 void bch2_btree_node_unlock_write(struct btree_trans *trans,
  60                         struct btree_path *path, struct btree *b)
  61 {
  62         bch2_btree_node_unlock_write_inlined(trans, path, b);
  63 }
  64
  65 /* lock */
  66
  67 /*
  68  * @trans wants to lock @b with type @type
  69  */
  70 struct trans_waiting_for_lock {
  71         struct btree_trans              *trans;
  72         struct btree_bkey_cached_common *node_want;
  73         enum six_lock_type              lock_want;
  74
  75         /* for iterating over held locks :*/
  76         u8                              path_idx;
  77         u8                              level;
  78         u64                             lock_start_time;
  79 };
  80
  81 struct lock_graph {
  82         struct trans_waiting_for_lock   g[8];
  83         unsigned                        nr;
  84 };
  85
  86 static noinline void print_cycle(struct printbuf *out, struct lock_graph *g)
  87 {
  88         struct trans_waiting_for_lock *i;
  89
  90         prt_printf(out, "Found lock cycle (%u entries):", g->nr);
  91         prt_newline(out);
  92
  93         for (i = g->g; i < g->g + g->nr; i++)
  94                 bch2_btree_trans_to_text(out, i->trans);
  95 }
  96
  97 static noinline void print_chain(struct printbuf *out, struct lock_graph *g)
  98 {
  99         struct trans_waiting_for_lock *i;
 100
 101         for (i = g->g; i != g->g + g->nr; i++) {
 102                 if (i != g->g)
 103                         prt_str(out, "<- ");
 104                 prt_printf(out, "%u ", i->trans->locking_wait.task->pid);
 105         }
 106         prt_newline(out);
 107 }
 108
 109 static void lock_graph_up(struct lock_graph *g)
 110 {
 111         closure_put(&g->g[--g->nr].trans->ref);
 112 }
 113
 114 static noinline void lock_graph_pop_all(struct lock_graph *g)
 115 {
 116         while (g->nr)
 117                 lock_graph_up(g);
 118 }
 119
 120 static void lock_graph_down(struct lock_graph *g, struct btree_trans *trans)
 121 {
 122         closure_get(&trans->ref);
 123
 124         g->g[g->nr++] = (struct trans_waiting_for_lock) {
 125                 .trans          = trans,
 126                 .node_want      = trans->locking,
 127                 .lock_want      = trans->locking_wait.lock_want,
 128         };
 129 }
 130
 131 static bool lock_graph_remove_non_waiters(struct lock_graph *g)
 132 {
 133         struct trans_waiting_for_lock *i;
 134
 135         for (i = g->g + 1; i < g->g + g->nr; i++)
 136                 if (i->trans->locking != i->node_want ||
 137                     i->trans->locking_wait.start_time != i[-1].lock_start_time) {
 138                         while (g->g + g->nr > i)
 139                                 lock_graph_up(g);
 140                         return true;
 141                 }
 142
 143         return false;
 144 }
 145
 146 static int abort_lock(struct lock_graph *g, struct trans_waiting_for_lock *i)
 147 {
 148         if (i == g->g) {
 149                 trace_and_count(i->trans->c, trans_restart_would_deadlock, i->trans, _RET_IP_);
 150                 return btree_trans_restart(i->trans, BCH_ERR_transaction_restart_would_deadlock);
 151         } else {
 152                 i->trans->lock_must_abort = true;
 153                 wake_up_process(i->trans->locking_wait.task);
 154                 return 0;
 155         }
 156 }
 157
 158 static int btree_trans_abort_preference(struct btree_trans *trans)
 159 {
 160         if (trans->lock_may_not_fail)
 161                 return 0;
 162         if (trans->locking_wait.lock_want == SIX_LOCK_write)
 163                 return 1;
 164         if (!trans->in_traverse_all)
 165                 return 2;
 166         return 3;
 167 }
 168
 169 static noinline int break_cycle(struct lock_graph *g, struct printbuf *cycle)
 170 {
 171         struct trans_waiting_for_lock *i, *abort = NULL;
 172         unsigned best = 0, pref;
 173         int ret;
 174
 175         if (lock_graph_remove_non_waiters(g))
 176                 return 0;
 177
 178         /* Only checking, for debugfs: */
 179         if (cycle) {
 180                 print_cycle(cycle, g);
 181                 ret = -1;
 182                 goto out;
 183         }
 184
 185         for (i = g->g; i < g->g + g->nr; i++) {
 186                 pref = btree_trans_abort_preference(i->trans);
 187                 if (pref > best) {
 188                         abort = i;
 189                         best = pref;
 190                 }
 191         }
 192
 193         if (unlikely(!best)) {
 194                 struct printbuf buf = PRINTBUF;
 195
 196                 prt_printf(&buf, bch2_fmt(g->g->trans->c, "cycle of nofail locks"));
 197
 198                 for (i = g->g; i < g->g + g->nr; i++) {
