1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Detect hard lockups on a system using perf
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(struct perf_event *, watchdog_ev);
24 static DEFINE_PER_CPU(struct perf_event *, dead_event);
25 static struct cpumask dead_events_mask;
26 
27 static atomic_t watchdog_cpus = ATOMIC_INIT(0);
28 
29 #ifdef CONFIG_HARDLOCKUP_CHECK_TIMESTAMP
30 static DEFINE_PER_CPU(ktime_t, last_timestamp);
31 static DEFINE_PER_CPU(unsigned int, nmi_rearmed);
32 static ktime_t watchdog_hrtimer_sample_threshold __read_mostly;
33 
watchdog_update_hrtimer_threshold(u64 period)34 void watchdog_update_hrtimer_threshold(u64 period)
35 {
36 	/*
37 	 * The hrtimer runs with a period of (watchdog_threshold * 2) / 5
38 	 *
39 	 * So it runs effectively with 2.5 times the rate of the NMI
40 	 * watchdog. That means the hrtimer should fire 2-3 times before
41 	 * the NMI watchdog expires. The NMI watchdog on x86 is based on
42 	 * unhalted CPU cycles, so if Turbo-Mode is enabled the CPU cycles
43 	 * might run way faster than expected and the NMI fires in a
44 	 * smaller period than the one deduced from the nominal CPU
45 	 * frequency. Depending on the Turbo-Mode factor this might be fast
46 	 * enough to get the NMI period smaller than the hrtimer watchdog
47 	 * period and trigger false positives.
48 	 *
49 	 * The sample threshold is used to check in the NMI handler whether
50 	 * the minimum time between two NMI samples has elapsed. That
51 	 * prevents false positives.
52 	 *
53 	 * Set this to 4/5 of the actual watchdog threshold period so the
54 	 * hrtimer is guaranteed to fire at least once within the real
55 	 * watchdog threshold.
56 	 */
57 	watchdog_hrtimer_sample_threshold = period * 2;
58 }
59 
watchdog_check_timestamp(void)60 static bool watchdog_check_timestamp(void)
61 {
62 	ktime_t delta, now = ktime_get_mono_fast_ns();
63 
64 	delta = now - __this_cpu_read(last_timestamp);
65 	if (delta < watchdog_hrtimer_sample_threshold) {
66 		/*
67 		 * If ktime is jiffies based, a stalled timer would prevent
68 		 * jiffies from being incremented and the filter would look
69 		 * at a stale timestamp and never trigger.
70 		 */
71 		if (__this_cpu_inc_return(nmi_rearmed) < 10)
72 			return false;
73 	}
74 	__this_cpu_write(nmi_rearmed, 0);
75 	__this_cpu_write(last_timestamp, now);
76 	return true;
77 }
78 #else
watchdog_check_timestamp(void)79 static inline bool watchdog_check_timestamp(void)
80 {
81 	return true;
82 }
83 #endif
84 
85 static struct perf_event_attr wd_hw_attr = {
86 	.type		= PERF_TYPE_HARDWARE,
87 	.config		= PERF_COUNT_HW_CPU_CYCLES,
88 	.size		= sizeof(struct perf_event_attr),
89 	.pinned		= 1,
90 	.disabled	= 1,
91 };
92 
93 /* Callback function for perf event subsystem */
watchdog_overflow_callback(struct perf_event * event,struct perf_sample_data * data,struct pt_regs * regs)94 static void watchdog_overflow_callback(struct perf_event *event,
95 				       struct perf_sample_data *data,
96 				       struct pt_regs *regs)
97 {
98 	/* Ensure the watchdog never gets throttled */
99 	event->hw.interrupts = 0;
100 
101 	if (!watchdog_check_timestamp())
102 		return;
103 
104 	watchdog_hardlockup_check(smp_processor_id(), regs);
105 }
106 
hardlockup_detector_event_create(void)107 static int hardlockup_detector_event_create(void)
108 {
109 	unsigned int cpu;
110 	struct perf_event_attr *wd_attr;
111 	struct perf_event *evt;
112 
113 	/*
114 	 * Preemption is not disabled because memory will be allocated.
115 	 * Ensure CPU-locality by calling this in per-CPU kthread.
116 	 */
117 	WARN_ON(!is_percpu_thread());
118 	cpu = raw_smp_processor_id();
119 	wd_attr = &wd_hw_attr;
120 	wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
121 
122 	/* Try to register using hardware perf events */
123 	evt = perf_event_create_kernel_counter(wd_attr, cpu, NULL,
124 					       watchdog_overflow_callback, NULL);
125 	if (IS_ERR(evt)) {
126 		pr_debug("Perf event create on CPU %d failed with %ld\n", cpu,
127 			 PTR_ERR(evt));
128 		return PTR_ERR(evt);
129 	}
130 	this_cpu_write(watchdog_ev, evt);
131 	return 0;
132 }
133 
134 /**
135  * watchdog_hardlockup_enable - Enable the local event
136  *
137  * @cpu: The CPU to enable hard lockup on.
