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, 21UL * 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
437 /*
438 * Print out diagnostic information for the specified stalled CPU.
439 *
440 * If the specified CPU is aware of the current RCU grace period, then
441 * print the number of scheduling clock interrupts the CPU has taken
442 * during the time that it has been aware. Otherwise, print the number
443 * of RCU grace periods that this CPU is ignorant of, for example, "1"
444 * if the CPU was aware of the previous grace period.
445 *
446 * Also print out idle info.
447 */
print_cpu_stall_info(int cpu)448 static void print_cpu_stall_info(int cpu)
449 {
450 unsigned long delta;
451 bool falsepositive;
452 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
453 char *ticks_title;
454 unsigned long ticks_value;
455 bool rcuc_starved;
456 unsigned long j;
457 char buf[32];
458
459 /*
460 * We could be printing a lot while holding a spinlock. Avoid
461 * triggering hard lockup.
462 */
463 touch_nmi_watchdog();
464
465 ticks_value = rcu_seq_ctr(rcu_state.gp_seq - rdp->gp_seq);
466 if (ticks_value) {
467 ticks_title = "GPs behind";
468 } else {
469 ticks_title = "ticks this GP";
470 ticks_value = rdp->ticks_this_gp;
471 }
472 delta = rcu_seq_ctr(rdp->mynode->gp_seq - rdp->rcu_iw_gp_seq);
473 falsepositive = rcu_is_gp_kthread_starving(NULL) &&
474 rcu_dynticks_in_eqs(rcu_dynticks_snap(cpu));
475 rcuc_starved = rcu_is_rcuc_kthread_starving(rdp, &j);
476 if (rcuc_starved)
477 sprintf(buf, " rcuc=%ld jiffies(starved)", j);
478 pr_err("\t%d-%c%c%c%c: (%lu %s) idle=%04x/%ld/%#lx softirq=%u/%u fqs=%ld%s%s\n",
479 cpu,
480 "O."[!!cpu_online(cpu)],
481 "o."[!!(rdp->grpmask & rdp->mynode->qsmaskinit)],
482 "N."[!!(rdp->grpmask & rdp->mynode->qsmaskinitnext)],
483 !IS_ENABLED(CONFIG_IRQ_WORK) ? '?' :
484 rdp->rcu_iw_pending ? (int)min(delta, 9UL) + '0' :
485 "!."[!delta],
486 ticks_value, ticks_title,
487 rcu_dynticks_snap(cpu) & 0xffff,
488 ct_dynticks_nesting_cpu(cpu), ct_dynticks_nmi_nesting_cpu(cpu),
489 rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
490 data_race(rcu_state.n_force_qs) - rcu_state.n_force_qs_gpstart,
491 rcuc_starved ? buf : "",
492 falsepositive ? " (false positive?)" : "");
493 }
494
495 /* Complain about starvation of grace-period kthread. */
rcu_check_gp_kthread_starvation(void)496 static void rcu_check_gp_kthread_starvation(void)
497 {
498 int cpu;
499 struct task_struct *gpk = rcu_state.gp_kthread;
500 unsigned long j;
501
502 if (rcu_is_gp_kthread_starving(&j)) {
503 cpu = gpk ? task_cpu(gpk) : -1;
504 pr_err("%s kthread starved for %ld jiffies! g%ld f%#x %s(%d) ->state=%#x ->cpu=%d\n",
505 rcu_state.name, j,
506 (long)rcu_seq_current(&rcu_state.gp_seq),
507 data_race(READ_ONCE(rcu_state.gp_flags)),
508 gp_state_getname(rcu_state.gp_state),
509 data_race(READ_ONCE(rcu_state.gp_state)),
510 gpk ? data_race(READ_ONCE(gpk->__state)) : ~0, cpu);
511 if (gpk) {
512 pr_err("\tUnless %s kthread gets sufficient CPU time, OOM is now expected behavior.\n", rcu_state.name);
513 pr_err("RCU grace-period kthread stack dump:\n");
514 sched_show_task(gpk);
515 if (cpu >= 0) {
516 if (cpu_is_offline(cpu)) {
517 pr_err("RCU GP kthread last ran on offline CPU %d.\n", cpu);
518 } else {
519 pr_err("Stack dump where RCU GP kthread last ran:\n");
520 dump_cpu_task(cpu);
521 }
522 }
523 wake_up_process(gpk);
524 }
525 }
526 }
527
528 /* Complain about missing wakeups from expired fqs wait timer */
rcu_check_gp_kthread_expired_fqs_timer(void)529 static void rcu_check_gp_kthread_expired_fqs_timer(void)
530 {
531 struct task_struct *gpk = rcu_state.gp_kthread;
532 short gp_state;
533 unsigned long jiffies_fqs;
534 int cpu;
535
536 /*
537 * Order reads of .gp_state and .jiffies_force_qs.
