1 // SPDX-License-Identifier: GPL-2.0+
3 * RCU CPU stall warnings for normal RCU grace periods
5 * Copyright IBM Corporation, 2019
7 * Author: Paul E. McKenney <paulmck@linux.ibm.com>
10 #include <linux/kvm_para.h>
12 //////////////////////////////////////////////////////////////////////////////
14 // Controlling CPU stall warnings, including delay calculation.
16 /* panic() on RCU Stall sysctl. */
17 int sysctl_panic_on_rcu_stall __read_mostly;
18 int sysctl_max_rcu_stall_to_panic __read_mostly;
20 #ifdef CONFIG_PROVE_RCU
21 #define RCU_STALL_DELAY_DELTA (5 * HZ)
23 #define RCU_STALL_DELAY_DELTA 0
25 #define RCU_STALL_MIGHT_DIV 8
26 #define RCU_STALL_MIGHT_MIN (2 * HZ)
28 int rcu_exp_jiffies_till_stall_check(void)
30 int cpu_stall_timeout = READ_ONCE(rcu_exp_cpu_stall_timeout);
31 int exp_stall_delay_delta = 0;
34 // Zero says to use rcu_cpu_stall_timeout, but in milliseconds.
35 if (!cpu_stall_timeout)
36 cpu_stall_timeout = jiffies_to_msecs(rcu_jiffies_till_stall_check());
38 // Limit check must be consistent with the Kconfig limits for
39 // CONFIG_RCU_EXP_CPU_STALL_TIMEOUT, so check the allowed range.
40 // The minimum clamped value is "2UL", because at least one full
41 // tick has to be guaranteed.
42 till_stall_check = clamp(msecs_to_jiffies(cpu_stall_timeout), 2UL, 300UL * HZ);
44 if (cpu_stall_timeout && jiffies_to_msecs(till_stall_check) != cpu_stall_timeout)
45 WRITE_ONCE(rcu_exp_cpu_stall_timeout, jiffies_to_msecs(till_stall_check));
47 #ifdef CONFIG_PROVE_RCU
48 /* Add extra ~25% out of till_stall_check. */
49 exp_stall_delay_delta = ((till_stall_check * 25) / 100) + 1;
52 return till_stall_check + exp_stall_delay_delta;
54 EXPORT_SYMBOL_GPL(rcu_exp_jiffies_till_stall_check);
56 /* Limit-check stall timeouts specified at boottime and runtime. */
57 int rcu_jiffies_till_stall_check(void)
59 int till_stall_check = READ_ONCE(rcu_cpu_stall_timeout);
62 * Limit check must be consistent with the Kconfig limits
63 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
65 if (till_stall_check < 3) {
66 WRITE_ONCE(rcu_cpu_stall_timeout, 3);
68 } else if (till_stall_check > 300) {
69 WRITE_ONCE(rcu_cpu_stall_timeout, 300);
70 till_stall_check = 300;
72 return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
74 EXPORT_SYMBOL_GPL(rcu_jiffies_till_stall_check);
77 * rcu_gp_might_be_stalled - Is it likely that the grace period is stalled?
79 * Returns @true if the current grace period is sufficiently old that
80 * it is reasonable to assume that it might be stalled. This can be
81 * useful when deciding whether to allocate memory to enable RCU-mediated
82 * freeing on the one hand or just invoking synchronize_rcu() on the other.
83 * The latter is preferable when the grace period is stalled.
85 * Note that sampling of the .gp_start and .gp_seq fields must be done
86 * carefully to avoid false positives at the beginnings and ends of
89 bool rcu_gp_might_be_stalled(void)
91 unsigned long d = rcu_jiffies_till_stall_check() / RCU_STALL_MIGHT_DIV;
92 unsigned long j = jiffies;
94 if (d < RCU_STALL_MIGHT_MIN)
95 d = RCU_STALL_MIGHT_MIN;
96 smp_mb(); // jiffies before .gp_seq to avoid false positives.
97 if (!rcu_gp_in_progress())
99 // Long delays at this point avoids false positive, but a delay
100 // of ULONG_MAX/4 jiffies voids your no-false-positive warranty.
