SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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 int __read_mostly softlockup_thresh = 60;
  29
  30 /*
  31  * Should we panic (and reboot, if panic_timeout= is set) when a
  32  * soft-lockup occurs:
  33  */
  34 unsigned int __read_mostly softlockup_panic =
  35                                 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
  36
  37 static int __init softlockup_panic_setup(char *str)
  38 {
  39         softlockup_panic = simple_strtoul(str, NULL, 0);
  40
  41         return 1;
  42 }
  43 __setup("softlockup_panic=", softlockup_panic_setup);
  44
  45 static int
  46 softlock_panic(struct notifier_block *this, unsigned long event, void *ptr)
  47 {
  48         did_panic = 1;
  49
  50         return NOTIFY_DONE;
  51 }
  52
  53 static struct notifier_block panic_block = {
  54         .notifier_call = softlock_panic,
  55 };
  56
  57 /*
  58  * Returns seconds, approximately.  We don't need nanosecond
  59  * resolution, and we don't need to waste time with a big divide when
  60  * 2^30ns == 1.074s.
  61  */
  62 static unsigned long get_timestamp(int this_cpu)
  63 {
  64         return cpu_clock(this_cpu) >> 30LL;  /* 2^30 ~= 10^9 */
  65 }
  66
  67 void touch_softlockup_watchdog(void)
  68 {
  69         int this_cpu = raw_smp_processor_id();
  70
  71         __raw_get_cpu_var(touch_timestamp) = get_timestamp(this_cpu);
  72 }
  73 EXPORT_SYMBOL(touch_softlockup_watchdog);
  74
  75 void touch_all_softlockup_watchdogs(void)
  76 {
  77         int cpu;
  78
  79         /* Cause each CPU to re-update its timestamp rather than complain */
  80         for_each_online_cpu(cpu)
  81                 per_cpu(touch_timestamp, cpu) = 0;
  82 }
  83 EXPORT_SYMBOL(touch_all_softlockup_watchdogs);
  84
  85 /*
  86  * This callback runs from the timer interrupt, and checks
  87  * whether the watchdog thread has hung or not:
  88  */
  89 void softlockup_tick(void)
  90 {
  91         int this_cpu = smp_processor_id();
  92         unsigned long touch_timestamp = per_cpu(touch_timestamp, this_cpu);
  93         unsigned long print_timestamp;
  94         struct pt_regs *regs = get_irq_regs();
  95         unsigned long now;
  96
  97         /* Is detection switched off? */
  98         if (!per_cpu(watchdog_task, this_cpu) || softlockup_thresh <= 0) {
  99                 /* Be sure we don't false trigger if switched back on */
 100                 if (touch_timestamp)
 101                         per_cpu(touch_timestamp, this_cpu) = 0;
 102                 return;
 103         }
 104
 105         if (touch_timestamp == 0) {
 106                 touch_softlockup_watchdog();
 107                 return;
 108         }
 109
 110         print_timestamp = per_cpu(print_timestamp, this_cpu);
 111
 112         /* report at most once a second */
 113         if ((print_timestamp >= touch_timestamp &&
 114                         print_timestamp < (touch_timestamp + 1)) ||
 115                         did_panic) {
 116                 return;
 117         }
 118
 119         /* do not print during early bootup: */
 120         if (unlikely(system_state != SYSTEM_RUNNING)) {
 121                 touch_softlockup_watchdog();
 122                 return;
 123         }
 124
 125         now = get_timestamp(this_cpu);
 126
 127         /* Wake up the high-prio watchdog task every second: */
 128         if (now > (touch_timestamp + 1))
 129                 wake_up_process(per_cpu(watchdog_task, this_cpu));
 130
 131         /* Warn about unreasonable delays: */
 132         if (now <= (touch_timestamp + softlockup_thresh))
 133                 return;
 134
 135         per_cpu(print_timestamp, this_cpu) = touch_timestamp;
 136
 137         spin_lock(&print_lock);
 138         printk(KERN_ERR "BUG: soft lockup - CPU#%d stuck for %lus! [%s:%d]\n",
 139                         this_cpu, now - touch_timestamp,
 140                         current->comm, task_pid_nr(current));
 141         if (regs)
 142                 show_regs(regs);
 143         else
 144                 dump_stack();
 145         spin_unlock(&print_lock);
 146
 147         if (softlockup_panic)
 148                 panic("softlockup: hung tasks");
 149 }
 150
 151 /*
 152  * Have a reasonable limit on the number of tasks checked:
 153  */
 154 unsigned long __read_mostly sysctl_hung_task_check_count = 1024;
 155
 156 /*
 157  * Zero means infinite timeout - no checking done:
 158  */
 159 unsigned long __read_mostly sysctl_hung_task_timeout_secs = 120;
 160
 161 unsigned long __read_mostly sysctl_hung_task_warnings = 10;
 162
 163 /*
 164  * Only do the hung-tasks check on one CPU:
 165  */
 166 static int check_cpu __read_mostly = -1;
 167
 168 static void check_hung_task(struct task_struct *t, unsigned long now)
 169 {
 170         unsigned long switch_count = t->nvcsw + t->nivcsw;
 171
 172         if (t->flags & PF_FROZEN)
 173                 return;
 174
 175         if (switch_count != t->last_switch_count || !t->last_switch_timestamp) {
 176                 t->last_switch_count = switch_count;
 177                 t->last_switch_timestamp = now;
 178                 return;
 179         }
 180         if ((long)(now - t->last_switch_timestamp) <
 181                                         sysctl_hung_task_timeout_secs)
 182                 return;
 183         if (sysctl_hung_task_warnings < 0)
 184                 return;
 185         sysctl_hung_task_warnings--;
 186
 187         /*
 188          * Ok, the task did not get scheduled for more than 2 minutes,
 189          * complain:
 190          */
 191         printk(KERN_ERR "INFO: task %s:%d blocked for more than "
 192                         "%ld seconds.\n", t->comm, t->pid,
 193                         sysctl_hung_task_timeout_secs);
 194         printk(KERN_ERR "\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
 195                         " disables this message.\n");
 196         sched_show_task(t);
 197         __debug_show_held_locks(t);
 198
 199         t->last_switch_timestamp = now;
 200         touch_nmi_watchdog();
 201
 202         if (softlockup_panic)
 203                 panic("softlockup: blocked tasks");
 204 }
 205
 206 /*
 207  * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
 208  * a really long time (120 seconds). If that happens, print out
 209  * a warning.
