kernel/rcu/tree_stall.h

   1 // SPDX-License-Identifier: GPL-2.0+
   2 /*
   3  * RCU CPU stall warnings for normal RCU grace periods
   4  *
   5  * Copyright IBM Corporation, 2019
   6  *
   7  * Author: Paul E. McKenney <paulmck@linux.ibm.com>
   8  */
   9
  10 //////////////////////////////////////////////////////////////////////////////
  11 //
  12 // Controlling CPU stall warnings, including delay calculation.
  13
  14 /* panic() on RCU Stall sysctl. */
  15 int sysctl_panic_on_rcu_stall __read_mostly;
  16 int sysctl_max_rcu_stall_to_panic __read_mostly;
  17
  18 #ifdef CONFIG_PROVE_RCU
  19 #define RCU_STALL_DELAY_DELTA           (5 * HZ)
  20 #else
  21 #define RCU_STALL_DELAY_DELTA           0
  22 #endif
  23 #define RCU_STALL_MIGHT_DIV             8
  24 #define RCU_STALL_MIGHT_MIN             (2 * HZ)
  25
  26 /* Limit-check stall timeouts specified at boottime and runtime. */
  27 int rcu_jiffies_till_stall_check(void)
  28 {
  29         int till_stall_check = READ_ONCE(rcu_cpu_stall_timeout);
  30
  31         /*
  32          * Limit check must be consistent with the Kconfig limits
  33          * for CONFIG_RCU_CPU_STALL_TIMEOUT.
  34          */
  35         if (till_stall_check < 3) {
  36                 WRITE_ONCE(rcu_cpu_stall_timeout, 3);
  37                 till_stall_check = 3;
  38         } else if (till_stall_check > 300) {
  39                 WRITE_ONCE(rcu_cpu_stall_timeout, 300);
  40                 till_stall_check = 300;
  41         }
  42         return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
  43 }
  44 EXPORT_SYMBOL_GPL(rcu_jiffies_till_stall_check);
  45
  46 /**
  47  * rcu_gp_might_be_stalled - Is it likely that the grace period is stalled?
  48  *
  49  * Returns @true if the current grace period is sufficiently old that
  50  * it is reasonable to assume that it might be stalled.  This can be
  51  * useful when deciding whether to allocate memory to enable RCU-mediated
  52  * freeing on the one hand or just invoking synchronize_rcu() on the other.
  53  * The latter is preferable when the grace period is stalled.
  54  *
  55  * Note that sampling of the .gp_start and .gp_seq fields must be done
  56  * carefully to avoid false positives at the beginnings and ends of
  57  * grace periods.
  58  */
  59 bool rcu_gp_might_be_stalled(void)
  60 {
  61         unsigned long d = rcu_jiffies_till_stall_check() / RCU_STALL_MIGHT_DIV;
  62         unsigned long j = jiffies;
  63
  64         if (d < RCU_STALL_MIGHT_MIN)
  65                 d = RCU_STALL_MIGHT_MIN;
  66         smp_mb(); // jiffies before .gp_seq to avoid false positives.
  67         if (!rcu_gp_in_progress())
  68                 return false;
  69         // Long delays at this point avoids false positive, but a delay
  70         // of ULONG_MAX/4 jiffies voids your no-false-positive warranty.
  71         smp_mb(); // .gp_seq before second .gp_start
  72         // And ditto here.
  73         return !time_before(j, READ_ONCE(rcu_state.gp_start) + d);
  74 }
  75
  76 /* Don't do RCU CPU stall warnings during long sysrq printouts. */
  77 void rcu_sysrq_start(void)
  78 {
  79         if (!rcu_cpu_stall_suppress)
  80                 rcu_cpu_stall_suppress = 2;
  81 }
  82
  83 void rcu_sysrq_end(void)
  84 {
  85         if (rcu_cpu_stall_suppress == 2)
  86                 rcu_cpu_stall_suppress = 0;
  87 }
  88
  89 /* Don't print RCU CPU stall warnings during a kernel panic. */
  90 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
  91 {
  92         rcu_cpu_stall_suppress = 1;
  93         return NOTIFY_DONE;
  94 }
  95
  96 static struct notifier_block rcu_panic_block = {
  97         .notifier_call = rcu_panic,
  98 };
  99
 100 static int __init check_cpu_stall_init(void)
 101 {
 102         atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
 103         return 0;
 104 }
 105 early_initcall(check_cpu_stall_init);
 106
 107 /* If so specified via sysctl, panic, yielding cleaner stall-warning output. */
 108 static void panic_on_rcu_stall(void)
 109 {
 110         static int cpu_stall;
 111
 112         if (++cpu_stall < sysctl_max_rcu_stall_to_panic)
 113                 return;
 114
 115         if (sysctl_panic_on_rcu_stall)
 116                 panic("RCU Stall\n");
 117 }
 118
 119 /**
 120  * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
 121  *
 122  * Set the stall-warning timeout way off into the future, thus preventing
 123  * any RCU CPU stall-warning messages from appearing in the current set of
 124  * RCU grace periods.