101 smp_mb(); // .gp_seq before second .gp_start
103 return !time_before(j, READ_ONCE(rcu_state.gp_start) + d);
106 /* Don't do RCU CPU stall warnings during long sysrq printouts. */
107 void rcu_sysrq_start(void)
109 if (!rcu_cpu_stall_suppress)
110 rcu_cpu_stall_suppress = 2;
113 void rcu_sysrq_end(void)
115 if (rcu_cpu_stall_suppress == 2)
116 rcu_cpu_stall_suppress = 0;
119 /* Don't print RCU CPU stall warnings during a kernel panic. */
120 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
122 rcu_cpu_stall_suppress = 1;
126 static struct notifier_block rcu_panic_block = {
127 .notifier_call = rcu_panic,
130 static int __init check_cpu_stall_init(void)
132 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
135 early_initcall(check_cpu_stall_init);
137 /* If so specified via sysctl, panic, yielding cleaner stall-warning output. */
138 static void panic_on_rcu_stall(void)
140 static int cpu_stall;
142 if (++cpu_stall < sysctl_max_rcu_stall_to_panic)
145 if (sysctl_panic_on_rcu_stall)
146 panic("RCU Stall\n");
150 * rcu_cpu_stall_reset - restart stall-warning timeout for current grace period
152 * The caller must disable hard irqs.
154 void rcu_cpu_stall_reset(void)
156 WRITE_ONCE(rcu_state.jiffies_stall,
157 jiffies + rcu_jiffies_till_stall_check());
160 //////////////////////////////////////////////////////////////////////////////
162 // Interaction with RCU grace periods
164 /* Start of new grace period, so record stall time (and forcing times). */
165 static void record_gp_stall_check_time(void)
167 unsigned long j = jiffies;
170 WRITE_ONCE(rcu_state.gp_start, j);
171 j1 = rcu_jiffies_till_stall_check();
172 smp_mb(); // ->gp_start before ->jiffies_stall and caller's ->gp_seq.
173 WRITE_ONCE(rcu_state.jiffies_stall, j + j1);
174 rcu_state.jiffies_resched = j + j1 / 2;
175 rcu_state.n_force_qs_gpstart = READ_ONCE(rcu_state.n_force_qs);
178 /* Zero ->ticks_this_gp and snapshot the number of RCU softirq handlers. */
179 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
181 rdp->ticks_this_gp = 0;
182 rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
183 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
187 * If too much time has passed in the current grace period, and if
188 * so configured, go kick the relevant kthreads.
190 static void rcu_stall_kick_kthreads(void)
194 if (!READ_ONCE(rcu_kick_kthreads))
196 j = READ_ONCE(rcu_state.jiffies_kick_kthreads);
197 if (time_after(jiffies, j) && rcu_state.gp_kthread &&
198 (rcu_gp_in_progress() || READ_ONCE(rcu_state.gp_flags))) {
199 WARN_ONCE(1, "Kicking %s grace-period kthread\n",
201 rcu_ftrace_dump(DUMP_ALL);
202 wake_up_process(rcu_state.gp_kthread);
203 WRITE_ONCE(rcu_state.jiffies_kick_kthreads, j + HZ);
208 * Handler for the irq_work request posted about halfway into the RCU CPU
209 * stall timeout, and used to detect excessive irq disabling. Set state
210 * appropriately, but just complain if there is unexpected state on entry.
212 static void rcu_iw_handler(struct irq_work *iwp)
214 struct rcu_data *rdp;
215 struct rcu_node *rnp;
217 rdp = container_of(iwp, struct rcu_data, rcu_iw);
219 raw_spin_lock_rcu_node(rnp);
220 if (!WARN_ON_ONCE(!rdp->rcu_iw_pending)) {
221 rdp->rcu_iw_gp_seq = rnp->gp_seq;
222 rdp->rcu_iw_pending = false;
224 raw_spin_unlock_rcu_node(rnp);
227 //////////////////////////////////////////////////////////////////////////////
229 // Printing RCU CPU stall warnings
231 #ifdef CONFIG_PREEMPT_RCU
234 * Dump detailed information for all tasks blocking the current RCU
235 * grace period on the specified rcu_node structure.
237 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
240 struct task_struct *t;
242 raw_spin_lock_irqsave_rcu_node(rnp, flags);
243 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
244 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
247 t = list_entry(rnp->gp_tasks->prev,
248 struct task_struct, rcu_node_entry);
249 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
251 * We could be printing a lot while holding a spinlock.
252 * Avoid triggering hard lockup.
254 touch_nmi_watchdog();
257 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
260 // Communicate task state back to the RCU CPU stall warning request.
261 struct rcu_stall_chk_rdr {
263 union rcu_special rs;
268 * Report out the state of a not-running task that is stalling the
269 * current RCU grace period.
271 static int check_slow_task(struct task_struct *t, void *arg)
273 struct rcu_stall_chk_rdr *rscrp = arg;
276 return -EBUSY; // It is running, so decline to inspect it.
