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         __releases(rnp->lock)
 318 {
 319         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 320         return 0;
 321 }
 322 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
 323
 324 /*
 325  * Dump stacks of all tasks running on stalled CPUs.  First try using
 326  * NMIs, but fall back to manual remote stack tracing on architectures
 327  * that don't support NMI-based stack dumps.  The NMI-triggered stack
 328  * traces are more accurate because they are printed by the target CPU.
 329  */
 330 static void rcu_dump_cpu_stacks(void)
 331 {
 332         int cpu;
 333         unsigned long flags;
 334         struct rcu_node *rnp;
 335
 336         rcu_for_each_leaf_node(rnp) {
 337                 raw_spin_lock_irqsave_rcu_node(rnp, flags);
 338                 for_each_leaf_node_possible_cpu(rnp, cpu)
 339                         if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
 340                                 if (cpu_is_offline(cpu))
 341                                         pr_err("Offline CPU %d blocking current GP.\n", cpu);
 342                                 else if (!trigger_single_cpu_backtrace(cpu))
 343                                         dump_cpu_task(cpu);
 344                         }
 345                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 346         }
 347 }
 348
 349 #ifdef CONFIG_RCU_FAST_NO_HZ
 350
 351 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
 352 {
 353         struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
 354
 355         sprintf(cp, "last_accelerate: %04lx/%04lx dyntick_enabled: %d",
 356                 rdp->last_accelerate & 0xffff, jiffies & 0xffff,
 357                 !!rdp->tick_nohz_enabled_snap);
 358 }
 359
 360 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
 361
 362 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
 363 {
 364         *cp = '\0';
 365 }
 366
 367 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
 368
 369 static const char * const gp_state_names[] = {
 370         [RCU_GP_IDLE] = "RCU_GP_IDLE",
 371         [RCU_GP_WAIT_GPS] = "RCU_GP_WAIT_GPS",
 372         [RCU_GP_DONE_GPS] = "RCU_GP_DONE_GPS",
 373         [RCU_GP_ONOFF] = "RCU_GP_ONOFF",
 374         [RCU_GP_INIT] = "RCU_GP_INIT",
 375         [RCU_GP_WAIT_FQS] = "RCU_GP_WAIT_FQS",
 376         [RCU_GP_DOING_FQS] = "RCU_GP_DOING_FQS",
 377         [RCU_GP_CLEANUP] = "RCU_GP_CLEANUP",
 378         [RCU_GP_CLEANED] = "RCU_GP_CLEANED",
 379 };
 380
 381 /*
 382  * Convert a ->gp_state value to a character string.
 383  */
 384 static const char *gp_state_getname(short gs)
 385 {
 386         if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
 387                 return "???";
 388         return gp_state_names[gs];
 389 }
 390
 391 /* Is the RCU grace-period kthread being starved of CPU time? */
 392 static bool rcu_is_gp_kthread_starving(unsigned long *jp)
 393 {
 394         unsigned long j = jiffies - READ_ONCE(rcu_state.gp_activity);
 395
 396         if (jp)
 397                 *jp = j;
 398         return j > 2 * HZ;
 399 }
 400
 401 /*
 402  * Print out diagnostic information for the specified stalled CPU.
 403  *
 404  * If the specified CPU is aware of the current RCU grace period, then
 405  * print the number of scheduling clock interrupts the CPU has taken
 406  * during the time that it has been aware.  Otherwise, print the number
 407  * of RCU grace periods that this CPU is ignorant of, for example, "1"
 408  * if the CPU was aware of the previous grace period.
 409  *
 410  * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
 411  */
 412 static void print_cpu_stall_info(int cpu)
 413 {
 414         unsigned long delta;
 415         bool falsepositive;
 416         char fast_no_hz[72];
 417         struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
 418         char *ticks_title;
 419         unsigned long ticks_value;
 420
 421         /*
 422          * We could be printing a lot while holding a spinlock.  Avoid
 423          * triggering hard lockup.
