Loading...
1// SPDX-License-Identifier: GPL-2.0+
2/*
3 * RCU CPU stall warnings for normal RCU grace periods
4 *
5 * Copyright IBM Corporation, 2019
6 *
7 * Author: Paul E. McKenney <paulmck@linux.ibm.com>
8 */
9
10//////////////////////////////////////////////////////////////////////////////
11//
12// Controlling CPU stall warnings, including delay calculation.
13
14/* panic() on RCU Stall sysctl. */
15int sysctl_panic_on_rcu_stall __read_mostly;
16
17#ifdef CONFIG_PROVE_RCU
18#define RCU_STALL_DELAY_DELTA (5 * HZ)
19#else
20#define RCU_STALL_DELAY_DELTA 0
21#endif
22
23/* Limit-check stall timeouts specified at boottime and runtime. */
24int rcu_jiffies_till_stall_check(void)
25{
26 int till_stall_check = READ_ONCE(rcu_cpu_stall_timeout);
27
28 /*
29 * Limit check must be consistent with the Kconfig limits
30 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
31 */
32 if (till_stall_check < 3) {
33 WRITE_ONCE(rcu_cpu_stall_timeout, 3);
34 till_stall_check = 3;
35 } else if (till_stall_check > 300) {
36 WRITE_ONCE(rcu_cpu_stall_timeout, 300);
37 till_stall_check = 300;
38 }
39 return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
40}
41EXPORT_SYMBOL_GPL(rcu_jiffies_till_stall_check);
42
43/* Don't do RCU CPU stall warnings during long sysrq printouts. */
44void rcu_sysrq_start(void)
45{
46 if (!rcu_cpu_stall_suppress)
47 rcu_cpu_stall_suppress = 2;
48}
49
50void rcu_sysrq_end(void)
51{
52 if (rcu_cpu_stall_suppress == 2)
53 rcu_cpu_stall_suppress = 0;
54}
55
56/* Don't print RCU CPU stall warnings during a kernel panic. */
57static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
58{
59 rcu_cpu_stall_suppress = 1;
60 return NOTIFY_DONE;
61}
62
63static struct notifier_block rcu_panic_block = {
64 .notifier_call = rcu_panic,
65};
66
67static int __init check_cpu_stall_init(void)
68{
69 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
70 return 0;
71}
72early_initcall(check_cpu_stall_init);
73
74/* If so specified via sysctl, panic, yielding cleaner stall-warning output. */
75static void panic_on_rcu_stall(void)
76{
77 if (sysctl_panic_on_rcu_stall)
78 panic("RCU Stall\n");
79}
80
81/**
82 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
83 *
84 * Set the stall-warning timeout way off into the future, thus preventing
85 * any RCU CPU stall-warning messages from appearing in the current set of
86 * RCU grace periods.
87 *
88 * The caller must disable hard irqs.
89 */
90void rcu_cpu_stall_reset(void)
91{
92 WRITE_ONCE(rcu_state.jiffies_stall, jiffies + ULONG_MAX / 2);
93}
94
95//////////////////////////////////////////////////////////////////////////////
96//
97// Interaction with RCU grace periods
98
99/* Start of new grace period, so record stall time (and forcing times). */
100static void record_gp_stall_check_time(void)
101{
102 unsigned long j = jiffies;
103 unsigned long j1;
104
105 rcu_state.gp_start = j;
106 j1 = rcu_jiffies_till_stall_check();
107 /* Record ->gp_start before ->jiffies_stall. */
108 smp_store_release(&rcu_state.jiffies_stall, j + j1); /* ^^^ */
109 rcu_state.jiffies_resched = j + j1 / 2;
110 rcu_state.n_force_qs_gpstart = READ_ONCE(rcu_state.n_force_qs);
111}
112
113/* Zero ->ticks_this_gp and snapshot the number of RCU softirq handlers. */
114static void zero_cpu_stall_ticks(struct rcu_data *rdp)
115{
116 rdp->ticks_this_gp = 0;
117 rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
118 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
119}
120
121/*
122 * If too much time has passed in the current grace period, and if
123 * so configured, go kick the relevant kthreads.
124 */
125static void rcu_stall_kick_kthreads(void)
126{
127 unsigned long j;
128
129 if (!rcu_kick_kthreads)
130 return;
131 j = READ_ONCE(rcu_state.jiffies_kick_kthreads);
132 if (time_after(jiffies, j) && rcu_state.gp_kthread &&
133 (rcu_gp_in_progress() || READ_ONCE(rcu_state.gp_flags))) {
134 WARN_ONCE(1, "Kicking %s grace-period kthread\n",
135 rcu_state.name);
136 rcu_ftrace_dump(DUMP_ALL);
137 wake_up_process(rcu_state.gp_kthread);
138 WRITE_ONCE(rcu_state.jiffies_kick_kthreads, j + HZ);
139 }
140}
141
142/*
143 * Handler for the irq_work request posted about halfway into the RCU CPU
144 * stall timeout, and used to detect excessive irq disabling. Set state
145 * appropriately, but just complain if there is unexpected state on entry.
146 */
147static void rcu_iw_handler(struct irq_work *iwp)
148{
149 struct rcu_data *rdp;
150 struct rcu_node *rnp;
151
152 rdp = container_of(iwp, struct rcu_data, rcu_iw);
153 rnp = rdp->mynode;
154 raw_spin_lock_rcu_node(rnp);
155 if (!WARN_ON_ONCE(!rdp->rcu_iw_pending)) {
156 rdp->rcu_iw_gp_seq = rnp->gp_seq;
157 rdp->rcu_iw_pending = false;
158 }
159 raw_spin_unlock_rcu_node(rnp);
160}
161
162//////////////////////////////////////////////////////////////////////////////
163//
164// Printing RCU CPU stall warnings
165
166#ifdef CONFIG_PREEMPTION
167
168/*
169 * Dump detailed information for all tasks blocking the current RCU
170 * grace period on the specified rcu_node structure.
171 */
172static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
173{
174 unsigned long flags;
175 struct task_struct *t;
176
177 raw_spin_lock_irqsave_rcu_node(rnp, flags);
178 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
179 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
180 return;
181 }
182 t = list_entry(rnp->gp_tasks->prev,
183 struct task_struct, rcu_node_entry);
184 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
185 /*
186 * We could be printing a lot while holding a spinlock.
187 * Avoid triggering hard lockup.
188 */
189 touch_nmi_watchdog();
190 sched_show_task(t);
191 }
192 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
193}
194
195/*
196 * Scan the current list of tasks blocked within RCU read-side critical
197 * sections, printing out the tid of each.
198 */
199static int rcu_print_task_stall(struct rcu_node *rnp)
200{
201 struct task_struct *t;
202 int ndetected = 0;
203
204 if (!rcu_preempt_blocked_readers_cgp(rnp))
205 return 0;
206 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
207 rnp->level, rnp->grplo, rnp->grphi);
208 t = list_entry(rnp->gp_tasks->prev,
209 struct task_struct, rcu_node_entry);
210 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
211 pr_cont(" P%d", t->pid);
212 ndetected++;
213 }
214 pr_cont("\n");
215 return ndetected;
216}
217
218#else /* #ifdef CONFIG_PREEMPTION */
219
220/*
221 * Because preemptible RCU does not exist, we never have to check for
222 * tasks blocked within RCU read-side critical sections.
