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 
rcu_exp_jiffies_till_stall_check(void)28 int rcu_exp_jiffies_till_stall_check(void)
29 {
30 	int cpu_stall_timeout = READ_ONCE(rcu_exp_cpu_stall_timeout);
31 	int exp_stall_delay_delta = 0;
32 	int till_stall_check;
33 
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());
37 
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, 21UL * HZ);
43 
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));
46 
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;
50 #endif
51 
52 	return till_stall_check + exp_stall_delay_delta;
53 }
54 EXPORT_SYMBOL_GPL(rcu_exp_jiffies_till_stall_check);
55 
56 /* Limit-check stall timeouts specified at boottime and runtime. */
rcu_jiffies_till_stall_check(void)57 int rcu_jiffies_till_stall_check(void)
58 {
59 	int till_stall_check = READ_ONCE(rcu_cpu_stall_timeout);
60 
61 	/*
62 	 * Limit check must be consistent with the Kconfig limits
63 	 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
64 	 */
65 	if (till_stall_check < 3) {
66 		WRITE_ONCE(rcu_cpu_stall_timeout, 3);
67 		till_stall_check = 3;
68 	} else if (till_stall_check > 300) {
69 		WRITE_ONCE(rcu_cpu_stall_timeout, 300);
70 		till_stall_check = 300;
71 	}
72 	return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
73 }
74 EXPORT_SYMBOL_GPL(rcu_jiffies_till_stall_check);
75 
76 /**
77  * rcu_gp_might_be_stalled - Is it likely that the grace period is stalled?
78  *
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.
84  *
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
87  * grace periods.
88  */
rcu_gp_might_be_stalled(void)89 bool rcu_gp_might_be_stalled(void)
90 {
91 	unsigned long d = rcu_jiffies_till_stall_check() / RCU_STALL_MIGHT_DIV;
92 	unsigned long j = jiffies;
93 
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())
98 		return false;
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
102 	// And ditto here.
103 	return !time_before(j, READ_ONCE(rcu_state.gp_start) + d);
104 }
105 
106 /* Don't do RCU CPU stall warnings during long sysrq printouts. */
rcu_sysrq_start(void)107 void rcu_sysrq_start(void)
108 {
109 	if (!rcu_cpu_stall_suppress)
110 		rcu_cpu_stall_suppress = 2;
111 }
112 
rcu_sysrq_end(void)113 void rcu_sysrq_end(void)
114 {
115 	if (rcu_cpu_stall_suppress == 2)
116 		rcu_cpu_stall_suppress = 0;
117 }
118 
119 /* Don't print RCU CPU stall warnings during a kernel panic. */
rcu_panic(struct notifier_block * this,unsigned long ev,void * ptr)120 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
121 {
122 	rcu_cpu_stall_suppress = 1;
123 	return NOTIFY_DONE;
124 }
125 
126 static struct notifier_block rcu_panic_block = {
127 	.notifier_call = rcu_panic,
128 };
129 
check_cpu_stall_init(void)130 static int __init check_cpu_stall_init(void)
131 {
132 	atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
133 	return 0;
134 }
135 early_initcall(check_cpu_stall_init);
136 
137 /* If so specified via sysctl, panic, yielding cleaner stall-warning output. */
panic_on_rcu_stall(void)138 static void panic_on_rcu_stall(void)
139 {
140 	static int cpu_stall;
141 
142 	if (++cpu_stall < sysctl_max_rcu_stall_to_panic)
143 		return;
144 
145 	if (sysctl_panic_on_rcu_stall)
146 		panic("RCU Stall\n");
147 }
148 
149 /**
150  * rcu_cpu_stall_reset - restart stall-warning timeout for current grace period
151  *
152  * The caller must disable hard irqs.
153  */
rcu_cpu_stall_reset(void)154 void rcu_cpu_stall_reset(void)
155 {
156 	WRITE_ONCE(rcu_state.jiffies_stall,
157 		   jiffies + rcu_jiffies_till_stall_check());
158 }
159 
160 //////////////////////////////////////////////////////////////////////////////
161 //
162 // Interaction with RCU grace periods
163 
164 /* Start of new grace period, so record stall time (and forcing times). */
record_gp_stall_check_time(void)165 static void record_gp_stall_check_time(void)
166 {
167 	unsigned long j = jiffies;
168 	unsigned long j1;
169 
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);
176 }
177 
178 /* Zero ->ticks_this_gp and snapshot the number of RCU softirq handlers. */
zero_cpu_stall_ticks(struct rcu_data * rdp)179 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
180 {
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);
184 }
185 
186 /*
187  * If too much time has passed in the current grace period, and if
188  * so configured, go kick the relevant kthreads.
