git.oblomov.eu Git - 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/lockdep.h>
  17 #include <linux/notifier.h>
  18 #include <linux/module.h>
  19
  20 #include <asm/irq_regs.h>
  21
  22 static DEFINE_SPINLOCK(print_lock);
  23
  24 static DEFINE_PER_CPU(unsigned long, touch_timestamp);
  25 static DEFINE_PER_CPU(unsigned long, print_timestamp);
  26 static DEFINE_PER_CPU(struct task_struct *, watchdog_task);
  27
  28 static int __read_mostly did_panic;
  29 int __read_mostly softlockup_thresh = 60;
  30
  31 /*
  32  * Should we panic (and reboot, if panic_timeout= is set) when a
  33  * soft-lockup occurs:
  34  */
  35 unsigned int __read_mostly softlockup_panic =
  36                                 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
  37
  38 static int __init softlockup_panic_setup(char *str)
  39 {
  40         softlockup_panic = simple_strtoul(str, NULL, 0);
  41
  42         return 1;
  43 }
  44 __setup("softlockup_panic=", softlockup_panic_setup);
  45
  46 static int
  47 softlock_panic(struct notifier_block *this, unsigned long event, void *ptr)
  48 {
  49         did_panic = 1;
  50
  51         return NOTIFY_DONE;
  52 }
  53
  54 static struct notifier_block panic_block = {
  55         .notifier_call = softlock_panic,
  56 };
  57
  58 /*
  59  * Returns seconds, approximately.  We don't need nanosecond
  60  * resolution, and we don't need to waste time with a big divide when
  61  * 2^30ns == 1.074s.
  62  */
  63 static unsigned long get_timestamp(int this_cpu)
  64 {
  65         return cpu_clock(this_cpu) >> 30LL;  /* 2^30 ~= 10^9 */
  66 }
  67
  68 static void __touch_softlockup_watchdog(void)
  69 {
  70         int this_cpu = raw_smp_processor_id();
  71
  72         __raw_get_cpu_var(touch_timestamp) = get_timestamp(this_cpu);
  73 }
  74
  75 void touch_softlockup_watchdog(void)
  76 {
  77         __raw_get_cpu_var(touch_timestamp) = 0;
  78 }
  79 EXPORT_SYMBOL(touch_softlockup_watchdog);
  80
  81 void touch_all_softlockup_watchdogs(void)
  82 {
  83         int cpu;
  84
  85         /* Cause each CPU to re-update its timestamp rather than complain */
  86         for_each_online_cpu(cpu)
  87                 per_cpu(touch_timestamp, cpu) = 0;
  88 }
  89 EXPORT_SYMBOL(touch_all_softlockup_watchdogs);
  90
  91 /*
  92  * This callback runs from the timer interrupt, and checks
  93  * whether the watchdog thread has hung or not:
  94  */
  95 void softlockup_tick(void)
  96 {
  97         int this_cpu = smp_processor_id();
  98         unsigned long touch_timestamp = per_cpu(touch_timestamp, this_cpu);
  99         unsigned long print_timestamp;
 100         struct pt_regs *regs = get_irq_regs();
 101         unsigned long now;
 102
 103         /* Is detection switched off? */
 104         if (!per_cpu(watchdog_task, this_cpu) || softlockup_thresh <= 0) {
 105                 /* Be sure we don't false trigger if switched back on */
 106                 if (touch_timestamp)
 107                         per_cpu(touch_timestamp, this_cpu) = 0;
 108                 return;
 109         }
 110
 111         if (touch_timestamp == 0) {
 112                 __touch_softlockup_watchdog();
 113                 return;
 114         }
 115
 116         print_timestamp = per_cpu(print_timestamp, this_cpu);
 117
 118         /* report at most once a second */
 119         if (print_timestamp == touch_timestamp || did_panic)
 120                 return;
 121
 122         /* do not print during early bootup: */
 123         if (unlikely(system_state != SYSTEM_RUNNING)) {
 124                 __touch_softlockup_watchdog();
 125                 return;
 126         }
 127
 128         now = get_timestamp(this_cpu);
 129
 130         /*
 131          * Wake up the high-prio watchdog task twice per
 132          * threshold timespan.