223 */
224static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
225{
226}
227
228/*
229 * Because preemptible RCU does not exist, we never have to check for
230 * tasks blocked within RCU read-side critical sections.
231 */
232static int rcu_print_task_stall(struct rcu_node *rnp)
233{
234 return 0;
235}
236#endif /* #else #ifdef CONFIG_PREEMPTION */
237
238/*
239 * Dump stacks of all tasks running on stalled CPUs. First try using
240 * NMIs, but fall back to manual remote stack tracing on architectures
241 * that don't support NMI-based stack dumps. The NMI-triggered stack
242 * traces are more accurate because they are printed by the target CPU.
243 */
244static void rcu_dump_cpu_stacks(void)
245{
246 int cpu;
247 unsigned long flags;
248 struct rcu_node *rnp;
249
250 rcu_for_each_leaf_node(rnp) {
251 raw_spin_lock_irqsave_rcu_node(rnp, flags);
252 for_each_leaf_node_possible_cpu(rnp, cpu)
253 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu))
254 if (!trigger_single_cpu_backtrace(cpu))
255 dump_cpu_task(cpu);
256 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
257 }
258}
259
260#ifdef CONFIG_RCU_FAST_NO_HZ
261
262static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
263{
264 struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
265
266 sprintf(cp, "last_accelerate: %04lx/%04lx, Nonlazy posted: %c%c%c",
267 rdp->last_accelerate & 0xffff, jiffies & 0xffff,
268 ".l"[rdp->all_lazy],
269 ".L"[!rcu_segcblist_n_nonlazy_cbs(&rdp->cblist)],
270 ".D"[!!rdp->tick_nohz_enabled_snap]);
271}
272
273#else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
274
275static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
276{
277 *cp = '\0';
278}
279
280#endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
281
282/*
283 * Print out diagnostic information for the specified stalled CPU.
284 *
285 * If the specified CPU is aware of the current RCU grace period, then
286 * print the number of scheduling clock interrupts the CPU has taken
287 * during the time that it has been aware. Otherwise, print the number
288 * of RCU grace periods that this CPU is ignorant of, for example, "1"
289 * if the CPU was aware of the previous grace period.
290 *
291 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
292 */
293static void print_cpu_stall_info(int cpu)
294{
295 unsigned long delta;
296 char fast_no_hz[72];
297 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
298 char *ticks_title;
299 unsigned long ticks_value;
300
301 /*
302 * We could be printing a lot while holding a spinlock. Avoid
303 * triggering hard lockup.
304 */
305 touch_nmi_watchdog();
306
307 ticks_value = rcu_seq_ctr(rcu_state.gp_seq - rdp->gp_seq);
308 if (ticks_value) {
309 ticks_title = "GPs behind";
310 } else {
311 ticks_title = "ticks this GP";
312 ticks_value = rdp->ticks_this_gp;
313 }
314 print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
315 delta = rcu_seq_ctr(rdp->mynode->gp_seq - rdp->rcu_iw_gp_seq);
316 pr_err("\t%d-%c%c%c%c: (%lu %s) idle=%03x/%ld/%#lx softirq=%u/%u fqs=%ld %s\n",
317 cpu,
318 "O."[!!cpu_online(cpu)],
319 "o."[!!(rdp->grpmask & rdp->mynode->qsmaskinit)],
320 "N."[!!(rdp->grpmask & rdp->mynode->qsmaskinitnext)],
321 !IS_ENABLED(CONFIG_IRQ_WORK) ? '?' :
322 rdp->rcu_iw_pending ? (int)min(delta, 9UL) + '0' :
323 "!."[!delta],
324 ticks_value, ticks_title,
325 rcu_dynticks_snap(rdp) & 0xfff,
326 rdp->dynticks_nesting, rdp->dynticks_nmi_nesting,
327 rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
328 READ_ONCE(rcu_state.n_force_qs) - rcu_state.n_force_qs_gpstart,
329 fast_no_hz);
330}
331
332/* Complain about starvation of grace-period kthread. */
333static void rcu_check_gp_kthread_starvation(void)
334{
335 struct task_struct *gpk = rcu_state.gp_kthread;
336 unsigned long j;
337
338 j = jiffies - READ_ONCE(rcu_state.gp_activity);
339 if (j > 2 * HZ) {
340 pr_err("%s kthread starved for %ld jiffies! g%ld f%#x %s(%d) ->state=%#lx ->cpu=%d\n",
341 rcu_state.name, j,
342 (long)rcu_seq_current(&rcu_state.gp_seq),
343 READ_ONCE(rcu_state.gp_flags),
344 gp_state_getname(rcu_state.gp_state), rcu_state.gp_state,
345 gpk ? gpk->state : ~0, gpk ? task_cpu(gpk) : -1);
346 if (gpk) {
347 pr_err("RCU grace-period kthread stack dump:\n");
348 sched_show_task(gpk);
349 wake_up_process(gpk);
350 }
351 }
352}
353
354static void print_other_cpu_stall(unsigned long gp_seq)
355{
356 int cpu;
357 unsigned long flags;
358 unsigned long gpa;
359 unsigned long j;
360 int ndetected = 0;
361 struct rcu_node *rnp;
362 long totqlen = 0;
363
364 /* Kick and suppress, if so configured. */
365 rcu_stall_kick_kthreads();
366 if (rcu_cpu_stall_suppress)
367 return;
368
369 /*
370 * OK, time to rat on our buddy...
371 * See Documentation/RCU/stallwarn.txt for info on how to debug
372 * RCU CPU stall warnings.
373 */
374 pr_err("INFO: %s detected stalls on CPUs/tasks:\n", rcu_state.name);
375 rcu_for_each_leaf_node(rnp) {
376 raw_spin_lock_irqsave_rcu_node(rnp, flags);
377 ndetected += rcu_print_task_stall(rnp);
378 if (rnp->qsmask != 0) {
379 for_each_leaf_node_possible_cpu(rnp, cpu)
380 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
381 print_cpu_stall_info(cpu);
382 ndetected++;
383 }
384 }
385 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
386 }
387
388 for_each_possible_cpu(cpu)
389 totqlen += rcu_get_n_cbs_cpu(cpu);
390 pr_cont("\t(detected by %d, t=%ld jiffies, g=%ld, q=%lu)\n",
391 smp_processor_id(), (long)(jiffies - rcu_state.gp_start),
392 (long)rcu_seq_current(&rcu_state.gp_seq), totqlen);
393 if (ndetected) {
394 rcu_dump_cpu_stacks();
395
396 /* Complain about tasks blocking the grace period. */
397 rcu_for_each_leaf_node(rnp)
398 rcu_print_detail_task_stall_rnp(rnp);
399 } else {
400 if (rcu_seq_current(&rcu_state.gp_seq) != gp_seq) {
401 pr_err("INFO: Stall ended before state dump start\n");
402 } else {
403 j = jiffies;
404 gpa = READ_ONCE(rcu_state.gp_activity);
405 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
406 rcu_state.name, j - gpa, j, gpa,
407 READ_ONCE(jiffies_till_next_fqs),
408 rcu_get_root()->qsmask);
409 /* In this case, the current CPU might be at fault. */
410 sched_show_task(current);
411 }
412 }
413 /* Rewrite if needed in case of slow consoles. */
414 if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
415 WRITE_ONCE(rcu_state.jiffies_stall,
416 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
417
418 rcu_check_gp_kthread_starvation();
419
420 panic_on_rcu_stall();
421
422 rcu_force_quiescent_state(); /* Kick them all. */
423}
424
425static void print_cpu_stall(void)
426{
427 int cpu;
428 unsigned long flags;
429 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
430 struct rcu_node *rnp = rcu_get_root();
431 long totqlen = 0;
432
433 /* Kick and suppress, if so configured. */
434 rcu_stall_kick_kthreads();
435 if (rcu_cpu_stall_suppress)
436 return;
437
438 /*
439 * OK, time to rat on ourselves...
