git.sesse.net Git - bcachefs-tools-debian/blob - c_src/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         lockdep_set_novalidate_class(&b->lock);
  14 }
  15
  16 #ifdef CONFIG_LOCKDEP
  17 void bch2_assert_btree_nodes_not_locked(void)
  18 {
  19 #if 0
  20         //Re-enable when lock_class_is_held() is merged:
  21         BUG_ON(lock_class_is_held(&bch2_btree_node_lock_key));
  22 #endif
  23 }
  24 #endif
  25
  26 /* Btree node locking: */
  27
  28 struct six_lock_count bch2_btree_node_lock_counts(struct btree_trans *trans,
  29                                                   struct btree_path *skip,
  30                                                   struct btree_bkey_cached_common *b,
  31                                                   unsigned level)
  32 {
  33         struct btree_path *path;
  34         struct six_lock_count ret;
  35         unsigned i;
  36
  37         memset(&ret, 0, sizeof(ret));
  38
  39         if (IS_ERR_OR_NULL(b))
  40                 return ret;
  41
  42         trans_for_each_path(trans, path, i)
  43                 if (path != skip && &path->l[level].b->c == b) {
  44                         int t = btree_node_locked_type(path, level);
  45
  46                         if (t != BTREE_NODE_UNLOCKED)
  47                                 ret.n[t]++;
  48                 }
  49
  50         return ret;
  51 }
  52
  53 /* unlock */
  54
  55 void bch2_btree_node_unlock_write(struct btree_trans *trans,
  56                         struct btree_path *path, struct btree *b)
  57 {
  58         bch2_btree_node_unlock_write_inlined(trans, path, b);
  59 }
  60
  61 /* lock */
  62
  63 /*
  64  * @trans wants to lock @b with type @type
  65  */
  66 struct trans_waiting_for_lock {
  67         struct btree_trans              *trans;
  68         struct btree_bkey_cached_common *node_want;
  69         enum six_lock_type              lock_want;
  70
  71         /* for iterating over held locks :*/
  72         u8                              path_idx;
  73         u8                              level;
  74         u64                             lock_start_time;
  75 };
  76
  77 struct lock_graph {
  78         struct trans_waiting_for_lock   g[8];
  79         unsigned                        nr;
  80 };
  81
  82 static noinline void print_cycle(struct printbuf *out, struct lock_graph *g)
  83 {
  84         struct trans_waiting_for_lock *i;
  85
  86         prt_printf(out, "Found lock cycle (%u entries):", g->nr);
  87         prt_newline(out);
  88
  89         for (i = g->g; i < g->g + g->nr; i++) {
  90                 struct task_struct *task = READ_ONCE(i->trans->locking_wait.task);
  91                 if (!task)
  92                         continue;
  93
  94                 bch2_btree_trans_to_text(out, i->trans);
  95                 bch2_prt_task_backtrace(out, task, i == g->g ? 5 : 1);
  96         }
  97 }
  98
  99 static noinline void print_chain(struct printbuf *out, struct lock_graph *g)
 100 {
 101         struct trans_waiting_for_lock *i;
 102
 103         for (i = g->g; i != g->g + g->nr; i++) {
 104                 struct task_struct *task = i->trans->locking_wait.task;
 105                 if (i != g->g)
 106                         prt_str(out, "<- ");
 107                 prt_printf(out, "%u ", task ?task->pid : 0);
 108         }
 109         prt_newline(out);
 110 }
 111
 112 static void lock_graph_up(struct lock_graph *g)
 113 {
 114         closure_put(&g->g[--g->nr].trans->ref);
 115 }
 116
 117 static noinline void lock_graph_pop_all(struct lock_graph *g)
 118 {
 119         while (g->nr)
 120                 lock_graph_up(g);
 121 }
 122
 123 static void __lock_graph_down(struct lock_graph *g, struct btree_trans *trans)
 124 {
 125         g->g[g->nr++] = (struct trans_waiting_for_lock) {
 126                 .trans          = trans,
 127                 .node_want      = trans->locking,
 128                 .lock_want      = trans->locking_wait.lock_want,
 129         };
 130 }
 131
 132 static void lock_graph_down(struct lock_graph *g, struct btree_trans *trans)
 133 {
 134         closure_get(&trans->ref);
 135         __lock_graph_down(g, trans);