 199                         struct btree_trans *trans = i->trans;
 200
 201                         bch2_btree_trans_to_text(&buf, trans);
 202
 203                         prt_printf(&buf, "backtrace:");
 204                         prt_newline(&buf);
 205                         printbuf_indent_add(&buf, 2);
 206                         bch2_prt_task_backtrace(&buf, trans->locking_wait.task);
 207                         printbuf_indent_sub(&buf, 2);
 208                         prt_newline(&buf);
 209                 }
 210
 211                 bch2_print_string_as_lines(KERN_ERR, buf.buf);
 212                 printbuf_exit(&buf);
 213                 BUG();
 214         }
 215
 216         ret = abort_lock(g, abort);
 217 out:
 218         if (ret)
 219                 while (g->nr)
 220                         lock_graph_up(g);
 221         return ret;
 222 }
 223
 224 static int lock_graph_descend(struct lock_graph *g, struct btree_trans *trans,
 225                               struct printbuf *cycle)
 226 {
 227         struct btree_trans *orig_trans = g->g->trans;
 228         struct trans_waiting_for_lock *i;
 229
 230         for (i = g->g; i < g->g + g->nr; i++)
 231                 if (i->trans == trans)
 232                         return break_cycle(g, cycle);
 233
 234         if (g->nr == ARRAY_SIZE(g->g)) {
 235                 if (orig_trans->lock_may_not_fail)
 236                         return 0;
 237
 238                 while (g->nr)
 239                         lock_graph_up(g);
 240
 241                 if (cycle)
 242                         return 0;
 243
 244                 trace_and_count(trans->c, trans_restart_would_deadlock_recursion_limit, trans, _RET_IP_);
 245                 return btree_trans_restart(orig_trans, BCH_ERR_transaction_restart_deadlock_recursion_limit);
 246         }
 247
 248         lock_graph_down(g, trans);
 249         return 0;
 250 }
 251
 252 static bool lock_type_conflicts(enum six_lock_type t1, enum six_lock_type t2)
 253 {
 254         return t1 + t2 > 1;
 255 }
 256
 257 int bch2_check_for_deadlock(struct btree_trans *trans, struct printbuf *cycle)
 258 {
 259         struct lock_graph g;
 260         struct trans_waiting_for_lock *top;
 261         struct btree_bkey_cached_common *b;
 262         struct btree_path *path;
 263         int ret;
 264
 265         if (trans->lock_must_abort) {
 266                 if (cycle)
 267                         return -1;
 268
 269                 trace_and_count(trans->c, trans_restart_would_deadlock, trans, _RET_IP_);
 270                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock);
 271         }
 272
 273         g.nr = 0;
 274         lock_graph_down(&g, trans);
 275 next:
 276         if (!g.nr)
 277                 return 0;
 278
 279         top = &g.g[g.nr - 1];
 280
 281         trans_for_each_path_from(top->trans, path, top->path_idx) {
 282                 if (!path->nodes_locked)
 283                         continue;
 284
 285                 if (top->path_idx != path->idx) {
 286                         top->path_idx           = path->idx;
 287                         top->level              = 0;
 288                         top->lock_start_time    = 0;
 289                 }
 290
 291                 for (;
 292                      top->level < BTREE_MAX_DEPTH;
 293                      top->level++, top->lock_start_time = 0) {
 294                         int lock_held = btree_node_locked_type(path, top->level);
 295
 296                         if (lock_held == BTREE_NODE_UNLOCKED)
 297                                 continue;
 298
 299                         b = &READ_ONCE(path->l[top->level].b)->c;
 300
 301                         if (IS_ERR_OR_NULL(b)) {
 302                                 /*
 303                                  * If we get here, it means we raced with the
 304                                  * other thread updating its btree_path
 305                                  * structures - which means it can't be blocked
 306                                  * waiting on a lock:
 307                                  */
 308                                 if (!lock_graph_remove_non_waiters(&g)) {
 309                                         /*
 310                                          * If lock_graph_remove_non_waiters()
 311                                          * didn't do anything, it must be
 312                                          * because we're being called by debugfs
 313                                          * checking for lock cycles, which
 314                                          * invokes us on btree_transactions that
 315                                          * aren't actually waiting on anything.
