Loading...
Note: File does not exist in v3.15.
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. */
17int sysctl_panic_on_rcu_stall __read_mostly;
18int sysctl_max_rcu_stall_to_panic __read_mostly;
19
20#ifdef CONFIG_PROVE_RCU
21#define RCU_STALL_DELAY_DELTA (5 * HZ)
22#else
23#define RCU_STALL_DELAY_DELTA 0
24#endif
25#define RCU_STALL_MIGHT_DIV 8
26#define RCU_STALL_MIGHT_MIN (2 * HZ)
27
28/* Limit-check stall timeouts specified at boottime and runtime. */
29int 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}
46EXPORT_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 */
61bool 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. */
79void rcu_sysrq_start(void)
80{
81 if (!rcu_cpu_stall_suppress)
82 rcu_cpu_stall_suppress = 2;
83}
84
85void 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. */
92static 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
98static struct notifier_block rcu_panic_block = {
99 .notifier_call = rcu_panic,
100};
101
102static int __init check_cpu_stall_init(void)
103{
104 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
105 return 0;
106}
107early_initcall(check_cpu_stall_init);
108
109/* If so specified via sysctl, panic, yielding cleaner stall-warning output. */
110static void panic_on_rcu_stall(void)
111{
112 static int cpu_stall;
113
114 if (++cpu_stall < sysctl_max_rcu_stall_to_panic)
115 return;
116
117 if (sysctl_panic_on_rcu_stall)
118 panic("RCU Stall\n");
119}
120
121/**
122 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
123 *
124 * Set the stall-warning timeout way off into the future, thus preventing
125 * any RCU CPU stall-warning messages from appearing in the current set of
126 * RCU grace periods.
127 *
128 * The caller must disable hard irqs.
129 */
130void rcu_cpu_stall_reset(void)
131{
132 WRITE_ONCE(rcu_state.jiffies_stall, jiffies + ULONG_MAX / 2);
133}
134
135//////////////////////////////////////////////////////////////////////////////
136//
137// Interaction with RCU grace periods
138
139/* Start of new grace period, so record stall time (and forcing times). */
140static void record_gp_stall_check_time(void)
141{
142 unsigned long j = jiffies;
143 unsigned long j1;
144
145 WRITE_ONCE(rcu_state.gp_start, j);
146 j1 = rcu_jiffies_till_stall_check();
147 smp_mb(); // ->gp_start before ->jiffies_stall and caller's ->gp_seq.
148 WRITE_ONCE(rcu_state.jiffies_stall, j + j1);
149 rcu_state.jiffies_resched = j + j1 / 2;
150 rcu_state.n_force_qs_gpstart = READ_ONCE(rcu_state.n_force_qs);
151}
152
153/* Zero ->ticks_this_gp and snapshot the number of RCU softirq handlers. */
154static void zero_cpu_stall_ticks(struct rcu_data *rdp)
155{
156 rdp->ticks_this_gp = 0;
157 rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
158 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
159}
160
161/*
162 * If too much time has passed in the current grace period, and if
163 * so configured, go kick the relevant kthreads.
164 */
165static void rcu_stall_kick_kthreads(void)
166{
167 unsigned long j;
168
169 if (!READ_ONCE(rcu_kick_kthreads))
170 return;
171 j = READ_ONCE(rcu_state.jiffies_kick_kthreads);
172 if (time_after(jiffies, j) && rcu_state.gp_kthread &&
173 (rcu_gp_in_progress() || READ_ONCE(rcu_state.gp_flags))) {
174 WARN_ONCE(1, "Kicking %s grace-period kthread\n",
175 rcu_state.name);
176 rcu_ftrace_dump(DUMP_ALL);
177 wake_up_process(rcu_state.gp_kthread);
178 WRITE_ONCE(rcu_state.jiffies_kick_kthreads, j + HZ);
179 }
180}
181
182/*
183 * Handler for the irq_work request posted about halfway into the RCU CPU
184 * stall timeout, and used to detect excessive irq disabling. Set state
185 * appropriately, but just complain if there is unexpected state on entry.
