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 bch_fs *c = g->g->trans->c;
 177                 struct printbuf buf = PRINTBUF;
 178
 179                 bch_err(c, "cycle of nofail locks");
 180
 181                 for (i = g->g; i < g->g + g->nr; i++) {
 182                         struct btree_trans *trans = i->trans;
 183
 184                         bch2_btree_trans_to_text(&buf, trans);
 185
 186                         prt_printf(&buf, "backtrace:");
 187                         prt_newline(&buf);
 188                         printbuf_indent_add(&buf, 2);
 189                         bch2_prt_backtrace(&buf, trans->locking_wait.task);
 190                         printbuf_indent_sub(&buf, 2);
 191                         prt_newline(&buf);
 192                 }
 193
 194                 bch2_print_string_as_lines(KERN_ERR, buf.buf);
 195                 printbuf_exit(&buf);
 196                 BUG();
 197         }
 198
 199         ret = abort_lock(g, abort);
 200 out:
 201         if (ret)
 202                 while (g->nr)
 203                         lock_graph_up(g);
 204         return ret;
 205 }
 206
 207 static int lock_graph_descend(struct lock_graph *g, struct btree_trans *trans,
 208                               struct printbuf *cycle)
 209 {
 210         struct btree_trans *orig_trans = g->g->trans;
 211         struct trans_waiting_for_lock *i;
 212
 213         for (i = g->g; i < g->g + g->nr; i++)
 214                 if (i->trans == trans)
 215                         return break_cycle(g, cycle);
 216
 217         if (g->nr == ARRAY_SIZE(g->g)) {
 218                 if (orig_trans->lock_may_not_fail)
 219                         return 0;
 220
 221                 while (g->nr)
 222                         lock_graph_up(g);
 223                 trace_and_count(trans->c, trans_restart_would_deadlock_recursion_limit, trans, _RET_IP_);
 224                 return btree_trans_restart(orig_trans, BCH_ERR_transaction_restart_deadlock_recursion_limit);
 225         }
 226
 227         lock_graph_down(g, trans);
 228         return 0;
 229 }
 230
 231 static bool lock_type_conflicts(enum six_lock_type t1, enum six_lock_type t2)
 232 {
 233         return t1 + t2 > 1;
 234 }
 235
 236 int bch2_check_for_deadlock(struct btree_trans *trans, struct printbuf *cycle)
 237 {
 238         struct lock_graph g;
 239         struct trans_waiting_for_lock *top;
 240         struct btree_bkey_cached_common *b;
 241         struct btree_path *path;
 242         int ret;
 243
 244         if (trans->lock_must_abort) {
 245                 trace_and_count(trans->c, trans_restart_would_deadlock, trans, _RET_IP_);
 246                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock);
 247         }
 248
 249         g.nr = 0;
 250         lock_graph_down(&g, trans);
 251 next:
 252         if (!g.nr)
 253                 return 0;
 254
 255         top = &g.g[g.nr - 1];
 256
 257         trans_for_each_path_from(top->trans, path, top->path_idx) {
 258                 if (!path->nodes_locked)
 259                         continue;
 260
 261                 if (top->path_idx != path->idx) {
 262                         top->path_idx           = path->idx;
 263                         top->level              = 0;
 264                         top->lock_start_time    = 0;
 265                 }
 266
 267                 for (;
 268                      top->level < BTREE_MAX_DEPTH;
 269                      top->level++, top->lock_start_time = 0) {
 270                         int lock_held = btree_node_locked_type(path, top->level);
 271
 272                         if (lock_held == BTREE_NODE_UNLOCKED)
 273                                 continue;
 274
 275                         b = &READ_ONCE(path->l[top->level].b)->c;
 276
 277                         if (unlikely(IS_ERR_OR_NULL(b))) {
 278                                 BUG_ON(!lock_graph_remove_non_waiters(&g));
 279                                 goto next;
 280                         }
 281
 282                         if (list_empty_careful(&b->lock.wait_list))
 283                                 continue;
 284
 285                         raw_spin_lock(&b->lock.wait_lock);
 286                         list_for_each_entry(trans, &b->lock.wait_list, locking_wait.list) {
 287                                 BUG_ON(b != trans->locking);
 288
 289                                 if (top->lock_start_time &&
 290                                     time_after_eq64(top->lock_start_time, trans->locking_wait.start_time))
 291                                         continue;
 292
 293                                 top->lock_start_time = trans->locking_wait.start_time;
 294
 295                                 /* Don't check for self deadlock: */
 296                                 if (trans == top->trans ||
