1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Detect hard lockups on a system
4  *
5  * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc.
6  *
7  * Note: Most of this code is borrowed heavily from the original softlockup
8  * detector, so thanks to Ingo for the initial implementation.
9  * Some chunks also taken from the old x86-specific nmi watchdog code, thanks
10  * to those contributors as well.
11  */
12 
13 #define pr_fmt(fmt) "NMI watchdog: " fmt
14 
15 #include <linux/nmi.h>
16 #include <linux/atomic.h>
17 #include <linux/module.h>
18 #include <linux/sched/debug.h>
19 
20 #include <asm/irq_regs.h>
21 #include <linux/perf_event.h>
22 
23 static DEFINE_PER_CPU(bool, hard_watchdog_warn);
24 static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
25 static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
26 static DEFINE_PER_CPU(struct perf_event *, dead_event);
27 static struct cpumask dead_events_mask;
28 
29 static unsigned long hardlockup_allcpu_dumped;
30 static atomic_t watchdog_cpus = ATOMIC_INIT(0);
31 
arch_touch_nmi_watchdog(void)32 notrace void arch_touch_nmi_watchdog(void)
33 {
34 	/*
35 	 * Using __raw here because some code paths have
36 	 * preemption enabled.  If preemption is enabled
37 	 * then interrupts should be enabled too, in which
38 	 * case we shouldn't have to worry about the watchdog
39 	 * going off.
40 	 */
41 	raw_cpu_write(watchdog_nmi_touch, true);
42 }
43 EXPORT_SYMBOL(arch_touch_nmi_watchdog);
44 
45 #ifdef CONFIG_HARDLOCKUP_CHECK_TIMESTAMP
46 static DEFINE_PER_CPU(ktime_t, last_timestamp);
47 static DEFINE_PER_CPU(unsigned int, nmi_rearmed);
48 static ktime_t watchdog_hrtimer_sample_threshold __read_mostly;
49 
watchdog_update_hrtimer_threshold(u64 period)50 void watchdog_update_hrtimer_threshold(u64 period)
51 {
52 	/*
53 	 * The hrtimer runs with a period of (watchdog_threshold * 2) / 5
54 	 *
55 	 * So it runs effectively with 2.5 times the rate of the NMI
56 	 * watchdog. That means the hrtimer should fire 2-3 times before
57 	 * the NMI watchdog expires. The NMI watchdog on x86 is based on
58 	 * unhalted CPU cycles, so if Turbo-Mode is enabled the CPU cycles
59 	 * might run way faster than expected and the NMI fires in a
60 	 * smaller period than the one deduced from the nominal CPU
61 	 * frequency. Depending on the Turbo-Mode factor this might be fast
62 	 * enough to get the NMI period smaller than the hrtimer watchdog
63 	 * period and trigger false positives.
64 	 *
65 	 * The sample threshold is used to check in the NMI handler whether
66 	 * the minimum time between two NMI samples has elapsed. That
67 	 * prevents false positives.
68 	 *
69 	 * Set this to 4/5 of the actual watchdog threshold period so the
70 	 * hrtimer is guaranteed to fire at least once within the real
71 	 * watchdog threshold.
72 	 */
73 	watchdog_hrtimer_sample_threshold = period * 2;
74 }
75 
watchdog_check_timestamp(void)76 static bool watchdog_check_timestamp(void)
77 {
78 	ktime_t delta, now = ktime_get_mono_fast_ns();
79 
80 	delta = now - __this_cpu_read(last_timestamp);
81 	if (delta < watchdog_hrtimer_sample_threshold) {
82 		/*
83 		 * If ktime is jiffies based, a stalled timer would prevent
84 		 * jiffies from being incremented and the filter would look
85 		 * at a stale timestamp and never trigger.
86 		 */
87 		if (__this_cpu_inc_return(nmi_rearmed) < 10)
88 			return false;
89 	}
90 	__this_cpu_write(nmi_rearmed, 0);
91 	__this_cpu_write(last_timestamp, now);
92 	return true;
93 }
94 #else
watchdog_check_timestamp(void)95 static inline bool watchdog_check_timestamp(void)
96 {
97 	return true;
98 }
99 #endif
100 
101 static struct perf_event_attr wd_hw_attr = {
102 	.type		= PERF_TYPE_HARDWARE,
103 	.config		= PERF_COUNT_HW_CPU_CYCLES,
104 	.size		= sizeof(struct perf_event_attr),
105 	.pinned		= 1,
106 	.disabled	= 1,
107 };
108 
109 /* Callback function for perf event subsystem */
watchdog_overflow_callback(struct perf_event * event,struct perf_sample_data * data,struct pt_regs * regs)110 static void watchdog_overflow_callback(struct perf_event *event,
111 				       struct perf_sample_data *data,
112 				       struct pt_regs *regs)
113 {
114 	/* Ensure the watchdog never gets throttled */
115 	event->hw.interrupts = 0;
116 
117 	if (__this_cpu_read(watchdog_nmi_touch) == true) {
118 		__this_cpu_write(watchdog_nmi_touch, false);
119 		return;
120 	}
121 
122 	if (!watchdog_check_timestamp())
123 		return;
124 
125 	/* check for a hardlockup
126 	 * This is done by making sure our timer interrupt
127 	 * is incrementing.  The timer interrupt should have
128 	 * fired multiple times before we overflow'd.  If it hasn't
129 	 * then this is a good indication the cpu is stuck
