xref: /openbmc/linux/arch/powerpc/kernel/watchdog.c (revision aa0dc6a7)
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 gobal "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 cpumask_t wd_smp_cpus_pending;
89 static cpumask_t wd_smp_cpus_stuck;
90 static u64 wd_smp_last_reset_tb;
91 
92 static inline void wd_smp_lock(unsigned long *flags)
93 {
94 	/*
95 	 * Avoid locking layers if possible.
96 	 * This may be called from low level interrupt handlers at some
97 	 * point in future.
98 	 */
99 	raw_local_irq_save(*flags);
100 	hard_irq_disable(); /* Make it soft-NMI safe */
101 	while (unlikely(test_and_set_bit_lock(0, &__wd_smp_lock))) {
102 		raw_local_irq_restore(*flags);
103 		spin_until_cond(!test_bit(0, &__wd_smp_lock));
104 		raw_local_irq_save(*flags);
105 		hard_irq_disable();
106 	}
107 }
108 
109 static inline void wd_smp_unlock(unsigned long *flags)
110 {
111 	clear_bit_unlock(0, &__wd_smp_lock);
112 	raw_local_irq_restore(*flags);
113 }
114 
115 static void wd_lockup_ipi(struct pt_regs *regs)
116 {
117 	int cpu = raw_smp_processor_id();
118 	u64 tb = get_tb();
119 
120 	pr_emerg("CPU %d Hard LOCKUP\n", cpu);
121 	pr_emerg("CPU %d TB:%lld, last heartbeat TB:%lld (%lldms ago)\n",
122 		 cpu, tb, per_cpu(wd_timer_tb, cpu),
123 		 tb_to_ns(tb - per_cpu(wd_timer_tb, cpu)) / 1000000);
124 	print_modules();
125 	print_irqtrace_events(current);
126 	if (regs)
127 		show_regs(regs);
128 	else
129 		dump_stack();
130 
131 	/* Do not panic from here because that can recurse into NMI IPI layer */
132 }
133 
134 static void set_cpumask_stuck(const struct cpumask *cpumask, u64 tb)
135 {
136 	cpumask_or(&wd_smp_cpus_stuck, &wd_smp_cpus_stuck, cpumask);
137 	cpumask_andnot(&wd_smp_cpus_pending, &wd_smp_cpus_pending, cpumask);
138 	if (cpumask_empty(&wd_smp_cpus_pending)) {
139 		wd_smp_last_reset_tb = tb;
140 		cpumask_andnot(&wd_smp_cpus_pending,
141 				&wd_cpus_enabled,
142 				&wd_smp_cpus_stuck);
143 	}
144 }
145 static void set_cpu_stuck(int cpu, u64 tb)
146 {
147 	set_cpumask_stuck(cpumask_of(cpu), tb);
148 }
149 
150 static void watchdog_smp_panic(int cpu, u64 tb)
151 {
152 	unsigned long flags;
153 	int c;
154 
155 	wd_smp_lock(&flags);
156 	/* Double check some things under lock */
157 	if ((s64)(tb - wd_smp_last_reset_tb) < (s64)wd_smp_panic_timeout_tb)
158 		goto out;
159 	if (cpumask_test_cpu(cpu, &wd_smp_cpus_pending))
160 		goto out;
161 	if (cpumask_weight(&wd_smp_cpus_pending) == 0)
162 		goto out;
163 
164 	pr_emerg("CPU %d detected hard LOCKUP on other CPUs %*pbl\n",
165 		 cpu, cpumask_pr_args(&wd_smp_cpus_pending));
166 	pr_emerg("CPU %d TB:%lld, last SMP heartbeat TB:%lld (%lldms ago)\n",
167 		 cpu, tb, wd_smp_last_reset_tb,
168 		 tb_to_ns(tb - wd_smp_last_reset_tb) / 1000000);
169 
170 	if (!sysctl_hardlockup_all_cpu_backtrace) {
171 		/*
172 		 * Try to trigger the stuck CPUs, unless we are going to
173 		 * get a backtrace on all of them anyway.
174 		 */
175 		for_each_cpu(c, &wd_smp_cpus_pending) {
176 			if (c == cpu)
177 				continue;
178 			smp_send_nmi_ipi(c, wd_lockup_ipi, 1000000);
179 		}
180 	}
181 
182 	/* Take the stuck CPUs out of the watch group */
183 	set_cpumask_stuck(&wd_smp_cpus_pending, tb);
184 
185 	wd_smp_unlock(&flags);
186 
187 	printk_safe_flush();
188 	/*
189 	 * printk_safe_flush() seems to require another print
190 	 * before anything actually goes out to console.