277 rscrp->nesting = t->rcu_read_lock_nesting;
278 rscrp->rs = t->rcu_read_unlock_special;
279 rscrp->on_blkd_list = !list_empty(&t->rcu_node_entry);
284 * Scan the current list of tasks blocked within RCU read-side critical
285 * sections, printing out the tid of each of the first few of them.
287 static int rcu_print_task_stall(struct rcu_node *rnp, unsigned long flags)
288 __releases(rnp->lock)
292 struct rcu_stall_chk_rdr rscr;
293 struct task_struct *t;
294 struct task_struct *ts[8];
296 lockdep_assert_irqs_disabled();
297 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
298 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
301 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
302 rnp->level, rnp->grplo, rnp->grphi);
303 t = list_entry(rnp->gp_tasks->prev,
304 struct task_struct, rcu_node_entry);
305 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
308 if (i >= ARRAY_SIZE(ts))
311 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
314 if (task_call_func(t, check_slow_task, &rscr))
315 pr_cont(" P%d", t->pid);
317 pr_cont(" P%d/%d:%c%c%c%c",
318 t->pid, rscr.nesting,
319 ".b"[rscr.rs.b.blocked],
320 ".q"[rscr.rs.b.need_qs],
321 ".e"[rscr.rs.b.exp_hint],
322 ".l"[rscr.on_blkd_list]);
323 lockdep_assert_irqs_disabled();
331 #else /* #ifdef CONFIG_PREEMPT_RCU */
334 * Because preemptible RCU does not exist, we never have to check for
335 * tasks blocked within RCU read-side critical sections.
337 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
342 * Because preemptible RCU does not exist, we never have to check for
343 * tasks blocked within RCU read-side critical sections.
345 static int rcu_print_task_stall(struct rcu_node *rnp, unsigned long flags)
346 __releases(rnp->lock)
348 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
351 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
354 * Dump stacks of all tasks running on stalled CPUs. First try using
355 * NMIs, but fall back to manual remote stack tracing on architectures
356 * that don't support NMI-based stack dumps. The NMI-triggered stack
357 * traces are more accurate because they are printed by the target CPU.
359 static void rcu_dump_cpu_stacks(void)
363 struct rcu_node *rnp;
365 rcu_for_each_leaf_node(rnp) {
366 raw_spin_lock_irqsave_rcu_node(rnp, flags);
367 for_each_leaf_node_possible_cpu(rnp, cpu)
368 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
369 if (cpu_is_offline(cpu))
370 pr_err("Offline CPU %d blocking current GP.\n", cpu);
374 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
378 static const char * const gp_state_names[] = {
379 [RCU_GP_IDLE] = "RCU_GP_IDLE",
380 [RCU_GP_WAIT_GPS] = "RCU_GP_WAIT_GPS",
381 [RCU_GP_DONE_GPS] = "RCU_GP_DONE_GPS",
382 [RCU_GP_ONOFF] = "RCU_GP_ONOFF",
383 [RCU_GP_INIT] = "RCU_GP_INIT",
384 [RCU_GP_WAIT_FQS] = "RCU_GP_WAIT_FQS",
385 [RCU_GP_DOING_FQS] = "RCU_GP_DOING_FQS",
386 [RCU_GP_CLEANUP] = "RCU_GP_CLEANUP",
387 [RCU_GP_CLEANED] = "RCU_GP_CLEANED",
391 * Convert a ->gp_state value to a character string.
