xref: /OK3568_Linux_fs/kernel/kernel/rcu/tree_stall.h (revision 4882a59341e53eb6f0b4789bf948001014eff981)
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 = CONFIG_BOOTPARAM_RCU_STALL_PANIC_VALUE;
18 ATOMIC_NOTIFIER_HEAD(rcu_stall_notifier_list);
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. */
rcu_jiffies_till_stall_check(void)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  */
rcu_gp_might_be_stalled(void)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. */
rcu_sysrq_start(void)79 void rcu_sysrq_start(void)
80 {
81 	if (!rcu_cpu_stall_suppress)
82 		rcu_cpu_stall_suppress = 2;
83 }
84 
rcu_sysrq_end(void)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. */
rcu_panic(struct notifier_block * this,unsigned long ev,void * ptr)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 
check_cpu_stall_init(void)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. */
panic_on_rcu_stall(void)110 static void panic_on_rcu_stall(void)
111 {
112 	if (sysctl_panic_on_rcu_stall)
113 		panic("RCU Stall\n");
114 }
115 
116 /**
117  * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
118  *
119  * Set the stall-warning timeout way off into the future, thus preventing
120  * any RCU CPU stall-warning messages from appearing in the current set of
121  * RCU grace periods.
122  *
123  * The caller must disable hard irqs.
124  */
rcu_cpu_stall_reset(void)125 void rcu_cpu_stall_reset(void)
126 {
127 	WRITE_ONCE(rcu_state.jiffies_stall, jiffies + ULONG_MAX / 2);
128 }
129 
130 //////////////////////////////////////////////////////////////////////////////
131 //
132 // Interaction with RCU grace periods
133 
134 /* Start of new grace period, so record stall time (and forcing times). */
record_gp_stall_check_time(void)135 static void record_gp_stall_check_time(void)
136 {
137 	unsigned long j = jiffies;
138 	unsigned long j1;
139 
140 	WRITE_ONCE(rcu_state.gp_start, j);
141 	j1 = rcu_jiffies_till_stall_check();
142 	smp_mb(); // ->gp_start before ->jiffies_stall and caller's ->gp_seq.
143 	WRITE_ONCE(rcu_state.jiffies_stall, j + j1);
144 	rcu_state.jiffies_resched = j + j1 / 2;
145 	rcu_state.n_force_qs_gpstart = READ_ONCE(rcu_state.n_force_qs);
146 }
147 
148 /* Zero ->ticks_this_gp and snapshot the number of RCU softirq handlers. */
zero_cpu_stall_ticks(struct rcu_data * rdp)149 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
150 {
151 	rdp->ticks_this_gp = 0;
152 	rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
153 	WRITE_ONCE(rdp->last_fqs_resched, jiffies);
154 }
155 
156 /*
157  * If too much time has passed in the current grace period, and if
158  * so configured, go kick the relevant kthreads.
159  */
rcu_stall_kick_kthreads(void)160 static void rcu_stall_kick_kthreads(void)
161 {
162 	unsigned long j;
163 
164 	if (!READ_ONCE(rcu_kick_kthreads))
165 		return;
166 	j = READ_ONCE(rcu_state.jiffies_kick_kthreads);
167 	if (time_after(jiffies, j) && rcu_state.gp_kthread &&
168 	    (rcu_gp_in_progress() || READ_ONCE(rcu_state.gp_flags))) {
169 		WARN_ONCE(1, "Kicking %s grace-period kthread\n",
170 			  rcu_state.name);
171 		rcu_ftrace_dump(DUMP_ALL);
172 		wake_up_process(rcu_state.gp_kthread);
173 		WRITE_ONCE(rcu_state.jiffies_kick_kthreads, j + HZ);
174 	}
175 }
176 
177 /*
178  * Handler for the irq_work request posted about halfway into the RCU CPU
179  * stall timeout, and used to detect excessive irq disabling.  Set state
180  * appropriately, but just complain if there is unexpected state on entry.
