1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3 * RCU CPU stall warnings for normal RCU grace periods
4 *
5 * Copyright IBM Corporation, 2019
6 *
7 * Author: Paul E. McKenney <paulmck@linux.ibm.com>
8 */
9
10 #include <linux/kvm_para.h>
11
12 //////////////////////////////////////////////////////////////////////////////
13 //
14 // Controlling CPU stall warnings, including delay calculation.
15
16 /* panic() on RCU Stall sysctl. */
17 int sysctl_panic_on_rcu_stall __read_mostly;
18 int sysctl_max_rcu_stall_to_panic __read_mostly;
19
20 #ifdef CONFIG_PROVE_RCU
21 #define RCU_STALL_DELAY_DELTA (5 * HZ)
22 #else
23 #define RCU_STALL_DELAY_DELTA 0
24 #endif
25 #define RCU_STALL_MIGHT_DIV 8
26 #define RCU_STALL_MIGHT_MIN (2 * HZ)
27
rcu_exp_jiffies_till_stall_check(void)28 int rcu_exp_jiffies_till_stall_check(void)
29 {
30 int cpu_stall_timeout = READ_ONCE(rcu_exp_cpu_stall_timeout);
31 int exp_stall_delay_delta = 0;
32 int till_stall_check;
33
34 // Zero says to use rcu_cpu_stall_timeout, but in milliseconds.
35 if (!cpu_stall_timeout)
36 cpu_stall_timeout = jiffies_to_msecs(rcu_jiffies_till_stall_check());
37
38 // Limit check must be consistent with the Kconfig limits for
39 // CONFIG_RCU_EXP_CPU_STALL_TIMEOUT, so check the allowed range.
40 // The minimum clamped value is "2UL", because at least one full
41 // tick has to be guaranteed.
42 till_stall_check = clamp(msecs_to_jiffies(cpu_stall_timeout), 2UL, 300UL * HZ);
43
44 if (cpu_stall_timeout && jiffies_to_msecs(till_stall_check) != cpu_stall_timeout)
45 WRITE_ONCE(rcu_exp_cpu_stall_timeout, jiffies_to_msecs(till_stall_check));
46
47 #ifdef CONFIG_PROVE_RCU
48 /* Add extra ~25% out of till_stall_check. */
49 exp_stall_delay_delta = ((till_stall_check * 25) / 100) + 1;
50 #endif
51
52 return till_stall_check + exp_stall_delay_delta;
53 }
54 EXPORT_SYMBOL_GPL(rcu_exp_jiffies_till_stall_check);
55
56 /* Limit-check stall timeouts specified at boottime and runtime. */
rcu_jiffies_till_stall_check(void)57 int rcu_jiffies_till_stall_check(void)
58 {
59 int till_stall_check = READ_ONCE(rcu_cpu_stall_timeout);
60
61 /*
62 * Limit check must be consistent with the Kconfig limits
63 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
64 */
65 if (till_stall_check < 3) {
66 WRITE_ONCE(rcu_cpu_stall_timeout, 3);
67 till_stall_check = 3;
68 } else if (till_stall_check > 300) {
69 WRITE_ONCE(rcu_cpu_stall_timeout, 300);
70 till_stall_check = 300;
71 }
72 return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
73 }
74 EXPORT_SYMBOL_GPL(rcu_jiffies_till_stall_check);
75
76 /**
77 * rcu_gp_might_be_stalled - Is it likely that the grace period is stalled?
78 *
79 * Returns @true if the current grace period is sufficiently old that
80 * it is reasonable to assume that it might be stalled. This can be
81 * useful when deciding whether to allocate memory to enable RCU-mediated
82 * freeing on the one hand or just invoking synchronize_rcu() on the other.
83 * The latter is preferable when the grace period is stalled.
84 *
85 * Note that sampling of the .gp_start and .gp_seq fields must be done
86 * carefully to avoid false positives at the beginnings and ends of
87 * grace periods.
88 */
rcu_gp_might_be_stalled(void)89 bool rcu_gp_might_be_stalled(void)
90 {
91 unsigned long d = rcu_jiffies_till_stall_check() / RCU_STALL_MIGHT_DIV;
92 unsigned long j = jiffies;
93
94 if (d < RCU_STALL_MIGHT_MIN)
95 d = RCU_STALL_MIGHT_MIN;
96 smp_mb(); // jiffies before .gp_seq to avoid false positives.
97 if (!rcu_gp_in_progress())
98 return false;
99 // Long delays at this point avoids false positive, but a delay
100 // of ULONG_MAX/4 jiffies voids your no-false-positive warranty.
101 smp_mb(); // .gp_seq before second .gp_start
102 // And ditto here.
103 return !time_before(j, READ_ONCE(rcu_state.gp_start) + d);
104 }
105
106 /* Don't do RCU CPU stall warnings during long sysrq printouts. */
rcu_sysrq_start(void)107 void rcu_sysrq_start(void)
108 {
109 if (!rcu_cpu_stall_suppress)
110 rcu_cpu_stall_suppress = 2;
111 }
112
rcu_sysrq_end(void)113 void rcu_sysrq_end(void)
114 {
115 if (rcu_cpu_stall_suppress == 2)
116 rcu_cpu_stall_suppress = 0;
117 }
118
119 /* Don't print RCU CPU stall warnings during a kernel panic. */
rcu_panic(struct notifier_block * this,unsigned long ev,void * ptr)120 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
121 {
122 rcu_cpu_stall_suppress = 1;
123 return NOTIFY_DONE;
124 }
125
126 static struct notifier_block rcu_panic_block = {
127 .notifier_call = rcu_panic,
128 };
129
check_cpu_stall_init(void)130 static int __init check_cpu_stall_init(void)
131 {
132 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
133 return 0;
134 }
135 early_initcall(check_cpu_stall_init);
136
137 /* If so specified via sysctl, panic, yielding cleaner stall-warning output. */
panic_on_rcu_stall(void)138 static void panic_on_rcu_stall(void)
139 {
140 static int cpu_stall;
141
142 if (++cpu_stall < sysctl_max_rcu_stall_to_panic)
143 return;
144
145 if (sysctl_panic_on_rcu_stall)
146 panic("RCU Stall\n");
147 }
148
149 /**
150 * rcu_cpu_stall_reset - restart stall-warning timeout for current grace period
151 *
152 * To perform the reset request from the caller, disable stall detection until
153 * 3 fqs loops have passed. This is required to ensure a fresh jiffies is
154 * loaded. It should be safe to do from the fqs loop as enough timer
155 * interrupts and context switches should have passed.
