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
23static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
24
25static atomic_t watchdog_cpus = ATOMIC_INIT(0);
26
27#ifdef CONFIG_HARDLOCKUP_CHECK_TIMESTAMP
28static DEFINE_PER_CPU(ktime_t, last_timestamp);
29static DEFINE_PER_CPU(unsigned int, nmi_rearmed);
30static ktime_t watchdog_hrtimer_sample_threshold __read_mostly;
31
32void watchdog_update_hrtimer_threshold(u64 period)
33{
34 /*
35 * The hrtimer runs with a period of (watchdog_threshold * 2) / 5
36 *
37 * So it runs effectively with 2.5 times the rate of the NMI
38 * watchdog. That means the hrtimer should fire 2-3 times before
39 * the NMI watchdog expires. The NMI watchdog on x86 is based on
40 * unhalted CPU cycles, so if Turbo-Mode is enabled the CPU cycles
41 * might run way faster than expected and the NMI fires in a
42 * smaller period than the one deduced from the nominal CPU
43 * frequency. Depending on the Turbo-Mode factor this might be fast
44 * enough to get the NMI period smaller than the hrtimer watchdog
45 * period and trigger false positives.
46 *
47 * The sample threshold is used to check in the NMI handler whether
48 * the minimum time between two NMI samples has elapsed. That
49 * prevents false positives.
50 *
51 * Set this to 4/5 of the actual watchdog threshold period so the
52 * hrtimer is guaranteed to fire at least once within the real
53 * watchdog threshold.
54 */
55 watchdog_hrtimer_sample_threshold = period * 2;
56}
57
58static bool watchdog_check_timestamp(void)
59{
60 ktime_t delta, now = ktime_get_mono_fast_ns();
61
62 delta = now - __this_cpu_read(last_timestamp);
63 if (delta < watchdog_hrtimer_sample_threshold) {
64 /*
65 * If ktime is jiffies based, a stalled timer would prevent
66 * jiffies from being incremented and the filter would look
67 * at a stale timestamp and never trigger.
68 */
69 if (__this_cpu_inc_return(nmi_rearmed) < 10)
70 return false;
71 }
72 __this_cpu_write(nmi_rearmed, 0);
73 __this_cpu_write(last_timestamp, now);
74 return true;
75}
76
77static void watchdog_init_timestamp(void)
78{
79 __this_cpu_write(nmi_rearmed, 0);
80 __this_cpu_write(last_timestamp, ktime_get_mono_fast_ns());
81}
82#else
83static inline bool watchdog_check_timestamp(void) { return true; }
84static inline void watchdog_init_timestamp(void) { }
85#endif
86
87static struct perf_event_attr wd_hw_attr = {
88 .type = PERF_TYPE_HARDWARE,
89 .config = PERF_COUNT_HW_CPU_CYCLES,
90 .size = sizeof(struct perf_event_attr),
91 .pinned = 1,
92 .disabled = 1,
93};
94
95static struct perf_event_attr fallback_wd_hw_attr = {
96 .type = PERF_TYPE_HARDWARE,
97 .config = PERF_COUNT_HW_CPU_CYCLES,
98 .size = sizeof(struct perf_event_attr),
99 .pinned = 1,
100 .disabled = 1,
101};
102
103/* Callback function for perf event subsystem */
104static void watchdog_overflow_callback(struct perf_event *event,
105 struct perf_sample_data *data,
106 struct pt_regs *regs)
107{
108 /* Ensure the watchdog never gets throttled */
109 event->hw.interrupts = 0;
110
111 if (!watchdog_check_timestamp())
112 return;
113
114 watchdog_hardlockup_check(smp_processor_id(), regs);
115}
116
117static int hardlockup_detector_event_create(void)
118{
119 unsigned int cpu;
120 struct perf_event_attr *wd_attr;
121 struct perf_event *evt;
122
123 /*
124 * Preemption is not disabled because memory will be allocated.
125 * Ensure CPU-locality by calling this in per-CPU kthread.