186 */
187static void rcu_iw_handler(struct irq_work *iwp)
188{
189 struct rcu_data *rdp;
190 struct rcu_node *rnp;
191
192 rdp = container_of(iwp, struct rcu_data, rcu_iw);
193 rnp = rdp->mynode;
194 raw_spin_lock_rcu_node(rnp);
195 if (!WARN_ON_ONCE(!rdp->rcu_iw_pending)) {
196 rdp->rcu_iw_gp_seq = rnp->gp_seq;
197 rdp->rcu_iw_pending = false;
198 }
199 raw_spin_unlock_rcu_node(rnp);
200}
201
202//////////////////////////////////////////////////////////////////////////////
203//
204// Printing RCU CPU stall warnings
205
206#ifdef CONFIG_PREEMPT_RCU
207
208/*
209 * Dump detailed information for all tasks blocking the current RCU
210 * grace period on the specified rcu_node structure.
211 */
212static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
213{
214 unsigned long flags;
215 struct task_struct *t;
216
217 raw_spin_lock_irqsave_rcu_node(rnp, flags);
218 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
219 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
220 return;
221 }
222 t = list_entry(rnp->gp_tasks->prev,
223 struct task_struct, rcu_node_entry);
224 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
225 /*
226 * We could be printing a lot while holding a spinlock.
227 * Avoid triggering hard lockup.
228 */
229 touch_nmi_watchdog();
230 sched_show_task(t);
231 }
232 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
233}
234
235// Communicate task state back to the RCU CPU stall warning request.
236struct rcu_stall_chk_rdr {
237 int nesting;
238 union rcu_special rs;
239 bool on_blkd_list;
240};
241
242/*
243 * Report out the state of a not-running task that is stalling the
244 * current RCU grace period.
245 */
246static bool check_slow_task(struct task_struct *t, void *arg)
247{
248 struct rcu_stall_chk_rdr *rscrp = arg;
249
250 if (task_curr(t))
251 return false; // It is running, so decline to inspect it.
252 rscrp->nesting = t->rcu_read_lock_nesting;
253 rscrp->rs = t->rcu_read_unlock_special;
254 rscrp->on_blkd_list = !list_empty(&t->rcu_node_entry);
255 return true;
256}
257
258/*
259 * Scan the current list of tasks blocked within RCU read-side critical
260 * sections, printing out the tid of each of the first few of them.
261 */
262static int rcu_print_task_stall(struct rcu_node *rnp, unsigned long flags)
263 __releases(rnp->lock)
264{
265 int i = 0;
266 int ndetected = 0;
267 struct rcu_stall_chk_rdr rscr;
268 struct task_struct *t;
269 struct task_struct *ts[8];
270
271 lockdep_assert_irqs_disabled();
272 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
273 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
274 return 0;
275 }
276 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
277 rnp->level, rnp->grplo, rnp->grphi);
278 t = list_entry(rnp->gp_tasks->prev,
279 struct task_struct, rcu_node_entry);
280 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
281 get_task_struct(t);
282 ts[i++] = t;
283 if (i >= ARRAY_SIZE(ts))
284 break;
285 }
286 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
287 while (i) {
288 t = ts[--i];
289 if (!try_invoke_on_locked_down_task(t, check_slow_task, &rscr))
290 pr_cont(" P%d", t->pid);
291 else
292 pr_cont(" P%d/%d:%c%c%c%c",
293 t->pid, rscr.nesting,
294 ".b"[rscr.rs.b.blocked],
295 ".q"[rscr.rs.b.need_qs],
296 ".e"[rscr.rs.b.exp_hint],
297 ".l"[rscr.on_blkd_list]);
298 lockdep_assert_irqs_disabled();
299 put_task_struct(t);
300 ndetected++;
301 }
302 pr_cont("\n");
303 return ndetected;
304}
305
306#else /* #ifdef CONFIG_PREEMPT_RCU */
307
308/*
309 * Because preemptible RCU does not exist, we never have to check for
310 * tasks blocked within RCU read-side critical sections.
311 */
312static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
313{
314}
315
316/*
317 * Because preemptible RCU does not exist, we never have to check for
318 * tasks blocked within RCU read-side critical sections.
319 */
320static int rcu_print_task_stall(struct rcu_node *rnp, unsigned long flags)
321 __releases(rnp->lock)
322{
323 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
324 return 0;
325}
326#endif /* #else #ifdef CONFIG_PREEMPT_RCU */
327
328/*
329 * Dump stacks of all tasks running on stalled CPUs. First try using
330 * NMIs, but fall back to manual remote stack tracing on architectures
331 * that don't support NMI-based stack dumps. The NMI-triggered stack
332 * traces are more accurate because they are printed by the target CPU.