156 *
157 * The caller must disable hard irqs.
158 */
rcu_cpu_stall_reset(void)159 void rcu_cpu_stall_reset(void)
160 {
161 WRITE_ONCE(rcu_state.nr_fqs_jiffies_stall, 3);
162 WRITE_ONCE(rcu_state.jiffies_stall, ULONG_MAX);
163 }
164
165 //////////////////////////////////////////////////////////////////////////////
166 //
167 // Interaction with RCU grace periods
168
169 /* Start of new grace period, so record stall time (and forcing times). */
record_gp_stall_check_time(void)170 static void record_gp_stall_check_time(void)
171 {
172 unsigned long j = jiffies;
173 unsigned long j1;
174
175 WRITE_ONCE(rcu_state.gp_start, j);
176 j1 = rcu_jiffies_till_stall_check();
177 smp_mb(); // ->gp_start before ->jiffies_stall and caller's ->gp_seq.
178 WRITE_ONCE(rcu_state.nr_fqs_jiffies_stall, 0);
179 WRITE_ONCE(rcu_state.jiffies_stall, j + j1);
180 rcu_state.jiffies_resched = j + j1 / 2;
181 rcu_state.n_force_qs_gpstart = READ_ONCE(rcu_state.n_force_qs);
182 }
183
184 /* Zero ->ticks_this_gp and snapshot the number of RCU softirq handlers. */
zero_cpu_stall_ticks(struct rcu_data * rdp)185 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
186 {
187 rdp->ticks_this_gp = 0;
188 rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
189 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
190 }
191
192 /*
193 * If too much time has passed in the current grace period, and if
194 * so configured, go kick the relevant kthreads.
195 */
rcu_stall_kick_kthreads(void)196 static void rcu_stall_kick_kthreads(void)
197 {
198 unsigned long j;
199
200 if (!READ_ONCE(rcu_kick_kthreads))
201 return;
202 j = READ_ONCE(rcu_state.jiffies_kick_kthreads);
203 if (time_after(jiffies, j) && rcu_state.gp_kthread &&
204 (rcu_gp_in_progress() || READ_ONCE(rcu_state.gp_flags))) {
205 WARN_ONCE(1, "Kicking %s grace-period kthread\n",
206 rcu_state.name);
207 rcu_ftrace_dump(DUMP_ALL);
208 wake_up_process(rcu_state.gp_kthread);
209 WRITE_ONCE(rcu_state.jiffies_kick_kthreads, j + HZ);
210 }
211 }
212
213 /*
214 * Handler for the irq_work request posted about halfway into the RCU CPU
215 * stall timeout, and used to detect excessive irq disabling. Set state
216 * appropriately, but just complain if there is unexpected state on entry.
217 */
rcu_iw_handler(struct irq_work * iwp)218 static void rcu_iw_handler(struct irq_work *iwp)
219 {
220 struct rcu_data *rdp;
221 struct rcu_node *rnp;
222
223 rdp = container_of(iwp, struct rcu_data, rcu_iw);
224 rnp = rdp->mynode;
225 raw_spin_lock_rcu_node(rnp);
226 if (!WARN_ON_ONCE(!rdp->rcu_iw_pending)) {
227 rdp->rcu_iw_gp_seq = rnp->gp_seq;
228 rdp->rcu_iw_pending = false;
229 }
230 raw_spin_unlock_rcu_node(rnp);
231 }
232
233 //////////////////////////////////////////////////////////////////////////////
234 //
235 // Printing RCU CPU stall warnings
236
237 #ifdef CONFIG_PREEMPT_RCU
238
239 /*
240 * Dump detailed information for all tasks blocking the current RCU
241 * grace period on the specified rcu_node structure.
242 */
rcu_print_detail_task_stall_rnp(struct rcu_node * rnp)243 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
244 {
245 unsigned long flags;
246 struct task_struct *t;
247
248 raw_spin_lock_irqsave_rcu_node(rnp, flags);
249 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
250 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
251 return;
252 }
253 t = list_entry(rnp->gp_tasks->prev,
254 struct task_struct, rcu_node_entry);
255 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
256 /*
257 * We could be printing a lot while holding a spinlock.
258 * Avoid triggering hard lockup.
259 */
260 touch_nmi_watchdog();
261 sched_show_task(t);
262 }
263 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
264 }
265
266 // Communicate task state back to the RCU CPU stall warning request.
267 struct rcu_stall_chk_rdr {
268 int nesting;
269 union rcu_special rs;
270 bool on_blkd_list;
271 };
272
273 /*
274 * Report out the state of a not-running task that is stalling the
275 * current RCU grace period.
276 */
check_slow_task(struct task_struct * t,void * arg)277 static int check_slow_task(struct task_struct *t, void *arg)
278 {
279 struct rcu_stall_chk_rdr *rscrp = arg;
280
281 if (task_curr(t))
282 return -EBUSY; // It is running, so decline to inspect it.