 125  *
 126  * The caller must disable hard irqs.
 127  */
 128 void rcu_cpu_stall_reset(void)
 129 {
 130         WRITE_ONCE(rcu_state.jiffies_stall, jiffies + ULONG_MAX / 2);
 131 }
 132
 133 //////////////////////////////////////////////////////////////////////////////
 134 //
 135 // Interaction with RCU grace periods
 136
 137 /* Start of new grace period, so record stall time (and forcing times). */
 138 static void record_gp_stall_check_time(void)
 139 {
 140         unsigned long j = jiffies;
 141         unsigned long j1;
 142
 143         WRITE_ONCE(rcu_state.gp_start, j);
 144         j1 = rcu_jiffies_till_stall_check();
 145         smp_mb(); // ->gp_start before ->jiffies_stall and caller's ->gp_seq.
 146         WRITE_ONCE(rcu_state.jiffies_stall, j + j1);
 147         rcu_state.jiffies_resched = j + j1 / 2;
 148         rcu_state.n_force_qs_gpstart = READ_ONCE(rcu_state.n_force_qs);
 149 }
 150
 151 /* Zero ->ticks_this_gp and snapshot the number of RCU softirq handlers. */
 152 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
 153 {
 154         rdp->ticks_this_gp = 0;
 155         rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
 156         WRITE_ONCE(rdp->last_fqs_resched, jiffies);
 157 }
 158
 159 /*
 160  * If too much time has passed in the current grace period, and if
 161  * so configured, go kick the relevant kthreads.
 162  */
 163 static void rcu_stall_kick_kthreads(void)
 164 {
 165         unsigned long j;
 166
 167         if (!READ_ONCE(rcu_kick_kthreads))
 168                 return;
 169         j = READ_ONCE(rcu_state.jiffies_kick_kthreads);
 170         if (time_after(jiffies, j) && rcu_state.gp_kthread &&
 171             (rcu_gp_in_progress() || READ_ONCE(rcu_state.gp_flags))) {
 172                 WARN_ONCE(1, "Kicking %s grace-period kthread\n",
 173                           rcu_state.name);
 174                 rcu_ftrace_dump(DUMP_ALL);
 175                 wake_up_process(rcu_state.gp_kthread);
 176                 WRITE_ONCE(rcu_state.jiffies_kick_kthreads, j + HZ);
 177         }
 178 }
 179
 180 /*
 181  * Handler for the irq_work request posted about halfway into the RCU CPU
 182  * stall timeout, and used to detect excessive irq disabling.  Set state
 183  * appropriately, but just complain if there is unexpected state on entry.
 184  */
 185 static void rcu_iw_handler(struct irq_work *iwp)
 186 {
 187         struct rcu_data *rdp;
 188         struct rcu_node *rnp;
 189
 190         rdp = container_of(iwp, struct rcu_data, rcu_iw);
 191         rnp = rdp->mynode;
 192         raw_spin_lock_rcu_node(rnp);
 193         if (!WARN_ON_ONCE(!rdp->rcu_iw_pending)) {
 194                 rdp->rcu_iw_gp_seq = rnp->gp_seq;
 195                 rdp->rcu_iw_pending = false;
 196         }
 197         raw_spin_unlock_rcu_node(rnp);
 198 }
 199
 200 //////////////////////////////////////////////////////////////////////////////
 201 //
 202 // Printing RCU CPU stall warnings
 203
 204 #ifdef CONFIG_PREEMPT_RCU
 205
 206 /*
 207  * Dump detailed information for all tasks blocking the current RCU
 208  * grace period on the specified rcu_node structure.
 209  */
 210 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
 211 {
 212         unsigned long flags;
 213         struct task_struct *t;
 214
 215         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 216         if (!rcu_preempt_blocked_readers_cgp(rnp)) {
 217                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 218                 return;
 219         }
 220         t = list_entry(rnp->gp_tasks->prev,
 221                        struct task_struct, rcu_node_entry);
 222         list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
 223                 /*
 224                  * We could be printing a lot while holding a spinlock.
 225                  * Avoid triggering hard lockup.
 226                  */
 227                 touch_nmi_watchdog();
 228                 sched_show_task(t);
 229         }
 230         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 231 }
 232
 233 // Communicate task state back to the RCU CPU stall warning request.