393 static const char *gp_state_getname(short gs)
395 if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
397 return gp_state_names[gs];
400 /* Is the RCU grace-period kthread being starved of CPU time? */
401 static bool rcu_is_gp_kthread_starving(unsigned long *jp)
403 unsigned long j = jiffies - READ_ONCE(rcu_state.gp_activity);
410 static bool rcu_is_rcuc_kthread_starving(struct rcu_data *rdp, unsigned long *jp)
413 struct task_struct *rcuc;
416 rcuc = rdp->rcu_cpu_kthread_task;
420 cpu = task_cpu(rcuc);
421 if (cpu_is_offline(cpu) || idle_cpu(cpu))
424 j = jiffies - READ_ONCE(rdp->rcuc_activity);
431 static void print_cpu_stat_info(int cpu)
433 struct rcu_snap_record rsr, *rsrp;
434 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
435 struct kernel_cpustat *kcsp = &kcpustat_cpu(cpu);
437 if (!rcu_cpu_stall_cputime)
440 rsrp = &rdp->snap_record;
441 if (rsrp->gp_seq != rdp->gp_seq)
444 rsr.cputime_irq = kcpustat_field(kcsp, CPUTIME_IRQ, cpu);
445 rsr.cputime_softirq = kcpustat_field(kcsp, CPUTIME_SOFTIRQ, cpu);
446 rsr.cputime_system = kcpustat_field(kcsp, CPUTIME_SYSTEM, cpu);
448 pr_err("\t hardirqs softirqs csw/system\n");
449 pr_err("\t number: %8ld %10d %12lld\n",
450 kstat_cpu_irqs_sum(cpu) - rsrp->nr_hardirqs,
451 kstat_cpu_softirqs_sum(cpu) - rsrp->nr_softirqs,
452 nr_context_switches_cpu(cpu) - rsrp->nr_csw);
453 pr_err("\tcputime: %8lld %10lld %12lld ==> %d(ms)\n",
454 div_u64(rsr.cputime_irq - rsrp->cputime_irq, NSEC_PER_MSEC),
455 div_u64(rsr.cputime_softirq - rsrp->cputime_softirq, NSEC_PER_MSEC),
456 div_u64(rsr.cputime_system - rsrp->cputime_system, NSEC_PER_MSEC),
457 jiffies_to_msecs(jiffies - rsrp->jiffies));
461 * Print out diagnostic information for the specified stalled CPU.
463 * If the specified CPU is aware of the current RCU grace period, then
464 * print the number of scheduling clock interrupts the CPU has taken
465 * during the time that it has been aware. Otherwise, print the number
466 * of RCU grace periods that this CPU is ignorant of, for example, "1"
467 * if the CPU was aware of the previous grace period.
469 * Also print out idle info.
471 static void print_cpu_stall_info(int cpu)
475 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
477 unsigned long ticks_value;
483 * We could be printing a lot while holding a spinlock. Avoid
484 * triggering hard lockup.
486 touch_nmi_watchdog();
488 ticks_value = rcu_seq_ctr(rcu_state.gp_seq - rdp->gp_seq);
490 ticks_title = "GPs behind";
492 ticks_title = "ticks this GP";
493 ticks_value = rdp->ticks_this_gp;
495 delta = rcu_seq_ctr(rdp->mynode->gp_seq - rdp->rcu_iw_gp_seq);
496 falsepositive = rcu_is_gp_kthread_starving(NULL) &&
497 rcu_dynticks_in_eqs(rcu_dynticks_snap(cpu));
498 rcuc_starved = rcu_is_rcuc_kthread_starving(rdp, &j);
500 sprintf(buf, " rcuc=%ld jiffies(starved)", j);
501 pr_err("\t%d-%c%c%c%c: (%lu %s) idle=%04x/%ld/%#lx softirq=%u/%u fqs=%ld%s%s\n",
503 "O."[!!cpu_online(cpu)],
504 "o."[!!(rdp->grpmask & rdp->mynode->qsmaskinit)],
505 "N."[!!(rdp->grpmask & rdp->mynode->qsmaskinitnext)],
506 !IS_ENABLED(CONFIG_IRQ_WORK) ? '?' :
507 rdp->rcu_iw_pending ? (int)min(delta, 9UL) + '0' :
509 ticks_value, ticks_title,
510 rcu_dynticks_snap(cpu) & 0xffff,
511 ct_dynticks_nesting_cpu(cpu), ct_dynticks_nmi_nesting_cpu(cpu),
512 rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
513 data_race(rcu_state.n_force_qs) - rcu_state.n_force_qs_gpstart,
514 rcuc_starved ? buf : "",
515 falsepositive ? " (false positive?)" : "");
517 print_cpu_stat_info(cpu);
520 /* Complain about starvation of grace-period kthread. */
521 static void rcu_check_gp_kthread_starvation(void)
524 struct task_struct *gpk = rcu_state.gp_kthread;
527 if (rcu_is_gp_kthread_starving(&j)) {
528 cpu = gpk ? task_cpu(gpk) : -1;
529 pr_err("%s kthread starved for %ld jiffies! g%ld f%#x %s(%d) ->state=%#x ->cpu=%d\n",
531 (long)rcu_seq_current(&rcu_state.gp_seq),
532 data_race(READ_ONCE(rcu_state.gp_flags)),
533 gp_state_getname(rcu_state.gp_state),
534 data_race(READ_ONCE(rcu_state.gp_state)),
535 gpk ? data_race(READ_ONCE(gpk->__state)) : ~0, cpu);
537 pr_err("\tUnless %s kthread gets sufficient CPU time, OOM is now expected behavior.\n", rcu_state.name);
538 pr_err("RCU grace-period kthread stack dump:\n");
539 sched_show_task(gpk);
541 if (cpu_is_offline(cpu)) {
542 pr_err("RCU GP kthread last ran on offline CPU %d.\n", cpu);
544 pr_err("Stack dump where RCU GP kthread last ran:\n");
548 wake_up_process(gpk);
553 /* Complain about missing wakeups from expired fqs wait timer */
554 static void rcu_check_gp_kthread_expired_fqs_timer(void)
556 struct task_struct *gpk = rcu_state.gp_kthread;
558 unsigned long jiffies_fqs;
562 * Order reads of .gp_state and .jiffies_force_qs.
