2 * Detect hard lockups on a system
4 * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc.
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.
12 #define pr_fmt(fmt) "NMI watchdog: " fmt
14 #include <linux/nmi.h>
15 #include <linux/module.h>
16 #include <linux/sched/debug.h>
18 #include <asm/irq_regs.h>
19 #include <linux/perf_event.h>
21 static DEFINE_PER_CPU(bool, hard_watchdog_warn
);
22 static DEFINE_PER_CPU(bool, watchdog_nmi_touch
);
23 static DEFINE_PER_CPU(struct perf_event
*, watchdog_ev
);
24 static struct cpumask dead_events_mask
;
26 static unsigned long hardlockup_allcpu_dumped
;
27 static unsigned int watchdog_cpus
;
29 void arch_touch_nmi_watchdog(void)
32 * Using __raw here because some code paths have
33 * preemption enabled. If preemption is enabled
34 * then interrupts should be enabled too, in which
35 * case we shouldn't have to worry about the watchdog
38 raw_cpu_write(watchdog_nmi_touch
, true);
40 EXPORT_SYMBOL(arch_touch_nmi_watchdog
);
42 #ifdef CONFIG_HARDLOCKUP_CHECK_TIMESTAMP
43 static DEFINE_PER_CPU(ktime_t
, last_timestamp
);
44 static DEFINE_PER_CPU(unsigned int, nmi_rearmed
);
45 static ktime_t watchdog_hrtimer_sample_threshold __read_mostly
;
47 void watchdog_update_hrtimer_threshold(u64 period
)
50 * The hrtimer runs with a period of (watchdog_threshold * 2) / 5
52 * So it runs effectively with 2.5 times the rate of the NMI
53 * watchdog. That means the hrtimer should fire 2-3 times before
54 * the NMI watchdog expires. The NMI watchdog on x86 is based on
55 * unhalted CPU cycles, so if Turbo-Mode is enabled the CPU cycles
56 * might run way faster than expected and the NMI fires in a
57 * smaller period than the one deduced from the nominal CPU
58 * frequency. Depending on the Turbo-Mode factor this might be fast
59 * enough to get the NMI period smaller than the hrtimer watchdog
60 * period and trigger false positives.
62 * The sample threshold is used to check in the NMI handler whether
63 * the minimum time between two NMI samples has elapsed. That
64 * prevents false positives.
66 * Set this to 4/5 of the actual watchdog threshold period so the
67 * hrtimer is guaranteed to fire at least once within the real
70 watchdog_hrtimer_sample_threshold
= period
* 2;
73 static bool watchdog_check_timestamp(void)
75 ktime_t delta
, now
= ktime_get_mono_fast_ns();
77 delta
= now
- __this_cpu_read(last_timestamp
);
78 if (delta
< watchdog_hrtimer_sample_threshold
) {
80 * If ktime is jiffies based, a stalled timer would prevent
81 * jiffies from being incremented and the filter would look
82 * at a stale timestamp and never trigger.
