kernel/nmi_watchdog.c

   1 /*
   2  * Detect Hard Lockups using the NMI
   3  *
   4  * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc.
   5  *
   6  * this code detects hard lockups: incidents in where on a CPU
   7  * the kernel does not respond to anything except NMI.
   8  *
   9  * Note: Most of this code is borrowed heavily from softlockup.c,
  10  * so thanks to Ingo for the initial implementation.
  11  * Some chunks also taken from arch/x86/kernel/apic/nmi.c, thanks
  12  * to those contributors as well.
  13  */
  14
  15 #include <linux/mm.h>
  16 #include <linux/cpu.h>
  17 #include <linux/nmi.h>
  18 #include <linux/init.h>
  19 #include <linux/delay.h>
  20 #include <linux/freezer.h>
  21 #include <linux/lockdep.h>
  22 #include <linux/notifier.h>
  23 #include <linux/module.h>
  24 #include <linux/sysctl.h>
  25
  26 #include <asm/irq_regs.h>
  27 #include <linux/perf_event.h>
  28
  29 static DEFINE_PER_CPU(struct perf_event *, nmi_watchdog_ev);
  30 static DEFINE_PER_CPU(int, nmi_watchdog_touch);
  31 static DEFINE_PER_CPU(long, alert_counter);
  32
  33 static int panic_on_timeout;
  34
  35 void touch_nmi_watchdog(void)
  36 {
  37         __raw_get_cpu_var(nmi_watchdog_touch) = 1;
  38         touch_softlockup_watchdog();
  39 }
  40 EXPORT_SYMBOL(touch_nmi_watchdog);
  41
  42 void touch_all_nmi_watchdog(void)
  43 {
  44         int cpu;
  45
  46         for_each_online_cpu(cpu)
  47                 per_cpu(nmi_watchdog_touch, cpu) = 1;
  48         touch_softlockup_watchdog();
  49 }
  50
  51 static int __init setup_nmi_watchdog(char *str)
  52 {
  53         if (!strncmp(str, "panic", 5)) {
  54                 panic_on_timeout = 1;
  55                 str = strchr(str, ',');
  56                 if (!str)
  57                         return 1;
  58                 ++str;
  59         }
  60         return 1;
  61 }
  62 __setup("nmi_watchdog=", setup_nmi_watchdog);
  63
  64 struct perf_event_attr wd_attr = {
  65         .type = PERF_TYPE_HARDWARE,
  66         .config = PERF_COUNT_HW_CPU_CYCLES,
  67         .size = sizeof(struct perf_event_attr),
  68         .pinned = 1,
  69         .disabled = 1,
  70 };
  71
  72 void wd_overflow(struct perf_event *event, int nmi,
  73                  struct perf_sample_data *data,
  74                  struct pt_regs *regs)
  75 {
  76         int cpu = smp_processor_id();
  77         int touched = 0;
  78
  79         if (__get_cpu_var(nmi_watchdog_touch)) {
  80                 per_cpu(nmi_watchdog_touch, cpu) = 0;
  81                 touched = 1;
  82         }
  83
  84         /* check to see if the cpu is doing anything */
  85         if (!touched && hw_nmi_is_cpu_stuck(regs)) {
  86                 /*
  87                  * Ayiee, looks like this CPU is stuck ...
  88                  * wait a few IRQs (5 seconds) before doing the oops ...
