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#include <linux/kvm_para.h>
11#include <linux/rcu_notifier.h>
12
13//////////////////////////////////////////////////////////////////////////////
14//
15// Controlling CPU stall warnings, including delay calculation.
16
17/* panic() on RCU Stall sysctl. */
18int sysctl_panic_on_rcu_stall __read_mostly;
19int sysctl_max_rcu_stall_to_panic __read_mostly;
20
21#ifdef CONFIG_PROVE_RCU
22#define RCU_STALL_DELAY_DELTA (5 * HZ)
23#else
24#define RCU_STALL_DELAY_DELTA 0
25#endif
26#define RCU_STALL_MIGHT_DIV 8
27#define RCU_STALL_MIGHT_MIN (2 * HZ)
28
29int rcu_exp_jiffies_till_stall_check(void)
30{
31 int cpu_stall_timeout = READ_ONCE(rcu_exp_cpu_stall_timeout);
32 int exp_stall_delay_delta = 0;
33 int till_stall_check;
34
35 // Zero says to use rcu_cpu_stall_timeout, but in milliseconds.
36 if (!cpu_stall_timeout)
37 cpu_stall_timeout = jiffies_to_msecs(rcu_jiffies_till_stall_check());
38
39 // Limit check must be consistent with the Kconfig limits for
40 // CONFIG_RCU_EXP_CPU_STALL_TIMEOUT, so check the allowed range.
41 // The minimum clamped value is "2UL", because at least one full
42 // tick has to be guaranteed.
43 till_stall_check = clamp(msecs_to_jiffies(cpu_stall_timeout), 2UL, 300UL * HZ);
44
45 if (cpu_stall_timeout && jiffies_to_msecs(till_stall_check) != cpu_stall_timeout)
46 WRITE_ONCE(rcu_exp_cpu_stall_timeout, jiffies_to_msecs(till_stall_check));
47
48#ifdef CONFIG_PROVE_RCU
49 /* Add extra ~25% out of till_stall_check. */
50 exp_stall_delay_delta = ((till_stall_check * 25) / 100) + 1;
51#endif
52
53 return till_stall_check + exp_stall_delay_delta;
54}
55EXPORT_SYMBOL_GPL(rcu_exp_jiffies_till_stall_check);
56
57/* Limit-check stall timeouts specified at boottime and runtime. */
58int rcu_jiffies_till_stall_check(void)
59{
60 int till_stall_check = READ_ONCE(rcu_cpu_stall_timeout);
61
62 /*
63 * Limit check must be consistent with the Kconfig limits
64 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
65 */
66 if (till_stall_check < 3) {
67 WRITE_ONCE(rcu_cpu_stall_timeout, 3);
68 till_stall_check = 3;
69 } else if (till_stall_check > 300) {
70 WRITE_ONCE(rcu_cpu_stall_timeout, 300);
71 till_stall_check = 300;
72 }
73 return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
74}
75EXPORT_SYMBOL_GPL(rcu_jiffies_till_stall_check);
76
77/**
78 * rcu_gp_might_be_stalled - Is it likely that the grace period is stalled?
79 *
80 * Returns @true if the current grace period is sufficiently old that
81 * it is reasonable to assume that it might be stalled. This can be
82 * useful when deciding whether to allocate memory to enable RCU-mediated
83 * freeing on the one hand or just invoking synchronize_rcu() on the other.
84 * The latter is preferable when the grace period is stalled.
85 *
86 * Note that sampling of the .gp_start and .gp_seq fields must be done
87 * carefully to avoid false positives at the beginnings and ends of
88 * grace periods.
89 */
90bool rcu_gp_might_be_stalled(void)
91{
92 unsigned long d = rcu_jiffies_till_stall_check() / RCU_STALL_MIGHT_DIV;
93 unsigned long j = jiffies;
94
95 if (d < RCU_STALL_MIGHT_MIN)
96 d = RCU_STALL_MIGHT_MIN;
97 smp_mb(); // jiffies before .gp_seq to avoid false positives.
98 if (!rcu_gp_in_progress())
99 return false;
100 // Long delays at this point avoids false positive, but a delay
101 // of ULONG_MAX/4 jiffies voids your no-false-positive warranty.
102 smp_mb(); // .gp_seq before second .gp_start
103 // And ditto here.
104 return !time_before(j, READ_ONCE(rcu_state.gp_start) + d);
105}
106
107/* Don't do RCU CPU stall warnings during long sysrq printouts. */
108void rcu_sysrq_start(void)
109{
110 if (!rcu_cpu_stall_suppress)
111 rcu_cpu_stall_suppress = 2;
112}
113
114void rcu_sysrq_end(void)
115{
116 if (rcu_cpu_stall_suppress == 2)
117 rcu_cpu_stall_suppress = 0;
118}
119
120/* Don't print RCU CPU stall warnings during a kernel panic. */
121static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
122{
123 rcu_cpu_stall_suppress = 1;
124 return NOTIFY_DONE;
125}
126
127static struct notifier_block rcu_panic_block = {
128 .notifier_call = rcu_panic,
129};
130
131static int __init check_cpu_stall_init(void)
132{
133 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
134 return 0;
135}
136early_initcall(check_cpu_stall_init);
137
138/* If so specified via sysctl, panic, yielding cleaner stall-warning output. */
139static void panic_on_rcu_stall(void)
140{
141 static int cpu_stall;
142
143 if (++cpu_stall < sysctl_max_rcu_stall_to_panic)
144 return;
145
146 if (sysctl_panic_on_rcu_stall)
147 panic("RCU Stall\n");
148}
149
150/**
151 * rcu_cpu_stall_reset - restart stall-warning timeout for current grace period
152 *
153 * To perform the reset request from the caller, disable stall detection until
154 * 3 fqs loops have passed. This is required to ensure a fresh jiffies is
155 * loaded. It should be safe to do from the fqs loop as enough timer
156 * interrupts and context switches should have passed.