333 */
334static void rcu_dump_cpu_stacks(void)
335{
336 int cpu;
337 unsigned long flags;
338 struct rcu_node *rnp;
339
340 rcu_for_each_leaf_node(rnp) {
341 raw_spin_lock_irqsave_rcu_node(rnp, flags);
342 for_each_leaf_node_possible_cpu(rnp, cpu)
343 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
344 if (cpu_is_offline(cpu))
345 pr_err("Offline CPU %d blocking current GP.\n", cpu);
346 else if (!trigger_single_cpu_backtrace(cpu))
347 dump_cpu_task(cpu);
348 }
349 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
350 }
351}
352
353#ifdef CONFIG_RCU_FAST_NO_HZ
354
355static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
356{
357 struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
358
359 sprintf(cp, "last_accelerate: %04lx/%04lx dyntick_enabled: %d",
360 rdp->last_accelerate & 0xffff, jiffies & 0xffff,
361 !!rdp->tick_nohz_enabled_snap);
362}
363
364#else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
365
366static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
367{
368 *cp = '\0';
369}
370
371#endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
372
373static const char * const gp_state_names[] = {
374 [RCU_GP_IDLE] = "RCU_GP_IDLE",
375 [RCU_GP_WAIT_GPS] = "RCU_GP_WAIT_GPS",
376 [RCU_GP_DONE_GPS] = "RCU_GP_DONE_GPS",
377 [RCU_GP_ONOFF] = "RCU_GP_ONOFF",
378 [RCU_GP_INIT] = "RCU_GP_INIT",
379 [RCU_GP_WAIT_FQS] = "RCU_GP_WAIT_FQS",
380 [RCU_GP_DOING_FQS] = "RCU_GP_DOING_FQS",
381 [RCU_GP_CLEANUP] = "RCU_GP_CLEANUP",
382 [RCU_GP_CLEANED] = "RCU_GP_CLEANED",
383};
384
385/*
386 * Convert a ->gp_state value to a character string.
387 */
388static const char *gp_state_getname(short gs)
389{
390 if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
391 return "???";
392 return gp_state_names[gs];
393}
394
395/* Is the RCU grace-period kthread being starved of CPU time? */
396static bool rcu_is_gp_kthread_starving(unsigned long *jp)
397{
398 unsigned long j = jiffies - READ_ONCE(rcu_state.gp_activity);
399
400 if (jp)
401 *jp = j;
402 return j > 2 * HZ;
403}
404
405/*
406 * Print out diagnostic information for the specified stalled CPU.
407 *
408 * If the specified CPU is aware of the current RCU grace period, then
409 * print the number of scheduling clock interrupts the CPU has taken
410 * during the time that it has been aware. Otherwise, print the number
411 * of RCU grace periods that this CPU is ignorant of, for example, "1"
412 * if the CPU was aware of the previous grace period.
413 *
414 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
415 */
416static void print_cpu_stall_info(int cpu)
417{
418 unsigned long delta;
419 bool falsepositive;
420 char fast_no_hz[72];
421 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
422 char *ticks_title;
423 unsigned long ticks_value;
424
425 /*
426 * We could be printing a lot while holding a spinlock. Avoid
427 * triggering hard lockup.
