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 struct lock_class_key bch2_btree_node_lock_key;
   8
   9 /* Btree node locking: */
  10
  11 static inline void six_lock_readers_add(struct six_lock *lock, int nr)
  12 {
  13         if (lock->readers)
  14                 this_cpu_add(*lock->readers, nr);
  15         else if (nr > 0)
  16                 atomic64_add(__SIX_VAL(read_lock, nr), &lock->state.counter);
  17         else
  18                 atomic64_sub(__SIX_VAL(read_lock, -nr), &lock->state.counter);
  19 }
  20
  21 struct six_lock_count bch2_btree_node_lock_counts(struct btree_trans *trans,
  22                                                   struct btree_path *skip,
  23                                                   struct btree_bkey_cached_common *b,
  24                                                   unsigned level)
  25 {
  26         struct btree_path *path;
  27         struct six_lock_count ret;
  28
  29         memset(&ret, 0, sizeof(ret));
  30
  31         if (IS_ERR_OR_NULL(b))
  32                 return ret;
  33
  34         trans_for_each_path(trans, path)
  35                 if (path != skip && &path->l[level].b->c == b) {
  36                         int t = btree_node_locked_type(path, level);
  37
  38                         if (t != BTREE_NODE_UNLOCKED)
  39                                 ret.n[t]++;
  40                 }
  41
  42         return ret;
  43 }
  44
  45 /* unlock */
  46
  47 void bch2_btree_node_unlock_write(struct btree_trans *trans,
  48                         struct btree_path *path, struct btree *b)
  49 {
  50         bch2_btree_node_unlock_write_inlined(trans, path, b);
  51 }
  52
  53 /* lock */
  54
  55 /*
  56  * @trans wants to lock @b with type @type
  57  */
  58 struct trans_waiting_for_lock {
  59         struct btree_trans              *trans;
  60         struct btree_bkey_cached_common *node_want;
  61         enum six_lock_type              lock_want;
  62
  63         /* for iterating over held locks :*/
  64         u8                              path_idx;
  65         u8                              level;
  66         u64                             lock_start_time;
  67 };
  68
  69 struct lock_graph {
  70         struct trans_waiting_for_lock   g[8];
  71         unsigned                        nr;
  72 };
  73
  74 static noinline void print_cycle(struct printbuf *out, struct lock_graph *g)
  75 {
  76         struct trans_waiting_for_lock *i;
  77
  78         prt_printf(out, "Found lock cycle (%u entries):", g->nr);
  79         prt_newline(out);
  80
  81         for (i = g->g; i < g->g + g->nr; i++)
  82                 bch2_btree_trans_to_text(out, i->trans);
  83 }
  84
  85 static noinline void print_chain(struct printbuf *out, struct lock_graph *g)
  86 {
  87         struct trans_waiting_for_lock *i;
  88
  89         for (i = g->g; i != g->g + g->nr; i++) {
  90                 if (i != g->g)
  91                         prt_str(out, "<- ");
  92                 prt_printf(out, "%u ", i->trans->locking_wait.task->pid);
  93         }
  94         prt_newline(out);
  95 }
  96
  97 static void lock_graph_up(struct lock_graph *g)
  98 {
  99         closure_put(&g->g[--g->nr].trans->ref);
 100 }
 101
 102 static void lock_graph_down(struct lock_graph *g, struct btree_trans *trans)
 103 {
 104         closure_get(&trans->ref);
 105
 106         g->g[g->nr++] = (struct trans_waiting_for_lock) {
 107                 .trans          = trans,
 108                 .node_want      = trans->locking,
 109                 .lock_want      = trans->locking_wait.lock_want,
 110         };
 111 }
 112
 113 static bool lock_graph_remove_non_waiters(struct lock_graph *g)
 114 {
 115         struct trans_waiting_for_lock *i;
 116
 117         for (i = g->g + 1; i < g->g + g->nr; i++)
 118                 if (i->trans->locking != i->node_want ||
 119                     i->trans->locking_wait.start_time != i[-1].lock_start_time) {
 120                         while (g->g + g->nr > i)
 121                                 lock_graph_up(g);
 122                         return true;
 123                 }