538 * Matching smp_wmb() is present in rcu_gp_fqs_loop().
539 */
540 gp_state = smp_load_acquire(&rcu_state.gp_state);
541 jiffies_fqs = READ_ONCE(rcu_state.jiffies_force_qs);
542
543 if (gp_state == RCU_GP_WAIT_FQS &&
544 time_after(jiffies, jiffies_fqs + RCU_STALL_MIGHT_MIN) &&
545 gpk && !READ_ONCE(gpk->on_rq)) {
546 cpu = task_cpu(gpk);
547 pr_err("%s kthread timer wakeup didn't happen for %ld jiffies! g%ld f%#x %s(%d) ->state=%#x\n",
548 rcu_state.name, (jiffies - jiffies_fqs),
549 (long)rcu_seq_current(&rcu_state.gp_seq),
550 data_race(rcu_state.gp_flags),
551 gp_state_getname(RCU_GP_WAIT_FQS), RCU_GP_WAIT_FQS,
552 data_race(READ_ONCE(gpk->__state)));
553 pr_err("\tPossible timer handling issue on cpu=%d timer-softirq=%u\n",
554 cpu, kstat_softirqs_cpu(TIMER_SOFTIRQ, cpu));
555 }
556 }
557
print_other_cpu_stall(unsigned long gp_seq,unsigned long gps)558 static void print_other_cpu_stall(unsigned long gp_seq, unsigned long gps)
559 {
560 int cpu;
561 unsigned long flags;
562 unsigned long gpa;
563 unsigned long j;
564 int ndetected = 0;
565 struct rcu_node *rnp;
566 long totqlen = 0;
567
568 lockdep_assert_irqs_disabled();
569
570 /* Kick and suppress, if so configured. */
571 rcu_stall_kick_kthreads();
572 if (rcu_stall_is_suppressed())
573 return;
574
575 /*
576 * OK, time to rat on our buddy...
577 * See Documentation/RCU/stallwarn.rst for info on how to debug
578 * RCU CPU stall warnings.
579 */
580 trace_rcu_stall_warning(rcu_state.name, TPS("StallDetected"));
581 pr_err("INFO: %s detected stalls on CPUs/tasks:\n", rcu_state.name);
582 rcu_for_each_leaf_node(rnp) {
583 raw_spin_lock_irqsave_rcu_node(rnp, flags);
584 if (rnp->qsmask != 0) {
585 for_each_leaf_node_possible_cpu(rnp, cpu)
586 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
587 print_cpu_stall_info(cpu);
588 ndetected++;
589 }
590 }
591 ndetected += rcu_print_task_stall(rnp, flags); // Releases rnp->lock.
