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/nmi.h>
  12 #include <linux/init.h>
  13 #include <linux/delay.h>
  14 #include <linux/freezer.h>
  15 #include <linux/kthread.h>
  16 #include <linux/notifier.h>
  17 #include <linux/module.h>
  18
  19 #include <asm/irq_regs.h>
  20
  21 static DEFINE_SPINLOCK(print_lock);
  22
  23 static DEFINE_PER_CPU(unsigned long, touch_timestamp);
  24 static DEFINE_PER_CPU(unsigned long, print_timestamp);
  25 static DEFINE_PER_CPU(struct task_struct *, watchdog_task);
  26
  27 static int __read_mostly did_panic;
  28 unsigned long __read_mostly softlockup_thresh = 60;
  29
  30 static int
  31 softlock_panic(struct notifier_block *this, unsigned long event, void *ptr)
  32 {
  33         did_panic = 1;
  34
  35         return NOTIFY_DONE;
  36 }
  37
  38 static struct notifier_block panic_block = {
  39         .notifier_call = softlock_panic,
  40 };
  41
  42 /*
  43  * Returns seconds, approximately.  We don't need nanosecond
  44  * resolution, and we don't need to waste time with a big divide when
  45  * 2^30ns == 1.074s.
  46  */
  47 static unsigned long get_timestamp(int this_cpu)
  48 {
  49         return cpu_clock(this_cpu) >> 30LL;  /* 2^30 ~= 10^9 */
  50 }
  51
  52 static void __touch_softlockup_watchdog(void)
  53 {
  54         int this_cpu = raw_smp_processor_id();
  55
  56         __raw_get_cpu_var(touch_timestamp) = get_timestamp(this_cpu);
  57 }
  58
  59 void touch_softlockup_watchdog(void)
  60 {
  61         __raw_get_cpu_var(touch_timestamp) = 0;
  62 }
  63 EXPORT_SYMBOL(touch_softlockup_watchdog);
  64
  65 void touch_all_softlockup_watchdogs(void)
  66 {
  67         int cpu;
  68
  69         /* Cause each CPU to re-update its timestamp rather than complain */
  70         for_each_online_cpu(cpu)
  71                 per_cpu(touch_timestamp, cpu) = 0;
  72 }
  73 EXPORT_SYMBOL(touch_all_softlockup_watchdogs);
  74
  75 /*
  76  * This callback runs from the timer interrupt, and checks
  77  * whether the watchdog thread has hung or not:
  78  */
  79 void softlockup_tick(void)
  80 {
  81         int this_cpu = smp_processor_id();
  82         unsigned long touch_timestamp = per_cpu(touch_timestamp, this_cpu);
  83         unsigned long print_timestamp;
  84         struct pt_regs *regs = get_irq_regs();
  85         unsigned long now;
  86
  87         if (touch_timestamp == 0) {
  88                 __touch_softlockup_watchdog();
  89                 return;
  90         }
  91
  92         print_timestamp = per_cpu(print_timestamp, this_cpu);
  93
  94         /* report at most once a second */
  95         if ((print_timestamp >= touch_timestamp &&
  96                         print_timestamp < (touch_timestamp + 1)) ||
  97                         did_panic || !per_cpu(watchdog_task, this_cpu)) {
  98                 return;
  99         }
 100
 101         /* do not print during early bootup: */
 102         if (unlikely(system_state != SYSTEM_RUNNING)) {
 103                 __touch_softlockup_watchdog();
 104                 return;
 105         }
 106
 107         now = get_timestamp(this_cpu);
 108
 109         /* Wake up the high-prio watchdog task every second: */
 110         if (now > (touch_timestamp + 1))
 111                 wake_up_process(per_cpu(watchdog_task, this_cpu));
 112
 113         /* Warn about unreasonable delays: */
 114         if (now <= (touch_timestamp + softlockup_thresh))
 115                 return;
 116
 117         per_cpu(print_timestamp, this_cpu) = touch_timestamp;
 118
 119         spin_lock(&print_lock);
 120         printk(KERN_ERR "BUG: soft lockup - CPU#%d stuck for %lus! [%s:%d]\n",
 121                         this_cpu, now - touch_timestamp,
 122                         current->comm, task_pid_nr(current));
 123         print_modules();
 124         if (regs)
 125                 show_regs(regs);
 126         else
 127                 dump_stack();
 128         spin_unlock(&print_lock);
 129 }
 130
 131 /*
 132  * Have a reasonable limit on the number of tasks checked:
 133  */
 134 unsigned long __read_mostly sysctl_hung_task_check_count = 1024;
 135
 136 /*
 137  * Zero means infinite timeout - no checking done:
 138  */
 139 unsigned long __read_mostly sysctl_hung_task_timeout_secs = 120;
 140
 141 unsigned long __read_mostly sysctl_hung_task_warnings = 10;
 142
 143 /*
 144  * Only do the hung-tasks check on one CPU:
 145  */
 146 static int check_cpu __read_mostly = -1;
 147
 148 static void check_hung_task(struct task_struct *t, unsigned long now)
 149 {
 150         unsigned long switch_count = t->nvcsw + t->nivcsw;
 151
 152         if (t->flags & PF_FROZEN)
 153                 return;
 154
 155         if (switch_count != t->last_switch_count || !t->last_switch_timestamp) {
 156                 t->last_switch_count = switch_count;
 157                 t->last_switch_timestamp = now;
 158                 return;
 159         }
 160         if ((long)(now - t->last_switch_timestamp) <
 161                                         sysctl_hung_task_timeout_secs)
 162                 return;
 163         if (sysctl_hung_task_warnings < 0)
 164                 return;
 165         sysctl_hung_task_warnings--;
 166
 167         /*
 168          * Ok, the task did not get scheduled for more than 2 minutes,
 169          * complain:
 170          */
 171         printk(KERN_ERR "INFO: task %s:%d blocked for more than "
 172                         "%ld seconds.\n", t->comm, t->pid,
 173                         sysctl_hung_task_timeout_secs);
 174         printk(KERN_ERR "\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
 175                         " disables this message.\n");
 176         sched_show_task(t);
 177         __debug_show_held_locks(t);
 178
 179         t->last_switch_timestamp = now;
 180         touch_nmi_watchdog();
 181 }
 182
 183 /*
 184  * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
 185  * a really long time (120 seconds). If that happens, print out
 186  * a warning.