181  */
rcu_iw_handler(struct irq_work * iwp)182 static void rcu_iw_handler(struct irq_work *iwp)
183 {
184 	struct rcu_data *rdp;
185 	struct rcu_node *rnp;
186 
187 	rdp = container_of(iwp, struct rcu_data, rcu_iw);
188 	rnp = rdp->mynode;
189 	raw_spin_lock_rcu_node(rnp);
190 	if (!WARN_ON_ONCE(!rdp->rcu_iw_pending)) {
191 		rdp->rcu_iw_gp_seq = rnp->gp_seq;
192 		rdp->rcu_iw_pending = false;
193 	}
194 	raw_spin_unlock_rcu_node(rnp);
195 }
196 
197 //////////////////////////////////////////////////////////////////////////////
198 //
199 // Printing RCU CPU stall warnings
200 
201 #ifdef CONFIG_PREEMPT_RCU
202 
203 /*
204  * Dump detailed information for all tasks blocking the current RCU
205  * grace period on the specified rcu_node structure.
206  */
rcu_print_detail_task_stall_rnp(struct rcu_node * rnp)207 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
208 {
209 	unsigned long flags;
210 	struct task_struct *t;
211 
212 	raw_spin_lock_irqsave_rcu_node(rnp, flags);
213 	if (!rcu_preempt_blocked_readers_cgp(rnp)) {
214 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
215 		return;
216 	}
217 	t = list_entry(rnp->gp_tasks->prev,
218 		       struct task_struct, rcu_node_entry);
219 	list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
220 		/*
221 		 * We could be printing a lot while holding a spinlock.
222 		 * Avoid triggering hard lockup.
223 		 */
224 		touch_nmi_watchdog();
225 		sched_show_task(t);
226 	}
227 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
228 }
229 
230 // Communicate task state back to the RCU CPU stall warning request.
231 struct rcu_stall_chk_rdr {
232 	int nesting;
233 	union rcu_special rs;
234 	bool on_blkd_list;
235 };
236 
237 /*
238  * Report out the state of a not-running task that is stalling the
239  * current RCU grace period.
240  */
check_slow_task(struct task_struct * t,void * arg)241 static bool check_slow_task(struct task_struct *t, void *arg)
242 {
243 	struct rcu_stall_chk_rdr *rscrp = arg;
244 
245 	if (task_curr(t))
246 		return false; // It is running, so decline to inspect it.
247 	rscrp->nesting = t->rcu_read_lock_nesting;
248 	rscrp->rs = t->rcu_read_unlock_special;
249 	rscrp->on_blkd_list = !list_empty(&t->rcu_node_entry);
250 	return true;
251 }
252 
253 /*
254  * Scan the current list of tasks blocked within RCU read-side critical
255  * sections, printing out the tid of each of the first few of them.
256  */
rcu_print_task_stall(struct rcu_node * rnp,unsigned long flags)257 static int rcu_print_task_stall(struct rcu_node *rnp, unsigned long flags)
258 	__releases(rnp->lock)
259 {
260 	int i = 0;
261 	int ndetected = 0;
262 	struct rcu_stall_chk_rdr rscr;
263 	struct task_struct *t;
264 	struct task_struct *ts[8];
265 
266 	lockdep_assert_irqs_disabled();
267 	if (!rcu_preempt_blocked_readers_cgp(rnp)) {
268 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
269 		return 0;
270 	}
271 	pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
272 	       rnp->level, rnp->grplo, rnp->grphi);
273 	t = list_entry(rnp->gp_tasks->prev,
274 		       struct task_struct, rcu_node_entry);
275 	list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
276 		get_task_struct(t);
277 		ts[i++] = t;
278 		if (i >= ARRAY_SIZE(ts))
279 			break;
280 	}
281 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
282 	while (i) {
283 		t = ts[--i];
284 		if (!try_invoke_on_locked_down_task(t, check_slow_task, &rscr))
285 			pr_cont(" P%d", t->pid);
286 		else
287 			pr_cont(" P%d/%d:%c%c%c%c",
288 				t->pid, rscr.nesting,
289 				".b"[rscr.rs.b.blocked],
290 				".q"[rscr.rs.b.need_qs],
291 				".e"[rscr.rs.b.exp_hint],
292 				".l"[rscr.on_blkd_list]);
293 		lockdep_assert_irqs_disabled();
294 		put_task_struct(t);
295 		ndetected++;
296 	}
297 	pr_cont("\n");
298 	return ndetected;
299 }
300 
301 #else /* #ifdef CONFIG_PREEMPT_RCU */
302 
303 /*
304  * Because preemptible RCU does not exist, we never have to check for
305  * tasks blocked within RCU read-side critical sections.