 424          */
 425         touch_nmi_watchdog();
 426
 427         ticks_value = rcu_seq_ctr(rcu_state.gp_seq - rdp->gp_seq);
 428         if (ticks_value) {
 429                 ticks_title = "GPs behind";
 430         } else {
 431                 ticks_title = "ticks this GP";
 432                 ticks_value = rdp->ticks_this_gp;
 433         }
 434         print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
 435         delta = rcu_seq_ctr(rdp->mynode->gp_seq - rdp->rcu_iw_gp_seq);
 436         falsepositive = rcu_is_gp_kthread_starving(NULL) &&
 437                         rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp));
 438         pr_err("\t%d-%c%c%c%c: (%lu %s) idle=%03x/%ld/%#lx softirq=%u/%u fqs=%ld %s%s\n",
 439                cpu,
 440                "O."[!!cpu_online(cpu)],
 441                "o."[!!(rdp->grpmask & rdp->mynode->qsmaskinit)],
 442                "N."[!!(rdp->grpmask & rdp->mynode->qsmaskinitnext)],
 443                !IS_ENABLED(CONFIG_IRQ_WORK) ? '?' :
 444                         rdp->rcu_iw_pending ? (int)min(delta, 9UL) + '0' :
 445                                 "!."[!delta],
 446                ticks_value, ticks_title,
 447                rcu_dynticks_snap(rdp) & 0xfff,
 448                rdp->dynticks_nesting, rdp->dynticks_nmi_nesting,
 449                rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
 450                data_race(rcu_state.n_force_qs) - rcu_state.n_force_qs_gpstart,
 451                fast_no_hz,
 452                falsepositive ? " (false positive?)" : "");
 453 }
 454
 455 /* Complain about starvation of grace-period kthread.  */
 456 static void rcu_check_gp_kthread_starvation(void)
 457 {
 458         int cpu;
 459         struct task_struct *gpk = rcu_state.gp_kthread;
 460         unsigned long j;
 461
 462         if (rcu_is_gp_kthread_starving(&j)) {
 463                 cpu = gpk ? task_cpu(gpk) : -1;
 464                 pr_err("%s kthread starved for %ld jiffies! g%ld f%#x %s(%d) ->state=%#x ->cpu=%d\n",
 465                        rcu_state.name, j,
 466                        (long)rcu_seq_current(&rcu_state.gp_seq),
 467                        data_race(rcu_state.gp_flags),
 468                        gp_state_getname(rcu_state.gp_state), rcu_state.gp_state,
 469                        gpk ? gpk->__state : ~0, cpu);
 470                 if (gpk) {
 471                         pr_err("\tUnless %s kthread gets sufficient CPU time, OOM is now expected behavior.\n", rcu_state.name);
 472                         pr_err("RCU grace-period kthread stack dump:\n");
 473                         sched_show_task(gpk);
 474                         if (cpu >= 0) {
 475                                 if (cpu_is_offline(cpu)) {
 476                                         pr_err("RCU GP kthread last ran on offline CPU %d.\n", cpu);
 477                                 } else  {
 478                                         pr_err("Stack dump where RCU GP kthread last ran:\n");
 479                                         if (!trigger_single_cpu_backtrace(cpu))
 480                                                 dump_cpu_task(cpu);
 481                                 }
 482                         }
 483                         wake_up_process(gpk);
 484                 }
 485         }
 486 }
 487
 488 /* Complain about missing wakeups from expired fqs wait timer */
 489 static void rcu_check_gp_kthread_expired_fqs_timer(void)
 490 {
 491         struct task_struct *gpk = rcu_state.gp_kthread;
 492         short gp_state;
 493         unsigned long jiffies_fqs;
 494         int cpu;
 495
 496         /*
 497          * Order reads of .gp_state and .jiffies_force_qs.
 498          * Matching smp_wmb() is present in rcu_gp_fqs_loop().