189  */
rcu_stall_kick_kthreads(void)190 static void rcu_stall_kick_kthreads(void)
191 {
192 	unsigned long j;
193 
194 	if (!READ_ONCE(rcu_kick_kthreads))
195 		return;
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",
200 			  rcu_state.name);
201 		rcu_ftrace_dump(DUMP_ALL);
202 		wake_up_process(rcu_state.gp_kthread);
203 		WRITE_ONCE(rcu_state.jiffies_kick_kthreads, j + HZ);
204 	}
205 }
206 
207 /*
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.
211  */
rcu_iw_handler(struct irq_work * iwp)212 static void rcu_iw_handler(struct irq_work *iwp)
213 {
214 	struct rcu_data *rdp;
215 	struct rcu_node *rnp;
216 
217 	rdp = container_of(iwp, struct rcu_data, rcu_iw);
218 	rnp = rdp->mynode;
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;
223 	}
224 	raw_spin_unlock_rcu_node(rnp);
225 }
226 
227 //////////////////////////////////////////////////////////////////////////////
228 //
229 // Printing RCU CPU stall warnings
230 
231 #ifdef CONFIG_PREEMPT_RCU
232 
233 /*
234  * Dump detailed information for all tasks blocking the current RCU
235  * grace period on the specified rcu_node structure.
236  */
rcu_print_detail_task_stall_rnp(struct rcu_node * rnp)237 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
238 {
239 	unsigned long flags;
240 	struct task_struct *t;
241 
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);
245 		return;
246 	}
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) {
250 		/*
251 		 * We could be printing a lot while holding a spinlock.
252 		 * Avoid triggering hard lockup.
253 		 */
254 		touch_nmi_watchdog();
255 		sched_show_task(t);
256 	}
257 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
258 }
259 
260 // Communicate task state back to the RCU CPU stall warning request.
261 struct rcu_stall_chk_rdr {
262 	int nesting;
263 	union rcu_special rs;
264 	bool on_blkd_list;
265 };
266 
267 /*
268  * Report out the state of a not-running task that is stalling the
269  * current RCU grace period.
270  */
check_slow_task(struct task_struct * t,void * arg)271 static int check_slow_task(struct task_struct *t, void *arg)
272 {
273 	struct rcu_stall_chk_rdr *rscrp = arg;
274 
275 	if (task_curr(t))
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);
280 	return 0;
281 }
282 
283 /*
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.
286  */
rcu_print_task_stall(struct rcu_node * rnp,unsigned long flags)287 static int rcu_print_task_stall(struct rcu_node *rnp, unsigned long flags)
288 	__releases(rnp->lock)
289 {
290 	int i = 0;
291 	int ndetected = 0;
292 	struct rcu_stall_chk_rdr rscr;
293 	struct task_struct *t;
294 	struct task_struct *ts[8];
295 
296 	lockdep_assert_irqs_disabled();
297 	if (!rcu_preempt_blocked_readers_cgp(rnp)) {
298 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
299 		return 0;
300 	}
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) {
306 		get_task_struct(t);
307 		ts[i++] = t;
308 		if (i >= ARRAY_SIZE(ts))
309 			break;
310 	}
311 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
312 	while (i) {
313 		t = ts[--i];
314 		if (task_call_func(t, check_slow_task, &rscr))
315 			pr_cont(" P%d", t->pid);
316 		else
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();
324 		put_task_struct(t);
325 		ndetected++;
326 	}
327 	pr_cont("\n");
328 	return ndetected;
329 }
330 
331 #else /* #ifdef CONFIG_PREEMPT_RCU */
332 
333 /*
334  * Because preemptible RCU does not exist, we never have to check for
335  * tasks blocked within RCU read-side critical sections.
336  */
rcu_print_detail_task_stall_rnp(struct rcu_node * rnp)337 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
338 {
339 }
340 
341 /*
342  * Because preemptible RCU does not exist, we never have to check for
343  * tasks blocked within RCU read-side critical sections.
344  */
rcu_print_task_stall(struct rcu_node * rnp,unsigned long flags)345 static int rcu_print_task_stall(struct rcu_node *rnp, unsigned long flags)
346 	__releases(rnp->lock)
347 {
348 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
349 	return 0;
350 }
351 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
352 
353 /*
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.
358  */
rcu_dump_cpu_stacks(void)359 static void rcu_dump_cpu_stacks(void)
360 {
361 	int cpu;
362 	unsigned long flags;
363 	struct rcu_node *rnp;
364 
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);
371 				else if (!trigger_single_cpu_backtrace(cpu))
372 					dump_cpu_task(cpu);
373 			}
374 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
375 	}
376 }
377 
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",
388 };
389 
390 /*
391  * Convert a ->gp_state value to a character string.