 234 struct rcu_stall_chk_rdr {
 235         int nesting;
 236         union rcu_special rs;
 237         bool on_blkd_list;
 238 };
 239
 240 /*
 241  * Report out the state of a not-running task that is stalling the
 242  * current RCU grace period.
 243  */
 244 static bool check_slow_task(struct task_struct *t, void *arg)
 245 {
 246         struct rcu_stall_chk_rdr *rscrp = arg;
 247
 248         if (task_curr(t))
 249                 return false; // It is running, so decline to inspect it.
 250         rscrp->nesting = t->rcu_read_lock_nesting;
 251         rscrp->rs = t->rcu_read_unlock_special;
 252         rscrp->on_blkd_list = !list_empty(&t->rcu_node_entry);
 253         return true;
 254 }
 255
 256 /*
 257  * Scan the current list of tasks blocked within RCU read-side critical
 258  * sections, printing out the tid of each of the first few of them.
 259  */
 260 static int rcu_print_task_stall(struct rcu_node *rnp, unsigned long flags)
 261         __releases(rnp->lock)
 262 {
 263         int i = 0;
 264         int ndetected = 0;
 265         struct rcu_stall_chk_rdr rscr;
 266         struct task_struct *t;
 267         struct task_struct *ts[8];
 268
 269         lockdep_assert_irqs_disabled();
 270         if (!rcu_preempt_blocked_readers_cgp(rnp))
 271                 return 0;
 272         pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
 273                rnp->level, rnp->grplo, rnp->grphi);
 274         t = list_entry(rnp->gp_tasks->prev,
 275                        struct task_struct, rcu_node_entry);
 276         list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
 277                 get_task_struct(t);
 278                 ts[i++] = t;
 279                 if (i >= ARRAY_SIZE(ts))
 280                         break;
 281         }
 282         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 283         for (i--; i; i--) {
 284                 t = ts[i];
 285                 if (!try_invoke_on_locked_down_task(t, check_slow_task, &rscr))
 286                         pr_cont(" P%d", t->pid);
 287                 else
 288                         pr_cont(" P%d/%d:%c%c%c%c",
 289                                 t->pid, rscr.nesting,
 290                                 ".b"[rscr.rs.b.blocked],
 291                                 ".q"[rscr.rs.b.need_qs],
 292                                 ".e"[rscr.rs.b.exp_hint],
 293                                 ".l"[rscr.on_blkd_list]);
 294                 lockdep_assert_irqs_disabled();
 295                 put_task_struct(t);
 296                 ndetected++;
 297         }
 298         pr_cont("\n");
 299         return ndetected;
 300 }
 301
 302 #else /* #ifdef CONFIG_PREEMPT_RCU */
 303
 304 /*
 305  * Because preemptible RCU does not exist, we never have to check for
 306  * tasks blocked within RCU read-side critical sections.
 307  */
 308 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
 309 {
 310 }
 311
 312 /*
 313  * Because preemptible RCU does not exist, we never have to check for
 314  * tasks blocked within RCU read-side critical sections.
 315  */
 316 static int rcu_print_task_stall(struct rcu_node *rnp, unsigned long flags)
 317 {
 318         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 319         return 0;
 320 }
 321 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
 322
 323 /*
 324  * Dump stacks of all tasks running on stalled CPUs.  First try using
 325  * NMIs, but fall back to manual remote stack tracing on architectures
 326  * that don't support NMI-based stack dumps.  The NMI-triggered stack
 327  * traces are more accurate because they are printed by the target CPU.
 328  */
 329 static void rcu_dump_cpu_stacks(void)
 330 {
 331         int cpu;
 332         unsigned long flags;
 333         struct rcu_node *rnp;
 334
 335         rcu_for_each_leaf_node(rnp) {
 336                 raw_spin_lock_irqsave_rcu_node(rnp, flags);
 337                 for_each_leaf_node_possible_cpu(rnp, cpu)
 338                         if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
 339                                 if (cpu_is_offline(cpu))
 340                                         pr_err("Offline CPU %d blocking current GP.\n", cpu);
 341                                 else if (!trigger_single_cpu_backtrace(cpu))
 342                                         dump_cpu_task(cpu);
 343                         }
 344                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 345         }
 346 }
 347
 348 #ifdef CONFIG_RCU_FAST_NO_HZ
 349
 350 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
 351 {
 352         struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
 353
 354         sprintf(cp, "last_accelerate: %04lx/%04lx dyntick_enabled: %d",
 355                 rdp->last_accelerate & 0xffff, jiffies & 0xffff,
 356                 !!rdp->tick_nohz_enabled_snap);
 357 }
 358
 359 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
 360
 361 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
 362 {
 363         *cp = '\0';
 364 }
 365
 366 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
 367
 368 static const char * const gp_state_names[] = {
 369         [RCU_GP_IDLE] = "RCU_GP_IDLE",
 370         [RCU_GP_WAIT_GPS] = "RCU_GP_WAIT_GPS",
 371         [RCU_GP_DONE_GPS] = "RCU_GP_DONE_GPS",
 372         [RCU_GP_ONOFF] = "RCU_GP_ONOFF",
 373         [RCU_GP_INIT] = "RCU_GP_INIT",
 374         [RCU_GP_WAIT_FQS] = "RCU_GP_WAIT_FQS",
 375         [RCU_GP_DOING_FQS] = "RCU_GP_DOING_FQS",
 376         [RCU_GP_CLEANUP] = "RCU_GP_CLEANUP",
 377         [RCU_GP_CLEANED] = "RCU_GP_CLEANED",
 378 };
 379
 380 /*
 381  * Convert a ->gp_state value to a character string.
