xref: /openbmc/linux/kernel/hung_task.c (revision b868a02e)
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/panic_notifier.h>
19 #include <linux/sysctl.h>
20 #include <linux/suspend.h>
21 #include <linux/utsname.h>
22 #include <linux/sched/signal.h>
23 #include <linux/sched/debug.h>
24 #include <linux/sched/sysctl.h>
25 
26 #include <trace/events/sched.h>
27 
28 /*
29  * The number of tasks checked:
30  */
31 int __read_mostly sysctl_hung_task_check_count = PID_MAX_LIMIT;
32 
33 /*
34  * Limit number of tasks checked in a batch.
35  *
36  * This value controls the preemptibility of khungtaskd since preemption
37  * is disabled during the critical section. It also controls the size of
38  * the RCU grace period. So it needs to be upper-bound.
39  */
40 #define HUNG_TASK_LOCK_BREAK (HZ / 10)
41 
42 /*
43  * Zero means infinite timeout - no checking done:
44  */
45 unsigned long __read_mostly sysctl_hung_task_timeout_secs = CONFIG_DEFAULT_HUNG_TASK_TIMEOUT;
46 
47 /*
48  * Zero (default value) means use sysctl_hung_task_timeout_secs:
49  */
50 unsigned long __read_mostly sysctl_hung_task_check_interval_secs;
51 
52 int __read_mostly sysctl_hung_task_warnings = 10;
53 
54 static int __read_mostly did_panic;
55 static bool hung_task_show_lock;
56 static bool hung_task_call_panic;
57 static bool hung_task_show_all_bt;
58 
59 static struct task_struct *watchdog_task;
60 
61 #ifdef CONFIG_SMP
62 /*
63  * Should we dump all CPUs backtraces in a hung task event?
64  * Defaults to 0, can be changed via sysctl.
65  */
66 static unsigned int __read_mostly sysctl_hung_task_all_cpu_backtrace;
67 #else
68 #define sysctl_hung_task_all_cpu_backtrace 0
69 #endif /* CONFIG_SMP */
70 
71 /*
72  * Should we panic (and reboot, if panic_timeout= is set) when a
73  * hung task is detected:
74  */
75 unsigned int __read_mostly sysctl_hung_task_panic =
76 				IS_ENABLED(CONFIG_BOOTPARAM_HUNG_TASK_PANIC);
77 
78 static int
79 hung_task_panic(struct notifier_block *this, unsigned long event, void *ptr)
80 {
81 	did_panic = 1;
82 
83 	return NOTIFY_DONE;
84 }
85 
86 static struct notifier_block panic_block = {
87 	.notifier_call = hung_task_panic,
88 };
89 
90 static void check_hung_task(struct task_struct *t, unsigned long timeout)
91 {
92 	unsigned long switch_count = t->nvcsw + t->nivcsw;
93 
94 	/*
95 	 * Ensure the task is not frozen.
96 	 * Also, skip vfork and any other user process that freezer should skip.
97 	 */
98 	if (unlikely(READ_ONCE(t->__state) & TASK_FROZEN))
99 		return;
100 
101 	/*
102 	 * When a freshly created task is scheduled once, changes its state to
103 	 * TASK_UNINTERRUPTIBLE without having ever been switched out once, it
104 	 * musn't be checked.
105 	 */
106 	if (unlikely(!switch_count))
107 		return;
108 
109 	if (switch_count != t->last_switch_count) {
110 		t->last_switch_count = switch_count;
111 		t->last_switch_time = jiffies;
112 		return;
113 	}
114 	if (time_is_after_jiffies(t->last_switch_time + timeout * HZ))
115 		return;
116 
117 	trace_sched_process_hang(t);
118 
119 	if (sysctl_hung_task_panic) {
120 		console_verbose();
121 		hung_task_show_lock = true;
122 		hung_task_call_panic = true;
123 	}
124 
125 	/*
126 	 * Ok, the task did not get scheduled for more than 2 minutes,
127 	 * complain:
128 	 */
129 	if (sysctl_hung_task_warnings) {
130 		if (sysctl_hung_task_warnings > 0)
131 			sysctl_hung_task_warnings--;
132 		pr_err("INFO: task %s:%d blocked for more than %ld seconds.\n",
133 		       t->comm, t->pid, (jiffies - t->last_switch_time) / HZ);
134 		pr_err("      %s %s %.*s\n",
135 			print_tainted(), init_utsname()->release,
136 			(int)strcspn(init_utsname()->version, " "),
137 			init_utsname()->version);
138 		pr_err("\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
139 			" disables this message.\n");
140 		sched_show_task(t);
141 		hung_task_show_lock = true;
142 
143 		if (sysctl_hung_task_all_cpu_backtrace)
144 			hung_task_show_all_bt = true;
145 	}
146 
147 	touch_nmi_watchdog();
148 }
149 
150 /*
151  * To avoid extending the RCU grace period for an unbounded amount of time,
152  * periodically exit the critical section and enter a new one.