 187  */
 188 static void check_hung_uninterruptible_tasks(int this_cpu)
 189 {
 190         int max_count = sysctl_hung_task_check_count;
 191         unsigned long now = get_timestamp(this_cpu);
 192         struct task_struct *g, *t;
 193
 194         /*
 195          * If the system crashed already then all bets are off,
 196          * do not report extra hung tasks:
 197          */
 198         if ((tainted & TAINT_DIE) || did_panic)
 199                 return;
 200
 201         read_lock(&tasklist_lock);
 202         do_each_thread(g, t) {
 203                 if (!--max_count)
 204                         goto unlock;
 205                 if (t->state & TASK_UNINTERRUPTIBLE)
 206                         check_hung_task(t, now);
 207         } while_each_thread(g, t);
 208  unlock:
 209         read_unlock(&tasklist_lock);
 210 }
 211
 212 /*
 213  * The watchdog thread - runs every second and touches the timestamp.
 214  */
 215 static int watchdog(void *__bind_cpu)
 216 {
 217         struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
 218         int this_cpu = (long)__bind_cpu;
 219
 220         sched_setscheduler(current, SCHED_FIFO, &param);
 221
 222         /* initialize timestamp */
 223         __touch_softlockup_watchdog();
 224
 225         set_current_state(TASK_INTERRUPTIBLE);
 226         /*
 227          * Run briefly once per second to reset the softlockup timestamp.
 228          * If this gets delayed for more than 60 seconds then the
 229          * debug-printout triggers in softlockup_tick().
 230          */
 231         while (!kthread_should_stop()) {
 232                 __touch_softlockup_watchdog();
 233                 schedule();
 234
 235                 if (kthread_should_stop())
 236                         break;
 237
 238                 if (this_cpu == check_cpu) {
 239                         if (sysctl_hung_task_timeout_secs)
 240                                 check_hung_uninterruptible_tasks(this_cpu);
 241                 }
 242
 243                 set_current_state(TASK_INTERRUPTIBLE);
 244         }
 245         __set_current_state(TASK_RUNNING);
 246
 247         return 0;
 248 }
 249
 250 /*
 251  * Create/destroy watchdog threads as CPUs come and go:
 252  */
 253 static int __cpuinit
 254 cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
 255 {
 256         int hotcpu = (unsigned long)hcpu;
 257         struct task_struct *p;
 258
 259         switch (action) {
 260         case CPU_UP_PREPARE:
 261         case CPU_UP_PREPARE_FROZEN:
 262                 BUG_ON(per_cpu(watchdog_task, hotcpu));
 263                 p = kthread_create(watchdog, hcpu, "watchdog/%d", hotcpu);
 264                 if (IS_ERR(p)) {
 265                         printk(KERN_ERR "watchdog for %i failed\n", hotcpu);
 266                         return NOTIFY_BAD;
 267                 }
 268                 per_cpu(touch_timestamp, hotcpu) = 0;
 269                 per_cpu(watchdog_task, hotcpu) = p;
 270                 kthread_bind(p, hotcpu);
 271                 break;
 272         case CPU_ONLINE:
 273         case CPU_ONLINE_FROZEN:
 274                 check_cpu = any_online_cpu(cpu_online_map);
 275                 wake_up_process(per_cpu(watchdog_task, hotcpu));
 276                 break;
 277 #ifdef CONFIG_HOTPLUG_CPU
 278         case CPU_DOWN_PREPARE:
 279         case CPU_DOWN_PREPARE_FROZEN:
 280                 if (hotcpu == check_cpu) {
 281                         cpumask_t temp_cpu_online_map = cpu_online_map;
 282
 283                         cpu_clear(hotcpu, temp_cpu_online_map);
 284                         check_cpu = any_online_cpu(temp_cpu_online_map);
 285                 }
 286                 break;
 287
 288         case CPU_UP_CANCELED:
 289         case CPU_UP_CANCELED_FROZEN:
 290                 if (!per_cpu(watchdog_task, hotcpu))
 291                         break;
 292                 /* Unbind so it can run.  Fall thru. */
 293                 kthread_bind(per_cpu(watchdog_task, hotcpu),
 294                              any_online_cpu(cpu_online_map));
 295         case CPU_DEAD:
 296         case CPU_DEAD_FROZEN:
 297                 p = per_cpu(watchdog_task, hotcpu);
 298                 per_cpu(watchdog_task, hotcpu) = NULL;
 299                 kthread_stop(p);
 300                 break;
 301 #endif /* CONFIG_HOTPLUG_CPU */
 302         }
 303         return NOTIFY_OK;
 304 }
 305
 306 static struct notifier_block __cpuinitdata cpu_nfb = {
 307         .notifier_call = cpu_callback
 308 };
 309
 310 __init void spawn_softlockup_task(void)
 311 {
 312         void *cpu = (void *)(long)smp_processor_id();
 313         int err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
 314
 315         BUG_ON(err == NOTIFY_BAD);
 316         cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
 317         register_cpu_notifier(&cpu_nfb);
 318
 319         atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
 320 }