1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Watchdog support on powerpc systems.
4  *
5  * Copyright 2017, IBM Corporation.
6  *
7  * This uses code from arch/sparc/kernel/nmi.c and kernel/watchdog.c
8  */
9 
10 #define pr_fmt(fmt) "watchdog: " fmt
11 
12 #include <linux/kernel.h>
13 #include <linux/param.h>
14 #include <linux/init.h>
15 #include <linux/percpu.h>
16 #include <linux/cpu.h>
17 #include <linux/nmi.h>
18 #include <linux/module.h>
19 #include <linux/export.h>
20 #include <linux/kprobes.h>
21 #include <linux/hardirq.h>
22 #include <linux/reboot.h>
23 #include <linux/slab.h>
24 #include <linux/kdebug.h>
25 #include <linux/sched/debug.h>
26 #include <linux/delay.h>
27 #include <linux/processor.h>
28 #include <linux/smp.h>
29 
30 #include <asm/interrupt.h>
31 #include <asm/paca.h>
32 #include <asm/nmi.h>
33 
34 /*
35  * The powerpc watchdog ensures that each CPU is able to service timers.
36  * The watchdog sets up a simple timer on each CPU to run once per timer
37  * period, and updates a per-cpu timestamp and a "pending" cpumask. This is
38  * the heartbeat.
39  *
40  * Then there are two systems to check that the heartbeat is still running.
41  * The local soft-NMI, and the SMP checker.
42  *
43  * The soft-NMI checker can detect lockups on the local CPU. When interrupts
44  * are disabled with local_irq_disable(), platforms that use soft-masking
45  * can leave hardware interrupts enabled and handle them with a masked
46  * interrupt handler. The masked handler can send the timer interrupt to the
47  * watchdog's soft_nmi_interrupt(), which appears to Linux as an NMI
48  * interrupt, and can be used to detect CPUs stuck with IRQs disabled.
49  *
50  * The soft-NMI checker will compare the heartbeat timestamp for this CPU
51  * with the current time, and take action if the difference exceeds the
52  * watchdog threshold.
53  *
54  * The limitation of the soft-NMI watchdog is that it does not work when
55  * interrupts are hard disabled or otherwise not being serviced. This is
56  * solved by also having a SMP watchdog where all CPUs check all other
57  * CPUs heartbeat.
58  *
59  * The SMP checker can detect lockups on other CPUs. A global "pending"
60  * cpumask is kept, containing all CPUs which enable the watchdog. Each
61  * CPU clears their pending bit in their heartbeat timer. When the bitmask
62  * becomes empty, the last CPU to clear its pending bit updates a global
63  * timestamp and refills the pending bitmask.
64  *
65  * In the heartbeat timer, if any CPU notices that the global timestamp has
66  * not been updated for a period exceeding the watchdog threshold, then it
67  * means the CPU(s) with their bit still set in the pending mask have had
68  * their heartbeat stop, and action is taken.
69  *
70  * Some platforms implement true NMI IPIs, which can be used by the SMP
71  * watchdog to detect an unresponsive CPU and pull it out of its stuck
72  * state with the NMI IPI, to get crash/debug data from it. This way the
73  * SMP watchdog can detect hardware interrupts off lockups.
74  */
75 
76 static cpumask_t wd_cpus_enabled __read_mostly;
77 
78 static u64 wd_panic_timeout_tb __read_mostly; /* timebase ticks until panic */
79 static u64 wd_smp_panic_timeout_tb __read_mostly; /* panic other CPUs */
80 
81 static u64 wd_timer_period_ms __read_mostly;  /* interval between heartbeat */
82 
83 static DEFINE_PER_CPU(struct hrtimer, wd_hrtimer);
84 static DEFINE_PER_CPU(u64, wd_timer_tb);
85 
86 /* SMP checker bits */
87 static unsigned long __wd_smp_lock;
88 static unsigned long __wd_reporting;
89 static unsigned long __wd_nmi_output;
90 static cpumask_t wd_smp_cpus_pending;
91 static cpumask_t wd_smp_cpus_stuck;
92 static u64 wd_smp_last_reset_tb;
93 
94 #ifdef CONFIG_PPC_PSERIES
95 static u64 wd_timeout_pct;
96 #endif
97 
98 /*
99  * Try to take the exclusive watchdog action / NMI IPI / printing lock.