157 *
158 * The caller must disable hard irqs.
159 */
160void rcu_cpu_stall_reset(void)
161{
162 WRITE_ONCE(rcu_state.nr_fqs_jiffies_stall, 3);
163 WRITE_ONCE(rcu_state.jiffies_stall, ULONG_MAX);
164}
165
166//////////////////////////////////////////////////////////////////////////////
167//
168// Interaction with RCU grace periods
169
170/* Start of new grace period, so record stall time (and forcing times). */
171static void record_gp_stall_check_time(void)
172{
173 unsigned long j = jiffies;
174 unsigned long j1;
175
176 WRITE_ONCE(rcu_state.gp_start, j);
177 j1 = rcu_jiffies_till_stall_check();
178 smp_mb(); // ->gp_start before ->jiffies_stall and caller's ->gp_seq.
179 WRITE_ONCE(rcu_state.nr_fqs_jiffies_stall, 0);
180 WRITE_ONCE(rcu_state.jiffies_stall, j + j1);
181 rcu_state.jiffies_resched = j + j1 / 2;
182 rcu_state.n_force_qs_gpstart = READ_ONCE(rcu_state.n_force_qs);
183}
184
185/* Zero ->ticks_this_gp and snapshot the number of RCU softirq handlers. */
186static void zero_cpu_stall_ticks(struct rcu_data *rdp)
187{
188 rdp->ticks_this_gp = 0;
189 rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
190 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
191}
192
193/*
194 * If too much time has passed in the current grace period, and if
195 * so configured, go kick the relevant kthreads.
196 */
197static void rcu_stall_kick_kthreads(void)
198{
199 unsigned long j;
200
201 if (!READ_ONCE(rcu_kick_kthreads))
202 return;
203 j = READ_ONCE(rcu_state.jiffies_kick_kthreads);
204 if (time_after(jiffies, j) && rcu_state.gp_kthread &&
205 (rcu_gp_in_progress() || READ_ONCE(rcu_state.gp_flags))) {
206 WARN_ONCE(1, "Kicking %s grace-period kthread\n",
207 rcu_state.name);
208 rcu_ftrace_dump(DUMP_ALL);
209 wake_up_process(rcu_state.gp_kthread);
210 WRITE_ONCE(rcu_state.jiffies_kick_kthreads, j + HZ);
211 }
212}
213
214/*
215 * Handler for the irq_work request posted about halfway into the RCU CPU
216 * stall timeout, and used to detect excessive irq disabling. Set state
217 * appropriately, but just complain if there is unexpected state on entry.
218 */
219static void rcu_iw_handler(struct irq_work *iwp)
220{
221 struct rcu_data *rdp;
222 struct rcu_node *rnp;
223
224 rdp = container_of(iwp, struct rcu_data, rcu_iw);
225 rnp = rdp->mynode;
226 raw_spin_lock_rcu_node(rnp);
227 if (!WARN_ON_ONCE(!rdp->rcu_iw_pending)) {
228 rdp->rcu_iw_gp_seq = rnp->gp_seq;
229 rdp->rcu_iw_pending = false;
230 }
231 raw_spin_unlock_rcu_node(rnp);
232}
233
234//////////////////////////////////////////////////////////////////////////////
235//
236// Printing RCU CPU stall warnings
237
238#ifdef CONFIG_PREEMPT_RCU
239
240/*
241 * Dump detailed information for all tasks blocking the current RCU
242 * grace period on the specified rcu_node structure.
243 */
244static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
245{
246 unsigned long flags;
247 struct task_struct *t;
248
249 raw_spin_lock_irqsave_rcu_node(rnp, flags);
250 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
251 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
252 return;
253 }
254 t = list_entry(rnp->gp_tasks->prev,
255 struct task_struct, rcu_node_entry);
256 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
257 /*
258 * We could be printing a lot while holding a spinlock.
259 * Avoid triggering hard lockup.
260 */
261 touch_nmi_watchdog();
262 sched_show_task(t);
263 }
264 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
265}
266
267// Communicate task state back to the RCU CPU stall warning request.
268struct rcu_stall_chk_rdr {
269 int nesting;
270 union rcu_special rs;
271 bool on_blkd_list;
272};
273
274/*
275 * Report out the state of a not-running task that is stalling the
276 * current RCU grace period.
277 */
278static int check_slow_task(struct task_struct *t, void *arg)
279{
280 struct rcu_stall_chk_rdr *rscrp = arg;
281
282 if (task_curr(t))
283 return -EBUSY; // It is running, so decline to inspect it.
284 rscrp->nesting = t->rcu_read_lock_nesting;
285 rscrp->rs = t->rcu_read_unlock_special;
286 rscrp->on_blkd_list = !list_empty(&t->rcu_node_entry);
287 return 0;
288}
289
290/*
291 * Scan the current list of tasks blocked within RCU read-side critical
292 * sections, printing out the tid of each of the first few of them.