392  */
gp_state_getname(short gs)393 static const char *gp_state_getname(short gs)
394 {
395 	if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
396 		return "???";
397 	return gp_state_names[gs];
398 }
399 
400 /* Is the RCU grace-period kthread being starved of CPU time? */
rcu_is_gp_kthread_starving(unsigned long * jp)401 static bool rcu_is_gp_kthread_starving(unsigned long *jp)
402 {
403 	unsigned long j = jiffies - READ_ONCE(rcu_state.gp_activity);
404 
405 	if (jp)
406 		*jp = j;
407 	return j > 2 * HZ;
408 }
409 
rcu_is_rcuc_kthread_starving(struct rcu_data * rdp,unsigned long * jp)410 static bool rcu_is_rcuc_kthread_starving(struct rcu_data *rdp, unsigned long *jp)
411 {
412 	unsigned long j = jiffies - READ_ONCE(rdp->rcuc_activity);
413 
414 	if (jp)
415 		*jp = j;
416 	return j > 2 * HZ;
417 }
418 
419 /*
420  * Print out diagnostic information for the specified stalled CPU.
421  *
422  * If the specified CPU is aware of the current RCU grace period, then
423  * print the number of scheduling clock interrupts the CPU has taken
424  * during the time that it has been aware.  Otherwise, print the number
425  * of RCU grace periods that this CPU is ignorant of, for example, "1"
426  * if the CPU was aware of the previous grace period.
427  *
428  * Also print out idle info.
429  */
print_cpu_stall_info(int cpu)430 static void print_cpu_stall_info(int cpu)
431 {
432 	unsigned long delta;
433 	bool falsepositive;
434 	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
435 	char *ticks_title;
436 	unsigned long ticks_value;
437 
438 	/*
439 	 * We could be printing a lot while holding a spinlock.  Avoid
440 	 * triggering hard lockup.
441 	 */
442 	touch_nmi_watchdog();
443 
444 	ticks_value = rcu_seq_ctr(rcu_state.gp_seq - rdp->gp_seq);
445 	if (ticks_value) {
446 		ticks_title = "GPs behind";
447 	} else {
448 		ticks_title = "ticks this GP";
449 		ticks_value = rdp->ticks_this_gp;
450 	}
451 	delta = rcu_seq_ctr(rdp->mynode->gp_seq - rdp->rcu_iw_gp_seq);
452 	falsepositive = rcu_is_gp_kthread_starving(NULL) &&
453 			rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp));
454 	pr_err("\t%d-%c%c%c%c: (%lu %s) idle=%03x/%ld/%#lx softirq=%u/%u fqs=%ld %s\n",
455 	       cpu,
456 	       "O."[!!cpu_online(cpu)],
457 	       "o."[!!(rdp->grpmask & rdp->mynode->qsmaskinit)],
458 	       "N."[!!(rdp->grpmask & rdp->mynode->qsmaskinitnext)],
459 	       !IS_ENABLED(CONFIG_IRQ_WORK) ? '?' :
460 			rdp->rcu_iw_pending ? (int)min(delta, 9UL) + '0' :
461 				"!."[!delta],
462 	       ticks_value, ticks_title,
463 	       rcu_dynticks_snap(rdp) & 0xfff,
464 	       rdp->dynticks_nesting, rdp->dynticks_nmi_nesting,
465 	       rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
466 	       data_race(rcu_state.n_force_qs) - rcu_state.n_force_qs_gpstart,
467 	       falsepositive ? " (false positive?)" : "");
468 }
469 
rcuc_kthread_dump(struct rcu_data * rdp)470 static void rcuc_kthread_dump(struct rcu_data *rdp)
471 {
472 	int cpu;
473 	unsigned long j;
474 	struct task_struct *rcuc;
475 
476 	rcuc = rdp->rcu_cpu_kthread_task;
477 	if (!rcuc)
478 		return;
479 
480 	cpu = task_cpu(rcuc);
481 	if (cpu_is_offline(cpu) || idle_cpu(cpu))
482 		return;
483 
484 	if (!rcu_is_rcuc_kthread_starving(rdp, &j))
485 		return;
486 
487 	pr_err("%s kthread starved for %ld jiffies\n", rcuc->comm, j);
488 	sched_show_task(rcuc);
489 	if (!trigger_single_cpu_backtrace(cpu))
490 		dump_cpu_task(cpu);
491 }
492 
493 /* Complain about starvation of grace-period kthread.  */
rcu_check_gp_kthread_starvation(void)494 static void rcu_check_gp_kthread_starvation(void)
