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 static 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 static unsigned long __read_mostly sysctl_hung_task_check_interval_secs;
51 
52 static 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 static unsigned int __read_mostly sysctl_hung_task_panic =
76 	IS_ENABLED(CONFIG_BOOTPARAM_HUNG_TASK_PANIC);
77 
78 static int
hung_task_panic(struct notifier_block * this,unsigned long event,void * ptr)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 
check_hung_task(struct task_struct * t,unsigned long timeout)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 		if (!sysctl_hung_task_warnings)
146 			pr_info("Future hung task reports are suppressed, see sysctl kernel.hung_task_warnings\n");
147 	}
148 
149 	touch_nmi_watchdog();
150 }
151 
152 /*
153  * To avoid extending the RCU grace period for an unbounded amount of time,
154  * periodically exit the critical section and enter a new one.
155  *
156  * For preemptible RCU it is sufficient to call rcu_read_unlock in order
157  * to exit the grace period. For classic RCU, a reschedule is required.
158  */
rcu_lock_break(struct task_struct * g,struct task_struct * t)159 static bool rcu_lock_break(struct task_struct *g, struct task_struct *t)
160 {
161 	bool can_cont;
162 
163 	get_task_struct(g);
164 	get_task_struct(t);
165 	rcu_read_unlock();
166 	cond_resched();
167 	rcu_read_lock();
168 	can_cont = pid_alive(g) && pid_alive(t);
169 	put_task_struct(t);
170 	put_task_struct(g);
171 
172 	return can_cont;
173 }
174 
175 /*
176  * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
177  * a really long time (120 seconds). If that happens, print out
178  * a warning.
179  */
check_hung_uninterruptible_tasks(unsigned long timeout)180 static void check_hung_uninterruptible_tasks(unsigned long timeout)
181 {
182 	int max_count = sysctl_hung_task_check_count;
183 	unsigned long last_break = jiffies;
184 	struct task_struct *g, *t;
185 
186 	/*
187 	 * If the system crashed already then all bets are off,
188 	 * do not report extra hung tasks:
189 	 */
190 	if (test_taint(TAINT_DIE) || did_panic)
191 		return;
192 
193 	hung_task_show_lock = false;
194 	rcu_read_lock();
195 	for_each_process_thread(g, t) {
196 		unsigned int state;
197 
198 		if (!max_count--)
199 			goto unlock;
200 		if (time_after(jiffies, last_break + HUNG_TASK_LOCK_BREAK)) {
201 			if (!rcu_lock_break(g, t))
202 				goto unlock;
203 			last_break = jiffies;
204 		}
205 		/*
206 		 * skip the TASK_KILLABLE tasks -- these can be killed
207 		 * skip the TASK_IDLE tasks -- those are genuinely idle
208 		 */
209 		state = READ_ONCE(t->__state);
210 		if ((state & TASK_UNINTERRUPTIBLE) &&
211 		    !(state & TASK_WAKEKILL) &&
212 		    !(state & TASK_NOLOAD))
213 			check_hung_task(t, timeout);
214 	}
215  unlock:
216 	rcu_read_unlock();
217 	if (hung_task_show_lock)
218 		debug_show_all_locks();
219 
220 	if (hung_task_show_all_bt) {
221 		hung_task_show_all_bt = false;
222 		trigger_all_cpu_backtrace();
223 	}
224 
225 	if (hung_task_call_panic)
226 		panic("hung_task: blocked tasks");
227 }
228 
hung_timeout_jiffies(unsigned long last_checked,unsigned long timeout)229 static long hung_timeout_jiffies(unsigned long last_checked,
230 				 unsigned long timeout)
231 {
232 	/* timeout of 0 will disable the watchdog */
233 	return timeout ? last_checked - jiffies + timeout * HZ :
234 		MAX_SCHEDULE_TIMEOUT;
235 }
236 
237 #ifdef CONFIG_SYSCTL
238 /*
239  * Process updating of timeout sysctl
240  */
proc_dohung_task_timeout_secs(struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)241 static int proc_dohung_task_timeout_secs(struct ctl_table *table, int write,
242 				  void *buffer,
243 				  size_t *lenp, loff_t *ppos)
244 {
245 	int ret;
246 
247 	ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
248 
249 	if (ret || !write)
250 		goto out;
251 
252 	wake_up_process(watchdog_task);
253 
254  out:
255 	return ret;
256 }
257 
258 /*
259  * This is needed for proc_doulongvec_minmax of sysctl_hung_task_timeout_secs
260  * and hung_task_check_interval_secs