293 */
294static int rcu_print_task_stall(struct rcu_node *rnp, unsigned long flags)
295 __releases(rnp->lock)
296{
297 int i = 0;
298 int ndetected = 0;
299 struct rcu_stall_chk_rdr rscr;
300 struct task_struct *t;
301 struct task_struct *ts[8];
302
303 lockdep_assert_irqs_disabled();
304 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
305 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
306 return 0;
307 }
308 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
309 rnp->level, rnp->grplo, rnp->grphi);
310 t = list_entry(rnp->gp_tasks->prev,
311 struct task_struct, rcu_node_entry);
312 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
313 get_task_struct(t);
314 ts[i++] = t;
315 if (i >= ARRAY_SIZE(ts))
316 break;
317 }
318 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
319 while (i) {
320 t = ts[--i];
321 if (task_call_func(t, check_slow_task, &rscr))
322 pr_cont(" P%d", t->pid);
323 else
324 pr_cont(" P%d/%d:%c%c%c%c",
325 t->pid, rscr.nesting,
326 ".b"[rscr.rs.b.blocked],
327 ".q"[rscr.rs.b.need_qs],
328 ".e"[rscr.rs.b.exp_hint],
329 ".l"[rscr.on_blkd_list]);
330 lockdep_assert_irqs_disabled();
331 put_task_struct(t);
332 ndetected++;
333 }
334 pr_cont("\n");
335 return ndetected;
336}
337
338#else /* #ifdef CONFIG_PREEMPT_RCU */
339
340/*
341 * Because preemptible RCU does not exist, we never have to check for
342 * tasks blocked within RCU read-side critical sections.
343 */
344static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
345{
346}
347
348/*
349 * Because preemptible RCU does not exist, we never have to check for
350 * tasks blocked within RCU read-side critical sections.
351 */
352static int rcu_print_task_stall(struct rcu_node *rnp, unsigned long flags)
353 __releases(rnp->lock)
354{
355 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
356 return 0;
357}
358#endif /* #else #ifdef CONFIG_PREEMPT_RCU */
359
360/*
361 * Dump stacks of all tasks running on stalled CPUs. First try using
362 * NMIs, but fall back to manual remote stack tracing on architectures
363 * that don't support NMI-based stack dumps. The NMI-triggered stack
364 * traces are more accurate because they are printed by the target CPU.
365 */
366static void rcu_dump_cpu_stacks(void)
367{
368 int cpu;
369 unsigned long flags;
370 struct rcu_node *rnp;
371
372 rcu_for_each_leaf_node(rnp) {
373 raw_spin_lock_irqsave_rcu_node(rnp, flags);
374 for_each_leaf_node_possible_cpu(rnp, cpu)
375 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
376 if (cpu_is_offline(cpu))
377 pr_err("Offline CPU %d blocking current GP.\n", cpu);
378 else
379 dump_cpu_task(cpu);
380 }
381 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
382 }
383}
384
385static const char * const gp_state_names[] = {
386 [RCU_GP_IDLE] = "RCU_GP_IDLE",
387 [RCU_GP_WAIT_GPS] = "RCU_GP_WAIT_GPS",
388 [RCU_GP_DONE_GPS] = "RCU_GP_DONE_GPS",
389 [RCU_GP_ONOFF] = "RCU_GP_ONOFF",
390 [RCU_GP_INIT] = "RCU_GP_INIT",
391 [RCU_GP_WAIT_FQS] = "RCU_GP_WAIT_FQS",
392 [RCU_GP_DOING_FQS] = "RCU_GP_DOING_FQS",
393 [RCU_GP_CLEANUP] = "RCU_GP_CLEANUP",
394 [RCU_GP_CLEANED] = "RCU_GP_CLEANED",
395};
396
397/*
398 * Convert a ->gp_state value to a character string.
399 */
400static const char *gp_state_getname(short gs)
401{
402 if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
403 return "???";
404 return gp_state_names[gs];
405}
406
407/* Is the RCU grace-period kthread being starved of CPU time? */
408static bool rcu_is_gp_kthread_starving(unsigned long *jp)
409{
410 unsigned long j = jiffies - READ_ONCE(rcu_state.gp_activity);
411
412 if (jp)
413 *jp = j;
414 return j > 2 * HZ;
415}
416
417static bool rcu_is_rcuc_kthread_starving(struct rcu_data *rdp, unsigned long *jp)
418{
419 int cpu;
420 struct task_struct *rcuc;
421 unsigned long j;
422
423 rcuc = rdp->rcu_cpu_kthread_task;
424 if (!rcuc)
425 return false;
426
427 cpu = task_cpu(rcuc);
428 if (cpu_is_offline(cpu) || idle_cpu(cpu))
429 return false;
430
431 j = jiffies - READ_ONCE(rdp->rcuc_activity);
432
433 if (jp)
434 *jp = j;
435 return j > 2 * HZ;
436}
437
438static void print_cpu_stat_info(int cpu)
439{
440 struct rcu_snap_record rsr, *rsrp;
441 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
442 struct kernel_cpustat *kcsp = &kcpustat_cpu(cpu);
443
444 if (!rcu_cpu_stall_cputime)
445 return;
446
447 rsrp = &rdp->snap_record;
448 if (rsrp->gp_seq != rdp->gp_seq)
449 return;
450
451 rsr.cputime_irq = kcpustat_field(kcsp, CPUTIME_IRQ, cpu);
452 rsr.cputime_softirq = kcpustat_field(kcsp, CPUTIME_SOFTIRQ, cpu);
453 rsr.cputime_system = kcpustat_field(kcsp, CPUTIME_SYSTEM, cpu);
454
455 pr_err("\t hardirqs softirqs csw/system\n");
456 pr_err("\t number: %8ld %10d %12lld\n",
457 kstat_cpu_irqs_sum(cpu) - rsrp->nr_hardirqs,
458 kstat_cpu_softirqs_sum(cpu) - rsrp->nr_softirqs,
459 nr_context_switches_cpu(cpu) - rsrp->nr_csw);
460 pr_err("\tcputime: %8lld %10lld %12lld ==> %d(ms)\n",
461 div_u64(rsr.cputime_irq - rsrp->cputime_irq, NSEC_PER_MSEC),
462 div_u64(rsr.cputime_softirq - rsrp->cputime_softirq, NSEC_PER_MSEC),
463 div_u64(rsr.cputime_system - rsrp->cputime_system, NSEC_PER_MSEC),
464 jiffies_to_msecs(jiffies - rsrp->jiffies));
465}
466
467/*
468 * Print out diagnostic information for the specified stalled CPU.