 136 }
 137
 138 static bool lock_graph_remove_non_waiters(struct lock_graph *g)
 139 {
 140         struct trans_waiting_for_lock *i;
 141
 142         for (i = g->g + 1; i < g->g + g->nr; i++)
 143                 if (i->trans->locking != i->node_want ||
 144                     i->trans->locking_wait.start_time != i[-1].lock_start_time) {
 145                         while (g->g + g->nr > i)
 146                                 lock_graph_up(g);
 147                         return true;
 148                 }
 149
 150         return false;
 151 }
 152
 153 static void trace_would_deadlock(struct lock_graph *g, struct btree_trans *trans)
 154 {
 155         struct bch_fs *c = trans->c;
 156
 157         count_event(c, trans_restart_would_deadlock);
 158
 159         if (trace_trans_restart_would_deadlock_enabled()) {
 160                 struct printbuf buf = PRINTBUF;
 161
 162                 buf.atomic++;
 163                 print_cycle(&buf, g);
 164
 165                 trace_trans_restart_would_deadlock(trans, buf.buf);
 166                 printbuf_exit(&buf);
 167         }
 168 }
 169
 170 static int abort_lock(struct lock_graph *g, struct trans_waiting_for_lock *i)
 171 {
 172         if (i == g->g) {
 173                 trace_would_deadlock(g, i->trans);
 174                 return btree_trans_restart(i->trans, BCH_ERR_transaction_restart_would_deadlock);
 175         } else {
 176                 i->trans->lock_must_abort = true;
 177                 wake_up_process(i->trans->locking_wait.task);
 178                 return 0;
 179         }
 180 }
 181
 182 static int btree_trans_abort_preference(struct btree_trans *trans)
 183 {
 184         if (trans->lock_may_not_fail)
 185                 return 0;
 186         if (trans->locking_wait.lock_want == SIX_LOCK_write)
 187                 return 1;
 188         if (!trans->in_traverse_all)
 189                 return 2;
 190         return 3;
 191 }
 192
 193 static noinline int break_cycle(struct lock_graph *g, struct printbuf *cycle)
 194 {
 195         struct trans_waiting_for_lock *i, *abort = NULL;
 196         unsigned best = 0, pref;
 197         int ret;
 198
 199         if (lock_graph_remove_non_waiters(g))
 200                 return 0;
 201
 202         /* Only checking, for debugfs: */
 203         if (cycle) {
 204                 print_cycle(cycle, g);
 205                 ret = -1;
 206                 goto out;
 207         }
 208
 209         for (i = g->g; i < g->g + g->nr; i++) {
 210                 pref = btree_trans_abort_preference(i->trans);
 211                 if (pref > best) {
 212                         abort = i;
 213                         best = pref;
 214                 }
 215         }
 216
 217         if (unlikely(!best)) {
 218                 struct printbuf buf = PRINTBUF;
 219
 220                 prt_printf(&buf, bch2_fmt(g->g->trans->c, "cycle of nofail locks"));
 221
 222                 for (i = g->g; i < g->g + g->nr; i++) {
 223                         struct btree_trans *trans = i->trans;
 224
 225                         bch2_btree_trans_to_text(&buf, trans);
 226
 227                         prt_printf(&buf, "backtrace:");
 228                         prt_newline(&buf);
 229                         printbuf_indent_add(&buf, 2);
 230                         bch2_prt_task_backtrace(&buf, trans->locking_wait.task, 2);
 231                         printbuf_indent_sub(&buf, 2);
 232                         prt_newline(&buf);
 233                 }
 234
 235                 bch2_print_string_as_lines(KERN_ERR, buf.buf);
 236                 printbuf_exit(&buf);
 237                 BUG();
 238         }
 239
 240         ret = abort_lock(g, abort);
 241 out:
 242         if (ret)
 243                 while (g->nr)
 244                         lock_graph_up(g);
 245         return ret;
 246 }
 247
 248 static int lock_graph_descend(struct lock_graph *g, struct btree_trans *trans,
 249                               struct printbuf *cycle)
 250 {
 251         struct btree_trans *orig_trans = g->g->trans;
 252         struct trans_waiting_for_lock *i;
 253
 254         for (i = g->g; i < g->g + g->nr; i++)
