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
hung_task_panic(struct notifier_block * this,unsigned long event,void * ptr)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 
97 #ifdef CONFIG_DETECT_HUNG_TASK_BLOCKER
debug_show_blocker(struct task_struct * task)98 static void debug_show_blocker(struct task_struct *task)
99 {
100 	struct task_struct *g, *t;
101 	unsigned long owner;
102 	struct mutex *lock;
103 
104 	RCU_LOCKDEP_WARN(!rcu_read_lock_held(), "No rcu lock held");
105 
106 	lock = READ_ONCE(task->blocker_mutex);
107 	if (!lock)
108 		return;
109 
110 	owner = mutex_get_owner(lock);
111 	if (unlikely(!owner)) {
112 		pr_err("INFO: task %s:%d is blocked on a mutex, but the owner is not found.\n",
113 			task->comm, task->pid);
114 		return;
115 	}
116 
117 	/* Ensure the owner information is correct. */
118 	for_each_process_thread(g, t) {
119 		if ((unsigned long)t == owner) {
120 			pr_err("INFO: task %s:%d is blocked on a mutex likely owned by task %s:%d.\n",
121 				task->comm, task->pid, t->comm, t->pid);
122 			sched_show_task(t);
123 			return;
124 		}
125 	}
126 }
127 #else
debug_show_blocker(struct task_struct * task)128 static inline void debug_show_blocker(struct task_struct *task)
129 {
130 }
131 #endif
132 
check_hung_task(struct task_struct * t,unsigned long timeout)133 static void check_hung_task(struct task_struct *t, unsigned long timeout)
134 {
135 	unsigned long switch_count = t->nvcsw + t->nivcsw;
136 
137 	/*
138 	 * Ensure the task is not frozen.
139 	 * Also, skip vfork and any other user process that freezer should skip.
140 	 */
141 	if (unlikely(READ_ONCE(t->__state) & TASK_FROZEN))
142 		return;
143 
144 	/*
145 	 * When a freshly created task is scheduled once, changes its state to
146 	 * TASK_UNINTERRUPTIBLE without having ever been switched out once, it
147 	 * musn't be checked.
148 	 */
149 	if (unlikely(!switch_count))
150 		return;
151 
152 	if (switch_count != t->last_switch_count) {
153 		t->last_switch_count = switch_count;
154 		t->last_switch_time = jiffies;
155 		return;
156 	}
157 	if (time_is_after_jiffies(t->last_switch_time + timeout * HZ))
158 		return;
159 
160 	/*
161 	 * This counter tracks the total number of tasks detected as hung
162 	 * since boot.
163 	 */
164 	sysctl_hung_task_detect_count++;
165 
166 	trace_sched_process_hang(t);
167 
168 	if (sysctl_hung_task_panic) {
169 		console_verbose();
170 		hung_task_show_lock = true;
171 		hung_task_call_panic = true;
172 	}
173 
174 	/*
175 	 * Ok, the task did not get scheduled for more than 2 minutes,
176 	 * complain:
177 	 */
178 	if (sysctl_hung_task_warnings || hung_task_call_panic) {
179 		if (sysctl_hung_task_warnings > 0)
180 			sysctl_hung_task_warnings--;
181 		pr_err("INFO: task %s:%d blocked for more than %ld seconds.\n",
182 		       t->comm, t->pid, (jiffies - t->last_switch_time) / HZ);
183 		pr_err("      %s %s %.*s\n",
184 			print_tainted(), init_utsname()->release,
185 			(int)strcspn(init_utsname()->version, " "),
186 			init_utsname()->version);
187 		if (t->flags & PF_POSTCOREDUMP)
188 			pr_err("      Blocked by coredump.\n");
189 		pr_err("\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
190 			" disables this message.\n");
191 		sched_show_task(t);
192 		debug_show_blocker(t);
193 		hung_task_show_lock = true;
194 
195 		if (sysctl_hung_task_all_cpu_backtrace)
196 			hung_task_show_all_bt = true;
197 		if (!sysctl_hung_task_warnings)
198 			pr_info("Future hung task reports are suppressed, see sysctl kernel.hung_task_warnings\n");
199 	}
200 
201 	touch_nmi_watchdog();
202 }
203 
204 /*
205  * To avoid extending the RCU grace period for an unbounded amount of time,
206  * periodically exit the critical section and enter a new one.
207  *
208  * For preemptible RCU it is sufficient to call rcu_read_unlock in order
209  * to exit the grace period. For classic RCU, a reschedule is required.