306  */
rcu_print_detail_task_stall_rnp(struct rcu_node * rnp)307 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
308 {
309 }
310 
311 /*
312  * Because preemptible RCU does not exist, we never have to check for
313  * tasks blocked within RCU read-side critical sections.
314  */
rcu_print_task_stall(struct rcu_node * rnp,unsigned long flags)315 static int rcu_print_task_stall(struct rcu_node *rnp, unsigned long flags)
316 {
317 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
318 	return 0;
319 }
320 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
321 
322 /*
323  * Dump stacks of all tasks running on stalled CPUs.  First try using
324  * NMIs, but fall back to manual remote stack tracing on architectures
325  * that don't support NMI-based stack dumps.  The NMI-triggered stack
326  * traces are more accurate because they are printed by the target CPU.
327  */
rcu_dump_cpu_stacks(void)328 static void rcu_dump_cpu_stacks(void)
329 {
330 	int cpu;
331 	unsigned long flags;
332 	struct rcu_node *rnp;
333 
334 	rcu_for_each_leaf_node(rnp) {
335 		raw_spin_lock_irqsave_rcu_node(rnp, flags);
336 		for_each_leaf_node_possible_cpu(rnp, cpu)
337 			if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu))
338 				if (!trigger_single_cpu_backtrace(cpu))
339 					dump_cpu_task(cpu);
340 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
341 	}
342 }
343 
344 #ifdef CONFIG_RCU_FAST_NO_HZ
345 
print_cpu_stall_fast_no_hz(char * cp,int cpu)346 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
347 {
348 	struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
349 
350 	sprintf(cp, "last_accelerate: %04lx/%04lx dyntick_enabled: %d",
351 		rdp->last_accelerate & 0xffff, jiffies & 0xffff,
352 		!!rdp->tick_nohz_enabled_snap);
353 }
354 
355 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
356 
print_cpu_stall_fast_no_hz(char * cp,int cpu)357 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
358 {
359 	*cp = '\0';
360 }
361 
362 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
363 
364 static const char * const gp_state_names[] = {
365 	[RCU_GP_IDLE] = "RCU_GP_IDLE",
366 	[RCU_GP_WAIT_GPS] = "RCU_GP_WAIT_GPS",
367 	[RCU_GP_DONE_GPS] = "RCU_GP_DONE_GPS",
368 	[RCU_GP_ONOFF] = "RCU_GP_ONOFF",
369 	[RCU_GP_INIT] = "RCU_GP_INIT",
370 	[RCU_GP_WAIT_FQS] = "RCU_GP_WAIT_FQS",
371 	[RCU_GP_DOING_FQS] = "RCU_GP_DOING_FQS",
372 	[RCU_GP_CLEANUP] = "RCU_GP_CLEANUP",
373 	[RCU_GP_CLEANED] = "RCU_GP_CLEANED",
374 };
375 
376 /*
377  * Convert a ->gp_state value to a character string.
378  */
gp_state_getname(short gs)379 static const char *gp_state_getname(short gs)
380 {
381 	if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
382 		return "???";
383 	return gp_state_names[gs];
384 }
385 
386 /* Is the RCU grace-period kthread being starved of CPU time? */
rcu_is_gp_kthread_starving(unsigned long * jp)387 static bool rcu_is_gp_kthread_starving(unsigned long *jp)
388 {
389 	unsigned long j = jiffies - READ_ONCE(rcu_state.gp_activity);
390 
391 	if (jp)
392 		*jp = j;
393 	return j > 2 * HZ;
394 }
395 
396 /*
397  * Print out diagnostic information for the specified stalled CPU.
398  *
399  * If the specified CPU is aware of the current RCU grace period, then
400  * print the number of scheduling clock interrupts the CPU has taken
401  * during the time that it has been aware.  Otherwise, print the number
402  * of RCU grace periods that this CPU is ignorant of, for example, "1"
403  * if the CPU was aware of the previous grace period.