 316                                          * Just bail out:
 317                                          */
 318                                         lock_graph_pop_all(&g);
 319                                 }
 320
 321                                 goto next;
 322                         }
 323
 324                         if (list_empty_careful(&b->lock.wait_list))
 325                                 continue;
 326
 327                         raw_spin_lock(&b->lock.wait_lock);
 328                         list_for_each_entry(trans, &b->lock.wait_list, locking_wait.list) {
 329                                 BUG_ON(b != trans->locking);
 330
 331                                 if (top->lock_start_time &&
 332                                     time_after_eq64(top->lock_start_time, trans->locking_wait.start_time))
 333                                         continue;
 334
 335                                 top->lock_start_time = trans->locking_wait.start_time;
 336
 337                                 /* Don't check for self deadlock: */
 338                                 if (trans == top->trans ||
 339                                     !lock_type_conflicts(lock_held, trans->locking_wait.lock_want))
 340                                         continue;
 341
 342                                 ret = lock_graph_descend(&g, trans, cycle);
 343                                 raw_spin_unlock(&b->lock.wait_lock);
 344
 345                                 if (ret)
 346                                         return ret;
 347                                 goto next;
 348
 349                         }
 350                         raw_spin_unlock(&b->lock.wait_lock);
 351                 }
 352         }
 353
 354         if (g.nr > 1 && cycle)
 355                 print_chain(cycle, &g);
 356         lock_graph_up(&g);
 357         goto next;
 358 }
 359
 360 int bch2_six_check_for_deadlock(struct six_lock *lock, void *p)
 361 {
 362         struct btree_trans *trans = p;
 363
 364         return bch2_check_for_deadlock(trans, NULL);
 365 }
 366
 367 int __bch2_btree_node_lock_write(struct btree_trans *trans, struct btree_path *path,
 368                                  struct btree_bkey_cached_common *b,
 369                                  bool lock_may_not_fail)
 370 {
 371         int readers = bch2_btree_node_lock_counts(trans, NULL, b, b->level).n[SIX_LOCK_read];
 372         int ret;
 373
 374         /*
 375          * Must drop our read locks before calling six_lock_write() -
 376          * six_unlock() won't do wakeups until the reader count
 377          * goes to 0, and it's safe because we have the node intent
 378          * locked:
 379          */
 380         six_lock_readers_add(&b->lock, -readers);
 381         ret = __btree_node_lock_nopath(trans, b, SIX_LOCK_write,
 382                                        lock_may_not_fail, _RET_IP_);
 383         six_lock_readers_add(&b->lock, readers);
 384
 385         if (ret)
 386                 mark_btree_node_locked_noreset(path, b->level, SIX_LOCK_intent);
 387
 388         return ret;
 389 }
 390
 391 void bch2_btree_node_lock_write_nofail(struct btree_trans *trans,
 392                                        struct btree_path *path,
 393                                        struct btree_bkey_cached_common *b)
 394 {
 395         struct btree_path *linked;
 396         unsigned i;
 397         int ret;
 398
 399         /*
 400          * XXX BIG FAT NOTICE
 401          *
 402          * Drop all read locks before taking a write lock:
 403          *
 404          * This is a hack, because bch2_btree_node_lock_write_nofail() is a
 405          * hack - but by dropping read locks first, this should never fail, and
 406          * we only use this in code paths where whatever read locks we've
 407          * already taken are no longer needed:
 408          */
 409
 410         trans_for_each_path(trans, linked) {
 411                 if (!linked->nodes_locked)
 412                         continue;
 413
 414                 for (i = 0; i < BTREE_MAX_DEPTH; i++)
 415                         if (btree_node_read_locked(linked, i)) {