 124
 125         return false;
 126 }
 127
 128 static int abort_lock(struct lock_graph *g, struct trans_waiting_for_lock *i)
 129 {
 130         if (i == g->g) {
 131                 trace_and_count(i->trans->c, trans_restart_would_deadlock, i->trans, _RET_IP_);
 132                 return btree_trans_restart(i->trans, BCH_ERR_transaction_restart_would_deadlock);
 133         } else {
 134                 i->trans->lock_must_abort = true;
 135                 wake_up_process(i->trans->locking_wait.task);
 136                 return 0;
 137         }
 138 }
 139
 140 static int btree_trans_abort_preference(struct btree_trans *trans)
 141 {
 142         if (trans->lock_may_not_fail)
 143                 return 0;
 144         if (trans->locking_wait.lock_want == SIX_LOCK_write)
 145                 return 1;
 146         if (!trans->in_traverse_all)
 147                 return 2;
 148         return 3;
 149 }
 150
 151 static noinline int break_cycle(struct lock_graph *g, struct printbuf *cycle)
 152 {
 153         struct trans_waiting_for_lock *i, *abort = NULL;
 154         unsigned best = 0, pref;
 155         int ret;
 156
 157         if (lock_graph_remove_non_waiters(g))
 158                 return 0;
 159
 160         /* Only checking, for debugfs: */
 161         if (cycle) {
 162                 print_cycle(cycle, g);
 163                 ret = -1;
 164                 goto out;
 165         }
 166
 167         for (i = g->g; i < g->g + g->nr; i++) {
 168                 pref = btree_trans_abort_preference(i->trans);
 169                 if (pref > best) {
 170                         abort = i;
 171                         best = pref;
 172                 }
 173         }
 174
 175         if (unlikely(!best)) {
 176                 struct printbuf buf = PRINTBUF;
 177
 178                 prt_printf(&buf, bch2_fmt(g->g->trans->c, "cycle of nofail locks"));
 179
 180                 for (i = g->g; i < g->g + g->nr; i++) {
 181                         struct btree_trans *trans = i->trans;
 182
 183                         bch2_btree_trans_to_text(&buf, trans);
 184
 185                         prt_printf(&buf, "backtrace:");
 186                         prt_newline(&buf);
 187                         printbuf_indent_add(&buf, 2);
 188                         bch2_prt_backtrace(&buf, trans->locking_wait.task);
 189                         printbuf_indent_sub(&buf, 2);
 190                         prt_newline(&buf);
 191                 }
 192
 193                 bch2_print_string_as_lines(KERN_ERR, buf.buf);
 194                 printbuf_exit(&buf);
 195                 BUG();
 196         }
 197
 198         ret = abort_lock(g, abort);
 199 out:
 200         if (ret)
 201                 while (g->nr)
 202                         lock_graph_up(g);
 203         return ret;
 204 }
 205
 206 static int lock_graph_descend(struct lock_graph *g, struct btree_trans *trans,
 207                               struct printbuf *cycle)
 208 {
 209         struct btree_trans *orig_trans = g->g->trans;
 210         struct trans_waiting_for_lock *i;
 211
 212         for (i = g->g; i < g->g + g->nr; i++)
 213                 if (i->trans == trans)
 214                         return break_cycle(g, cycle);
 215
 216         if (g->nr == ARRAY_SIZE(g->g)) {
 217                 if (orig_trans->lock_may_not_fail)
 218                         return 0;
 219
 220                 while (g->nr)
 221                         lock_graph_up(g);
 222                 trace_and_count(trans->c, trans_restart_would_deadlock_recursion_limit, trans, _RET_IP_);
 223                 return btree_trans_restart(orig_trans, BCH_ERR_transaction_restart_deadlock_recursion_limit);
 224         }
 225
 226         lock_graph_down(g, trans);
 227         return 0;
 228 }
 229
 230 static bool lock_type_conflicts(enum six_lock_type t1, enum six_lock_type t2)
 231 {
 232         return t1 + t2 > 1;
 233 }
 234
 235 int bch2_check_for_deadlock(struct btree_trans *trans, struct printbuf *cycle)
 236 {
 237         struct lock_graph g;
 238         struct trans_waiting_for_lock *top;
 239         struct btree_bkey_cached_common *b;