428 */
429 touch_nmi_watchdog();
430
431 ticks_value = rcu_seq_ctr(rcu_state.gp_seq - rdp->gp_seq);
432 if (ticks_value) {
433 ticks_title = "GPs behind";
434 } else {
435 ticks_title = "ticks this GP";
436 ticks_value = rdp->ticks_this_gp;
437 }
438 print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
439 delta = rcu_seq_ctr(rdp->mynode->gp_seq - rdp->rcu_iw_gp_seq);
440 falsepositive = rcu_is_gp_kthread_starving(NULL) &&
441 rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp));
442 pr_err("\t%d-%c%c%c%c: (%lu %s) idle=%03x/%ld/%#lx softirq=%u/%u fqs=%ld %s%s\n",
443 cpu,
444 "O."[!!cpu_online(cpu)],
445 "o."[!!(rdp->grpmask & rdp->mynode->qsmaskinit)],
446 "N."[!!(rdp->grpmask & rdp->mynode->qsmaskinitnext)],
447 !IS_ENABLED(CONFIG_IRQ_WORK) ? '?' :
448 rdp->rcu_iw_pending ? (int)min(delta, 9UL) + '0' :
449 "!."[!delta],
450 ticks_value, ticks_title,
451 rcu_dynticks_snap(rdp) & 0xfff,
452 rdp->dynticks_nesting, rdp->dynticks_nmi_nesting,
453 rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
454 data_race(rcu_state.n_force_qs) - rcu_state.n_force_qs_gpstart,
455 fast_no_hz,
456 falsepositive ? " (false positive?)" : "");
457}
458
459/* Complain about starvation of grace-period kthread. */
460static void rcu_check_gp_kthread_starvation(void)
461{
462 int cpu;
463 struct task_struct *gpk = rcu_state.gp_kthread;
464 unsigned long j;
465
466 if (rcu_is_gp_kthread_starving(&j)) {
467 cpu = gpk ? task_cpu(gpk) : -1;
468 pr_err("%s kthread starved for %ld jiffies! g%ld f%#x %s(%d) ->state=%#x ->cpu=%d\n",
469 rcu_state.name, j,
470 (long)rcu_seq_current(&rcu_state.gp_seq),
471 data_race(rcu_state.gp_flags),
472 gp_state_getname(rcu_state.gp_state), rcu_state.gp_state,
473 gpk ? gpk->__state : ~0, cpu);
474 if (gpk) {
475 pr_err("\tUnless %s kthread gets sufficient CPU time, OOM is now expected behavior.\n", rcu_state.name);
476 pr_err("RCU grace-period kthread stack dump:\n");
477 sched_show_task(gpk);
478 if (cpu >= 0) {
479 if (cpu_is_offline(cpu)) {
480 pr_err("RCU GP kthread last ran on offline CPU %d.\n", cpu);
481 } else {
482 pr_err("Stack dump where RCU GP kthread last ran:\n");
483 if (!trigger_single_cpu_backtrace(cpu))
484 dump_cpu_task(cpu);
485 }
486 }
487 wake_up_process(gpk);
488 }
489 }
490}
491
492/* Complain about missing wakeups from expired fqs wait timer */
493static void rcu_check_gp_kthread_expired_fqs_timer(void)
494{
495 struct task_struct *gpk = rcu_state.gp_kthread;
496 short gp_state;
497 unsigned long jiffies_fqs;
498 int cpu;
499
500 /*
501 * Order reads of .gp_state and .jiffies_force_qs.
502 * Matching smp_wmb() is present in rcu_gp_fqs_loop().
503 */
504 gp_state = smp_load_acquire(&rcu_state.gp_state);
505 jiffies_fqs = READ_ONCE(rcu_state.jiffies_force_qs);
506
507 if (gp_state == RCU_GP_WAIT_FQS &&
508 time_after(jiffies, jiffies_fqs + RCU_STALL_MIGHT_MIN) &&
509 gpk && !READ_ONCE(gpk->on_rq)) {
510 cpu = task_cpu(gpk);
511 pr_err("%s kthread timer wakeup didn't happen for %ld jiffies! g%ld f%#x %s(%d) ->state=%#x\n",
512 rcu_state.name, (jiffies - jiffies_fqs),
513 (long)rcu_seq_current(&rcu_state.gp_seq),
514 data_race(rcu_state.gp_flags),
515 gp_state_getname(RCU_GP_WAIT_FQS), RCU_GP_WAIT_FQS,
516 gpk->__state);
517 pr_err("\tPossible timer handling issue on cpu=%d timer-softirq=%u\n",
518 cpu, kstat_softirqs_cpu(TIMER_SOFTIRQ, cpu));
519 }
520}
521
522static void print_other_cpu_stall(unsigned long gp_seq, unsigned long gps)
523{
524 int cpu;
525 unsigned long flags;
526 unsigned long gpa;
527 unsigned long j;
528 int ndetected = 0;
529 struct rcu_node *rnp;
530 long totqlen = 0;
531
532 lockdep_assert_irqs_disabled();
533
534 /* Kick and suppress, if so configured. */
535 rcu_stall_kick_kthreads();
536 if (rcu_stall_is_suppressed())
537 return;
538
539 /*
540 * OK, time to rat on our buddy...
541 * See Documentation/RCU/stallwarn.rst for info on how to debug
542 * RCU CPU stall warnings.