440 * See Documentation/RCU/stallwarn.txt for info on how to debug
441 * RCU CPU stall warnings.
442 */
443 pr_err("INFO: %s self-detected stall on CPU\n", rcu_state.name);
444 raw_spin_lock_irqsave_rcu_node(rdp->mynode, flags);
445 print_cpu_stall_info(smp_processor_id());
446 raw_spin_unlock_irqrestore_rcu_node(rdp->mynode, flags);
447 for_each_possible_cpu(cpu)
448 totqlen += rcu_get_n_cbs_cpu(cpu);
449 pr_cont("\t(t=%lu jiffies g=%ld q=%lu)\n",
450 jiffies - rcu_state.gp_start,
451 (long)rcu_seq_current(&rcu_state.gp_seq), totqlen);
452
453 rcu_check_gp_kthread_starvation();
454
455 rcu_dump_cpu_stacks();
456
457 raw_spin_lock_irqsave_rcu_node(rnp, flags);
458 /* Rewrite if needed in case of slow consoles. */
459 if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
460 WRITE_ONCE(rcu_state.jiffies_stall,
461 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
462 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
463
464 panic_on_rcu_stall();
465
466 /*
467 * Attempt to revive the RCU machinery by forcing a context switch.
468 *
469 * A context switch would normally allow the RCU state machine to make
470 * progress and it could be we're stuck in kernel space without context
471 * switches for an entirely unreasonable amount of time.
472 */
473 set_tsk_need_resched(current);
474 set_preempt_need_resched();
475}
476
477static void check_cpu_stall(struct rcu_data *rdp)
478{
479 unsigned long gs1;
480 unsigned long gs2;
481 unsigned long gps;
482 unsigned long j;
483 unsigned long jn;
484 unsigned long js;
485 struct rcu_node *rnp;
486
487 if ((rcu_cpu_stall_suppress && !rcu_kick_kthreads) ||
488 !rcu_gp_in_progress())
489 return;
490 rcu_stall_kick_kthreads();
491 j = jiffies;
492
493 /*
494 * Lots of memory barriers to reject false positives.
495 *
496 * The idea is to pick up rcu_state.gp_seq, then
497 * rcu_state.jiffies_stall, then rcu_state.gp_start, and finally
498 * another copy of rcu_state.gp_seq. These values are updated in
499 * the opposite order with memory barriers (or equivalent) during
500 * grace-period initialization and cleanup. Now, a false positive
501 * can occur if we get an new value of rcu_state.gp_start and a old
502 * value of rcu_state.jiffies_stall. But given the memory barriers,
503 * the only way that this can happen is if one grace period ends
504 * and another starts between these two fetches. This is detected
505 * by comparing the second fetch of rcu_state.gp_seq with the
506 * previous fetch from rcu_state.gp_seq.
507 *
508 * Given this check, comparisons of jiffies, rcu_state.jiffies_stall,
509 * and rcu_state.gp_start suffice to forestall false positives.
510 */
511 gs1 = READ_ONCE(rcu_state.gp_seq);
512 smp_rmb(); /* Pick up ->gp_seq first... */
513 js = READ_ONCE(rcu_state.jiffies_stall);
514 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
515 gps = READ_ONCE(rcu_state.gp_start);
516 smp_rmb(); /* ...and finally ->gp_start before ->gp_seq again. */
517 gs2 = READ_ONCE(rcu_state.gp_seq);
518 if (gs1 != gs2 ||
519 ULONG_CMP_LT(j, js) ||
520 ULONG_CMP_GE(gps, js))
521 return; /* No stall or GP completed since entering function. */
522 rnp = rdp->mynode;
523 jn = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
524 if (rcu_gp_in_progress() &&
525 (READ_ONCE(rnp->qsmask) & rdp->grpmask) &&
526 cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
527
528 /* We haven't checked in, so go dump stack. */
529 print_cpu_stall();
530 if (rcu_cpu_stall_ftrace_dump)
531 rcu_ftrace_dump(DUMP_ALL);
532
533 } else if (rcu_gp_in_progress() &&
534 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY) &&
535 cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
536
537 /* They had a few time units to dump stack, so complain. */
538 print_other_cpu_stall(gs2);
539 if (rcu_cpu_stall_ftrace_dump)
540 rcu_ftrace_dump(DUMP_ALL);
541 }
542}
543
544//////////////////////////////////////////////////////////////////////////////
545//
546// RCU forward-progress mechanisms, including of callback invocation.
547
548
549/*
550 * Show the state of the grace-period kthreads.
551 */
552void show_rcu_gp_kthreads(void)
553{
554 int cpu;
555 unsigned long j;
556 unsigned long ja;
557 unsigned long jr;
558 unsigned long jw;
559 struct rcu_data *rdp;
560 struct rcu_node *rnp;
561
562 j = jiffies;
563 ja = j - READ_ONCE(rcu_state.gp_activity);
564 jr = j - READ_ONCE(rcu_state.gp_req_activity);
565 jw = j - READ_ONCE(rcu_state.gp_wake_time);
566 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",
567 rcu_state.name, gp_state_getname(rcu_state.gp_state),
568 rcu_state.gp_state,
569 rcu_state.gp_kthread ? rcu_state.gp_kthread->state : 0x1ffffL,
570 ja, jr, jw, (long)READ_ONCE(rcu_state.gp_wake_seq),
571 (long)READ_ONCE(rcu_state.gp_seq),
572 (long)READ_ONCE(rcu_get_root()->gp_seq_needed),
573 READ_ONCE(rcu_state.gp_flags));
574 rcu_for_each_node_breadth_first(rnp) {
575 if (ULONG_CMP_GE(rcu_state.gp_seq, rnp->gp_seq_needed))
576 continue;
577 pr_info("\trcu_node %d:%d ->gp_seq %ld ->gp_seq_needed %ld\n",
578 rnp->grplo, rnp->grphi, (long)rnp->gp_seq,
579 (long)rnp->gp_seq_needed);
580 if (!rcu_is_leaf_node(rnp))
581 continue;
582 for_each_leaf_node_possible_cpu(rnp, cpu) {
583 rdp = per_cpu_ptr(&rcu_data, cpu);
584 if (rdp->gpwrap ||
585 ULONG_CMP_GE(rcu_state.gp_seq,
586 rdp->gp_seq_needed))
587 continue;
588 pr_info("\tcpu %d ->gp_seq_needed %ld\n",
589 cpu, (long)rdp->gp_seq_needed);
590 }
591 }
592 for_each_possible_cpu(cpu) {
593 rdp = per_cpu_ptr(&rcu_data, cpu);
594 if (rcu_segcblist_is_offloaded(&rdp->cblist))
595 show_rcu_nocb_state(rdp);
596 }
597 /* sched_show_task(rcu_state.gp_kthread); */
598}
599EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
600
601/*
602 * This function checks for grace-period requests that fail to motivate
603 * RCU to come out of its idle mode.