592 lockdep_assert_irqs_disabled();
593 }
594
595 for_each_possible_cpu(cpu)
596 totqlen += rcu_get_n_cbs_cpu(cpu);
597 pr_cont("\t(detected by %d, t=%ld jiffies, g=%ld, q=%lu ncpus=%d)\n",
598 smp_processor_id(), (long)(jiffies - gps),
599 (long)rcu_seq_current(&rcu_state.gp_seq), totqlen, rcu_state.n_online_cpus);
600 if (ndetected) {
601 rcu_dump_cpu_stacks();
602
603 /* Complain about tasks blocking the grace period. */
604 rcu_for_each_leaf_node(rnp)
605 rcu_print_detail_task_stall_rnp(rnp);
606 } else {
607 if (rcu_seq_current(&rcu_state.gp_seq) != gp_seq) {
608 pr_err("INFO: Stall ended before state dump start\n");
609 } else {
610 j = jiffies;
611 gpa = data_race(READ_ONCE(rcu_state.gp_activity));
612 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
613 rcu_state.name, j - gpa, j, gpa,
614 data_race(READ_ONCE(jiffies_till_next_fqs)),
615 data_race(READ_ONCE(rcu_get_root()->qsmask)));
616 }
617 }
618 /* Rewrite if needed in case of slow consoles. */
619 if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
620 WRITE_ONCE(rcu_state.jiffies_stall,
621 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
622
623 rcu_check_gp_kthread_expired_fqs_timer();
624 rcu_check_gp_kthread_starvation();
625
626 panic_on_rcu_stall();
627
628 rcu_force_quiescent_state(); /* Kick them all. */
629 }
630
print_cpu_stall(unsigned long gps)631 static void print_cpu_stall(unsigned long gps)
632 {
633 int cpu;
634 unsigned long flags;
635 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
636 struct rcu_node *rnp = rcu_get_root();
637 long totqlen = 0;
638
639 lockdep_assert_irqs_disabled();
640
641 /* Kick and suppress, if so configured. */
642 rcu_stall_kick_kthreads();
643 if (rcu_stall_is_suppressed())
644 return;
645
646 /*
647 * OK, time to rat on ourselves...
648 * See Documentation/RCU/stallwarn.rst for info on how to debug
649 * RCU CPU stall warnings.
650 */
651 trace_rcu_stall_warning(rcu_state.name, TPS("SelfDetected"));
652 pr_err("INFO: %s self-detected stall on CPU\n", rcu_state.name);
653 raw_spin_lock_irqsave_rcu_node(rdp->mynode, flags);
654 print_cpu_stall_info(smp_processor_id());
655 raw_spin_unlock_irqrestore_rcu_node(rdp->mynode, flags);
656 for_each_possible_cpu(cpu)
657 totqlen += rcu_get_n_cbs_cpu(cpu);
658 pr_cont("\t(t=%lu jiffies g=%ld q=%lu ncpus=%d)\n",
659 jiffies - gps,
660 (long)rcu_seq_current(&rcu_state.gp_seq), totqlen, rcu_state.n_online_cpus);
661
662 rcu_check_gp_kthread_expired_fqs_timer();
663 rcu_check_gp_kthread_starvation();
664
665 rcu_dump_cpu_stacks();
666
667 raw_spin_lock_irqsave_rcu_node(rnp, flags);
668 /* Rewrite if needed in case of slow consoles. */
669 if (ULONG_CMP_GE(jiffies, READ_ONCE(rcu_state.jiffies_stall)))
670 WRITE_ONCE(rcu_state.jiffies_stall,
671 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
672 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
673
674 panic_on_rcu_stall();
675
676 /*
677 * Attempt to revive the RCU machinery by forcing a context switch.
678 *
679 * A context switch would normally allow the RCU state machine to make
680 * progress and it could be we're stuck in kernel space without context
681 * switches for an entirely unreasonable amount of time.