191 	 */
192 	if (sysctl_hardlockup_all_cpu_backtrace)
193 		trigger_allbutself_cpu_backtrace();
194 
195 	if (hardlockup_panic)
196 		nmi_panic(NULL, "Hard LOCKUP");
197 
198 	return;
199 
200 out:
201 	wd_smp_unlock(&flags);
202 }
203 
204 static void wd_smp_clear_cpu_pending(int cpu, u64 tb)
205 {
206 	if (!cpumask_test_cpu(cpu, &wd_smp_cpus_pending)) {
207 		if (unlikely(cpumask_test_cpu(cpu, &wd_smp_cpus_stuck))) {
208 			struct pt_regs *regs = get_irq_regs();
209 			unsigned long flags;
210 
211 			wd_smp_lock(&flags);
212 
213 			pr_emerg("CPU %d became unstuck TB:%lld\n",
214 				 cpu, tb);
215 			print_irqtrace_events(current);
216 			if (regs)
217 				show_regs(regs);
218 			else
219 				dump_stack();
220 
221 			cpumask_clear_cpu(cpu, &wd_smp_cpus_stuck);
222 			wd_smp_unlock(&flags);
223 		}
224 		return;
225 	}
226 	cpumask_clear_cpu(cpu, &wd_smp_cpus_pending);
227 	if (cpumask_empty(&wd_smp_cpus_pending)) {
228 		unsigned long flags;
229 
230 		wd_smp_lock(&flags);
231 		if (cpumask_empty(&wd_smp_cpus_pending)) {
232 			wd_smp_last_reset_tb = tb;
233 			cpumask_andnot(&wd_smp_cpus_pending,
234 					&wd_cpus_enabled,
235 					&wd_smp_cpus_stuck);
236 		}
237 		wd_smp_unlock(&flags);
238 	}
239 }
240 
241 static void watchdog_timer_interrupt(int cpu)
242 {
243 	u64 tb = get_tb();
244 
245 	per_cpu(wd_timer_tb, cpu) = tb;
246 
247 	wd_smp_clear_cpu_pending(cpu, tb);
248 
249 	if ((s64)(tb - wd_smp_last_reset_tb) >= (s64)wd_smp_panic_timeout_tb)
250 		watchdog_smp_panic(cpu, tb);
251 }
252 
253 DEFINE_INTERRUPT_HANDLER_NMI(soft_nmi_interrupt)
254 {
255 	unsigned long flags;
256 	int cpu = raw_smp_processor_id();
257 	u64 tb;
258 
259 	/* should only arrive from kernel, with irqs disabled */
260 	WARN_ON_ONCE(!arch_irq_disabled_regs(regs));
261 
262 	if (!cpumask_test_cpu(cpu, &wd_cpus_enabled))
263 		return 0;
264 
265 	__this_cpu_inc(irq_stat.soft_nmi_irqs);
266 
267 	tb = get_tb();
268 	if (tb - per_cpu(wd_timer_tb, cpu) >= wd_panic_timeout_tb) {
269 		wd_smp_lock(&flags);
270 		if (cpumask_test_cpu(cpu, &wd_smp_cpus_stuck)) {
271 			wd_smp_unlock(&flags);
272 			return 0;
273 		}
274 		set_cpu_stuck(cpu, tb);
275 
276 		pr_emerg("CPU %d self-detected hard LOCKUP @ %pS\n",
277 			 cpu, (void *)regs->nip);
278 		pr_emerg("CPU %d TB:%lld, last heartbeat TB:%lld (%lldms ago)\n",
279 			 cpu, tb, per_cpu(wd_timer_tb, cpu),
280 			 tb_to_ns(tb - per_cpu(wd_timer_tb, cpu)) / 1000000);
281 		print_modules();
282 		print_irqtrace_events(current);
283 		show_regs(regs);
284 
285 		wd_smp_unlock(&flags);
286 
287 		if (sysctl_hardlockup_all_cpu_backtrace)
288 			trigger_allbutself_cpu_backtrace();
289 
290 		if (hardlockup_panic)
291 			nmi_panic(regs, "Hard LOCKUP");
292 	}
293 	if (wd_panic_timeout_tb < 0x7fffffff)
294 		mtspr(SPRN_DEC, wd_panic_timeout_tb);
295 
296 	return 0;
297 }
298 
299 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
300 {
301 	int cpu = smp_processor_id();
302 
303 	if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
304 		return HRTIMER_NORESTART;
305 
306 	if (!cpumask_test_cpu(cpu, &watchdog_cpumask))
307 		return HRTIMER_NORESTART;
308 
309 	watchdog_timer_interrupt(cpu);
310 
311 	hrtimer_forward_now(hrtimer, ms_to_ktime(wd_timer_period_ms));