 499          */
 500         gp_state = smp_load_acquire(&rcu_state.gp_state);
 501         jiffies_fqs = READ_ONCE(rcu_state.jiffies_force_qs);
 502
 503         if (gp_state == RCU_GP_WAIT_FQS &&
 504             time_after(jiffies, jiffies_fqs + RCU_STALL_MIGHT_MIN) &&
 505             gpk && !READ_ONCE(gpk->on_rq)) {
 506                 cpu = task_cpu(gpk);
 507                 pr_err("%s kthread timer wakeup didn't happen for %ld jiffies! g%ld f%#x %s(%d) ->state=%#x\n",
 508                        rcu_state.name, (jiffies - jiffies_fqs),
 509                        (long)rcu_seq_current(&rcu_state.gp_seq),
 510                        data_race(rcu_state.gp_flags),
 511                        gp_state_getname(RCU_GP_WAIT_FQS), RCU_GP_WAIT_FQS,
 512                        gpk->__state);
 513                 pr_err("\tPossible timer handling issue on cpu=%d timer-softirq=%u\n",
 514                        cpu, kstat_softirqs_cpu(TIMER_SOFTIRQ, cpu));
 515         }
 516 }
 517
 518 static void print_other_cpu_stall(unsigned long gp_seq, unsigned long gps)
 519 {
 520         int cpu;
 521         unsigned long flags;
 522         unsigned long gpa;
 523         unsigned long j;
 524         int ndetected = 0;
 525         struct rcu_node *rnp;
 526         long totqlen = 0;
 527
 528         lockdep_assert_irqs_disabled();
 529
 530         /* Kick and suppress, if so configured. */
 531         rcu_stall_kick_kthreads();
 532         if (rcu_stall_is_suppressed())
 533                 return;
 534
 535         /*
 536          * OK, time to rat on our buddy...
 537          * See Documentation/RCU/stallwarn.rst for info on how to debug
 538          * RCU CPU stall warnings.
 539          */
 540         trace_rcu_stall_warning(rcu_state.name, TPS("StallDetected"));
 541         pr_err("INFO: %s detected stalls on CPUs/tasks:\n", rcu_state.name);
 542         rcu_for_each_leaf_node(rnp) {
 543                 raw_spin_lock_irqsave_rcu_node(rnp, flags);
 544                 if (rnp->qsmask != 0) {
 545                         for_each_leaf_node_possible_cpu(rnp, cpu)
 546                                 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
 547                                         print_cpu_stall_info(cpu);
 548                                         ndetected++;
 549                                 }
 550                 }
 551                 ndetected += rcu_print_task_stall(rnp, flags); // Releases rnp->lock.
 552                 lockdep_assert_irqs_disabled();
 553         }
 554
 555         for_each_possible_cpu(cpu)
 556                 totqlen += rcu_get_n_cbs_cpu(cpu);
 557         pr_cont("\t(detected by %d, t=%ld jiffies, g=%ld, q=%lu)\n",
 558                smp_processor_id(), (long)(jiffies - gps),
 559                (long)rcu_seq_current(&rcu_state.gp_seq), totqlen);
 560         if (ndetected) {
 561                 rcu_dump_cpu_stacks();
 562
 563                 /* Complain about tasks blocking the grace period. */
 564                 rcu_for_each_leaf_node(rnp)
 565                         rcu_print_detail_task_stall_rnp(rnp);
 566         } else {
 567                 if (rcu_seq_current(&rcu_state.gp_seq) != gp_seq) {
 568                         pr_err("INFO: Stall ended before state dump start\n");
 569                 } else {
 570                         j = jiffies;
 571                         gpa = data_race(rcu_state.gp_activity);
 572                         pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
 573                                rcu_state.name, j - gpa, j, gpa,
 574                                data_race(jiffies_till_next_fqs),
 575                                rcu_get_root()->qsmask);
 576                 }
 577         }
 578         /* Rewrite if needed in case of slow consoles. */
 579         if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
 580                 WRITE_ONCE(rcu_state.jiffies_stall,
 581                            jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
 582
 583         rcu_check_gp_kthread_expired_fqs_timer();
 584         rcu_check_gp_kthread_starvation();
 585
 586         panic_on_rcu_stall();
 587
 588         rcu_force_quiescent_state();  /* Kick them all. */
 589 }
 590
 591 static void print_cpu_stall(unsigned long gps)
 592 {
 593         int cpu;
 594         unsigned long flags;
 595         struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 596         struct rcu_node *rnp = rcu_get_root();
 597         long totqlen = 0;
 598
 599         lockdep_assert_irqs_disabled();
 600
 601         /* Kick and suppress, if so configured. */
 602         rcu_stall_kick_kthreads();
 603         if (rcu_stall_is_suppressed())
 604                 return;
 605
 606         /*
 607          * OK, time to rat on ourselves...