469 *
470 * If the specified CPU is aware of the current RCU grace period, then
471 * print the number of scheduling clock interrupts the CPU has taken
472 * during the time that it has been aware. Otherwise, print the number
473 * of RCU grace periods that this CPU is ignorant of, for example, "1"
474 * if the CPU was aware of the previous grace period.
475 *
476 * Also print out idle info.
477 */
478static void print_cpu_stall_info(int cpu)
479{
480 unsigned long delta;
481 bool falsepositive;
482 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
483 char *ticks_title;
484 unsigned long ticks_value;
485 bool rcuc_starved;
486 unsigned long j;
487 char buf[32];
488
489 /*
490 * We could be printing a lot while holding a spinlock. Avoid
491 * triggering hard lockup.
492 */
493 touch_nmi_watchdog();
494
495 ticks_value = rcu_seq_ctr(rcu_state.gp_seq - rdp->gp_seq);
496 if (ticks_value) {
497 ticks_title = "GPs behind";
498 } else {
499 ticks_title = "ticks this GP";
500 ticks_value = rdp->ticks_this_gp;
501 }
502 delta = rcu_seq_ctr(rdp->mynode->gp_seq - rdp->rcu_iw_gp_seq);
503 falsepositive = rcu_is_gp_kthread_starving(NULL) &&
504 rcu_dynticks_in_eqs(rcu_dynticks_snap(cpu));
505 rcuc_starved = rcu_is_rcuc_kthread_starving(rdp, &j);
506 if (rcuc_starved)
507 // Print signed value, as negative values indicate a probable bug.
508 snprintf(buf, sizeof(buf), " rcuc=%ld jiffies(starved)", j);
509 pr_err("\t%d-%c%c%c%c: (%lu %s) idle=%04x/%ld/%#lx softirq=%u/%u fqs=%ld%s%s\n",
510 cpu,
511 "O."[!!cpu_online(cpu)],
512 "o."[!!(rdp->grpmask & rdp->mynode->qsmaskinit)],
513 "N."[!!(rdp->grpmask & rdp->mynode->qsmaskinitnext)],
514 !IS_ENABLED(CONFIG_IRQ_WORK) ? '?' :
515 rdp->rcu_iw_pending ? (int)min(delta, 9UL) + '0' :
516 "!."[!delta],
517 ticks_value, ticks_title,
518 rcu_dynticks_snap(cpu) & 0xffff,
519 ct_dynticks_nesting_cpu(cpu), ct_dynticks_nmi_nesting_cpu(cpu),
520 rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
521 data_race(rcu_state.n_force_qs) - rcu_state.n_force_qs_gpstart,
522 rcuc_starved ? buf : "",
523 falsepositive ? " (false positive?)" : "");
524
525 print_cpu_stat_info(cpu);
526}
527
528/* Complain about starvation of grace-period kthread. */
529static void rcu_check_gp_kthread_starvation(void)
530{
531 int cpu;
532 struct task_struct *gpk = rcu_state.gp_kthread;
533 unsigned long j;
534
535 if (rcu_is_gp_kthread_starving(&j)) {
536 cpu = gpk ? task_cpu(gpk) : -1;
537 pr_err("%s kthread starved for %ld jiffies! g%ld f%#x %s(%d) ->state=%#x ->cpu=%d\n",
538 rcu_state.name, j,
539 (long)rcu_seq_current(&rcu_state.gp_seq),
540 data_race(READ_ONCE(rcu_state.gp_flags)),
541 gp_state_getname(rcu_state.gp_state),
542 data_race(READ_ONCE(rcu_state.gp_state)),
543 gpk ? data_race(READ_ONCE(gpk->__state)) : ~0, cpu);
544 if (gpk) {
545 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
546
547 pr_err("\tUnless %s kthread gets sufficient CPU time, OOM is now expected behavior.\n", rcu_state.name);
548 pr_err("RCU grace-period kthread stack dump:\n");
549 sched_show_task(gpk);
550 if (cpu_is_offline(cpu)) {
551 pr_err("RCU GP kthread last ran on offline CPU %d.\n", cpu);
552 } else if (!(data_race(READ_ONCE(rdp->mynode->qsmask)) & rdp->grpmask)) {
553 pr_err("Stack dump where RCU GP kthread last ran:\n");
554 dump_cpu_task(cpu);
555 }
556 wake_up_process(gpk);
557 }
558 }
559}
560
561/* Complain about missing wakeups from expired fqs wait timer */
562static void rcu_check_gp_kthread_expired_fqs_timer(void)
563{
564 struct task_struct *gpk = rcu_state.gp_kthread;
565 short gp_state;
566 unsigned long jiffies_fqs;
567 int cpu;
568
569 /*
570 * Order reads of .gp_state and .jiffies_force_qs.
571 * Matching smp_wmb() is present in rcu_gp_fqs_loop().