283 rscrp->nesting = t->rcu_read_lock_nesting;
284 rscrp->rs = t->rcu_read_unlock_special;
285 rscrp->on_blkd_list = !list_empty(&t->rcu_node_entry);
286 return 0;
287 }
288
289 /*
290 * Scan the current list of tasks blocked within RCU read-side critical
291 * sections, printing out the tid of each of the first few of them.
292 */
rcu_print_task_stall(struct rcu_node * rnp,unsigned long flags)293 static int rcu_print_task_stall(struct rcu_node *rnp, unsigned long flags)
294 __releases(rnp->lock)
295 {
296 int i = 0;
297 int ndetected = 0;
298 struct rcu_stall_chk_rdr rscr;
299 struct task_struct *t;
300 struct task_struct *ts[8];
301
302 lockdep_assert_irqs_disabled();
303 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
304 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
305 return 0;
306 }
307 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
308 rnp->level, rnp->grplo, rnp->grphi);
309 t = list_entry(rnp->gp_tasks->prev,
310 struct task_struct, rcu_node_entry);
311 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
312 get_task_struct(t);
313 ts[i++] = t;
314 if (i >= ARRAY_SIZE(ts))
315 break;
316 }
317 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
318 while (i) {
319 t = ts[--i];
320 if (task_call_func(t, check_slow_task, &rscr))
321 pr_cont(" P%d", t->pid);
322 else
323 pr_cont(" P%d/%d:%c%c%c%c",
324 t->pid, rscr.nesting,
325 ".b"[rscr.rs.b.blocked],
326 ".q"[rscr.rs.b.need_qs],
327 ".e"[rscr.rs.b.exp_hint],
328 ".l"[rscr.on_blkd_list]);
329 lockdep_assert_irqs_disabled();
330 put_task_struct(t);
331 ndetected++;
332 }
333 pr_cont("\n");
334 return ndetected;
335 }
336
337 #else /* #ifdef CONFIG_PREEMPT_RCU */
338
339 /*
340 * Because preemptible RCU does not exist, we never have to check for
341 * tasks blocked within RCU read-side critical sections.
342 */
rcu_print_detail_task_stall_rnp(struct rcu_node * rnp)343 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
344 {
345 }
346
347 /*
348 * Because preemptible RCU does not exist, we never have to check for
349 * tasks blocked within RCU read-side critical sections.
350 */
rcu_print_task_stall(struct rcu_node * rnp,unsigned long flags)351 static int rcu_print_task_stall(struct rcu_node *rnp, unsigned long flags)
352 __releases(rnp->lock)
353 {
354 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
355 return 0;
356 }
357 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
358
359 /*
360 * Dump stacks of all tasks running on stalled CPUs. First try using
361 * NMIs, but fall back to manual remote stack tracing on architectures
362 * that don't support NMI-based stack dumps. The NMI-triggered stack
363 * traces are more accurate because they are printed by the target CPU.
364 */
rcu_dump_cpu_stacks(void)365 static void rcu_dump_cpu_stacks(void)
366 {
367 int cpu;
368 unsigned long flags;
369 struct rcu_node *rnp;
370
371 rcu_for_each_leaf_node(rnp) {
372 raw_spin_lock_irqsave_rcu_node(rnp, flags);
373 for_each_leaf_node_possible_cpu(rnp, cpu)
374 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
375 if (cpu_is_offline(cpu))
376 pr_err("Offline CPU %d blocking current GP.\n", cpu);
377 else
378 dump_cpu_task(cpu);
379 }
380 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
381 }
382 }
383
384 static const char * const gp_state_names[] = {
385 [RCU_GP_IDLE] = "RCU_GP_IDLE",
386 [RCU_GP_WAIT_GPS] = "RCU_GP_WAIT_GPS",
387 [RCU_GP_DONE_GPS] = "RCU_GP_DONE_GPS",
388 [RCU_GP_ONOFF] = "RCU_GP_ONOFF",
389 [RCU_GP_INIT] = "RCU_GP_INIT",
390 [RCU_GP_WAIT_FQS] = "RCU_GP_WAIT_FQS",
391 [RCU_GP_DOING_FQS] = "RCU_GP_DOING_FQS",
392 [RCU_GP_CLEANUP] = "RCU_GP_CLEANUP",
393 [RCU_GP_CLEANED] = "RCU_GP_CLEANED",
394 };
395
396 /*
397 * Convert a ->gp_state value to a character string.
