xref: /linux/kernel/hung_task.c (revision 7f71507851fc7764b36a3221839607d3a45c2025)
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  * Total number of tasks detected as hung since boot:
35  */
36 static unsigned long __read_mostly sysctl_hung_task_detect_count;
37 
38 /*
39  * Limit number of tasks checked in a batch.
40  *
41  * This value controls the preemptibility of khungtaskd since preemption
42  * is disabled during the critical section. It also controls the size of
43  * the RCU grace period. So it needs to be upper-bound.
44  */
45 #define HUNG_TASK_LOCK_BREAK (HZ / 10)
46 
47 /*
48  * Zero means infinite timeout - no checking done:
49  */
50 unsigned long __read_mostly sysctl_hung_task_timeout_secs = CONFIG_DEFAULT_HUNG_TASK_TIMEOUT;
51 EXPORT_SYMBOL_GPL(sysctl_hung_task_timeout_secs);
52 
53 /*
54  * Zero (default value) means use sysctl_hung_task_timeout_secs:
55  */
56 static unsigned long __read_mostly sysctl_hung_task_check_interval_secs;
57 
58 static int __read_mostly sysctl_hung_task_warnings = 10;
59 
60 static int __read_mostly did_panic;
61 static bool hung_task_show_lock;
62 static bool hung_task_call_panic;
63 static bool hung_task_show_all_bt;
64 
65 static struct task_struct *watchdog_task;
66 
67 #ifdef CONFIG_SMP
68 /*
69  * Should we dump all CPUs backtraces in a hung task event?
70  * Defaults to 0, can be changed via sysctl.
71  */
72 static unsigned int __read_mostly sysctl_hung_task_all_cpu_backtrace;
73 #else
74 #define sysctl_hung_task_all_cpu_backtrace 0
75 #endif /* CONFIG_SMP */
76 
77 /*
78  * Should we panic (and reboot, if panic_timeout= is set) when a
79  * hung task is detected:
80  */
81 static unsigned int __read_mostly sysctl_hung_task_panic =
82 	IS_ENABLED(CONFIG_BOOTPARAM_HUNG_TASK_PANIC);
83 
84 static int
85 hung_task_panic(struct notifier_block *this, unsigned long event, void *ptr)
86 {
87 	did_panic = 1;
88 
89 	return NOTIFY_DONE;
90 }
91 
92 static struct notifier_block panic_block = {
93 	.notifier_call = hung_task_panic,
94 };
95 
96 static void check_hung_task(struct task_struct *t, unsigned long timeout)
97 {
98 	unsigned long switch_count = t->nvcsw + t->nivcsw;
99 
100 	/*
101 	 * Ensure the task is not frozen.
102 	 * Also, skip vfork and any other user process that freezer should skip.
103 	 */
104 	if (unlikely(READ_ONCE(t->__state) & TASK_FROZEN))
105 		return;
106 
107 	/*
108 	 * When a freshly created task is scheduled once, changes its state to
109 	 * TASK_UNINTERRUPTIBLE without having ever been switched out once, it
110 	 * musn't be checked.
111 	 */
112 	if (unlikely(!switch_count))
113 		return;
114 
115 	if (switch_count != t->last_switch_count) {
116 		t->last_switch_count = switch_count;
117 		t->last_switch_time = jiffies;
118 		return;
119 	}
120 	if (time_is_after_jiffies(t->last_switch_time + timeout * HZ))
121 		return;
122 
123 	/*
124 	 * This counter tracks the total number of tasks detected as hung
125 	 * since boot.
126 	 */
127 	sysctl_hung_task_detect_count++;
128 
129 	trace_sched_process_hang(t);
130 
131 	if (sysctl_hung_task_panic) {
132 		console_verbose();
133 		hung_task_show_lock = true;
134 		hung_task_call_panic = true;
135 	}
136 
137 	/*
138 	 * Ok, the task did not get scheduled for more than 2 minutes,
139 	 * complain:
140 	 */
141 	if (sysctl_hung_task_warnings || hung_task_call_panic) {
142 		if (sysctl_hung_task_warnings > 0)
143 			sysctl_hung_task_warnings--;
144 		pr_err("INFO: task %s:%d blocked for more than %ld seconds.\n",
145 		       t->comm, t->pid, (jiffies - t->last_switch_time) / HZ);
146 		pr_err("      %s %s %.*s\n",
147 			print_tainted(), init_utsname()->release,
148 			(int)strcspn(init_utsname()->version, " "),
149 			init_utsname()->version);
150 		pr_err("\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
151 			" disables this message.\n");
152 		sched_show_task(t);
153 		hung_task_show_lock = true;
154 
155 		if (sysctl_hung_task_all_cpu_backtrace)
156 			hung_task_show_all_bt = true;
157 		if (!sysctl_hung_task_warnings)
158 			pr_info("Future hung task reports are suppressed, see sysctl kernel.hung_task_warnings\n");
159 	}
160 
161 	touch_nmi_watchdog();
162 }
163 
164 /*
165  * To avoid extending the RCU grace period for an unbounded amount of time,
166  * periodically exit the critical section and enter a new one.