 382  */
 383 static const char *gp_state_getname(short gs)
 384 {
 385         if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
 386                 return "???";
 387         return gp_state_names[gs];
 388 }
 389
 390 /* Is the RCU grace-period kthread being starved of CPU time? */
 391 static bool rcu_is_gp_kthread_starving(unsigned long *jp)
 392 {
 393         unsigned long j = jiffies - READ_ONCE(rcu_state.gp_activity);
 394
 395         if (jp)
 396                 *jp = j;
 397         return j > 2 * HZ;
 398 }
 399
 400 /*
 401  * Print out diagnostic information for the specified stalled CPU.
 402  *
 403  * If the specified CPU is aware of the current RCU grace period, then
 404  * print the number of scheduling clock interrupts the CPU has taken
 405  * during the time that it has been aware.  Otherwise, print the number
 406  * of RCU grace periods that this CPU is ignorant of, for example, "1"
 407  * if the CPU was aware of the previous grace period.
 408  *
 409  * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
 410  */
 411 static void print_cpu_stall_info(int cpu)
 412 {
 413         unsigned long delta;
 414         bool falsepositive;
 415         char fast_no_hz[72];
 416         struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
 417         char *ticks_title;
 418         unsigned long ticks_value;
 419
 420         /*
 421          * We could be printing a lot while holding a spinlock.  Avoid
 422          * triggering hard lockup.
 423          */
 424         touch_nmi_watchdog();
 425
 426         ticks_value = rcu_seq_ctr(rcu_state.gp_seq - rdp->gp_seq);
 427         if (ticks_value) {
 428                 ticks_title = "GPs behind";
 429         } else {
 430                 ticks_title = "ticks this GP";
 431                 ticks_value = rdp->ticks_this_gp;
 432         }
 433         print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
 434         delta = rcu_seq_ctr(rdp->mynode->gp_seq - rdp->rcu_iw_gp_seq);
 435         falsepositive = rcu_is_gp_kthread_starving(NULL) &&
 436                         rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp));
 437         pr_err("\t%d-%c%c%c%c: (%lu %s) idle=%03x/%ld/%#lx softirq=%u/%u fqs=%ld %s%s\n",
 438                cpu,
 439                "O."[!!cpu_online(cpu)],
 440                "o."[!!(rdp->grpmask & rdp->mynode->qsmaskinit)],
 441                "N."[!!(rdp->grpmask & rdp->mynode->qsmaskinitnext)],
 442                !IS_ENABLED(CONFIG_IRQ_WORK) ? '?' :
 443                         rdp->rcu_iw_pending ? (int)min(delta, 9UL) + '0' :
 444                                 "!."[!delta],
 445                ticks_value, ticks_title,
 446                rcu_dynticks_snap(rdp) & 0xfff,
 447                rdp->dynticks_nesting, rdp->dynticks_nmi_nesting,
 448                rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
 449                data_race(rcu_state.n_force_qs) - rcu_state.n_force_qs_gpstart,
 450                fast_no_hz,
 451                falsepositive ? " (false positive?)" : "");
 452 }
 453
 454 /* Complain about starvation of grace-period kthread.  */
 455 static void rcu_check_gp_kthread_starvation(void)
 456 {
 457         int cpu;
 458         struct task_struct *gpk = rcu_state.gp_kthread;
 459         unsigned long j;
 460
 461         if (rcu_is_gp_kthread_starving(&j)) {
 462                 cpu = gpk ? task_cpu(gpk) : -1;
 463                 pr_err("%s kthread starved for %ld jiffies! g%ld f%#x %s(%d) ->state=%#lx ->cpu=%d\n",
 464                        rcu_state.name, j,
 465                        (long)rcu_seq_current(&rcu_state.gp_seq),
 466                        data_race(rcu_state.gp_flags),
 467                        gp_state_getname(rcu_state.gp_state), rcu_state.gp_state,
 468                        gpk ? gpk->state : ~0, cpu);
 469                 if (gpk) {
 470                         pr_err("\tUnless %s kthread gets sufficient CPU time, OOM is now expected behavior.\n", rcu_state.name);
 471                         pr_err("RCU grace-period kthread stack dump:\n");
 472                         sched_show_task(gpk);
 473                         if (cpu >= 0) {
 474                                 if (cpu_is_offline(cpu)) {
 475                                         pr_err("RCU GP kthread last ran on offline CPU %d.\n", cpu);
 476                                 } else  {
 477                                         pr_err("Stack dump where RCU GP kthread last ran:\n");
 478                                         if (!trigger_single_cpu_backtrace(cpu))
 479                                                 dump_cpu_task(cpu);
 480                                 }
 481                         }
 482                         wake_up_process(gpk);
 483                 }
 484         }
 485 }
 486
 487 /* Complain about missing wakeups from expired fqs wait timer */
 488 static void rcu_check_gp_kthread_expired_fqs_timer(void)
 489 {
 490         struct task_struct *gpk = rcu_state.gp_kthread;
 491         short gp_state;
 492         unsigned long jiffies_fqs;
 493         int cpu;
 494
 495         /*
 496          * Order reads of .gp_state and .jiffies_force_qs.