398 */
gp_state_getname(short gs)399 static const char *gp_state_getname(short gs)
400 {
401 if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
402 return "???";
403 return gp_state_names[gs];
404 }
405
406 /* Is the RCU grace-period kthread being starved of CPU time? */
rcu_is_gp_kthread_starving(unsigned long * jp)407 static bool rcu_is_gp_kthread_starving(unsigned long *jp)
408 {
409 unsigned long j = jiffies - READ_ONCE(rcu_state.gp_activity);
410
411 if (jp)
412 *jp = j;
413 return j > 2 * HZ;
414 }
415
rcu_is_rcuc_kthread_starving(struct rcu_data * rdp,unsigned long * jp)416 static bool rcu_is_rcuc_kthread_starving(struct rcu_data *rdp, unsigned long *jp)
417 {
418 int cpu;
419 struct task_struct *rcuc;
420 unsigned long j;
421
422 rcuc = rdp->rcu_cpu_kthread_task;
423 if (!rcuc)
424 return false;
425
426 cpu = task_cpu(rcuc);
427 if (cpu_is_offline(cpu) || idle_cpu(cpu))
428 return false;
429
430 j = jiffies - READ_ONCE(rdp->rcuc_activity);
431
432 if (jp)
433 *jp = j;
434 return j > 2 * HZ;
435 }
436
print_cpu_stat_info(int cpu)437 static void print_cpu_stat_info(int cpu)
438 {
439 struct rcu_snap_record rsr, *rsrp;
440 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
441 struct kernel_cpustat *kcsp = &kcpustat_cpu(cpu);
442
443 if (!rcu_cpu_stall_cputime)
444 return;
445
446 rsrp = &rdp->snap_record;
447 if (rsrp->gp_seq != rdp->gp_seq)
448 return;
449
450 rsr.cputime_irq = kcpustat_field(kcsp, CPUTIME_IRQ, cpu);
451 rsr.cputime_softirq = kcpustat_field(kcsp, CPUTIME_SOFTIRQ, cpu);
452 rsr.cputime_system = kcpustat_field(kcsp, CPUTIME_SYSTEM, cpu);
453
454 pr_err("\t hardirqs softirqs csw/system\n");
455 pr_err("\t number: %8ld %10d %12lld\n",
456 kstat_cpu_irqs_sum(cpu) - rsrp->nr_hardirqs,
457 kstat_cpu_softirqs_sum(cpu) - rsrp->nr_softirqs,
458 nr_context_switches_cpu(cpu) - rsrp->nr_csw);
459 pr_err("\tcputime: %8lld %10lld %12lld ==> %d(ms)\n",
460 div_u64(rsr.cputime_irq - rsrp->cputime_irq, NSEC_PER_MSEC),
461 div_u64(rsr.cputime_softirq - rsrp->cputime_softirq, NSEC_PER_MSEC),
462 div_u64(rsr.cputime_system - rsrp->cputime_system, NSEC_PER_MSEC),
463 jiffies_to_msecs(jiffies - rsrp->jiffies));
464 }
465
466 /*
467 * Print out diagnostic information for the specified stalled CPU.
468 *
469 * If the specified CPU is aware of the current RCU grace period, then
470 * print the number of scheduling clock interrupts the CPU has taken
471 * during the time that it has been aware. Otherwise, print the number
472 * of RCU grace periods that this CPU is ignorant of, for example, "1"
473 * if the CPU was aware of the previous grace period.
474 *
475 * Also print out idle info.
476 */
print_cpu_stall_info(int cpu)477 static void print_cpu_stall_info(int cpu)
478 {
479 unsigned long delta;
480 bool falsepositive;
481 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
482 char *ticks_title;
483 unsigned long ticks_value;
484 bool rcuc_starved;
485 unsigned long j;
486 char buf[32];
487
488 /*
489 * We could be printing a lot while holding a spinlock. Avoid
490 * triggering hard lockup.
491 */
492 touch_nmi_watchdog();
493
494 ticks_value = rcu_seq_ctr(rcu_state.gp_seq - rdp->gp_seq);
495 if (ticks_value) {
496 ticks_title = "GPs behind";
497 } else {
498 ticks_title = "ticks this GP";
499 ticks_value = rdp->ticks_this_gp;
500 }
501 delta = rcu_seq_ctr(rdp->mynode->gp_seq - rdp->rcu_iw_gp_seq);
502 falsepositive = rcu_is_gp_kthread_starving(NULL) &&
503 rcu_dynticks_in_eqs(rcu_dynticks_snap(cpu));
504 rcuc_starved = rcu_is_rcuc_kthread_starving(rdp, &j);
505 if (rcuc_starved)
506 // Print signed value, as negative values indicate a probable bug.
507 snprintf(buf, sizeof(buf), " rcuc=%ld jiffies(starved)", j);
508 pr_err("\t%d-%c%c%c%c: (%lu %s) idle=%04x/%ld/%#lx softirq=%u/%u fqs=%ld%s%s\n",
509 cpu,
510 "O."[!!cpu_online(cpu)],
511 "o."[!!(rdp->grpmask & rdp->mynode->qsmaskinit)],
512 "N."[!!(rdp->grpmask & rdp->mynode->qsmaskinitnext)],
513 !IS_ENABLED(CONFIG_IRQ_WORK) ? '?' :
514 rdp->rcu_iw_pending ? (int)min(delta, 9UL) + '0' :
515 "!."[!delta],
516 ticks_value, ticks_title,
517 rcu_dynticks_snap(cpu) & 0xffff,
518 ct_dynticks_nesting_cpu(cpu), ct_dynticks_nmi_nesting_cpu(cpu),
519 rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
520 data_race(rcu_state.n_force_qs) - rcu_state.n_force_qs_gpstart,
521 rcuc_starved ? buf : "",
522 falsepositive ? " (false positive?)" : "");
523
524 print_cpu_stat_info(cpu);
525 }
526
527 /* Complain about starvation of grace-period kthread. */
rcu_check_gp_kthread_starvation(void)528 static void rcu_check_gp_kthread_starvation(void)
529 {
530 int cpu;
531 struct task_struct *gpk = rcu_state.gp_kthread;
532 unsigned long j;
533
534 if (rcu_is_gp_kthread_starving(&j)) {
535 cpu = gpk ? task_cpu(gpk) : -1;
536 pr_err("%s kthread starved for %ld jiffies! g%ld f%#x %s(%d) ->state=%#x ->cpu=%d\n",
537 rcu_state.name, j,
538 (long)rcu_seq_current(&rcu_state.gp_seq),
539 data_race(READ_ONCE(rcu_state.gp_flags)),
540 gp_state_getname(rcu_state.gp_state),
541 data_race(READ_ONCE(rcu_state.gp_state)),
542 gpk ? data_race(READ_ONCE(gpk->__state)) : ~0, cpu);
543 if (gpk) {
544 pr_err("\tUnless %s kthread gets sufficient CPU time, OOM is now expected behavior.\n", rcu_state.name);
545 pr_err("RCU grace-period kthread stack dump:\n");
546 sched_show_task(gpk);
547 if (cpu >= 0) {
548 if (cpu_is_offline(cpu)) {
549 pr_err("RCU GP kthread last ran on offline CPU %d.\n", cpu);
550 } else {
551 pr_err("Stack dump where RCU GP kthread last ran:\n");
552 dump_cpu_task(cpu);
553 }
554 }
555 wake_up_process(gpk);
556 }
557 }
558 }
559
560 /* Complain about missing wakeups from expired fqs wait timer */
rcu_check_gp_kthread_expired_fqs_timer(void)561 static void rcu_check_gp_kthread_expired_fqs_timer(void)
562 {
563 struct task_struct *gpk = rcu_state.gp_kthread;
564 short gp_state;
565 unsigned long jiffies_fqs;
566 int cpu;
567
568 /*
569 * Order reads of .gp_state and .jiffies_force_qs.