84 if (__this_cpu_inc_return(nmi_rearmed
) < 10)
87 __this_cpu_write(nmi_rearmed
, 0);
88 __this_cpu_write(last_timestamp
, now
);
92 static inline bool watchdog_check_timestamp(void)
98 static struct perf_event_attr wd_hw_attr
= {
99 .type
= PERF_TYPE_HARDWARE
,
100 .config
= PERF_COUNT_HW_CPU_CYCLES
,
101 .size
= sizeof(struct perf_event_attr
),
106 /* Callback function for perf event subsystem */
107 static void watchdog_overflow_callback(struct perf_event
*event
,
108 struct perf_sample_data
*data
,
109 struct pt_regs
*regs
)
111 /* Ensure the watchdog never gets throttled */
112 event
->hw
.interrupts
= 0;
114 if (__this_cpu_read(watchdog_nmi_touch
) == true) {
115 __this_cpu_write(watchdog_nmi_touch
, false);
119 if (!watchdog_check_timestamp())
122 /* check for a hardlockup
123 * This is done by making sure our timer interrupt
124 * is incrementing. The timer interrupt should have
125 * fired multiple times before we overflow'd. If it hasn't
126 * then this is a good indication the cpu is stuck
128 if (is_hardlockup()) {
129 int this_cpu
= smp_processor_id();
131 /* only print hardlockups once */
132 if (__this_cpu_read(hard_watchdog_warn
) == true)
135 pr_emerg("Watchdog detected hard LOCKUP on cpu %d", this_cpu
);
137 print_irqtrace_events(current
);
144 * Perform all-CPU dump only once to avoid multiple hardlockups
145 * generating interleaving traces
147 if (sysctl_hardlockup_all_cpu_backtrace
&&
148 !test_and_set_bit(0, &hardlockup_allcpu_dumped
))
149 trigger_allbutself_cpu_backtrace();
151 if (hardlockup_panic
)
152 nmi_panic(regs
, "Hard LOCKUP");
154 __this_cpu_write(hard_watchdog_warn
, true);
158 __this_cpu_write(hard_watchdog_warn
, false);
162 static int hardlockup_detector_event_create(void)
164 unsigned int cpu
= smp_processor_id();
165 struct perf_event_attr
*wd_attr
;
166 struct perf_event
*evt
;
168 wd_attr
= &wd_hw_attr
;
169 wd_attr
->sample_period
= hw_nmi_get_sample_period(watchdog_thresh
);
171 /* Try to register using hardware perf events */
172 evt
= perf_event_create_kernel_counter(wd_attr
, cpu
, NULL
,
173 watchdog_overflow_callback
, NULL
);
175 pr_info("Perf event create on CPU %d failed with %ld\n", cpu
,
179 this_cpu_write(watchdog_ev
, evt
);
184 * hardlockup_detector_perf_enable - Enable the local event
186 void hardlockup_detector_perf_enable(void)
188 if (hardlockup_detector_event_create())
191 if (!watchdog_cpus
++)
192 pr_info("Enabled. Permanently consumes one hw-PMU counter.\n");
194 perf_event_enable(this_cpu_read(watchdog_ev
));
198 * hardlockup_detector_perf_disable - Disable the local event
200 void hardlockup_detector_perf_disable(void)
202 struct perf_event
*event
= this_cpu_read(watchdog_ev
);
205 perf_event_disable(event
);
206 cpumask_set_cpu(smp_processor_id(), &dead_events_mask
);
212 * hardlockup_detector_perf_cleanup - Cleanup disabled events and destroy them
214 * Called from lockup_detector_cleanup(). Serialized by the caller.
216 void hardlockup_detector_perf_cleanup(void)
220 for_each_cpu(cpu
, &dead_events_mask
) {
221 struct perf_event
*event
= per_cpu(watchdog_ev
, cpu
);
224 * Required because for_each_cpu() reports unconditionally
225 * CPU0 as set on UP kernels. Sigh.
228 perf_event_release_kernel(event
);
229 per_cpu(watchdog_ev
, cpu
) = NULL
;
231 cpumask_clear(&dead_events_mask
);
235 * hardlockup_detector_perf_stop - Globally stop watchdog events
237 * Special interface for x86 to handle the perf HT bug.
239 void __init
hardlockup_detector_perf_stop(void)
243 lockdep_assert_cpus_held();
245 for_each_online_cpu(cpu
) {
246 struct perf_event
*event
= per_cpu(watchdog_ev
, cpu
);
249 perf_event_disable(event
);
254 * hardlockup_detector_perf_restart - Globally restart watchdog events
256 * Special interface for x86 to handle the perf HT bug.
258 void __init
hardlockup_detector_perf_restart(void)
262 lockdep_assert_cpus_held();
264 if (!(watchdog_enabled
& NMI_WATCHDOG_ENABLED
))
267 for_each_online_cpu(cpu
) {
268 struct perf_event
*event
= per_cpu(watchdog_ev
, cpu
);
271 perf_event_enable(event
);
276 * hardlockup_detector_perf_init - Probe whether NMI event is available at all
278 int __init
hardlockup_detector_perf_init(void)
280 int ret
= hardlockup_detector_event_create();
283 pr_info("Perf NMI watchdog permanently disabled\n");
285 perf_event_release_kernel(this_cpu_read(watchdog_ev
));
286 this_cpu_write(watchdog_ev
, NULL
);