495 {
496 	int cpu;
497 	struct task_struct *gpk = rcu_state.gp_kthread;
498 	unsigned long j;
499 
500 	if (rcu_is_gp_kthread_starving(&j)) {
501 		cpu = gpk ? task_cpu(gpk) : -1;
502 		pr_err("%s kthread starved for %ld jiffies! g%ld f%#x %s(%d) ->state=%#x ->cpu=%d\n",
503 		       rcu_state.name, j,
504 		       (long)rcu_seq_current(&rcu_state.gp_seq),
505 		       data_race(READ_ONCE(rcu_state.gp_flags)),
506 		       gp_state_getname(rcu_state.gp_state),
507 		       data_race(READ_ONCE(rcu_state.gp_state)),
508 		       gpk ? data_race(READ_ONCE(gpk->__state)) : ~0, cpu);
509 		if (gpk) {
510 			pr_err("\tUnless %s kthread gets sufficient CPU time, OOM is now expected behavior.\n", rcu_state.name);
511 			pr_err("RCU grace-period kthread stack dump:\n");
512 			sched_show_task(gpk);
513 			if (cpu >= 0) {
514 				if (cpu_is_offline(cpu)) {
515 					pr_err("RCU GP kthread last ran on offline CPU %d.\n", cpu);
516 				} else  {
517 					pr_err("Stack dump where RCU GP kthread last ran:\n");
518 					if (!trigger_single_cpu_backtrace(cpu))
519 						dump_cpu_task(cpu);
520 				}
521 			}
522 			wake_up_process(gpk);
523 		}
524 	}
525 }
526 
527 /* Complain about missing wakeups from expired fqs wait timer */
rcu_check_gp_kthread_expired_fqs_timer(void)528 static void rcu_check_gp_kthread_expired_fqs_timer(void)
529 {
530 	struct task_struct *gpk = rcu_state.gp_kthread;
531 	short gp_state;
532 	unsigned long jiffies_fqs;
533 	int cpu;
534 
535 	/*
536 	 * Order reads of .gp_state and .jiffies_force_qs.
537 	 * Matching smp_wmb() is present in rcu_gp_fqs_loop().
538 	 */
539 	gp_state = smp_load_acquire(&rcu_state.gp_state);
540 	jiffies_fqs = READ_ONCE(rcu_state.jiffies_force_qs);
541 
542 	if (gp_state == RCU_GP_WAIT_FQS &&
543 	    time_after(jiffies, jiffies_fqs + RCU_STALL_MIGHT_MIN) &&
544 	    gpk && !READ_ONCE(gpk->on_rq)) {
545 		cpu = task_cpu(gpk);
546 		pr_err("%s kthread timer wakeup didn't happen for %ld jiffies! g%ld f%#x %s(%d) ->state=%#x\n",
547 		       rcu_state.name, (jiffies - jiffies_fqs),
548 		       (long)rcu_seq_current(&rcu_state.gp_seq),
549 		       data_race(rcu_state.gp_flags),
550 		       gp_state_getname(RCU_GP_WAIT_FQS), RCU_GP_WAIT_FQS,
551 		       data_race(READ_ONCE(gpk->__state)));
552 		pr_err("\tPossible timer handling issue on cpu=%d timer-softirq=%u\n",
553 		       cpu, kstat_softirqs_cpu(TIMER_SOFTIRQ, cpu));
554 	}
555 }
556 
print_other_cpu_stall(unsigned long gp_seq,unsigned long gps)557 static void print_other_cpu_stall(unsigned long gp_seq, unsigned long gps)
558 {
559 	int cpu;
560 	unsigned long flags;
561 	unsigned long gpa;
562 	unsigned long j;
563 	int ndetected = 0;
564 	struct rcu_node *rnp;
565 	long totqlen = 0;
566 
567 	lockdep_assert_irqs_disabled();
568 
569 	/* Kick and suppress, if so configured. */
570 	rcu_stall_kick_kthreads();
571 	if (rcu_stall_is_suppressed())
572 		return;
573 
574 	/*
575 	 * OK, time to rat on our buddy...
576 	 * See Documentation/RCU/stallwarn.rst for info on how to debug
577 	 * RCU CPU stall warnings.
578 	 */
579 	trace_rcu_stall_warning(rcu_state.name, TPS("StallDetected"));
580 	pr_err("INFO: %s detected stalls on CPUs/tasks:\n", rcu_state.name);
581 	rcu_for_each_leaf_node(rnp) {
582 		raw_spin_lock_irqsave_rcu_node(rnp, flags);
583 		if (rnp->qsmask != 0) {
584 			for_each_leaf_node_possible_cpu(rnp, cpu)
585 				if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
586 					print_cpu_stall_info(cpu);
587 					ndetected++;
588 				}
589 		}
590 		ndetected += rcu_print_task_stall(rnp, flags); // Releases rnp->lock.