 240         struct btree_path *path;
 241         int ret;
 242
 243         if (trans->lock_must_abort) {
 244                 trace_and_count(trans->c, trans_restart_would_deadlock, trans, _RET_IP_);
 245                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock);
 246         }
 247
 248         g.nr = 0;
 249         lock_graph_down(&g, trans);
 250 next:
 251         if (!g.nr)
 252                 return 0;
 253
 254         top = &g.g[g.nr - 1];
 255
 256         trans_for_each_path_from(top->trans, path, top->path_idx) {
 257                 if (!path->nodes_locked)
 258                         continue;
 259
 260                 if (top->path_idx != path->idx) {
 261                         top->path_idx           = path->idx;
 262                         top->level              = 0;
 263                         top->lock_start_time    = 0;
 264                 }
 265
 266                 for (;
 267                      top->level < BTREE_MAX_DEPTH;
 268                      top->level++, top->lock_start_time = 0) {
 269                         int lock_held = btree_node_locked_type(path, top->level);
 270
 271                         if (lock_held == BTREE_NODE_UNLOCKED)
 272                                 continue;
 273
 274                         b = &READ_ONCE(path->l[top->level].b)->c;
 275
 276                         if (IS_ERR_OR_NULL(b)) {
 277                                 BUG_ON(!lock_graph_remove_non_waiters(&g));
 278                                 goto next;
 279                         }
 280
 281                         if (list_empty_careful(&b->lock.wait_list))
 282                                 continue;
 283
 284                         raw_spin_lock(&b->lock.wait_lock);
 285                         list_for_each_entry(trans, &b->lock.wait_list, locking_wait.list) {
 286                                 BUG_ON(b != trans->locking);
 287
 288                                 if (top->lock_start_time &&
 289                                     time_after_eq64(top->lock_start_time, trans->locking_wait.start_time))
 290                                         continue;
 291
 292                                 top->lock_start_time = trans->locking_wait.start_time;
 293
 294                                 /* Don't check for self deadlock: */
 295                                 if (trans == top->trans ||
 296                                     !lock_type_conflicts(lock_held, trans->locking_wait.lock_want))
 297                                         continue;
 298
 299                                 ret = lock_graph_descend(&g, trans, cycle);
 300                                 raw_spin_unlock(&b->lock.wait_lock);
 301
 302                                 if (ret)
 303                                         return ret;
 304                                 goto next;
 305
 306                         }
 307                         raw_spin_unlock(&b->lock.wait_lock);
 308                 }
 309         }
 310
 311         if (g.nr > 1 && cycle)
 312                 print_chain(cycle, &g);
 313         lock_graph_up(&g);
 314         goto next;
 315 }
 316
 317 int bch2_six_check_for_deadlock(struct six_lock *lock, void *p)
 318 {
 319         struct btree_trans *trans = p;
 320
 321         return bch2_check_for_deadlock(trans, NULL);
 322 }
 323
 324 int __bch2_btree_node_lock_write(struct btree_trans *trans, struct btree_path *path,
 325                                  struct btree_bkey_cached_common *b,
 326                                  bool lock_may_not_fail)
 327 {
 328         int readers = bch2_btree_node_lock_counts(trans, NULL, b, b->level).n[SIX_LOCK_read];
 329         int ret;
 330
 331         /*
 332          * Must drop our read locks before calling six_lock_write() -
 333          * six_unlock() won't do wakeups until the reader count
 334          * goes to 0, and it's safe because we have the node intent
 335          * locked:
 336          */
 337         six_lock_readers_add(&b->lock, -readers);
 338         ret = __btree_node_lock_nopath(trans, b, SIX_LOCK_write, lock_may_not_fail);
