xref: /linux/kernel/rcu/tree_stall.h (revision 79e03f29fc84db76f78834b1f73ed329b377d753)
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 #include <linux/kvm_para.h>
11 
12 //////////////////////////////////////////////////////////////////////////////
13 //
14 // Controlling CPU stall warnings, including delay calculation.
15 
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;
19 
20 #ifdef CONFIG_PROVE_RCU
21 #define RCU_STALL_DELAY_DELTA		(5 * HZ)
22 #else
23 #define RCU_STALL_DELAY_DELTA		0
24 #endif
25 #define RCU_STALL_MIGHT_DIV		8
26 #define RCU_STALL_MIGHT_MIN		(2 * HZ)
27 
28 /* Limit-check stall timeouts specified at boottime and runtime. */
29 int rcu_jiffies_till_stall_check(void)
30 {
31 	int till_stall_check = READ_ONCE(rcu_cpu_stall_timeout);
32 
33 	/*
34 	 * Limit check must be consistent with the Kconfig limits
35 	 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
36 	 */
37 	if (till_stall_check < 3) {
38 		WRITE_ONCE(rcu_cpu_stall_timeout, 3);
39 		till_stall_check = 3;
40 	} else if (till_stall_check > 300) {
41 		WRITE_ONCE(rcu_cpu_stall_timeout, 300);
42 		till_stall_check = 300;
43 	}
44 	return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
45 }
46 EXPORT_SYMBOL_GPL(rcu_jiffies_till_stall_check);
47 
48 /**
49  * rcu_gp_might_be_stalled - Is it likely that the grace period is stalled?
50  *
51  * Returns @true if the current grace period is sufficiently old that
52  * it is reasonable to assume that it might be stalled.  This can be
53  * useful when deciding whether to allocate memory to enable RCU-mediated
54  * freeing on the one hand or just invoking synchronize_rcu() on the other.
55  * The latter is preferable when the grace period is stalled.
56  *
57  * Note that sampling of the .gp_start and .gp_seq fields must be done
58  * carefully to avoid false positives at the beginnings and ends of
59  * grace periods.
60  */
61 bool rcu_gp_might_be_stalled(void)
62 {
63 	unsigned long d = rcu_jiffies_till_stall_check() / RCU_STALL_MIGHT_DIV;
64 	unsigned long j = jiffies;
65 
66 	if (d < RCU_STALL_MIGHT_MIN)
67 		d = RCU_STALL_MIGHT_MIN;
68 	smp_mb(); // jiffies before .gp_seq to avoid false positives.
69 	if (!rcu_gp_in_progress())
70 		return false;
71 	// Long delays at this point avoids false positive, but a delay
72 	// of ULONG_MAX/4 jiffies voids your no-false-positive warranty.
73 	smp_mb(); // .gp_seq before second .gp_start
74 	// And ditto here.
75 	return !time_before(j, READ_ONCE(rcu_state.gp_start) + d);
76 }
77 
78 /* Don't do RCU CPU stall warnings during long sysrq printouts. */
79 void rcu_sysrq_start(void)
80 {
81 	if (!rcu_cpu_stall_suppress)
82 		rcu_cpu_stall_suppress = 2;
83 }
84 
85 void rcu_sysrq_end(void)
86 {
87 	if (rcu_cpu_stall_suppress == 2)
88 		rcu_cpu_stall_suppress = 0;
89 }
90 
91 /* Don't print RCU CPU stall warnings during a kernel panic. */
92 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
93 {
94 	rcu_cpu_stall_suppress = 1;
95 	return NOTIFY_DONE;
96 }
97 
98 static struct notifier_block rcu_panic_block = {
99 	.notifier_call = rcu_panic,
100 };
101 
102 static int __init check_cpu_stall_init(void)
103 {
104 	atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
105 	return 0;
106 }
107 early_initcall(check_cpu_stall_init);
108 
109 /* If so specified via sysctl, panic, yielding cleaner stall-warning output. */
110 static void panic_on_rcu_stall(void)
111 {
112 	static int cpu_stall;
113 
114 	if (++cpu_stall < sysctl_max_rcu_stall_to_panic)
115 		return;
116 
117 	if (sysctl_panic_on_rcu_stall)
118 		panic("RCU Stall\n");
119 }
120 
121 /**
122  * rcu_cpu_stall_reset - restart stall-warning timeout for current grace period
123  *
124  * The caller must disable hard irqs.
125  */
126 void rcu_cpu_stall_reset(void)
127 {
128 	WRITE_ONCE(rcu_state.jiffies_stall,
129 		   jiffies + rcu_jiffies_till_stall_check());
130 }
131 
132 //////////////////////////////////////////////////////////////////////////////
133 //
134 // Interaction with RCU grace periods
135 
136 /* Start of new grace period, so record stall time (and forcing times). */
137 static void record_gp_stall_check_time(void)
138 {
139 	unsigned long j = jiffies;
140 	unsigned long j1;
141 
142 	WRITE_ONCE(rcu_state.gp_start, j);
143 	j1 = rcu_jiffies_till_stall_check();
144 	smp_mb(); // ->gp_start before ->jiffies_stall and caller's ->gp_seq.