100  * wd_smp_lock must be held. If this fails, we should return and wait
101  * for the watchdog to kick in again (or another CPU to trigger it).
102  *
103  * Importantly, if hardlockup_panic is set, wd_try_report failure should
104  * not delay the panic, because whichever other CPU is reporting will
105  * call panic.
106  */
wd_try_report(void)107 static bool wd_try_report(void)
108 {
109 	if (__wd_reporting)
110 		return false;
111 	__wd_reporting = 1;
112 	return true;
113 }
114 
115 /* End printing after successful wd_try_report. wd_smp_lock not required. */
wd_end_reporting(void)116 static void wd_end_reporting(void)
117 {
118 	smp_mb(); /* End printing "critical section" */
119 	WARN_ON_ONCE(__wd_reporting == 0);
120 	WRITE_ONCE(__wd_reporting, 0);
121 }
122 
wd_smp_lock(unsigned long * flags)123 static inline void wd_smp_lock(unsigned long *flags)
124 {
125 	/*
126 	 * Avoid locking layers if possible.
127 	 * This may be called from low level interrupt handlers at some
128 	 * point in future.
129 	 */
130 	raw_local_irq_save(*flags);
131 	hard_irq_disable(); /* Make it soft-NMI safe */
132 	while (unlikely(test_and_set_bit_lock(0, &__wd_smp_lock))) {
133 		raw_local_irq_restore(*flags);
134 		spin_until_cond(!test_bit(0, &__wd_smp_lock));
135 		raw_local_irq_save(*flags);
136 		hard_irq_disable();
137 	}
138 }
139 
wd_smp_unlock(unsigned long * flags)140 static inline void wd_smp_unlock(unsigned long *flags)
141 {
142 	clear_bit_unlock(0, &__wd_smp_lock);
143 	raw_local_irq_restore(*flags);
144 }
145 
wd_lockup_ipi(struct pt_regs * regs)146 static void wd_lockup_ipi(struct pt_regs *regs)
147 {
148 	int cpu = raw_smp_processor_id();
149 	u64 tb = get_tb();
150 
151 	pr_emerg("CPU %d Hard LOCKUP\n", cpu);
152 	pr_emerg("CPU %d TB:%lld, last heartbeat TB:%lld (%lldms ago)\n",
153 		 cpu, tb, per_cpu(wd_timer_tb, cpu),
154 		 tb_to_ns(tb - per_cpu(wd_timer_tb, cpu)) / 1000000);
155 	print_modules();
156 	print_irqtrace_events(current);
157 	if (regs)
158 		show_regs(regs);
159 	else
160 		dump_stack();
161 
162 	/*
163 	 * __wd_nmi_output must be set after we printk from NMI context.
164 	 *
165 	 * printk from NMI context defers printing to the console to irq_work.
166 	 * If that NMI was taken in some code that is hard-locked, then irqs
167 	 * are disabled so irq_work will never fire. That can result in the
168 	 * hard lockup messages being delayed (indefinitely, until something
169 	 * else kicks the console drivers).
170 	 *
171 	 * Setting __wd_nmi_output will cause another CPU to notice and kick
172 	 * the console drivers for us.
173 	 *
174 	 * xchg is not needed here (it could be a smp_mb and store), but xchg
175 	 * gives the memory ordering and atomicity required.
