xref: /openbmc/linux/kernel/hung_task.c (revision 1406f0f3)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Detect Hung Task
4  *
5  * kernel/hung_task.c - kernel thread for detecting tasks stuck in D state
6  *
7  */
8 
9 #include <linux/mm.h>
10 #include <linux/cpu.h>
11 #include <linux/nmi.h>
12 #include <linux/init.h>
13 #include <linux/delay.h>
14 #include <linux/freezer.h>
15 #include <linux/kthread.h>
16 #include <linux/lockdep.h>
17 #include <linux/export.h>
18 #include <linux/sysctl.h>
19 #include <linux/suspend.h>
20 #include <linux/utsname.h>
21 #include <linux/sched/signal.h>
22 #include <linux/sched/debug.h>
23 #include <linux/sched/sysctl.h>
24 
25 #include <trace/events/sched.h>
26 
27 /*
28  * The number of tasks checked:
29  */
30 int __read_mostly sysctl_hung_task_check_count = PID_MAX_LIMIT;
31 
32 /*
33  * Limit number of tasks checked in a batch.
34  *
35  * This value controls the preemptibility of khungtaskd since preemption
36  * is disabled during the critical section. It also controls the size of
37  * the RCU grace period. So it needs to be upper-bound.
38  */
39 #define HUNG_TASK_LOCK_BREAK (HZ / 10)
40 
41 /*
42  * Zero means infinite timeout - no checking done:
43  */
44 unsigned long __read_mostly sysctl_hung_task_timeout_secs = CONFIG_DEFAULT_HUNG_TASK_TIMEOUT;
45 
46 /*
47  * Zero (default value) means use sysctl_hung_task_timeout_secs:
48  */
49 unsigned long __read_mostly sysctl_hung_task_check_interval_secs;
50 
51 int __read_mostly sysctl_hung_task_warnings = 10;
52 
53 static int __read_mostly did_panic;
54 static bool hung_task_show_lock;
55 static bool hung_task_call_panic;
56 static bool hung_task_show_all_bt;
57 
58 static struct task_struct *watchdog_task;
59 
60 #ifdef CONFIG_SMP
61 /*
62  * Should we dump all CPUs backtraces in a hung task event?
63  * Defaults to 0, can be changed via sysctl.
64  */
65 unsigned int __read_mostly sysctl_hung_task_all_cpu_backtrace;
66 #endif /* CONFIG_SMP */
67 
68 /*
69  * Should we panic (and reboot, if panic_timeout= is set) when a
70  * hung task is detected:
71  */
72 unsigned int __read_mostly sysctl_hung_task_panic =
73 				CONFIG_BOOTPARAM_HUNG_TASK_PANIC_VALUE;
74 
75 static int
76 hung_task_panic(struct notifier_block *this, unsigned long event, void *ptr)
77 {
78 	did_panic = 1;
79 
80 	return NOTIFY_DONE;
81 }
82 
83 static struct notifier_block panic_block = {
84 	.notifier_call = hung_task_panic,
85 };
86 
87 static void check_hung_task(struct task_struct *t, unsigned long timeout)
88 {
89 	unsigned long switch_count = t->nvcsw + t->nivcsw;
90 
91 	/*
92 	 * Ensure the task is not frozen.
93 	 * Also, skip vfork and any other user process that freezer should skip.
94 	 */
95 	if (unlikely(t->flags & (PF_FROZEN | PF_FREEZER_SKIP)))
96 	    return;
97 
98 	/*
99 	 * When a freshly created task is scheduled once, changes its state to
100 	 * TASK_UNINTERRUPTIBLE without having ever been switched out once, it
101 	 * musn't be checked.
102 	 */
103 	if (unlikely(!switch_count))
104 		return;
105 
106 	if (switch_count != t->last_switch_count) {
107 		t->last_switch_count = switch_count;
108 		t->last_switch_time = jiffies;
109 		return;
110 	}
111 	if (time_is_after_jiffies(t->last_switch_time + timeout * HZ))
112 		return;
113 
114 	trace_sched_process_hang(t);
115 
116 	if (sysctl_hung_task_panic) {
117 		console_verbose();
118 		hung_task_show_lock = true;
119 		hung_task_call_panic = true;
120 	}
121 
122 	/*
123 	 * Ok, the task did not get scheduled for more than 2 minutes,
124 	 * complain:
125 	 */
126 	if (sysctl_hung_task_warnings) {
127 		if (sysctl_hung_task_warnings > 0)
128 			sysctl_hung_task_warnings--;
129 		pr_err("INFO: task %s:%d blocked for more than %ld seconds.\n",
130 		       t->comm, t->pid, (jiffies - t->last_switch_time) / HZ);
131 		pr_err("      %s %s %.*s\n",
132 			print_tainted(), init_utsname()->release,
133 			(int)strcspn(init_utsname()->version, " "),
134 			init_utsname()->version);
135 		pr_err("\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
136 			" disables this message.\n");
137 		sched_show_task(t);
138 		hung_task_show_lock = true;
139 
140 		if (sysctl_hung_task_all_cpu_backtrace)
141 			hung_task_show_all_bt = true;
142 	}
143 
144 	touch_nmi_watchdog();
145 }
146 
147 /*
148  * To avoid extending the RCU grace period for an unbounded amount of time,
149  * periodically exit the critical section and enter a new one.