312 
313 	return HRTIMER_RESTART;
314 }
315 
316 void arch_touch_nmi_watchdog(void)
317 {
318 	unsigned long ticks = tb_ticks_per_usec * wd_timer_period_ms * 1000;
319 	int cpu = smp_processor_id();
320 	u64 tb = get_tb();
321 
322 	if (tb - per_cpu(wd_timer_tb, cpu) >= ticks) {
323 		per_cpu(wd_timer_tb, cpu) = tb;
324 		wd_smp_clear_cpu_pending(cpu, tb);
325 	}
326 }
327 EXPORT_SYMBOL(arch_touch_nmi_watchdog);
328 
329 static void start_watchdog(void *arg)
330 {
331 	struct hrtimer *hrtimer = this_cpu_ptr(&wd_hrtimer);
332 	int cpu = smp_processor_id();
333 	unsigned long flags;
334 
335 	if (cpumask_test_cpu(cpu, &wd_cpus_enabled)) {
336 		WARN_ON(1);
337 		return;
338 	}
339 
340 	if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
341 		return;
342 
343 	if (!cpumask_test_cpu(cpu, &watchdog_cpumask))
344 		return;
345 
346 	wd_smp_lock(&flags);
347 	cpumask_set_cpu(cpu, &wd_cpus_enabled);
348 	if (cpumask_weight(&wd_cpus_enabled) == 1) {
349 		cpumask_set_cpu(cpu, &wd_smp_cpus_pending);
350 		wd_smp_last_reset_tb = get_tb();
351 	}
352 	wd_smp_unlock(&flags);
353 
354 	*this_cpu_ptr(&wd_timer_tb) = get_tb();
355 
356 	hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
357 	hrtimer->function = watchdog_timer_fn;
358 	hrtimer_start(hrtimer, ms_to_ktime(wd_timer_period_ms),
359 		      HRTIMER_MODE_REL_PINNED);
360 }
361 
362 static int start_watchdog_on_cpu(unsigned int cpu)
363 {
364 	return smp_call_function_single(cpu, start_watchdog, NULL, true);
365 }
366 
367 static void stop_watchdog(void *arg)
368 {
369 	struct hrtimer *hrtimer = this_cpu_ptr(&wd_hrtimer);
370 	int cpu = smp_processor_id();
371 	unsigned long flags;
372 
373 	if (!cpumask_test_cpu(cpu, &wd_cpus_enabled))
374 		return; /* Can happen in CPU unplug case */
375 
376 	hrtimer_cancel(hrtimer);
377 
378 	wd_smp_lock(&flags);
379 	cpumask_clear_cpu(cpu, &wd_cpus_enabled);
380 	wd_smp_unlock(&flags);
381 
382 	wd_smp_clear_cpu_pending(cpu, get_tb());
383 }
384 
385 static int stop_watchdog_on_cpu(unsigned int cpu)
386 {
387 	return smp_call_function_single(cpu, stop_watchdog, NULL, true);
388 }
389 
390 static void watchdog_calc_timeouts(void)
391 {
392 	wd_panic_timeout_tb = watchdog_thresh * ppc_tb_freq;
393 
394 	/* Have the SMP detector trigger a bit later */
395 	wd_smp_panic_timeout_tb = wd_panic_timeout_tb * 3 / 2;
396 
397 	/* 2/5 is the factor that the perf based detector uses */
398 	wd_timer_period_ms = watchdog_thresh * 1000 * 2 / 5;
399 }
400 
401 void watchdog_nmi_stop(void)
402 {
403 	int cpu;
404 
405 	for_each_cpu(cpu, &wd_cpus_enabled)
406 		stop_watchdog_on_cpu(cpu);
407 }
408 
409 void watchdog_nmi_start(void)
410 {
411 	int cpu;
412 
413 	watchdog_calc_timeouts();
414 	for_each_cpu_and(cpu, cpu_online_mask, &watchdog_cpumask)
415 		start_watchdog_on_cpu(cpu);
416 }
417 
418 /*
419  * Invoked from core watchdog init.
420  */
421 int __init watchdog_nmi_probe(void)
422 {
423 	int err;
424 
425 	err = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
426 					"powerpc/watchdog:online",
427 					start_watchdog_on_cpu,
428 					stop_watchdog_on_cpu);
429 	if (err < 0) {
430 		pr_warn("could not be initialized");
431 		return err;
432 	}
433 	return 0;
434 }
435