210  */
rcu_lock_break(struct task_struct * g,struct task_struct * t)211 static bool rcu_lock_break(struct task_struct *g, struct task_struct *t)
212 {
213 	bool can_cont;
214 
215 	get_task_struct(g);
216 	get_task_struct(t);
217 	rcu_read_unlock();
218 	cond_resched();
219 	rcu_read_lock();
220 	can_cont = pid_alive(g) && pid_alive(t);
221 	put_task_struct(t);
222 	put_task_struct(g);
223 
224 	return can_cont;
225 }
226 
227 /*
228  * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
229  * a really long time (120 seconds). If that happens, print out
230  * a warning.
231  */
check_hung_uninterruptible_tasks(unsigned long timeout)232 static void check_hung_uninterruptible_tasks(unsigned long timeout)
233 {
234 	int max_count = sysctl_hung_task_check_count;
235 	unsigned long last_break = jiffies;
236 	struct task_struct *g, *t;
237 
238 	/*
239 	 * If the system crashed already then all bets are off,
240 	 * do not report extra hung tasks:
241 	 */
242 	if (test_taint(TAINT_DIE) || did_panic)
243 		return;
244 
245 	hung_task_show_lock = false;
246 	rcu_read_lock();
247 	for_each_process_thread(g, t) {
248 		unsigned int state;
249 
250 		if (!max_count--)
251 			goto unlock;
252 		if (time_after(jiffies, last_break + HUNG_TASK_LOCK_BREAK)) {
253 			if (!rcu_lock_break(g, t))
254 				goto unlock;
255 			last_break = jiffies;
256 		}
257 		/*
258 		 * skip the TASK_KILLABLE tasks -- these can be killed
259 		 * skip the TASK_IDLE tasks -- those are genuinely idle
260 		 */
261 		state = READ_ONCE(t->__state);
262 		if ((state & TASK_UNINTERRUPTIBLE) &&
263 		    !(state & TASK_WAKEKILL) &&
264 		    !(state & TASK_NOLOAD))
265 			check_hung_task(t, timeout);
266 	}
267  unlock:
268 	rcu_read_unlock();
269 	if (hung_task_show_lock)
270 		debug_show_all_locks();
271 
272 	if (hung_task_show_all_bt) {
273 		hung_task_show_all_bt = false;
274 		trigger_all_cpu_backtrace();
275 	}
276 
277 	if (hung_task_call_panic)
278 		panic("hung_task: blocked tasks");
279 }
280 
hung_timeout_jiffies(unsigned long last_checked,unsigned long timeout)281 static long hung_timeout_jiffies(unsigned long last_checked,
282 				 unsigned long timeout)
283 {
284 	/* timeout of 0 will disable the watchdog */
285 	return timeout ? last_checked - jiffies + timeout * HZ :
286 		MAX_SCHEDULE_TIMEOUT;
287 }
288 
289 #ifdef CONFIG_SYSCTL
290 /*
291  * Process updating of timeout sysctl
292  */
proc_dohung_task_timeout_secs(const struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)293 static int proc_dohung_task_timeout_secs(const struct ctl_table *table, int write,
294 				  void *buffer,
295 				  size_t *lenp, loff_t *ppos)
296 {
297 	int ret;
298 
299 	ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
300 
301 	if (ret || !write)
302 		goto out;
303 
304 	wake_up_process(watchdog_task);
305 
306  out:
307 	return ret;
308 }
309 
310 /*
311  * This is needed for proc_doulongvec_minmax of sysctl_hung_task_timeout_secs
312  * and hung_task_check_interval_secs
313  */
314 static const unsigned long hung_task_timeout_max = (LONG_MAX / HZ);
315 static const struct ctl_table hung_task_sysctls[] = {
316 #ifdef CONFIG_SMP
317 	{
318 		.procname	= "hung_task_all_cpu_backtrace",
319 		.data		= &sysctl_hung_task_all_cpu_backtrace,
320 		.maxlen		= sizeof(int),
321 		.mode		= 0644,
322 		.proc_handler	= proc_dointvec_minmax,
323 		.extra1		= SYSCTL_ZERO,
324 		.extra2		= SYSCTL_ONE,
325 	},
326 #endif /* CONFIG_SMP */
327 	{
328 		.procname	= "hung_task_panic",
329 		.data		= &sysctl_hung_task_panic,
330 		.maxlen		= sizeof(int),
331 		.mode		= 0644,
332 		.proc_handler	= proc_dointvec_minmax,