150  *
151  * For preemptible RCU it is sufficient to call rcu_read_unlock in order
152  * to exit the grace period. For classic RCU, a reschedule is required.
153  */
154 static bool rcu_lock_break(struct task_struct *g, struct task_struct *t)
155 {
156 	bool can_cont;
157 
158 	get_task_struct(g);
159 	get_task_struct(t);
160 	rcu_read_unlock();
161 	cond_resched();
162 	rcu_read_lock();
163 	can_cont = pid_alive(g) && pid_alive(t);
164 	put_task_struct(t);
165 	put_task_struct(g);
166 
167 	return can_cont;
168 }
169 
170 /*
171  * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
172  * a really long time (120 seconds). If that happens, print out
173  * a warning.
174  */
175 static void check_hung_uninterruptible_tasks(unsigned long timeout)
176 {
177 	int max_count = sysctl_hung_task_check_count;
178 	unsigned long last_break = jiffies;
179 	struct task_struct *g, *t;
180 
181 	/*
182 	 * If the system crashed already then all bets are off,
183 	 * do not report extra hung tasks:
184 	 */
185 	if (test_taint(TAINT_DIE) || did_panic)
186 		return;
187 
188 	hung_task_show_lock = false;
189 	rcu_read_lock();
190 	for_each_process_thread(g, t) {
191 		if (!max_count--)
192 			goto unlock;
193 		if (time_after(jiffies, last_break + HUNG_TASK_LOCK_BREAK)) {
194 			if (!rcu_lock_break(g, t))
195 				goto unlock;
196 			last_break = jiffies;
197 		}
198 		/* use "==" to skip the TASK_KILLABLE tasks waiting on NFS */
199 		if (t->state == TASK_UNINTERRUPTIBLE)
200 			check_hung_task(t, timeout);
201 	}
202  unlock:
203 	rcu_read_unlock();
204 	if (hung_task_show_lock)
205 		debug_show_all_locks();
206 
207 	if (hung_task_show_all_bt) {
208 		hung_task_show_all_bt = false;
209 		trigger_all_cpu_backtrace();
210 	}
211 
212 	if (hung_task_call_panic)
213 		panic("hung_task: blocked tasks");
214 }
215 
216 static long hung_timeout_jiffies(unsigned long last_checked,
217 				 unsigned long timeout)
218 {
219 	/* timeout of 0 will disable the watchdog */
220 	return timeout ? last_checked - jiffies + timeout * HZ :
221 		MAX_SCHEDULE_TIMEOUT;
222 }
223 
224 /*
225  * Process updating of timeout sysctl
226  */
227 int proc_dohung_task_timeout_secs(struct ctl_table *table, int write,
228 				  void *buffer, size_t *lenp, loff_t *ppos)
229 {
230 	int ret;
231 
232 	ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
233 
234 	if (ret || !write)
235 		goto out;
236 
237 	wake_up_process(watchdog_task);
238 
239  out:
240 	return ret;
241 }
242 
243 static atomic_t reset_hung_task = ATOMIC_INIT(0);
244 
245 void reset_hung_task_detector(void)
246 {
247 	atomic_set(&reset_hung_task, 1);
248 }
249 EXPORT_SYMBOL_GPL(reset_hung_task_detector);
250 
251 static bool hung_detector_suspended;
252 
253 static int hungtask_pm_notify(struct notifier_block *self,
254 			      unsigned long action, void *hcpu)
255 {
256 	switch (action) {
257 	case PM_SUSPEND_PREPARE:
258 	case PM_HIBERNATION_PREPARE:
259 	case PM_RESTORE_PREPARE:
260 		hung_detector_suspended = true;
261 		break;
262 	case PM_POST_SUSPEND:
263 	case PM_POST_HIBERNATION:
264 	case PM_POST_RESTORE:
265 		hung_detector_suspended = false;
266 		break;
267 	default:
268 		break;
269 	}
270 	return NOTIFY_OK;
271 }
272 
273 /*
274  * kthread which checks for tasks stuck in D state
275  */
276 static int watchdog(void *dummy)
277 {
278 	unsigned long hung_last_checked = jiffies;
279 
280 	set_user_nice(current, 0);
281 
282 	for ( ; ; ) {
283 		unsigned long timeout = sysctl_hung_task_timeout_secs;
284 		unsigned long interval = sysctl_hung_task_check_interval_secs;
285 		long t;
286 
287 		if (interval == 0)
288 			interval = timeout;
289 		interval = min_t(unsigned long, interval, timeout);
290 		t = hung_timeout_jiffies(hung_last_checked, interval);
291 		if (t <= 0) {
292 			if (!atomic_xchg(&reset_hung_task, 0) &&
293 			    !hung_detector_suspended)
294 				check_hung_uninterruptible_tasks(timeout);
295 			hung_last_checked = jiffies;
296 			continue;
297 		}
298 		schedule_timeout_interruptible(t);
299 	}
300 
301 	return 0;
302 }
303 
304 static int __init hung_task_init(void)
305 {
306 	atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
307 
308 	/* Disable hung task detector on suspend */
309 	pm_notifier(hungtask_pm_notify, 0);
310 
311 	watchdog_task = kthread_run(watchdog, NULL, "khungtaskd");
312 
313 	return 0;
314 }
315 subsys_initcall(hung_task_init);
316