682 */
683 set_tsk_need_resched(current);
684 set_preempt_need_resched();
685 }
686
check_cpu_stall(struct rcu_data * rdp)687 static void check_cpu_stall(struct rcu_data *rdp)
688 {
689 bool didstall = false;
690 unsigned long gs1;
691 unsigned long gs2;
692 unsigned long gps;
693 unsigned long j;
694 unsigned long jn;
695 unsigned long js;
696 struct rcu_node *rnp;
697
698 lockdep_assert_irqs_disabled();
699 if ((rcu_stall_is_suppressed() && !READ_ONCE(rcu_kick_kthreads)) ||
700 !rcu_gp_in_progress())
701 return;
702 rcu_stall_kick_kthreads();
703
704 /*
705 * Check if it was requested (via rcu_cpu_stall_reset()) that the FQS
706 * loop has to set jiffies to ensure a non-stale jiffies value. This
707 * is required to have good jiffies value after coming out of long
708 * breaks of jiffies updates. Not doing so can cause false positives.
709 */
710 if (READ_ONCE(rcu_state.nr_fqs_jiffies_stall) > 0)
711 return;
712
713 j = jiffies;
714
715 /*
716 * Lots of memory barriers to reject false positives.
717 *
718 * The idea is to pick up rcu_state.gp_seq, then
719 * rcu_state.jiffies_stall, then rcu_state.gp_start, and finally
720 * another copy of rcu_state.gp_seq. These values are updated in
721 * the opposite order with memory barriers (or equivalent) during
722 * grace-period initialization and cleanup. Now, a false positive
723 * can occur if we get an new value of rcu_state.gp_start and a old
724 * value of rcu_state.jiffies_stall. But given the memory barriers,
725 * the only way that this can happen is if one grace period ends
726 * and another starts between these two fetches. This is detected
727 * by comparing the second fetch of rcu_state.gp_seq with the
728 * previous fetch from rcu_state.gp_seq.
729 *
730 * Given this check, comparisons of jiffies, rcu_state.jiffies_stall,
731 * and rcu_state.gp_start suffice to forestall false positives.
732 */
733 gs1 = READ_ONCE(rcu_state.gp_seq);
734 smp_rmb(); /* Pick up ->gp_seq first... */
735 js = READ_ONCE(rcu_state.jiffies_stall);
736 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
737 gps = READ_ONCE(rcu_state.gp_start);
738 smp_rmb(); /* ...and finally ->gp_start before ->gp_seq again. */
739 gs2 = READ_ONCE(rcu_state.gp_seq);
740 if (gs1 != gs2 ||
741 ULONG_CMP_LT(j, js) ||
742 ULONG_CMP_GE(gps, js))
743 return; /* No stall or GP completed since entering function. */
744 rnp = rdp->mynode;
745 jn = jiffies + ULONG_MAX / 2;
746 if (rcu_gp_in_progress() &&
747 (READ_ONCE(rnp->qsmask) & rdp->grpmask) &&
748 cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
749
750 /*
751 * If a virtual machine is stopped by the host it can look to
752 * the watchdog like an RCU stall. Check to see if the host
753 * stopped the vm.
754 */
755 if (kvm_check_and_clear_guest_paused())
756 return;
757
758 /* We haven't checked in, so go dump stack. */
759 print_cpu_stall(gps);
760 if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
761 rcu_ftrace_dump(DUMP_ALL);
762 didstall = true;
763
764 } else if (rcu_gp_in_progress() &&
765 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY) &&
766 cmpxchg(&rcu_state.jiffies_stall, js, jn) == js) {
767
768 /*
769 * If a virtual machine is stopped by the host it can look to
770 * the watchdog like an RCU stall. Check to see if the host
771 * stopped the vm.
772 */
773 if (kvm_check_and_clear_guest_paused())
774 return;
775
776 /* They had a few time units to dump stack, so complain. */
777 print_other_cpu_stall(gs2, gps);
778 if (READ_ONCE(rcu_cpu_stall_ftrace_dump))
779 rcu_ftrace_dump(DUMP_ALL);
780 didstall = true;
781 }
782 if (didstall && READ_ONCE(rcu_state.jiffies_stall) == jn) {
783 jn = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
784 WRITE_ONCE(rcu_state.jiffies_stall, jn);
785 }
786 }
787
788 //////////////////////////////////////////////////////////////////////////////
789 //
790 // RCU forward-progress mechanisms, including of callback invocation.