572 */
573 gp_state = smp_load_acquire(&rcu_state.gp_state);
574 jiffies_fqs = READ_ONCE(rcu_state.jiffies_force_qs);
575
576 if (gp_state == RCU_GP_WAIT_FQS &&
577 time_after(jiffies, jiffies_fqs + RCU_STALL_MIGHT_MIN) &&
578 gpk && !READ_ONCE(gpk->on_rq)) {
579 cpu = task_cpu(gpk);
580 pr_err("%s kthread timer wakeup didn't happen for %ld jiffies! g%ld f%#x %s(%d) ->state=%#x\n",
581 rcu_state.name, (jiffies - jiffies_fqs),
582 (long)rcu_seq_current(&rcu_state.gp_seq),
583 data_race(rcu_state.gp_flags),
584 gp_state_getname(RCU_GP_WAIT_FQS), RCU_GP_WAIT_FQS,
585 data_race(READ_ONCE(gpk->__state)));
586 pr_err("\tPossible timer handling issue on cpu=%d timer-softirq=%u\n",
587 cpu, kstat_softirqs_cpu(TIMER_SOFTIRQ, cpu));
588 }
589}
590
591static void print_other_cpu_stall(unsigned long gp_seq, unsigned long gps)
592{
593 int cpu;
594 unsigned long flags;
595 unsigned long gpa;
596 unsigned long j;
597 int ndetected = 0;
598 struct rcu_node *rnp;
599 long totqlen = 0;
600
601 lockdep_assert_irqs_disabled();
602
603 /* Kick and suppress, if so configured. */
604 rcu_stall_kick_kthreads();
605 if (rcu_stall_is_suppressed())
606 return;
607
608 /*
609 * OK, time to rat on our buddy...
610 * See Documentation/RCU/stallwarn.rst for info on how to debug
611 * RCU CPU stall warnings.
612 */
613 trace_rcu_stall_warning(rcu_state.name, TPS("StallDetected"));
614 pr_err("INFO: %s detected stalls on CPUs/tasks:\n", rcu_state.name);
615 rcu_for_each_leaf_node(rnp) {
616 raw_spin_lock_irqsave_rcu_node(rnp, flags);
617 if (rnp->qsmask != 0) {
618 for_each_leaf_node_possible_cpu(rnp, cpu)
619 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
620 print_cpu_stall_info(cpu);
621 ndetected++;
622 }
623 }
624 ndetected += rcu_print_task_stall(rnp, flags); // Releases rnp->lock.
625 lockdep_assert_irqs_disabled();
626 }
627
628 for_each_possible_cpu(cpu)
629 totqlen += rcu_get_n_cbs_cpu(cpu);
630 pr_err("\t(detected by %d, t=%ld jiffies, g=%ld, q=%lu ncpus=%d)\n",
631 smp_processor_id(), (long)(jiffies - gps),
632 (long)rcu_seq_current(&rcu_state.gp_seq), totqlen, rcu_state.n_online_cpus);
633 if (ndetected) {
634 rcu_dump_cpu_stacks();
635
636 /* Complain about tasks blocking the grace period. */
637 rcu_for_each_leaf_node(rnp)
638 rcu_print_detail_task_stall_rnp(rnp);
639 } else {
640 if (rcu_seq_current(&rcu_state.gp_seq) != gp_seq) {
641 pr_err("INFO: Stall ended before state dump start\n");
642 } else {
643 j = jiffies;
644 gpa = data_race(READ_ONCE(rcu_state.gp_activity));
645 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
646 rcu_state.name, j - gpa, j, gpa,
647 data_race(READ_ONCE(jiffies_till_next_fqs)),
648 data_race(READ_ONCE(rcu_get_root()->qsmask)));
649 }
650 }
651 /* Rewrite if needed in case of slow consoles. */
652 if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
653 WRITE_ONCE(rcu_state.jiffies_stall,
654 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
655
656 rcu_check_gp_kthread_expired_fqs_timer();
657 rcu_check_gp_kthread_starvation();
658
659 panic_on_rcu_stall();
660
661 rcu_force_quiescent_state(); /* Kick them all. */
662}
663
664static void print_cpu_stall(unsigned long gps)
665{
666 int cpu;
667 unsigned long flags;
668 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
669 struct rcu_node *rnp = rcu_get_root();
670 long totqlen = 0;
671
672 lockdep_assert_irqs_disabled();
673
674 /* Kick and suppress, if so configured. */
675 rcu_stall_kick_kthreads();
676 if (rcu_stall_is_suppressed())
677 return;
678
679 /*
680 * OK, time to rat on ourselves...
681 * See Documentation/RCU/stallwarn.rst for info on how to debug
682 * RCU CPU stall warnings.
683 */
684 trace_rcu_stall_warning(rcu_state.name, TPS("SelfDetected"));
685 pr_err("INFO: %s self-detected stall on CPU\n", rcu_state.name);
686 raw_spin_lock_irqsave_rcu_node(rdp->mynode, flags);
687 print_cpu_stall_info(smp_processor_id());
688 raw_spin_unlock_irqrestore_rcu_node(rdp->mynode, flags);
689 for_each_possible_cpu(cpu)
690 totqlen += rcu_get_n_cbs_cpu(cpu);
691 pr_err("\t(t=%lu jiffies g=%ld q=%lu ncpus=%d)\n",
692 jiffies - gps,
693 (long)rcu_seq_current(&rcu_state.gp_seq), totqlen, rcu_state.n_online_cpus);
694
695 rcu_check_gp_kthread_expired_fqs_timer();
696 rcu_check_gp_kthread_starvation();
697
698 rcu_dump_cpu_stacks();
699
700 raw_spin_lock_irqsave_rcu_node(rnp, flags);
701 /* Rewrite if needed in case of slow consoles. */
702 if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
703 WRITE_ONCE(rcu_state.jiffies_stall,
704 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
705 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
706
707 panic_on_rcu_stall();
708
709 /*
710 * Attempt to revive the RCU machinery by forcing a context switch.
711 *
712 * A context switch would normally allow the RCU state machine to make
713 * progress and it could be we're stuck in kernel space without context
714 * switches for an entirely unreasonable amount of time.