543 */
544 trace_rcu_stall_warning(rcu_state.name, TPS("StallDetected"));
545 pr_err("INFO: %s detected stalls on CPUs/tasks:\n", rcu_state.name);
546 rcu_for_each_leaf_node(rnp) {
547 raw_spin_lock_irqsave_rcu_node(rnp, flags);
548 if (rnp->qsmask != 0) {
549 for_each_leaf_node_possible_cpu(rnp, cpu)
550 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
551 print_cpu_stall_info(cpu);
552 ndetected++;
553 }
554 }
555 ndetected += rcu_print_task_stall(rnp, flags); // Releases rnp->lock.
556 lockdep_assert_irqs_disabled();
557 }
558
559 for_each_possible_cpu(cpu)
560 totqlen += rcu_get_n_cbs_cpu(cpu);
561 pr_cont("\t(detected by %d, t=%ld jiffies, g=%ld, q=%lu)\n",
562 smp_processor_id(), (long)(jiffies - gps),
563 (long)rcu_seq_current(&rcu_state.gp_seq), totqlen);
564 if (ndetected) {
565 rcu_dump_cpu_stacks();
566
567 /* Complain about tasks blocking the grace period. */
568 rcu_for_each_leaf_node(rnp)
569 rcu_print_detail_task_stall_rnp(rnp);
570 } else {
571 if (rcu_seq_current(&rcu_state.gp_seq) != gp_seq) {
572 pr_err("INFO: Stall ended before state dump start\n");
573 } else {
574 j = jiffies;
575 gpa = data_race(rcu_state.gp_activity);
576 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
577 rcu_state.name, j - gpa, j, gpa,
578 data_race(jiffies_till_next_fqs),
579 rcu_get_root()->qsmask);
580 }
581 }
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
587 rcu_check_gp_kthread_expired_fqs_timer();
588 rcu_check_gp_kthread_starvation();
589
590 panic_on_rcu_stall();
591
592 rcu_force_quiescent_state(); /* Kick them all. */
593}
594
595static void print_cpu_stall(unsigned long gps)
596{
597 int cpu;
598 unsigned long flags;
599 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
600 struct rcu_node *rnp = rcu_get_root();
601 long totqlen = 0;
602
603 lockdep_assert_irqs_disabled();
604
605 /* Kick and suppress, if so configured. */
606 rcu_stall_kick_kthreads();
607 if (rcu_stall_is_suppressed())
608 return;
609
610 /*
611 * OK, time to rat on ourselves...
612 * See Documentation/RCU/stallwarn.rst for info on how to debug
613 * RCU CPU stall warnings.
614 */
615 trace_rcu_stall_warning(rcu_state.name, TPS("SelfDetected"));
616 pr_err("INFO: %s self-detected stall on CPU\n", rcu_state.name);
617 raw_spin_lock_irqsave_rcu_node(rdp->mynode, flags);
618 print_cpu_stall_info(smp_processor_id());
619 raw_spin_unlock_irqrestore_rcu_node(rdp->mynode, flags);
620 for_each_possible_cpu(cpu)
621 totqlen += rcu_get_n_cbs_cpu(cpu);
622 pr_cont("\t(t=%lu jiffies g=%ld q=%lu)\n",
623 jiffies - gps,
624 (long)rcu_seq_current(&rcu_state.gp_seq), totqlen);
625
626 rcu_check_gp_kthread_expired_fqs_timer();
627 rcu_check_gp_kthread_starvation();
628
629 rcu_dump_cpu_stacks();
630
631 raw_spin_lock_irqsave_rcu_node(rnp, flags);
632 /* Rewrite if needed in case of slow consoles. */
633 if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
634 WRITE_ONCE(rcu_state.jiffies_stall,
635 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
636 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
637
638 panic_on_rcu_stall();
639
640 /*
641 * Attempt to revive the RCU machinery by forcing a context switch.
642 *
643 * A context switch would normally allow the RCU state machine to make
644 * progress and it could be we're stuck in kernel space without context
645 * switches for an entirely unreasonable amount of time.