176 	 */
177 	xchg(&__wd_nmi_output, 1);
178 
179 	/* Do not panic from here because that can recurse into NMI IPI layer */
180 }
181 
set_cpu_stuck(int cpu)182 static bool set_cpu_stuck(int cpu)
183 {
184 	cpumask_set_cpu(cpu, &wd_smp_cpus_stuck);
185 	cpumask_clear_cpu(cpu, &wd_smp_cpus_pending);
186 	/*
187 	 * See wd_smp_clear_cpu_pending()
188 	 */
189 	smp_mb();
190 	if (cpumask_empty(&wd_smp_cpus_pending)) {
191 		wd_smp_last_reset_tb = get_tb();
192 		cpumask_andnot(&wd_smp_cpus_pending,
193 				&wd_cpus_enabled,
194 				&wd_smp_cpus_stuck);
195 		return true;
196 	}
197 	return false;
198 }
199 
watchdog_smp_panic(int cpu)200 static void watchdog_smp_panic(int cpu)
201 {
202 	static cpumask_t wd_smp_cpus_ipi; // protected by reporting
203 	unsigned long flags;
204 	u64 tb, last_reset;
205 	int c;
206 
207 	wd_smp_lock(&flags);
208 	/* Double check some things under lock */
209 	tb = get_tb();
210 	last_reset = wd_smp_last_reset_tb;
211 	if ((s64)(tb - last_reset) < (s64)wd_smp_panic_timeout_tb)
212 		goto out;
213 	if (cpumask_test_cpu(cpu, &wd_smp_cpus_pending))
214 		goto out;
215 	if (!wd_try_report())
216 		goto out;
217 	for_each_online_cpu(c) {
218 		if (!cpumask_test_cpu(c, &wd_smp_cpus_pending))
219 			continue;
220 		if (c == cpu)
221 			continue; // should not happen
222 
223 		__cpumask_set_cpu(c, &wd_smp_cpus_ipi);
224 		if (set_cpu_stuck(c))
225 			break;
226 	}
227 	if (cpumask_empty(&wd_smp_cpus_ipi)) {
228 		wd_end_reporting();
229 		goto out;
230 	}
231 	wd_smp_unlock(&flags);
232 
233 	pr_emerg("CPU %d detected hard LOCKUP on other CPUs %*pbl\n",
234 		 cpu, cpumask_pr_args(&wd_smp_cpus_ipi));
235 	pr_emerg("CPU %d TB:%lld, last SMP heartbeat TB:%lld (%lldms ago)\n",
236 		 cpu, tb, last_reset, tb_to_ns(tb - last_reset) / 1000000);
237 
238 	if (!sysctl_hardlockup_all_cpu_backtrace) {
239 		/*
240 		 * Try to trigger the stuck CPUs, unless we are going to
241 		 * get a backtrace on all of them anyway.
242 		 */
243 		for_each_cpu(c, &wd_smp_cpus_ipi) {
244 			smp_send_nmi_ipi(c, wd_lockup_ipi, 1000000);
245 			__cpumask_clear_cpu(c, &wd_smp_cpus_ipi);
246 		}
247 	} else {
248 		trigger_allbutself_cpu_backtrace();
249 		cpumask_clear(&wd_smp_cpus_ipi);
250 	}
251 
252 	if (hardlockup_panic)
253 		nmi_panic(NULL, "Hard LOCKUP");
254 
255 	wd_end_reporting();
256 
257 	return;
258 
259 out:
260 	wd_smp_unlock(&flags);
261 }
262 
wd_smp_clear_cpu_pending(int cpu)263 static void wd_smp_clear_cpu_pending(int cpu)
264 {
265 	if (!cpumask_test_cpu(cpu, &wd_smp_cpus_pending)) {
266 		if (unlikely(cpumask_test_cpu(cpu, &wd_smp_cpus_stuck))) {
267 			struct pt_regs *regs = get_irq_regs();
268 			unsigned long flags;
269 
270 			pr_emerg("CPU %d became unstuck TB:%lld\n",
271 				 cpu, get_tb());
272 			print_irqtrace_events(current);
273 			if (regs)
274 				show_regs(regs);
275 			else
276 				dump_stack();
277 
278 			wd_smp_lock(&flags);
279 			cpumask_clear_cpu(cpu, &wd_smp_cpus_stuck);
280 			wd_smp_unlock(&flags);
281 		} else {
282 			/*
283 			 * The last CPU to clear pending should have reset the
284 			 * watchdog so we generally should not find it empty
285 			 * here if our CPU was clear. However it could happen
286 			 * due to a rare race with another CPU taking the
287 			 * last CPU out of the mask concurrently.