404  *
405  * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
406  */
print_cpu_stall_info(int cpu)407 static void print_cpu_stall_info(int cpu)
408 {
409 	unsigned long delta;
410 	bool falsepositive;
411 	char fast_no_hz[72];
412 	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
413 	char *ticks_title;
414 	unsigned long ticks_value;
415 
416 	/*
417 	 * We could be printing a lot while holding a spinlock.  Avoid
418 	 * triggering hard lockup.
419 	 */
420 	touch_nmi_watchdog();
421 
422 	ticks_value = rcu_seq_ctr(rcu_state.gp_seq - rdp->gp_seq);
423 	if (ticks_value) {
424 		ticks_title = "GPs behind";
425 	} else {
426 		ticks_title = "ticks this GP";
427 		ticks_value = rdp->ticks_this_gp;
428 	}
429 	print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
430 	delta = rcu_seq_ctr(rdp->mynode->gp_seq - rdp->rcu_iw_gp_seq);
431 	falsepositive = rcu_is_gp_kthread_starving(NULL) &&
432 			rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp));
433 	pr_err("\t%d-%c%c%c%c: (%lu %s) idle=%03x/%ld/%#lx softirq=%u/%u fqs=%ld %s%s\n",
434 	       cpu,
435 	       "O."[!!cpu_online(cpu)],
436 	       "o."[!!(rdp->grpmask & rdp->mynode->qsmaskinit)],
437 	       "N."[!!(rdp->grpmask & rdp->mynode->qsmaskinitnext)],
438 	       !IS_ENABLED(CONFIG_IRQ_WORK) ? '?' :
439 			rdp->rcu_iw_pending ? (int)min(delta, 9UL) + '0' :
440 				"!."[!delta],
441 	       ticks_value, ticks_title,
442 	       rcu_dynticks_snap(rdp) & 0xfff,
443 	       rdp->dynticks_nesting, rdp->dynticks_nmi_nesting,
444 	       rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
445 	       data_race(rcu_state.n_force_qs) - rcu_state.n_force_qs_gpstart,
446 	       fast_no_hz,
447 	       falsepositive ? " (false positive?)" : "");
448 }
449 
450 /* Complain about starvation of grace-period kthread.  */
rcu_check_gp_kthread_starvation(void)451 static void rcu_check_gp_kthread_starvation(void)
452 {
453 	struct task_struct *gpk = rcu_state.gp_kthread;
454 	unsigned long j;
455 
456 	if (rcu_is_gp_kthread_starving(&j)) {
457 		pr_err("%s kthread starved for %ld jiffies! g%ld f%#x %s(%d) ->state=%#lx ->cpu=%d\n",
458 		       rcu_state.name, j,
459 		       (long)rcu_seq_current(&rcu_state.gp_seq),
460 		       data_race(rcu_state.gp_flags),
461 		       gp_state_getname(rcu_state.gp_state), rcu_state.gp_state,
462 		       gpk ? gpk->state : ~0, gpk ? task_cpu(gpk) : -1);
463 		if (gpk) {
464 			pr_err("\tUnless %s kthread gets sufficient CPU time, OOM is now expected behavior.\n", rcu_state.name);
465 			pr_err("RCU grace-period kthread stack dump:\n");
466 			sched_show_task(gpk);
467 			wake_up_process(gpk);
468 		}
469 	}
470 }
471 
print_other_cpu_stall(unsigned long gp_seq,unsigned long gps)472 static void print_other_cpu_stall(unsigned long gp_seq, unsigned long gps)
473 {
474 	int cpu;
475 	unsigned long flags;
476 	unsigned long gpa;
477 	unsigned long j;
478 	int ndetected = 0;
479 	struct rcu_node *rnp;
480 	long totqlen = 0;
481 
482 	lockdep_assert_irqs_disabled();
483 
484 	/* Kick and suppress, if so configured. */
485 	rcu_stall_kick_kthreads();
486 	if (rcu_stall_is_suppressed())
487 		return;
488 
489 	/*
490 	 * OK, time to rat on our buddy...