 608          * See Documentation/RCU/stallwarn.rst for info on how to debug
 609          * RCU CPU stall warnings.
 610          */
 611         trace_rcu_stall_warning(rcu_state.name, TPS("SelfDetected"));
 612         pr_err("INFO: %s self-detected stall on CPU\n", rcu_state.name);
 613         raw_spin_lock_irqsave_rcu_node(rdp->mynode, flags);
 614         print_cpu_stall_info(smp_processor_id());
 615         raw_spin_unlock_irqrestore_rcu_node(rdp->mynode, flags);
 616         for_each_possible_cpu(cpu)
 617                 totqlen += rcu_get_n_cbs_cpu(cpu);
 618         pr_cont("\t(t=%lu jiffies g=%ld q=%lu)\n",
 619                 jiffies - gps,
 620                 (long)rcu_seq_current(&rcu_state.gp_seq), totqlen);
 621
 622         rcu_check_gp_kthread_expired_fqs_timer();
 623         rcu_check_gp_kthread_starvation();
 624
 625         rcu_dump_cpu_stacks();
 626
 627         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 628         /* Rewrite if needed in case of slow consoles. */
 629         if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
 630                 WRITE_ONCE(rcu_state.jiffies_stall,
 631                            jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
 632         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 633
 634         panic_on_rcu_stall();
 635
 636         /*
 637          * Attempt to revive the RCU machinery by forcing a context switch.
 638          *
 639          * A context switch would normally allow the RCU state machine to make
 640          * progress and it could be we're stuck in kernel space without context
 641          * switches for an entirely unreasonable amount of time.
 642          */
 643         set_tsk_need_resched(current);
 644         set_preempt_need_resched();
 645 }
 646
 647 static void check_cpu_stall(struct rcu_data *rdp)
 648 {
 649         unsigned long gs1;
 650         unsigned long gs2;
 651         unsigned long gps;
 652         unsigned long j;
 653         unsigned long jn;
 654         unsigned long js;
 655         struct rcu_node *rnp;
 656
 657         lockdep_assert_irqs_disabled();
 658         if ((rcu_stall_is_suppressed() && !READ_ONCE(rcu_kick_kthreads)) ||
 659             !rcu_gp_in_progress())
 660                 return;
 661         rcu_stall_kick_kthreads();
 662         j = jiffies;
 663
 664         /*
 665          * Lots of memory barriers to reject false positives.
 666          *
 667          * The idea is to pick up rcu_state.gp_seq, then
 668          * rcu_state.jiffies_stall, then rcu_state.gp_start, and finally
 669          * another copy of rcu_state.gp_seq.  These values are updated in
 670          * the opposite order with memory barriers (or equivalent) during
 671          * grace-period initialization and cleanup.  Now, a false positive
 672          * can occur if we get an new value of rcu_state.gp_start and a old
 673          * value of rcu_state.jiffies_stall.  But given the memory barriers,
 674          * the only way that this can happen is if one grace period ends
 675          * and another starts between these two fetches.  This is detected
 676          * by comparing the second fetch of rcu_state.gp_seq with the
 677          * previous fetch from rcu_state.gp_seq.
 678          *
 679          * Given this check, comparisons of jiffies, rcu_state.jiffies_stall,
 680          * and rcu_state.gp_start suffice to forestall false positives.