646 */
647 set_tsk_need_resched(current);
648 set_preempt_need_resched();
649}
650
651static void check_cpu_stall(struct rcu_data *rdp)
652{
653 unsigned long gs1;
654 unsigned long gs2;
655 unsigned long gps;
656 unsigned long j;
657 unsigned long jn;
658 unsigned long js;
659 struct rcu_node *rnp;
660
661 lockdep_assert_irqs_disabled();
662 if ((rcu_stall_is_suppressed() && !READ_ONCE(rcu_kick_kthreads)) ||
663 !rcu_gp_in_progress())
664 return;
665 rcu_stall_kick_kthreads();
666 j = jiffies;
667
668 /*
669 * Lots of memory barriers to reject false positives.
670 *
671 * The idea is to pick up rcu_state.gp_seq, then
672 * rcu_state.jiffies_stall, then rcu_state.gp_start, and finally
673 * another copy of rcu_state.gp_seq. These values are updated in
674 * the opposite order with memory barriers (or equivalent) during
675 * grace-period initialization and cleanup. Now, a false positive
676 * can occur if we get an new value of rcu_state.gp_start and a old
677 * value of rcu_state.jiffies_stall. But given the memory barriers,
678 * the only way that this can happen is if one grace period ends
679 * and another starts between these two fetches. This is detected
680 * by comparing the second fetch of rcu_state.gp_seq with the
681 * previous fetch from rcu_state.gp_seq.
682 *
683 * Given this check, comparisons of jiffies, rcu_state.jiffies_stall,
684 * and rcu_state.gp_start suffice to forestall false positives.
685 */
686 gs1 = READ_ONCE(rcu_state.gp_seq);
687 smp_rmb(); /* Pick up ->gp_seq first... */
688 js = READ_ONCE(rcu_state.jiffies_stall);
689 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
690 gps = READ_ONCE(rcu_state.gp_start);
691 smp_rmb(); /* ...and finally ->gp_start before ->gp_seq again. */
692 gs2 = READ_ONCE(rcu_state.gp_seq);
693 if (gs1 != gs2 ||
694 ULONG_CMP_LT(j, js) ||
695 ULONG_CMP_GE(gps, js))
696 return; /* No stall or GP completed since entering function. */
697 rnp = rdp->mynode;
698 jn = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
699 if (rcu_gp_in_progress() &&
700 (READ_ONCE(rnp->qsmask) & rdp->grpmask) &&
701 cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
702
703 /*
704 * If a virtual machine is stopped by the host it can look to
705 * the watchdog like an RCU stall. Check to see if the host
706 * stopped the vm.
707 */
708 if (kvm_check_and_clear_guest_paused())
709 return;
710
711 /* We haven't checked in, so go dump stack. */
712 print_cpu_stall(gps);
713 if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
714 rcu_ftrace_dump(DUMP_ALL);
715
716 } else if (rcu_gp_in_progress() &&
717 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY) &&
718 cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
719
720 /*
721 * If a virtual machine is stopped by the host it can look to
722 * the watchdog like an RCU stall. Check to see if the host
723 * stopped the vm.
724 */
725 if (kvm_check_and_clear_guest_paused())
726 return;
727
728 /* They had a few time units to dump stack, so complain. */
729 print_other_cpu_stall(gs2, gps);
730 if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
731 rcu_ftrace_dump(DUMP_ALL);
732 }
733}
734
735//////////////////////////////////////////////////////////////////////////////
736//
737// RCU forward-progress mechanisms, including of callback invocation.
738
739
740/*
741 * Check to see if a failure to end RCU priority inversion was due to
742 * a CPU not passing through a quiescent state. When this happens, there
743 * is nothing that RCU priority boosting can do to help, so we shouldn't
744 * count this as an RCU priority boosting failure. A return of true says
745 * RCU priority boosting is to blame, and false says otherwise. If false
746 * is returned, the first of the CPUs to blame is stored through cpup.
747 * If there was no CPU blocking the current grace period, but also nothing
748 * in need of being boosted, *cpup is set to -1. This can happen in case
749 * of vCPU preemption while the last CPU is reporting its quiscent state,
750 * for example.
751 *
752 * If cpup is NULL, then a lockless quick check is carried out, suitable
753 * for high-rate usage. On the other hand, if cpup is non-NULL, each
754 * rcu_node structure's ->lock is acquired, ruling out high-rate usage.