591 		lockdep_assert_irqs_disabled();
592 	}
593 
594 	for_each_possible_cpu(cpu)
595 		totqlen += rcu_get_n_cbs_cpu(cpu);
596 	pr_cont("\t(detected by %d, t=%ld jiffies, g=%ld, q=%lu ncpus=%d)\n",
597 	       smp_processor_id(), (long)(jiffies - gps),
598 	       (long)rcu_seq_current(&rcu_state.gp_seq), totqlen, rcu_state.n_online_cpus);
599 	if (ndetected) {
600 		rcu_dump_cpu_stacks();
601 
602 		/* Complain about tasks blocking the grace period. */
603 		rcu_for_each_leaf_node(rnp)
604 			rcu_print_detail_task_stall_rnp(rnp);
605 	} else {
606 		if (rcu_seq_current(&rcu_state.gp_seq) != gp_seq) {
607 			pr_err("INFO: Stall ended before state dump start\n");
608 		} else {
609 			j = jiffies;
610 			gpa = data_race(READ_ONCE(rcu_state.gp_activity));
611 			pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
612 			       rcu_state.name, j - gpa, j, gpa,
613 			       data_race(READ_ONCE(jiffies_till_next_fqs)),
614 			       data_race(READ_ONCE(rcu_get_root()->qsmask)));
615 		}
616 	}
617 	/* Rewrite if needed in case of slow consoles. */
618 	if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
619 		WRITE_ONCE(rcu_state.jiffies_stall,
620 			   jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
621 
622 	rcu_check_gp_kthread_expired_fqs_timer();
623 	rcu_check_gp_kthread_starvation();
624 
625 	panic_on_rcu_stall();
626 
627 	rcu_force_quiescent_state();  /* Kick them all. */
628 }
629 
print_cpu_stall(unsigned long gps)630 static void print_cpu_stall(unsigned long gps)
631 {
632 	int cpu;
633 	unsigned long flags;
634 	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
635 	struct rcu_node *rnp = rcu_get_root();
636 	long totqlen = 0;
637 
638 	lockdep_assert_irqs_disabled();
639 
640 	/* Kick and suppress, if so configured. */
641 	rcu_stall_kick_kthreads();
642 	if (rcu_stall_is_suppressed())
643 		return;
644 
645 	/*
646 	 * OK, time to rat on ourselves...
647 	 * See Documentation/RCU/stallwarn.rst for info on how to debug
648 	 * RCU CPU stall warnings.
649 	 */
650 	trace_rcu_stall_warning(rcu_state.name, TPS("SelfDetected"));
651 	pr_err("INFO: %s self-detected stall on CPU\n", rcu_state.name);
652 	raw_spin_lock_irqsave_rcu_node(rdp->mynode, flags);
653 	print_cpu_stall_info(smp_processor_id());
654 	raw_spin_unlock_irqrestore_rcu_node(rdp->mynode, flags);
655 	for_each_possible_cpu(cpu)
656 		totqlen += rcu_get_n_cbs_cpu(cpu);
657 	pr_cont("\t(t=%lu jiffies g=%ld q=%lu ncpus=%d)\n",
658 		jiffies - gps,
659 		(long)rcu_seq_current(&rcu_state.gp_seq), totqlen, rcu_state.n_online_cpus);
660 
661 	rcu_check_gp_kthread_expired_fqs_timer();
662 	rcu_check_gp_kthread_starvation();
663 
664 	if (!use_softirq)
665 		rcuc_kthread_dump(rdp);
666 
667 	rcu_dump_cpu_stacks();
668 
669 	raw_spin_lock_irqsave_rcu_node(rnp, flags);
670 	/* Rewrite if needed in case of slow consoles. */
671 	if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
672 		WRITE_ONCE(rcu_state.jiffies_stall,
673 			   jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
674 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
675 
676 	panic_on_rcu_stall();
677 
678 	/*
679 	 * Attempt to revive the RCU machinery by forcing a context switch.
680 	 *
681 	 * A context switch would normally allow the RCU state machine to make
682 	 * progress and it could be we're stuck in kernel space without context
683 	 * switches for an entirely unreasonable amount of time.
684 	 */
685 	set_tsk_need_resched(current);
686 	set_preempt_need_resched();
687 }
688 
check_cpu_stall(struct rcu_data * rdp)689 static void check_cpu_stall(struct rcu_data *rdp)
690 {
691 	bool didstall = false;
692 	unsigned long gs1;
693 	unsigned long gs2;
694 	unsigned long gps;
695 	unsigned long j;
696 	unsigned long jn;
697 	unsigned long js;
698 	struct rcu_node *rnp;
699 
700 	lockdep_assert_irqs_disabled();
701 	if ((rcu_stall_is_suppressed() && !READ_ONCE(rcu_kick_kthreads)) ||
702 	    !rcu_gp_in_progress())
703 		return;
704 	rcu_stall_kick_kthreads();
705 	j = jiffies;
706 
707 	/*
708 	 * Lots of memory barriers to reject false positives.