570 * Matching smp_wmb() is present in rcu_gp_fqs_loop().
571 */
572 gp_state = smp_load_acquire(&rcu_state.gp_state);
573 jiffies_fqs = READ_ONCE(rcu_state.jiffies_force_qs);
574
575 if (gp_state == RCU_GP_WAIT_FQS &&
576 time_after(jiffies, jiffies_fqs + RCU_STALL_MIGHT_MIN) &&
577 gpk && !READ_ONCE(gpk->on_rq)) {
578 cpu = task_cpu(gpk);
579 pr_err("%s kthread timer wakeup didn't happen for %ld jiffies! g%ld f%#x %s(%d) ->state=%#x\n",
580 rcu_state.name, (jiffies - jiffies_fqs),
581 (long)rcu_seq_current(&rcu_state.gp_seq),
582 data_race(rcu_state.gp_flags),
583 gp_state_getname(RCU_GP_WAIT_FQS), RCU_GP_WAIT_FQS,
584 data_race(READ_ONCE(gpk->__state)));
585 pr_err("\tPossible timer handling issue on cpu=%d timer-softirq=%u\n",
586 cpu, kstat_softirqs_cpu(TIMER_SOFTIRQ, cpu));
587 }
588 }
589
print_other_cpu_stall(unsigned long gp_seq,unsigned long gps)590 static void print_other_cpu_stall(unsigned long gp_seq, unsigned long gps)
591 {
592 int cpu;
593 unsigned long flags;
594 unsigned long gpa;
595 unsigned long j;
596 int ndetected = 0;
597 struct rcu_node *rnp;
598 long totqlen = 0;
599
600 lockdep_assert_irqs_disabled();
601
602 /* Kick and suppress, if so configured. */
603 rcu_stall_kick_kthreads();
604 if (rcu_stall_is_suppressed())
605 return;
606
607 /*
608 * OK, time to rat on our buddy...
609 * See Documentation/RCU/stallwarn.rst for info on how to debug
610 * RCU CPU stall warnings.
611 */
612 trace_rcu_stall_warning(rcu_state.name, TPS("StallDetected"));
613 pr_err("INFO: %s detected stalls on CPUs/tasks:\n", rcu_state.name);
614 rcu_for_each_leaf_node(rnp) {
615 raw_spin_lock_irqsave_rcu_node(rnp, flags);
616 if (rnp->qsmask != 0) {
617 for_each_leaf_node_possible_cpu(rnp, cpu)
618 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
619 print_cpu_stall_info(cpu);
620 ndetected++;
621 }
622 }
623 ndetected += rcu_print_task_stall(rnp, flags); // Releases rnp->lock.
624 lockdep_assert_irqs_disabled();
625 }
626
627 for_each_possible_cpu(cpu)
628 totqlen += rcu_get_n_cbs_cpu(cpu);
629 pr_err("\t(detected by %d, t=%ld jiffies, g=%ld, q=%lu ncpus=%d)\n",
630 smp_processor_id(), (long)(jiffies - gps),
631 (long)rcu_seq_current(&rcu_state.gp_seq), totqlen, rcu_state.n_online_cpus);
632 if (ndetected) {
633 rcu_dump_cpu_stacks();
634
635 /* Complain about tasks blocking the grace period. */
636 rcu_for_each_leaf_node(rnp)
637 rcu_print_detail_task_stall_rnp(rnp);
638 } else {
639 if (rcu_seq_current(&rcu_state.gp_seq) != gp_seq) {
640 pr_err("INFO: Stall ended before state dump start\n");
641 } else {
642 j = jiffies;
643 gpa = data_race(READ_ONCE(rcu_state.gp_activity));
644 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
645 rcu_state.name, j - gpa, j, gpa,
646 data_race(READ_ONCE(jiffies_till_next_fqs)),
647 data_race(READ_ONCE(rcu_get_root()->qsmask)));
648 }
649 }
650 /* Rewrite if needed in case of slow consoles. */
651 if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
652 WRITE_ONCE(rcu_state.jiffies_stall,
653 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
654
655 rcu_check_gp_kthread_expired_fqs_timer();
656 rcu_check_gp_kthread_starvation();
657
658 panic_on_rcu_stall();
659
660 rcu_force_quiescent_state(); /* Kick them all. */
661 }
662
print_cpu_stall(unsigned long gps)663 static void print_cpu_stall(unsigned long gps)
664 {
665 int cpu;
666 unsigned long flags;
667 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
668 struct rcu_node *rnp = rcu_get_root();
669 long totqlen = 0;
670
671 lockdep_assert_irqs_disabled();
672
673 /* Kick and suppress, if so configured. */
674 rcu_stall_kick_kthreads();
675 if (rcu_stall_is_suppressed())
676 return;
677
678 /*
679 * OK, time to rat on ourselves...