491 	 * See Documentation/RCU/stallwarn.rst for info on how to debug
492 	 * RCU CPU stall warnings.
493 	 */
494 	trace_rcu_stall_warning(rcu_state.name, TPS("StallDetected"));
495 	pr_err("INFO: %s detected stalls on CPUs/tasks:\n", rcu_state.name);
496 	rcu_for_each_leaf_node(rnp) {
497 		raw_spin_lock_irqsave_rcu_node(rnp, flags);
498 		if (rnp->qsmask != 0) {
499 			for_each_leaf_node_possible_cpu(rnp, cpu)
500 				if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
501 					print_cpu_stall_info(cpu);
502 					ndetected++;
503 				}
504 		}
505 		ndetected += rcu_print_task_stall(rnp, flags); // Releases rnp->lock.
506 		lockdep_assert_irqs_disabled();
507 	}
508 
509 	for_each_possible_cpu(cpu)
510 		totqlen += rcu_get_n_cbs_cpu(cpu);
511 	pr_cont("\t(detected by %d, t=%ld jiffies, g=%ld, q=%lu)\n",
512 	       smp_processor_id(), (long)(jiffies - gps),
513 	       (long)rcu_seq_current(&rcu_state.gp_seq), totqlen);
514 	if (ndetected) {
515 		rcu_dump_cpu_stacks();
516 
517 		/* Complain about tasks blocking the grace period. */
518 		rcu_for_each_leaf_node(rnp)
519 			rcu_print_detail_task_stall_rnp(rnp);
520 	} else {
521 		if (rcu_seq_current(&rcu_state.gp_seq) != gp_seq) {
522 			pr_err("INFO: Stall ended before state dump start\n");
523 		} else {
524 			j = jiffies;
525 			gpa = data_race(rcu_state.gp_activity);
526 			pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
527 			       rcu_state.name, j - gpa, j, gpa,
528 			       data_race(jiffies_till_next_fqs),
529 			       rcu_get_root()->qsmask);
530 		}
531 	}
532 	/* Rewrite if needed in case of slow consoles. */
533 	if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
534 		WRITE_ONCE(rcu_state.jiffies_stall,
535 			   jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
536 
537 	rcu_check_gp_kthread_starvation();
538 
539 	atomic_notifier_call_chain(&rcu_stall_notifier_list, 0, NULL);
540 
541 	panic_on_rcu_stall();
542 
543 	rcu_force_quiescent_state();  /* Kick them all. */
544 }
545 
print_cpu_stall(unsigned long gps)546 static void print_cpu_stall(unsigned long gps)
547 {
548 	int cpu;
549 	unsigned long flags;
550 	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
551 	struct rcu_node *rnp = rcu_get_root();
552 	long totqlen = 0;
553 
554 	lockdep_assert_irqs_disabled();
555 
556 	/* Kick and suppress, if so configured. */
557 	rcu_stall_kick_kthreads();
558 	if (rcu_stall_is_suppressed())
559 		return;
560 
561 	/*
562 	 * OK, time to rat on ourselves...
563 	 * See Documentation/RCU/stallwarn.rst for info on how to debug
564 	 * RCU CPU stall warnings.
565 	 */
566 	trace_rcu_stall_warning(rcu_state.name, TPS("SelfDetected"));
567 	pr_err("INFO: %s self-detected stall on CPU\n", rcu_state.name);
568 	raw_spin_lock_irqsave_rcu_node(rdp->mynode, flags);
569 	print_cpu_stall_info(smp_processor_id());
570 	raw_spin_unlock_irqrestore_rcu_node(rdp->mynode, flags);
571 	for_each_possible_cpu(cpu)
572 		totqlen += rcu_get_n_cbs_cpu(cpu);
573 	pr_cont("\t(t=%lu jiffies g=%ld q=%lu)\n",
574 		jiffies - gps,
575 		(long)rcu_seq_current(&rcu_state.gp_seq), totqlen);
576 
577 	rcu_check_gp_kthread_starvation();
578 
579 	rcu_dump_cpu_stacks();
580 
581 	raw_spin_lock_irqsave_rcu_node(rnp, flags);
582 	/* Rewrite if needed in case of slow consoles. */
583 	if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
584 		WRITE_ONCE(rcu_state.jiffies_stall,
585 			   jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
586 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
587 
588 	panic_on_rcu_stall();
589 
590 	/*
591 	 * Attempt to revive the RCU machinery by forcing a context switch.