130 	 */
131 	if (is_hardlockup()) {
132 		int this_cpu = smp_processor_id();
133 
134 		/* only print hardlockups once */
135 		if (__this_cpu_read(hard_watchdog_warn) == true)
136 			return;
137 
138 		pr_emerg("Watchdog detected hard LOCKUP on cpu %d\n",
139 			 this_cpu);
140 		print_modules();
141 		print_irqtrace_events(current);
142 		if (regs)
143 			show_regs(regs);
144 		else
145 			dump_stack();
146 
147 		/*
148 		 * Perform all-CPU dump only once to avoid multiple hardlockups
149 		 * generating interleaving traces
150 		 */
151 		if (sysctl_hardlockup_all_cpu_backtrace &&
152 				!test_and_set_bit(0, &hardlockup_allcpu_dumped))
153 			trigger_allbutself_cpu_backtrace();
154 
155 		if (hardlockup_panic)
156 			nmi_panic(regs, "Hard LOCKUP");
157 
158 		__this_cpu_write(hard_watchdog_warn, true);
159 		return;
160 	}
161 
162 	__this_cpu_write(hard_watchdog_warn, false);
163 	return;
164 }
165 
hardlockup_detector_event_create(void)166 static int hardlockup_detector_event_create(void)
167 {
168 	unsigned int cpu = smp_processor_id();
169 	struct perf_event_attr *wd_attr;
170 	struct perf_event *evt;
171 
172 	wd_attr = &wd_hw_attr;
173 	wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
174 
175 	/* Try to register using hardware perf events */
176 	evt = perf_event_create_kernel_counter(wd_attr, cpu, NULL,
177 					       watchdog_overflow_callback, NULL);
178 	if (IS_ERR(evt)) {
179 		pr_debug("Perf event create on CPU %d failed with %ld\n", cpu,
180 			 PTR_ERR(evt));
181 		return PTR_ERR(evt);
182 	}
183 	this_cpu_write(watchdog_ev, evt);
184 	return 0;
185 }
186 
187 /**
188  * hardlockup_detector_perf_enable - Enable the local event
189  */
hardlockup_detector_perf_enable(void)190 void hardlockup_detector_perf_enable(void)
191 {
192 	if (hardlockup_detector_event_create())
193 		return;
194 
195 	/* use original value for check */
196 	if (!atomic_fetch_inc(&watchdog_cpus))
197 		pr_info("Enabled. Permanently consumes one hw-PMU counter.\n");
198 
199 	perf_event_enable(this_cpu_read(watchdog_ev));
200 }
201 
202 /**
203  * hardlockup_detector_perf_disable - Disable the local event
204  */
hardlockup_detector_perf_disable(void)205 void hardlockup_detector_perf_disable(void)
206 {
207 	struct perf_event *event = this_cpu_read(watchdog_ev);
208 
209 	if (event) {
210 		perf_event_disable(event);
211 		this_cpu_write(watchdog_ev, NULL);
212 		this_cpu_write(dead_event, event);
213 		cpumask_set_cpu(smp_processor_id(), &dead_events_mask);
214 		atomic_dec(&watchdog_cpus);
215 	}
216 }
217 
218 /**
219  * hardlockup_detector_perf_cleanup - Cleanup disabled events and destroy them
220  *
221  * Called from lockup_detector_cleanup(). Serialized by the caller.
222  */
hardlockup_detector_perf_cleanup(void)223 void hardlockup_detector_perf_cleanup(void)
224 {
225 	int cpu;
226 
227 	for_each_cpu(cpu, &dead_events_mask) {
228 		struct perf_event *event = per_cpu(dead_event, cpu);
229 
230 		/*
231 		 * Required because for_each_cpu() reports  unconditionally
232 		 * CPU0 as set on UP kernels. Sigh.
233 		 */
234 		if (event)
235 			perf_event_release_kernel(event);
236 		per_cpu(dead_event, cpu) = NULL;
237 	}
238 	cpumask_clear(&dead_events_mask);
239 }
240 
241 /**
242  * hardlockup_detector_perf_stop - Globally stop watchdog events
243  *
244  * Special interface for x86 to handle the perf HT bug.
245  */
hardlockup_detector_perf_stop(void)246 void __init hardlockup_detector_perf_stop(void)
247 {
248 	int cpu;
249 
250 	lockdep_assert_cpus_held();
251 
252 	for_each_online_cpu(cpu) {
253 		struct perf_event *event = per_cpu(watchdog_ev, cpu);
254 
255 		if (event)
256 			perf_event_disable(event);
257 	}
258 }
259 
260 /**
261  * hardlockup_detector_perf_restart - Globally restart watchdog events
262  *
263  * Special interface for x86 to handle the perf HT bug.
264  */
hardlockup_detector_perf_restart(void)265 void __init hardlockup_detector_perf_restart(void)
266 {
267 	int cpu;
268 
269 	lockdep_assert_cpus_held();
270 
271 	if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
272 		return;
273 
274 	for_each_online_cpu(cpu) {
275 		struct perf_event *event = per_cpu(watchdog_ev, cpu);
276 
277 		if (event)
278 			perf_event_enable(event);
279 	}
280 }
281 
282 /**
283  * hardlockup_detector_perf_init - Probe whether NMI event is available at all
284  */
hardlockup_detector_perf_init(void)285 int __init hardlockup_detector_perf_init(void)
286 {
287 	int ret = hardlockup_detector_event_create();
288 
289 	if (ret) {
290 		pr_info("Perf NMI watchdog permanently disabled\n");
291 	} else {
292 		perf_event_release_kernel(this_cpu_read(watchdog_ev));
293 		this_cpu_write(watchdog_ev, NULL);
294 	}
295 	return ret;
296 }
297