288 			 *
289 			 * We can't add a warning for it. But just in case
290 			 * there is a problem with the watchdog that is causing
291 			 * the mask to not be reset, try to kick it along here.
292 			 */
293 			if (unlikely(cpumask_empty(&wd_smp_cpus_pending)))
294 				goto none_pending;
295 		}
296 		return;
297 	}
298 
299 	/*
300 	 * All other updates to wd_smp_cpus_pending are performed under
301 	 * wd_smp_lock. All of them are atomic except the case where the
302 	 * mask becomes empty and is reset. This will not happen here because
303 	 * cpu was tested to be in the bitmap (above), and a CPU only clears
304 	 * its own bit. _Except_ in the case where another CPU has detected a
305 	 * hard lockup on our CPU and takes us out of the pending mask. So in
306 	 * normal operation there will be no race here, no problem.
307 	 *
308 	 * In the lockup case, this atomic clear-bit vs a store that refills
309 	 * other bits in the accessed word wll not be a problem. The bit clear
310 	 * is atomic so it will not cause the store to get lost, and the store
311 	 * will never set this bit so it will not overwrite the bit clear. The
312 	 * only way for a stuck CPU to return to the pending bitmap is to
313 	 * become unstuck itself.
314 	 */
315 	cpumask_clear_cpu(cpu, &wd_smp_cpus_pending);
316 
317 	/*
318 	 * Order the store to clear pending with the load(s) to check all
319 	 * words in the pending mask to check they are all empty. This orders
320 	 * with the same barrier on another CPU. This prevents two CPUs
321 	 * clearing the last 2 pending bits, but neither seeing the other's
322 	 * store when checking if the mask is empty, and missing an empty
323 	 * mask, which ends with a false positive.
324 	 */
325 	smp_mb();
326 	if (cpumask_empty(&wd_smp_cpus_pending)) {
327 		unsigned long flags;
328 
329 none_pending:
330 		/*
331 		 * Double check under lock because more than one CPU could see
332 		 * a clear mask with the lockless check after clearing their
333 		 * pending bits.
334 		 */
335 		wd_smp_lock(&flags);
336 		if (cpumask_empty(&wd_smp_cpus_pending)) {
337 			wd_smp_last_reset_tb = get_tb();
338 			cpumask_andnot(&wd_smp_cpus_pending,
339 					&wd_cpus_enabled,
340 					&wd_smp_cpus_stuck);
341 		}
342 		wd_smp_unlock(&flags);
343 	}
344 }
345 
watchdog_timer_interrupt(int cpu)346 static void watchdog_timer_interrupt(int cpu)
347 {
348 	u64 tb = get_tb();
349 
350 	per_cpu(wd_timer_tb, cpu) = tb;
351 
352 	wd_smp_clear_cpu_pending(cpu);
353 
354 	if ((s64)(tb - wd_smp_last_reset_tb) >= (s64)wd_smp_panic_timeout_tb)
355 		watchdog_smp_panic(cpu);
356 
357 	if (__wd_nmi_output && xchg(&__wd_nmi_output, 0)) {
358 		/*
359 		 * Something has called printk from NMI context. It might be
360 		 * stuck, so this triggers a flush that will get that
361 		 * printk output to the console.
362 		 *
363 		 * See wd_lockup_ipi.
364 		 */
365 		printk_trigger_flush();
366 	}
367 }
368 
DEFINE_INTERRUPT_HANDLER_NMI(soft_nmi_interrupt)369 DEFINE_INTERRUPT_HANDLER_NMI(soft_nmi_interrupt)
370 {
371 	unsigned long flags;
372 	int cpu = raw_smp_processor_id();
373 	u64 tb;
374 
375 	/* should only arrive from kernel, with irqs disabled */
376 	WARN_ON_ONCE(!arch_irq_disabled_regs(regs));
377 
378 	if (!cpumask_test_cpu(cpu, &wd_cpus_enabled))
379 		return 0;
380 
381 	__this_cpu_inc(irq_stat.soft_nmi_irqs);
382 
383 	tb = get_tb();
384 	if (tb - per_cpu(wd_timer_tb, cpu) >= wd_panic_timeout_tb) {
385 		/*
386 		 * Taking wd_smp_lock here means it is a soft-NMI lock, which
387 		 * means we can't take any regular or irqsafe spin locks while
388 		 * holding this lock. This is why timers can't printk while
389 		 * holding the lock.