604 */
605static void rcu_check_gp_start_stall(struct rcu_node *rnp, struct rcu_data *rdp,
606 const unsigned long gpssdelay)
607{
608 unsigned long flags;
609 unsigned long j;
610 struct rcu_node *rnp_root = rcu_get_root();
611 static atomic_t warned = ATOMIC_INIT(0);
612
613 if (!IS_ENABLED(CONFIG_PROVE_RCU) || rcu_gp_in_progress() ||
614 ULONG_CMP_GE(rnp_root->gp_seq, rnp_root->gp_seq_needed))
615 return;
616 j = jiffies; /* Expensive access, and in common case don't get here. */
617 if (time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
618 time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
619 atomic_read(&warned))
620 return;
621
622 raw_spin_lock_irqsave_rcu_node(rnp, flags);
623 j = jiffies;
624 if (rcu_gp_in_progress() ||
625 ULONG_CMP_GE(rnp_root->gp_seq, rnp_root->gp_seq_needed) ||
626 time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
627 time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
628 atomic_read(&warned)) {
629 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
630 return;
631 }
632 /* Hold onto the leaf lock to make others see warned==1. */
633
634 if (rnp_root != rnp)
635 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
636 j = jiffies;
637 if (rcu_gp_in_progress() ||
638 ULONG_CMP_GE(rnp_root->gp_seq, rnp_root->gp_seq_needed) ||
639 time_before(j, rcu_state.gp_req_activity + gpssdelay) ||
640 time_before(j, rcu_state.gp_activity + gpssdelay) ||
641 atomic_xchg(&warned, 1)) {
642 if (rnp_root != rnp)
643 /* irqs remain disabled. */
644 raw_spin_unlock_rcu_node(rnp_root);
645 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
646 return;
647 }
648 WARN_ON(1);
649 if (rnp_root != rnp)
650 raw_spin_unlock_rcu_node(rnp_root);
651 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
652 show_rcu_gp_kthreads();
653}
654
655/*
656 * Do a forward-progress check for rcutorture. This is normally invoked
657 * due to an OOM event. The argument "j" gives the time period during
658 * which rcutorture would like progress to have been made.
659 */
660void rcu_fwd_progress_check(unsigned long j)
661{
662 unsigned long cbs;
663 int cpu;
664 unsigned long max_cbs = 0;
665 int max_cpu = -1;
666 struct rcu_data *rdp;
667
668 if (rcu_gp_in_progress()) {
669 pr_info("%s: GP age %lu jiffies\n",
670 __func__, jiffies - rcu_state.gp_start);
671 show_rcu_gp_kthreads();
672 } else {
673 pr_info("%s: Last GP end %lu jiffies ago\n",
674 __func__, jiffies - rcu_state.gp_end);
675 preempt_disable();
676 rdp = this_cpu_ptr(&rcu_data);
677 rcu_check_gp_start_stall(rdp->mynode, rdp, j);
678 preempt_enable();
679 }
680 for_each_possible_cpu(cpu) {
681 cbs = rcu_get_n_cbs_cpu(cpu);
682 if (!cbs)
683 continue;
684 if (max_cpu < 0)
685 pr_info("%s: callbacks", __func__);
686 pr_cont(" %d: %lu", cpu, cbs);
687 if (cbs <= max_cbs)
688 continue;
689 max_cbs = cbs;
690 max_cpu = cpu;
691 }
692 if (max_cpu >= 0)
693 pr_cont("\n");
694}
695EXPORT_SYMBOL_GPL(rcu_fwd_progress_check);
696
697/* Commandeer a sysrq key to dump RCU's tree. */
698static bool sysrq_rcu;
699module_param(sysrq_rcu, bool, 0444);
700
701/* Dump grace-period-request information due to commandeered sysrq. */
702static void sysrq_show_rcu(int key)
703{
704 show_rcu_gp_kthreads();
705}
706
707static struct sysrq_key_op sysrq_rcudump_op = {
708 .handler = sysrq_show_rcu,
709 .help_msg = "show-rcu(y)",
710 .action_msg = "Show RCU tree",
711 .enable_mask = SYSRQ_ENABLE_DUMP,
712};
713
714static int __init rcu_sysrq_init(void)
715{
716 if (sysrq_rcu)
717 return register_sysrq_key('y', &sysrq_rcudump_op);
718 return 0;
719}
720early_initcall(rcu_sysrq_init);
1// SPDX-License-Identifier: GPL-2.0+
2/*
3 * RCU CPU stall warnings for normal RCU grace periods
4 *
5 * Copyright IBM Corporation, 2019
6 *
7 * Author: Paul E. McKenney <paulmck@linux.ibm.com>
8 */
9
10//////////////////////////////////////////////////////////////////////////////
11//
12// Controlling CPU stall warnings, including delay calculation.
13
14/* panic() on RCU Stall sysctl. */
15int sysctl_panic_on_rcu_stall __read_mostly;
16
17#ifdef CONFIG_PROVE_RCU
18#define RCU_STALL_DELAY_DELTA (5 * HZ)
19#else
20#define RCU_STALL_DELAY_DELTA 0
21#endif
22#define RCU_STALL_MIGHT_DIV 8
23#define RCU_STALL_MIGHT_MIN (2 * HZ)
24
25/* Limit-check stall timeouts specified at boottime and runtime. */
26int rcu_jiffies_till_stall_check(void)
27{
28 int till_stall_check = READ_ONCE(rcu_cpu_stall_timeout);
29
30 /*
31 * Limit check must be consistent with the Kconfig limits
32 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
33 */
34 if (till_stall_check < 3) {
35 WRITE_ONCE(rcu_cpu_stall_timeout, 3);
36 till_stall_check = 3;
37 } else if (till_stall_check > 300) {
38 WRITE_ONCE(rcu_cpu_stall_timeout, 300);
39 till_stall_check = 300;
40 }
41 return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
42}
43EXPORT_SYMBOL_GPL(rcu_jiffies_till_stall_check);
44
45/**
46 * rcu_gp_might_be_stalled - Is it likely that the grace period is stalled?
47 *
48 * Returns @true if the current grace period is sufficiently old that
49 * it is reasonable to assume that it might be stalled. This can be
50 * useful when deciding whether to allocate memory to enable RCU-mediated
51 * freeing on the one hand or just invoking synchronize_rcu() on the other.
52 * The latter is preferable when the grace period is stalled.
