kernel/softlockup.c

   1 /*
   2  * Detect Soft Lockups
   3  *
   4  * started by Ingo Molnar, Copyright (C) 2005, 2006 Red Hat, Inc.
   5  *
   6  * this code detects soft lockups: incidents in where on a CPU
   7  * the kernel does not reschedule for 10 seconds or more.
   8  */
   9 #include <linux/mm.h>
  10 #include <linux/cpu.h>
  11 #include <linux/init.h>
  12 #include <linux/delay.h>
  13 #include <linux/kthread.h>
  14 #include <linux/notifier.h>
  15 #include <linux/module.h>
  16
  17 static DEFINE_SPINLOCK(print_lock);
  18
  19 static DEFINE_PER_CPU(unsigned long, touch_timestamp);
  20 static DEFINE_PER_CPU(unsigned long, print_timestamp);
  21 static DEFINE_PER_CPU(struct task_struct *, watchdog_task);
  22
  23 static int did_panic = 0;
  24
  25 static int
  26 softlock_panic(struct notifier_block *this, unsigned long event, void *ptr)
  27 {
  28         did_panic = 1;
  29
  30         return NOTIFY_DONE;
  31 }
  32
  33 static struct notifier_block panic_block = {
  34         .notifier_call = softlock_panic,
  35 };
  36
  37 void touch_softlockup_watchdog(void)
  38 {
  39         per_cpu(touch_timestamp, raw_smp_processor_id()) = jiffies;
  40 }
  41 EXPORT_SYMBOL(touch_softlockup_watchdog);
  42
  43 /*
  44  * This callback runs from the timer interrupt, and checks
  45  * whether the watchdog thread has hung or not:
  46  */
  47 void softlockup_tick(void)
  48 {
  49         int this_cpu = smp_processor_id();
  50         unsigned long touch_timestamp = per_cpu(touch_timestamp, this_cpu);
  51
  52         /* prevent double reports: */
  53         if (per_cpu(print_timestamp, this_cpu) == touch_timestamp ||
  54                 did_panic ||
  55                         !per_cpu(watchdog_task, this_cpu))
  56                 return;
  57
  58         /* do not print during early bootup: */
  59         if (unlikely(system_state != SYSTEM_RUNNING)) {
  60                 touch_softlockup_watchdog();
  61                 return;
  62         }
  63
  64         /* Wake up the high-prio watchdog task every second: */
  65         if (time_after(jiffies, touch_timestamp + HZ))
  66                 wake_up_process(per_cpu(watchdog_task, this_cpu));
  67
  68         /* Warn about unreasonable 10+ seconds delays: */
  69         if (time_after(jiffies, touch_timestamp + 10*HZ)) {
  70                 per_cpu(print_timestamp, this_cpu) = touch_timestamp;
  71
  72                 spin_lock(&print_lock);
  73                 printk(KERN_ERR "BUG: soft lockup detected on CPU#%d!\n",
  74                         this_cpu);
  75                 dump_stack();
  76                 spin_unlock(&print_lock);
  77         }
  78 }
  79
  80 /*
  81  * The watchdog thread - runs every second and touches the timestamp.
  82  */
  83 static int watchdog(void * __bind_cpu)
  84 {
  85         struct sched_param param = { .sched_priority = 99 };
  86
  87         sched_setscheduler(current, SCHED_FIFO, &param);
  88         current->flags |= PF_NOFREEZE;
  89
  90         /*
  91          * Run briefly once per second to reset the softlockup timestamp.
  92          * If this gets delayed for more than 10 seconds then the
  93          * debug-printout triggers in softlockup_tick().
  94          */
  95         while (!kthread_should_stop()) {
  96                 set_current_state(TASK_INTERRUPTIBLE);
  97                 touch_softlockup_watchdog();
  98                 schedule();
  99         }
 100
 101         return 0;
 102 }
 103
 104 /*
 105  * Create/destroy watchdog threads as CPUs come and go:
 106  */
 107 static int __devinit
 108 cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
 109 {
 110         int hotcpu = (unsigned long)hcpu;
 111         struct task_struct *p;
 112
 113         switch (action) {
 114         case CPU_UP_PREPARE:
 115                 BUG_ON(per_cpu(watchdog_task, hotcpu));
 116                 p = kthread_create(watchdog, hcpu, "watchdog/%d", hotcpu);
 117                 if (IS_ERR(p)) {
 118                         printk("watchdog for %i failed\n", hotcpu);
 119                         return NOTIFY_BAD;
 120                 }
 121                 per_cpu(touch_timestamp, hotcpu) = jiffies;
 122                 per_cpu(watchdog_task, hotcpu) = p;
 123                 kthread_bind(p, hotcpu);
 124                 break;
 125         case CPU_ONLINE:
 126                 wake_up_process(per_cpu(watchdog_task, hotcpu));
 127                 break;
 128 #ifdef CONFIG_HOTPLUG_CPU
 129         case CPU_UP_CANCELED:
 130                 /* Unbind so it can run.  Fall thru. */
 131                 kthread_bind(per_cpu(watchdog_task, hotcpu),
 132                              any_online_cpu(cpu_online_map));
 133         case CPU_DEAD:
 134                 p = per_cpu(watchdog_task, hotcpu);
 135                 per_cpu(watchdog_task, hotcpu) = NULL;
 136                 kthread_stop(p);
 137                 break;
 138 #endif /* CONFIG_HOTPLUG_CPU */
 139         }
 140         return NOTIFY_OK;
 141 }
 142
 143 static struct notifier_block __devinitdata cpu_nfb = {
 144         .notifier_call = cpu_callback
 145 };
 146
 147 __init void spawn_softlockup_task(void)
 148 {
 149         void *cpu = (void *)(long)smp_processor_id();
 150
 151         cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
 152         cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
 153         register_cpu_notifier(&cpu_nfb);
 154
 155         atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
 156 }