 297                                     !lock_type_conflicts(lock_held, trans->locking_wait.lock_want))
 298                                         continue;
 299
 300                                 ret = lock_graph_descend(&g, trans, cycle);
 301                                 raw_spin_unlock(&b->lock.wait_lock);
 302
 303                                 if (ret)
 304                                         return ret;
 305                                 goto next;
 306
 307                         }
 308                         raw_spin_unlock(&b->lock.wait_lock);
 309                 }
 310         }
 311
 312         if (g.nr > 1 && cycle)
 313                 print_chain(cycle, &g);
 314         lock_graph_up(&g);
 315         goto next;
 316 }
 317
 318 int bch2_six_check_for_deadlock(struct six_lock *lock, void *p)
 319 {
 320         struct btree_trans *trans = p;
 321
 322         return bch2_check_for_deadlock(trans, NULL);
 323 }
 324
 325 int __bch2_btree_node_lock_write(struct btree_trans *trans, struct btree_path *path,
 326                                  struct btree_bkey_cached_common *b,
 327                                  bool lock_may_not_fail)
 328 {
 329         int readers = bch2_btree_node_lock_counts(trans, NULL, b, b->level).n[SIX_LOCK_read];
 330         int ret;
 331
 332         /*
 333          * Must drop our read locks before calling six_lock_write() -
 334          * six_unlock() won't do wakeups until the reader count
 335          * goes to 0, and it's safe because we have the node intent
 336          * locked:
 337          */
 338         six_lock_readers_add(&b->lock, -readers);
 339         ret = __btree_node_lock_nopath(trans, b, SIX_LOCK_write, lock_may_not_fail);
 340         six_lock_readers_add(&b->lock, readers);
 341
 342         if (ret)
 343                 mark_btree_node_locked_noreset(path, b->level, SIX_LOCK_intent);
 344
 345         return ret;
 346 }
 347
 348 /* relock */
 349
 350 static inline bool btree_path_get_locks(struct btree_trans *trans,
 351                                         struct btree_path *path,
 352                                         bool upgrade)
 353 {
 354         unsigned l = path->level;
 355         int fail_idx = -1;
 356
 357         do {
 358                 if (!btree_path_node(path, l))
 359                         break;
 360
 361                 if (!(upgrade
 362                       ? bch2_btree_node_upgrade(trans, path, l)
 363                       : bch2_btree_node_relock(trans, path, l)))
 364                         fail_idx = l;
 365
 366                 l++;
 367         } while (l < path->locks_want);
 368
 369         /*
 370          * When we fail to get a lock, we have to ensure that any child nodes
 371          * can't be relocked so bch2_btree_path_traverse has to walk back up to
 372          * the node that we failed to relock:
 373          */
 374         if (fail_idx >= 0) {
 375                 __bch2_btree_path_unlock(trans, path);
 376                 btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 377
 378                 do {
 379                         path->l[fail_idx].b = upgrade
 380                                 ? ERR_PTR(-BCH_ERR_no_btree_node_upgrade)
 381                                 : ERR_PTR(-BCH_ERR_no_btree_node_relock);
 382                         --fail_idx;
 383                 } while (fail_idx >= 0);
 384         }
 385
 386         if (path->uptodate == BTREE_ITER_NEED_RELOCK)
 387                 path->uptodate = BTREE_ITER_UPTODATE;
 388
 389         bch2_trans_verify_locks(trans);
 390
 391         return path->uptodate < BTREE_ITER_NEED_RELOCK;
 392 }
 393
 394 bool __bch2_btree_node_relock(struct btree_trans *trans,
 395                               struct btree_path *path, unsigned level,
 396                               bool trace)
 397 {
 398         struct btree *b = btree_path_node(path, level);
 399         int want = __btree_lock_want(path, level);
 400
 401         if (race_fault())
 402                 goto fail;
 403
 404         if (six_relock_type(&b->c.lock, want, path->l[level].lock_seq) ||
 405             (btree_node_lock_seq_matches(path, b, level) &&
 406              btree_node_lock_increment(trans, &b->c, level, want))) {
 407                 mark_btree_node_locked(trans, path, level, want);
 408                 return true;
 409         }
 410 fail:
 411         if (trace)
 412                 trace_and_count(trans->c, btree_path_relock_fail, trans, _RET_IP_, path, level);
 413         return false;
 414 }
 415
 416 /* upgrade */
 417
 418 bool bch2_btree_node_upgrade(struct btree_trans *trans,
 419                              struct btree_path *path, unsigned level)
 420 {
 421         struct btree *b = path->l[level].b;
 422         struct six_lock_count count = bch2_btree_node_lock_counts(trans, path, &b->c, level);
 423
 424         if (!is_btree_node(path, level))