 339         six_lock_readers_add(&b->lock, readers);
 340
 341         if (ret)
 342                 mark_btree_node_locked_noreset(path, b->level, SIX_LOCK_intent);
 343
 344         return ret;
 345 }
 346
 347 /* relock */
 348
 349 static inline bool btree_path_get_locks(struct btree_trans *trans,
 350                                         struct btree_path *path,
 351                                         bool upgrade)
 352 {
 353         unsigned l = path->level;
 354         int fail_idx = -1;
 355
 356         do {
 357                 if (!btree_path_node(path, l))
 358                         break;
 359
 360                 if (!(upgrade
 361                       ? bch2_btree_node_upgrade(trans, path, l)
 362                       : bch2_btree_node_relock(trans, path, l)))
 363                         fail_idx = l;
 364
 365                 l++;
 366         } while (l < path->locks_want);
 367
 368         /*
 369          * When we fail to get a lock, we have to ensure that any child nodes
 370          * can't be relocked so bch2_btree_path_traverse has to walk back up to
 371          * the node that we failed to relock:
 372          */
 373         if (fail_idx >= 0) {
 374                 __bch2_btree_path_unlock(trans, path);
 375                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 376
 377                 do {
 378                         path->l[fail_idx].b = upgrade
 379                                 ? ERR_PTR(-BCH_ERR_no_btree_node_upgrade)
 380                                 : ERR_PTR(-BCH_ERR_no_btree_node_relock);
 381                         --fail_idx;
 382                 } while (fail_idx >= 0);
 383         }
 384
 385         if (path->uptodate == BTREE_ITER_NEED_RELOCK)
 386                 path->uptodate = BTREE_ITER_UPTODATE;
 387
 388         bch2_trans_verify_locks(trans);
 389
 390         return path->uptodate < BTREE_ITER_NEED_RELOCK;
 391 }
 392
 393 bool __bch2_btree_node_relock(struct btree_trans *trans,
 394                               struct btree_path *path, unsigned level,
 395                               bool trace)
 396 {
 397         struct btree *b = btree_path_node(path, level);
 398         int want = __btree_lock_want(path, level);
 399
 400         if (race_fault())
 401                 goto fail;
 402
 403         if (six_relock_type(&b->c.lock, want, path->l[level].lock_seq) ||
 404             (btree_node_lock_seq_matches(path, b, level) &&
 405              btree_node_lock_increment(trans, &b->c, level, want))) {
 406                 mark_btree_node_locked(trans, path, level, want);
 407                 return true;
 408         }
 409 fail:
 410         if (trace)
 411                 trace_and_count(trans->c, btree_path_relock_fail, trans, _RET_IP_, path, level);
 412         return false;
 413 }
 414
 415 /* upgrade */
 416
 417 bool bch2_btree_node_upgrade(struct btree_trans *trans,
 418                              struct btree_path *path, unsigned level)
 419 {
 420         struct btree *b = path->l[level].b;
 421         struct six_lock_count count = bch2_btree_node_lock_counts(trans, path, &b->c, level);
 422
 423         if (!is_btree_node(path, level))
 424                 return false;
 425
 426         switch (btree_lock_want(path, level)) {
 427         case BTREE_NODE_UNLOCKED:
 428                 BUG_ON(btree_node_locked(path, level));
 429                 return true;
 430         case BTREE_NODE_READ_LOCKED:
 431                 BUG_ON(btree_node_intent_locked(path, level));
 432                 return bch2_btree_node_relock(trans, path, level);
 433         case BTREE_NODE_INTENT_LOCKED:
 434                 break;
 435         case BTREE_NODE_WRITE_LOCKED:
 436                 BUG();
 437         }
 438
 439         if (btree_node_intent_locked(path, level))
 440                 return true;
 441
 442         if (race_fault())
 443                 return false;
 444
 445         if (btree_node_locked(path, level)) {
 446                 bool ret;
 447
 448                 six_lock_readers_add(&b->c.lock, -count.n[SIX_LOCK_read]);
 449                 ret = six_lock_tryupgrade(&b->c.lock);