755 */
756bool rcu_check_boost_fail(unsigned long gp_state, int *cpup)
757{
758 bool atb = false;
759 int cpu;
760 unsigned long flags;
761 struct rcu_node *rnp;
762
763 rcu_for_each_leaf_node(rnp) {
764 if (!cpup) {
765 if (READ_ONCE(rnp->qsmask)) {
766 return false;
767 } else {
768 if (READ_ONCE(rnp->gp_tasks))
769 atb = true;
770 continue;
771 }
772 }
773 *cpup = -1;
774 raw_spin_lock_irqsave_rcu_node(rnp, flags);
775 if (rnp->gp_tasks)
776 atb = true;
777 if (!rnp->qsmask) {
778 // No CPUs without quiescent states for this rnp.
779 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
780 continue;
781 }
782 // Find the first holdout CPU.
783 for_each_leaf_node_possible_cpu(rnp, cpu) {
784 if (rnp->qsmask & (1UL << (cpu - rnp->grplo))) {
785 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
786 *cpup = cpu;
787 return false;
788 }
789 }
790 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
791 }
792 // Can't blame CPUs, so must blame RCU priority boosting.
793 return atb;
794}
795EXPORT_SYMBOL_GPL(rcu_check_boost_fail);
796
797/*
798 * Show the state of the grace-period kthreads.
799 */
800void show_rcu_gp_kthreads(void)
801{
802 unsigned long cbs = 0;
803 int cpu;
804 unsigned long j;
805 unsigned long ja;
806 unsigned long jr;
807 unsigned long js;
808 unsigned long jw;
809 struct rcu_data *rdp;
810 struct rcu_node *rnp;
811 struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
812
813 j = jiffies;
814 ja = j - data_race(rcu_state.gp_activity);
815 jr = j - data_race(rcu_state.gp_req_activity);
816 js = j - data_race(rcu_state.gp_start);
817 jw = j - data_race(rcu_state.gp_wake_time);
818 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",
819 rcu_state.name, gp_state_getname(rcu_state.gp_state),
820 rcu_state.gp_state, t ? t->__state : 0x1ffff, t ? t->rt_priority : 0xffU,
821 js, ja, jr, jw, (long)data_race(rcu_state.gp_wake_seq),
822 (long)data_race(rcu_state.gp_seq),
823 (long)data_race(rcu_get_root()->gp_seq_needed),
824 data_race(rcu_state.gp_max),
825 data_race(rcu_state.gp_flags));
826 rcu_for_each_node_breadth_first(rnp) {
827 if (ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq), READ_ONCE(rnp->gp_seq_needed)) &&
828 !data_race(rnp->qsmask) && !data_race(rnp->boost_tasks) &&
829 !data_race(rnp->exp_tasks) && !data_race(rnp->gp_tasks))
830 continue;
831 pr_info("\trcu_node %d:%d ->gp_seq %ld ->gp_seq_needed %ld ->qsmask %#lx %c%c%c%c ->n_boosts %ld\n",
832 rnp->grplo, rnp->grphi,
833 (long)data_race(rnp->gp_seq), (long)data_race(rnp->gp_seq_needed),
834 data_race(rnp->qsmask),
835 ".b"[!!data_race(rnp->boost_kthread_task)],
836 ".B"[!!data_race(rnp->boost_tasks)],
837 ".E"[!!data_race(rnp->exp_tasks)],
838 ".G"[!!data_race(rnp->gp_tasks)],
839 data_race(rnp->n_boosts));
840 if (!rcu_is_leaf_node(rnp))
841 continue;
842 for_each_leaf_node_possible_cpu(rnp, cpu) {
843 rdp = per_cpu_ptr(&rcu_data, cpu);
844 if (READ_ONCE(rdp->gpwrap) ||
845 ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq),
846 READ_ONCE(rdp->gp_seq_needed)))
847 continue;
848 pr_info("\tcpu %d ->gp_seq_needed %ld\n",
849 cpu, (long)data_race(rdp->gp_seq_needed));
850 }
851 }
852 for_each_possible_cpu(cpu) {
853 rdp = per_cpu_ptr(&rcu_data, cpu);
854 cbs += data_race(rdp->n_cbs_invoked);
855 if (rcu_segcblist_is_offloaded(&rdp->cblist))
856 show_rcu_nocb_state(rdp);
857 }
858 pr_info("RCU callbacks invoked since boot: %lu\n", cbs);
859 show_rcu_tasks_gp_kthreads();
860}
861EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
862
863/*
864 * This function checks for grace-period requests that fail to motivate
865 * RCU to come out of its idle mode.