715 */
716 set_tsk_need_resched(current);
717 set_preempt_need_resched();
718}
719
720static void check_cpu_stall(struct rcu_data *rdp)
721{
722 bool self_detected;
723 unsigned long gs1;
724 unsigned long gs2;
725 unsigned long gps;
726 unsigned long j;
727 unsigned long jn;
728 unsigned long js;
729 struct rcu_node *rnp;
730
731 lockdep_assert_irqs_disabled();
732 if ((rcu_stall_is_suppressed() && !READ_ONCE(rcu_kick_kthreads)) ||
733 !rcu_gp_in_progress())
734 return;
735 rcu_stall_kick_kthreads();
736
737 /*
738 * Check if it was requested (via rcu_cpu_stall_reset()) that the FQS
739 * loop has to set jiffies to ensure a non-stale jiffies value. This
740 * is required to have good jiffies value after coming out of long
741 * breaks of jiffies updates. Not doing so can cause false positives.
742 */
743 if (READ_ONCE(rcu_state.nr_fqs_jiffies_stall) > 0)
744 return;
745
746 j = jiffies;
747
748 /*
749 * Lots of memory barriers to reject false positives.
750 *
751 * The idea is to pick up rcu_state.gp_seq, then
752 * rcu_state.jiffies_stall, then rcu_state.gp_start, and finally
753 * another copy of rcu_state.gp_seq. These values are updated in
754 * the opposite order with memory barriers (or equivalent) during
755 * grace-period initialization and cleanup. Now, a false positive
756 * can occur if we get an new value of rcu_state.gp_start and a old
757 * value of rcu_state.jiffies_stall. But given the memory barriers,
758 * the only way that this can happen is if one grace period ends
759 * and another starts between these two fetches. This is detected
760 * by comparing the second fetch of rcu_state.gp_seq with the
761 * previous fetch from rcu_state.gp_seq.
762 *
763 * Given this check, comparisons of jiffies, rcu_state.jiffies_stall,
764 * and rcu_state.gp_start suffice to forestall false positives.
765 */
766 gs1 = READ_ONCE(rcu_state.gp_seq);
767 smp_rmb(); /* Pick up ->gp_seq first... */
768 js = READ_ONCE(rcu_state.jiffies_stall);
769 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
770 gps = READ_ONCE(rcu_state.gp_start);
771 smp_rmb(); /* ...and finally ->gp_start before ->gp_seq again. */
772 gs2 = READ_ONCE(rcu_state.gp_seq);
773 if (gs1 != gs2 ||
774 ULONG_CMP_LT(j, js) ||
775 ULONG_CMP_GE(gps, js))
776 return; /* No stall or GP completed since entering function. */
777 rnp = rdp->mynode;
778 jn = jiffies + ULONG_MAX / 2;
779 self_detected = READ_ONCE(rnp->qsmask) & rdp->grpmask;
780 if (rcu_gp_in_progress() &&
781 (self_detected || ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) &&
782 cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
783 /*
784 * If a virtual machine is stopped by the host it can look to
785 * the watchdog like an RCU stall. Check to see if the host
786 * stopped the vm.
787 */
788 if (kvm_check_and_clear_guest_paused())
789 return;
790
791 rcu_stall_notifier_call_chain(RCU_STALL_NOTIFY_NORM, (void *)j - gps);
792 if (self_detected) {
793 /* We haven't checked in, so go dump stack. */
794 print_cpu_stall(gps);
795 } else {
796 /* They had a few time units to dump stack, so complain. */
797 print_other_cpu_stall(gs2, gps);
798 }
799
800 if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
801 rcu_ftrace_dump(DUMP_ALL);
802
803 if (READ_ONCE(rcu_state.jiffies_stall) == jn) {
804 jn = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
805 WRITE_ONCE(rcu_state.jiffies_stall, jn);
806 }
807 }
808}
809
810//////////////////////////////////////////////////////////////////////////////
811//
812// RCU forward-progress mechanisms, including for callback invocation.
813
814
815/*
816 * Check to see if a failure to end RCU priority inversion was due to
817 * a CPU not passing through a quiescent state. When this happens, there
818 * is nothing that RCU priority boosting can do to help, so we shouldn't
819 * count this as an RCU priority boosting failure. A return of true says
820 * RCU priority boosting is to blame, and false says otherwise. If false
821 * is returned, the first of the CPUs to blame is stored through cpup.
822 * If there was no CPU blocking the current grace period, but also nothing
823 * in need of being boosted, *cpup is set to -1. This can happen in case
824 * of vCPU preemption while the last CPU is reporting its quiscent state,
825 * for example.
826 *
827 * If cpup is NULL, then a lockless quick check is carried out, suitable
828 * for high-rate usage. On the other hand, if cpup is non-NULL, each
829 * rcu_node structure's ->lock is acquired, ruling out high-rate usage.
830 */
831bool rcu_check_boost_fail(unsigned long gp_state, int *cpup)
832{
833 bool atb = false;
834 int cpu;
835 unsigned long flags;
836 struct rcu_node *rnp;
837
838 rcu_for_each_leaf_node(rnp) {
839 if (!cpup) {
840 if (data_race(READ_ONCE(rnp->qsmask))) {
841 return false;
842 } else {
843 if (READ_ONCE(rnp->gp_tasks))
844 atb = true;
845 continue;
846 }
847 }
848 *cpup = -1;
849 raw_spin_lock_irqsave_rcu_node(rnp, flags);
850 if (rnp->gp_tasks)
851 atb = true;
852 if (!rnp->qsmask) {
853 // No CPUs without quiescent states for this rnp.
854 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
855 continue;
856 }
857 // Find the first holdout CPU.
858 for_each_leaf_node_possible_cpu(rnp, cpu) {
859 if (rnp->qsmask & (1UL << (cpu - rnp->grplo))) {
860 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
861 *cpup = cpu;
862 return false;
863 }
864 }
865 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
866 }
867 // Can't blame CPUs, so must blame RCU priority boosting.
