SAFE public projects git trees. - safe/jmp/linux-2.6/blob - kernel/hung_task.c

   1 /*
   2  * Detect Hung Task
   3  *
   4  * kernel/hung_task.c - kernel thread for detecting tasks stuck in D state
   5  *
   6  */
   7
   8 #include <linux/mm.h>
   9 #include <linux/cpu.h>
  10 #include <linux/nmi.h>
  11 #include <linux/init.h>
  12 #include <linux/delay.h>
  13 #include <linux/freezer.h>
  14 #include <linux/kthread.h>
  15 #include <linux/lockdep.h>
  16 #include <linux/module.h>
  17 #include <linux/sysctl.h>
  18
  19 /*
  20  * Have a reasonable limit on the number of tasks checked:
  21  */
  22 unsigned long __read_mostly sysctl_hung_task_check_count = 1024;
  23
  24 /*
  25  * Zero means infinite timeout - no checking done:
  26  */
  27 unsigned long __read_mostly sysctl_hung_task_timeout_secs = 120;
  28 static unsigned long __read_mostly hung_task_poll_jiffies;
  29
  30 unsigned long __read_mostly sysctl_hung_task_warnings = 10;
  31
  32 static int __read_mostly did_panic;
  33
  34 static struct task_struct *watchdog_task;
  35
  36 /*
  37  * Should we panic (and reboot, if panic_timeout= is set) when a
  38  * hung task is detected:
  39  */
  40 unsigned int __read_mostly sysctl_hung_task_panic =
  41                                 CONFIG_BOOTPARAM_HUNG_TASK_PANIC_VALUE;
  42
  43 static int __init hung_task_panic_setup(char *str)
  44 {
  45         sysctl_hung_task_panic = simple_strtoul(str, NULL, 0);
  46
  47         return 1;
  48 }
  49 __setup("hung_task_panic=", hung_task_panic_setup);
  50
  51 static int
  52 hung_task_panic(struct notifier_block *this, unsigned long event, void *ptr)
  53 {
  54         did_panic = 1;
  55
  56         return NOTIFY_DONE;
  57 }
  58
  59 static struct notifier_block panic_block = {
  60         .notifier_call = hung_task_panic,
  61 };
  62
  63 /*
  64  * Returns seconds, approximately.  We don't need nanosecond
  65  * resolution, and we don't need to waste time with a big divide when
  66  * 2^30ns == 1.074s.
  67  */
  68 static unsigned long get_timestamp(void)
  69 {
  70         int this_cpu = raw_smp_processor_id();
  71
  72         return cpu_clock(this_cpu) >> 30LL;  /* 2^30 ~= 10^9 */
  73 }
  74
  75 static void check_hung_task(struct task_struct *t, unsigned long now,
  76                             unsigned long timeout)
  77 {
  78         unsigned long switch_count = t->nvcsw + t->nivcsw;
  79
  80         if (t->flags & PF_FROZEN)
  81                 return;
  82
  83         if (switch_count != t->last_switch_count || !t->last_switch_timestamp) {
  84                 t->last_switch_count = switch_count;
  85                 t->last_switch_timestamp = now;
  86                 return;
  87         }
  88         if ((long)(now - t->last_switch_timestamp) < timeout)
  89                 return;
  90         if (!sysctl_hung_task_warnings)
  91                 return;
  92         sysctl_hung_task_warnings--;
  93
  94         /*
  95          * Ok, the task did not get scheduled for more than 2 minutes,
  96          * complain:
  97          */
  98         printk(KERN_ERR "INFO: task %s:%d blocked for more than "
  99                         "%ld seconds.\n", t->comm, t->pid, timeout);
 100         printk(KERN_ERR "\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
 101                         " disables this message.\n");
 102         sched_show_task(t);
 103         __debug_show_held_locks(t);
 104
 105         t->last_switch_timestamp = now;
 106         touch_nmi_watchdog();
 107
 108         if (sysctl_hung_task_panic)
 109                 panic("hung_task: blocked tasks");
 110 }
 111
 112 /*
 113  * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
 114  * a really long time (120 seconds). If that happens, print out
 115  * a warning.
 116  */
 117 static void check_hung_uninterruptible_tasks(unsigned long timeout)
 118 {
 119         int max_count = sysctl_hung_task_check_count;
 120         unsigned long now = get_timestamp();
 121         struct task_struct *g, *t;
 122
 123         /*
 124          * If the system crashed already then all bets are off,
 125          * do not report extra hung tasks:
 126          */
 127         if (test_taint(TAINT_DIE) || did_panic)
 128                 return;
 129
 130         read_lock(&tasklist_lock);
 131         do_each_thread(g, t) {
 132                 if (!--max_count)
 133                         goto unlock;
 134                 /* use "==" to skip the TASK_KILLABLE tasks waiting on NFS */
 135                 if (t->state == TASK_UNINTERRUPTIBLE)
 136                         check_hung_task(t, now, timeout);
 137         } while_each_thread(g, t);
 138  unlock:
 139         read_unlock(&tasklist_lock);
 140 }
 141
 142 static void update_poll_jiffies(void)
 143 {
 144         /* timeout of 0 will disable the watchdog */
 145         if (sysctl_hung_task_timeout_secs == 0)
 146                 hung_task_poll_jiffies = MAX_SCHEDULE_TIMEOUT;
 147         else
 148                 hung_task_poll_jiffies = sysctl_hung_task_timeout_secs * HZ / 2;
 149 }
 150
 151 /*
 152  * Process updating of timeout sysctl
 153  */
 154 int proc_dohung_task_timeout_secs(struct ctl_table *table, int write,
 155                                   struct file *filp, void __user *buffer,
 156                                   size_t *lenp, loff_t *ppos)
 157 {
 158         int ret;
 159
 160         ret = proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos);
 161
 162         if (ret || !write)
 163                 goto out;
 164
 165         update_poll_jiffies();
 166
 167         wake_up_process(watchdog_task);
 168
 169  out:
 170         return ret;
 171 }
 172
 173 /*
 174  * kthread which checks for tasks stuck in D state
 175  */
 176 static int watchdog(void *dummy)
 177 {
 178         set_user_nice(current, 0);
 179         update_poll_jiffies();
 180
 181         for ( ; ; ) {
 182                 unsigned long timeout;
 183
 184                 while (schedule_timeout_interruptible(hung_task_poll_jiffies));
 185
 186                 /*
 187                  * Need to cache timeout here to avoid timeout being set
 188                  * to 0 via sysctl while inside check_hung_*_tasks().
 189                  */
 190                 timeout = sysctl_hung_task_timeout_secs;
 191                 if (timeout)
 192                         check_hung_uninterruptible_tasks(timeout);
 193         }
 194
 195         return 0;
 196 }
 197
 198 static int __init hung_task_init(void)
 199 {
 200         atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
 201         watchdog_task = kthread_run(watchdog, NULL, "khungtaskd");
 202
 203         return 0;
 204 }
 205
 206 module_init(hung_task_init);