 210  */
 211 static void check_hung_uninterruptible_tasks(int this_cpu)
 212 {
 213         int max_count = sysctl_hung_task_check_count;
 214         unsigned long now = get_timestamp(this_cpu);
 215         struct task_struct *g, *t;
 216
 217         /*
 218          * If the system crashed already then all bets are off,
 219          * do not report extra hung tasks:
 220          */
 221         if ((tainted & TAINT_DIE) || did_panic)
 222                 return;
 223
 224         read_lock(&tasklist_lock);
 225         do_each_thread(g, t) {
 226                 if (!--max_count)
 227                         goto unlock;
 228                 if (t->state & TASK_UNINTERRUPTIBLE)
 229                         check_hung_task(t, now);
 230         } while_each_thread(g, t);
 231  unlock:
 232         read_unlock(&tasklist_lock);
 233 }
 234
 235 /*
 236  * The watchdog thread - runs every second and touches the timestamp.
 237  */
 238 static int watchdog(void *__bind_cpu)
 239 {
 240         struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
 241         int this_cpu = (long)__bind_cpu;
 242
 243         sched_setscheduler(current, SCHED_FIFO, &param);
 244
 245         /* initialize timestamp */
 246         touch_softlockup_watchdog();
 247
 248         set_current_state(TASK_INTERRUPTIBLE);
 249         /*
 250          * Run briefly once per second to reset the softlockup timestamp.
 251          * If this gets delayed for more than 60 seconds then the
 252          * debug-printout triggers in softlockup_tick().
 253          */
 254         while (!kthread_should_stop()) {
 255                 touch_softlockup_watchdog();
 256                 schedule();
 257
 258                 if (kthread_should_stop())
 259                         break;
 260
 261                 if (this_cpu == check_cpu) {
 262                         if (sysctl_hung_task_timeout_secs)
 263                                 check_hung_uninterruptible_tasks(this_cpu);
 264                 }
 265
 266                 set_current_state(TASK_INTERRUPTIBLE);
 267         }
 268         __set_current_state(TASK_RUNNING);
 269
 270         return 0;
 271 }
 272
 273 /*
 274  * Create/destroy watchdog threads as CPUs come and go:
 275  */
 276 static int __cpuinit
 277 cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
 278 {
 279         int hotcpu = (unsigned long)hcpu;
 280         struct task_struct *p;
 281
 282         switch (action) {
 283         case CPU_UP_PREPARE:
 284         case CPU_UP_PREPARE_FROZEN:
 285                 BUG_ON(per_cpu(watchdog_task, hotcpu));
 286                 p = kthread_create(watchdog, hcpu, "watchdog/%d", hotcpu);
 287                 if (IS_ERR(p)) {
 288                         printk(KERN_ERR "watchdog for %i failed\n", hotcpu);
 289                         return NOTIFY_BAD;
 290                 }
 291                 per_cpu(touch_timestamp, hotcpu) = 0;
 292                 per_cpu(watchdog_task, hotcpu) = p;
 293                 kthread_bind(p, hotcpu);
 294                 break;
 295         case CPU_ONLINE:
 296         case CPU_ONLINE_FROZEN:
 297                 check_cpu = any_online_cpu(cpu_online_map);
 298                 wake_up_process(per_cpu(watchdog_task, hotcpu));
 299                 break;
 300 #ifdef CONFIG_HOTPLUG_CPU
 301         case CPU_DOWN_PREPARE:
 302         case CPU_DOWN_PREPARE_FROZEN:
 303                 if (hotcpu == check_cpu) {
 304                         cpumask_t temp_cpu_online_map = cpu_online_map;
 305
 306                         cpu_clear(hotcpu, temp_cpu_online_map);
 307                         check_cpu = any_online_cpu(temp_cpu_online_map);
 308                 }
 309                 break;
 310
 311         case CPU_UP_CANCELED:
 312         case CPU_UP_CANCELED_FROZEN:
 313                 if (!per_cpu(watchdog_task, hotcpu))
 314                         break;
 315                 /* Unbind so it can run.  Fall thru. */
 316                 kthread_bind(per_cpu(watchdog_task, hotcpu),
 317                              any_online_cpu(cpu_online_map));
 318         case CPU_DEAD:
 319         case CPU_DEAD_FROZEN:
 320                 p = per_cpu(watchdog_task, hotcpu);
 321                 per_cpu(watchdog_task, hotcpu) = NULL;
 322                 kthread_stop(p);
 323                 break;
 324 #endif /* CONFIG_HOTPLUG_CPU */
 325         }
 326         return NOTIFY_OK;
 327 }
 328
 329 static struct notifier_block __cpuinitdata cpu_nfb = {
 330         .notifier_call = cpu_callback
 331 };
 332
 333 __init void spawn_softlockup_task(void)
 334 {
 335         void *cpu = (void *)(long)smp_processor_id();
 336         int err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
 337
 338         BUG_ON(err == NOTIFY_BAD);
 339         cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
 340         register_cpu_notifier(&cpu_nfb);
 341
 342         atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
 343 }