53 *
54 * Note that sampling of the .gp_start and .gp_seq fields must be done
55 * carefully to avoid false positives at the beginnings and ends of
56 * grace periods.
57 */
58bool rcu_gp_might_be_stalled(void)
59{
60 unsigned long d = rcu_jiffies_till_stall_check() / RCU_STALL_MIGHT_DIV;
61 unsigned long j = jiffies;
62
63 if (d < RCU_STALL_MIGHT_MIN)
64 d = RCU_STALL_MIGHT_MIN;
65 smp_mb(); // jiffies before .gp_seq to avoid false positives.
66 if (!rcu_gp_in_progress())
67 return false;
68 // Long delays at this point avoids false positive, but a delay
69 // of ULONG_MAX/4 jiffies voids your no-false-positive warranty.
70 smp_mb(); // .gp_seq before second .gp_start
71 // And ditto here.
72 return !time_before(j, READ_ONCE(rcu_state.gp_start) + d);
73}
74
75/* Don't do RCU CPU stall warnings during long sysrq printouts. */
76void rcu_sysrq_start(void)
77{
78 if (!rcu_cpu_stall_suppress)
79 rcu_cpu_stall_suppress = 2;
80}
81
82void rcu_sysrq_end(void)
83{
84 if (rcu_cpu_stall_suppress == 2)
85 rcu_cpu_stall_suppress = 0;
86}
87
88/* Don't print RCU CPU stall warnings during a kernel panic. */
89static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
90{
91 rcu_cpu_stall_suppress = 1;
92 return NOTIFY_DONE;
93}
94
95static struct notifier_block rcu_panic_block = {
96 .notifier_call = rcu_panic,
97};
98
99static int __init check_cpu_stall_init(void)
100{
101 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
102 return 0;
103}
104early_initcall(check_cpu_stall_init);
105
106/* If so specified via sysctl, panic, yielding cleaner stall-warning output. */
107static void panic_on_rcu_stall(void)
108{
109 if (sysctl_panic_on_rcu_stall)
110 panic("RCU Stall\n");
111}
112
113/**
114 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
115 *
116 * Set the stall-warning timeout way off into the future, thus preventing
117 * any RCU CPU stall-warning messages from appearing in the current set of
118 * RCU grace periods.
119 *
120 * The caller must disable hard irqs.
121 */
122void rcu_cpu_stall_reset(void)
123{
124 WRITE_ONCE(rcu_state.jiffies_stall, jiffies + ULONG_MAX / 2);
125}
126
127//////////////////////////////////////////////////////////////////////////////
128//
129// Interaction with RCU grace periods
130
131/* Start of new grace period, so record stall time (and forcing times). */
132static void record_gp_stall_check_time(void)
133{
134 unsigned long j = jiffies;
135 unsigned long j1;
136
137 WRITE_ONCE(rcu_state.gp_start, j);
138 j1 = rcu_jiffies_till_stall_check();
139 smp_mb(); // ->gp_start before ->jiffies_stall and caller's ->gp_seq.
140 WRITE_ONCE(rcu_state.jiffies_stall, j + j1);
141 rcu_state.jiffies_resched = j + j1 / 2;
142 rcu_state.n_force_qs_gpstart = READ_ONCE(rcu_state.n_force_qs);
143}
144
145/* Zero ->ticks_this_gp and snapshot the number of RCU softirq handlers. */
146static void zero_cpu_stall_ticks(struct rcu_data *rdp)
147{
148 rdp->ticks_this_gp = 0;
149 rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
150 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
151}
152
153/*
154 * If too much time has passed in the current grace period, and if
155 * so configured, go kick the relevant kthreads.
156 */
157static void rcu_stall_kick_kthreads(void)
158{
159 unsigned long j;
160
161 if (!rcu_kick_kthreads)
162 return;
163 j = READ_ONCE(rcu_state.jiffies_kick_kthreads);
164 if (time_after(jiffies, j) && rcu_state.gp_kthread &&
165 (rcu_gp_in_progress() || READ_ONCE(rcu_state.gp_flags))) {
166 WARN_ONCE(1, "Kicking %s grace-period kthread\n",
167 rcu_state.name);
168 rcu_ftrace_dump(DUMP_ALL);
169 wake_up_process(rcu_state.gp_kthread);
170 WRITE_ONCE(rcu_state.jiffies_kick_kthreads, j + HZ);
171 }
172}
173
174/*
175 * Handler for the irq_work request posted about halfway into the RCU CPU
176 * stall timeout, and used to detect excessive irq disabling. Set state
177 * appropriately, but just complain if there is unexpected state on entry.
178 */
179static void rcu_iw_handler(struct irq_work *iwp)
180{
181 struct rcu_data *rdp;
182 struct rcu_node *rnp;
183
184 rdp = container_of(iwp, struct rcu_data, rcu_iw);
185 rnp = rdp->mynode;
186 raw_spin_lock_rcu_node(rnp);
187 if (!WARN_ON_ONCE(!rdp->rcu_iw_pending)) {
188 rdp->rcu_iw_gp_seq = rnp->gp_seq;
189 rdp->rcu_iw_pending = false;
190 }
191 raw_spin_unlock_rcu_node(rnp);
192}
193
194//////////////////////////////////////////////////////////////////////////////
195//
196// Printing RCU CPU stall warnings
197
198#ifdef CONFIG_PREEMPT_RCU
199
200/*
201 * Dump detailed information for all tasks blocking the current RCU
202 * grace period on the specified rcu_node structure.
203 */
204static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
205{
206 unsigned long flags;
207 struct task_struct *t;
208
209 raw_spin_lock_irqsave_rcu_node(rnp, flags);
210 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
211 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
212 return;
213 }
214 t = list_entry(rnp->gp_tasks->prev,
215 struct task_struct, rcu_node_entry);
216 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
217 /*
218 * We could be printing a lot while holding a spinlock.
219 * Avoid triggering hard lockup.
220 */
221 touch_nmi_watchdog();
222 sched_show_task(t);
223 }
224 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
225}
226
227// Communicate task state back to the RCU CPU stall warning request.
228struct rcu_stall_chk_rdr {
229 int nesting;
230 union rcu_special rs;
231 bool on_blkd_list;
232};
233
234/*
235 * Report out the state of a not-running task that is stalling the
236 * current RCU grace period.
237 */
238static bool check_slow_task(struct task_struct *t, void *arg)
239{
240 struct rcu_stall_chk_rdr *rscrp = arg;
241
242 if (task_curr(t))
243 return false; // It is running, so decline to inspect it.
244 rscrp->nesting = t->rcu_read_lock_nesting;
245 rscrp->rs = t->rcu_read_unlock_special;
246 rscrp->on_blkd_list = !list_empty(&t->rcu_node_entry);
247 return true;
248}
249
250/*
251 * Scan the current list of tasks blocked within RCU read-side critical
252 * sections, printing out the tid of each.