791
792
793 /*
794 * Check to see if a failure to end RCU priority inversion was due to
795 * a CPU not passing through a quiescent state. When this happens, there
796 * is nothing that RCU priority boosting can do to help, so we shouldn't
797 * count this as an RCU priority boosting failure. A return of true says
798 * RCU priority boosting is to blame, and false says otherwise. If false
799 * is returned, the first of the CPUs to blame is stored through cpup.
800 * If there was no CPU blocking the current grace period, but also nothing
801 * in need of being boosted, *cpup is set to -1. This can happen in case
802 * of vCPU preemption while the last CPU is reporting its quiscent state,
803 * for example.
804 *
805 * If cpup is NULL, then a lockless quick check is carried out, suitable
806 * for high-rate usage. On the other hand, if cpup is non-NULL, each
807 * rcu_node structure's ->lock is acquired, ruling out high-rate usage.
808 */
rcu_check_boost_fail(unsigned long gp_state,int * cpup)809 bool rcu_check_boost_fail(unsigned long gp_state, int *cpup)
810 {
811 bool atb = false;
812 int cpu;
813 unsigned long flags;
814 struct rcu_node *rnp;
815
816 rcu_for_each_leaf_node(rnp) {
817 if (!cpup) {
818 if (data_race(READ_ONCE(rnp->qsmask))) {
819 return false;
820 } else {
821 if (READ_ONCE(rnp->gp_tasks))
822 atb = true;
823 continue;
824 }
825 }
826 *cpup = -1;
827 raw_spin_lock_irqsave_rcu_node(rnp, flags);
828 if (rnp->gp_tasks)
829 atb = true;
830 if (!rnp->qsmask) {
831 // No CPUs without quiescent states for this rnp.
832 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
833 continue;
834 }
835 // Find the first holdout CPU.
836 for_each_leaf_node_possible_cpu(rnp, cpu) {
837 if (rnp->qsmask & (1UL << (cpu - rnp->grplo))) {
838 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
839 *cpup = cpu;
840 return false;
841 }
842 }
843 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
844 }
845 // Can't blame CPUs, so must blame RCU priority boosting.
846 return atb;
847 }
848 EXPORT_SYMBOL_GPL(rcu_check_boost_fail);
849
850 /*
851 * Show the state of the grace-period kthreads.
852 */
show_rcu_gp_kthreads(void)853 void show_rcu_gp_kthreads(void)
854 {
855 unsigned long cbs = 0;
856 int cpu;
857 unsigned long j;
858 unsigned long ja;
859 unsigned long jr;
860 unsigned long js;
861 unsigned long jw;
862 struct rcu_data *rdp;
863 struct rcu_node *rnp;
864 struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
865
866 j = jiffies;
867 ja = j - data_race(READ_ONCE(rcu_state.gp_activity));
868 jr = j - data_race(READ_ONCE(rcu_state.gp_req_activity));
869 js = j - data_race(READ_ONCE(rcu_state.gp_start));
870 jw = j - data_race(READ_ONCE(rcu_state.gp_wake_time));
871 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",