 255                 if (i->trans == trans) {
 256                         closure_put(&trans->ref);
 257                         return break_cycle(g, cycle);
 258                 }
 259
 260         if (g->nr == ARRAY_SIZE(g->g)) {
 261                 closure_put(&trans->ref);
 262
 263                 if (orig_trans->lock_may_not_fail)
 264                         return 0;
 265
 266                 while (g->nr)
 267                         lock_graph_up(g);
 268
 269                 if (cycle)
 270                         return 0;
 271
 272                 trace_and_count(trans->c, trans_restart_would_deadlock_recursion_limit, trans, _RET_IP_);
 273                 return btree_trans_restart(orig_trans, BCH_ERR_transaction_restart_deadlock_recursion_limit);
 274         }
 275
 276         __lock_graph_down(g, trans);
 277         return 0;
 278 }
 279
 280 static bool lock_type_conflicts(enum six_lock_type t1, enum six_lock_type t2)
 281 {
 282         return t1 + t2 > 1;
 283 }
 284
 285 int bch2_check_for_deadlock(struct btree_trans *trans, struct printbuf *cycle)
 286 {
 287         struct lock_graph g;
 288         struct trans_waiting_for_lock *top;
 289         struct btree_bkey_cached_common *b;
 290         btree_path_idx_t path_idx;
 291         int ret = 0;
 292
 293         g.nr = 0;
 294
 295         if (trans->lock_must_abort) {
 296                 if (cycle)
 297                         return -1;
 298
 299                 trace_would_deadlock(&g, trans);
 300                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock);
 301         }
 302
 303         lock_graph_down(&g, trans);
 304
 305         /* trans->paths is rcu protected vs. freeing */
 306         rcu_read_lock();
 307         if (cycle)
 308                 cycle->atomic++;
 309 next:
 310         if (!g.nr)
 311                 goto out;
 312
 313         top = &g.g[g.nr - 1];
 314
 315         struct btree_path *paths = rcu_dereference(top->trans->paths);
 316         if (!paths)
 317                 goto up;
 318
 319         unsigned long *paths_allocated = trans_paths_allocated(paths);
 320
 321         trans_for_each_path_idx_from(paths_allocated, *trans_paths_nr(paths),
 322                                      path_idx, top->path_idx) {
 323                 struct btree_path *path = paths + path_idx;
 324                 if (!path->nodes_locked)
 325                         continue;
 326
 327                 if (path_idx != top->path_idx) {
 328                         top->path_idx           = path_idx;
 329                         top->level              = 0;
 330                         top->lock_start_time    = 0;
 331                 }
 332
 333                 for (;
 334                      top->level < BTREE_MAX_DEPTH;
 335                      top->level++, top->lock_start_time = 0) {
 336                         int lock_held = btree_node_locked_type(path, top->level);
 337
 338                         if (lock_held == BTREE_NODE_UNLOCKED)
 339                                 continue;
 340
 341                         b = &READ_ONCE(path->l[top->level].b)->c;
 342
 343                         if (IS_ERR_OR_NULL(b)) {
 344                                 /*
 345                                  * If we get here, it means we raced with the
 346                                  * other thread updating its btree_path
 347                                  * structures - which means it can't be blocked
 348                                  * waiting on a lock:
 349                                  */
 350                                 if (!lock_graph_remove_non_waiters(&g)) {
 351                                         /*
 352                                          * If lock_graph_remove_non_waiters()
 353                                          * didn't do anything, it must be
 354                                          * because we're being called by debugfs
 355                                          * checking for lock cycles, which
 356                                          * invokes us on btree_transactions that
 357                                          * aren't actually waiting on anything.