390 		 */
391 		wd_smp_lock(&flags);
392 		if (cpumask_test_cpu(cpu, &wd_smp_cpus_stuck)) {
393 			wd_smp_unlock(&flags);
394 			return 0;
395 		}
396 		if (!wd_try_report()) {
397 			wd_smp_unlock(&flags);
398 			/* Couldn't report, try again in 100ms */
399 			mtspr(SPRN_DEC, 100 * tb_ticks_per_usec * 1000);
400 			return 0;
401 		}
402 
403 		set_cpu_stuck(cpu);
404 
405 		wd_smp_unlock(&flags);
406 
407 		pr_emerg("CPU %d self-detected hard LOCKUP @ %pS\n",
408 			 cpu, (void *)regs->nip);
409 		pr_emerg("CPU %d TB:%lld, last heartbeat TB:%lld (%lldms ago)\n",
410 			 cpu, tb, per_cpu(wd_timer_tb, cpu),
411 			 tb_to_ns(tb - per_cpu(wd_timer_tb, cpu)) / 1000000);
412 		print_modules();
413 		print_irqtrace_events(current);
414 		show_regs(regs);
415 
416 		xchg(&__wd_nmi_output, 1); // see wd_lockup_ipi
417 
418 		if (sysctl_hardlockup_all_cpu_backtrace)
419 			trigger_allbutself_cpu_backtrace();
420 
421 		if (hardlockup_panic)
422 			nmi_panic(regs, "Hard LOCKUP");
423 
424 		wd_end_reporting();
425 	}
426 	/*
427 	 * We are okay to change DEC in soft_nmi_interrupt because the masked
428 	 * handler has marked a DEC as pending, so the timer interrupt will be
429 	 * replayed as soon as local irqs are enabled again.
430 	 */
431 	if (wd_panic_timeout_tb < 0x7fffffff)
432 		mtspr(SPRN_DEC, wd_panic_timeout_tb);
433 
434 	return 0;
435 }
436 
watchdog_timer_fn(struct hrtimer * hrtimer)437 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
438 {
439 	int cpu = smp_processor_id();
440 
441 	if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
442 		return HRTIMER_NORESTART;
443 
444 	if (!cpumask_test_cpu(cpu, &watchdog_cpumask))
445 		return HRTIMER_NORESTART;
446 
447 	watchdog_timer_interrupt(cpu);
448 
449 	hrtimer_forward_now(hrtimer, ms_to_ktime(wd_timer_period_ms));
450 
451 	return HRTIMER_RESTART;
452 }
453 
arch_touch_nmi_watchdog(void)454 void arch_touch_nmi_watchdog(void)
455 {
456 	unsigned long ticks = tb_ticks_per_usec * wd_timer_period_ms * 1000;
457 	int cpu = smp_processor_id();
458 	u64 tb;
459 
460 	if (!cpumask_test_cpu(cpu, &watchdog_cpumask))
461 		return;
462 
463 	tb = get_tb();
464 	if (tb - per_cpu(wd_timer_tb, cpu) >= ticks) {
465 		per_cpu(wd_timer_tb, cpu) = tb;
466 		wd_smp_clear_cpu_pending(cpu);
467 	}
468 }
469 EXPORT_SYMBOL(arch_touch_nmi_watchdog);
470 
start_watchdog(void * arg)471 static void start_watchdog(void *arg)
472 {
473 	struct hrtimer *hrtimer = this_cpu_ptr(&wd_hrtimer);
474 	int cpu = smp_processor_id();
475 	unsigned long flags;
476 
477 	if (cpumask_test_cpu(cpu, &wd_cpus_enabled)) {
478 		WARN_ON(1);
479 		return;
480 	}
481 
482 	if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
483 		return;
484 
485 	if (!cpumask_test_cpu(cpu, &watchdog_cpumask))
486 		return;
487 
488 	wd_smp_lock(&flags);
489 	cpumask_set_cpu(cpu, &wd_cpus_enabled);
490 	if (cpumask_weight(&wd_cpus_enabled) == 1) {