145 	WRITE_ONCE(rcu_state.jiffies_stall, j + j1);
146 	rcu_state.jiffies_resched = j + j1 / 2;
147 	rcu_state.n_force_qs_gpstart = READ_ONCE(rcu_state.n_force_qs);
148 }
149 
150 /* Zero ->ticks_this_gp and snapshot the number of RCU softirq handlers. */
151 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
152 {
153 	rdp->ticks_this_gp = 0;
154 	rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
155 	WRITE_ONCE(rdp->last_fqs_resched, jiffies);
156 }
157 
158 /*
159  * If too much time has passed in the current grace period, and if
160  * so configured, go kick the relevant kthreads.
161  */
162 static void rcu_stall_kick_kthreads(void)
163 {
164 	unsigned long j;
165 
166 	if (!READ_ONCE(rcu_kick_kthreads))
167 		return;
168 	j = READ_ONCE(rcu_state.jiffies_kick_kthreads);
169 	if (time_after(jiffies, j) && rcu_state.gp_kthread &&
170 	    (rcu_gp_in_progress() || READ_ONCE(rcu_state.gp_flags))) {
171 		WARN_ONCE(1, "Kicking %s grace-period kthread\n",
172 			  rcu_state.name);
173 		rcu_ftrace_dump(DUMP_ALL);
174 		wake_up_process(rcu_state.gp_kthread);
175 		WRITE_ONCE(rcu_state.jiffies_kick_kthreads, j + HZ);
176 	}
177 }
178 
179 /*
180  * Handler for the irq_work request posted about halfway into the RCU CPU
181  * stall timeout, and used to detect excessive irq disabling.  Set state
182  * appropriately, but just complain if there is unexpected state on entry.
183  */
184 static void rcu_iw_handler(struct irq_work *iwp)
185 {
186 	struct rcu_data *rdp;
187 	struct rcu_node *rnp;
188 
189 	rdp = container_of(iwp, struct rcu_data, rcu_iw);
190 	rnp = rdp->mynode;
191 	raw_spin_lock_rcu_node(rnp);
192 	if (!WARN_ON_ONCE(!rdp->rcu_iw_pending)) {
193 		rdp->rcu_iw_gp_seq = rnp->gp_seq;
194 		rdp->rcu_iw_pending = false;
195 	}
196 	raw_spin_unlock_rcu_node(rnp);
197 }
198 
199 //////////////////////////////////////////////////////////////////////////////
200 //
201 // Printing RCU CPU stall warnings
202 
203 #ifdef CONFIG_PREEMPT_RCU
204 
205 /*
206  * Dump detailed information for all tasks blocking the current RCU
207  * grace period on the specified rcu_node structure.
208  */
209 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
210 {
211 	unsigned long flags;
212 	struct task_struct *t;
213 
214 	raw_spin_lock_irqsave_rcu_node(rnp, flags);
215 	if (!rcu_preempt_blocked_readers_cgp(rnp)) {
216 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
217 		return;
218 	}
219 	t = list_entry(rnp->gp_tasks->prev,
220 		       struct task_struct, rcu_node_entry);
221 	list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
222 		/*
223 		 * We could be printing a lot while holding a spinlock.
224 		 * Avoid triggering hard lockup.
225 		 */
226 		touch_nmi_watchdog();
227 		sched_show_task(t);
228 	}
229 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
230 }
231 
232 // Communicate task state back to the RCU CPU stall warning request.
233 struct rcu_stall_chk_rdr {
234 	int nesting;
235 	union rcu_special rs;
236 	bool on_blkd_list;
237 };
238 
239 /*
240  * Report out the state of a not-running task that is stalling the
241  * current RCU grace period.
242  */
243 static int check_slow_task(struct task_struct *t, void *arg)
244 {
245 	struct rcu_stall_chk_rdr *rscrp = arg;
246 
247 	if (task_curr(t))
248 		return -EBUSY; // It is running, so decline to inspect it.
249 	rscrp->nesting = t->rcu_read_lock_nesting;
250 	rscrp->rs = t->rcu_read_unlock_special;
251 	rscrp->on_blkd_list = !list_empty(&t->rcu_node_entry);
252 	return 0;
253 }
254 
255 /*
256  * Scan the current list of tasks blocked within RCU read-side critical
257  * sections, printing out the tid of each of the first few of them.