592 	 *
593 	 * A context switch would normally allow the RCU state machine to make
594 	 * progress and it could be we're stuck in kernel space without context
595 	 * switches for an entirely unreasonable amount of time.
596 	 */
597 	set_tsk_need_resched(current);
598 	set_preempt_need_resched();
599 }
600 
check_cpu_stall(struct rcu_data * rdp)601 static void check_cpu_stall(struct rcu_data *rdp)
602 {
603 	unsigned long gs1;
604 	unsigned long gs2;
605 	unsigned long gps;
606 	unsigned long j;
607 	unsigned long jn;
608 	unsigned long js;
609 	struct rcu_node *rnp;
610 
611 	lockdep_assert_irqs_disabled();
612 	if ((rcu_stall_is_suppressed() && !READ_ONCE(rcu_kick_kthreads)) ||
613 	    !rcu_gp_in_progress())
614 		return;
615 	rcu_stall_kick_kthreads();
616 	j = jiffies;
617 
618 	/*
619 	 * Lots of memory barriers to reject false positives.
620 	 *
621 	 * The idea is to pick up rcu_state.gp_seq, then
622 	 * rcu_state.jiffies_stall, then rcu_state.gp_start, and finally
623 	 * another copy of rcu_state.gp_seq.  These values are updated in
624 	 * the opposite order with memory barriers (or equivalent) during
625 	 * grace-period initialization and cleanup.  Now, a false positive
626 	 * can occur if we get an new value of rcu_state.gp_start and a old
627 	 * value of rcu_state.jiffies_stall.  But given the memory barriers,
628 	 * the only way that this can happen is if one grace period ends
629 	 * and another starts between these two fetches.  This is detected
630 	 * by comparing the second fetch of rcu_state.gp_seq with the
631 	 * previous fetch from rcu_state.gp_seq.
632 	 *
633 	 * Given this check, comparisons of jiffies, rcu_state.jiffies_stall,
634 	 * and rcu_state.gp_start suffice to forestall false positives.
635 	 */
636 	gs1 = READ_ONCE(rcu_state.gp_seq);
637 	smp_rmb(); /* Pick up ->gp_seq first... */
638 	js = READ_ONCE(rcu_state.jiffies_stall);
639 	smp_rmb(); /* ...then ->jiffies_stall before the rest... */
640 	gps = READ_ONCE(rcu_state.gp_start);
641 	smp_rmb(); /* ...and finally ->gp_start before ->gp_seq again. */
642 	gs2 = READ_ONCE(rcu_state.gp_seq);
643 	if (gs1 != gs2 ||
644 	    ULONG_CMP_LT(j, js) ||
645 	    ULONG_CMP_GE(gps, js))
646 		return; /* No stall or GP completed since entering function. */
647 	rnp = rdp->mynode;
648 	jn = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
649 	if (rcu_gp_in_progress() &&
650 	    (READ_ONCE(rnp->qsmask) & rdp->grpmask) &&
651 	    cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
652 
653 		/*
654 		 * If a virtual machine is stopped by the host it can look to
655 		 * the watchdog like an RCU stall. Check to see if the host
656 		 * stopped the vm.
657 		 */
658 		if (kvm_check_and_clear_guest_paused())
659 			return;
660 
661 		/* We haven't checked in, so go dump stack. */
662 		print_cpu_stall(gps);
663 		if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
664 			rcu_ftrace_dump(DUMP_ALL);
665 
666 	} else if (rcu_gp_in_progress() &&
667 		   ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY) &&
668 		   cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
669 
670 		/*
671 		 * If a virtual machine is stopped by the host it can look to
672 		 * the watchdog like an RCU stall. Check to see if the host
673 		 * stopped the vm.
