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 /*
  38  * Returns seconds, approximately.  We don't need nanosecond
  39  * resolution, and we don't need to waste time with a big divide when
  40  * 2^30ns == 1.074s.
  41  */
  42 static unsigned long get_timestamp(void)
  43 {
  44         return sched_clock() >> 30;  /* 2^30 ~= 10^9 */
  45 }
  46
  47 void touch_softlockup_watchdog(void)
  48 {
  49         __raw_get_cpu_var(touch_timestamp) = get_timestamp();
  50 }
  51 EXPORT_SYMBOL(touch_softlockup_watchdog);
  52
  53 /*
  54  * This callback runs from the timer interrupt, and checks
  55  * whether the watchdog thread has hung or not:
  56  */
  57 void softlockup_tick(void)
  58 {
  59         int this_cpu = smp_processor_id();
  60         unsigned long touch_timestamp = per_cpu(touch_timestamp, this_cpu);
  61         unsigned long print_timestamp;
  62         unsigned long now;
  63
  64         /* watchdog task hasn't updated timestamp yet */
  65         if (touch_timestamp == 0)
  66                 return;
  67
  68         print_timestamp = per_cpu(print_timestamp, this_cpu);
  69
  70         /* report at most once a second */
  71         if (print_timestamp < (touch_timestamp + 1) ||
  72                 did_panic ||
  73                         !per_cpu(watchdog_task, this_cpu))
  74                 return;
  75
  76         /* do not print during early bootup: */
  77         if (unlikely(system_state != SYSTEM_RUNNING)) {
  78                 touch_softlockup_watchdog();
  79                 return;
  80         }
  81
  82         now = get_timestamp();
  83
  84         /* Wake up the high-prio watchdog task every second: */
  85         if (now > (touch_timestamp + 1))
  86                 wake_up_process(per_cpu(watchdog_task, this_cpu));
  87
  88         /* Warn about unreasonable 10+ seconds delays: */
  89         if (now > (touch_timestamp + 10)) {
  90                 per_cpu(print_timestamp, this_cpu) = touch_timestamp;
  91
  92                 spin_lock(&print_lock);
  93                 printk(KERN_ERR "BUG: soft lockup detected on CPU#%d!\n",
  94                         this_cpu);
  95                 dump_stack();
  96                 spin_unlock(&print_lock);
  97         }
  98 }
  99
 100 /*
 101  * The watchdog thread - runs every second and touches the timestamp.
 102  */
 103 static int watchdog(void * __bind_cpu)
 104 {
 105         struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
 106
 107         sched_setscheduler(current, SCHED_FIFO, &param);
 108         current->flags |= PF_NOFREEZE;
 109
 110         /* initialize timestamp */
 111         touch_softlockup_watchdog();
 112
 113         /*
 114          * Run briefly once per second to reset the softlockup timestamp.
 115          * If this gets delayed for more than 10 seconds then the
 116          * debug-printout triggers in softlockup_tick().
 117          */
 118         while (!kthread_should_stop()) {
 119                 set_current_state(TASK_INTERRUPTIBLE);
 120                 touch_softlockup_watchdog();
 121                 schedule();
 122         }
 123
 124         return 0;
 125 }
 126
 127 /*
 128  * Create/destroy watchdog threads as CPUs come and go:
 129  */
 130 static int __cpuinit
 131 cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
 132 {
 133         int hotcpu = (unsigned long)hcpu;
 134         struct task_struct *p;
 135
 136         switch (action) {
 137         case CPU_UP_PREPARE:
 138                 BUG_ON(per_cpu(watchdog_task, hotcpu));
 139                 p = kthread_create(watchdog, hcpu, "watchdog/%d", hotcpu);
 140                 if (IS_ERR(p)) {
 141                         printk("watchdog for %i failed\n", hotcpu);
 142                         return NOTIFY_BAD;
 143                 }
 144                 per_cpu(touch_timestamp, hotcpu) = 0;
 145                 per_cpu(watchdog_task, hotcpu) = p;
 146                 kthread_bind(p, hotcpu);
 147                 break;
 148         case CPU_ONLINE:
 149                 wake_up_process(per_cpu(watchdog_task, hotcpu));
 150                 break;
 151 #ifdef CONFIG_HOTPLUG_CPU
 152         case CPU_UP_CANCELED:
 153                 if (!per_cpu(watchdog_task, hotcpu))
 154                         break;
 155                 /* Unbind so it can run.  Fall thru. */
 156                 kthread_bind(per_cpu(watchdog_task, hotcpu),
 157                              any_online_cpu(cpu_online_map));
 158         case CPU_DEAD:
 159                 p = per_cpu(watchdog_task, hotcpu);
 160                 per_cpu(watchdog_task, hotcpu) = NULL;
 161                 kthread_stop(p);
 162                 break;
 163 #endif /* CONFIG_HOTPLUG_CPU */
 164         }
 165         return NOTIFY_OK;
 166 }
 167
 168 static struct notifier_block __cpuinitdata cpu_nfb = {
 169         .notifier_call = cpu_callback
 170 };
 171
 172 __init void spawn_softlockup_task(void)
 173 {
 174         void *cpu = (void *)(long)smp_processor_id();
 175         int err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
 176
 177         BUG_ON(err == NOTIFY_BAD);
 178         cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
 179         register_cpu_notifier(&cpu_nfb);
 180
 181         atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
 182 }