153  *
154  * For preemptible RCU it is sufficient to call rcu_read_unlock in order
155  * to exit the grace period. For classic RCU, a reschedule is required.
156  */
157 static bool rcu_lock_break(struct task_struct *g, struct task_struct *t)
158 {
159 	bool can_cont;
160 
161 	get_task_struct(g);
162 	get_task_struct(t);
163 	rcu_read_unlock();
164 	cond_resched();
165 	rcu_read_lock();
166 	can_cont = pid_alive(g) && pid_alive(t);
167 	put_task_struct(t);
168 	put_task_struct(g);
169 
170 	return can_cont;
171 }
172 
173 /*
174  * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
175  * a really long time (120 seconds). If that happens, print out
176  * a warning.
177  */
178 static void check_hung_uninterruptible_tasks(unsigned long timeout)
179 {
180 	int max_count = sysctl_hung_task_check_count;
181 	unsigned long last_break = jiffies;
182 	struct task_struct *g, *t;
183 
184 	/*
185 	 * If the system crashed already then all bets are off,
186 	 * do not report extra hung tasks:
187 	 */
188 	if (test_taint(TAINT_DIE) || did_panic)
189 		return;
190 
191 	hung_task_show_lock = false;
192 	rcu_read_lock();
193 	for_each_process_thread(g, t) {
194 		unsigned int state;
195 
196 		if (!max_count--)
197 			goto unlock;
198 		if (time_after(jiffies, last_break + HUNG_TASK_LOCK_BREAK)) {
199 			if (!rcu_lock_break(g, t))
200 				goto unlock;
201 			last_break = jiffies;
202 		}
203 		/*
204 		 * skip the TASK_KILLABLE tasks -- these can be killed
205 		 * skip the TASK_IDLE tasks -- those are genuinely idle
206 		 */
207 		state = READ_ONCE(t->__state);
208 		if ((state & TASK_UNINTERRUPTIBLE) &&
209 		    !(state & TASK_WAKEKILL) &&
210 		    !(state & TASK_NOLOAD))
211 			check_hung_task(t, timeout);
212 	}
213  unlock:
214 	rcu_read_unlock();
215 	if (hung_task_show_lock)
216 		debug_show_all_locks();
217 
218 	if (hung_task_show_all_bt) {
219 		hung_task_show_all_bt = false;
220 		trigger_all_cpu_backtrace();
221 	}
222 
223 	if (hung_task_call_panic)
224 		panic("hung_task: blocked tasks");
225 }
226 
227 static long hung_timeout_jiffies(unsigned long last_checked,
228 				 unsigned long timeout)
229 {
230 	/* timeout of 0 will disable the watchdog */
231 	return timeout ? last_checked - jiffies + timeout * HZ :
232 		MAX_SCHEDULE_TIMEOUT;
233 }
234 
235 #ifdef CONFIG_SYSCTL
236 /*
237  * Process updating of timeout sysctl
238  */
239 static int proc_dohung_task_timeout_secs(struct ctl_table *table, int write,
240 				  void *buffer,
241 				  size_t *lenp, loff_t *ppos)
242 {
243 	int ret;
244 
245 	ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
246 
247 	if (ret || !write)
248 		goto out;
249 
250 	wake_up_process(watchdog_task);
251 
252  out:
253 	return ret;
254 }
255 
256 /*
257  * This is needed for proc_doulongvec_minmax of sysctl_hung_task_timeout_secs
258  * and hung_task_check_interval_secs
259  */
260 static const unsigned long hung_task_timeout_max = (LONG_MAX / HZ);
261 static struct ctl_table hung_task_sysctls[] = {
262 #ifdef CONFIG_SMP
263 	{
264 		.procname	= "hung_task_all_cpu_backtrace",
265 		.data		= &sysctl_hung_task_all_cpu_backtrace,
266 		.maxlen		= sizeof(int),
267 		.mode		= 0644,
268 		.proc_handler	= proc_dointvec_minmax,
269 		.extra1		= SYSCTL_ZERO,
270 		.extra2		= SYSCTL_ONE,
271 	},
272 #endif /* CONFIG_SMP */
273 	{
274 		.procname	= "hung_task_panic",