709 	 *
710 	 * The idea is to pick up rcu_state.gp_seq, then
711 	 * rcu_state.jiffies_stall, then rcu_state.gp_start, and finally
712 	 * another copy of rcu_state.gp_seq.  These values are updated in
713 	 * the opposite order with memory barriers (or equivalent) during
714 	 * grace-period initialization and cleanup.  Now, a false positive
715 	 * can occur if we get an new value of rcu_state.gp_start and a old
716 	 * value of rcu_state.jiffies_stall.  But given the memory barriers,
717 	 * the only way that this can happen is if one grace period ends
718 	 * and another starts between these two fetches.  This is detected
719 	 * by comparing the second fetch of rcu_state.gp_seq with the
720 	 * previous fetch from rcu_state.gp_seq.
721 	 *
722 	 * Given this check, comparisons of jiffies, rcu_state.jiffies_stall,
723 	 * and rcu_state.gp_start suffice to forestall false positives.
724 	 */
725 	gs1 = READ_ONCE(rcu_state.gp_seq);
726 	smp_rmb(); /* Pick up ->gp_seq first... */
727 	js = READ_ONCE(rcu_state.jiffies_stall);
728 	smp_rmb(); /* ...then ->jiffies_stall before the rest... */
729 	gps = READ_ONCE(rcu_state.gp_start);
730 	smp_rmb(); /* ...and finally ->gp_start before ->gp_seq again. */
731 	gs2 = READ_ONCE(rcu_state.gp_seq);
732 	if (gs1 != gs2 ||
733 	    ULONG_CMP_LT(j, js) ||
734 	    ULONG_CMP_GE(gps, js))
735 		return; /* No stall or GP completed since entering function. */
736 	rnp = rdp->mynode;
737 	jn = jiffies + ULONG_MAX / 2;
738 	if (rcu_gp_in_progress() &&
739 	    (READ_ONCE(rnp->qsmask) & rdp->grpmask) &&
740 	    cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
741 
742 		/*
743 		 * If a virtual machine is stopped by the host it can look to
744 		 * the watchdog like an RCU stall. Check to see if the host
745 		 * stopped the vm.
746 		 */
747 		if (kvm_check_and_clear_guest_paused())
748 			return;
749 
750 		/* We haven't checked in, so go dump stack. */
751 		print_cpu_stall(gps);
752 		if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
753 			rcu_ftrace_dump(DUMP_ALL);
754 		didstall = true;
755 
756 	} else if (rcu_gp_in_progress() &&
757 		   ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY) &&
758 		   cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
759 
760 		/*
761 		 * If a virtual machine is stopped by the host it can look to
762 		 * the watchdog like an RCU stall. Check to see if the host
763 		 * stopped the vm.
764 		 */
765 		if (kvm_check_and_clear_guest_paused())
766 			return;
767 
768 		/* They had a few time units to dump stack, so complain. */
769 		print_other_cpu_stall(gs2, gps);
770 		if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
771 			rcu_ftrace_dump(DUMP_ALL);
772 		didstall = true;
773 	}
774 	if (didstall && READ_ONCE(rcu_state.jiffies_stall) == jn) {
775 		jn = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
776 		WRITE_ONCE(rcu_state.jiffies_stall, jn);
777 	}
778 }
779 
780 //////////////////////////////////////////////////////////////////////////////
781 //
782 // RCU forward-progress mechanisms, including of callback invocation.
783 
784 
785 /*
786  * Check to see if a failure to end RCU priority inversion was due to
787  * a CPU not passing through a quiescent state.  When this happens, there
788  * is nothing that RCU priority boosting can do to help, so we shouldn't
789  * count this as an RCU priority boosting failure.  A return of true says
790  * RCU priority boosting is to blame, and false says otherwise.  If false
791  * is returned, the first of the CPUs to blame is stored through cpup.
792  * If there was no CPU blocking the current grace period, but also nothing
793  * in need of being boosted, *cpup is set to -1.  This can happen in case
794  * of vCPU preemption while the last CPU is reporting its quiscent state,
795  * for example.
796  *
797  * If cpup is NULL, then a lockless quick check is carried out, suitable
798  * for high-rate usage.  On the other hand, if cpup is non-NULL, each
799  * rcu_node structure's ->lock is acquired, ruling out high-rate usage.
800  */
rcu_check_boost_fail(unsigned long gp_state,int * cpup)801 bool rcu_check_boost_fail(unsigned long gp_state, int *cpup)
802 {
803 	bool atb = false;
804 	int cpu;
805 	unsigned long flags;
806 	struct rcu_node *rnp;
807 
808 	rcu_for_each_leaf_node(rnp) {
809 		if (!cpup) {
810 			if (data_race(READ_ONCE(rnp->qsmask))) {
811 				return false;
812 			} else {
813 				if (READ_ONCE(rnp->gp_tasks))
814 					atb = true;
815 				continue;
816 			}
817 		}
818 		*cpup = -1;
819 		raw_spin_lock_irqsave_rcu_node(rnp, flags);
820 		if (rnp->gp_tasks)
821 			atb = true;
822 		if (!rnp->qsmask) {
823 			// No CPUs without quiescent states for this rnp.