258  */
259 static int rcu_print_task_stall(struct rcu_node *rnp, unsigned long flags)
260 	__releases(rnp->lock)
261 {
262 	int i = 0;
263 	int ndetected = 0;
264 	struct rcu_stall_chk_rdr rscr;
265 	struct task_struct *t;
266 	struct task_struct *ts[8];
267 
268 	lockdep_assert_irqs_disabled();
269 	if (!rcu_preempt_blocked_readers_cgp(rnp)) {
270 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
271 		return 0;
272 	}
273 	pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
274 	       rnp->level, rnp->grplo, rnp->grphi);
275 	t = list_entry(rnp->gp_tasks->prev,
276 		       struct task_struct, rcu_node_entry);
277 	list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
278 		get_task_struct(t);
279 		ts[i++] = t;
280 		if (i >= ARRAY_SIZE(ts))
281 			break;
282 	}
283 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
284 	while (i) {
285 		t = ts[--i];
286 		if (task_call_func(t, check_slow_task, &rscr))
287 			pr_cont(" P%d", t->pid);
288 		else
289 			pr_cont(" P%d/%d:%c%c%c%c",
290 				t->pid, rscr.nesting,
291 				".b"[rscr.rs.b.blocked],
292 				".q"[rscr.rs.b.need_qs],
293 				".e"[rscr.rs.b.exp_hint],
294 				".l"[rscr.on_blkd_list]);
295 		lockdep_assert_irqs_disabled();
296 		put_task_struct(t);
297 		ndetected++;
298 	}
299 	pr_cont("\n");
300 	return ndetected;
301 }
302 
303 #else /* #ifdef CONFIG_PREEMPT_RCU */
304 
305 /*
306  * Because preemptible RCU does not exist, we never have to check for
307  * tasks blocked within RCU read-side critical sections.
308  */
309 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
310 {
311 }
312 
313 /*
314  * Because preemptible RCU does not exist, we never have to check for
315  * tasks blocked within RCU read-side critical sections.
316  */
317 static int rcu_print_task_stall(struct rcu_node *rnp, unsigned long flags)
318 	__releases(rnp->lock)
319 {
320 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
321 	return 0;
322 }
323 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
324 
325 /*
326  * Dump stacks of all tasks running on stalled CPUs.  First try using
327  * NMIs, but fall back to manual remote stack tracing on architectures
328  * that don't support NMI-based stack dumps.  The NMI-triggered stack
329  * traces are more accurate because they are printed by the target CPU.
330  */
331 static void rcu_dump_cpu_stacks(void)
332 {
333 	int cpu;
334 	unsigned long flags;
335 	struct rcu_node *rnp;
336 
337 	rcu_for_each_leaf_node(rnp) {
338 		raw_spin_lock_irqsave_rcu_node(rnp, flags);
339 		for_each_leaf_node_possible_cpu(rnp, cpu)
340 			if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
341 				if (cpu_is_offline(cpu))
342 					pr_err("Offline CPU %d blocking current GP.\n", cpu);
343 				else if (!trigger_single_cpu_backtrace(cpu))
344 					dump_cpu_task(cpu);
345 			}
346 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
347 	}
348 }
349 
350 static const char * const gp_state_names[] = {
351 	[RCU_GP_IDLE] = "RCU_GP_IDLE",
352 	[RCU_GP_WAIT_GPS] = "RCU_GP_WAIT_GPS",
353 	[RCU_GP_DONE_GPS] = "RCU_GP_DONE_GPS",
354 	[RCU_GP_ONOFF] = "RCU_GP_ONOFF",
355 	[RCU_GP_INIT] = "RCU_GP_INIT",
356 	[RCU_GP_WAIT_FQS] = "RCU_GP_WAIT_FQS",
357 	[RCU_GP_DOING_FQS] = "RCU_GP_DOING_FQS",
358 	[RCU_GP_CLEANUP] = "RCU_GP_CLEANUP",
359 	[RCU_GP_CLEANED] = "RCU_GP_CLEANED",
360 };
361 
362 /*
363  * Convert a ->gp_state value to a character string.
364  */
365 static const char *gp_state_getname(short gs)
366 {
367 	if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
368 		return "???";
369 	return gp_state_names[gs];
370 }
371 
372 /* Is the RCU grace-period kthread being starved of CPU time? */
373 static bool rcu_is_gp_kthread_starving(unsigned long *jp)
374 {
375 	unsigned long j = jiffies - READ_ONCE(rcu_state.gp_activity);
376 
377 	if (jp)
378 		*jp = j;
379 	return j > 2 * HZ;
380 }
381 
382 /*
383  * Print out diagnostic information for the specified stalled CPU.
384  *
385  * If the specified CPU is aware of the current RCU grace period, then
386  * print the number of scheduling clock interrupts the CPU has taken
387  * during the time that it has been aware.  Otherwise, print the number
388  * of RCU grace periods that this CPU is ignorant of, for example, "1"
389  * if the CPU was aware of the previous grace period.
390  *
391  * Also print out idle info.
392  */
393 static void print_cpu_stall_info(int cpu)
394 {
395 	unsigned long delta;
396 	bool falsepositive;
397 	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
398 	char *ticks_title;
399 	unsigned long ticks_value;
400 
401 	/*
402 	 * We could be printing a lot while holding a spinlock.  Avoid
403 	 * triggering hard lockup.