167  *
168  * For preemptible RCU it is sufficient to call rcu_read_unlock in order
169  * to exit the grace period. For classic RCU, a reschedule is required.
170  */
171 static bool rcu_lock_break(struct task_struct *g, struct task_struct *t)
172 {
173 	bool can_cont;
174 
175 	get_task_struct(g);
176 	get_task_struct(t);
177 	rcu_read_unlock();
178 	cond_resched();
179 	rcu_read_lock();
180 	can_cont = pid_alive(g) && pid_alive(t);
181 	put_task_struct(t);
182 	put_task_struct(g);
183 
184 	return can_cont;
185 }
186 
187 /*
188  * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
189  * a really long time (120 seconds). If that happens, print out
190  * a warning.
191  */
192 static void check_hung_uninterruptible_tasks(unsigned long timeout)
193 {
194 	int max_count = sysctl_hung_task_check_count;
195 	unsigned long last_break = jiffies;
196 	struct task_struct *g, *t;
197 
198 	/*
199 	 * If the system crashed already then all bets are off,
200 	 * do not report extra hung tasks:
201 	 */
202 	if (test_taint(TAINT_DIE) || did_panic)
203 		return;
204 
205 	hung_task_show_lock = false;
206 	rcu_read_lock();
207 	for_each_process_thread(g, t) {
208 		unsigned int state;
209 
210 		if (!max_count--)
211 			goto unlock;
212 		if (time_after(jiffies, last_break + HUNG_TASK_LOCK_BREAK)) {
213 			if (!rcu_lock_break(g, t))
214 				goto unlock;
215 			last_break = jiffies;
216 		}
217 		/*
218 		 * skip the TASK_KILLABLE tasks -- these can be killed
219 		 * skip the TASK_IDLE tasks -- those are genuinely idle
220 		 */
221 		state = READ_ONCE(t->__state);
222 		if ((state & TASK_UNINTERRUPTIBLE) &&
223 		    !(state & TASK_WAKEKILL) &&
224 		    !(state & TASK_NOLOAD))
225 			check_hung_task(t, timeout);
226 	}
227  unlock:
228 	rcu_read_unlock();
229 	if (hung_task_show_lock)
230 		debug_show_all_locks();
231 
232 	if (hung_task_show_all_bt) {
233 		hung_task_show_all_bt = false;
234 		trigger_all_cpu_backtrace();
235 	}
236 
237 	if (hung_task_call_panic)
238 		panic("hung_task: blocked tasks");
239 }
240 
241 static long hung_timeout_jiffies(unsigned long last_checked,
242 				 unsigned long timeout)
243 {
244 	/* timeout of 0 will disable the watchdog */
245 	return timeout ? last_checked - jiffies + timeout * HZ :
246 		MAX_SCHEDULE_TIMEOUT;
247 }
248 
249 #ifdef CONFIG_SYSCTL
250 /*
251  * Process updating of timeout sysctl
252  */
253 static int proc_dohung_task_timeout_secs(const struct ctl_table *table, int write,
254 				  void *buffer,
255 				  size_t *lenp, loff_t *ppos)
256 {
257 	int ret;
258 
259 	ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
260 
261 	if (ret || !write)
262 		goto out;
263 
264 	wake_up_process(watchdog_task);
265 
266  out:
267 	return ret;
268 }
269 
270 /*
271  * This is needed for proc_doulongvec_minmax of sysctl_hung_task_timeout_secs
272  * and hung_task_check_interval_secs
273  */
274 static const unsigned long hung_task_timeout_max = (LONG_MAX / HZ);
275 static struct ctl_table hung_task_sysctls[] = {
276 #ifdef CONFIG_SMP
277 	{
278 		.procname	= "hung_task_all_cpu_backtrace",
279 		.data		= &sysctl_hung_task_all_cpu_backtrace,
280 		.maxlen		= sizeof(int),
281 		.mode		= 0644,
282 		.proc_handler	= proc_dointvec_minmax,
283 		.extra1		= SYSCTL_ZERO,
284 		.extra2		= SYSCTL_ONE,
285 	},
286 #endif /* CONFIG_SMP */
287 	{
288 		.procname	= "hung_task_panic",
289 		.data		= &sysctl_hung_task_panic,
290 		.maxlen		= sizeof(int),
291 		.mode		= 0644,
292 		.proc_handler	= proc_dointvec_minmax,