 358                                          * Just bail out:
 359                                          */
 360                                         lock_graph_pop_all(&g);
 361                                 }
 362
 363                                 goto next;
 364                         }
 365
 366                         if (list_empty_careful(&b->lock.wait_list))
 367                                 continue;
 368
 369                         raw_spin_lock(&b->lock.wait_lock);
 370                         list_for_each_entry(trans, &b->lock.wait_list, locking_wait.list) {
 371                                 BUG_ON(b != trans->locking);
 372
 373                                 if (top->lock_start_time &&
 374                                     time_after_eq64(top->lock_start_time, trans->locking_wait.start_time))
 375                                         continue;
 376
 377                                 top->lock_start_time = trans->locking_wait.start_time;
 378
 379                                 /* Don't check for self deadlock: */
 380                                 if (trans == top->trans ||
 381                                     !lock_type_conflicts(lock_held, trans->locking_wait.lock_want))
 382                                         continue;
 383
 384                                 closure_get(&trans->ref);
 385                                 raw_spin_unlock(&b->lock.wait_lock);
 386
 387                                 ret = lock_graph_descend(&g, trans, cycle);
 388                                 if (ret)
 389                                         goto out;
 390                                 goto next;
 391
 392                         }
 393                         raw_spin_unlock(&b->lock.wait_lock);
 394                 }
 395         }
 396 up:
 397         if (g.nr > 1 && cycle)
 398                 print_chain(cycle, &g);
 399         lock_graph_up(&g);
 400         goto next;
 401 out:
 402         if (cycle)
 403                 --cycle->atomic;
 404         rcu_read_unlock();
 405         return ret;
 406 }
 407
 408 int bch2_six_check_for_deadlock(struct six_lock *lock, void *p)
 409 {
 410         struct btree_trans *trans = p;
 411
 412         return bch2_check_for_deadlock(trans, NULL);
 413 }
 414
 415 int __bch2_btree_node_lock_write(struct btree_trans *trans, struct btree_path *path,
 416                                  struct btree_bkey_cached_common *b,
 417                                  bool lock_may_not_fail)
 418 {
 419         int readers = bch2_btree_node_lock_counts(trans, NULL, b, b->level).n[SIX_LOCK_read];
 420         int ret;
 421
 422         /*
 423          * Must drop our read locks before calling six_lock_write() -
 424          * six_unlock() won't do wakeups until the reader count
 425          * goes to 0, and it's safe because we have the node intent
 426          * locked:
 427          */
 428         six_lock_readers_add(&b->lock, -readers);
 429         ret = __btree_node_lock_nopath(trans, b, SIX_LOCK_write,
 430                                        lock_may_not_fail, _RET_IP_);
 431         six_lock_readers_add(&b->lock, readers);
 432
 433         if (ret)
 434                 mark_btree_node_locked_noreset(path, b->level, BTREE_NODE_INTENT_LOCKED);
 435
 436         return ret;
 437 }
 438
 439 void bch2_btree_node_lock_write_nofail(struct btree_trans *trans,
 440                                        struct btree_path *path,
 441                                        struct btree_bkey_cached_common *b)
 442 {
 443         struct btree_path *linked;
 444         unsigned i, iter;
 445         int ret;
 446
 447         /*
 448          * XXX BIG FAT NOTICE
 449          *
 450          * Drop all read locks before taking a write lock:
 451          *
 452          * This is a hack, because bch2_btree_node_lock_write_nofail() is a
 453          * hack - but by dropping read locks first, this should never fail, and
 454          * we only use this in code paths where whatever read locks we've
 455          * already taken are no longer needed:
 456          */
 457
 458         trans_for_each_path(trans, linked, iter) {
 459                 if (!linked->nodes_locked)
 460                         continue;
 461
 462                 for (i = 0; i < BTREE_MAX_DEPTH; i++)
 463                         if (btree_node_read_locked(linked, i)) {
 464                                 btree_node_unlock(trans, linked, i);
 465                                 btree_path_set_dirty(linked, BTREE_ITER_NEED_RELOCK);
 466                         }
 467         }
 468
 469         ret = __btree_node_lock_write(trans, path, b, true);
 470         BUG_ON(ret);
 471 }
 472
 473 /* relock */
 474
 475 static inline bool btree_path_get_locks(struct btree_trans *trans,
 476                                         struct btree_path *path,