 450                 six_lock_readers_add(&b->c.lock, count.n[SIX_LOCK_read]);
 451
 452                 if (ret)
 453                         goto success;
 454         } else {
 455                 if (six_relock_type(&b->c.lock, SIX_LOCK_intent, path->l[level].lock_seq))
 456                         goto success;
 457         }
 458
 459         /*
 460          * Do we already have an intent lock via another path? If so, just bump
 461          * lock count:
 462          */
 463         if (btree_node_lock_seq_matches(path, b, level) &&
 464             btree_node_lock_increment(trans, &b->c, level, BTREE_NODE_INTENT_LOCKED)) {
 465                 btree_node_unlock(trans, path, level);
 466                 goto success;
 467         }
 468
 469         trace_and_count(trans->c, btree_path_upgrade_fail, trans, _RET_IP_, path, level);
 470         return false;
 471 success:
 472         mark_btree_node_locked_noreset(path, level, SIX_LOCK_intent);
 473         return true;
 474 }
 475
 476 /* Btree path locking: */
 477
 478 /*
 479  * Only for btree_cache.c - only relocks intent locks
 480  */
 481 int bch2_btree_path_relock_intent(struct btree_trans *trans,
 482                                   struct btree_path *path)
 483 {
 484         unsigned l;
 485
 486         for (l = path->level;
 487              l < path->locks_want && btree_path_node(path, l);
 488              l++) {
 489                 if (!bch2_btree_node_relock(trans, path, l)) {
 490                         __bch2_btree_path_unlock(trans, path);
 491                         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 492                         trace_and_count(trans->c, trans_restart_relock_path_intent, trans, _RET_IP_, path);
 493                         return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_path_intent);
 494                 }
 495         }
 496
 497         return 0;
 498 }
 499
 500 __flatten
 501 bool bch2_btree_path_relock_norestart(struct btree_trans *trans,
 502                         struct btree_path *path, unsigned long trace_ip)
 503 {
 504         return btree_path_get_locks(trans, path, false);
 505 }
 506
 507 __flatten
 508 bool bch2_btree_path_upgrade_norestart(struct btree_trans *trans,
 509                         struct btree_path *path, unsigned long trace_ip)
 510 {
 511         return btree_path_get_locks(trans, path, true);
 512 }
 513
 514 bool bch2_btree_path_upgrade_noupgrade_sibs(struct btree_trans *trans,
 515                                struct btree_path *path,
 516                                unsigned new_locks_want)
 517 {
 518         EBUG_ON(path->locks_want >= new_locks_want);
 519
 520         path->locks_want = new_locks_want;
 521
 522         return btree_path_get_locks(trans, path, true);
 523 }
 524
 525 bool __bch2_btree_path_upgrade(struct btree_trans *trans,
 526                                struct btree_path *path,
 527                                unsigned new_locks_want)
 528 {
 529         struct btree_path *linked;
 530
 531         if (bch2_btree_path_upgrade_noupgrade_sibs(trans, path, new_locks_want))
 532                 return true;
 533
 534         /*
 535          * XXX: this is ugly - we'd prefer to not be mucking with other
 536          * iterators in the btree_trans here.
 537          *
 538          * On failure to upgrade the iterator, setting iter->locks_want and
 539          * calling get_locks() is sufficient to make bch2_btree_path_traverse()
 540          * get the locks we want on transaction restart.
 541          *
 542          * But if this iterator was a clone, on transaction restart what we did
 543          * to this iterator isn't going to be preserved.
 544          *
 545          * Possibly we could add an iterator field for the parent iterator when
 546          * an iterator is a copy - for now, we'll just upgrade any other
 547          * iterators with the same btree id.
 548          *
 549          * The code below used to be needed to ensure ancestor nodes get locked
 550          * before interior nodes - now that's handled by
 551          * bch2_btree_path_traverse_all().