275 		.data		= &sysctl_hung_task_panic,
276 		.maxlen		= sizeof(int),
277 		.mode		= 0644,
278 		.proc_handler	= proc_dointvec_minmax,
279 		.extra1		= SYSCTL_ZERO,
280 		.extra2		= SYSCTL_ONE,
281 	},
282 	{
283 		.procname	= "hung_task_check_count",
284 		.data		= &sysctl_hung_task_check_count,
285 		.maxlen		= sizeof(int),
286 		.mode		= 0644,
287 		.proc_handler	= proc_dointvec_minmax,
288 		.extra1		= SYSCTL_ZERO,
289 	},
290 	{
291 		.procname	= "hung_task_timeout_secs",
292 		.data		= &sysctl_hung_task_timeout_secs,
293 		.maxlen		= sizeof(unsigned long),
294 		.mode		= 0644,
295 		.proc_handler	= proc_dohung_task_timeout_secs,
296 		.extra2		= (void *)&hung_task_timeout_max,
297 	},
298 	{
299 		.procname	= "hung_task_check_interval_secs",
300 		.data		= &sysctl_hung_task_check_interval_secs,
301 		.maxlen		= sizeof(unsigned long),
302 		.mode		= 0644,
303 		.proc_handler	= proc_dohung_task_timeout_secs,
304 		.extra2		= (void *)&hung_task_timeout_max,
305 	},
306 	{
307 		.procname	= "hung_task_warnings",
308 		.data		= &sysctl_hung_task_warnings,
309 		.maxlen		= sizeof(int),
310 		.mode		= 0644,
311 		.proc_handler	= proc_dointvec_minmax,
312 		.extra1		= SYSCTL_NEG_ONE,
313 	},
314 	{}
315 };
316 
317 static void __init hung_task_sysctl_init(void)
318 {
319 	register_sysctl_init("kernel", hung_task_sysctls);
320 }
321 #else
322 #define hung_task_sysctl_init() do { } while (0)
323 #endif /* CONFIG_SYSCTL */
324 
325 
326 static atomic_t reset_hung_task = ATOMIC_INIT(0);
327 
328 void reset_hung_task_detector(void)
329 {
330 	atomic_set(&reset_hung_task, 1);
331 }
332 EXPORT_SYMBOL_GPL(reset_hung_task_detector);
333 
334 static bool hung_detector_suspended;
335 
336 static int hungtask_pm_notify(struct notifier_block *self,
337 			      unsigned long action, void *hcpu)
338 {
339 	switch (action) {
340 	case PM_SUSPEND_PREPARE:
341 	case PM_HIBERNATION_PREPARE:
342 	case PM_RESTORE_PREPARE:
343 		hung_detector_suspended = true;
344 		break;
345 	case PM_POST_SUSPEND:
346 	case PM_POST_HIBERNATION:
347 	case PM_POST_RESTORE:
348 		hung_detector_suspended = false;
349 		break;
350 	default:
351 		break;
352 	}
353 	return NOTIFY_OK;
354 }
355 
356 /*
357  * kthread which checks for tasks stuck in D state
358  */
359 static int watchdog(void *dummy)
360 {
361 	unsigned long hung_last_checked = jiffies;
362 
363 	set_user_nice(current, 0);
364 
365 	for ( ; ; ) {
366 		unsigned long timeout = sysctl_hung_task_timeout_secs;
367 		unsigned long interval = sysctl_hung_task_check_interval_secs;
368 		long t;
369 
370 		if (interval == 0)
371 			interval = timeout;
372 		interval = min_t(unsigned long, interval, timeout);
373 		t = hung_timeout_jiffies(hung_last_checked, interval);
374 		if (t <= 0) {
375 			if (!atomic_xchg(&reset_hung_task, 0) &&
376 			    !hung_detector_suspended)
377 				check_hung_uninterruptible_tasks(timeout);
378 			hung_last_checked = jiffies;
379 			continue;
380 		}
381 		schedule_timeout_interruptible(t);
382 	}
383 
384 	return 0;
385 }
386 
387 static int __init hung_task_init(void)
388 {
389 	atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
390 
391 	/* Disable hung task detector on suspend */
392 	pm_notifier(hungtask_pm_notify, 0);
393 
394 	watchdog_task = kthread_run(watchdog, NULL, "khungtaskd");
395 	hung_task_sysctl_init();
396 
397 	return 0;
398 }
399 subsys_initcall(hung_task_init);
400