404 	 */
405 	touch_nmi_watchdog();
406 
407 	ticks_value = rcu_seq_ctr(rcu_state.gp_seq - rdp->gp_seq);
408 	if (ticks_value) {
409 		ticks_title = "GPs behind";
410 	} else {
411 		ticks_title = "ticks this GP";
412 		ticks_value = rdp->ticks_this_gp;
413 	}
414 	delta = rcu_seq_ctr(rdp->mynode->gp_seq - rdp->rcu_iw_gp_seq);
415 	falsepositive = rcu_is_gp_kthread_starving(NULL) &&
416 			rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp));
417 	pr_err("\t%d-%c%c%c%c: (%lu %s) idle=%03x/%ld/%#lx softirq=%u/%u fqs=%ld %s\n",
418 	       cpu,
419 	       "O."[!!cpu_online(cpu)],
420 	       "o."[!!(rdp->grpmask & rdp->mynode->qsmaskinit)],
421 	       "N."[!!(rdp->grpmask & rdp->mynode->qsmaskinitnext)],
422 	       !IS_ENABLED(CONFIG_IRQ_WORK) ? '?' :
423 			rdp->rcu_iw_pending ? (int)min(delta, 9UL) + '0' :
424 				"!."[!delta],
425 	       ticks_value, ticks_title,
426 	       rcu_dynticks_snap(rdp) & 0xfff,
427 	       rdp->dynticks_nesting, rdp->dynticks_nmi_nesting,
428 	       rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
429 	       data_race(rcu_state.n_force_qs) - rcu_state.n_force_qs_gpstart,
430 	       falsepositive ? " (false positive?)" : "");
431 }
432 
433 /* Complain about starvation of grace-period kthread.  */
434 static void rcu_check_gp_kthread_starvation(void)
435 {
436 	int cpu;
437 	struct task_struct *gpk = rcu_state.gp_kthread;
438 	unsigned long j;
439 
440 	if (rcu_is_gp_kthread_starving(&j)) {
441 		cpu = gpk ? task_cpu(gpk) : -1;
442 		pr_err("%s kthread starved for %ld jiffies! g%ld f%#x %s(%d) ->state=%#x ->cpu=%d\n",
443 		       rcu_state.name, j,
444 		       (long)rcu_seq_current(&rcu_state.gp_seq),
445 		       data_race(READ_ONCE(rcu_state.gp_flags)),
446 		       gp_state_getname(rcu_state.gp_state),
447 		       data_race(READ_ONCE(rcu_state.gp_state)),
448 		       gpk ? data_race(READ_ONCE(gpk->__state)) : ~0, cpu);
449 		if (gpk) {
450 			pr_err("\tUnless %s kthread gets sufficient CPU time, OOM is now expected behavior.\n", rcu_state.name);
451 			pr_err("RCU grace-period kthread stack dump:\n");
452 			sched_show_task(gpk);
453 			if (cpu >= 0) {
454 				if (cpu_is_offline(cpu)) {
455 					pr_err("RCU GP kthread last ran on offline CPU %d.\n", cpu);
456 				} else  {
457 					pr_err("Stack dump where RCU GP kthread last ran:\n");
458 					if (!trigger_single_cpu_backtrace(cpu))
459 						dump_cpu_task(cpu);
460 				}
461 			}
462 			wake_up_process(gpk);
463 		}
464 	}
465 }
466 
467 /* Complain about missing wakeups from expired fqs wait timer */
468 static void rcu_check_gp_kthread_expired_fqs_timer(void)
469 {
470 	struct task_struct *gpk = rcu_state.gp_kthread;
471 	short gp_state;
472 	unsigned long jiffies_fqs;
473 	int cpu;
474 
475 	/*
476 	 * Order reads of .gp_state and .jiffies_force_qs.
477 	 * Matching smp_wmb() is present in rcu_gp_fqs_loop().
478 	 */
479 	gp_state = smp_load_acquire(&rcu_state.gp_state);
480 	jiffies_fqs = READ_ONCE(rcu_state.jiffies_force_qs);
481 
482 	if (gp_state == RCU_GP_WAIT_FQS &&
483 	    time_after(jiffies, jiffies_fqs + RCU_STALL_MIGHT_MIN) &&
484 	    gpk && !READ_ONCE(gpk->on_rq)) {
485 		cpu = task_cpu(gpk);
486 		pr_err("%s kthread timer wakeup didn't happen for %ld jiffies! g%ld f%#x %s(%d) ->state=%#x\n",
487 		       rcu_state.name, (jiffies - jiffies_fqs),
488 		       (long)rcu_seq_current(&rcu_state.gp_seq),
489 		       data_race(rcu_state.gp_flags),
490 		       gp_state_getname(RCU_GP_WAIT_FQS), RCU_GP_WAIT_FQS,
491 		       data_race(READ_ONCE(gpk->__state)));
492 		pr_err("\tPossible timer handling issue on cpu=%d timer-softirq=%u\n",
493 		       cpu, kstat_softirqs_cpu(TIMER_SOFTIRQ, cpu));
494 	}
495 }
496 
497 static void print_other_cpu_stall(unsigned long gp_seq, unsigned long gps)
498 {
499 	int cpu;
500 	unsigned long flags;
501 	unsigned long gpa;
502 	unsigned long j;
503 	int ndetected = 0;
504 	struct rcu_node *rnp;
505 	long totqlen = 0;
506 
507 	lockdep_assert_irqs_disabled();
508 
509 	/* Kick and suppress, if so configured. */
510 	rcu_stall_kick_kthreads();
511 	if (rcu_stall_is_suppressed())
512 		return;
513 
514 	/*
515 	 * OK, time to rat on our buddy...