 416                                 btree_node_unlock(trans, linked, i);
 417                                 btree_path_set_dirty(linked, BTREE_ITER_NEED_RELOCK);
 418                         }
 419         }
 420
 421         ret = __btree_node_lock_write(trans, path, b, true);
 422         BUG_ON(ret);
 423 }
 424
 425 /* relock */
 426
 427 static inline bool btree_path_get_locks(struct btree_trans *trans,
 428                                         struct btree_path *path,
 429                                         bool upgrade)
 430 {
 431         unsigned l = path->level;
 432         int fail_idx = -1;
 433
 434         do {
 435                 if (!btree_path_node(path, l))
 436                         break;
 437
 438                 if (!(upgrade
 439                       ? bch2_btree_node_upgrade(trans, path, l)
 440                       : bch2_btree_node_relock(trans, path, l)))
 441                         fail_idx = l;
 442
 443                 l++;
 444         } while (l < path->locks_want);
 445
 446         /*
 447          * When we fail to get a lock, we have to ensure that any child nodes
 448          * can't be relocked so bch2_btree_path_traverse has to walk back up to
 449          * the node that we failed to relock:
 450          */
 451         if (fail_idx >= 0) {
 452                 __bch2_btree_path_unlock(trans, path);
 453                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 454
 455                 do {
 456                         path->l[fail_idx].b = upgrade
 457                                 ? ERR_PTR(-BCH_ERR_no_btree_node_upgrade)
 458                                 : ERR_PTR(-BCH_ERR_no_btree_node_relock);
 459                         --fail_idx;
 460                 } while (fail_idx >= 0);
 461         }
 462
 463         if (path->uptodate == BTREE_ITER_NEED_RELOCK)
 464                 path->uptodate = BTREE_ITER_UPTODATE;
 465
 466         bch2_trans_verify_locks(trans);
 467
 468         return path->uptodate < BTREE_ITER_NEED_RELOCK;
 469 }
 470
 471 bool __bch2_btree_node_relock(struct btree_trans *trans,
 472                               struct btree_path *path, unsigned level,
 473                               bool trace)
 474 {
 475         struct btree *b = btree_path_node(path, level);
 476         int want = __btree_lock_want(path, level);
 477
 478         if (race_fault())
 479                 goto fail;
 480
 481         if (six_relock_type(&b->c.lock, want, path->l[level].lock_seq) ||
 482             (btree_node_lock_seq_matches(path, b, level) &&
 483              btree_node_lock_increment(trans, &b->c, level, want))) {
 484                 mark_btree_node_locked(trans, path, level, want);
 485                 return true;
 486         }
 487 fail:
 488         if (trace && !trans->notrace_relock_fail)
 489                 trace_and_count(trans->c, btree_path_relock_fail, trans, _RET_IP_, path, level);
 490         return false;
 491 }
 492
 493 /* upgrade */
 494
 495 bool bch2_btree_node_upgrade(struct btree_trans *trans,
 496                              struct btree_path *path, unsigned level)
 497 {
 498         struct btree *b = path->l[level].b;
 499         struct six_lock_count count = bch2_btree_node_lock_counts(trans, path, &b->c, level);
 500
 501         if (!is_btree_node(path, level))
 502                 return false;
 503
 504         switch (btree_lock_want(path, level)) {
 505         case BTREE_NODE_UNLOCKED:
 506                 BUG_ON(btree_node_locked(path, level));
 507                 return true;
 508         case BTREE_NODE_READ_LOCKED:
 509                 BUG_ON(btree_node_intent_locked(path, level));
 510                 return bch2_btree_node_relock(trans, path, level);
 511         case BTREE_NODE_INTENT_LOCKED:
 512                 break;
 513         case BTREE_NODE_WRITE_LOCKED:
 514                 BUG();
 515         }
 516
 517         if (btree_node_intent_locked(path, level))
 518                 return true;
 519
 520         if (race_fault())
 521                 return false;
 522
 523         if (btree_node_locked(path, level)) {
 524                 bool ret;
 525
 526                 six_lock_readers_add(&b->c.lock, -count.n[SIX_LOCK_read]);