126 */
127 WARN_ON(!is_percpu_thread());
128 cpu = raw_smp_processor_id();
129 wd_attr = &wd_hw_attr;
130 wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
131
132 /* Try to register using hardware perf events */
133 evt = perf_event_create_kernel_counter(attr: wd_attr, cpu, NULL,
134 callback: watchdog_overflow_callback, NULL);
135 if (IS_ERR(ptr: evt)) {
136 wd_attr = &fallback_wd_hw_attr;
137 wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
138 evt = perf_event_create_kernel_counter(attr: wd_attr, cpu, NULL,
139 callback: watchdog_overflow_callback, NULL);
140 }
141
142 if (IS_ERR(ptr: evt)) {
143 pr_debug("Perf event create on CPU %d failed with %ld\n", cpu,
144 PTR_ERR(evt));
145 return PTR_ERR(ptr: evt);
146 }
147 WARN_ONCE(this_cpu_read(watchdog_ev), "unexpected watchdog_ev leak");
148 this_cpu_write(watchdog_ev, evt);
149 return 0;
150}
151
152/**
153 * watchdog_hardlockup_enable - Enable the local event
154 * @cpu: The CPU to enable hard lockup on.
155 */
156void watchdog_hardlockup_enable(unsigned int cpu)
157{
158 WARN_ON_ONCE(cpu != smp_processor_id());
159
160 if (hardlockup_detector_event_create())
161 return;
162
163 /* use original value for check */
164 if (!atomic_fetch_inc(v: &watchdog_cpus))
165 pr_info("Enabled. Permanently consumes one hw-PMU counter.\n");
166
167 watchdog_init_timestamp();
168 perf_event_enable(this_cpu_read(watchdog_ev));
169}
170
171/**
172 * watchdog_hardlockup_disable - Disable the local event
173 * @cpu: The CPU to enable hard lockup on.
174 */
175void watchdog_hardlockup_disable(unsigned int cpu)
176{
177 struct perf_event *event = this_cpu_read(watchdog_ev);
178
179 WARN_ON_ONCE(cpu != smp_processor_id());
180
181 if (event) {
182 perf_event_disable(event);
183 perf_event_release_kernel(event);
184 this_cpu_write(watchdog_ev, NULL);
185 atomic_dec(v: &watchdog_cpus);
186 }
187}
188
189/**
190 * hardlockup_detector_perf_stop - Globally stop watchdog events
191 *
192 * Special interface for x86 to handle the perf HT bug.
193 */
194void __init hardlockup_detector_perf_stop(void)
195{
196 int cpu;
197
198 lockdep_assert_cpus_held();
199
200 for_each_online_cpu(cpu) {
201 struct perf_event *event = per_cpu(watchdog_ev, cpu);
202
203 if (event)
204 perf_event_disable(event);
205 }
206}
207
208/**
209 * hardlockup_detector_perf_restart - Globally restart watchdog events
210 *
211 * Special interface for x86 to handle the perf HT bug.
212 */
213void __init hardlockup_detector_perf_restart(void)
214{
215 int cpu;
216
217 lockdep_assert_cpus_held();
218
219 if (!(watchdog_enabled & WATCHDOG_HARDLOCKUP_ENABLED))
220 return;
221
222 for_each_online_cpu(cpu) {
223 struct perf_event *event = per_cpu(watchdog_ev, cpu);
224
225 if (event)
226 perf_event_enable(event);
227 }
228}
229
230bool __weak __init arch_perf_nmi_is_available(void)
231{
232 return true;
233}
234
235/**
236 * watchdog_hardlockup_probe - Probe whether NMI event is available at all
237 */
238int __init watchdog_hardlockup_probe(void)
239{
240 int ret;
241
242 if (!arch_perf_nmi_is_available())
243 return -ENODEV;
244
245 ret = hardlockup_detector_event_create();
246
247 if (ret) {
248 pr_info("Perf NMI watchdog permanently disabled\n");
249 } else {
250 perf_event_release_kernel(this_cpu_read(watchdog_ev));
251 this_cpu_write(watchdog_ev, NULL);
252 }
253 return ret;
254}
255
256/**
257 * hardlockup_config_perf_event - Overwrite config of wd_hw_attr.
258 * @str: number which identifies the raw perf event to use
259 */
260void __init hardlockup_config_perf_event(const char *str)
261{
262 u64 config;
263 char buf[24];
264 char *comma = strchr(str, ',');
265
266 if (!comma) {
267 if (kstrtoull(s: str, base: 16, res: &config))
268 return;
269 } else {
270 unsigned int len = comma - str;
271
272 if (len > sizeof(buf))
273 return;
274
275 strscpy(buf, str, len);
276 if (kstrtoull(s: buf, base: 16, res: &config))
277 return;
278 }
279
280 wd_hw_attr.type = PERF_TYPE_RAW;
281 wd_hw_attr.config = config;
282}
283

source code of linux/kernel/watchdog_perf.c