 133          */
 134         if (now > touch_timestamp + softlockup_thresh/2)
 135                 wake_up_process(per_cpu(watchdog_task, this_cpu));
 136
 137         /* Warn about unreasonable delays: */
 138         if (now <= (touch_timestamp + softlockup_thresh))
 139                 return;
 140
 141         per_cpu(print_timestamp, this_cpu) = touch_timestamp;
 142
 143         spin_lock(&print_lock);
 144         printk(KERN_ERR "BUG: soft lockup - CPU#%d stuck for %lus! [%s:%d]\n",
 145                         this_cpu, now - touch_timestamp,
 146                         current->comm, task_pid_nr(current));
 147         print_modules();
 148         print_irqtrace_events(current);
 149         if (regs)
 150                 show_regs(regs);
 151         else
 152                 dump_stack();
 153         spin_unlock(&print_lock);
 154
 155         if (softlockup_panic)
 156                 panic("softlockup: hung tasks");
 157 }
 158
 159 /*
 160  * Have a reasonable limit on the number of tasks checked:
 161  */
 162 unsigned long __read_mostly sysctl_hung_task_check_count = 1024;
 163
 164 /*
 165  * Zero means infinite timeout - no checking done:
 166  */
 167 unsigned long __read_mostly sysctl_hung_task_timeout_secs = 480;
 168
 169 unsigned long __read_mostly sysctl_hung_task_warnings = 10;
 170
 171 /*
 172  * Only do the hung-tasks check on one CPU:
 173  */
 174 static int check_cpu __read_mostly = -1;
 175
 176 static void check_hung_task(struct task_struct *t, unsigned long now)
 177 {
 178         unsigned long switch_count = t->nvcsw + t->nivcsw;
 179
 180         if (t->flags & PF_FROZEN)
 181                 return;
 182
 183         if (switch_count != t->last_switch_count || !t->last_switch_timestamp) {
 184                 t->last_switch_count = switch_count;
 185                 t->last_switch_timestamp = now;
 186                 return;
 187         }
 188         if ((long)(now - t->last_switch_timestamp) <
 189                                         sysctl_hung_task_timeout_secs)
 190                 return;
 191         if (!sysctl_hung_task_warnings)
 192                 return;
 193         sysctl_hung_task_warnings--;
 194
 195         /*
 196          * Ok, the task did not get scheduled for more than 2 minutes,
 197          * complain:
 198          */
 199         printk(KERN_ERR "INFO: task %s:%d blocked for more than "
 200                         "%ld seconds.\n", t->comm, t->pid,
 201                         sysctl_hung_task_timeout_secs);
 202         printk(KERN_ERR "\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
 203                         " disables this message.\n");
 204         sched_show_task(t);
 205         __debug_show_held_locks(t);
 206
 207         t->last_switch_timestamp = now;
 208         touch_nmi_watchdog();
 209
 210         if (softlockup_panic)
 211                 panic("softlockup: blocked tasks");
 212 }
 213
 214 /*
 215  * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
 216  * a really long time (120 seconds). If that happens, print out
 217  * a warning.
 218  */
 219 static void check_hung_uninterruptible_tasks(int this_cpu)
 220 {
 221         int max_count = sysctl_hung_task_check_count;
 222         unsigned long now = get_timestamp(this_cpu);
 223         struct task_struct *g, *t;
 224
 225         /*
 226          * If the system crashed already then all bets are off,
 227          * do not report extra hung tasks:
 228          */
 229         if (test_taint(TAINT_DIE) || did_panic)
 230                 return;
 231
 232         read_lock(&tasklist_lock);
 233         do_each_thread(g, t) {
 234                 if (!--max_count)
 235                         goto unlock;
 236                 /* use "==" to skip the TASK_KILLABLE tasks waiting on NFS */
 237                 if (t->state == TASK_UNINTERRUPTIBLE)
 238                         check_hung_task(t, now);
 239         } while_each_thread(g, t);
 240  unlock:
 241         read_unlock(&tasklist_lock);
 242 }
 243
 244 /*
 245  * The watchdog thread - runs every second and touches the timestamp.
 246  */
 247 static int watchdog(void *__bind_cpu)
 248 {
 249         struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
 250         int this_cpu = (long)__bind_cpu;
 251
 252         sched_setscheduler(current, SCHED_FIFO, &param);
 253
 254         /* initialize timestamp */
 255         __touch_softlockup_watchdog();
 256
 257         set_current_state(TASK_INTERRUPTIBLE);
 258         /*
 259          * Run briefly once per second to reset the softlockup timestamp.