 497          * Matching smp_wmb() is present in rcu_gp_fqs_loop().
 498          */
 499         gp_state = smp_load_acquire(&rcu_state.gp_state);
 500         jiffies_fqs = READ_ONCE(rcu_state.jiffies_force_qs);
 501
 502         if (gp_state == RCU_GP_WAIT_FQS &&
 503             time_after(jiffies, jiffies_fqs + RCU_STALL_MIGHT_MIN) &&
 504             gpk && !READ_ONCE(gpk->on_rq)) {
 505                 cpu = task_cpu(gpk);
 506                 pr_err("%s kthread timer wakeup didn't happen for %ld jiffies! g%ld f%#x %s(%d) ->state=%#lx\n",
 507                        rcu_state.name, (jiffies - jiffies_fqs),
 508                        (long)rcu_seq_current(&rcu_state.gp_seq),
 509                        data_race(rcu_state.gp_flags),
 510                        gp_state_getname(RCU_GP_WAIT_FQS), RCU_GP_WAIT_FQS,
 511                        gpk->state);
 512                 pr_err("\tPossible timer handling issue on cpu=%d timer-softirq=%u\n",
 513                        cpu, kstat_softirqs_cpu(TIMER_SOFTIRQ, cpu));
 514         }
 515 }
 516
 517 static void print_other_cpu_stall(unsigned long gp_seq, unsigned long gps)
 518 {
 519         int cpu;
 520         unsigned long flags;
 521         unsigned long gpa;
 522         unsigned long j;
 523         int ndetected = 0;
 524         struct rcu_node *rnp;
 525         long totqlen = 0;
 526
 527         lockdep_assert_irqs_disabled();
 528
 529         /* Kick and suppress, if so configured. */
 530         rcu_stall_kick_kthreads();
 531         if (rcu_stall_is_suppressed())
 532                 return;
 533
 534         /*
 535          * OK, time to rat on our buddy...
 536          * See Documentation/RCU/stallwarn.rst for info on how to debug
 537          * RCU CPU stall warnings.
 538          */
 539         trace_rcu_stall_warning(rcu_state.name, TPS("StallDetected"));
 540         pr_err("INFO: %s detected stalls on CPUs/tasks:\n", rcu_state.name);
 541         rcu_for_each_leaf_node(rnp) {
 542                 raw_spin_lock_irqsave_rcu_node(rnp, flags);
 543                 if (rnp->qsmask != 0) {
 544                         for_each_leaf_node_possible_cpu(rnp, cpu)
 545                                 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
 546                                         print_cpu_stall_info(cpu);
 547                                         ndetected++;
 548                                 }
 549                 }
 550                 ndetected += rcu_print_task_stall(rnp, flags); // Releases rnp->lock.