516 	 * See Documentation/RCU/stallwarn.rst for info on how to debug
517 	 * RCU CPU stall warnings.
518 	 */
519 	trace_rcu_stall_warning(rcu_state.name, TPS("StallDetected"));
520 	pr_err("INFO: %s detected stalls on CPUs/tasks:\n", rcu_state.name);
521 	rcu_for_each_leaf_node(rnp) {
522 		raw_spin_lock_irqsave_rcu_node(rnp, flags);
523 		if (rnp->qsmask != 0) {
524 			for_each_leaf_node_possible_cpu(rnp, cpu)
525 				if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
526 					print_cpu_stall_info(cpu);
527 					ndetected++;
528 				}
529 		}
530 		ndetected += rcu_print_task_stall(rnp, flags); // Releases rnp->lock.
531 		lockdep_assert_irqs_disabled();
532 	}
533 
534 	for_each_possible_cpu(cpu)
535 		totqlen += rcu_get_n_cbs_cpu(cpu);
536 	pr_cont("\t(detected by %d, t=%ld jiffies, g=%ld, q=%lu)\n",
537 	       smp_processor_id(), (long)(jiffies - gps),
538 	       (long)rcu_seq_current(&rcu_state.gp_seq), totqlen);
539 	if (ndetected) {
540 		rcu_dump_cpu_stacks();
541 
542 		/* Complain about tasks blocking the grace period. */
543 		rcu_for_each_leaf_node(rnp)
544 			rcu_print_detail_task_stall_rnp(rnp);
545 	} else {
546 		if (rcu_seq_current(&rcu_state.gp_seq) != gp_seq) {
547 			pr_err("INFO: Stall ended before state dump start\n");
548 		} else {
549 			j = jiffies;
550 			gpa = data_race(READ_ONCE(rcu_state.gp_activity));
551 			pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
552 			       rcu_state.name, j - gpa, j, gpa,
553 			       data_race(READ_ONCE(jiffies_till_next_fqs)),
554 			       data_race(READ_ONCE(rcu_get_root()->qsmask)));
555 		}
556 	}
557 	/* Rewrite if needed in case of slow consoles. */
558 	if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
559 		WRITE_ONCE(rcu_state.jiffies_stall,
560 			   jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
561 
562 	rcu_check_gp_kthread_expired_fqs_timer();
563 	rcu_check_gp_kthread_starvation();
564 
565 	panic_on_rcu_stall();
566 
567 	rcu_force_quiescent_state();  /* Kick them all. */
568 }
569 
570 static void print_cpu_stall(unsigned long gps)
571 {
572 	int cpu;
573 	unsigned long flags;
574 	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
575 	struct rcu_node *rnp = rcu_get_root();
576 	long totqlen = 0;
577 
578 	lockdep_assert_irqs_disabled();
579 
580 	/* Kick and suppress, if so configured. */
581 	rcu_stall_kick_kthreads();
582 	if (rcu_stall_is_suppressed())
583 		return;
584 
585 	/*
586 	 * OK, time to rat on ourselves...
587 	 * See Documentation/RCU/stallwarn.rst for info on how to debug
588 	 * RCU CPU stall warnings.
589 	 */
590 	trace_rcu_stall_warning(rcu_state.name, TPS("SelfDetected"));
591 	pr_err("INFO: %s self-detected stall on CPU\n", rcu_state.name);
592 	raw_spin_lock_irqsave_rcu_node(rdp->mynode, flags);
593 	print_cpu_stall_info(smp_processor_id());
594 	raw_spin_unlock_irqrestore_rcu_node(rdp->mynode, flags);
595 	for_each_possible_cpu(cpu)
596 		totqlen += rcu_get_n_cbs_cpu(cpu);
597 	pr_cont("\t(t=%lu jiffies g=%ld q=%lu)\n",
598 		jiffies - gps,
599 		(long)rcu_seq_current(&rcu_state.gp_seq), totqlen);
600 
601 	rcu_check_gp_kthread_expired_fqs_timer();
602 	rcu_check_gp_kthread_starvation();
603 
604 	rcu_dump_cpu_stacks();
605 
606 	raw_spin_lock_irqsave_rcu_node(rnp, flags);
607 	/* Rewrite if needed in case of slow consoles. */
608 	if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
609 		WRITE_ONCE(rcu_state.jiffies_stall,
610 			   jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
611 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
612 
613 	panic_on_rcu_stall();
614 
615 	/*
616 	 * Attempt to revive the RCU machinery by forcing a context switch.
617 	 *
618 	 * A context switch would normally allow the RCU state machine to make
619 	 * progress and it could be we're stuck in kernel space without context
620 	 * switches for an entirely unreasonable amount of time.