674 		 */
675 		if (kvm_check_and_clear_guest_paused())
676 			return;
677 
678 		/* They had a few time units to dump stack, so complain. */
679 		print_other_cpu_stall(gs2, gps);
680 		if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
681 			rcu_ftrace_dump(DUMP_ALL);
682 	}
683 }
684 
685 //////////////////////////////////////////////////////////////////////////////
686 //
687 // RCU forward-progress mechanisms, including of callback invocation.
688 
689 
690 /*
691  * Show the state of the grace-period kthreads.
692  */
show_rcu_gp_kthreads(void)693 void show_rcu_gp_kthreads(void)
694 {
695 	unsigned long cbs = 0;
696 	int cpu;
697 	unsigned long j;
698 	unsigned long ja;
699 	unsigned long jr;
700 	unsigned long jw;
701 	struct rcu_data *rdp;
702 	struct rcu_node *rnp;
703 	struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
704 
705 	j = jiffies;
706 	ja = j - data_race(rcu_state.gp_activity);
707 	jr = j - data_race(rcu_state.gp_req_activity);
708 	jw = j - data_race(rcu_state.gp_wake_time);
709 	pr_info("%s: wait state: %s(%d) ->state: %#lx delta ->gp_activity %lu ->gp_req_activity %lu ->gp_wake_time %lu ->gp_wake_seq %ld ->gp_seq %ld ->gp_seq_needed %ld ->gp_flags %#x\n",
710 		rcu_state.name, gp_state_getname(rcu_state.gp_state),
711 		rcu_state.gp_state, t ? t->state : 0x1ffffL,
712 		ja, jr, jw, (long)data_race(rcu_state.gp_wake_seq),
713 		(long)data_race(rcu_state.gp_seq),
714 		(long)data_race(rcu_get_root()->gp_seq_needed),
715 		data_race(rcu_state.gp_flags));
716 	rcu_for_each_node_breadth_first(rnp) {
717 		if (ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq),
718 				 READ_ONCE(rnp->gp_seq_needed)))
719 			continue;
720 		pr_info("\trcu_node %d:%d ->gp_seq %ld ->gp_seq_needed %ld\n",
721 			rnp->grplo, rnp->grphi, (long)data_race(rnp->gp_seq),
722 			(long)data_race(rnp->gp_seq_needed));
723 		if (!rcu_is_leaf_node(rnp))
724 			continue;
725 		for_each_leaf_node_possible_cpu(rnp, cpu) {
726 			rdp = per_cpu_ptr(&rcu_data, cpu);
727 			if (READ_ONCE(rdp->gpwrap) ||
728 			    ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq),
729 					 READ_ONCE(rdp->gp_seq_needed)))
730 				continue;
731 			pr_info("\tcpu %d ->gp_seq_needed %ld\n",
732 				cpu, (long)data_race(rdp->gp_seq_needed));
733 		}
734 	}
735 	for_each_possible_cpu(cpu) {
736 		rdp = per_cpu_ptr(&rcu_data, cpu);
737 		cbs += data_race(rdp->n_cbs_invoked);
738 		if (rcu_segcblist_is_offloaded(&rdp->cblist))
739 			show_rcu_nocb_state(rdp);
740 	}
741 	pr_info("RCU callbacks invoked since boot: %lu\n", cbs);
742 	show_rcu_tasks_gp_kthreads();
743 }
744 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
745 
746 /*
747  * This function checks for grace-period requests that fail to motivate
748  * RCU to come out of its idle mode.
749  */
rcu_check_gp_start_stall(struct rcu_node * rnp,struct rcu_data * rdp,const unsigned long gpssdelay)750 static void rcu_check_gp_start_stall(struct rcu_node *rnp, struct rcu_data *rdp,
751 				     const unsigned long gpssdelay)
752 {
753 	unsigned long flags;
754 	unsigned long j;
755 	struct rcu_node *rnp_root = rcu_get_root();
756 	static atomic_t warned = ATOMIC_INIT(0);
757 
758 	if (!IS_ENABLED(CONFIG_PROVE_RCU) || rcu_gp_in_progress() ||
759 	    ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
760 			 READ_ONCE(rnp_root->gp_seq_needed)) ||
761 	    !smp_load_acquire(&rcu_state.gp_kthread)) // Get stable kthread.