253 */
254static int rcu_print_task_stall(struct rcu_node *rnp)
255{
256 int ndetected = 0;
257 struct rcu_stall_chk_rdr rscr;
258 struct task_struct *t;
259
260 if (!rcu_preempt_blocked_readers_cgp(rnp))
261 return 0;
262 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
263 rnp->level, rnp->grplo, rnp->grphi);
264 t = list_entry(rnp->gp_tasks->prev,
265 struct task_struct, rcu_node_entry);
266 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
267 if (!try_invoke_on_locked_down_task(t, check_slow_task, &rscr))
268 pr_cont(" P%d", t->pid);
269 else
270 pr_cont(" P%d/%d:%c%c%c%c",
271 t->pid, rscr.nesting,
272 ".b"[rscr.rs.b.blocked],
273 ".q"[rscr.rs.b.need_qs],
274 ".e"[rscr.rs.b.exp_hint],
275 ".l"[rscr.on_blkd_list]);
276 ndetected++;
277 }
278 pr_cont("\n");
279 return ndetected;
280}
281
282#else /* #ifdef CONFIG_PREEMPT_RCU */
283
284/*
285 * Because preemptible RCU does not exist, we never have to check for
286 * tasks blocked within RCU read-side critical sections.
287 */
288static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
289{
290}
291
292/*
293 * Because preemptible RCU does not exist, we never have to check for
294 * tasks blocked within RCU read-side critical sections.
295 */
296static int rcu_print_task_stall(struct rcu_node *rnp)
297{
298 return 0;
299}
300#endif /* #else #ifdef CONFIG_PREEMPT_RCU */
301
302/*
303 * Dump stacks of all tasks running on stalled CPUs. First try using
304 * NMIs, but fall back to manual remote stack tracing on architectures
305 * that don't support NMI-based stack dumps. The NMI-triggered stack
306 * traces are more accurate because they are printed by the target CPU.
307 */
308static void rcu_dump_cpu_stacks(void)
309{
310 int cpu;
311 unsigned long flags;
312 struct rcu_node *rnp;
313
314 rcu_for_each_leaf_node(rnp) {
315 raw_spin_lock_irqsave_rcu_node(rnp, flags);
316 for_each_leaf_node_possible_cpu(rnp, cpu)
317 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu))
318 if (!trigger_single_cpu_backtrace(cpu))
319 dump_cpu_task(cpu);
320 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
321 }
322}
323
324#ifdef CONFIG_RCU_FAST_NO_HZ
325
326static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
327{
328 struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
329
330 sprintf(cp, "last_accelerate: %04lx/%04lx dyntick_enabled: %d",
331 rdp->last_accelerate & 0xffff, jiffies & 0xffff,
332 !!rdp->tick_nohz_enabled_snap);
333}
334
335#else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
336
337static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
338{
339 *cp = '\0';
340}
341
342#endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
343
344static const char * const gp_state_names[] = {
345 [RCU_GP_IDLE] = "RCU_GP_IDLE",
346 [RCU_GP_WAIT_GPS] = "RCU_GP_WAIT_GPS",
347 [RCU_GP_DONE_GPS] = "RCU_GP_DONE_GPS",
348 [RCU_GP_ONOFF] = "RCU_GP_ONOFF",
349 [RCU_GP_INIT] = "RCU_GP_INIT",
350 [RCU_GP_WAIT_FQS] = "RCU_GP_WAIT_FQS",
351 [RCU_GP_DOING_FQS] = "RCU_GP_DOING_FQS",
352 [RCU_GP_CLEANUP] = "RCU_GP_CLEANUP",
353 [RCU_GP_CLEANED] = "RCU_GP_CLEANED",
354};
355
356/*
357 * Convert a ->gp_state value to a character string.
358 */
359static const char *gp_state_getname(short gs)
360{
361 if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
362 return "???";
363 return gp_state_names[gs];
364}
365
366/* Is the RCU grace-period kthread being starved of CPU time? */
367static bool rcu_is_gp_kthread_starving(unsigned long *jp)
368{
369 unsigned long j = jiffies - READ_ONCE(rcu_state.gp_activity);
370
371 if (jp)
372 *jp = j;
373 return j > 2 * HZ;
374}
375
376/*
377 * Print out diagnostic information for the specified stalled CPU.
378 *
379 * If the specified CPU is aware of the current RCU grace period, then
380 * print the number of scheduling clock interrupts the CPU has taken
381 * during the time that it has been aware. Otherwise, print the number
382 * of RCU grace periods that this CPU is ignorant of, for example, "1"
383 * if the CPU was aware of the previous grace period.
384 *
385 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
386 */
387static void print_cpu_stall_info(int cpu)
388{
389 unsigned long delta;
390 bool falsepositive;
391 char fast_no_hz[72];
392 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
393 char *ticks_title;
394 unsigned long ticks_value;
395
396 /*
397 * We could be printing a lot while holding a spinlock. Avoid
398 * triggering hard lockup.
399 */
400 touch_nmi_watchdog();
401
402 ticks_value = rcu_seq_ctr(rcu_state.gp_seq - rdp->gp_seq);
403 if (ticks_value) {
404 ticks_title = "GPs behind";
405 } else {
406 ticks_title = "ticks this GP";
407 ticks_value = rdp->ticks_this_gp;
408 }
409 print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
410 delta = rcu_seq_ctr(rdp->mynode->gp_seq - rdp->rcu_iw_gp_seq);
411 falsepositive = rcu_is_gp_kthread_starving(NULL) &&
412 rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp));
413 pr_err("\t%d-%c%c%c%c: (%lu %s) idle=%03x/%ld/%#lx softirq=%u/%u fqs=%ld %s%s\n",
414 cpu,
415 "O."[!!cpu_online(cpu)],
416 "o."[!!(rdp->grpmask & rdp->mynode->qsmaskinit)],
417 "N."[!!(rdp->grpmask & rdp->mynode->qsmaskinitnext)],
418 !IS_ENABLED(CONFIG_IRQ_WORK) ? '?' :
419 rdp->rcu_iw_pending ? (int)min(delta, 9UL) + '0' :
420 "!."[!delta],
421 ticks_value, ticks_title,
422 rcu_dynticks_snap(rdp) & 0xfff,
423 rdp->dynticks_nesting, rdp->dynticks_nmi_nesting,
424 rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
425 data_race(rcu_state.n_force_qs) - rcu_state.n_force_qs_gpstart,
426 fast_no_hz,
427 falsepositive ? " (false positive?)" : "");
428}
429
430/* Complain about starvation of grace-period kthread. */
431static void rcu_check_gp_kthread_starvation(void)
432{
433 struct task_struct *gpk = rcu_state.gp_kthread;
434 unsigned long j;
435
436 if (rcu_is_gp_kthread_starving(&j)) {
437 pr_err("%s kthread starved for %ld jiffies! g%ld f%#x %s(%d) ->state=%#lx ->cpu=%d\n",
438 rcu_state.name, j,
439 (long)rcu_seq_current(&rcu_state.gp_seq),
440 data_race(rcu_state.gp_flags),
441 gp_state_getname(rcu_state.gp_state), rcu_state.gp_state,
442 gpk ? gpk->state : ~0, gpk ? task_cpu(gpk) : -1);
443 if (gpk) {
444 pr_err("\tUnless %s kthread gets sufficient CPU time, OOM is now expected behavior.\n", rcu_state.name);
445 pr_err("RCU grace-period kthread stack dump:\n");
446 sched_show_task(gpk);
447 wake_up_process(gpk);
448 }
449 }
450}
451
452static void print_other_cpu_stall(unsigned long gp_seq, unsigned long gps)
453{
454 int cpu;
455 unsigned long flags;
456 unsigned long gpa;
457 unsigned long j;
458 int ndetected = 0;
459 struct rcu_node *rnp;
460 long totqlen = 0;
461
462 /* Kick and suppress, if so configured. */
463 rcu_stall_kick_kthreads();
464 if (rcu_stall_is_suppressed())
465 return;
466
467 /*
468 * OK, time to rat on our buddy...