563 * Matching smp_wmb() is present in rcu_gp_fqs_loop().
565 gp_state = smp_load_acquire(&rcu_state.gp_state);
566 jiffies_fqs = READ_ONCE(rcu_state.jiffies_force_qs);
568 if (gp_state == RCU_GP_WAIT_FQS &&
569 time_after(jiffies, jiffies_fqs + RCU_STALL_MIGHT_MIN) &&
570 gpk && !READ_ONCE(gpk->on_rq)) {
572 pr_err("%s kthread timer wakeup didn't happen for %ld jiffies! g%ld f%#x %s(%d) ->state=%#x\n",
573 rcu_state.name, (jiffies - jiffies_fqs),
574 (long)rcu_seq_current(&rcu_state.gp_seq),
575 data_race(rcu_state.gp_flags),
576 gp_state_getname(RCU_GP_WAIT_FQS), RCU_GP_WAIT_FQS,
577 data_race(READ_ONCE(gpk->__state)));
578 pr_err("\tPossible timer handling issue on cpu=%d timer-softirq=%u\n",
579 cpu, kstat_softirqs_cpu(TIMER_SOFTIRQ, cpu));
583 static void print_other_cpu_stall(unsigned long gp_seq, unsigned long gps)
590 struct rcu_node *rnp;
593 lockdep_assert_irqs_disabled();
595 /* Kick and suppress, if so configured. */
596 rcu_stall_kick_kthreads();
597 if (rcu_stall_is_suppressed())
601 * OK, time to rat on our buddy...
602 * See Documentation/RCU/stallwarn.rst for info on how to debug
603 * RCU CPU stall warnings.
605 trace_rcu_stall_warning(rcu_state.name, TPS("StallDetected"));
606 pr_err("INFO: %s detected stalls on CPUs/tasks:\n", rcu_state.name);
607 rcu_for_each_leaf_node(rnp) {
608 raw_spin_lock_irqsave_rcu_node(rnp, flags);
609 if (rnp->qsmask != 0) {
610 for_each_leaf_node_possible_cpu(rnp, cpu)
611 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
612 print_cpu_stall_info(cpu);
616 ndetected += rcu_print_task_stall(rnp, flags); // Releases rnp->lock.
617 lockdep_assert_irqs_disabled();
620 for_each_possible_cpu(cpu)
621 totqlen += rcu_get_n_cbs_cpu(cpu);
622 pr_err("\t(detected by %d, t=%ld jiffies, g=%ld, q=%lu ncpus=%d)\n",
623 smp_processor_id(), (long)(jiffies - gps),
624 (long)rcu_seq_current(&rcu_state.gp_seq), totqlen, rcu_state.n_online_cpus);
626 rcu_dump_cpu_stacks();
628 /* Complain about tasks blocking the grace period. */
629 rcu_for_each_leaf_node(rnp)
630 rcu_print_detail_task_stall_rnp(rnp);
632 if (rcu_seq_current(&rcu_state.gp_seq) != gp_seq) {
633 pr_err("INFO: Stall ended before state dump start\n");
636 gpa = data_race(READ_ONCE(rcu_state.gp_activity));
637 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
638 rcu_state.name, j - gpa, j, gpa,
639 data_race(READ_ONCE(jiffies_till_next_fqs)),
640 data_race(READ_ONCE(rcu_get_root()->qsmask)));
643 /* Rewrite if needed in case of slow consoles. */
644 if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
645 WRITE_ONCE(rcu_state.jiffies_stall,
646 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
648 rcu_check_gp_kthread_expired_fqs_timer();
649 rcu_check_gp_kthread_starvation();
651 panic_on_rcu_stall();
653 rcu_force_quiescent_state(); /* Kick them all. */
656 static void print_cpu_stall(unsigned long gps)
660 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
661 struct rcu_node *rnp = rcu_get_root();
664 lockdep_assert_irqs_disabled();
666 /* Kick and suppress, if so configured. */
667 rcu_stall_kick_kthreads();
668 if (rcu_stall_is_suppressed())
672 * OK, time to rat on ourselves...