 681          */
 682         gs1 = READ_ONCE(rcu_state.gp_seq);
 683         smp_rmb(); /* Pick up ->gp_seq first... */
 684         js = READ_ONCE(rcu_state.jiffies_stall);
 685         smp_rmb(); /* ...then ->jiffies_stall before the rest... */
 686         gps = READ_ONCE(rcu_state.gp_start);
 687         smp_rmb(); /* ...and finally ->gp_start before ->gp_seq again. */
 688         gs2 = READ_ONCE(rcu_state.gp_seq);
 689         if (gs1 != gs2 ||
 690             ULONG_CMP_LT(j, js) ||
 691             ULONG_CMP_GE(gps, js))
 692                 return; /* No stall or GP completed since entering function. */
 693         rnp = rdp->mynode;
 694         jn = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
 695         if (rcu_gp_in_progress() &&
 696             (READ_ONCE(rnp->qsmask) & rdp->grpmask) &&
 697             cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
 698
 699                 /* We haven't checked in, so go dump stack. */
 700                 print_cpu_stall(gps);
 701                 if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
 702                         rcu_ftrace_dump(DUMP_ALL);
 703
 704         } else if (rcu_gp_in_progress() &&
 705                    ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY) &&
 706                    cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
 707
 708                 /* They had a few time units to dump stack, so complain. */
 709                 print_other_cpu_stall(gs2, gps);
 710                 if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
 711                         rcu_ftrace_dump(DUMP_ALL);
 712         }
 713 }
 714
 715 //////////////////////////////////////////////////////////////////////////////
 716 //
 717 // RCU forward-progress mechanisms, including of callback invocation.
 718
 719
 720 /*
 721  * Check to see if a failure to end RCU priority inversion was due to
 722  * a CPU not passing through a quiescent state.  When this happens, there
 723  * is nothing that RCU priority boosting can do to help, so we shouldn't
 724  * count this as an RCU priority boosting failure.  A return of true says
 725  * RCU priority boosting is to blame, and false says otherwise.  If false
 726  * is returned, the first of the CPUs to blame is stored through cpup.
 727  * If there was no CPU blocking the current grace period, but also nothing
 728  * in need of being boosted, *cpup is set to -1.  This can happen in case
 729  * of vCPU preemption while the last CPU is reporting its quiscent state,
 730  * for example.
 731  *
 732  * If cpup is NULL, then a lockless quick check is carried out, suitable
 733  * for high-rate usage.  On the other hand, if cpup is non-NULL, each
 734  * rcu_node structure's ->lock is acquired, ruling out high-rate usage.
 735  */
 736 bool rcu_check_boost_fail(unsigned long gp_state, int *cpup)
 737 {
 738         bool atb = false;
 739         int cpu;
 740         unsigned long flags;
 741         struct rcu_node *rnp;
 742
 743         rcu_for_each_leaf_node(rnp) {
 744                 if (!cpup) {
 745                         if (READ_ONCE(rnp->qsmask)) {
 746                                 return false;
 747                         } else {
 748                                 if (READ_ONCE(rnp->gp_tasks))
 749                                         atb = true;
 750                                 continue;
 751                         }
 752                 }
 753                 *cpup = -1;
 754                 raw_spin_lock_irqsave_rcu_node(rnp, flags);
 755                 if (rnp->gp_tasks)
 756                         atb = true;
 757                 if (!rnp->qsmask) {
 758                         // No CPUs without quiescent states for this rnp.
 759                         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 760                         continue;
 761                 }
 762                 // Find the first holdout CPU.
 763                 for_each_leaf_node_possible_cpu(rnp, cpu) {
 764                         if (rnp->qsmask & (1UL << (cpu - rnp->grplo))) {
 765                                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 766                                 *cpup = cpu;
 767                                 return false;
 768                         }
 769                 }
 770                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 771         }
 772         // Can't blame CPUs, so must blame RCU priority boosting.
 773         return atb;
 774 }
 775 EXPORT_SYMBOL_GPL(rcu_check_boost_fail);
 776
 777 /*
 778  * Show the state of the grace-period kthreads.