 477                                         bool upgrade,
 478                                         struct get_locks_fail *f)
 479 {
 480         unsigned l = path->level;
 481         int fail_idx = -1;
 482
 483         do {
 484                 if (!btree_path_node(path, l))
 485                         break;
 486
 487                 if (!(upgrade
 488                       ? bch2_btree_node_upgrade(trans, path, l)
 489                       : bch2_btree_node_relock(trans, path, l))) {
 490                         fail_idx        = l;
 491
 492                         if (f) {
 493                                 f->l    = l;
 494                                 f->b    = path->l[l].b;
 495                         }
 496                 }
 497
 498                 l++;
 499         } while (l < path->locks_want);
 500
 501         /*
 502          * When we fail to get a lock, we have to ensure that any child nodes
 503          * can't be relocked so bch2_btree_path_traverse has to walk back up to
 504          * the node that we failed to relock:
 505          */
 506         if (fail_idx >= 0) {
 507                 __bch2_btree_path_unlock(trans, path);
 508                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 509
 510                 do {
 511                         path->l[fail_idx].b = upgrade
 512                                 ? ERR_PTR(-BCH_ERR_no_btree_node_upgrade)
 513                                 : ERR_PTR(-BCH_ERR_no_btree_node_relock);
 514                         --fail_idx;
 515                 } while (fail_idx >= 0);
 516         }
 517
 518         if (path->uptodate == BTREE_ITER_NEED_RELOCK)
 519                 path->uptodate = BTREE_ITER_UPTODATE;
 520
 521         bch2_trans_verify_locks(trans);
 522
 523         return path->uptodate < BTREE_ITER_NEED_RELOCK;
 524 }
 525
 526 bool __bch2_btree_node_relock(struct btree_trans *trans,
 527                               struct btree_path *path, unsigned level,
 528                               bool trace)
 529 {
 530         struct btree *b = btree_path_node(path, level);
 531         int want = __btree_lock_want(path, level);
 532
 533         if (race_fault())
 534                 goto fail;
 535
 536         if (six_relock_type(&b->c.lock, want, path->l[level].lock_seq) ||
 537             (btree_node_lock_seq_matches(path, b, level) &&
 538              btree_node_lock_increment(trans, &b->c, level, want))) {
 539                 mark_btree_node_locked(trans, path, level, want);
 540                 return true;
 541         }
 542 fail:
 543         if (trace && !trans->notrace_relock_fail)
 544                 trace_and_count(trans->c, btree_path_relock_fail, trans, _RET_IP_, path, level);
 545         return false;
 546 }
 547
 548 /* upgrade */
 549
 550 bool bch2_btree_node_upgrade(struct btree_trans *trans,
 551                              struct btree_path *path, unsigned level)
 552 {
 553         struct btree *b = path->l[level].b;
 554         struct six_lock_count count = bch2_btree_node_lock_counts(trans, path, &b->c, level);
 555
 556         if (!is_btree_node(path, level))
 557                 return false;
 558
 559         switch (btree_lock_want(path, level)) {
 560         case BTREE_NODE_UNLOCKED:
 561                 BUG_ON(btree_node_locked(path, level));
 562                 return true;
 563         case BTREE_NODE_READ_LOCKED:
 564                 BUG_ON(btree_node_intent_locked(path, level));
 565                 return bch2_btree_node_relock(trans, path, level);
 566         case BTREE_NODE_INTENT_LOCKED:
 567                 break;
 568         case BTREE_NODE_WRITE_LOCKED:
 569                 BUG();
 570         }
 571
 572         if (btree_node_intent_locked(path, level))
 573                 return true;
 574
 575         if (race_fault())
 576                 return false;
 577
 578         if (btree_node_locked(path, level)) {
 579                 bool ret;
 580
 581                 six_lock_readers_add(&b->c.lock, -count.n[SIX_LOCK_read]);
 582                 ret = six_lock_tryupgrade(&b->c.lock);
 583                 six_lock_readers_add(&b->c.lock, count.n[SIX_LOCK_read]);
 584
 585                 if (ret)
 586                         goto success;
 587         } else {
 588                 if (six_relock_type(&b->c.lock, SIX_LOCK_intent, path->l[level].lock_seq))
 589                         goto success;
 590         }
 591
 592         /*
 593          * Do we already have an intent lock via another path? If so, just bump
 594          * lock count:
 595          */
 596         if (btree_node_lock_seq_matches(path, b, level) &&
 597             btree_node_lock_increment(trans, &b->c, level, BTREE_NODE_INTENT_LOCKED)) {
 598                 btree_node_unlock(trans, path, level);
 599                 goto success;
 600         }