 425                 return false;
 426
 427         switch (btree_lock_want(path, level)) {
 428         case BTREE_NODE_UNLOCKED:
 429                 BUG_ON(btree_node_locked(path, level));
 430                 return true;
 431         case BTREE_NODE_READ_LOCKED:
 432                 BUG_ON(btree_node_intent_locked(path, level));
 433                 return bch2_btree_node_relock(trans, path, level);
 434         case BTREE_NODE_INTENT_LOCKED:
 435                 break;
 436         case BTREE_NODE_WRITE_LOCKED:
 437                 BUG();
 438         }
 439
 440         if (btree_node_intent_locked(path, level))
 441                 return true;
 442
 443         if (race_fault())
 444                 return false;
 445
 446         if (btree_node_locked(path, level)) {
 447                 bool ret;
 448
 449                 six_lock_readers_add(&b->c.lock, -count.n[SIX_LOCK_read]);
 450                 ret = six_lock_tryupgrade(&b->c.lock);
 451                 six_lock_readers_add(&b->c.lock, count.n[SIX_LOCK_read]);
 452
 453                 if (ret)
 454                         goto success;
 455         } else {
 456                 if (six_relock_type(&b->c.lock, SIX_LOCK_intent, path->l[level].lock_seq))
 457                         goto success;
 458         }
 459
 460         /*
 461          * Do we already have an intent lock via another path? If so, just bump
 462          * lock count:
 463          */
 464         if (btree_node_lock_seq_matches(path, b, level) &&
 465             btree_node_lock_increment(trans, &b->c, level, BTREE_NODE_INTENT_LOCKED)) {
 466                 btree_node_unlock(trans, path, level);
 467                 goto success;
 468         }
 469
 470         trace_and_count(trans->c, btree_path_upgrade_fail, trans, _RET_IP_, path, level);
 471         return false;
 472 success:
 473         mark_btree_node_locked_noreset(path, level, SIX_LOCK_intent);
 474         return true;
 475 }
 476
 477 /* Btree path locking: */
 478
 479 /*
 480  * Only for btree_cache.c - only relocks intent locks
 481  */
 482 int bch2_btree_path_relock_intent(struct btree_trans *trans,
 483                                   struct btree_path *path)
 484 {
 485         unsigned l;
 486
 487         for (l = path->level;
 488              l < path->locks_want && btree_path_node(path, l);
 489              l++) {
 490                 if (!bch2_btree_node_relock(trans, path, l)) {
 491                         __bch2_btree_path_unlock(trans, path);
 492                         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 493                         trace_and_count(trans->c, trans_restart_relock_path_intent, trans, _RET_IP_, path);
 494                         return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_path_intent);
 495                 }
 496         }
 497
 498         return 0;
 499 }
 500
 501 __flatten
 502 bool bch2_btree_path_relock_norestart(struct btree_trans *trans,
 503                         struct btree_path *path, unsigned long trace_ip)
 504 {
 505         return btree_path_get_locks(trans, path, false);
 506 }
 507
 508 __flatten
 509 bool bch2_btree_path_upgrade_norestart(struct btree_trans *trans,
 510                         struct btree_path *path, unsigned long trace_ip)
 511 {
 512         return btree_path_get_locks(trans, path, true);
 513 }
 514
 515 bool bch2_btree_path_upgrade_noupgrade_sibs(struct btree_trans *trans,
 516                                struct btree_path *path,
 517                                unsigned new_locks_want)
 518 {
 519         EBUG_ON(path->locks_want >= new_locks_want);
 520
 521         path->locks_want = new_locks_want;
 522
 523         return btree_path_get_locks(trans, path, true);
 524 }
 525
 526 bool __bch2_btree_path_upgrade(struct btree_trans *trans,
 527                                struct btree_path *path,
 528                                unsigned new_locks_want)
 529 {
 530         struct btree_path *linked;
 531
 532         if (bch2_btree_path_upgrade_noupgrade_sibs(trans, path, new_locks_want))
 533                 return true;
 534
 535         /*
 536          * XXX: this is ugly - we'd prefer to not be mucking with other
 537          * iterators in the btree_trans here.
 538          *
 539          * On failure to upgrade the iterator, setting iter->locks_want and
 540          * calling get_locks() is sufficient to make bch2_btree_path_traverse()
 541          * get the locks we want on transaction restart.
 542          *
 543          * But if this iterator was a clone, on transaction restart what we did
 544          * to this iterator isn't going to be preserved.
 545          *
 546          * Possibly we could add an iterator field for the parent iterator when
 547          * an iterator is a copy - for now, we'll just upgrade any other
 548          * iterators with the same btree id.
