Loading...
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
23static DEFINE_PER_CPU(bool, hard_watchdog_warn);
24static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
25static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
26static DEFINE_PER_CPU(struct perf_event *, dead_event);
27static struct cpumask dead_events_mask;
28
29static unsigned long hardlockup_allcpu_dumped;
30static atomic_t watchdog_cpus = ATOMIC_INIT(0);
31
32notrace 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}
43EXPORT_SYMBOL(arch_touch_nmi_watchdog);
44
45#ifdef CONFIG_HARDLOCKUP_CHECK_TIMESTAMP
46static DEFINE_PER_CPU(ktime_t, last_timestamp);
47static DEFINE_PER_CPU(unsigned int, nmi_rearmed);
48static ktime_t watchdog_hrtimer_sample_threshold __read_mostly;
49
50void 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
76static 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
95static inline bool watchdog_check_timestamp(void)
96{
97 return true;
98}
99#endif
100
101static 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 */
110static 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
166static 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 */
190void 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 */
205void 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 */
223void 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 */
246void __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 */
265void __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 */
285int __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}
1/*
2 * Detect hard lockups on a system
3 *
4 * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc.
5 *
6 * Note: Most of this code is borrowed heavily from the original softlockup
7 * detector, so thanks to Ingo for the initial implementation.
8 * Some chunks also taken from the old x86-specific nmi watchdog code, thanks
9 * to those contributors as well.
10 */
11
12#define pr_fmt(fmt) "NMI watchdog: " fmt
13
14#include <linux/nmi.h>
15#include <linux/module.h>
16#include <asm/irq_regs.h>
17#include <linux/perf_event.h>
18
19static DEFINE_PER_CPU(bool, hard_watchdog_warn);
20static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
21static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
22
23/* boot commands */
24/*
25 * Should we panic when a soft-lockup or hard-lockup occurs:
26 */
27unsigned int __read_mostly hardlockup_panic =
28 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
29static unsigned long hardlockup_allcpu_dumped;
30/*
31 * We may not want to enable hard lockup detection by default in all cases,
32 * for example when running the kernel as a guest on a hypervisor. In these
33 * cases this function can be called to disable hard lockup detection. This
34 * function should only be executed once by the boot processor before the
35 * kernel command line parameters are parsed, because otherwise it is not
36 * possible to override this in hardlockup_panic_setup().
37 */
38void hardlockup_detector_disable(void)
39{
40 watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
41}
42
43static int __init hardlockup_panic_setup(char *str)
44{
45 if (!strncmp(str, "panic", 5))
46 hardlockup_panic = 1;
47 else if (!strncmp(str, "nopanic", 7))
48 hardlockup_panic = 0;
49 else if (!strncmp(str, "0", 1))
50 watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
51 else if (!strncmp(str, "1", 1))
52 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
53 return 1;
54}
55__setup("nmi_watchdog=", hardlockup_panic_setup);
56
57void touch_nmi_watchdog(void)
58{
59 /*
60 * Using __raw here because some code paths have
61 * preemption enabled. If preemption is enabled
62 * then interrupts should be enabled too, in which
63 * case we shouldn't have to worry about the watchdog
64 * going off.
