xref: /openbmc/linux/kernel/hung_task.c (revision da2ef666)
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 #include <linux/sched/signal.h>
20 #include <linux/sched/debug.h>
21 
22 #include <trace/events/sched.h>
23 
24 /*
25  * The number of tasks checked:
26  */
27 int __read_mostly sysctl_hung_task_check_count = PID_MAX_LIMIT;
28 
29 /*
30  * Limit number of tasks checked in a batch.
31  *
32  * This value controls the preemptibility of khungtaskd since preemption
33  * is disabled during the critical section. It also controls the size of
34  * the RCU grace period. So it needs to be upper-bound.
35  */
36 #define HUNG_TASK_BATCHING 1024
37 
38 /*
39  * Zero means infinite timeout - no checking done:
40  */
41 unsigned long __read_mostly sysctl_hung_task_timeout_secs = CONFIG_DEFAULT_HUNG_TASK_TIMEOUT;
42 
43 /*
44  * Zero (default value) means use sysctl_hung_task_timeout_secs:
45  */
46 unsigned long __read_mostly sysctl_hung_task_check_interval_secs;
47 
48 int __read_mostly sysctl_hung_task_warnings = 10;
49 
50 static int __read_mostly did_panic;
51 static bool hung_task_show_lock;
52 static bool hung_task_call_panic;
53 
54 static struct task_struct *watchdog_task;
55 
56 /*
57  * Should we panic (and reboot, if panic_timeout= is set) when a
58  * hung task is detected:
59  */
60 unsigned int __read_mostly sysctl_hung_task_panic =
61 				CONFIG_BOOTPARAM_HUNG_TASK_PANIC_VALUE;
62 
63 static int __init hung_task_panic_setup(char *str)
64 {
65 	int rc = kstrtouint(str, 0, &sysctl_hung_task_panic);
66 
67 	if (rc)
68 		return rc;
69 	return 1;
70 }
71 __setup("hung_task_panic=", hung_task_panic_setup);
72 
73 static int
74 hung_task_panic(struct notifier_block *this, unsigned long event, void *ptr)
75 {
76 	did_panic = 1;
77 
78 	return NOTIFY_DONE;
79 }
80 
81 static struct notifier_block panic_block = {
82 	.notifier_call = hung_task_panic,
83 };
84 
85 static void check_hung_task(struct task_struct *t, unsigned long timeout)
86 {
87 	unsigned long switch_count = t->nvcsw + t->nivcsw;
88 
89 	/*
90 	 * Ensure the task is not frozen.
91 	 * Also, skip vfork and any other user process that freezer should skip.
92 	 */
93 	if (unlikely(t->flags & (PF_FROZEN | PF_FREEZER_SKIP)))
94 	    return;
95 
96 	/*
97 	 * When a freshly created task is scheduled once, changes its state to
98 	 * TASK_UNINTERRUPTIBLE without having ever been switched out once, it
99 	 * musn't be checked.
100 	 */
101 	if (unlikely(!switch_count))
102 		return;
103 
104 	if (switch_count != t->last_switch_count) {
105 		t->last_switch_count = switch_count;
106 		t->last_switch_time = jiffies;
107 		return;
108 	}
109 	if (time_is_after_jiffies(t->last_switch_time + timeout * HZ))
110 		return;
111 
112 	trace_sched_process_hang(t);
113 
114 	if (!sysctl_hung_task_warnings && !sysctl_hung_task_panic)
115 		return;
116 
117 	/*
118 	 * Ok, the task did not get scheduled for more than 2 minutes,
119 	 * complain:
120 	 */
121 	if (sysctl_hung_task_warnings) {
122 		if (sysctl_hung_task_warnings > 0)
123 			sysctl_hung_task_warnings--;
124 		pr_err("INFO: task %s:%d blocked for more than %ld seconds.\n",
125 			t->comm, t->pid, timeout);
126 		pr_err("      %s %s %.*s\n",
127 			print_tainted(), init_utsname()->release,
128 			(int)strcspn(init_utsname()->version, " "),
129 			init_utsname()->version);
130 		pr_err("\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
131 			" disables this message.\n");
132 		sched_show_task(t);
133 		hung_task_show_lock = true;
134 	}
135 
136 	touch_nmi_watchdog();
137 
138 	if (sysctl_hung_task_panic) {
139 		hung_task_show_lock = true;
140 		hung_task_call_panic = true;
141 	}
142 }
143 
144 /*
145  * To avoid extending the RCU grace period for an unbounded amount of time,
146  * periodically exit the critical section and enter a new one.