673 * See Documentation/RCU/stallwarn.rst for info on how to debug
674 * RCU CPU stall warnings.
676 trace_rcu_stall_warning(rcu_state.name, TPS("SelfDetected"));
677 pr_err("INFO: %s self-detected stall on CPU\n", rcu_state.name);
678 raw_spin_lock_irqsave_rcu_node(rdp->mynode, flags);
679 print_cpu_stall_info(smp_processor_id());
680 raw_spin_unlock_irqrestore_rcu_node(rdp->mynode, flags);
681 for_each_possible_cpu(cpu)
682 totqlen += rcu_get_n_cbs_cpu(cpu);
683 pr_err("\t(t=%lu jiffies g=%ld q=%lu ncpus=%d)\n",
685 (long)rcu_seq_current(&rcu_state.gp_seq), totqlen, rcu_state.n_online_cpus);
687 rcu_check_gp_kthread_expired_fqs_timer();
688 rcu_check_gp_kthread_starvation();
690 rcu_dump_cpu_stacks();
692 raw_spin_lock_irqsave_rcu_node(rnp, flags);
693 /* Rewrite if needed in case of slow consoles. */
694 if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
695 WRITE_ONCE(rcu_state.jiffies_stall,
696 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
697 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
699 panic_on_rcu_stall();
702 * Attempt to revive the RCU machinery by forcing a context switch.
704 * A context switch would normally allow the RCU state machine to make
705 * progress and it could be we're stuck in kernel space without context
706 * switches for an entirely unreasonable amount of time.
708 set_tsk_need_resched(current);
709 set_preempt_need_resched();
712 static void check_cpu_stall(struct rcu_data *rdp)
714 bool didstall = false;
721 struct rcu_node *rnp;
723 lockdep_assert_irqs_disabled();
724 if ((rcu_stall_is_suppressed() && !READ_ONCE(rcu_kick_kthreads)) ||
725 !rcu_gp_in_progress())
727 rcu_stall_kick_kthreads();
731 * Lots of memory barriers to reject false positives.
733 * The idea is to pick up rcu_state.gp_seq, then
734 * rcu_state.jiffies_stall, then rcu_state.gp_start, and finally
735 * another copy of rcu_state.gp_seq. These values are updated in
736 * the opposite order with memory barriers (or equivalent) during
737 * grace-period initialization and cleanup. Now, a false positive
738 * can occur if we get an new value of rcu_state.gp_start and a old
739 * value of rcu_state.jiffies_stall. But given the memory barriers,
740 * the only way that this can happen is if one grace period ends
741 * and another starts between these two fetches. This is detected
742 * by comparing the second fetch of rcu_state.gp_seq with the
743 * previous fetch from rcu_state.gp_seq.
745 * Given this check, comparisons of jiffies, rcu_state.jiffies_stall,
746 * and rcu_state.gp_start suffice to forestall false positives.
748 gs1 = READ_ONCE(rcu_state.gp_seq);
749 smp_rmb(); /* Pick up ->gp_seq first... */
750 js = READ_ONCE(rcu_state.jiffies_stall);
751 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
752 gps = READ_ONCE(rcu_state.gp_start);
753 smp_rmb(); /* ...and finally ->gp_start before ->gp_seq again. */
754 gs2 = READ_ONCE(rcu_state.gp_seq);
756 ULONG_CMP_LT(j, js) ||
757 ULONG_CMP_GE(gps, js))
758 return; /* No stall or GP completed since entering function. */
760 jn = jiffies + ULONG_MAX / 2;
761 if (rcu_gp_in_progress() &&
762 (READ_ONCE(rnp->qsmask) & rdp->grpmask) &&
763 cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
766 * If a virtual machine is stopped by the host it can look to
767 * the watchdog like an RCU stall. Check to see if the host
770 if (kvm_check_and_clear_guest_paused())
773 /* We haven't checked in, so go dump stack. */
774 print_cpu_stall(gps);
775 if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
776 rcu_ftrace_dump(DUMP_ALL);
779 } else if (rcu_gp_in_progress() &&
780 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY) &&
781 cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
784 * If a virtual machine is stopped by the host it can look to
785 * the watchdog like an RCU stall. Check to see if the host
788 if (kvm_check_and_clear_guest_paused())
791 /* They had a few time units to dump stack, so complain. */
792 print_other_cpu_stall(gs2, gps);
793 if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
794 rcu_ftrace_dump(DUMP_ALL);
797 if (didstall && READ_ONCE(rcu_state.jiffies_stall) == jn) {
798 jn = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
799 WRITE_ONCE(rcu_state.jiffies_stall, jn);
803 //////////////////////////////////////////////////////////////////////////////
805 // RCU forward-progress mechanisms, including of callback invocation.