 260          * If this gets delayed for more than 60 seconds then the
 261          * debug-printout triggers in softlockup_tick().
 262          */
 263         while (!kthread_should_stop()) {
 264                 __touch_softlockup_watchdog();
 265                 schedule();
 266
 267                 if (kthread_should_stop())
 268                         break;
 269
 270                 if (this_cpu == check_cpu) {
 271                         if (sysctl_hung_task_timeout_secs)
 272                                 check_hung_uninterruptible_tasks(this_cpu);
 273                 }
 274
 275                 set_current_state(TASK_INTERRUPTIBLE);
 276         }
 277         __set_current_state(TASK_RUNNING);
 278
 279         return 0;
 280 }
 281
 282 /*
 283  * Create/destroy watchdog threads as CPUs come and go:
 284  */
 285 static int __cpuinit
 286 cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
 287 {
 288         int hotcpu = (unsigned long)hcpu;
 289         struct task_struct *p;
 290
 291         switch (action) {
 292         case CPU_UP_PREPARE:
 293         case CPU_UP_PREPARE_FROZEN:
 294                 BUG_ON(per_cpu(watchdog_task, hotcpu));
 295                 p = kthread_create(watchdog, hcpu, "watchdog/%d", hotcpu);
 296                 if (IS_ERR(p)) {
 297                         printk(KERN_ERR "watchdog for %i failed\n", hotcpu);
 298                         return NOTIFY_BAD;
 299                 }
 300                 per_cpu(touch_timestamp, hotcpu) = 0;
 301                 per_cpu(watchdog_task, hotcpu) = p;
 302                 kthread_bind(p, hotcpu);
 303                 break;
 304         case CPU_ONLINE:
 305         case CPU_ONLINE_FROZEN:
 306                 check_cpu = any_online_cpu(cpu_online_map);
 307                 wake_up_process(per_cpu(watchdog_task, hotcpu));
 308                 break;
 309 #ifdef CONFIG_HOTPLUG_CPU
 310         case CPU_DOWN_PREPARE:
 311         case CPU_DOWN_PREPARE_FROZEN:
 312                 if (hotcpu == check_cpu) {
 313                         cpumask_t temp_cpu_online_map = cpu_online_map;
 314
 315                         cpu_clear(hotcpu, temp_cpu_online_map);
 316                         check_cpu = any_online_cpu(temp_cpu_online_map);
 317                 }
 318                 break;
 319
 320         case CPU_UP_CANCELED:
 321         case CPU_UP_CANCELED_FROZEN:
 322                 if (!per_cpu(watchdog_task, hotcpu))
 323                         break;
 324                 /* Unbind so it can run.  Fall thru. */
 325                 kthread_bind(per_cpu(watchdog_task, hotcpu),
 326                              any_online_cpu(cpu_online_map));
 327         case CPU_DEAD:
 328         case CPU_DEAD_FROZEN:
 329                 p = per_cpu(watchdog_task, hotcpu);
 330                 per_cpu(watchdog_task, hotcpu) = NULL;
 331                 kthread_stop(p);
 332                 break;
 333 #endif /* CONFIG_HOTPLUG_CPU */
 334         }
 335         return NOTIFY_OK;
 336 }
 337
 338 static struct notifier_block __cpuinitdata cpu_nfb = {
 339         .notifier_call = cpu_callback
 340 };
 341
 342 static int __initdata nosoftlockup;
 343
 344 static int __init nosoftlockup_setup(char *str)
 345 {
 346         nosoftlockup = 1;
 347         return 1;
 348 }
 349 __setup("nosoftlockup", nosoftlockup_setup);
 350
 351 static int __init spawn_softlockup_task(void)
 352 {
 353         void *cpu = (void *)(long)smp_processor_id();
 354         int err;
 355
 356         if (nosoftlockup)
 357                 return 0;
 358
 359         err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
 360         if (err == NOTIFY_BAD) {
 361                 BUG();
 362                 return 1;
 363         }
 364         cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
 365         register_cpu_notifier(&cpu_nfb);
 366
 367         atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
 368
 369         return 0;
 370 }
 371 early_initcall(spawn_softlockup_task);