 779  */
 780 void show_rcu_gp_kthreads(void)
 781 {
 782         unsigned long cbs = 0;
 783         int cpu;
 784         unsigned long j;
 785         unsigned long ja;
 786         unsigned long jr;
 787         unsigned long js;
 788         unsigned long jw;
 789         struct rcu_data *rdp;
 790         struct rcu_node *rnp;
 791         struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
 792
 793         j = jiffies;
 794         ja = j - data_race(rcu_state.gp_activity);
 795         jr = j - data_race(rcu_state.gp_req_activity);
 796         js = j - data_race(rcu_state.gp_start);
 797         jw = j - data_race(rcu_state.gp_wake_time);
 798         pr_info("%s: wait state: %s(%d) ->state: %#x ->rt_priority %u delta ->gp_start %lu ->gp_activity %lu ->gp_req_activity %lu ->gp_wake_time %lu ->gp_wake_seq %ld ->gp_seq %ld ->gp_seq_needed %ld ->gp_max %lu ->gp_flags %#x\n",
 799                 rcu_state.name, gp_state_getname(rcu_state.gp_state),
 800                 rcu_state.gp_state, t ? t->__state : 0x1ffff, t ? t->rt_priority : 0xffU,
 801                 js, ja, jr, jw, (long)data_race(rcu_state.gp_wake_seq),
 802                 (long)data_race(rcu_state.gp_seq),
 803                 (long)data_race(rcu_get_root()->gp_seq_needed),
 804                 data_race(rcu_state.gp_max),
 805                 data_race(rcu_state.gp_flags));
 806         rcu_for_each_node_breadth_first(rnp) {
 807                 if (ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq), READ_ONCE(rnp->gp_seq_needed)) &&
 808                     !data_race(rnp->qsmask) && !data_race(rnp->boost_tasks) &&
 809                     !data_race(rnp->exp_tasks) && !data_race(rnp->gp_tasks))
 810                         continue;
 811                 pr_info("\trcu_node %d:%d ->gp_seq %ld ->gp_seq_needed %ld ->qsmask %#lx %c%c%c%c ->n_boosts %ld\n",
 812                         rnp->grplo, rnp->grphi,
 813                         (long)data_race(rnp->gp_seq), (long)data_race(rnp->gp_seq_needed),
 814                         data_race(rnp->qsmask),
 815                         ".b"[!!data_race(rnp->boost_kthread_task)],
 816                         ".B"[!!data_race(rnp->boost_tasks)],
 817                         ".E"[!!data_race(rnp->exp_tasks)],
 818                         ".G"[!!data_race(rnp->gp_tasks)],
 819                         data_race(rnp->n_boosts));
 820                 if (!rcu_is_leaf_node(rnp))
 821                         continue;
 822                 for_each_leaf_node_possible_cpu(rnp, cpu) {
 823                         rdp = per_cpu_ptr(&rcu_data, cpu);
 824                         if (READ_ONCE(rdp->gpwrap) ||
 825                             ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq),
 826                                          READ_ONCE(rdp->gp_seq_needed)))
 827                                 continue;
 828                         pr_info("\tcpu %d ->gp_seq_needed %ld\n",
 829                                 cpu, (long)data_race(rdp->gp_seq_needed));
 830                 }
 831         }
 832         for_each_possible_cpu(cpu) {
 833                 rdp = per_cpu_ptr(&rcu_data, cpu);
 834                 cbs += data_race(rdp->n_cbs_invoked);
 835                 if (rcu_segcblist_is_offloaded(&rdp->cblist))
 836                         show_rcu_nocb_state(rdp);
 837         }
 838         pr_info("RCU callbacks invoked since boot: %lu\n", cbs);
 839         show_rcu_tasks_gp_kthreads();
 840 }
 841 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
 842
 843 /*
 844  * This function checks for grace-period requests that fail to motivate
 845  * RCU to come out of its idle mode.
 846  */
 847 static void rcu_check_gp_start_stall(struct rcu_node *rnp, struct rcu_data *rdp,
 848                                      const unsigned long gpssdelay)
 849 {
 850         unsigned long flags;
 851         unsigned long j;
 852         struct rcu_node *rnp_root = rcu_get_root();
 853         static atomic_t warned = ATOMIC_INIT(0);
 854
 855         if (!IS_ENABLED(CONFIG_PROVE_RCU) || rcu_gp_in_progress() ||
 856             ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
 857                          READ_ONCE(rnp_root->gp_seq_needed)) ||
 858             !smp_load_acquire(&rcu_state.gp_kthread)) // Get stable kthread.