621 	 */
622 	set_tsk_need_resched(current);
623 	set_preempt_need_resched();
624 }
625 
626 static void check_cpu_stall(struct rcu_data *rdp)
627 {
628 	bool didstall = false;
629 	unsigned long gs1;
630 	unsigned long gs2;
631 	unsigned long gps;
632 	unsigned long j;
633 	unsigned long jn;
634 	unsigned long js;
635 	struct rcu_node *rnp;
636 
637 	lockdep_assert_irqs_disabled();
638 	if ((rcu_stall_is_suppressed() && !READ_ONCE(rcu_kick_kthreads)) ||
639 	    !rcu_gp_in_progress())
640 		return;
641 	rcu_stall_kick_kthreads();
642 	j = jiffies;
643 
644 	/*
645 	 * Lots of memory barriers to reject false positives.
646 	 *
647 	 * The idea is to pick up rcu_state.gp_seq, then
648 	 * rcu_state.jiffies_stall, then rcu_state.gp_start, and finally
649 	 * another copy of rcu_state.gp_seq.  These values are updated in
650 	 * the opposite order with memory barriers (or equivalent) during
651 	 * grace-period initialization and cleanup.  Now, a false positive
652 	 * can occur if we get an new value of rcu_state.gp_start and a old
653 	 * value of rcu_state.jiffies_stall.  But given the memory barriers,
654 	 * the only way that this can happen is if one grace period ends
655 	 * and another starts between these two fetches.  This is detected
656 	 * by comparing the second fetch of rcu_state.gp_seq with the
657 	 * previous fetch from rcu_state.gp_seq.
658 	 *
659 	 * Given this check, comparisons of jiffies, rcu_state.jiffies_stall,
660 	 * and rcu_state.gp_start suffice to forestall false positives.
661 	 */
662 	gs1 = READ_ONCE(rcu_state.gp_seq);
663 	smp_rmb(); /* Pick up ->gp_seq first... */
664 	js = READ_ONCE(rcu_state.jiffies_stall);
665 	smp_rmb(); /* ...then ->jiffies_stall before the rest... */
666 	gps = READ_ONCE(rcu_state.gp_start);
667 	smp_rmb(); /* ...and finally ->gp_start before ->gp_seq again. */
668 	gs2 = READ_ONCE(rcu_state.gp_seq);
669 	if (gs1 != gs2 ||
670 	    ULONG_CMP_LT(j, js) ||
671 	    ULONG_CMP_GE(gps, js))
672 		return; /* No stall or GP completed since entering function. */
673 	rnp = rdp->mynode;
674 	jn = jiffies + ULONG_MAX / 2;
675 	if (rcu_gp_in_progress() &&
676 	    (READ_ONCE(rnp->qsmask) & rdp->grpmask) &&
677 	    cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
678 
679 		/*
680 		 * If a virtual machine is stopped by the host it can look to
681 		 * the watchdog like an RCU stall. Check to see if the host
682 		 * stopped the vm.
683 		 */
684 		if (kvm_check_and_clear_guest_paused())
685 			return;
686 
687 		/* We haven't checked in, so go dump stack. */
688 		print_cpu_stall(gps);
689 		if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
690 			rcu_ftrace_dump(DUMP_ALL);
691 		didstall = true;
692 
693 	} else if (rcu_gp_in_progress() &&
694 		   ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY) &&
695 		   cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
696 
697 		/*
698 		 * If a virtual machine is stopped by the host it can look to
699 		 * the watchdog like an RCU stall. Check to see if the host
700 		 * stopped the vm.
701 		 */
702 		if (kvm_check_and_clear_guest_paused())
703 			return;
704 
705 		/* They had a few time units to dump stack, so complain. */
706 		print_other_cpu_stall(gs2, gps);
707 		if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
708 			rcu_ftrace_dump(DUMP_ALL);
709 		didstall = true;
710 	}
711 	if (didstall && READ_ONCE(rcu_state.jiffies_stall) == jn) {
712 		jn = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
713 		WRITE_ONCE(rcu_state.jiffies_stall, jn);
714 	}
715 }
716 
717 //////////////////////////////////////////////////////////////////////////////
718 //
719 // RCU forward-progress mechanisms, including of callback invocation.
720 
721 
722 /*
723  * Check to see if a failure to end RCU priority inversion was due to
724  * a CPU not passing through a quiescent state.  When this happens, there
725  * is nothing that RCU priority boosting can do to help, so we shouldn't
726  * count this as an RCU priority boosting failure.  A return of true says
727  * RCU priority boosting is to blame, and false says otherwise.  If false
728  * is returned, the first of the CPUs to blame is stored through cpup.
729  * If there was no CPU blocking the current grace period, but also nothing
730  * in need of being boosted, *cpup is set to -1.  This can happen in case
731  * of vCPU preemption while the last CPU is reporting its quiscent state,
732  * for example.
733  *
734  * If cpup is NULL, then a lockless quick check is carried out, suitable
735  * for high-rate usage.  On the other hand, if cpup is non-NULL, each
736  * rcu_node structure's ->lock is acquired, ruling out high-rate usage.