469 * See Documentation/RCU/stallwarn.rst for info on how to debug
470 * RCU CPU stall warnings.
471 */
472 pr_err("INFO: %s detected stalls on CPUs/tasks:\n", rcu_state.name);
473 rcu_for_each_leaf_node(rnp) {
474 raw_spin_lock_irqsave_rcu_node(rnp, flags);
475 ndetected += rcu_print_task_stall(rnp);
476 if (rnp->qsmask != 0) {
477 for_each_leaf_node_possible_cpu(rnp, cpu)
478 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
479 print_cpu_stall_info(cpu);
480 ndetected++;
481 }
482 }
483 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
484 }
485
486 for_each_possible_cpu(cpu)
487 totqlen += rcu_get_n_cbs_cpu(cpu);
488 pr_cont("\t(detected by %d, t=%ld jiffies, g=%ld, q=%lu)\n",
489 smp_processor_id(), (long)(jiffies - gps),
490 (long)rcu_seq_current(&rcu_state.gp_seq), totqlen);
491 if (ndetected) {
492 rcu_dump_cpu_stacks();
493
494 /* Complain about tasks blocking the grace period. */
495 rcu_for_each_leaf_node(rnp)
496 rcu_print_detail_task_stall_rnp(rnp);
497 } else {
498 if (rcu_seq_current(&rcu_state.gp_seq) != gp_seq) {
499 pr_err("INFO: Stall ended before state dump start\n");
500 } else {
501 j = jiffies;
502 gpa = data_race(rcu_state.gp_activity);
503 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
504 rcu_state.name, j - gpa, j, gpa,
505 data_race(jiffies_till_next_fqs),
506 rcu_get_root()->qsmask);
507 }
508 }
509 /* Rewrite if needed in case of slow consoles. */
510 if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
511 WRITE_ONCE(rcu_state.jiffies_stall,
512 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
513
514 rcu_check_gp_kthread_starvation();
515
516 panic_on_rcu_stall();
517
518 rcu_force_quiescent_state(); /* Kick them all. */
519}
520
521static void print_cpu_stall(unsigned long gps)
522{
523 int cpu;
524 unsigned long flags;
525 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
526 struct rcu_node *rnp = rcu_get_root();
527 long totqlen = 0;
528
529 /* Kick and suppress, if so configured. */
530 rcu_stall_kick_kthreads();
531 if (rcu_stall_is_suppressed())
532 return;
533
534 /*
535 * OK, time to rat on ourselves...
536 * See Documentation/RCU/stallwarn.rst for info on how to debug
537 * RCU CPU stall warnings.
538 */
539 pr_err("INFO: %s self-detected stall on CPU\n", rcu_state.name);
540 raw_spin_lock_irqsave_rcu_node(rdp->mynode, flags);
541 print_cpu_stall_info(smp_processor_id());
542 raw_spin_unlock_irqrestore_rcu_node(rdp->mynode, flags);
543 for_each_possible_cpu(cpu)
544 totqlen += rcu_get_n_cbs_cpu(cpu);
545 pr_cont("\t(t=%lu jiffies g=%ld q=%lu)\n",
546 jiffies - gps,
547 (long)rcu_seq_current(&rcu_state.gp_seq), totqlen);
548
549 rcu_check_gp_kthread_starvation();
550
551 rcu_dump_cpu_stacks();
552
553 raw_spin_lock_irqsave_rcu_node(rnp, flags);
554 /* Rewrite if needed in case of slow consoles. */
555 if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
556 WRITE_ONCE(rcu_state.jiffies_stall,
557 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
558 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
559
560 panic_on_rcu_stall();
561
562 /*
563 * Attempt to revive the RCU machinery by forcing a context switch.
564 *
565 * A context switch would normally allow the RCU state machine to make
566 * progress and it could be we're stuck in kernel space without context
567 * switches for an entirely unreasonable amount of time.
568 */
569 set_tsk_need_resched(current);
570 set_preempt_need_resched();
571}
572
573static void check_cpu_stall(struct rcu_data *rdp)
574{
575 unsigned long gs1;
576 unsigned long gs2;
577 unsigned long gps;
578 unsigned long j;
579 unsigned long jn;
580 unsigned long js;
581 struct rcu_node *rnp;
582
583 if ((rcu_stall_is_suppressed() && !rcu_kick_kthreads) ||
584 !rcu_gp_in_progress())
585 return;
586 rcu_stall_kick_kthreads();
587 j = jiffies;
588
589 /*
590 * Lots of memory barriers to reject false positives.
591 *
592 * The idea is to pick up rcu_state.gp_seq, then
593 * rcu_state.jiffies_stall, then rcu_state.gp_start, and finally
594 * another copy of rcu_state.gp_seq. These values are updated in
595 * the opposite order with memory barriers (or equivalent) during
596 * grace-period initialization and cleanup. Now, a false positive
597 * can occur if we get an new value of rcu_state.gp_start and a old
598 * value of rcu_state.jiffies_stall. But given the memory barriers,
599 * the only way that this can happen is if one grace period ends
600 * and another starts between these two fetches. This is detected
601 * by comparing the second fetch of rcu_state.gp_seq with the
602 * previous fetch from rcu_state.gp_seq.
603 *
604 * Given this check, comparisons of jiffies, rcu_state.jiffies_stall,
605 * and rcu_state.gp_start suffice to forestall false positives.
606 */
607 gs1 = READ_ONCE(rcu_state.gp_seq);
608 smp_rmb(); /* Pick up ->gp_seq first... */
609 js = READ_ONCE(rcu_state.jiffies_stall);
610 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
611 gps = READ_ONCE(rcu_state.gp_start);
612 smp_rmb(); /* ...and finally ->gp_start before ->gp_seq again. */
613 gs2 = READ_ONCE(rcu_state.gp_seq);
614 if (gs1 != gs2 ||
615 ULONG_CMP_LT(j, js) ||
616 ULONG_CMP_GE(gps, js))
617 return; /* No stall or GP completed since entering function. */
618 rnp = rdp->mynode;
619 jn = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
620 if (rcu_gp_in_progress() &&
621 (READ_ONCE(rnp->qsmask) & rdp->grpmask) &&
622 cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
623
624 /* We haven't checked in, so go dump stack. */
625 print_cpu_stall(gps);
626 if (rcu_cpu_stall_ftrace_dump)
627 rcu_ftrace_dump(DUMP_ALL);
628
629 } else if (rcu_gp_in_progress() &&
630 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY) &&
631 cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
632
633 /* They had a few time units to dump stack, so complain. */
634 print_other_cpu_stall(gs2, gps);
635 if (rcu_cpu_stall_ftrace_dump)
636 rcu_ftrace_dump(DUMP_ALL);
637 }
638}
639
640//////////////////////////////////////////////////////////////////////////////
641//
642// RCU forward-progress mechanisms, including of callback invocation.