65 */
66 raw_cpu_write(watchdog_nmi_touch, true);
67 touch_softlockup_watchdog();
68}
69EXPORT_SYMBOL(touch_nmi_watchdog);
70
71static struct perf_event_attr wd_hw_attr = {
72 .type = PERF_TYPE_HARDWARE,
73 .config = PERF_COUNT_HW_CPU_CYCLES,
74 .size = sizeof(struct perf_event_attr),
75 .pinned = 1,
76 .disabled = 1,
77};
78
79/* Callback function for perf event subsystem */
80static void watchdog_overflow_callback(struct perf_event *event,
81 struct perf_sample_data *data,
82 struct pt_regs *regs)
83{
84 /* Ensure the watchdog never gets throttled */
85 event->hw.interrupts = 0;
86
87 if (atomic_read(&watchdog_park_in_progress) != 0)
88 return;
89
90 if (__this_cpu_read(watchdog_nmi_touch) == true) {
91 __this_cpu_write(watchdog_nmi_touch, false);
92 return;
93 }
94
95 /* check for a hardlockup
96 * This is done by making sure our timer interrupt
97 * is incrementing. The timer interrupt should have
98 * fired multiple times before we overflow'd. If it hasn't
99 * then this is a good indication the cpu is stuck
100 */
101 if (is_hardlockup()) {
102 int this_cpu = smp_processor_id();
103
104 /* only print hardlockups once */
105 if (__this_cpu_read(hard_watchdog_warn) == true)
106 return;
107
108 pr_emerg("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
109 print_modules();
110 print_irqtrace_events(current);
111 if (regs)
112 show_regs(regs);
113 else
114 dump_stack();
115
116 /*
117 * Perform all-CPU dump only once to avoid multiple hardlockups
118 * generating interleaving traces
119 */
120 if (sysctl_hardlockup_all_cpu_backtrace &&
121 !test_and_set_bit(0, &hardlockup_allcpu_dumped))
122 trigger_allbutself_cpu_backtrace();
123
124 if (hardlockup_panic)
125 nmi_panic(regs, "Hard LOCKUP");
126
127 __this_cpu_write(hard_watchdog_warn, true);
128 return;
129 }
130
131 __this_cpu_write(hard_watchdog_warn, false);
132 return;
133}
134
135/*
136 * People like the simple clean cpu node info on boot.
137 * Reduce the watchdog noise by only printing messages
138 * that are different from what cpu0 displayed.
139 */
140static unsigned long cpu0_err;
141
142int watchdog_nmi_enable(unsigned int cpu)
143{
144 struct perf_event_attr *wd_attr;
145 struct perf_event *event = per_cpu(watchdog_ev, cpu);
146
147 /* nothing to do if the hard lockup detector is disabled */
148 if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
149 goto out;
150
151 /* is it already setup and enabled? */
152 if (event && event->state > PERF_EVENT_STATE_OFF)
153 goto out;
154
155 /* it is setup but not enabled */
156 if (event != NULL)
157 goto out_enable;
158
159 wd_attr = &wd_hw_attr;
160 wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
161
162 /* Try to register using hardware perf events */
163 event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
164
165 /* save cpu0 error for future comparision */
166 if (cpu == 0 && IS_ERR(event))
167 cpu0_err = PTR_ERR(event);
168
169 if (!IS_ERR(event)) {
170 /* only print for cpu0 or different than cpu0 */
171 if (cpu == 0 || cpu0_err)
172 pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
173 goto out_save;
174 }
175
176 /*
177 * Disable the hard lockup detector if _any_ CPU fails to set up
178 * set up the hardware perf event. The watchdog() function checks
179 * the NMI_WATCHDOG_ENABLED bit periodically.
180 *
181 * The barriers are for syncing up watchdog_enabled across all the
182 * cpus, as clear_bit() does not use barriers.
183 */
184 smp_mb__before_atomic();
185 clear_bit(NMI_WATCHDOG_ENABLED_BIT, &watchdog_enabled);
186 smp_mb__after_atomic();
187
188 /* skip displaying the same error again */
189 if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
190 return PTR_ERR(event);
191
192 /* vary the KERN level based on the returned errno */
193 if (PTR_ERR(event) == -EOPNOTSUPP)
194 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
195 else if (PTR_ERR(event) == -ENOENT)
196 pr_warn("disabled (cpu%i): hardware events not enabled\n",
197 cpu);
198 else
199 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
200 cpu, PTR_ERR(event));
201
202 pr_info("Shutting down hard lockup detector on all cpus\n");
203
204 return PTR_ERR(event);
205
206 /* success path */
207out_save:
208 per_cpu(watchdog_ev, cpu) = event;
209out_enable:
210 perf_event_enable(per_cpu(watchdog_ev, cpu));
211out:
212 return 0;
213}
214
215void watchdog_nmi_disable(unsigned int cpu)
216{
217 struct perf_event *event = per_cpu(watchdog_ev, cpu);
218
219 if (event) {
220 perf_event_disable(event);
221 per_cpu(watchdog_ev, cpu) = NULL;
222
223 /* should be in cleanup, but blocks oprofile */
224 perf_event_release_kernel(event);
225 }
226 if (cpu == 0) {
227 /* watchdog_nmi_enable() expects this to be zero initially. */
228 cpu0_err = 0;
229 }
230}