147  *
148  * For preemptible RCU it is sufficient to call rcu_read_unlock in order
149  * to exit the grace period. For classic RCU, a reschedule is required.
150  */
151 static bool rcu_lock_break(struct task_struct *g, struct task_struct *t)
152 {
153 	bool can_cont;
154 
155 	get_task_struct(g);
156 	get_task_struct(t);
157 	rcu_read_unlock();
158 	cond_resched();
159 	rcu_read_lock();
160 	can_cont = pid_alive(g) && pid_alive(t);
161 	put_task_struct(t);
162 	put_task_struct(g);
163 
164 	return can_cont;
165 }
166 
167 /*
168  * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
169  * a really long time (120 seconds). If that happens, print out
170  * a warning.
171  */
172 static void check_hung_uninterruptible_tasks(unsigned long timeout)
173 {
174 	int max_count = sysctl_hung_task_check_count;
175 	int batch_count = HUNG_TASK_BATCHING;
176 	struct task_struct *g, *t;
177 
178 	/*
179 	 * If the system crashed already then all bets are off,
180 	 * do not report extra hung tasks:
181 	 */
182 	if (test_taint(TAINT_DIE) || did_panic)
183 		return;
184 
185 	hung_task_show_lock = false;
186 	rcu_read_lock();
187 	for_each_process_thread(g, t) {
188 		if (!max_count--)
189 			goto unlock;
190 		if (!--batch_count) {
191 			batch_count = HUNG_TASK_BATCHING;
192 			if (!rcu_lock_break(g, t))
193 				goto unlock;
194 		}
195 		/* use "==" to skip the TASK_KILLABLE tasks waiting on NFS */
196 		if (t->state == TASK_UNINTERRUPTIBLE)
197 			check_hung_task(t, timeout);
198 	}
199  unlock:
200 	rcu_read_unlock();
201 	if (hung_task_show_lock)
202 		debug_show_all_locks();
203 	if (hung_task_call_panic) {
204 		trigger_all_cpu_backtrace();
205 		panic("hung_task: blocked tasks");
206 	}
207 }
208 
209 static long hung_timeout_jiffies(unsigned long last_checked,
210 				 unsigned long timeout)
211 {
212 	/* timeout of 0 will disable the watchdog */
213 	return timeout ? last_checked - jiffies + timeout * HZ :
214 		MAX_SCHEDULE_TIMEOUT;
215 }
216 
217 /*
218  * Process updating of timeout sysctl
219  */
220 int proc_dohung_task_timeout_secs(struct ctl_table *table, int write,
221 				  void __user *buffer,
222 				  size_t *lenp, loff_t *ppos)
223 {
224 	int ret;
225 
226 	ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
227 
228 	if (ret || !write)
229 		goto out;
230 
231 	wake_up_process(watchdog_task);
232 
233  out:
234 	return ret;
235 }
236 
237 static atomic_t reset_hung_task = ATOMIC_INIT(0);
238 
239 void reset_hung_task_detector(void)
240 {
241 	atomic_set(&reset_hung_task, 1);
242 }
243 EXPORT_SYMBOL_GPL(reset_hung_task_detector);
244 
245 /*
246  * kthread which checks for tasks stuck in D state
247  */
248 static int watchdog(void *dummy)
249 {
250 	unsigned long hung_last_checked = jiffies;
251 
252 	set_user_nice(current, 0);
253 
254 	for ( ; ; ) {
255 		unsigned long timeout = sysctl_hung_task_timeout_secs;
256 		unsigned long interval = sysctl_hung_task_check_interval_secs;
257 		long t;
258 
259 		if (interval == 0)
260 			interval = timeout;
261 		interval = min_t(unsigned long, interval, timeout);
262 		t = hung_timeout_jiffies(hung_last_checked, interval);
263 		if (t <= 0) {
264 			if (!atomic_xchg(&reset_hung_task, 0))
265 				check_hung_uninterruptible_tasks(timeout);
266 			hung_last_checked = jiffies;
267 			continue;
268 		}
269 		schedule_timeout_interruptible(t);
270 	}
271 
272 	return 0;
273 }
274 
275 static int __init hung_task_init(void)
276 {
277 	atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
278 	watchdog_task = kthread_run(watchdog, NULL, "khungtaskd");
279 
280 	return 0;
281 }
282 subsys_initcall(hung_task_init);
283