643
644
645/*
646 * Show the state of the grace-period kthreads.
647 */
648void show_rcu_gp_kthreads(void)
649{
650 unsigned long cbs = 0;
651 int cpu;
652 unsigned long j;
653 unsigned long ja;
654 unsigned long jr;
655 unsigned long jw;
656 struct rcu_data *rdp;
657 struct rcu_node *rnp;
658 struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
659
660 j = jiffies;
661 ja = j - data_race(rcu_state.gp_activity);
662 jr = j - data_race(rcu_state.gp_req_activity);
663 jw = j - data_race(rcu_state.gp_wake_time);
664 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",
665 rcu_state.name, gp_state_getname(rcu_state.gp_state),
666 rcu_state.gp_state, t ? t->state : 0x1ffffL,
667 ja, jr, jw, (long)data_race(rcu_state.gp_wake_seq),
668 (long)data_race(rcu_state.gp_seq),
669 (long)data_race(rcu_get_root()->gp_seq_needed),
670 data_race(rcu_state.gp_flags));
671 rcu_for_each_node_breadth_first(rnp) {
672 if (ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq),
673 READ_ONCE(rnp->gp_seq_needed)))
674 continue;
675 pr_info("\trcu_node %d:%d ->gp_seq %ld ->gp_seq_needed %ld\n",
676 rnp->grplo, rnp->grphi, (long)data_race(rnp->gp_seq),
677 (long)data_race(rnp->gp_seq_needed));
678 if (!rcu_is_leaf_node(rnp))
679 continue;
680 for_each_leaf_node_possible_cpu(rnp, cpu) {
681 rdp = per_cpu_ptr(&rcu_data, cpu);
682 if (READ_ONCE(rdp->gpwrap) ||
683 ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq),
684 READ_ONCE(rdp->gp_seq_needed)))
685 continue;
686 pr_info("\tcpu %d ->gp_seq_needed %ld\n",
687 cpu, (long)data_race(rdp->gp_seq_needed));
688 }
689 }
690 for_each_possible_cpu(cpu) {
691 rdp = per_cpu_ptr(&rcu_data, cpu);
692 cbs += data_race(rdp->n_cbs_invoked);
693 if (rcu_segcblist_is_offloaded(&rdp->cblist))
694 show_rcu_nocb_state(rdp);
695 }
696 pr_info("RCU callbacks invoked since boot: %lu\n", cbs);
697 show_rcu_tasks_gp_kthreads();
698}
699EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
700
701/*
702 * This function checks for grace-period requests that fail to motivate
703 * RCU to come out of its idle mode.
704 */
705static void rcu_check_gp_start_stall(struct rcu_node *rnp, struct rcu_data *rdp,
706 const unsigned long gpssdelay)
707{
708 unsigned long flags;
709 unsigned long j;
710 struct rcu_node *rnp_root = rcu_get_root();
711 static atomic_t warned = ATOMIC_INIT(0);
712
713 if (!IS_ENABLED(CONFIG_PROVE_RCU) || rcu_gp_in_progress() ||
714 ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
715 READ_ONCE(rnp_root->gp_seq_needed)) ||
716 !smp_load_acquire(&rcu_state.gp_kthread)) // Get stable kthread.
717 return;
718 j = jiffies; /* Expensive access, and in common case don't get here. */
719 if (time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
720 time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
721 atomic_read(&warned))
722 return;
723
724 raw_spin_lock_irqsave_rcu_node(rnp, flags);
725 j = jiffies;
726 if (rcu_gp_in_progress() ||
727 ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
728 READ_ONCE(rnp_root->gp_seq_needed)) ||
729 time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
730 time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
731 atomic_read(&warned)) {
732 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
733 return;
734 }
735 /* Hold onto the leaf lock to make others see warned==1. */
736
737 if (rnp_root != rnp)
738 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
739 j = jiffies;
740 if (rcu_gp_in_progress() ||
741 ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
742 READ_ONCE(rnp_root->gp_seq_needed)) ||
743 time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
744 time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
745 atomic_xchg(&warned, 1)) {
746 if (rnp_root != rnp)
747 /* irqs remain disabled. */
748 raw_spin_unlock_rcu_node(rnp_root);
749 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
750 return;
751 }
752 WARN_ON(1);
753 if (rnp_root != rnp)
754 raw_spin_unlock_rcu_node(rnp_root);
755 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
756 show_rcu_gp_kthreads();
757}
758
759/*
760 * Do a forward-progress check for rcutorture. This is normally invoked
761 * due to an OOM event. The argument "j" gives the time period during
762 * which rcutorture would like progress to have been made.
763 */
764void rcu_fwd_progress_check(unsigned long j)
765{
766 unsigned long cbs;
767 int cpu;
768 unsigned long max_cbs = 0;
769 int max_cpu = -1;
770 struct rcu_data *rdp;
771
772 if (rcu_gp_in_progress()) {
773 pr_info("%s: GP age %lu jiffies\n",
774 __func__, jiffies - rcu_state.gp_start);
775 show_rcu_gp_kthreads();
776 } else {
777 pr_info("%s: Last GP end %lu jiffies ago\n",
778 __func__, jiffies - rcu_state.gp_end);
779 preempt_disable();
780 rdp = this_cpu_ptr(&rcu_data);
781 rcu_check_gp_start_stall(rdp->mynode, rdp, j);
782 preempt_enable();
783 }
784 for_each_possible_cpu(cpu) {
785 cbs = rcu_get_n_cbs_cpu(cpu);
786 if (!cbs)
787 continue;
788 if (max_cpu < 0)
789 pr_info("%s: callbacks", __func__);
790 pr_cont(" %d: %lu", cpu, cbs);
791 if (cbs <= max_cbs)
792 continue;
793 max_cbs = cbs;
794 max_cpu = cpu;
795 }
796 if (max_cpu >= 0)
797 pr_cont("\n");
798}
799EXPORT_SYMBOL_GPL(rcu_fwd_progress_check);
800
801/* Commandeer a sysrq key to dump RCU's tree. */
802static bool sysrq_rcu;
803module_param(sysrq_rcu, bool, 0444);
804
805/* Dump grace-period-request information due to commandeered sysrq. */
806static void sysrq_show_rcu(int key)
807{
808 show_rcu_gp_kthreads();
809}
810
811static const struct sysrq_key_op sysrq_rcudump_op = {
812 .handler = sysrq_show_rcu,
813 .help_msg = "show-rcu(y)",
814 .action_msg = "Show RCU tree",
815 .enable_mask = SYSRQ_ENABLE_DUMP,
816};
817
818static int __init rcu_sysrq_init(void)
819{
820 if (sysrq_rcu)
821 return register_sysrq_key('y', &sysrq_rcudump_op);
822 return 0;
823}
824early_initcall(rcu_sysrq_init);