293 		.extra1		= SYSCTL_ZERO,
294 		.extra2		= SYSCTL_ONE,
295 	},
296 	{
297 		.procname	= "hung_task_check_count",
298 		.data		= &sysctl_hung_task_check_count,
299 		.maxlen		= sizeof(int),
300 		.mode		= 0644,
301 		.proc_handler	= proc_dointvec_minmax,
302 		.extra1		= SYSCTL_ZERO,
303 	},
304 	{
305 		.procname	= "hung_task_timeout_secs",
306 		.data		= &sysctl_hung_task_timeout_secs,
307 		.maxlen		= sizeof(unsigned long),
308 		.mode		= 0644,
309 		.proc_handler	= proc_dohung_task_timeout_secs,
310 		.extra2		= (void *)&hung_task_timeout_max,
311 	},
312 	{
313 		.procname	= "hung_task_check_interval_secs",
314 		.data		= &sysctl_hung_task_check_interval_secs,
315 		.maxlen		= sizeof(unsigned long),
316 		.mode		= 0644,
317 		.proc_handler	= proc_dohung_task_timeout_secs,
318 		.extra2		= (void *)&hung_task_timeout_max,
319 	},
320 	{
321 		.procname	= "hung_task_warnings",
322 		.data		= &sysctl_hung_task_warnings,
323 		.maxlen		= sizeof(int),
324 		.mode		= 0644,
325 		.proc_handler	= proc_dointvec_minmax,
326 		.extra1		= SYSCTL_NEG_ONE,
327 	},
328 	{
329 		.procname	= "hung_task_detect_count",
330 		.data		= &sysctl_hung_task_detect_count,
331 		.maxlen		= sizeof(unsigned long),
332 		.mode		= 0444,
333 		.proc_handler	= proc_doulongvec_minmax,
334 	},
335 };
336 
337 static void __init hung_task_sysctl_init(void)
338 {
339 	register_sysctl_init("kernel", hung_task_sysctls);
340 }
341 #else
342 #define hung_task_sysctl_init() do { } while (0)
343 #endif /* CONFIG_SYSCTL */
344 
345 
346 static atomic_t reset_hung_task = ATOMIC_INIT(0);
347 
348 void reset_hung_task_detector(void)
349 {
350 	atomic_set(&reset_hung_task, 1);
351 }
352 EXPORT_SYMBOL_GPL(reset_hung_task_detector);
353 
354 static bool hung_detector_suspended;
355 
356 static int hungtask_pm_notify(struct notifier_block *self,
357 			      unsigned long action, void *hcpu)
358 {
359 	switch (action) {
360 	case PM_SUSPEND_PREPARE:
361 	case PM_HIBERNATION_PREPARE:
362 	case PM_RESTORE_PREPARE:
363 		hung_detector_suspended = true;
364 		break;
365 	case PM_POST_SUSPEND:
366 	case PM_POST_HIBERNATION:
367 	case PM_POST_RESTORE:
368 		hung_detector_suspended = false;
369 		break;
370 	default:
371 		break;
372 	}
373 	return NOTIFY_OK;
374 }
375 
376 /*
377  * kthread which checks for tasks stuck in D state
378  */
379 static int watchdog(void *dummy)
380 {
381 	unsigned long hung_last_checked = jiffies;
382 
383 	set_user_nice(current, 0);
384 
385 	for ( ; ; ) {
386 		unsigned long timeout = sysctl_hung_task_timeout_secs;
387 		unsigned long interval = sysctl_hung_task_check_interval_secs;
388 		long t;
389 
390 		if (interval == 0)
391 			interval = timeout;
392 		interval = min_t(unsigned long, interval, timeout);
393 		t = hung_timeout_jiffies(hung_last_checked, interval);
394 		if (t <= 0) {
395 			if (!atomic_xchg(&reset_hung_task, 0) &&
396 			    !hung_detector_suspended)
397 				check_hung_uninterruptible_tasks(timeout);
398 			hung_last_checked = jiffies;
399 			continue;
400 		}
401 		schedule_timeout_interruptible(t);
402 	}
403 
404 	return 0;
405 }
406 
407 static int __init hung_task_init(void)
408 {
409 	atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
410 
411 	/* Disable hung task detector on suspend */
412 	pm_notifier(hungtask_pm_notify, 0);
413 
414 	watchdog_task = kthread_run(watchdog, NULL, "khungtaskd");
415 	hung_task_sysctl_init();
416 
417 	return 0;
418 }
419 subsys_initcall(hung_task_init);
420