xref: /openbmc/linux/kernel/hung_task.c (revision 5bd8e16d)
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 #include <linux/utsname.h>
19 
20 /*
21  * The number of tasks checked:
22  */
23 unsigned long __read_mostly sysctl_hung_task_check_count = PID_MAX_LIMIT;
24 
25 /*
26  * Limit number of tasks checked in a batch.
27  *
28  * This value controls the preemptibility of khungtaskd since preemption
29  * is disabled during the critical section. It also controls the size of
30  * the RCU grace period. So it needs to be upper-bound.
31  */
32 #define HUNG_TASK_BATCHING 1024
33 
34 /*
35  * Zero means infinite timeout - no checking done:
36  */
37 unsigned long __read_mostly sysctl_hung_task_timeout_secs = CONFIG_DEFAULT_HUNG_TASK_TIMEOUT;
38 
39 unsigned long __read_mostly sysctl_hung_task_warnings = 10;
40 
41 static int __read_mostly did_panic;
42 
43 static struct task_struct *watchdog_task;
44 
45 /*
46  * Should we panic (and reboot, if panic_timeout= is set) when a
47  * hung task is detected:
48  */
49 unsigned int __read_mostly sysctl_hung_task_panic =
50 				CONFIG_BOOTPARAM_HUNG_TASK_PANIC_VALUE;
51 
52 static int __init hung_task_panic_setup(char *str)
53 {
54 	sysctl_hung_task_panic = simple_strtoul(str, NULL, 0);
55 
56 	return 1;
57 }
58 __setup("hung_task_panic=", hung_task_panic_setup);
59 
60 static int
61 hung_task_panic(struct notifier_block *this, unsigned long event, void *ptr)
62 {
63 	did_panic = 1;
64 
65 	return NOTIFY_DONE;
66 }
67 
68 static struct notifier_block panic_block = {
69 	.notifier_call = hung_task_panic,
70 };
71 
72 static void check_hung_task(struct task_struct *t, unsigned long timeout)
73 {
74 	unsigned long switch_count = t->nvcsw + t->nivcsw;
75 
76 	/*
77 	 * Ensure the task is not frozen.
78 	 * Also, skip vfork and any other user process that freezer should skip.
79 	 */
80 	if (unlikely(t->flags & (PF_FROZEN | PF_FREEZER_SKIP)))
81 	    return;
82 
83 	/*
84 	 * When a freshly created task is scheduled once, changes its state to
85 	 * TASK_UNINTERRUPTIBLE without having ever been switched out once, it
86 	 * musn't be checked.
87 	 */
88 	if (unlikely(!switch_count))
89 		return;
90 
91 	if (switch_count != t->last_switch_count) {
92 		t->last_switch_count = switch_count;
93 		return;
94 	}
95 	if (!sysctl_hung_task_warnings)
96 		return;
97 	sysctl_hung_task_warnings--;
98 
99 	/*
100 	 * Ok, the task did not get scheduled for more than 2 minutes,
101 	 * complain:
102 	 */
103 	pr_err("INFO: task %s:%d blocked for more than %ld seconds.\n",
104 		t->comm, t->pid, timeout);
105 	pr_err("      %s %s %.*s\n",
106 		print_tainted(), init_utsname()->release,
107 		(int)strcspn(init_utsname()->version, " "),
108 		init_utsname()->version);
109 	pr_err("\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
110 		" disables this message.\n");
111 	sched_show_task(t);
112 	debug_show_held_locks(t);
113 
114 	touch_nmi_watchdog();
115 
116 	if (sysctl_hung_task_panic) {
117 		trigger_all_cpu_backtrace();
118 		panic("hung_task: blocked tasks");
119 	}
120 }
121 
122 /*
123  * To avoid extending the RCU grace period for an unbounded amount of time,
124  * periodically exit the critical section and enter a new one.
125  *
126  * For preemptible RCU it is sufficient to call rcu_read_unlock in order
127  * to exit the grace period. For classic RCU, a reschedule is required.
128  */
129 static bool rcu_lock_break(struct task_struct *g, struct task_struct *t)
130 {
131 	bool can_cont;
132 
133 	get_task_struct(g);
134 	get_task_struct(t);
135 	rcu_read_unlock();
136 	cond_resched();
137 	rcu_read_lock();
138 	can_cont = pid_alive(g) && pid_alive(t);
139 	put_task_struct(t);
140 	put_task_struct(g);
141 
142 	return can_cont;
143 }
144 
145 /*
146  * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
147  * a really long time (120 seconds). If that happens, print out
148  * a warning.
149  */
150 static void check_hung_uninterruptible_tasks(unsigned long timeout)
151 {
152 	int max_count = sysctl_hung_task_check_count;
153 	int batch_count = HUNG_TASK_BATCHING;
154 	struct task_struct *g, *t;
155 
156 	/*
157 	 * If the system crashed already then all bets are off,
158 	 * do not report extra hung tasks:
159 	 */
160 	if (test_taint(TAINT_DIE) || did_panic)
161 		return;
162 
163 	rcu_read_lock();
164 	do_each_thread(g, t) {
165 		if (!max_count--)
166 			goto unlock;
167 		if (!--batch_count) {
168 			batch_count = HUNG_TASK_BATCHING;
169 			if (!rcu_lock_break(g, t))
170 				goto unlock;
171 		}
172 		/* use "==" to skip the TASK_KILLABLE tasks waiting on NFS */
173 		if (t->state == TASK_UNINTERRUPTIBLE)
174 			check_hung_task(t, timeout);
175 	} while_each_thread(g, t);
176  unlock:
177 	rcu_read_unlock();
178 }
179 
180 static unsigned long timeout_jiffies(unsigned long timeout)
181 {
182 	/* timeout of 0 will disable the watchdog */
183 	return timeout ? timeout * HZ : MAX_SCHEDULE_TIMEOUT;
184 }
185 
186 /*
187  * Process updating of timeout sysctl
188  */
189 int proc_dohung_task_timeout_secs(struct ctl_table *table, int write,
190 				  void __user *buffer,
191 				  size_t *lenp, loff_t *ppos)
192 {
193 	int ret;
194 
195 	ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
196 
197 	if (ret || !write)
198 		goto out;
199 
200 	wake_up_process(watchdog_task);
201 
202  out:
203 	return ret;
204 }
205 
206 /*
207  * kthread which checks for tasks stuck in D state
208  */
209 static int watchdog(void *dummy)
210 {
211 	set_user_nice(current, 0);
212 
213 	for ( ; ; ) {
214 		unsigned long timeout = sysctl_hung_task_timeout_secs;
215 
216 		while (schedule_timeout_interruptible(timeout_jiffies(timeout)))
217 			timeout = sysctl_hung_task_timeout_secs;
218 
219 		check_hung_uninterruptible_tasks(timeout);
220 	}
221 
222 	return 0;
223 }
224 
225 static int __init hung_task_init(void)
226 {
227 	atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
228 	watchdog_task = kthread_run(watchdog, NULL, "khungtaskd");
229 
230 	return 0;
231 }
232 
233 module_init(hung_task_init);
234