872 rcu_state.name, gp_state_getname(rcu_state.gp_state),
873 data_race(READ_ONCE(rcu_state.gp_state)),
874 t ? data_race(READ_ONCE(t->__state)) : 0x1ffff, t ? t->rt_priority : 0xffU,
875 js, ja, jr, jw, (long)data_race(READ_ONCE(rcu_state.gp_wake_seq)),
876 (long)data_race(READ_ONCE(rcu_state.gp_seq)),
877 (long)data_race(READ_ONCE(rcu_get_root()->gp_seq_needed)),
878 data_race(READ_ONCE(rcu_state.gp_max)),
879 data_race(READ_ONCE(rcu_state.gp_flags)));
880 rcu_for_each_node_breadth_first(rnp) {
881 if (ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq), READ_ONCE(rnp->gp_seq_needed)) &&
882 !data_race(READ_ONCE(rnp->qsmask)) && !data_race(READ_ONCE(rnp->boost_tasks)) &&
883 !data_race(READ_ONCE(rnp->exp_tasks)) && !data_race(READ_ONCE(rnp->gp_tasks)))
884 continue;
885 pr_info("\trcu_node %d:%d ->gp_seq %ld ->gp_seq_needed %ld ->qsmask %#lx %c%c%c%c ->n_boosts %ld\n",
886 rnp->grplo, rnp->grphi,
887 (long)data_race(READ_ONCE(rnp->gp_seq)),
888 (long)data_race(READ_ONCE(rnp->gp_seq_needed)),
889 data_race(READ_ONCE(rnp->qsmask)),
890 ".b"[!!data_race(READ_ONCE(rnp->boost_kthread_task))],
891 ".B"[!!data_race(READ_ONCE(rnp->boost_tasks))],
892 ".E"[!!data_race(READ_ONCE(rnp->exp_tasks))],
893 ".G"[!!data_race(READ_ONCE(rnp->gp_tasks))],
894 data_race(READ_ONCE(rnp->n_boosts)));
895 if (!rcu_is_leaf_node(rnp))
896 continue;
897 for_each_leaf_node_possible_cpu(rnp, cpu) {
898 rdp = per_cpu_ptr(&rcu_data, cpu);
899 if (READ_ONCE(rdp->gpwrap) ||
900 ULONG_CMP_GE(READ_ONCE(rcu_state.gp_seq),
901 READ_ONCE(rdp->gp_seq_needed)))
902 continue;
903 pr_info("\tcpu %d ->gp_seq_needed %ld\n",
904 cpu, (long)data_race(READ_ONCE(rdp->gp_seq_needed)));
905 }
906 }
907 for_each_possible_cpu(cpu) {
908 rdp = per_cpu_ptr(&rcu_data, cpu);
909 cbs += data_race(READ_ONCE(rdp->n_cbs_invoked));
910 if (rcu_segcblist_is_offloaded(&rdp->cblist))
911 show_rcu_nocb_state(rdp);
912 }
913 pr_info("RCU callbacks invoked since boot: %lu\n", cbs);
914 show_rcu_tasks_gp_kthreads();
915 }
916 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
917
918 /*
919 * This function checks for grace-period requests that fail to motivate
920 * RCU to come out of its idle mode.
921 */
rcu_check_gp_start_stall(struct rcu_node * rnp,struct rcu_data * rdp,const unsigned long gpssdelay)922 static void rcu_check_gp_start_stall(struct rcu_node *rnp, struct rcu_data *rdp,
923 const unsigned long gpssdelay)
924 {
925 unsigned long flags;
926 unsigned long j;
927 struct rcu_node *rnp_root = rcu_get_root();
928 static atomic_t warned = ATOMIC_INIT(0);
929
930 if (!IS_ENABLED(CONFIG_PROVE_RCU) || rcu_gp_in_progress() ||
931 ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
932 READ_ONCE(rnp_root->gp_seq_needed)) ||
933 !smp_load_acquire(&rcu_state.gp_kthread)) // Get stable kthread.