866 */
867static void rcu_check_gp_start_stall(struct rcu_node *rnp, struct rcu_data *rdp,
868 const unsigned long gpssdelay)
869{
870 unsigned long flags;
871 unsigned long j;
872 struct rcu_node *rnp_root = rcu_get_root();
873 static atomic_t warned = ATOMIC_INIT(0);
874
875 if (!IS_ENABLED(CONFIG_PROVE_RCU) || rcu_gp_in_progress() ||
876 ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
877 READ_ONCE(rnp_root->gp_seq_needed)) ||
878 !smp_load_acquire(&rcu_state.gp_kthread)) // Get stable kthread.
879 return;
880 j = jiffies; /* Expensive access, and in common case don't get here. */
881 if (time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
882 time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
883 atomic_read(&warned))
884 return;
885
886 raw_spin_lock_irqsave_rcu_node(rnp, flags);
887 j = jiffies;
888 if (rcu_gp_in_progress() ||
889 ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
890 READ_ONCE(rnp_root->gp_seq_needed)) ||
891 time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
892 time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
893 atomic_read(&warned)) {
894 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
895 return;
896 }
897 /* Hold onto the leaf lock to make others see warned==1. */
898
899 if (rnp_root != rnp)
900 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
901 j = jiffies;
902 if (rcu_gp_in_progress() ||
903 ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
904 READ_ONCE(rnp_root->gp_seq_needed)) ||
905 time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
906 time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
907 atomic_xchg(&warned, 1)) {
908 if (rnp_root != rnp)
909 /* irqs remain disabled. */
910 raw_spin_unlock_rcu_node(rnp_root);
911 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
912 return;
913 }
914 WARN_ON(1);
915 if (rnp_root != rnp)
916 raw_spin_unlock_rcu_node(rnp_root);
917 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
918 show_rcu_gp_kthreads();
919}
920
921/*
922 * Do a forward-progress check for rcutorture. This is normally invoked
923 * due to an OOM event. The argument "j" gives the time period during
924 * which rcutorture would like progress to have been made.
925 */
926void rcu_fwd_progress_check(unsigned long j)
927{
928 unsigned long cbs;
929 int cpu;
930 unsigned long max_cbs = 0;
931 int max_cpu = -1;
932 struct rcu_data *rdp;
933
934 if (rcu_gp_in_progress()) {
935 pr_info("%s: GP age %lu jiffies\n",
936 __func__, jiffies - rcu_state.gp_start);
937 show_rcu_gp_kthreads();
938 } else {
939 pr_info("%s: Last GP end %lu jiffies ago\n",
940 __func__, jiffies - rcu_state.gp_end);
941 preempt_disable();
942 rdp = this_cpu_ptr(&rcu_data);
943 rcu_check_gp_start_stall(rdp->mynode, rdp, j);
944 preempt_enable();
945 }
946 for_each_possible_cpu(cpu) {
947 cbs = rcu_get_n_cbs_cpu(cpu);
948 if (!cbs)
949 continue;
950 if (max_cpu < 0)
951 pr_info("%s: callbacks", __func__);
952 pr_cont(" %d: %lu", cpu, cbs);
953 if (cbs <= max_cbs)
954 continue;
955 max_cbs = cbs;
956 max_cpu = cpu;
957 }
958 if (max_cpu >= 0)
959 pr_cont("\n");
960}
961EXPORT_SYMBOL_GPL(rcu_fwd_progress_check);
962
963/* Commandeer a sysrq key to dump RCU's tree. */
964static bool sysrq_rcu;
965module_param(sysrq_rcu, bool, 0444);
966
967/* Dump grace-period-request information due to commandeered sysrq. */
968static void sysrq_show_rcu(int key)
969{
970 show_rcu_gp_kthreads();
971}
972
973static const struct sysrq_key_op sysrq_rcudump_op = {
974 .handler = sysrq_show_rcu,
975 .help_msg = "show-rcu(y)",
976 .action_msg = "Show RCU tree",
977 .enable_mask = SYSRQ_ENABLE_DUMP,
978};
979
980static int __init rcu_sysrq_init(void)
981{
982 if (sysrq_rcu)
983 return register_sysrq_key('y', &sysrq_rcudump_op);
984 return 0;
985}
986early_initcall(rcu_sysrq_init);