linux/kernel/hung_task.c
<<
>>
Prefs
   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 * The number of tasks checked:
  21 */
  22unsigned long __read_mostly sysctl_hung_task_check_count = PID_MAX_LIMIT;
  23
  24/*
  25 * Limit number of tasks checked in a batch.
  26 *
  27 * This value controls the preemptibility of khungtaskd since preemption
  28 * is disabled during the critical section. It also controls the size of
  29 * the RCU grace period. So it needs to be upper-bound.
  30 */
  31#define HUNG_TASK_BATCHING 1024
  32
  33/*
  34 * Zero means infinite timeout - no checking done:
  35 */
  36unsigned long __read_mostly sysctl_hung_task_timeout_secs = CONFIG_DEFAULT_HUNG_TASK_TIMEOUT;
  37
  38unsigned long __read_mostly sysctl_hung_task_warnings = 10;
  39
  40static int __read_mostly did_panic;
  41
  42static struct task_struct *watchdog_task;
  43
  44/*
  45 * Should we panic (and reboot, if panic_timeout= is set) when a
  46 * hung task is detected:
  47 */
  48unsigned int __read_mostly sysctl_hung_task_panic =
  49                                CONFIG_BOOTPARAM_HUNG_TASK_PANIC_VALUE;
  50
  51static int __init hung_task_panic_setup(char *str)
  52{
  53        sysctl_hung_task_panic = simple_strtoul(str, NULL, 0);
  54
  55        return 1;
  56}
  57__setup("hung_task_panic=", hung_task_panic_setup);
  58
  59static int
  60hung_task_panic(struct notifier_block *this, unsigned long event, void *ptr)
  61{
  62        did_panic = 1;
  63
  64        return NOTIFY_DONE;
  65}
  66
  67static struct notifier_block panic_block = {
  68        .notifier_call = hung_task_panic,
  69};
  70
  71static void check_hung_task(struct task_struct *t, unsigned long timeout)
  72{
  73        unsigned long switch_count = t->nvcsw + t->nivcsw;
  74
  75        /*
  76         * Ensure the task is not frozen.
  77         * Also, when a freshly created task is scheduled once, changes
  78         * its state to TASK_UNINTERRUPTIBLE without having ever been
  79         * switched out once, it musn't be checked.
  80         */
  81        if (unlikely(t->flags & PF_FROZEN || !switch_count))
  82                return;
  83
  84        if (switch_count != t->last_switch_count) {
  85                t->last_switch_count = switch_count;
  86                return;
  87        }
  88        if (!sysctl_hung_task_warnings)
  89                return;
  90        sysctl_hung_task_warnings--;
  91
  92        /*
  93         * Ok, the task did not get scheduled for more than 2 minutes,
  94         * complain:
  95         */
  96        printk(KERN_ERR "INFO: task %s:%d blocked for more than "
  97                        "%ld seconds.\n", t->comm, t->pid, timeout);
  98        printk(KERN_ERR "\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
  99                        " disables this message.\n");
 100        sched_show_task(t);
 101        debug_show_held_locks(t);
 102
 103        touch_nmi_watchdog();
 104
 105        if (sysctl_hung_task_panic)
 106                panic("hung_task: blocked tasks");
 107}
 108
 109/*
 110 * To avoid extending the RCU grace period for an unbounded amount of time,
 111 * periodically exit the critical section and enter a new one.
 112 *
 113 * For preemptible RCU it is sufficient to call rcu_read_unlock in order
 114 * to exit the grace period. For classic RCU, a reschedule is required.
 115 */
 116static void rcu_lock_break(struct task_struct *g, struct task_struct *t)
 117{
 118        get_task_struct(g);
 119        get_task_struct(t);
 120        rcu_read_unlock();
 121        cond_resched();
 122        rcu_read_lock();
 123        put_task_struct(t);
 124        put_task_struct(g);
 125}
 126
 127/*
 128 * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
 129 * a really long time (120 seconds). If that happens, print out
 130 * a warning.
 131 */
 132static void check_hung_uninterruptible_tasks(unsigned long timeout)
 133{
 134        int max_count = sysctl_hung_task_check_count;
 135        int batch_count = HUNG_TASK_BATCHING;
 136        struct task_struct *g, *t;
 137
 138        /*
 139         * If the system crashed already then all bets are off,
 140         * do not report extra hung tasks:
 141         */
 142        if (test_taint(TAINT_DIE) || did_panic)
 143                return;
 144
 145        rcu_read_lock();
 146        do_each_thread(g, t) {
 147                if (!max_count--)
 148                        goto unlock;
 149                if (!--batch_count) {
 150                        batch_count = HUNG_TASK_BATCHING;
 151                        rcu_lock_break(g, t);
 152                        /* Exit if t or g was unhashed during refresh. */
 153                        if (t->state == TASK_DEAD || g->state == TASK_DEAD)
 154                                goto unlock;
 155                }
 156                /* use "==" to skip the TASK_KILLABLE tasks waiting on NFS */
 157                if (t->state == TASK_UNINTERRUPTIBLE)
 158                        check_hung_task(t, timeout);
 159        } while_each_thread(g, t);
 160 unlock:
 161        rcu_read_unlock();
 162}
 163
 164static unsigned long timeout_jiffies(unsigned long timeout)
 165{
 166        /* timeout of 0 will disable the watchdog */
 167        return timeout ? timeout * HZ : MAX_SCHEDULE_TIMEOUT;
 168}
 169
 170/*
 171 * Process updating of timeout sysctl
 172 */
 173int proc_dohung_task_timeout_secs(struct ctl_table *table, int write,
 174                                  void __user *buffer,
 175                                  size_t *lenp, loff_t *ppos)
 176{
 177        int ret;
 178
 179        ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
 180
 181        if (ret || !write)
 182                goto out;
 183
 184        wake_up_process(watchdog_task);
 185
 186 out:
 187        return ret;
 188}
 189
 190/*
 191 * kthread which checks for tasks stuck in D state
 192 */
 193static int watchdog(void *dummy)
 194{
 195        set_user_nice(current, 0);
 196
 197        for ( ; ; ) {
 198                unsigned long timeout = sysctl_hung_task_timeout_secs;
 199
 200                while (schedule_timeout_interruptible(timeout_jiffies(timeout)))
 201                        timeout = sysctl_hung_task_timeout_secs;
 202
 203                check_hung_uninterruptible_tasks(timeout);
 204        }
 205
 206        return 0;
 207}
 208
 209static int __init hung_task_init(void)
 210{
 211        atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
 212        watchdog_task = kthread_run(watchdog, NULL, "khungtaskd");
 213
 214        return 0;
 215}
 216
 217module_init(hung_task_init);
 218
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.