868 return atb;
869}
870EXPORT_SYMBOL_GPL(rcu_check_boost_fail);
871
872/*
873 * Show the state of the grace-period kthreads.
874 */
875void show_rcu_gp_kthreads(void)
876{
877 unsigned long cbs = 0;
878 int cpu;
879 unsigned long j;
880 unsigned long ja;
881 unsigned long jr;
882 unsigned long js;
883 unsigned long jw;
884 struct rcu_data *rdp;
885 struct rcu_node *rnp;
886 struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
887
888 j = jiffies;
889 ja = j - data_race(READ_ONCE(rcu_state.gp_activity));
890 jr = j - data_race(READ_ONCE(rcu_state.gp_req_activity));
891 js = j - data_race(READ_ONCE(rcu_state.gp_start));
892 jw = j - data_race(READ_ONCE(rcu_state.gp_wake_time));
893 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",
894 rcu_state.name, gp_state_getname(rcu_state.gp_state),
895 data_race(READ_ONCE(rcu_state.gp_state)),
896 t ? data_race(READ_ONCE(t->__state)) : 0x1ffff, t ? t->rt_priority : 0xffU,
897 js, ja, jr, jw, (long)data_race(READ_ONCE(rcu_state.gp_wake_seq)),
898 (long)data_race(READ_ONCE(rcu_state.gp_seq)),
899 (long)data_race(READ_ONCE(rcu_get_root()->gp_seq_needed)),
900 data_race(READ_ONCE(rcu_state.gp_max)),
901 data_race(READ_ONCE(rcu_state.gp_flags)));
902 rcu_for_each_node_breadth_first(rnp) {
903 if (ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq), READ_ONCE(rnp->gp_seq_needed)) &&
904 !data_race(READ_ONCE(rnp->qsmask)) && !data_race(READ_ONCE(rnp->boost_tasks)) &&
905 !data_race(READ_ONCE(rnp->exp_tasks)) && !data_race(READ_ONCE(rnp->gp_tasks)))
906 continue;
907 pr_info("\trcu_node %d:%d ->gp_seq %ld ->gp_seq_needed %ld ->qsmask %#lx %c%c%c%c ->n_boosts %ld\n",
908 rnp->grplo, rnp->grphi,
909 (long)data_race(READ_ONCE(rnp->gp_seq)),
910 (long)data_race(READ_ONCE(rnp->gp_seq_needed)),
911 data_race(READ_ONCE(rnp->qsmask)),
912 ".b"[!!data_race(READ_ONCE(rnp->boost_kthread_task))],
913 ".B"[!!data_race(READ_ONCE(rnp->boost_tasks))],
914 ".E"[!!data_race(READ_ONCE(rnp->exp_tasks))],
915 ".G"[!!data_race(READ_ONCE(rnp->gp_tasks))],
916 data_race(READ_ONCE(rnp->n_boosts)));
917 if (!rcu_is_leaf_node(rnp))
918 continue;
919 for_each_leaf_node_possible_cpu(rnp, cpu) {
920 rdp = per_cpu_ptr(&rcu_data, cpu);
921 if (READ_ONCE(rdp->gpwrap) ||
922 ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq),
923 READ_ONCE(rdp->gp_seq_needed)))
924 continue;
925 pr_info("\tcpu %d ->gp_seq_needed %ld\n",
926 cpu, (long)data_race(READ_ONCE(rdp->gp_seq_needed)));
927 }
928 }
929 for_each_possible_cpu(cpu) {
930 rdp = per_cpu_ptr(&rcu_data, cpu);
931 cbs += data_race(READ_ONCE(rdp->n_cbs_invoked));
932 if (rcu_segcblist_is_offloaded(&rdp->cblist))
933 show_rcu_nocb_state(rdp);
934 }
935 pr_info("RCU callbacks invoked since boot: %lu\n", cbs);
936 show_rcu_tasks_gp_kthreads();
937}
938EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
939
940/*
941 * This function checks for grace-period requests that fail to motivate
942 * RCU to come out of its idle mode.
943 */
944static void rcu_check_gp_start_stall(struct rcu_node *rnp, struct rcu_data *rdp,
945 const unsigned long gpssdelay)
946{
947 unsigned long flags;
948 unsigned long j;
949 struct rcu_node *rnp_root = rcu_get_root();
950 static atomic_t warned = ATOMIC_INIT(0);
951
952 if (!IS_ENABLED(CONFIG_PROVE_RCU) || rcu_gp_in_progress() ||
953 ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
954 READ_ONCE(rnp_root->gp_seq_needed)) ||
955 !smp_load_acquire(&rcu_state.gp_kthread)) // Get stable kthread.
956 return;
957 j = jiffies; /* Expensive access, and in common case don't get here. */
958 if (time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
959 time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
960 atomic_read(&warned))
961 return;
962
963 raw_spin_lock_irqsave_rcu_node(rnp, flags);
964 j = jiffies;
965 if (rcu_gp_in_progress() ||
966 ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
967 READ_ONCE(rnp_root->gp_seq_needed)) ||
968 time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
969 time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
970 atomic_read(&warned)) {
971 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
972 return;
973 }
974 /* Hold onto the leaf lock to make others see warned==1. */
975
976 if (rnp_root != rnp)
977 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
978 j = jiffies;
979 if (rcu_gp_in_progress() ||
980 ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
981 READ_ONCE(rnp_root->gp_seq_needed)) ||
982 time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
983 time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
984 atomic_xchg(&warned, 1)) {
985 if (rnp_root != rnp)
986 /* irqs remain disabled. */
987 raw_spin_unlock_rcu_node(rnp_root);
988 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
989 return;
990 }
991 WARN_ON(1);
992 if (rnp_root != rnp)
993 raw_spin_unlock_rcu_node(rnp_root);
994 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
995 show_rcu_gp_kthreads();
996}
997
998/*
999 * Do a forward-progress check for rcutorture. This is normally invoked
1000 * due to an OOM event. The argument "j" gives the time period during
1001 * which rcutorture would like progress to have been made.
