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 <asm/irq_regs.h>
17 #include <linux/perf_event.h>
19 static DEFINE_PER_CPU(bool, hard_watchdog_warn);
20 static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
21 static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
25 * Should we panic when a soft-lockup or hard-lockup occurs:
27 unsigned int __read_mostly hardlockup_panic =
28 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
29 static unsigned long hardlockup_allcpu_dumped;
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().
38 void hardlockup_detector_disable(void)
40 watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
43 static int __init hardlockup_panic_setup(char *str)
45 if (!strncmp(str, "panic", 5))
47 else if (!strncmp(str, "nopanic", 7))
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;
55 __setup("nmi_watchdog=", hardlockup_panic_setup);
57 void touch_nmi_watchdog(void)
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
66 raw_cpu_write(watchdog_nmi_touch, true);
67 touch_softlockup_watchdog();
69 EXPORT_SYMBOL(touch_nmi_watchdog);
71 static 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),
79 /* Callback function for perf event subsystem */
80 static void watchdog_overflow_callback(struct perf_event *event,
81 struct perf_sample_data *data,
84 /* Ensure the watchdog never gets throttled */
85 event->hw.interrupts = 0;
87 if (atomic_read(&watchdog_park_in_progress) != 0)
90 if (__this_cpu_read(watchdog_nmi_touch) == true) {
91 __this_cpu_write(watchdog_nmi_touch, false);
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
101 if (is_hardlockup()) {
102 int this_cpu = smp_processor_id();
104 /* only print hardlockups once */
105 if (__this_cpu_read(hard_watchdog_warn) == true)
108 pr_emerg("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
110 print_irqtrace_events(current);
117 * Perform all-CPU dump only once to avoid multiple hardlockups
118 * generating interleaving traces
120 if (sysctl_hardlockup_all_cpu_backtrace &&
121 !test_and_set_bit(0, &hardlockup_allcpu_dumped))
122 trigger_allbutself_cpu_backtrace();
124 if (hardlockup_panic)
125 nmi_panic(regs, "Hard LOCKUP");
127 __this_cpu_write(hard_watchdog_warn, true);
131 __this_cpu_write(hard_watchdog_warn, false);
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.
140 static unsigned long firstcpu_err;
141 static atomic_t watchdog_cpus;
143 int watchdog_nmi_enable(unsigned int cpu)
145 struct perf_event_attr *wd_attr;
146 struct perf_event *event = per_cpu(watchdog_ev, cpu);
149 /* nothing to do if the hard lockup detector is disabled */
150 if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
153 /* is it already setup and enabled? */
154 if (event && event->state > PERF_EVENT_STATE_OFF)
157 /* it is setup but not enabled */
161 if (atomic_inc_return(&watchdog_cpus) == 1)
164 wd_attr = &wd_hw_attr;
165 wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
167 /* Try to register using hardware perf events */
168 event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
170 /* save the first cpu's error for future comparision */
171 if (firstcpu && IS_ERR(event))
172 firstcpu_err = PTR_ERR(event);
174 if (!IS_ERR(event)) {
175 /* only print for the first cpu initialized */
176 if (firstcpu || firstcpu_err)
177 pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
182 * Disable the hard lockup detector if _any_ CPU fails to set up
183 * set up the hardware perf event. The watchdog() function checks
184 * the NMI_WATCHDOG_ENABLED bit periodically.
186 * The barriers are for syncing up watchdog_enabled across all the
187 * cpus, as clear_bit() does not use barriers.
189 smp_mb__before_atomic();
190 clear_bit(NMI_WATCHDOG_ENABLED_BIT, &watchdog_enabled);
191 smp_mb__after_atomic();
193 /* skip displaying the same error again */
194 if (!firstcpu && (PTR_ERR(event) == firstcpu_err))
195 return PTR_ERR(event);
197 /* vary the KERN level based on the returned errno */
198 if (PTR_ERR(event) == -EOPNOTSUPP)
199 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
200 else if (PTR_ERR(event) == -ENOENT)
201 pr_warn("disabled (cpu%i): hardware events not enabled\n",
204 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
205 cpu, PTR_ERR(event));
207 pr_info("Shutting down hard lockup detector on all cpus\n");
209 return PTR_ERR(event);
213 per_cpu(watchdog_ev, cpu) = event;
215 perf_event_enable(per_cpu(watchdog_ev, cpu));
220 void watchdog_nmi_disable(unsigned int cpu)
222 struct perf_event *event = per_cpu(watchdog_ev, cpu);
225 perf_event_disable(event);
226 per_cpu(watchdog_ev, cpu) = NULL;
228 /* should be in cleanup, but blocks oprofile */
229 perf_event_release_kernel(event);
231 /* watchdog_nmi_enable() expects this to be zero initially. */
232 if (atomic_dec_and_test(&watchdog_cpus))