 552          */
 553         if (!path->cached && !trans->in_traverse_all)
 554                 trans_for_each_path(trans, linked)
 555                         if (linked != path &&
 556                             linked->cached == path->cached &&
 557                             linked->btree_id == path->btree_id &&
 558                             linked->locks_want < new_locks_want) {
 559                                 linked->locks_want = new_locks_want;
 560                                 btree_path_get_locks(trans, linked, true);
 561                         }
 562
 563         return false;
 564 }
 565
 566 void __bch2_btree_path_downgrade(struct btree_trans *trans,
 567                                  struct btree_path *path,
 568                                  unsigned new_locks_want)
 569 {
 570         unsigned l;
 571
 572         EBUG_ON(path->locks_want < new_locks_want);
 573
 574         path->locks_want = new_locks_want;
 575
 576         while (path->nodes_locked &&
 577                (l = btree_path_highest_level_locked(path)) >= path->locks_want) {
 578                 if (l > path->level) {
 579                         btree_node_unlock(trans, path, l);
 580                 } else {
 581                         if (btree_node_intent_locked(path, l)) {
 582                                 six_lock_downgrade(&path->l[l].b->c.lock);
 583                                 mark_btree_node_locked_noreset(path, l, SIX_LOCK_read);
 584                         }
 585                         break;
 586                 }
 587         }
 588
 589         bch2_btree_path_verify_locks(path);
 590 }
 591
 592 /* Btree transaction locking: */
 593
 594 void bch2_trans_downgrade(struct btree_trans *trans)
 595 {
 596         struct btree_path *path;
 597
 598         trans_for_each_path(trans, path)
 599                 bch2_btree_path_downgrade(trans, path);
 600 }
 601
 602 int bch2_trans_relock(struct btree_trans *trans)
 603 {
 604         struct btree_path *path;
 605
 606         if (unlikely(trans->restarted))
 607                 return -((int) trans->restarted);
 608
 609         trans_for_each_path(trans, path)
 610                 if (path->should_be_locked &&
 611                     !bch2_btree_path_relock_norestart(trans, path, _RET_IP_)) {
 612                         trace_and_count(trans->c, trans_restart_relock, trans, _RET_IP_, path);
 613                         return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
 614                 }
 615         return 0;
 616 }
 617
 618 void bch2_trans_unlock(struct btree_trans *trans)
 619 {
 620         struct btree_path *path;
 621
 622         trans_for_each_path(trans, path)
 623                 __bch2_btree_path_unlock(trans, path);
 624
 625         /*
 626          * bch2_gc_btree_init_recurse() doesn't use btree iterators for walking
 627          * btree nodes, it implements its own walking:
 628          */
 629         EBUG_ON(!trans->is_initial_gc &&
 630                 lock_class_is_held(&bch2_btree_node_lock_key));
 631 }
 632
 633 bool bch2_trans_locked(struct btree_trans *trans)
 634 {
 635         struct btree_path *path;
 636
 637         trans_for_each_path(trans, path)
 638                 if (path->nodes_locked)
 639                         return true;
 640         return false;
 641 }
 642
 643 /* Debug */
 644
 645 #ifdef CONFIG_BCACHEFS_DEBUG
 646
 647 void bch2_btree_path_verify_locks(struct btree_path *path)
 648 {
 649         unsigned l;
 650
 651         if (!path->nodes_locked) {
 652                 BUG_ON(path->uptodate == BTREE_ITER_UPTODATE &&
 653                        btree_path_node(path, path->level));
 654                 return;
 655         }
 656
 657         for (l = 0; l < BTREE_MAX_DEPTH; l++) {
 658                 int want = btree_lock_want(path, l);
 659                 int have = btree_node_locked_type(path, l);
 660
 661                 BUG_ON(!is_btree_node(path, l) && have != BTREE_NODE_UNLOCKED);
 662
 663                 BUG_ON(is_btree_node(path, l) &&
 664                        (want == BTREE_NODE_UNLOCKED ||
 665                         have != BTREE_NODE_WRITE_LOCKED) &&
 666                        want != have);
 667         }
 668 }
 669
 670 void bch2_trans_verify_locks(struct btree_trans *trans)
 671 {
 672         struct btree_path *path;
 673
 674         trans_for_each_path(trans, path)
 675                 bch2_btree_path_verify_locks(path);
 676 }
 677
 678 #endif