824 			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
825 			continue;
826 		}
827 		// Find the first holdout CPU.
828 		for_each_leaf_node_possible_cpu(rnp, cpu) {
829 			if (rnp->qsmask & (1UL << (cpu - rnp->grplo))) {
830 				raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
831 				*cpup = cpu;
832 				return false;
833 			}
834 		}
835 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
836 	}
837 	// Can't blame CPUs, so must blame RCU priority boosting.
838 	return atb;
839 }
840 EXPORT_SYMBOL_GPL(rcu_check_boost_fail);
841 
842 /*
843  * Show the state of the grace-period kthreads.
844  */
show_rcu_gp_kthreads(void)845 void show_rcu_gp_kthreads(void)
846 {
847 	unsigned long cbs = 0;
848 	int cpu;
849 	unsigned long j;
850 	unsigned long ja;
851 	unsigned long jr;
852 	unsigned long js;
853 	unsigned long jw;
854 	struct rcu_data *rdp;
855 	struct rcu_node *rnp;
856 	struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
857 
858 	j = jiffies;
859 	ja = j - data_race(READ_ONCE(rcu_state.gp_activity));
860 	jr = j - data_race(READ_ONCE(rcu_state.gp_req_activity));
861 	js = j - data_race(READ_ONCE(rcu_state.gp_start));
862 	jw = j - data_race(READ_ONCE(rcu_state.gp_wake_time));
863 	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",
864 		rcu_state.name, gp_state_getname(rcu_state.gp_state),
865 		data_race(READ_ONCE(rcu_state.gp_state)),
866 		t ? data_race(READ_ONCE(t->__state)) : 0x1ffff, t ? t->rt_priority : 0xffU,
867 		js, ja, jr, jw, (long)data_race(READ_ONCE(rcu_state.gp_wake_seq)),
868 		(long)data_race(READ_ONCE(rcu_state.gp_seq)),
869 		(long)data_race(READ_ONCE(rcu_get_root()->gp_seq_needed)),
870 		data_race(READ_ONCE(rcu_state.gp_max)),
871 		data_race(READ_ONCE(rcu_state.gp_flags)));
872 	rcu_for_each_node_breadth_first(rnp) {
873 		if (ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq), READ_ONCE(rnp->gp_seq_needed)) &&
874 		    !data_race(READ_ONCE(rnp->qsmask)) && !data_race(READ_ONCE(rnp->boost_tasks)) &&
875 		    !data_race(READ_ONCE(rnp->exp_tasks)) && !data_race(READ_ONCE(rnp->gp_tasks)))
876 			continue;
877 		pr_info("\trcu_node %d:%d ->gp_seq %ld ->gp_seq_needed %ld ->qsmask %#lx %c%c%c%c ->n_boosts %ld\n",
878 			rnp->grplo, rnp->grphi,
879 			(long)data_race(READ_ONCE(rnp->gp_seq)),
880 			(long)data_race(READ_ONCE(rnp->gp_seq_needed)),
881 			data_race(READ_ONCE(rnp->qsmask)),
882 			".b"[!!data_race(READ_ONCE(rnp->boost_kthread_task))],
883 			".B"[!!data_race(READ_ONCE(rnp->boost_tasks))],
884 			".E"[!!data_race(READ_ONCE(rnp->exp_tasks))],
885 			".G"[!!data_race(READ_ONCE(rnp->gp_tasks))],
886 			data_race(READ_ONCE(rnp->n_boosts)));
887 		if (!rcu_is_leaf_node(rnp))
888 			continue;
889 		for_each_leaf_node_possible_cpu(rnp, cpu) {
890 			rdp = per_cpu_ptr(&rcu_data, cpu);
891 			if (READ_ONCE(rdp->gpwrap) ||
892 			    ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq),
893 					 READ_ONCE(rdp->gp_seq_needed)))
894 				continue;
895 			pr_info("\tcpu %d ->gp_seq_needed %ld\n",
896 				cpu, (long)data_race(READ_ONCE(rdp->gp_seq_needed)));
897 		}
898 	}
899 	for_each_possible_cpu(cpu) {
900 		rdp = per_cpu_ptr(&rcu_data, cpu);
901 		cbs += data_race(READ_ONCE(rdp->n_cbs_invoked));
902 		if (rcu_segcblist_is_offloaded(&rdp->cblist))
903 			show_rcu_nocb_state(rdp);
904 	}
905 	pr_info("RCU callbacks invoked since boot: %lu\n", cbs);
906 	show_rcu_tasks_gp_kthreads();
907 }
908 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
909 
910 /*
911  * This function checks for grace-period requests that fail to motivate
912  * RCU to come out of its idle mode.