 551                 lockdep_assert_irqs_disabled();
 552         }
 553
 554         for_each_possible_cpu(cpu)
 555                 totqlen += rcu_get_n_cbs_cpu(cpu);
 556         pr_cont("\t(detected by %d, t=%ld jiffies, g=%ld, q=%lu)\n",
 557                smp_processor_id(), (long)(jiffies - gps),
 558                (long)rcu_seq_current(&rcu_state.gp_seq), totqlen);
 559         if (ndetected) {
 560                 rcu_dump_cpu_stacks();
 561
 562                 /* Complain about tasks blocking the grace period. */
 563                 rcu_for_each_leaf_node(rnp)
 564                         rcu_print_detail_task_stall_rnp(rnp);
 565         } else {
 566                 if (rcu_seq_current(&rcu_state.gp_seq) != gp_seq) {
 567                         pr_err("INFO: Stall ended before state dump start\n");
 568                 } else {
 569                         j = jiffies;
 570                         gpa = data_race(rcu_state.gp_activity);
 571                         pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
 572                                rcu_state.name, j - gpa, j, gpa,
 573                                data_race(jiffies_till_next_fqs),
 574                                rcu_get_root()->qsmask);
 575                 }
 576         }
 577         /* Rewrite if needed in case of slow consoles. */
 578         if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
 579                 WRITE_ONCE(rcu_state.jiffies_stall,
 580                            jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
 581
 582         rcu_check_gp_kthread_expired_fqs_timer();
 583         rcu_check_gp_kthread_starvation();
 584
 585         panic_on_rcu_stall();
 586
 587         rcu_force_quiescent_state();  /* Kick them all. */
 588 }
 589
 590 static void print_cpu_stall(unsigned long gps)
 591 {
 592         int cpu;
 593         unsigned long flags;
 594         struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 595         struct rcu_node *rnp = rcu_get_root();
 596         long totqlen = 0;
 597
 598         lockdep_assert_irqs_disabled();
 599
 600         /* Kick and suppress, if so configured. */
 601         rcu_stall_kick_kthreads();
 602         if (rcu_stall_is_suppressed())
 603                 return;
 604
 605         /*
 606          * OK, time to rat on ourselves...
 607          * See Documentation/RCU/stallwarn.rst for info on how to debug
 608          * RCU CPU stall warnings.
 609          */
 610         trace_rcu_stall_warning(rcu_state.name, TPS("SelfDetected"));
 611         pr_err("INFO: %s self-detected stall on CPU\n", rcu_state.name);
 612         raw_spin_lock_irqsave_rcu_node(rdp->mynode, flags);
 613         print_cpu_stall_info(smp_processor_id());
 614         raw_spin_unlock_irqrestore_rcu_node(rdp->mynode, flags);
 615         for_each_possible_cpu(cpu)
 616                 totqlen += rcu_get_n_cbs_cpu(cpu);
 617         pr_cont("\t(t=%lu jiffies g=%ld q=%lu)\n",
 618                 jiffies - gps,
 619                 (long)rcu_seq_current(&rcu_state.gp_seq), totqlen);
 620
 621         rcu_check_gp_kthread_expired_fqs_timer();
 622         rcu_check_gp_kthread_starvation();
 623
 624         rcu_dump_cpu_stacks();
 625
 626         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 627         /* Rewrite if needed in case of slow consoles. */
 628         if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
 629                 WRITE_ONCE(rcu_state.jiffies_stall,
 630                            jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
 631         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 632
 633         panic_on_rcu_stall();
 634
 635         /*
 636          * Attempt to revive the RCU machinery by forcing a context switch.
 637          *
 638          * A context switch would normally allow the RCU state machine to make
 639          * progress and it could be we're stuck in kernel space without context
 640          * switches for an entirely unreasonable amount of time.
 641          */
 642         set_tsk_need_resched(current);
 643         set_preempt_need_resched();
 644 }
 645
 646 static void check_cpu_stall(struct rcu_data *rdp)
 647 {
 648         unsigned long gs1;
 649         unsigned long gs2;
 650         unsigned long gps;
 651         unsigned long j;
 652         unsigned long jn;
 653         unsigned long js;
 654         struct rcu_node *rnp;
 655
 656         lockdep_assert_irqs_disabled();
 657         if ((rcu_stall_is_suppressed() && !READ_ONCE(rcu_kick_kthreads)) ||
 658             !rcu_gp_in_progress())
 659                 return;
 660         rcu_stall_kick_kthreads();
 661         j = jiffies;
 662
 663         /*
 664          * Lots of memory barriers to reject false positives.
 665          *
 666          * The idea is to pick up rcu_state.gp_seq, then
 667          * rcu_state.jiffies_stall, then rcu_state.gp_start, and finally
 668          * another copy of rcu_state.gp_seq.  These values are updated in
 669          * the opposite order with memory barriers (or equivalent) during
 670          * grace-period initialization and cleanup.  Now, a false positive
 671          * can occur if we get an new value of rcu_state.gp_start and a old
 672          * value of rcu_state.jiffies_stall.  But given the memory barriers,
 673          * the only way that this can happen is if one grace period ends
 674          * and another starts between these two fetches.  This is detected
 675          * by comparing the second fetch of rcu_state.gp_seq with the
 676          * previous fetch from rcu_state.gp_seq.
 677          *
 678          * Given this check, comparisons of jiffies, rcu_state.jiffies_stall,
 679          * and rcu_state.gp_start suffice to forestall false positives.