261  */
262 static const unsigned long hung_task_timeout_max = (LONG_MAX / HZ);
263 static struct ctl_table hung_task_sysctls[] = {
264 #ifdef CONFIG_SMP
265 	{
266 		.procname	= "hung_task_all_cpu_backtrace",
267 		.data		= &sysctl_hung_task_all_cpu_backtrace,
268 		.maxlen		= sizeof(int),
269 		.mode		= 0644,
270 		.proc_handler	= proc_dointvec_minmax,
271 		.extra1		= SYSCTL_ZERO,
272 		.extra2		= SYSCTL_ONE,
273 	},
274 #endif /* CONFIG_SMP */
275 	{
276 		.procname	= "hung_task_panic",
277 		.data		= &sysctl_hung_task_panic,
278 		.maxlen		= sizeof(int),
279 		.mode		= 0644,
280 		.proc_handler	= proc_dointvec_minmax,
281 		.extra1		= SYSCTL_ZERO,
282 		.extra2		= SYSCTL_ONE,
283 	},
284 	{
285 		.procname	= "hung_task_check_count",
286 		.data		= &sysctl_hung_task_check_count,
287 		.maxlen		= sizeof(int),
288 		.mode		= 0644,
289 		.proc_handler	= proc_dointvec_minmax,
290 		.extra1		= SYSCTL_ZERO,
291 	},
292 	{
293 		.procname	= "hung_task_timeout_secs",
294 		.data		= &sysctl_hung_task_timeout_secs,
295 		.maxlen		= sizeof(unsigned long),
296 		.mode		= 0644,
297 		.proc_handler	= proc_dohung_task_timeout_secs,
298 		.extra2		= (void *)&hung_task_timeout_max,
299 	},
300 	{
301 		.procname	= "hung_task_check_interval_secs",
302 		.data		= &sysctl_hung_task_check_interval_secs,
303 		.maxlen		= sizeof(unsigned long),
304 		.mode		= 0644,
305 		.proc_handler	= proc_dohung_task_timeout_secs,
306 		.extra2		= (void *)&hung_task_timeout_max,
307 	},
308 	{
309 		.procname	= "hung_task_warnings",
310 		.data		= &sysctl_hung_task_warnings,
311 		.maxlen		= sizeof(int),
312 		.mode		= 0644,
313 		.proc_handler	= proc_dointvec_minmax,
314 		.extra1		= SYSCTL_NEG_ONE,
315 	},
316 	{}
317 };
318 
hung_task_sysctl_init(void)319 static void __init hung_task_sysctl_init(void)
320 {
321 	register_sysctl_init("kernel", hung_task_sysctls);
322 }
323 #else
324 #define hung_task_sysctl_init() do { } while (0)
325 #endif /* CONFIG_SYSCTL */
326 
327 
328 static atomic_t reset_hung_task = ATOMIC_INIT(0);
329 
reset_hung_task_detector(void)330 void reset_hung_task_detector(void)
331 {
332 	atomic_set(&reset_hung_task, 1);
333 }
334 EXPORT_SYMBOL_GPL(reset_hung_task_detector);
335 
336 static bool hung_detector_suspended;
337 
hungtask_pm_notify(struct notifier_block * self,unsigned long action,void * hcpu)338 static int hungtask_pm_notify(struct notifier_block *self,
339 			      unsigned long action, void *hcpu)
340 {
341 	switch (action) {
342 	case PM_SUSPEND_PREPARE:
343 	case PM_HIBERNATION_PREPARE:
344 	case PM_RESTORE_PREPARE:
345 		hung_detector_suspended = true;
346 		break;
347 	case PM_POST_SUSPEND:
348 	case PM_POST_HIBERNATION:
349 	case PM_POST_RESTORE:
350 		hung_detector_suspended = false;
351 		break;
352 	default:
353 		break;
354 	}
355 	return NOTIFY_OK;
356 }
357 
358 /*
359  * kthread which checks for tasks stuck in D state
360  */
watchdog(void * dummy)361 static int watchdog(void *dummy)
362 {
363 	unsigned long hung_last_checked = jiffies;
364 
365 	set_user_nice(current, 0);
366 
367 	for ( ; ; ) {
368 		unsigned long timeout = sysctl_hung_task_timeout_secs;
369 		unsigned long interval = sysctl_hung_task_check_interval_secs;
370 		long t;
371 
372 		if (interval == 0)
373 			interval = timeout;
374 		interval = min_t(unsigned long, interval, timeout);
375 		t = hung_timeout_jiffies(hung_last_checked, interval);
376 		if (t <= 0) {
377 			if (!atomic_xchg(&reset_hung_task, 0) &&
378 			    !hung_detector_suspended)
379 				check_hung_uninterruptible_tasks(timeout);
380 			hung_last_checked = jiffies;
381 			continue;
382 		}
383 		schedule_timeout_interruptible(t);
384 	}
385 
386 	return 0;
387 }
388 
hung_task_init(void)389 static int __init hung_task_init(void)
390 {
391 	atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
392 
393 	/* Disable hung task detector on suspend */
394 	pm_notifier(hungtask_pm_notify, 0);
395 
396 	watchdog_task = kthread_run(watchdog, NULL, "khungtaskd");
397 	hung_task_sysctl_init();
398 
399 	return 0;
400 }
401 subsys_initcall(hung_task_init);
402