762 		return;
763 	j = jiffies; /* Expensive access, and in common case don't get here. */
764 	if (time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
765 	    time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
766 	    atomic_read(&warned))
767 		return;
768 
769 	raw_spin_lock_irqsave_rcu_node(rnp, flags);
770 	j = jiffies;
771 	if (rcu_gp_in_progress() ||
772 	    ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
773 			 READ_ONCE(rnp_root->gp_seq_needed)) ||
774 	    time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
775 	    time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
776 	    atomic_read(&warned)) {
777 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
778 		return;
779 	}
780 	/* Hold onto the leaf lock to make others see warned==1. */
781 
782 	if (rnp_root != rnp)
783 		raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
784 	j = jiffies;
785 	if (rcu_gp_in_progress() ||
786 	    ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
787 			 READ_ONCE(rnp_root->gp_seq_needed)) ||
788 	    time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
789 	    time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
790 	    atomic_xchg(&warned, 1)) {
791 		if (rnp_root != rnp)
792 			/* irqs remain disabled. */
793 			raw_spin_unlock_rcu_node(rnp_root);
794 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
795 		return;
796 	}
797 	WARN_ON(1);
798 	if (rnp_root != rnp)
799 		raw_spin_unlock_rcu_node(rnp_root);
800 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
801 	show_rcu_gp_kthreads();
802 }
803 
804 /*
805  * Do a forward-progress check for rcutorture.  This is normally invoked
806  * due to an OOM event.  The argument "j" gives the time period during
807  * which rcutorture would like progress to have been made.
808  */
rcu_fwd_progress_check(unsigned long j)809 void rcu_fwd_progress_check(unsigned long j)
810 {
811 	unsigned long cbs;
812 	int cpu;
813 	unsigned long max_cbs = 0;
814 	int max_cpu = -1;
815 	struct rcu_data *rdp;
816 
817 	if (rcu_gp_in_progress()) {
818 		pr_info("%s: GP age %lu jiffies\n",
819 			__func__, jiffies - rcu_state.gp_start);
820 		show_rcu_gp_kthreads();
821 	} else {
822 		pr_info("%s: Last GP end %lu jiffies ago\n",
823 			__func__, jiffies - rcu_state.gp_end);
824 		preempt_disable();
825 		rdp = this_cpu_ptr(&rcu_data);
826 		rcu_check_gp_start_stall(rdp->mynode, rdp, j);
827 		preempt_enable();
828 	}
829 	for_each_possible_cpu(cpu) {
830 		cbs = rcu_get_n_cbs_cpu(cpu);
831 		if (!cbs)
832 			continue;
833 		if (max_cpu < 0)
834 			pr_info("%s: callbacks", __func__);
835 		pr_cont(" %d: %lu", cpu, cbs);
836 		if (cbs <= max_cbs)
837 			continue;
838 		max_cbs = cbs;
839 		max_cpu = cpu;
840 	}
841 	if (max_cpu >= 0)
842 		pr_cont("\n");
843 }
844 EXPORT_SYMBOL_GPL(rcu_fwd_progress_check);
845 
846 /* Commandeer a sysrq key to dump RCU's tree. */
847 static bool sysrq_rcu;
848 module_param(sysrq_rcu, bool, 0444);
849 
850 /* Dump grace-period-request information due to commandeered sysrq. */
sysrq_show_rcu(int key)851 static void sysrq_show_rcu(int key)
852 {
853 	show_rcu_gp_kthreads();
854 }
855 
856 static const struct sysrq_key_op sysrq_rcudump_op = {
857 	.handler = sysrq_show_rcu,
858 	.help_msg = "show-rcu(y)",
859 	.action_msg = "Show RCU tree",
860 	.enable_mask = SYSRQ_ENABLE_DUMP,
861 };
862 
rcu_sysrq_init(void)863 static int __init rcu_sysrq_init(void)
864 {
865 	if (sysrq_rcu)
866 		return register_sysrq_key('y', &sysrq_rcudump_op);
867 	return 0;
868 }
869 early_initcall(rcu_sysrq_init);
870