809 * Check to see if a failure to end RCU priority inversion was due to
810 * a CPU not passing through a quiescent state. When this happens, there
811 * is nothing that RCU priority boosting can do to help, so we shouldn't
812 * count this as an RCU priority boosting failure. A return of true says
813 * RCU priority boosting is to blame, and false says otherwise. If false
814 * is returned, the first of the CPUs to blame is stored through cpup.
815 * If there was no CPU blocking the current grace period, but also nothing
816 * in need of being boosted, *cpup is set to -1. This can happen in case
817 * of vCPU preemption while the last CPU is reporting its quiscent state,
820 * If cpup is NULL, then a lockless quick check is carried out, suitable
821 * for high-rate usage. On the other hand, if cpup is non-NULL, each
822 * rcu_node structure's ->lock is acquired, ruling out high-rate usage.
824 bool rcu_check_boost_fail(unsigned long gp_state, int *cpup)
829 struct rcu_node *rnp;
831 rcu_for_each_leaf_node(rnp) {
833 if (data_race(READ_ONCE(rnp->qsmask))) {
836 if (READ_ONCE(rnp->gp_tasks))
842 raw_spin_lock_irqsave_rcu_node(rnp, flags);
846 // No CPUs without quiescent states for this rnp.
847 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
850 // Find the first holdout CPU.
851 for_each_leaf_node_possible_cpu(rnp, cpu) {
852 if (rnp->qsmask & (1UL << (cpu - rnp->grplo))) {
853 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
858 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
860 // Can't blame CPUs, so must blame RCU priority boosting.
863 EXPORT_SYMBOL_GPL(rcu_check_boost_fail);
866 * Show the state of the grace-period kthreads.
868 void show_rcu_gp_kthreads(void)
870 unsigned long cbs = 0;
877 struct rcu_data *rdp;
878 struct rcu_node *rnp;
879 struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
882 ja = j - data_race(READ_ONCE(rcu_state.gp_activity));
883 jr = j - data_race(READ_ONCE(rcu_state.gp_req_activity));
884 js = j - data_race(READ_ONCE(rcu_state.gp_start));
885 jw = j - data_race(READ_ONCE(rcu_state.gp_wake_time));
886 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",
887 rcu_state.name, gp_state_getname(rcu_state.gp_state),
888 data_race(READ_ONCE(rcu_state.gp_state)),
889 t ? data_race(READ_ONCE(t->__state)) : 0x1ffff, t ? t->rt_priority : 0xffU,
890 js, ja, jr, jw, (long)data_race(READ_ONCE(rcu_state.gp_wake_seq)),
891 (long)data_race(READ_ONCE(rcu_state.gp_seq)),
892 (long)data_race(READ_ONCE(rcu_get_root()->gp_seq_needed)),
893 data_race(READ_ONCE(rcu_state.gp_max)),
894 data_race(READ_ONCE(rcu_state.gp_flags)));
895 rcu_for_each_node_breadth_first(rnp) {
896 if (ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq), READ_ONCE(rnp->gp_seq_needed)) &&
897 !data_race(READ_ONCE(rnp->qsmask)) && !data_race(READ_ONCE(rnp->boost_tasks)) &&
898 !data_race(READ_ONCE(rnp->exp_tasks)) && !data_race(READ_ONCE(rnp->gp_tasks)))
900 pr_info("\trcu_node %d:%d ->gp_seq %ld ->gp_seq_needed %ld ->qsmask %#lx %c%c%c%c ->n_boosts %ld\n",
901 rnp->grplo, rnp->grphi,
902 (long)data_race(READ_ONCE(rnp->gp_seq)),
903 (long)data_race(READ_ONCE(rnp->gp_seq_needed)),
904 data_race(READ_ONCE(rnp->qsmask)),
905 ".b"[!!data_race(READ_ONCE(rnp->boost_kthread_task))],
906 ".B"[!!data_race(READ_ONCE(rnp->boost_tasks))],
907 ".E"[!!data_race(READ_ONCE(rnp->exp_tasks))],
908 ".G"[!!data_race(READ_ONCE(rnp->gp_tasks))],
909 data_race(READ_ONCE(rnp->n_boosts)));
910 if (!rcu_is_leaf_node(rnp))
912 for_each_leaf_node_possible_cpu(rnp, cpu) {
913 rdp = per_cpu_ptr(&rcu_data, cpu);
914 if (READ_ONCE(rdp->gpwrap) ||
915 ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq),
916 READ_ONCE(rdp->gp_seq_needed)))
918 pr_info("\tcpu %d ->gp_seq_needed %ld\n",
919 cpu, (long)data_race(READ_ONCE(rdp->gp_seq_needed)));
922 for_each_possible_cpu(cpu) {
923 rdp = per_cpu_ptr(&rcu_data, cpu);
924 cbs += data_race(READ_ONCE(rdp->n_cbs_invoked));
925 if (rcu_segcblist_is_offloaded(&rdp->cblist))
926 show_rcu_nocb_state(rdp);
928 pr_info("RCU callbacks invoked since boot: %lu\n", cbs);
929 show_rcu_tasks_gp_kthreads();
931 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
934 * This function checks for grace-period requests that fail to motivate
935 * RCU to come out of its idle mode.