  89                  */
  90                 per_cpu(alert_counter,cpu) += 1;
  91                 if (per_cpu(alert_counter,cpu) == 5) {
  92                         if (panic_on_timeout) {
  93                                 panic("NMI Watchdog detected LOCKUP on cpu %d", cpu);
  94                         } else {
  95                                 WARN(1, "NMI Watchdog detected LOCKUP on cpu %d", cpu);
  96                         }
  97                 }
  98         } else {
  99                 per_cpu(alert_counter,cpu) = 0;
 100         }
 101
 102         return;
 103 }
 104
 105 static int enable_nmi_watchdog(int cpu)
 106 {
 107         struct perf_event *event;
 108
 109         event = per_cpu(nmi_watchdog_ev, cpu);
 110         if (event && event->state > PERF_EVENT_STATE_OFF)
 111                 return 0;
 112
 113         if (event == NULL) {
 114                 /* Try to register using hardware perf events first */
 115                 wd_attr.sample_period = hw_nmi_get_sample_period();
 116                 event = perf_event_create_kernel_counter(&wd_attr, cpu, -1, wd_overflow);
 117                 if (IS_ERR(event)) {
 118                         wd_attr.type = PERF_TYPE_SOFTWARE;
 119                         event = perf_event_create_kernel_counter(&wd_attr, cpu, -1, wd_overflow);
 120                         if (IS_ERR(event)) {
 121                                 printk(KERN_ERR "nmi watchdog failed to create perf event on %i: %p\n", cpu, event);
 122                                 return -1;
 123                         }
 124                 }
 125                 per_cpu(nmi_watchdog_ev, cpu) = event;
 126         }
 127         perf_event_enable(per_cpu(nmi_watchdog_ev, cpu));
 128         return 0;
 129 }
 130
 131 static void disable_nmi_watchdog(int cpu)
 132 {
 133         struct perf_event *event;
 134
 135         event = per_cpu(nmi_watchdog_ev, cpu);
 136         if (event) {
 137                 perf_event_disable(per_cpu(nmi_watchdog_ev, cpu));
 138                 per_cpu(nmi_watchdog_ev, cpu) = NULL;
 139                 perf_event_release_kernel(event);
 140         }
 141 }
 142
 143 #ifdef CONFIG_SYSCTL
 144 /*
 145  * proc handler for /proc/sys/kernel/nmi_watchdog
 146  */
 147 int nmi_watchdog_enabled;
 148
 149 int proc_nmi_enabled(struct ctl_table *table, int write,
 150                      void __user *buffer, size_t *length, loff_t *ppos)
 151 {
 152         int cpu;
 153
 154         if (!write) {
 155                 struct perf_event *event;
 156                 for_each_online_cpu(cpu) {
 157                         event = per_cpu(nmi_watchdog_ev, cpu);
 158                         if (event && event->state > PERF_EVENT_STATE_OFF) {
 159                                 nmi_watchdog_enabled = 1;
 160                                 break;
 161                         }
 162                 }
 163                 proc_dointvec(table, write, buffer, length, ppos);
 164                 return 0;
 165         }
 166
 167         touch_all_nmi_watchdog();
 168         proc_dointvec(table, write, buffer, length, ppos);
 169         if (nmi_watchdog_enabled) {
 170                 for_each_online_cpu(cpu)
 171                         if (enable_nmi_watchdog(cpu)) {
 172                                 printk("NMI watchdog failed configuration, "
 173                                         " can not be enabled\n");
 174                         }
 175         } else {
 176                 for_each_online_cpu(cpu)
 177                         disable_nmi_watchdog(cpu);
 178         }
 179         return 0;
 180 }
 181
 182 #endif /* CONFIG_SYSCTL */
 183
 184 /*
 185  * Create/destroy watchdog threads as CPUs come and go:
 186  */
 187 static int __cpuinit
 188 cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
 189 {
 190         int hotcpu = (unsigned long)hcpu;
 191
 192         switch (action) {
 193         case CPU_UP_PREPARE:
 194         case CPU_UP_PREPARE_FROZEN:
 195                 per_cpu(nmi_watchdog_touch, hotcpu) = 0;
 196                 break;
 197         case CPU_ONLINE:
 198         case CPU_ONLINE_FROZEN:
 199                 if (enable_nmi_watchdog(hotcpu))
 200                         return NOTIFY_BAD;
 201                 break;
 202 #ifdef CONFIG_HOTPLUG_CPU
 203         case CPU_UP_CANCELED:
 204         case CPU_UP_CANCELED_FROZEN:
 205                 disable_nmi_watchdog(hotcpu);
 206         case CPU_DEAD:
 207         case CPU_DEAD_FROZEN:
 208                 break;
 209 #endif /* CONFIG_HOTPLUG_CPU */
 210         }
 211         return NOTIFY_OK;
 212 }
 213
 214 static struct notifier_block __cpuinitdata cpu_nfb = {
 215         .notifier_call = cpu_callback
 216 };
 217
 218 static int __initdata nonmi_watchdog;
 219
 220 static int __init nonmi_watchdog_setup(char *str)
 221 {
 222         nonmi_watchdog = 1;
 223         return 1;
 224 }
 225 __setup("nonmi_watchdog", nonmi_watchdog_setup);
 226
 227 static int __init spawn_nmi_watchdog_task(void)
 228 {
 229         void *cpu = (void *)(long)smp_processor_id();
 230         int err;
 231
 232         if (nonmi_watchdog)
 233                 return 0;
 234
 235         err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
 236         if (err == NOTIFY_BAD) {
 237                 BUG();
 238                 return 1;
 239         }
 240         cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
 241         register_cpu_notifier(&cpu_nfb);
 242
 243         return 0;
 244 }
 245 early_initcall(spawn_nmi_watchdog_task);