934 return;
935 j = jiffies; /* Expensive access, and in common case don't get here. */
936 if (time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
937 time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
938 atomic_read(&warned))
939 return;
940
941 raw_spin_lock_irqsave_rcu_node(rnp, flags);
942 j = jiffies;
943 if (rcu_gp_in_progress() ||
944 ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
945 READ_ONCE(rnp_root->gp_seq_needed)) ||
946 time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
947 time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
948 atomic_read(&warned)) {
949 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
950 return;
951 }
952 /* Hold onto the leaf lock to make others see warned==1. */
953
954 if (rnp_root != rnp)
955 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
956 j = jiffies;
957 if (rcu_gp_in_progress() ||
958 ULONG_CMP_GE(READ_ONCE(rnp_root->gp_seq),
959 READ_ONCE(rnp_root->gp_seq_needed)) ||
960 time_before(j, READ_ONCE(rcu_state.gp_req_activity) + gpssdelay) ||
961 time_before(j, READ_ONCE(rcu_state.gp_activity) + gpssdelay) ||
962 atomic_xchg(&warned, 1)) {
963 if (rnp_root != rnp)
964 /* irqs remain disabled. */
965 raw_spin_unlock_rcu_node(rnp_root);
966 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
967 return;
968 }
969 WARN_ON(1);
970 if (rnp_root != rnp)
971 raw_spin_unlock_rcu_node(rnp_root);
972 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
973 show_rcu_gp_kthreads();
974 }
975
976 /*
977 * Do a forward-progress check for rcutorture. This is normally invoked
978 * due to an OOM event. The argument "j" gives the time period during
979 * which rcutorture would like progress to have been made.
980 */
rcu_fwd_progress_check(unsigned long j)981 void rcu_fwd_progress_check(unsigned long j)
982 {
983 unsigned long cbs;
984 int cpu;
985 unsigned long max_cbs = 0;
986 int max_cpu = -1;
987 struct rcu_data *rdp;
988
989 if (rcu_gp_in_progress()) {
990 pr_info("%s: GP age %lu jiffies\n",
991 __func__, jiffies - data_race(READ_ONCE(rcu_state.gp_start)));
992 show_rcu_gp_kthreads();
993 } else {
994 pr_info("%s: Last GP end %lu jiffies ago\n",
995 __func__, jiffies - data_race(READ_ONCE(rcu_state.gp_end)));
996 preempt_disable();
997 rdp = this_cpu_ptr(&rcu_data);
998 rcu_check_gp_start_stall(rdp->mynode, rdp, j);
999 preempt_enable();
1000 }
1001 for_each_possible_cpu(cpu) {
1002 cbs = rcu_get_n_cbs_cpu(cpu);
1003 if (!cbs)
1004 continue;
1005 if (max_cpu < 0)
1006 pr_info("%s: callbacks", __func__);
1007 pr_cont(" %d: %lu", cpu, cbs);
1008 if (cbs <= max_cbs)
1009 continue;
1010 max_cbs = cbs;
1011 max_cpu = cpu;
1012 }
1013 if (max_cpu >= 0)
1014 pr_cont("\n");
1015 }
1016 EXPORT_SYMBOL_GPL(rcu_fwd_progress_check);
1017
1018 /* Commandeer a sysrq key to dump RCU's tree. */
1019 static bool sysrq_rcu;
1020 module_param(sysrq_rcu, bool, 0444);
1021
1022 /* Dump grace-period-request information due to commandeered sysrq. */
sysrq_show_rcu(int key)1023 static void sysrq_show_rcu(int key)
1024 {
1025 show_rcu_gp_kthreads();
1026 }
1027
1028 static const struct sysrq_key_op sysrq_rcudump_op = {
1029 .handler = sysrq_show_rcu,
1030 .help_msg = "show-rcu(y)",
1031 .action_msg = "Show RCU tree",
1032 .enable_mask = SYSRQ_ENABLE_DUMP,
1033 };
1034
rcu_sysrq_init(void)1035 static int __init rcu_sysrq_init(void)
1036 {
1037 if (sysrq_rcu)
1038 return register_sysrq_key('y', &sysrq_rcudump_op);
1039 return 0;
1040 }
1041 early_initcall(rcu_sysrq_init);
1042