737  */
738 bool rcu_check_boost_fail(unsigned long gp_state, int *cpup)
739 {
740 	bool atb = false;
741 	int cpu;
742 	unsigned long flags;
743 	struct rcu_node *rnp;
744 
745 	rcu_for_each_leaf_node(rnp) {
746 		if (!cpup) {
747 			if (data_race(READ_ONCE(rnp->qsmask))) {
748 				return false;
749 			} else {
750 				if (READ_ONCE(rnp->gp_tasks))
751 					atb = true;
752 				continue;
753 			}
754 		}
755 		*cpup = -1;
756 		raw_spin_lock_irqsave_rcu_node(rnp, flags);
757 		if (rnp->gp_tasks)
758 			atb = true;
759 		if (!rnp->qsmask) {
760 			// No CPUs without quiescent states for this rnp.
761 			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
762 			continue;
763 		}
764 		// Find the first holdout CPU.
765 		for_each_leaf_node_possible_cpu(rnp, cpu) {
766 			if (rnp->qsmask & (1UL << (cpu - rnp->grplo))) {
767 				raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
768 				*cpup = cpu;
769 				return false;
770 			}
771 		}
772 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
773 	}
774 	// Can't blame CPUs, so must blame RCU priority boosting.
775 	return atb;
776 }
777 EXPORT_SYMBOL_GPL(rcu_check_boost_fail);
778 
779 /*
780  * Show the state of the grace-period kthreads.
781  */
782 void show_rcu_gp_kthreads(void)
783 {
784 	unsigned long cbs = 0;
785 	int cpu;
786 	unsigned long j;
787 	unsigned long ja;
788 	unsigned long jr;
789 	unsigned long js;
790 	unsigned long jw;
791 	struct rcu_data *rdp;
792 	struct rcu_node *rnp;
793 	struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
794 
795 	j = jiffies;
796 	ja = j - data_race(READ_ONCE(rcu_state.gp_activity));
797 	jr = j - data_race(READ_ONCE(rcu_state.gp_req_activity));
798 	js = j - data_race(READ_ONCE(rcu_state.gp_start));
799 	jw = j - data_race(READ_ONCE(rcu_state.gp_wake_time));
800 	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",
801 		rcu_state.name, gp_state_getname(rcu_state.gp_state),
802 		data_race(READ_ONCE(rcu_state.gp_state)),
803 		t ? data_race(READ_ONCE(t->__state)) : 0x1ffff, t ? t->rt_priority : 0xffU,
804 		js, ja, jr, jw, (long)data_race(READ_ONCE(rcu_state.gp_wake_seq)),
805 		(long)data_race(READ_ONCE(rcu_state.gp_seq)),
806 		(long)data_race(READ_ONCE(rcu_get_root()->gp_seq_needed)),
807 		data_race(READ_ONCE(rcu_state.gp_max)),
808 		data_race(READ_ONCE(rcu_state.gp_flags)));
809 	rcu_for_each_node_breadth_first(rnp) {
810 		if (ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq), READ_ONCE(rnp->gp_seq_needed)) &&
811 		    !data_race(READ_ONCE(rnp->qsmask)) && !data_race(READ_ONCE(rnp->boost_tasks)) &&
812 		    !data_race(READ_ONCE(rnp->exp_tasks)) && !data_race(READ_ONCE(rnp->gp_tasks)))
813 			continue;
814 		pr_info("\trcu_node %d:%d ->gp_seq %ld ->gp_seq_needed %ld ->qsmask %#lx %c%c%c%c ->n_boosts %ld\n",
815 			rnp->grplo, rnp->grphi,
816 			(long)data_race(READ_ONCE(rnp->gp_seq)),
817 			(long)data_race(READ_ONCE(rnp->gp_seq_needed)),
818 			data_race(READ_ONCE(rnp->qsmask)),
819 			".b"[!!data_race(READ_ONCE(rnp->boost_kthread_task))],
820 			".B"[!!data_race(READ_ONCE(rnp->boost_tasks))],
821 			".E"[!!data_race(READ_ONCE(rnp->exp_tasks))],
822 			".G"[!!data_race(READ_ONCE(rnp->gp_tasks))],
823 			data_race(READ_ONCE(rnp->n_boosts)));
824 		if (!rcu_is_leaf_node(rnp))
825 			continue;
826 		for_each_leaf_node_possible_cpu(rnp, cpu) {
827 			rdp = per_cpu_ptr(&rcu_data, cpu);
828 			if (READ_ONCE(rdp->gpwrap) ||
829 			    ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq),
830 					 READ_ONCE(rdp->gp_seq_needed)))
831 				continue;
832 			pr_info("\tcpu %d ->gp_seq_needed %ld\n",
833 				cpu, (long)data_race(READ_ONCE(rdp->gp_seq_needed)));
834 		}
835 	}
836 	for_each_possible_cpu(cpu) {
837 		rdp = per_cpu_ptr(&rcu_data, cpu);
838 		cbs += data_race(READ_ONCE(rdp->n_cbs_invoked));
839 		if (rcu_segcblist_is_offloaded(&rdp->cblist))
840 			show_rcu_nocb_state(rdp);
841 	}
842 	pr_info("RCU callbacks invoked since boot: %lu\n", cbs);
843 	show_rcu_tasks_gp_kthreads();
844 }
845 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
846 
847 /*
848  * This function checks for grace-period requests that fail to motivate
849  * RCU to come out of its idle mode.