 601
 602         trace_and_count(trans->c, btree_path_upgrade_fail, trans, _RET_IP_, path, level);
 603         return false;
 604 success:
 605         mark_btree_node_locked_noreset(path, level, BTREE_NODE_INTENT_LOCKED);
 606         return true;
 607 }
 608
 609 /* Btree path locking: */
 610
 611 /*
 612  * Only for btree_cache.c - only relocks intent locks
 613  */
 614 int bch2_btree_path_relock_intent(struct btree_trans *trans,
 615                                   struct btree_path *path)
 616 {
 617         unsigned l;
 618
 619         for (l = path->level;
 620              l < path->locks_want && btree_path_node(path, l);
 621              l++) {
 622                 if (!bch2_btree_node_relock(trans, path, l)) {
 623                         __bch2_btree_path_unlock(trans, path);
 624                         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 625                         trace_and_count(trans->c, trans_restart_relock_path_intent, trans, _RET_IP_, path);
 626                         return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_path_intent);
 627                 }
 628         }
 629
 630         return 0;
 631 }
 632
 633 __flatten
 634 bool bch2_btree_path_relock_norestart(struct btree_trans *trans,
 635                         struct btree_path *path, unsigned long trace_ip)
 636 {
 637         struct get_locks_fail f;
 638
 639         return btree_path_get_locks(trans, path, false, &f);
 640 }
 641
 642 int __bch2_btree_path_relock(struct btree_trans *trans,
 643                         struct btree_path *path, unsigned long trace_ip)
 644 {
 645         if (!bch2_btree_path_relock_norestart(trans, path, trace_ip)) {
 646                 trace_and_count(trans->c, trans_restart_relock_path, trans, trace_ip, path);
 647                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_path);
 648         }
 649
 650         return 0;
 651 }
 652
 653 bool bch2_btree_path_upgrade_noupgrade_sibs(struct btree_trans *trans,
 654                                struct btree_path *path,
 655                                unsigned new_locks_want,
 656                                struct get_locks_fail *f)
 657 {
 658         EBUG_ON(path->locks_want >= new_locks_want);
 659
 660         path->locks_want = new_locks_want;
 661
 662         return btree_path_get_locks(trans, path, true, f);
 663 }
 664
 665 bool __bch2_btree_path_upgrade(struct btree_trans *trans,
 666                                struct btree_path *path,
 667                                unsigned new_locks_want,
 668                                struct get_locks_fail *f)
 669 {
 670         if (bch2_btree_path_upgrade_noupgrade_sibs(trans, path, new_locks_want, f))
 671                 return true;
 672
 673         /*
 674          * XXX: this is ugly - we'd prefer to not be mucking with other
 675          * iterators in the btree_trans here.
 676          *
 677          * On failure to upgrade the iterator, setting iter->locks_want and
 678          * calling get_locks() is sufficient to make bch2_btree_path_traverse()
 679          * get the locks we want on transaction restart.
 680          *
 681          * But if this iterator was a clone, on transaction restart what we did
 682          * to this iterator isn't going to be preserved.
 683          *
 684          * Possibly we could add an iterator field for the parent iterator when
 685          * an iterator is a copy - for now, we'll just upgrade any other
 686          * iterators with the same btree id.
 687          *
 688          * The code below used to be needed to ensure ancestor nodes get locked
 689          * before interior nodes - now that's handled by
 690          * bch2_btree_path_traverse_all().
 691          */
 692         if (!path->cached && !trans->in_traverse_all) {
 693                 struct btree_path *linked;
 694                 unsigned i;
 695
 696                 trans_for_each_path(trans, linked, i)
 697                         if (linked != path &&
 698                             linked->cached == path->cached &&
 699                             linked->btree_id == path->btree_id &&
 700                             linked->locks_want < new_locks_want) {
 701                                 linked->locks_want = new_locks_want;
 702                                 btree_path_get_locks(trans, linked, true, NULL);
 703                         }
 704         }
 705
 706         return false;
 707 }
 708
 709 void __bch2_btree_path_downgrade(struct btree_trans *trans,
 710                                  struct btree_path *path,
 711                                  unsigned new_locks_want)
 712 {
 713         unsigned l, old_locks_want = path->locks_want;
 714
 715         if (trans->restarted)
 716                 return;
 717
 718         EBUG_ON(path->locks_want < new_locks_want);