913  */
rcu_check_gp_start_stall(struct rcu_node * rnp,struct rcu_data * rdp,const unsigned long gpssdelay)914 static void rcu_check_gp_start_stall(struct rcu_node *rnp, struct rcu_data *rdp,
915 				     const unsigned long gpssdelay)
916 {
917 	unsigned long flags;
918 	unsigned long j;
919 	struct rcu_node *rnp_root = rcu_get_root();
920 	static atomic_t warned = ATOMIC_INIT(0);
921 
922 	if (!IS_ENABLED(CONFIG_PROVE_RCU) || rcu_gp_in_progress() ||
923 	    ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
924 			 READ_ONCE(rnp_root->gp_seq_needed)) ||
925 	    !smp_load_acquire(&rcu_state.gp_kthread)) // Get stable kthread.
926 		return;
927 	j = jiffies; /* Expensive access, and in common case don't get here. */
928 	if (time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
929 	    time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
930 	    atomic_read(&warned))
931 		return;
932 
933 	raw_spin_lock_irqsave_rcu_node(rnp, flags);
934 	j = jiffies;
935 	if (rcu_gp_in_progress() ||
936 	    ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
937 			 READ_ONCE(rnp_root->gp_seq_needed)) ||
938 	    time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
939 	    time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
940 	    atomic_read(&warned)) {
941 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
942 		return;
943 	}
944 	/* Hold onto the leaf lock to make others see warned==1. */
945 
946 	if (rnp_root != rnp)
947 		raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
948 	j = jiffies;
949 	if (rcu_gp_in_progress() ||
950 	    ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
951 			 READ_ONCE(rnp_root->gp_seq_needed)) ||
952 	    time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
953 	    time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
954 	    atomic_xchg(&warned, 1)) {
955 		if (rnp_root != rnp)
956 			/* irqs remain disabled. */
957 			raw_spin_unlock_rcu_node(rnp_root);
958 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
959 		return;
960 	}
961 	WARN_ON(1);
962 	if (rnp_root != rnp)
963 		raw_spin_unlock_rcu_node(rnp_root);
964 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
965 	show_rcu_gp_kthreads();
966 }
967 
968 /*
969  * Do a forward-progress check for rcutorture.  This is normally invoked
970  * due to an OOM event.  The argument "j" gives the time period during
971  * which rcutorture would like progress to have been made.
972  */
rcu_fwd_progress_check(unsigned long j)973 void rcu_fwd_progress_check(unsigned long j)
974 {
975 	unsigned long cbs;
976 	int cpu;
977 	unsigned long max_cbs = 0;
978 	int max_cpu = -1;
979 	struct rcu_data *rdp;
980 
981 	if (rcu_gp_in_progress()) {
982 		pr_info("%s: GP age %lu jiffies\n",
983 			__func__, jiffies - data_race(READ_ONCE(rcu_state.gp_start)));
984 		show_rcu_gp_kthreads();
985 	} else {
986 		pr_info("%s: Last GP end %lu jiffies ago\n",
987 			__func__, jiffies - data_race(READ_ONCE(rcu_state.gp_end)));
988 		preempt_disable();
989 		rdp = this_cpu_ptr(&rcu_data);
990 		rcu_check_gp_start_stall(rdp->mynode, rdp, j);
991 		preempt_enable();
992 	}
993 	for_each_possible_cpu(cpu) {
994 		cbs = rcu_get_n_cbs_cpu(cpu);
995 		if (!cbs)
996 			continue;
997 		if (max_cpu < 0)
998 			pr_info("%s: callbacks", __func__);
999 		pr_cont(" %d: %lu", cpu, cbs);
1000 		if (cbs <= max_cbs)
1001 			continue;
1002 		max_cbs = cbs;
1003 		max_cpu = cpu;
1004 	}
1005 	if (max_cpu >= 0)
1006 		pr_cont("\n");
1007 }
1008 EXPORT_SYMBOL_GPL(rcu_fwd_progress_check);
1009 
1010 /* Commandeer a sysrq key to dump RCU's tree. */
1011 static bool sysrq_rcu;
1012 module_param(sysrq_rcu, bool, 0444);
1013 
1014 /* Dump grace-period-request information due to commandeered sysrq. */
sysrq_show_rcu(int key)1015 static void sysrq_show_rcu(int key)
1016 {
1017 	show_rcu_gp_kthreads();
1018 }
1019 
1020 static const struct sysrq_key_op sysrq_rcudump_op = {
1021 	.handler = sysrq_show_rcu,
1022 	.help_msg = "show-rcu(y)",
1023 	.action_msg = "Show RCU tree",
1024 	.enable_mask = SYSRQ_ENABLE_DUMP,
1025 };
1026 
rcu_sysrq_init(void)1027 static int __init rcu_sysrq_init(void)
1028 {
1029 	if (sysrq_rcu)
1030 		return register_sysrq_key('y', &sysrq_rcudump_op);
1031 	return 0;
1032 }
1033 early_initcall(rcu_sysrq_init);
1034