 680          */
 681         gs1 = READ_ONCE(rcu_state.gp_seq);
 682         smp_rmb(); /* Pick up ->gp_seq first... */
 683         js = READ_ONCE(rcu_state.jiffies_stall);
 684         smp_rmb(); /* ...then ->jiffies_stall before the rest... */
 685         gps = READ_ONCE(rcu_state.gp_start);
 686         smp_rmb(); /* ...and finally ->gp_start before ->gp_seq again. */
 687         gs2 = READ_ONCE(rcu_state.gp_seq);
 688         if (gs1 != gs2 ||
 689             ULONG_CMP_LT(j, js) ||
 690             ULONG_CMP_GE(gps, js))
 691                 return; /* No stall or GP completed since entering function. */
 692         rnp = rdp->mynode;
 693         jn = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
 694         if (rcu_gp_in_progress() &&
 695             (READ_ONCE(rnp->qsmask) & rdp->grpmask) &&
 696             cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
 697
 698                 /* We haven't checked in, so go dump stack. */
 699                 print_cpu_stall(gps);
 700                 if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
 701                         rcu_ftrace_dump(DUMP_ALL);
 702
 703         } else if (rcu_gp_in_progress() &&
 704                    ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY) &&
 705                    cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
 706
 707                 /* They had a few time units to dump stack, so complain. */
 708                 print_other_cpu_stall(gs2, gps);
 709                 if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
 710                         rcu_ftrace_dump(DUMP_ALL);
 711         }
 712 }
 713
 714 //////////////////////////////////////////////////////////////////////////////
 715 //
 716 // RCU forward-progress mechanisms, including of callback invocation.
 717
 718
 719 /*
 720  * Show the state of the grace-period kthreads.
 721  */
 722 void show_rcu_gp_kthreads(void)
 723 {
 724         unsigned long cbs = 0;
 725         int cpu;
 726         unsigned long j;
 727         unsigned long ja;
 728         unsigned long jr;
 729         unsigned long jw;
 730         struct rcu_data *rdp;
 731         struct rcu_node *rnp;
 732         struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
 733
 734         j = jiffies;
 735         ja = j - data_race(rcu_state.gp_activity);
 736         jr = j - data_race(rcu_state.gp_req_activity);
 737         jw = j - data_race(rcu_state.gp_wake_time);
 738         pr_info("%s: wait state: %s(%d) ->state: %#lx delta ->gp_activity %lu ->gp_req_activity %lu ->gp_wake_time %lu ->gp_wake_seq %ld ->gp_seq %ld ->gp_seq_needed %ld ->gp_flags %#x\n",
 739                 rcu_state.name, gp_state_getname(rcu_state.gp_state),
 740                 rcu_state.gp_state, t ? t->state : 0x1ffffL,
 741                 ja, jr, jw, (long)data_race(rcu_state.gp_wake_seq),
 742                 (long)data_race(rcu_state.gp_seq),
 743                 (long)data_race(rcu_get_root()->gp_seq_needed),
 744                 data_race(rcu_state.gp_flags));
 745         rcu_for_each_node_breadth_first(rnp) {
 746                 if (ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq),
 747                                  READ_ONCE(rnp->gp_seq_needed)))
 748                         continue;
 749                 pr_info("\trcu_node %d:%d ->gp_seq %ld ->gp_seq_needed %ld\n",
 750                         rnp->grplo, rnp->grphi, (long)data_race(rnp->gp_seq),
 751                         (long)data_race(rnp->gp_seq_needed));
 752                 if (!rcu_is_leaf_node(rnp))
 753                         continue;
 754                 for_each_leaf_node_possible_cpu(rnp, cpu) {
 755                         rdp = per_cpu_ptr(&rcu_data, cpu);
 756                         if (READ_ONCE(rdp->gpwrap) ||
 757                             ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq),
 758                                          READ_ONCE(rdp->gp_seq_needed)))
 759                                 continue;
 760                         pr_info("\tcpu %d ->gp_seq_needed %ld\n",
 761                                 cpu, (long)data_race(rdp->gp_seq_needed));
 762                 }
 763         }
 764         for_each_possible_cpu(cpu) {
 765                 rdp = per_cpu_ptr(&rcu_data, cpu);
 766                 cbs += data_race(rdp->n_cbs_invoked);
 767                 if (rcu_segcblist_is_offloaded(&rdp->cblist))
 768                         show_rcu_nocb_state(rdp);
 769         }
 770         pr_info("RCU callbacks invoked since boot: %lu\n", cbs);
 771         show_rcu_tasks_gp_kthreads();
 772 }
 773 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
 774
 775 /*
 776  * This function checks for grace-period requests that fail to motivate
 777  * RCU to come out of its idle mode.