 719
 720         path->locks_want = new_locks_want;
 721
 722         while (path->nodes_locked &&
 723                (l = btree_path_highest_level_locked(path)) >= path->locks_want) {
 724                 if (l > path->level) {
 725                         btree_node_unlock(trans, path, l);
 726                 } else {
 727                         if (btree_node_intent_locked(path, l)) {
 728                                 six_lock_downgrade(&path->l[l].b->c.lock);
 729                                 mark_btree_node_locked_noreset(path, l, BTREE_NODE_READ_LOCKED);
 730                         }
 731                         break;
 732                 }
 733         }
 734
 735         bch2_btree_path_verify_locks(path);
 736
 737         trace_path_downgrade(trans, _RET_IP_, path, old_locks_want);
 738 }
 739
 740 /* Btree transaction locking: */
 741
 742 void bch2_trans_downgrade(struct btree_trans *trans)
 743 {
 744         struct btree_path *path;
 745         unsigned i;
 746
 747         if (trans->restarted)
 748                 return;
 749
 750         trans_for_each_path(trans, path, i)
 751                 bch2_btree_path_downgrade(trans, path);
 752 }
 753
 754 int bch2_trans_relock(struct btree_trans *trans)
 755 {
 756         struct btree_path *path;
 757         unsigned i;
 758
 759         if (unlikely(trans->restarted))
 760                 return -((int) trans->restarted);
 761
 762         trans_for_each_path(trans, path, i)
 763                 if (path->should_be_locked &&
 764                     !bch2_btree_path_relock_norestart(trans, path, _RET_IP_)) {
 765                         trace_and_count(trans->c, trans_restart_relock, trans, _RET_IP_, path);
 766                         return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
 767                 }
 768         return 0;
 769 }
 770
 771 int bch2_trans_relock_notrace(struct btree_trans *trans)
 772 {
 773         struct btree_path *path;
 774         unsigned i;
 775
 776         if (unlikely(trans->restarted))
 777                 return -((int) trans->restarted);
 778
 779         trans_for_each_path(trans, path, i)
 780                 if (path->should_be_locked &&
 781                     !bch2_btree_path_relock_norestart(trans, path, _RET_IP_)) {
 782                         return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
 783                 }
 784         return 0;
 785 }
 786
 787 void bch2_trans_unlock_noassert(struct btree_trans *trans)
 788 {
 789         struct btree_path *path;
 790         unsigned i;
 791
 792         trans_for_each_path(trans, path, i)
 793                 __bch2_btree_path_unlock(trans, path);
 794 }
 795
 796 void bch2_trans_unlock(struct btree_trans *trans)
 797 {
 798         struct btree_path *path;
 799         unsigned i;
 800
 801         trans_for_each_path(trans, path, i)
 802                 __bch2_btree_path_unlock(trans, path);
 803 }
 804
 805 void bch2_trans_unlock_long(struct btree_trans *trans)
 806 {
 807         bch2_trans_unlock(trans);
 808         bch2_trans_srcu_unlock(trans);
 809 }
 810
 811 bool bch2_trans_locked(struct btree_trans *trans)
 812 {
 813         struct btree_path *path;
 814         unsigned i;
 815
 816         trans_for_each_path(trans, path, i)
 817                 if (path->nodes_locked)
 818                         return true;
 819         return false;
 820 }
 821
 822 int __bch2_trans_mutex_lock(struct btree_trans *trans,
 823                             struct mutex *lock)
 824 {
 825         int ret = drop_locks_do(trans, (mutex_lock(lock), 0));
 826
 827         if (ret)
 828                 mutex_unlock(lock);
 829         return ret;
 830 }
 831
 832 /* Debug */
 833
 834 #ifdef CONFIG_BCACHEFS_DEBUG
 835
 836 void bch2_btree_path_verify_locks(struct btree_path *path)
 837 {
 838         unsigned l;
 839
 840         if (!path->nodes_locked) {
 841                 BUG_ON(path->uptodate == BTREE_ITER_UPTODATE &&
 842                        btree_path_node(path, path->level));
 843                 return;
 844         }
 845
 846         for (l = 0; l < BTREE_MAX_DEPTH; l++) {
 847                 int want = btree_lock_want(path, l);
 848                 int have = btree_node_locked_type(path, l);
 849
 850                 BUG_ON(!is_btree_node(path, l) && have != BTREE_NODE_UNLOCKED);
 851
 852                 BUG_ON(is_btree_node(path, l) &&
 853                        (want == BTREE_NODE_UNLOCKED ||
 854                         have != BTREE_NODE_WRITE_LOCKED) &&
 855                        want != have);
 856         }
 857 }
 858
 859 void bch2_trans_verify_locks(struct btree_trans *trans)
 860 {
 861         struct btree_path *path;
 862         unsigned i;
 863
 864         trans_for_each_path(trans, path, i)
 865                 bch2_btree_path_verify_locks(path);
 866 }
 867
 868 #endif