680 * See Documentation/RCU/stallwarn.rst for info on how to debug
681 * RCU CPU stall warnings.
682 */
683 trace_rcu_stall_warning(rcu_state.name, TPS("SelfDetected"));
684 pr_err("INFO: %s self-detected stall on CPU\n", rcu_state.name);
685 raw_spin_lock_irqsave_rcu_node(rdp->mynode, flags);
686 print_cpu_stall_info(smp_processor_id());
687 raw_spin_unlock_irqrestore_rcu_node(rdp->mynode, flags);
688 for_each_possible_cpu(cpu)
689 totqlen += rcu_get_n_cbs_cpu(cpu);
690 pr_err("\t(t=%lu jiffies g=%ld q=%lu ncpus=%d)\n",
691 jiffies - gps,
692 (long)rcu_seq_current(&rcu_state.gp_seq), totqlen, rcu_state.n_online_cpus);
693
694 rcu_check_gp_kthread_expired_fqs_timer();
695 rcu_check_gp_kthread_starvation();
696
697 rcu_dump_cpu_stacks();
698
699 raw_spin_lock_irqsave_rcu_node(rnp, flags);
700 /* Rewrite if needed in case of slow consoles. */
701 if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
702 WRITE_ONCE(rcu_state.jiffies_stall,
703 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
704 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
705
706 panic_on_rcu_stall();
707
708 /*
709 * Attempt to revive the RCU machinery by forcing a context switch.
710 *
711 * A context switch would normally allow the RCU state machine to make
712 * progress and it could be we're stuck in kernel space without context
713 * switches for an entirely unreasonable amount of time.
714 */
715 set_tsk_need_resched(current);
716 set_preempt_need_resched();
717 }
718
check_cpu_stall(struct rcu_data * rdp)719 static void check_cpu_stall(struct rcu_data *rdp)
720 {
721 bool didstall = false;
722 unsigned long gs1;
723 unsigned long gs2;
724 unsigned long gps;
725 unsigned long j;
726 unsigned long jn;
727 unsigned long js;
728 struct rcu_node *rnp;
729
730 lockdep_assert_irqs_disabled();
731 if ((rcu_stall_is_suppressed() && !READ_ONCE(rcu_kick_kthreads)) ||
732 !rcu_gp_in_progress())
733 return;
734 rcu_stall_kick_kthreads();
735
736 /*
737 * Check if it was requested (via rcu_cpu_stall_reset()) that the FQS
738 * loop has to set jiffies to ensure a non-stale jiffies value. This
739 * is required to have good jiffies value after coming out of long
740 * breaks of jiffies updates. Not doing so can cause false positives.
741 */
742 if (READ_ONCE(rcu_state.nr_fqs_jiffies_stall) > 0)
743 return;
744
745 j = jiffies;
746
747 /*
748 * Lots of memory barriers to reject false positives.
749 *
750 * The idea is to pick up rcu_state.gp_seq, then
751 * rcu_state.jiffies_stall, then rcu_state.gp_start, and finally
752 * another copy of rcu_state.gp_seq. These values are updated in
753 * the opposite order with memory barriers (or equivalent) during
754 * grace-period initialization and cleanup. Now, a false positive
755 * can occur if we get an new value of rcu_state.gp_start and a old
756 * value of rcu_state.jiffies_stall. But given the memory barriers,
757 * the only way that this can happen is if one grace period ends
758 * and another starts between these two fetches. This is detected
759 * by comparing the second fetch of rcu_state.gp_seq with the
760 * previous fetch from rcu_state.gp_seq.
761 *
762 * Given this check, comparisons of jiffies, rcu_state.jiffies_stall,
763 * and rcu_state.gp_start suffice to forestall false positives.
764 */
765 gs1 = READ_ONCE(rcu_state.gp_seq);
766 smp_rmb(); /* Pick up ->gp_seq first... */
767 js = READ_ONCE(rcu_state.jiffies_stall);
768 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
769 gps = READ_ONCE(rcu_state.gp_start);
770 smp_rmb(); /* ...and finally ->gp_start before ->gp_seq again. */
771 gs2 = READ_ONCE(rcu_state.gp_seq);
772 if (gs1 != gs2 ||
773 ULONG_CMP_LT(j, js) ||
774 ULONG_CMP_GE(gps, js))
775 return; /* No stall or GP completed since entering function. */
776 rnp = rdp->mynode;
777 jn = jiffies + ULONG_MAX / 2;
778 if (rcu_gp_in_progress() &&
779 (READ_ONCE(rnp->qsmask) & rdp->grpmask) &&
780 cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
781
782 /*
783 * If a virtual machine is stopped by the host it can look to
784 * the watchdog like an RCU stall. Check to see if the host
785 * stopped the vm.
786 */
787 if (kvm_check_and_clear_guest_paused())
788 return;
789
790 /* We haven't checked in, so go dump stack. */
791 print_cpu_stall(gps);
792 if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
793 rcu_ftrace_dump(DUMP_ALL);
794 didstall = true;
795
796 } else if (rcu_gp_in_progress() &&
797 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY) &&
798 cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
799
800 /*
801 * If a virtual machine is stopped by the host it can look to
802 * the watchdog like an RCU stall. Check to see if the host
803 * stopped the vm.