 527                 ret = six_lock_tryupgrade(&b->c.lock);
 528                 six_lock_readers_add(&b->c.lock, count.n[SIX_LOCK_read]);
 529
 530                 if (ret)
 531                         goto success;
 532         } else {
 533                 if (six_relock_type(&b->c.lock, SIX_LOCK_intent, path->l[level].lock_seq))
 534                         goto success;
 535         }
 536
 537         /*
 538          * Do we already have an intent lock via another path? If so, just bump
 539          * lock count:
 540          */
 541         if (btree_node_lock_seq_matches(path, b, level) &&
 542             btree_node_lock_increment(trans, &b->c, level, BTREE_NODE_INTENT_LOCKED)) {
 543                 btree_node_unlock(trans, path, level);
 544                 goto success;
 545         }
 546
 547         trace_and_count(trans->c, btree_path_upgrade_fail, trans, _RET_IP_, path, level);
 548         return false;
 549 success:
 550         mark_btree_node_locked_noreset(path, level, SIX_LOCK_intent);
 551         return true;
 552 }
 553
 554 /* Btree path locking: */
 555
 556 /*
 557  * Only for btree_cache.c - only relocks intent locks
 558  */
 559 int bch2_btree_path_relock_intent(struct btree_trans *trans,
 560                                   struct btree_path *path)
 561 {
 562         unsigned l;
 563
 564         for (l = path->level;
 565              l < path->locks_want && btree_path_node(path, l);
 566              l++) {
 567                 if (!bch2_btree_node_relock(trans, path, l)) {
 568                         __bch2_btree_path_unlock(trans, path);
 569                         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 570                         trace_and_count(trans->c, trans_restart_relock_path_intent, trans, _RET_IP_, path);
 571                         return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_path_intent);
 572                 }
 573         }
 574
 575         return 0;
 576 }
 577
 578 __flatten
 579 bool bch2_btree_path_relock_norestart(struct btree_trans *trans,
 580                         struct btree_path *path, unsigned long trace_ip)
 581 {
 582         return btree_path_get_locks(trans, path, false);
 583 }
 584
 585 int __bch2_btree_path_relock(struct btree_trans *trans,
 586                         struct btree_path *path, unsigned long trace_ip)
 587 {
 588         if (!bch2_btree_path_relock_norestart(trans, path, trace_ip)) {
 589                 trace_and_count(trans->c, trans_restart_relock_path, trans, trace_ip, path);
 590                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_path);
 591         }
 592
 593         return 0;
 594 }
 595
 596 __flatten
 597 bool bch2_btree_path_upgrade_norestart(struct btree_trans *trans,
 598                         struct btree_path *path, unsigned long trace_ip)
 599 {
 600         return btree_path_get_locks(trans, path, true);
 601 }
 602
 603 bool bch2_btree_path_upgrade_noupgrade_sibs(struct btree_trans *trans,
 604                                struct btree_path *path,
 605                                unsigned new_locks_want)
 606 {
 607         EBUG_ON(path->locks_want >= new_locks_want);
 608
 609         path->locks_want = new_locks_want;
 610
 611         return btree_path_get_locks(trans, path, true);
 612 }
 613
 614 bool __bch2_btree_path_upgrade(struct btree_trans *trans,
 615                                struct btree_path *path,
 616                                unsigned new_locks_want)
 617 {
 618         struct btree_path *linked;
 619
 620         if (bch2_btree_path_upgrade_noupgrade_sibs(trans, path, new_locks_want))
 621                 return true;
 622
 623         /*
 624          * XXX: this is ugly - we'd prefer to not be mucking with other
 625          * iterators in the btree_trans here.
 626          *
 627          * On failure to upgrade the iterator, setting iter->locks_want and
 628          * calling get_locks() is sufficient to make bch2_btree_path_traverse()
 629          * get the locks we want on transaction restart.
 630          *
 631          * But if this iterator was a clone, on transaction restart what we did
 632          * to this iterator isn't going to be preserved.