 778  */
 779 static void rcu_check_gp_start_stall(struct rcu_node *rnp, struct rcu_data *rdp,
 780                                      const unsigned long gpssdelay)
 781 {
 782         unsigned long flags;
 783         unsigned long j;
 784         struct rcu_node *rnp_root = rcu_get_root();
 785         static atomic_t warned = ATOMIC_INIT(0);
 786
 787         if (!IS_ENABLED(CONFIG_PROVE_RCU) || rcu_gp_in_progress() ||
 788             ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
 789                          READ_ONCE(rnp_root->gp_seq_needed)) ||
 790             !smp_load_acquire(&rcu_state.gp_kthread)) // Get stable kthread.
 791                 return;
 792         j = jiffies; /* Expensive access, and in common case don't get here. */
 793         if (time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
 794             time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
 795             atomic_read(&warned))
 796                 return;
 797
 798         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 799         j = jiffies;
 800         if (rcu_gp_in_progress() ||
 801             ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
 802                          READ_ONCE(rnp_root->gp_seq_needed)) ||
 803             time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
 804             time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
 805             atomic_read(&warned)) {
 806                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 807                 return;
 808         }
 809         /* Hold onto the leaf lock to make others see warned==1. */
 810
 811         if (rnp_root != rnp)
 812                 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
 813         j = jiffies;
 814         if (rcu_gp_in_progress() ||
 815             ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
 816                          READ_ONCE(rnp_root->gp_seq_needed)) ||
 817             time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
 818             time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
 819             atomic_xchg(&warned, 1)) {
 820                 if (rnp_root != rnp)
 821                         /* irqs remain disabled. */
 822                         raw_spin_unlock_rcu_node(rnp_root);
 823                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 824                 return;
 825         }
 826         WARN_ON(1);
 827         if (rnp_root != rnp)
 828                 raw_spin_unlock_rcu_node(rnp_root);
 829         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 830         show_rcu_gp_kthreads();
 831 }
 832
 833 /*
 834  * Do a forward-progress check for rcutorture.  This is normally invoked
 835  * due to an OOM event.  The argument "j" gives the time period during
 836  * which rcutorture would like progress to have been made.
 837  */
 838 void rcu_fwd_progress_check(unsigned long j)
 839 {
 840         unsigned long cbs;
 841         int cpu;
 842         unsigned long max_cbs = 0;
 843         int max_cpu = -1;
 844         struct rcu_data *rdp;
 845
 846         if (rcu_gp_in_progress()) {
 847                 pr_info("%s: GP age %lu jiffies\n",
 848                         __func__, jiffies - rcu_state.gp_start);
 849                 show_rcu_gp_kthreads();
 850         } else {
 851                 pr_info("%s: Last GP end %lu jiffies ago\n",
 852                         __func__, jiffies - rcu_state.gp_end);
 853                 preempt_disable();
 854                 rdp = this_cpu_ptr(&rcu_data);
 855                 rcu_check_gp_start_stall(rdp->mynode, rdp, j);
 856                 preempt_enable();
 857         }
 858         for_each_possible_cpu(cpu) {
 859                 cbs = rcu_get_n_cbs_cpu(cpu);
 860                 if (!cbs)
 861                         continue;
 862                 if (max_cpu < 0)
 863                         pr_info("%s: callbacks", __func__);
 864                 pr_cont(" %d: %lu", cpu, cbs);
 865                 if (cbs <= max_cbs)
 866                         continue;
 867                 max_cbs = cbs;
 868                 max_cpu = cpu;
 869         }
 870         if (max_cpu >= 0)
 871                 pr_cont("\n");
 872 }
 873 EXPORT_SYMBOL_GPL(rcu_fwd_progress_check);
 874
 875 /* Commandeer a sysrq key to dump RCU's tree. */
 876 static bool sysrq_rcu;
 877 module_param(sysrq_rcu, bool, 0444);
 878
 879 /* Dump grace-period-request information due to commandeered sysrq. */
 880 static void sysrq_show_rcu(int key)
 881 {
 882         show_rcu_gp_kthreads();
 883 }
 884
 885 static const struct sysrq_key_op sysrq_rcudump_op = {
 886         .handler = sysrq_show_rcu,
 887         .help_msg = "show-rcu(y)",
 888         .action_msg = "Show RCU tree",
 889         .enable_mask = SYSRQ_ENABLE_DUMP,
 890 };
 891
 892 static int __init rcu_sysrq_init(void)
 893 {
 894         if (sysrq_rcu)
 895                 return register_sysrq_key('y', &sysrq_rcudump_op);
 896         return 0;
 897 }
 898 early_initcall(rcu_sysrq_init);