138  */
watchdog_hardlockup_enable(unsigned int cpu)139 void watchdog_hardlockup_enable(unsigned int cpu)
140 {
141 	WARN_ON_ONCE(cpu != smp_processor_id());
142 
143 	if (hardlockup_detector_event_create())
144 		return;
145 
146 	/* use original value for check */
147 	if (!atomic_fetch_inc(&watchdog_cpus))
148 		pr_info("Enabled. Permanently consumes one hw-PMU counter.\n");
149 
150 	perf_event_enable(this_cpu_read(watchdog_ev));
151 }
152 
153 /**
154  * watchdog_hardlockup_disable - Disable the local event
155  *
156  * @cpu: The CPU to enable hard lockup on.
157  */
watchdog_hardlockup_disable(unsigned int cpu)158 void watchdog_hardlockup_disable(unsigned int cpu)
159 {
160 	struct perf_event *event = this_cpu_read(watchdog_ev);
161 
162 	WARN_ON_ONCE(cpu != smp_processor_id());
163 
164 	if (event) {
165 		perf_event_disable(event);
166 		this_cpu_write(watchdog_ev, NULL);
167 		this_cpu_write(dead_event, event);
168 		cpumask_set_cpu(smp_processor_id(), &dead_events_mask);
169 		atomic_dec(&watchdog_cpus);
170 	}
171 }
172 
173 /**
174  * hardlockup_detector_perf_cleanup - Cleanup disabled events and destroy them
175  *
176  * Called from lockup_detector_cleanup(). Serialized by the caller.
177  */
hardlockup_detector_perf_cleanup(void)178 void hardlockup_detector_perf_cleanup(void)
179 {
180 	int cpu;
181 
182 	for_each_cpu(cpu, &dead_events_mask) {
183 		struct perf_event *event = per_cpu(dead_event, cpu);
184 
185 		/*
186 		 * Required because for_each_cpu() reports  unconditionally
187 		 * CPU0 as set on UP kernels. Sigh.
188 		 */
189 		if (event)
190 			perf_event_release_kernel(event);
191 		per_cpu(dead_event, cpu) = NULL;
192 	}
193 	cpumask_clear(&dead_events_mask);
194 }
195 
196 /**
197  * hardlockup_detector_perf_stop - Globally stop watchdog events
198  *
199  * Special interface for x86 to handle the perf HT bug.
200  */
hardlockup_detector_perf_stop(void)201 void __init hardlockup_detector_perf_stop(void)
202 {
203 	int cpu;
204 
205 	lockdep_assert_cpus_held();
206 
207 	for_each_online_cpu(cpu) {
208 		struct perf_event *event = per_cpu(watchdog_ev, cpu);
209 
210 		if (event)
211 			perf_event_disable(event);
212 	}
213 }
214 
215 /**
216  * hardlockup_detector_perf_restart - Globally restart watchdog events
217  *
218  * Special interface for x86 to handle the perf HT bug.
219  */
hardlockup_detector_perf_restart(void)220 void __init hardlockup_detector_perf_restart(void)
221 {
222 	int cpu;
223 
224 	lockdep_assert_cpus_held();
225 
226 	if (!(watchdog_enabled & WATCHDOG_HARDLOCKUP_ENABLED))
227 		return;
228 
229 	for_each_online_cpu(cpu) {
230 		struct perf_event *event = per_cpu(watchdog_ev, cpu);
231 
232 		if (event)
233 			perf_event_enable(event);
234 	}
235 }
236 
arch_perf_nmi_is_available(void)237 bool __weak __init arch_perf_nmi_is_available(void)
238 {
239 	return true;
240 }
241 
242 /**
243  * watchdog_hardlockup_probe - Probe whether NMI event is available at all
244  */
watchdog_hardlockup_probe(void)245 int __init watchdog_hardlockup_probe(void)
246 {
247 	int ret;
248 
249 	if (!arch_perf_nmi_is_available())
250 		return -ENODEV;
251 
252 	ret = hardlockup_detector_event_create();
253 
254 	if (ret) {
255 		pr_info("Perf NMI watchdog permanently disabled\n");
256 	} else {
257 		perf_event_release_kernel(this_cpu_read(watchdog_ev));
258 		this_cpu_write(watchdog_ev, NULL);
259 	}
260 	return ret;
261 }
262