 549          *
 550          * The code below used to be needed to ensure ancestor nodes get locked
 551          * before interior nodes - now that's handled by
 552          * bch2_btree_path_traverse_all().
 553          */
 554         if (!path->cached && !trans->in_traverse_all)
 555                 trans_for_each_path(trans, linked)
 556                         if (linked != path &&
 557                             linked->cached == path->cached &&
 558                             linked->btree_id == path->btree_id &&
 559                             linked->locks_want < new_locks_want) {
 560                                 linked->locks_want = new_locks_want;
 561                                 btree_path_get_locks(trans, linked, true);
 562                         }
 563
 564         return false;
 565 }
 566
 567 void __bch2_btree_path_downgrade(struct btree_trans *trans,
 568                                  struct btree_path *path,
 569                                  unsigned new_locks_want)
 570 {
 571         unsigned l;
 572
 573         EBUG_ON(path->locks_want < new_locks_want);
 574
 575         path->locks_want = new_locks_want;
 576
 577         while (path->nodes_locked &&
 578                (l = btree_path_highest_level_locked(path)) >= path->locks_want) {
 579                 if (l > path->level) {
 580                         btree_node_unlock(trans, path, l);
 581                 } else {
 582                         if (btree_node_intent_locked(path, l)) {
 583                                 six_lock_downgrade(&path->l[l].b->c.lock);
 584                                 mark_btree_node_locked_noreset(path, l, SIX_LOCK_read);
 585                         }
 586                         break;
 587                 }
 588         }
 589
 590         bch2_btree_path_verify_locks(path);
 591 }
 592
 593 /* Btree transaction locking: */
 594
 595 void bch2_trans_downgrade(struct btree_trans *trans)
 596 {
 597         struct btree_path *path;
 598
 599         trans_for_each_path(trans, path)
 600                 bch2_btree_path_downgrade(trans, path);
 601 }
 602
 603 int bch2_trans_relock(struct btree_trans *trans)
 604 {
 605         struct btree_path *path;
 606
 607         if (unlikely(trans->restarted))
 608                 return - ((int) trans->restarted);
 609
 610         trans_for_each_path(trans, path)
 611                 if (path->should_be_locked &&
 612                     !bch2_btree_path_relock_norestart(trans, path, _RET_IP_)) {
 613                         trace_and_count(trans->c, trans_restart_relock, trans, _RET_IP_, path);
 614                         return btree_trans_restart(trans, BCH_ERR_transaction_restart_relock);
 615                 }
 616         return 0;
 617 }
 618
 619 void bch2_trans_unlock(struct btree_trans *trans)
 620 {
 621         struct btree_path *path;
 622
 623         trans_for_each_path(trans, path)
 624                 __bch2_btree_path_unlock(trans, path);
 625
 626         /*
 627          * bch2_gc_btree_init_recurse() doesn't use btree iterators for walking
 628          * btree nodes, it implements its own walking:
 629          */
 630         EBUG_ON(!trans->is_initial_gc &&
 631                 lock_class_is_held(&bch2_btree_node_lock_key));
 632 }
 633
 634 bool bch2_trans_locked(struct btree_trans *trans)
 635 {
 636         struct btree_path *path;
 637
 638         trans_for_each_path(trans, path)
 639                 if (path->nodes_locked)
 640                         return true;
 641         return false;
 642 }
 643
 644 /* Debug */
 645
 646 #ifdef CONFIG_BCACHEFS_DEBUG
 647
 648 void bch2_btree_path_verify_locks(struct btree_path *path)
 649 {
 650         unsigned l;
 651
 652         if (!path->nodes_locked) {
 653                 BUG_ON(path->uptodate == BTREE_ITER_UPTODATE &&
 654                        btree_path_node(path, path->level));
 655                 return;
 656         }
 657
 658         for (l = 0; l < BTREE_MAX_DEPTH; l++) {
 659                 int want = btree_lock_want(path, l);
 660                 int have = btree_node_locked_type(path, l);
 661
 662                 BUG_ON(!is_btree_node(path, l) && have != BTREE_NODE_UNLOCKED);
 663
 664                 BUG_ON(is_btree_node(path, l) &&
 665                        (want == BTREE_NODE_UNLOCKED ||
 666                         have != BTREE_NODE_WRITE_LOCKED) &&
 667                        want != have);
 668         }
 669 }
 670
 671 void bch2_trans_verify_locks(struct btree_trans *trans)
 672 {
 673         struct btree_path *path;
 674
 675         trans_for_each_path(trans, path)
 676                 bch2_btree_path_verify_locks(path);
 677 }
 678
 679 #endif