Root/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/export.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, skip vfork and any other user process that freezer should skip.
78     */
79    if (unlikely(t->flags & (PF_FROZEN | PF_FREEZER_SKIP)))
80        return;
81
82    /*
83     * When a freshly created task is scheduled once, changes its state to
84     * TASK_UNINTERRUPTIBLE without having ever been switched out once, it
85     * musn't be checked.
86     */
87    if (unlikely(!switch_count))
88        return;
89
90    if (switch_count != t->last_switch_count) {
91        t->last_switch_count = switch_count;
92        return;
93    }
94    if (!sysctl_hung_task_warnings)
95        return;
96    sysctl_hung_task_warnings--;
97
98    /*
99     * Ok, the task did not get scheduled for more than 2 minutes,
100     * complain:
101     */
102    printk(KERN_ERR "INFO: task %s:%d blocked for more than "
103            "%ld seconds.\n", t->comm, t->pid, timeout);
104    printk(KERN_ERR "\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
105            " disables this message.\n");
106    sched_show_task(t);
107    debug_show_held_locks(t);
108
109    touch_nmi_watchdog();
110
111    if (sysctl_hung_task_panic) {
112        trigger_all_cpu_backtrace();
113        panic("hung_task: blocked tasks");
114    }
115}
116
117/*
118 * To avoid extending the RCU grace period for an unbounded amount of time,
119 * periodically exit the critical section and enter a new one.
120 *
121 * For preemptible RCU it is sufficient to call rcu_read_unlock in order
122 * to exit the grace period. For classic RCU, a reschedule is required.
123 */
124static bool rcu_lock_break(struct task_struct *g, struct task_struct *t)
125{
126    bool can_cont;
127
128    get_task_struct(g);
129    get_task_struct(t);
130    rcu_read_unlock();
131    cond_resched();
132    rcu_read_lock();
133    can_cont = pid_alive(g) && pid_alive(t);
134    put_task_struct(t);
135    put_task_struct(g);
136
137    return can_cont;
138}
139
140/*
141 * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
142 * a really long time (120 seconds). If that happens, print out
143 * a warning.
144 */
145static void check_hung_uninterruptible_tasks(unsigned long timeout)
146{
147    int max_count = sysctl_hung_task_check_count;
148    int batch_count = HUNG_TASK_BATCHING;
149    struct task_struct *g, *t;
150
151    /*
152     * If the system crashed already then all bets are off,
153     * do not report extra hung tasks:
154     */
155    if (test_taint(TAINT_DIE) || did_panic)
156        return;
157
158    rcu_read_lock();
159    do_each_thread(g, t) {
160        if (!max_count--)
161            goto unlock;
162        if (!--batch_count) {
163            batch_count = HUNG_TASK_BATCHING;
164            if (!rcu_lock_break(g, t))
165                goto unlock;
166        }
167        /* use "==" to skip the TASK_KILLABLE tasks waiting on NFS */
168        if (t->state == TASK_UNINTERRUPTIBLE)
169            check_hung_task(t, timeout);
170    } while_each_thread(g, t);
171 unlock:
172    rcu_read_unlock();
173}
174
175static unsigned long timeout_jiffies(unsigned long timeout)
176{
177    /* timeout of 0 will disable the watchdog */
178    return timeout ? timeout * HZ : MAX_SCHEDULE_TIMEOUT;
179}
180
181/*
182 * Process updating of timeout sysctl
183 */
184int proc_dohung_task_timeout_secs(struct ctl_table *table, int write,
185                  void __user *buffer,
186                  size_t *lenp, loff_t *ppos)
187{
188    int ret;
189
190    ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
191
192    if (ret || !write)
193        goto out;
194
195    wake_up_process(watchdog_task);
196
197 out:
198    return ret;
199}
200
201/*
202 * kthread which checks for tasks stuck in D state
203 */
204static int watchdog(void *dummy)
205{
206    set_user_nice(current, 0);
207
208    for ( ; ; ) {
209        unsigned long timeout = sysctl_hung_task_timeout_secs;
210
211        while (schedule_timeout_interruptible(timeout_jiffies(timeout)))
212            timeout = sysctl_hung_task_timeout_secs;
213
214        check_hung_uninterruptible_tasks(timeout);
215    }
216
217    return 0;
218}
219
220static int __init hung_task_init(void)
221{
222    atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
223    watchdog_task = kthread_run(watchdog, NULL, "khungtaskd");
224
225    return 0;
226}
227
228module_init(hung_task_init);
229

Archive Download this file



interactive