 633          *
 634          * Possibly we could add an iterator field for the parent iterator when
 635          * an iterator is a copy - for now, we'll just upgrade any other
 636          * iterators with the same btree id.
 637          *
 638          * The code below used to be needed to ensure ancestor nodes get locked
 639          * before interior nodes - now that's handled by
 640          * bch2_btree_path_traverse_all().
 641          */
 642         if (!path->cached && !trans->in_traverse_all)
 643                 trans_for_each_path(trans, linked)
 644                         if (linked != path &&
 645                             linked->cached == path->cached &&
 646                             linked->btree_id == path->btree_id &&
 647                             linked->locks_want < new_locks_want) {
 648                                 linked->locks_want = new_locks_want;
 649                                 btree_path_get_locks(trans, linked, true);
 650                         }
 651
 652         return false;
 653 }
 654
 655 void __bch2_btree_path_downgrade(struct btree_trans *trans,
 656                                  struct btree_path *path,
 657                                  unsigned new_locks_want)
 658 {
 659         unsigned l;
 660
 661         EBUG_ON(path->locks_want < new_locks_want);
 662
 663         path->locks_want = new_locks_want;
 664
 665         while (path->nodes_locked &&
 666                (l = btree_path_highest_level_locked(path)) >= path->locks_want) {
 667                 if (l > path->level) {
 668                         btree_node_unlock(trans, path, l);
 669                 } else {
 670                         if (btree_node_intent_locked(path, l)) {
 671                                 six_lock_downgrade(&path->l[l].b->c.lock);
 672                                 mark_btree_node_locked_noreset(path, l, SIX_LOCK_read);
 673                         }
 674                         break;
 675                 }
 676         }
 677
 678         bch2_btree_path_verify_locks(path);
 679 }
 680
 681 /* Btree transaction locking: */
 682
 683 void bch2_trans_downgrade(struct btree_trans *trans)
 684 {
 685         struct btree_path *path;
 686
 687         trans_for_each_path(trans, path)
 688                 bch2_btree_path_downgrade(trans, path);
 689 }
 690
 691 int bch2_trans_relock(struct btree_trans *trans)
 692 {
 693         struct btree_path *path;
 694
 695         if (unlikely(trans->restarted))
 696                 return -((int) trans->restarted);
 697
 698         trans_for_each_path(trans, path)
 699                 if (path->should_be_locked &&
 700                     !bch2_btree_path_relock_norestart(trans, path, _RET_IP_)) {
 701                         trace_and_count(trans->c, trans_restart_relock, trans, _RET_IP_, path);
 702                         return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
 703                 }
 704         return 0;
 705 }
 706
 707 int bch2_trans_relock_notrace(struct btree_trans *trans)
 708 {
 709         struct btree_path *path;
 710
 711         if (unlikely(trans->restarted))
 712                 return -((int) trans->restarted);
 713
 714         trans_for_each_path(trans, path)
 715                 if (path->should_be_locked &&
 716                     !bch2_btree_path_relock_norestart(trans, path, _RET_IP_)) {
 717                         return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
 718                 }
 719         return 0;
 720 }
 721
 722 void bch2_trans_unlock(struct btree_trans *trans)
 723 {
 724         struct btree_path *path;
 725
 726         trans_for_each_path(trans, path)
 727                 __bch2_btree_path_unlock(trans, path);
 728
 729         /*
 730          * bch2_gc_btree_init_recurse() doesn't use btree iterators for walking
 731          * btree nodes, it implements its own walking:
 732          */
 733         if (!trans->is_initial_gc)
 734                 bch2_assert_btree_nodes_not_locked();
 735 }
 736
 737 bool bch2_trans_locked(struct btree_trans *trans)
 738 {
 739         struct btree_path *path;
 740
 741         trans_for_each_path(trans, path)
 742                 if (path->nodes_locked)
 743                         return true;
 744         return false;
 745 }
 746
 747 int __bch2_trans_mutex_lock(struct btree_trans *trans,
 748                             struct mutex *lock)
 749 {
 750         int ret;
 751
 752         bch2_trans_unlock(trans);
 753         mutex_lock(lock);
 754         ret = bch2_trans_relock(trans);
 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