850  */
851 static void rcu_check_gp_start_stall(struct rcu_node *rnp, struct rcu_data *rdp,
852 				     const unsigned long gpssdelay)
853 {
854 	unsigned long flags;
855 	unsigned long j;
856 	struct rcu_node *rnp_root = rcu_get_root();
857 	static atomic_t warned = ATOMIC_INIT(0);
858 
859 	if (!IS_ENABLED(CONFIG_PROVE_RCU) || rcu_gp_in_progress() ||
860 	    ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
861 			 READ_ONCE(rnp_root->gp_seq_needed)) ||
862 	    !smp_load_acquire(&rcu_state.gp_kthread)) // Get stable kthread.
863 		return;
864 	j = jiffies; /* Expensive access, and in common case don't get here. */
865 	if (time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
866 	    time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
867 	    atomic_read(&warned))
868 		return;
869 
870 	raw_spin_lock_irqsave_rcu_node(rnp, flags);
871 	j = jiffies;
872 	if (rcu_gp_in_progress() ||
873 	    ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
874 			 READ_ONCE(rnp_root->gp_seq_needed)) ||
875 	    time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
876 	    time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
877 	    atomic_read(&warned)) {
878 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
879 		return;
880 	}
881 	/* Hold onto the leaf lock to make others see warned==1. */
882 
883 	if (rnp_root != rnp)
884 		raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
885 	j = jiffies;
886 	if (rcu_gp_in_progress() ||
887 	    ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
888 			 READ_ONCE(rnp_root->gp_seq_needed)) ||
889 	    time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
890 	    time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
891 	    atomic_xchg(&warned, 1)) {
892 		if (rnp_root != rnp)
893 			/* irqs remain disabled. */
894 			raw_spin_unlock_rcu_node(rnp_root);
895 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
896 		return;
897 	}
898 	WARN_ON(1);
899 	if (rnp_root != rnp)
900 		raw_spin_unlock_rcu_node(rnp_root);
901 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
902 	show_rcu_gp_kthreads();
903 }
904 
905 /*
906  * Do a forward-progress check for rcutorture.  This is normally invoked
907  * due to an OOM event.  The argument "j" gives the time period during
908  * which rcutorture would like progress to have been made.
909  */
910 void rcu_fwd_progress_check(unsigned long j)
911 {
912 	unsigned long cbs;
913 	int cpu;
914 	unsigned long max_cbs = 0;
915 	int max_cpu = -1;
916 	struct rcu_data *rdp;
917 
918 	if (rcu_gp_in_progress()) {
919 		pr_info("%s: GP age %lu jiffies\n",
920 			__func__, jiffies - data_race(READ_ONCE(rcu_state.gp_start)));
921 		show_rcu_gp_kthreads();
922 	} else {
923 		pr_info("%s: Last GP end %lu jiffies ago\n",
924 			__func__, jiffies - data_race(READ_ONCE(rcu_state.gp_end)));
925 		preempt_disable();
926 		rdp = this_cpu_ptr(&rcu_data);
927 		rcu_check_gp_start_stall(rdp->mynode, rdp, j);
928 		preempt_enable();
929 	}
930 	for_each_possible_cpu(cpu) {
931 		cbs = rcu_get_n_cbs_cpu(cpu);
932 		if (!cbs)
933 			continue;
934 		if (max_cpu < 0)
935 			pr_info("%s: callbacks", __func__);
936 		pr_cont(" %d: %lu", cpu, cbs);
937 		if (cbs <= max_cbs)
938 			continue;
939 		max_cbs = cbs;
940 		max_cpu = cpu;
941 	}
942 	if (max_cpu >= 0)
943 		pr_cont("\n");
944 }
945 EXPORT_SYMBOL_GPL(rcu_fwd_progress_check);
946 
947 /* Commandeer a sysrq key to dump RCU's tree. */
948 static bool sysrq_rcu;
949 module_param(sysrq_rcu, bool, 0444);
950 
951 /* Dump grace-period-request information due to commandeered sysrq. */
952 static void sysrq_show_rcu(int key)
953 {
954 	show_rcu_gp_kthreads();
955 }
956 
957 static const struct sysrq_key_op sysrq_rcudump_op = {
958 	.handler = sysrq_show_rcu,
959 	.help_msg = "show-rcu(y)",
960 	.action_msg = "Show RCU tree",
961 	.enable_mask = SYSRQ_ENABLE_DUMP,
962 };
963 
964 static int __init rcu_sysrq_init(void)
965 {
966 	if (sysrq_rcu)
967 		return register_sysrq_key('y', &sysrq_rcudump_op);
968 	return 0;
969 }
970 early_initcall(rcu_sysrq_init);
971