333 		.extra1		= SYSCTL_ZERO,
334 		.extra2		= SYSCTL_ONE,
335 	},
336 	{
337 		.procname	= "hung_task_check_count",
338 		.data		= &sysctl_hung_task_check_count,
339 		.maxlen		= sizeof(int),
340 		.mode		= 0644,
341 		.proc_handler	= proc_dointvec_minmax,
342 		.extra1		= SYSCTL_ZERO,
343 	},
344 	{
345 		.procname	= "hung_task_timeout_secs",
346 		.data		= &sysctl_hung_task_timeout_secs,
347 		.maxlen		= sizeof(unsigned long),
348 		.mode		= 0644,
349 		.proc_handler	= proc_dohung_task_timeout_secs,
350 		.extra2		= (void *)&hung_task_timeout_max,
351 	},
352 	{
353 		.procname	= "hung_task_check_interval_secs",
354 		.data		= &sysctl_hung_task_check_interval_secs,
355 		.maxlen		= sizeof(unsigned long),
356 		.mode		= 0644,
357 		.proc_handler	= proc_dohung_task_timeout_secs,
358 		.extra2		= (void *)&hung_task_timeout_max,
359 	},
360 	{
361 		.procname	= "hung_task_warnings",
362 		.data		= &sysctl_hung_task_warnings,
363 		.maxlen		= sizeof(int),
364 		.mode		= 0644,
365 		.proc_handler	= proc_dointvec_minmax,
366 		.extra1		= SYSCTL_NEG_ONE,
367 	},
368 	{
369 		.procname	= "hung_task_detect_count",
370 		.data		= &sysctl_hung_task_detect_count,
371 		.maxlen		= sizeof(unsigned long),
372 		.mode		= 0444,
373 		.proc_handler	= proc_doulongvec_minmax,
374 	},
375 };
376 
hung_task_sysctl_init(void)377 static void __init hung_task_sysctl_init(void)
378 {
379 	register_sysctl_init("kernel", hung_task_sysctls);
380 }
381 #else
382 #define hung_task_sysctl_init() do { } while (0)
383 #endif /* CONFIG_SYSCTL */
384 
385 
386 static atomic_t reset_hung_task = ATOMIC_INIT(0);
387 
reset_hung_task_detector(void)388 void reset_hung_task_detector(void)
389 {
390 	atomic_set(&reset_hung_task, 1);
391 }
392 EXPORT_SYMBOL_GPL(reset_hung_task_detector);
393 
394 static bool hung_detector_suspended;
395 
hungtask_pm_notify(struct notifier_block * self,unsigned long action,void * hcpu)396 static int hungtask_pm_notify(struct notifier_block *self,
397 			      unsigned long action, void *hcpu)
398 {
399 	switch (action) {
400 	case PM_SUSPEND_PREPARE:
401 	case PM_HIBERNATION_PREPARE:
402 	case PM_RESTORE_PREPARE:
403 		hung_detector_suspended = true;
404 		break;
405 	case PM_POST_SUSPEND:
406 	case PM_POST_HIBERNATION:
407 	case PM_POST_RESTORE:
408 		hung_detector_suspended = false;
409 		break;
410 	default:
411 		break;
412 	}
413 	return NOTIFY_OK;
414 }
415 
416 /*
417  * kthread which checks for tasks stuck in D state
418  */
watchdog(void * dummy)419 static int watchdog(void *dummy)
420 {
421 	unsigned long hung_last_checked = jiffies;
422 
423 	set_user_nice(current, 0);
424 
425 	for ( ; ; ) {
426 		unsigned long timeout = sysctl_hung_task_timeout_secs;
427 		unsigned long interval = sysctl_hung_task_check_interval_secs;
428 		long t;
429 
430 		if (interval == 0)
431 			interval = timeout;
432 		interval = min_t(unsigned long, interval, timeout);
433 		t = hung_timeout_jiffies(hung_last_checked, interval);
434 		if (t <= 0) {
435 			if (!atomic_xchg(&reset_hung_task, 0) &&
436 			    !hung_detector_suspended)
437 				check_hung_uninterruptible_tasks(timeout);
438 			hung_last_checked = jiffies;
439 			continue;
440 		}
441 		schedule_timeout_interruptible(t);
442 	}
443 
444 	return 0;
445 }
446 
hung_task_init(void)447 static int __init hung_task_init(void)
448 {
449 	atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
450 
451 	/* Disable hung task detector on suspend */
452 	pm_notifier(hungtask_pm_notify, 0);
453 
454 	watchdog_task = kthread_run(watchdog, NULL, "khungtaskd");
455 	hung_task_sysctl_init();
456 
457 	return 0;
458 }
459 subsys_initcall(hung_task_init);
460