804 */
805 if (kvm_check_and_clear_guest_paused())
806 return;
807
808 /* They had a few time units to dump stack, so complain. */
809 print_other_cpu_stall(gs2, gps);
810 if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
811 rcu_ftrace_dump(DUMP_ALL);
812 didstall = true;
813 }
814 if (didstall && READ_ONCE(rcu_state.jiffies_stall) == jn) {
815 jn = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
816 WRITE_ONCE(rcu_state.jiffies_stall, jn);
817 }
818 }
819
820 //////////////////////////////////////////////////////////////////////////////
821 //
822 // RCU forward-progress mechanisms, including of callback invocation.
823
824
825 /*
826 * Check to see if a failure to end RCU priority inversion was due to
827 * a CPU not passing through a quiescent state. When this happens, there
828 * is nothing that RCU priority boosting can do to help, so we shouldn't
829 * count this as an RCU priority boosting failure. A return of true says
830 * RCU priority boosting is to blame, and false says otherwise. If false
831 * is returned, the first of the CPUs to blame is stored through cpup.
832 * If there was no CPU blocking the current grace period, but also nothing
833 * in need of being boosted, *cpup is set to -1. This can happen in case
834 * of vCPU preemption while the last CPU is reporting its quiscent state,
835 * for example.
836 *
837 * If cpup is NULL, then a lockless quick check is carried out, suitable
838 * for high-rate usage. On the other hand, if cpup is non-NULL, each
839 * rcu_node structure's ->lock is acquired, ruling out high-rate usage.
840 */
rcu_check_boost_fail(unsigned long gp_state,int * cpup)841 bool rcu_check_boost_fail(unsigned long gp_state, int *cpup)
842 {
843 bool atb = false;
844 int cpu;
845 unsigned long flags;
846 struct rcu_node *rnp;
847
848 rcu_for_each_leaf_node(rnp) {
849 if (!cpup) {
850 if (data_race(READ_ONCE(rnp->qsmask))) {
851 return false;
852 } else {
853 if (READ_ONCE(rnp->gp_tasks))
854 atb = true;
855 continue;
856 }
857 }
858 *cpup = -1;
859 raw_spin_lock_irqsave_rcu_node(rnp, flags);
860 if (rnp->gp_tasks)
861 atb = true;
862 if (!rnp->qsmask) {
863 // No CPUs without quiescent states for this rnp.
864 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
865 continue;
866 }
867 // Find the first holdout CPU.
868 for_each_leaf_node_possible_cpu(rnp, cpu) {
869 if (rnp->qsmask & (1UL << (cpu - rnp->grplo))) {
870 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
871 *cpup = cpu;
872 return false;
873 }
874 }
875 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
876 }
877 // Can't blame CPUs, so must blame RCU priority boosting.
878 return atb;
879 }
880 EXPORT_SYMBOL_GPL(rcu_check_boost_fail);
881
882 /*
883 * Show the state of the grace-period kthreads.
884 */
show_rcu_gp_kthreads(void)885 void show_rcu_gp_kthreads(void)
886 {
887 unsigned long cbs = 0;
888 int cpu;
889 unsigned long j;
890 unsigned long ja;
891 unsigned long jr;
892 unsigned long js;
893 unsigned long jw;
894 struct rcu_data *rdp;
895 struct rcu_node *rnp;
896 struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
897
898 j = jiffies;
899 ja = j - data_race(READ_ONCE(rcu_state.gp_activity));
900 jr = j - data_race(READ_ONCE(rcu_state.gp_req_activity));
901 js = j - data_race(READ_ONCE(rcu_state.gp_start));
902 jw = j - data_race(READ_ONCE(rcu_state.gp_wake_time));
903 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",
904 rcu_state.name, gp_state_getname(rcu_state.gp_state),
905 data_race(READ_ONCE(rcu_state.gp_state)),
906 t ? data_race(READ_ONCE(t->__state)) : 0x1ffff, t ? t->rt_priority : 0xffU,
907 js, ja, jr, jw, (long)data_race(READ_ONCE(rcu_state.gp_wake_seq)),
908 (long)data_race(READ_ONCE(rcu_state.gp_seq)),
909 (long)data_race(READ_ONCE(rcu_get_root()->gp_seq_needed)),
910 data_race(READ_ONCE(rcu_state.gp_max)),
911 data_race(READ_ONCE(rcu_state.gp_flags)));
912 rcu_for_each_node_breadth_first(rnp) {
913 if (ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq), READ_ONCE(rnp->gp_seq_needed)) &&
914 !data_race(READ_ONCE(rnp->qsmask)) && !data_race(READ_ONCE(rnp->boost_tasks)) &&
915 !data_race(READ_ONCE(rnp->exp_tasks)) && !data_race(READ_ONCE(rnp->gp_tasks)))
916 continue;
917 pr_info("\trcu_node %d:%d ->gp_seq %ld ->gp_seq_needed %ld ->qsmask %#lx %c%c%c%c ->n_boosts %ld\n",
918 rnp->grplo, rnp->grphi,
919 (long)data_race(READ_ONCE(rnp->gp_seq)),
920 (long)data_race(READ_ONCE(rnp->gp_seq_needed)),
921 data_race(READ_ONCE(rnp->qsmask)),
922 ".b"[!!data_race(READ_ONCE(rnp->boost_kthread_task))],
923 ".B"[!!data_race(READ_ONCE(rnp->boost_tasks))],
924 ".E"[!!data_race(READ_ONCE(rnp->exp_tasks))],
925 ".G"[!!data_race(READ_ONCE(rnp->gp_tasks))],
926 data_race(READ_ONCE(rnp->n_boosts)));
927 if (!rcu_is_leaf_node(rnp))
928 continue;
929 for_each_leaf_node_possible_cpu(rnp, cpu) {
930 rdp = per_cpu_ptr(&rcu_data, cpu);
931 if (READ_ONCE(rdp->gpwrap) ||
932 ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq),
933 READ_ONCE(rdp->gp_seq_needed)))
934 continue;
935 pr_info("\tcpu %d ->gp_seq_needed %ld\n",
936 cpu, (long)data_race(READ_ONCE(rdp->gp_seq_needed)));
937 }
938 }
939 for_each_possible_cpu(cpu) {
940 rdp = per_cpu_ptr(&rcu_data, cpu);
941 cbs += data_race(READ_ONCE(rdp->n_cbs_invoked));
942 if (rcu_segcblist_is_offloaded(&rdp->cblist))
943 show_rcu_nocb_state(rdp);
944 }
945 pr_info("RCU callbacks invoked since boot: %lu\n", cbs);
946 show_rcu_tasks_gp_kthreads();
947 }
948 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
949
950 /*
951 * This function checks for grace-period requests that fail to motivate
952 * RCU to come out of its idle mode.