1002 */
1003void rcu_fwd_progress_check(unsigned long j)
1004{
1005 unsigned long cbs;
1006 int cpu;
1007 unsigned long max_cbs = 0;
1008 int max_cpu = -1;
1009 struct rcu_data *rdp;
1010
1011 if (rcu_gp_in_progress()) {
1012 pr_info("%s: GP age %lu jiffies\n",
1013 __func__, jiffies - data_race(READ_ONCE(rcu_state.gp_start)));
1014 show_rcu_gp_kthreads();
1015 } else {
1016 pr_info("%s: Last GP end %lu jiffies ago\n",
1017 __func__, jiffies - data_race(READ_ONCE(rcu_state.gp_end)));
1018 preempt_disable();
1019 rdp = this_cpu_ptr(&rcu_data);
1020 rcu_check_gp_start_stall(rdp->mynode, rdp, j);
1021 preempt_enable();
1022 }
1023 for_each_possible_cpu(cpu) {
1024 cbs = rcu_get_n_cbs_cpu(cpu);
1025 if (!cbs)
1026 continue;
1027 if (max_cpu < 0)
1028 pr_info("%s: callbacks", __func__);
1029 pr_cont(" %d: %lu", cpu, cbs);
1030 if (cbs <= max_cbs)
1031 continue;
1032 max_cbs = cbs;
1033 max_cpu = cpu;
1034 }
1035 if (max_cpu >= 0)
1036 pr_cont("\n");
1037}
1038EXPORT_SYMBOL_GPL(rcu_fwd_progress_check);
1039
1040/* Commandeer a sysrq key to dump RCU's tree. */
1041static bool sysrq_rcu;
1042module_param(sysrq_rcu, bool, 0444);
1043
1044/* Dump grace-period-request information due to commandeered sysrq. */
1045static void sysrq_show_rcu(u8 key)
1046{
1047 show_rcu_gp_kthreads();
1048}
1049
1050static const struct sysrq_key_op sysrq_rcudump_op = {
1051 .handler = sysrq_show_rcu,
1052 .help_msg = "show-rcu(y)",
1053 .action_msg = "Show RCU tree",
1054 .enable_mask = SYSRQ_ENABLE_DUMP,
1055};
1056
1057static int __init rcu_sysrq_init(void)
1058{
1059 if (sysrq_rcu)
1060 return register_sysrq_key('y', &sysrq_rcudump_op);
1061 return 0;
1062}
1063early_initcall(rcu_sysrq_init);
1064
1065#ifdef CONFIG_RCU_CPU_STALL_NOTIFIER
1066
1067//////////////////////////////////////////////////////////////////////////////
1068//
1069// RCU CPU stall-warning notifiers
1070
1071static ATOMIC_NOTIFIER_HEAD(rcu_cpu_stall_notifier_list);
1072
1073/**
1074 * rcu_stall_chain_notifier_register - Add an RCU CPU stall notifier
1075 * @n: Entry to add.
1076 *
1077 * Adds an RCU CPU stall notifier to an atomic notifier chain.
1078 * The @action passed to a notifier will be @RCU_STALL_NOTIFY_NORM or
1079 * friends. The @data will be the duration of the stalled grace period,
1080 * in jiffies, coerced to a void* pointer.
1081 *
1082 * Returns 0 on success, %-EEXIST on error.
1083 */
1084int rcu_stall_chain_notifier_register(struct notifier_block *n)
1085{
1086 int rcsn = rcu_cpu_stall_notifiers;
1087
1088 WARN(1, "Adding %pS() to RCU stall notifier list (%s).\n", n->notifier_call,
1089 rcsn ? "possibly suppressing RCU CPU stall warnings" : "failed, so all is well");
1090 if (rcsn)
1091 return atomic_notifier_chain_register(&rcu_cpu_stall_notifier_list, n);
1092 return -EEXIST;
1093}
1094EXPORT_SYMBOL_GPL(rcu_stall_chain_notifier_register);
1095
1096/**
1097 * rcu_stall_chain_notifier_unregister - Remove an RCU CPU stall notifier
1098 * @n: Entry to add.
1099 *
1100 * Removes an RCU CPU stall notifier from an atomic notifier chain.
1101 *
1102 * Returns zero on success, %-ENOENT on failure.
1103 */
1104int rcu_stall_chain_notifier_unregister(struct notifier_block *n)
1105{
1106 return atomic_notifier_chain_unregister(&rcu_cpu_stall_notifier_list, n);
1107}
1108EXPORT_SYMBOL_GPL(rcu_stall_chain_notifier_unregister);
1109
1110/*
1111 * rcu_stall_notifier_call_chain - Call functions in an RCU CPU stall notifier chain
1112 * @val: Value passed unmodified to notifier function
1113 * @v: Pointer passed unmodified to notifier function
1114 *
1115 * Calls each function in the RCU CPU stall notifier chain in turn, which
1116 * is an atomic call chain. See atomic_notifier_call_chain() for more
1117 * information.
1118 *
1119 * This is for use within RCU, hence the omission of the extra asterisk
1120 * to indicate a non-kerneldoc format header comment.
1121 */
1122int rcu_stall_notifier_call_chain(unsigned long val, void *v)
1123{
1124 return atomic_notifier_call_chain(&rcu_cpu_stall_notifier_list, val, v);
1125}
1126
1127#endif // #ifdef CONFIG_RCU_CPU_STALL_NOTIFIER
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);