 859                 return;
 860         j = jiffies; /* Expensive access, and in common case don't get here. */
 861         if (time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
 862             time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
 863             atomic_read(&warned))
 864                 return;
 865
 866         raw_spin_lock_irqsave_rcu_node(rnp, flags);
 867         j = jiffies;
 868         if (rcu_gp_in_progress() ||
 869             ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
 870                          READ_ONCE(rnp_root->gp_seq_needed)) ||
 871             time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
 872             time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
 873             atomic_read(&warned)) {
 874                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 875                 return;
 876         }
 877         /* Hold onto the leaf lock to make others see warned==1. */
 878
 879         if (rnp_root != rnp)
 880                 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
 881         j = jiffies;
 882         if (rcu_gp_in_progress() ||
 883             ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
 884                          READ_ONCE(rnp_root->gp_seq_needed)) ||
 885             time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
 886             time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
 887             atomic_xchg(&warned, 1)) {
 888                 if (rnp_root != rnp)
 889                         /* irqs remain disabled. */
 890                         raw_spin_unlock_rcu_node(rnp_root);
 891                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 892                 return;
 893         }
 894         WARN_ON(1);
 895         if (rnp_root != rnp)
 896                 raw_spin_unlock_rcu_node(rnp_root);
 897         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
 898         show_rcu_gp_kthreads();
 899 }
 900
 901 /*
 902  * Do a forward-progress check for rcutorture.  This is normally invoked
 903  * due to an OOM event.  The argument "j" gives the time period during
 904  * which rcutorture would like progress to have been made.
 905  */
 906 void rcu_fwd_progress_check(unsigned long j)
 907 {
 908         unsigned long cbs;
 909         int cpu;
 910         unsigned long max_cbs = 0;
 911         int max_cpu = -1;
 912         struct rcu_data *rdp;
 913
 914         if (rcu_gp_in_progress()) {
 915                 pr_info("%s: GP age %lu jiffies\n",
 916                         __func__, jiffies - rcu_state.gp_start);
 917                 show_rcu_gp_kthreads();
 918         } else {
 919                 pr_info("%s: Last GP end %lu jiffies ago\n",
 920                         __func__, jiffies - rcu_state.gp_end);
 921                 preempt_disable();
 922                 rdp = this_cpu_ptr(&rcu_data);
 923                 rcu_check_gp_start_stall(rdp->mynode, rdp, j);
 924                 preempt_enable();
 925         }
 926         for_each_possible_cpu(cpu) {
 927                 cbs = rcu_get_n_cbs_cpu(cpu);
 928                 if (!cbs)
 929                         continue;
 930                 if (max_cpu < 0)
 931                         pr_info("%s: callbacks", __func__);
 932                 pr_cont(" %d: %lu", cpu, cbs);
 933                 if (cbs <= max_cbs)
 934                         continue;
 935                 max_cbs = cbs;
 936                 max_cpu = cpu;
 937         }
 938         if (max_cpu >= 0)
 939                 pr_cont("\n");
 940 }
 941 EXPORT_SYMBOL_GPL(rcu_fwd_progress_check);
 942
 943 /* Commandeer a sysrq key to dump RCU's tree. */
 944 static bool sysrq_rcu;
 945 module_param(sysrq_rcu, bool, 0444);
 946
 947 /* Dump grace-period-request information due to commandeered sysrq. */
 948 static void sysrq_show_rcu(int key)
 949 {
 950         show_rcu_gp_kthreads();
 951 }
 952
 953 static const struct sysrq_key_op sysrq_rcudump_op = {
 954         .handler = sysrq_show_rcu,
 955         .help_msg = "show-rcu(y)",
 956         .action_msg = "Show RCU tree",
 957         .enable_mask = SYSRQ_ENABLE_DUMP,
 958 };
 959
 960 static int __init rcu_sysrq_init(void)
 961 {
 962         if (sysrq_rcu)
 963                 return register_sysrq_key('y', &sysrq_rcudump_op);
 964         return 0;
 965 }
 966 early_initcall(rcu_sysrq_init);