953 */
rcu_check_gp_start_stall(struct rcu_node * rnp,struct rcu_data * rdp,const unsigned long gpssdelay)954 static void rcu_check_gp_start_stall(struct rcu_node *rnp, struct rcu_data *rdp,
955 const unsigned long gpssdelay)
956 {
957 unsigned long flags;
958 unsigned long j;
959 struct rcu_node *rnp_root = rcu_get_root();
960 static atomic_t warned = ATOMIC_INIT(0);
961
962 if (!IS_ENABLED(CONFIG_PROVE_RCU) || rcu_gp_in_progress() ||
963 ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
964 READ_ONCE(rnp_root->gp_seq_needed)) ||
965 !smp_load_acquire(&rcu_state.gp_kthread)) // Get stable kthread.
966 return;
967 j = jiffies; /* Expensive access, and in common case don't get here. */
968 if (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 return;
972
973 raw_spin_lock_irqsave_rcu_node(rnp, flags);
974 j = jiffies;
975 if (rcu_gp_in_progress() ||
976 ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
977 READ_ONCE(rnp_root->gp_seq_needed)) ||
978 time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
979 time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
980 atomic_read(&warned)) {
981 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
982 return;
983 }
984 /* Hold onto the leaf lock to make others see warned==1. */
985
986 if (rnp_root != rnp)
987 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
988 j = jiffies;
989 if (rcu_gp_in_progress() ||
990 ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
991 READ_ONCE(rnp_root->gp_seq_needed)) ||
992 time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
993 time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
994 atomic_xchg(&warned, 1)) {
995 if (rnp_root != rnp)
996 /* irqs remain disabled. */
997 raw_spin_unlock_rcu_node(rnp_root);
998 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
999 return;
1000 }
1001 WARN_ON(1);
1002 if (rnp_root != rnp)
1003 raw_spin_unlock_rcu_node(rnp_root);
1004 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1005 show_rcu_gp_kthreads();
1006 }
1007
1008 /*
1009 * Do a forward-progress check for rcutorture. This is normally invoked
1010 * due to an OOM event. The argument "j" gives the time period during
1011 * which rcutorture would like progress to have been made.
1012 */
rcu_fwd_progress_check(unsigned long j)1013 void rcu_fwd_progress_check(unsigned long j)
1014 {
1015 unsigned long cbs;
1016 int cpu;
1017 unsigned long max_cbs = 0;
1018 int max_cpu = -1;
1019 struct rcu_data *rdp;
1020
1021 if (rcu_gp_in_progress()) {
1022 pr_info("%s: GP age %lu jiffies\n",
1023 __func__, jiffies - data_race(READ_ONCE(rcu_state.gp_start)));
1024 show_rcu_gp_kthreads();
1025 } else {
1026 pr_info("%s: Last GP end %lu jiffies ago\n",
1027 __func__, jiffies - data_race(READ_ONCE(rcu_state.gp_end)));
1028 preempt_disable();
1029 rdp = this_cpu_ptr(&rcu_data);
1030 rcu_check_gp_start_stall(rdp->mynode, rdp, j);
1031 preempt_enable();
1032 }
1033 for_each_possible_cpu(cpu) {
1034 cbs = rcu_get_n_cbs_cpu(cpu);
1035 if (!cbs)
1036 continue;
1037 if (max_cpu < 0)
1038 pr_info("%s: callbacks", __func__);
1039 pr_cont(" %d: %lu", cpu, cbs);
1040 if (cbs <= max_cbs)
1041 continue;
1042 max_cbs = cbs;
1043 max_cpu = cpu;
1044 }
1045 if (max_cpu >= 0)
1046 pr_cont("\n");
1047 }
1048 EXPORT_SYMBOL_GPL(rcu_fwd_progress_check);
1049
1050 /* Commandeer a sysrq key to dump RCU's tree. */
1051 static bool sysrq_rcu;
1052 module_param(sysrq_rcu, bool, 0444);
1053
1054 /* Dump grace-period-request information due to commandeered sysrq. */
sysrq_show_rcu(u8 key)1055 static void sysrq_show_rcu(u8 key)
1056 {
1057 show_rcu_gp_kthreads();
1058 }
1059
1060 static const struct sysrq_key_op sysrq_rcudump_op = {
1061 .handler = sysrq_show_rcu,
1062 .help_msg = "show-rcu(y)",
1063 .action_msg = "Show RCU tree",
1064 .enable_mask = SYSRQ_ENABLE_DUMP,
1065 };
1066
rcu_sysrq_init(void)1067 static int __init rcu_sysrq_init(void)
1068 {
1069 if (sysrq_rcu)
1070 return register_sysrq_key('y', &sysrq_rcudump_op);
1071 return 0;
1072 }
1073 early_initcall(rcu_sysrq_init);
1074