937 static void rcu_check_gp_start_stall(struct rcu_node *rnp, struct rcu_data *rdp,
938 const unsigned long gpssdelay)
942 struct rcu_node *rnp_root = rcu_get_root();
943 static atomic_t warned = ATOMIC_INIT(0);
945 if (!IS_ENABLED(CONFIG_PROVE_RCU) || rcu_gp_in_progress() ||
946 ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
947 READ_ONCE(rnp_root->gp_seq_needed)) ||
948 !smp_load_acquire(&rcu_state.gp_kthread)) // Get stable kthread.
950 j = jiffies; /* Expensive access, and in common case don't get here. */
951 if (time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
952 time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
953 atomic_read(&warned))
956 raw_spin_lock_irqsave_rcu_node(rnp, flags);
958 if (rcu_gp_in_progress() ||
959 ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
960 READ_ONCE(rnp_root->gp_seq_needed)) ||
961 time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
962 time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
963 atomic_read(&warned)) {
964 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
967 /* Hold onto the leaf lock to make others see warned==1. */
970 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
972 if (rcu_gp_in_progress() ||
973 ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
974 READ_ONCE(rnp_root->gp_seq_needed)) ||
975 time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
976 time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
977 atomic_xchg(&warned, 1)) {
979 /* irqs remain disabled. */
980 raw_spin_unlock_rcu_node(rnp_root);
981 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
986 raw_spin_unlock_rcu_node(rnp_root);
987 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
988 show_rcu_gp_kthreads();
992 * Do a forward-progress check for rcutorture. This is normally invoked
993 * due to an OOM event. The argument "j" gives the time period during
994 * which rcutorture would like progress to have been made.
996 void rcu_fwd_progress_check(unsigned long j)
1000 unsigned long max_cbs = 0;
1002 struct rcu_data *rdp;
1004 if (rcu_gp_in_progress()) {
1005 pr_info("%s: GP age %lu jiffies\n",
1006 __func__, jiffies - data_race(READ_ONCE(rcu_state.gp_start)));
1007 show_rcu_gp_kthreads();
1009 pr_info("%s: Last GP end %lu jiffies ago\n",
1010 __func__, jiffies - data_race(READ_ONCE(rcu_state.gp_end)));
1012 rdp = this_cpu_ptr(&rcu_data);
1013 rcu_check_gp_start_stall(rdp->mynode, rdp, j);
1016 for_each_possible_cpu(cpu) {
1017 cbs = rcu_get_n_cbs_cpu(cpu);
1021 pr_info("%s: callbacks", __func__);
1022 pr_cont(" %d: %lu", cpu, cbs);
1031 EXPORT_SYMBOL_GPL(rcu_fwd_progress_check);
1033 /* Commandeer a sysrq key to dump RCU's tree. */
1034 static bool sysrq_rcu;
1035 module_param(sysrq_rcu, bool, 0444);
1037 /* Dump grace-period-request information due to commandeered sysrq. */
1038 static void sysrq_show_rcu(int key)
1040 show_rcu_gp_kthreads();
1043 static const struct sysrq_key_op sysrq_rcudump_op = {
1044 .handler = sysrq_show_rcu,
1045 .help_msg = "show-rcu(y)",
1046 .action_msg = "Show RCU tree",
1047 .enable_mask = SYSRQ_ENABLE_DUMP,
1050 static int __init rcu_sysrq_init(void)
1053 return register_sysrq_key('y', &sysrq_rcudump_op);
1056 early_initcall(rcu_sysrq_init);