491 		cpumask_set_cpu(cpu, &wd_smp_cpus_pending);
492 		wd_smp_last_reset_tb = get_tb();
493 	}
494 	wd_smp_unlock(&flags);
495 
496 	*this_cpu_ptr(&wd_timer_tb) = get_tb();
497 
498 	hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
499 	hrtimer->function = watchdog_timer_fn;
500 	hrtimer_start(hrtimer, ms_to_ktime(wd_timer_period_ms),
501 		      HRTIMER_MODE_REL_PINNED);
502 }
503 
start_watchdog_on_cpu(unsigned int cpu)504 static int start_watchdog_on_cpu(unsigned int cpu)
505 {
506 	return smp_call_function_single(cpu, start_watchdog, NULL, true);
507 }
508 
stop_watchdog(void * arg)509 static void stop_watchdog(void *arg)
510 {
511 	struct hrtimer *hrtimer = this_cpu_ptr(&wd_hrtimer);
512 	int cpu = smp_processor_id();
513 	unsigned long flags;
514 
515 	if (!cpumask_test_cpu(cpu, &wd_cpus_enabled))
516 		return; /* Can happen in CPU unplug case */
517 
518 	hrtimer_cancel(hrtimer);
519 
520 	wd_smp_lock(&flags);
521 	cpumask_clear_cpu(cpu, &wd_cpus_enabled);
522 	wd_smp_unlock(&flags);
523 
524 	wd_smp_clear_cpu_pending(cpu);
525 }
526 
stop_watchdog_on_cpu(unsigned int cpu)527 static int stop_watchdog_on_cpu(unsigned int cpu)
528 {
529 	return smp_call_function_single(cpu, stop_watchdog, NULL, true);
530 }
531 
watchdog_calc_timeouts(void)532 static void watchdog_calc_timeouts(void)
533 {
534 	u64 threshold = watchdog_thresh;
535 
536 #ifdef CONFIG_PPC_PSERIES
537 	threshold += (READ_ONCE(wd_timeout_pct) * threshold) / 100;
538 #endif
539 
540 	wd_panic_timeout_tb = threshold * ppc_tb_freq;
541 
542 	/* Have the SMP detector trigger a bit later */
543 	wd_smp_panic_timeout_tb = wd_panic_timeout_tb * 3 / 2;
544 
545 	/* 2/5 is the factor that the perf based detector uses */
546 	wd_timer_period_ms = watchdog_thresh * 1000 * 2 / 5;
547 }
548 
watchdog_nmi_stop(void)549 void watchdog_nmi_stop(void)
550 {
551 	int cpu;
552 
553 	for_each_cpu(cpu, &wd_cpus_enabled)
554 		stop_watchdog_on_cpu(cpu);
555 }
556 
watchdog_nmi_start(void)557 void watchdog_nmi_start(void)
558 {
559 	int cpu;
560 
561 	watchdog_calc_timeouts();
562 	for_each_cpu_and(cpu, cpu_online_mask, &watchdog_cpumask)
563 		start_watchdog_on_cpu(cpu);
564 }
565 
566 /*
567  * Invoked from core watchdog init.
568  */
watchdog_nmi_probe(void)569 int __init watchdog_nmi_probe(void)
570 {
571 	int err;
572 
573 	err = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
574 					"powerpc/watchdog:online",
575 					start_watchdog_on_cpu,
576 					stop_watchdog_on_cpu);
577 	if (err < 0) {
578 		pr_warn("could not be initialized");
579 		return err;
580 	}
581 	return 0;
582 }
583 
584 #ifdef CONFIG_PPC_PSERIES
watchdog_nmi_set_timeout_pct(u64 pct)585 void watchdog_nmi_set_timeout_pct(u64 pct)
586 {
587 	pr_info("Set the NMI watchdog timeout factor to %llu%%\n", pct);
588 	WRITE_ONCE(wd_timeout_pct, pct);
589 	lockup_detector_reconfigure();
590 }
591 #endif
592