xref: /linux/kernel/hung_task.c (revision e991acf1bce7a428794514cbbe216973c9c0a3c8)
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 #include <linux/hung_task.h>
26 #include <linux/rwsem.h>
27 
28 #include <trace/events/sched.h>
29 
30 /*
31  * The number of tasks checked:
32  */
33 static int __read_mostly sysctl_hung_task_check_count = PID_MAX_LIMIT;
34 
35 /*
36  * Total number of tasks detected as hung since boot:
37  */
38 static unsigned long __read_mostly sysctl_hung_task_detect_count;
39 
40 /*
41  * Limit number of tasks checked in a batch.
42  *
43  * This value controls the preemptibility of khungtaskd since preemption
44  * is disabled during the critical section. It also controls the size of
45  * the RCU grace period. So it needs to be upper-bound.
46  */
47 #define HUNG_TASK_LOCK_BREAK (HZ / 10)
48 
49 /*
50  * Zero means infinite timeout - no checking done:
51  */
52 unsigned long __read_mostly sysctl_hung_task_timeout_secs = CONFIG_DEFAULT_HUNG_TASK_TIMEOUT;
53 EXPORT_SYMBOL_GPL(sysctl_hung_task_timeout_secs);
54 
55 /*
56  * Zero (default value) means use sysctl_hung_task_timeout_secs:
57  */
58 static unsigned long __read_mostly sysctl_hung_task_check_interval_secs;
59 
60 static int __read_mostly sysctl_hung_task_warnings = 10;
61 
62 static int __read_mostly did_panic;
63 static bool hung_task_show_lock;
64 static bool hung_task_call_panic;
65 static bool hung_task_show_all_bt;
66 
67 static struct task_struct *watchdog_task;
68 
69 #ifdef CONFIG_SMP
70 /*
71  * Should we dump all CPUs backtraces in a hung task event?
72  * Defaults to 0, can be changed via sysctl.
73  */
74 static unsigned int __read_mostly sysctl_hung_task_all_cpu_backtrace;
75 #else
76 #define sysctl_hung_task_all_cpu_backtrace 0
77 #endif /* CONFIG_SMP */
78 
79 /*
80  * Should we panic (and reboot, if panic_timeout= is set) when a
81  * hung task is detected:
82  */
83 static unsigned int __read_mostly sysctl_hung_task_panic =
84 	IS_ENABLED(CONFIG_BOOTPARAM_HUNG_TASK_PANIC);
85 
86 static int
hung_task_panic(struct notifier_block * this,unsigned long event,void * ptr)87 hung_task_panic(struct notifier_block *this, unsigned long event, void *ptr)
88 {
89 	did_panic = 1;
90 
91 	return NOTIFY_DONE;
92 }
93 
94 static struct notifier_block panic_block = {
95 	.notifier_call = hung_task_panic,
96 };
97 
98 
99 #ifdef CONFIG_DETECT_HUNG_TASK_BLOCKER
debug_show_blocker(struct task_struct * task)100 static void debug_show_blocker(struct task_struct *task)
101 {
102 	struct task_struct *g, *t;
103 	unsigned long owner, blocker, blocker_type;
104 	const char *rwsem_blocked_by, *rwsem_blocked_as;
105 
106 	RCU_LOCKDEP_WARN(!rcu_read_lock_held(), "No rcu lock held");
107 
108 	blocker = READ_ONCE(task->blocker);
109 	if (!blocker)
110 		return;
111 
112 	blocker_type = hung_task_get_blocker_type(blocker);
113 
114 	switch (blocker_type) {
115 	case BLOCKER_TYPE_MUTEX:
116 		owner = mutex_get_owner(hung_task_blocker_to_lock(blocker));
117 		break;
118 	case BLOCKER_TYPE_SEM:
119 		owner = sem_last_holder(hung_task_blocker_to_lock(blocker));
120 		break;
121 	case BLOCKER_TYPE_RWSEM_READER:
122 	case BLOCKER_TYPE_RWSEM_WRITER:
123 		owner = (unsigned long)rwsem_owner(
124 					hung_task_blocker_to_lock(blocker));
125 		rwsem_blocked_as = (blocker_type == BLOCKER_TYPE_RWSEM_READER) ?
126 					"reader" : "writer";
127 		rwsem_blocked_by = is_rwsem_reader_owned(
128 					hung_task_blocker_to_lock(blocker)) ?
129 					"reader" : "writer";
130 		break;
131 	default:
132 		WARN_ON_ONCE(1);
133 		return;
134 	}
135 
136 
137 	if (unlikely(!owner)) {
138 		switch (blocker_type) {
139 		case BLOCKER_TYPE_MUTEX:
140 			pr_err("INFO: task %s:%d is blocked on a mutex, but the owner is not found.\n",
141 			       task->comm, task->pid);
142 			break;
143 		case BLOCKER_TYPE_SEM:
144 			pr_err("INFO: task %s:%d is blocked on a semaphore, but the last holder is not found.\n",
145 			       task->comm, task->pid);
146 			break;
147 		case BLOCKER_TYPE_RWSEM_READER:
148 		case BLOCKER_TYPE_RWSEM_WRITER:
149 			pr_err("INFO: task %s:%d is blocked on an rw-semaphore, but the owner is not found.\n",
150 			       task->comm, task->pid);
151 			break;
152 		}
153 		return;
154 	}
155 
156 	/* Ensure the owner information is correct. */
157 	for_each_process_thread(g, t) {
158 		if ((unsigned long)t != owner)
159 			continue;
160 
161 		switch (blocker_type) {
162 		case BLOCKER_TYPE_MUTEX:
163 			pr_err("INFO: task %s:%d is blocked on a mutex likely owned by task %s:%d.\n",
164 			       task->comm, task->pid, t->comm, t->pid);
165 			break;
166 		case BLOCKER_TYPE_SEM:
167 			pr_err("INFO: task %s:%d blocked on a semaphore likely last held by task %s:%d\n",
168 			       task->comm, task->pid, t->comm, t->pid);
169 			break;
170 		case BLOCKER_TYPE_RWSEM_READER:
171 		case BLOCKER_TYPE_RWSEM_WRITER:
172 			pr_err("INFO: task %s:%d <%s> blocked on an rw-semaphore likely owned by task %s:%d <%s>\n",
173 			       task->comm, task->pid, rwsem_blocked_as, t->comm,
174 			       t->pid, rwsem_blocked_by);
175 			break;
176 		}
177 		sched_show_task(t);
178 		return;
179 	}
180 }
181 #else
debug_show_blocker(struct task_struct * task)182 static inline void debug_show_blocker(struct task_struct *task)
183 {
184 }
185 #endif
186 
check_hung_task(struct task_struct * t,unsigned long timeout)187 static void check_hung_task(struct task_struct *t, unsigned long timeout)
188 {
189 	unsigned long switch_count = t->nvcsw + t->nivcsw;
190 
191 	/*
192 	 * Ensure the task is not frozen.
193 	 * Also, skip vfork and any other user process that freezer should skip.
194 	 */
195 	if (unlikely(READ_ONCE(t->__state) & TASK_FROZEN))
196 		return;
197 
198 	/*
199 	 * When a freshly created task is scheduled once, changes its state to
200 	 * TASK_UNINTERRUPTIBLE without having ever been switched out once, it
201 	 * musn't be checked.
202 	 */
203 	if (unlikely(!switch_count))
204 		return;
205 
206 	if (switch_count != t->last_switch_count) {
207 		t->last_switch_count = switch_count;
208 		t->last_switch_time = jiffies;
209 		return;
210 	}
211 	if (time_is_after_jiffies(t->last_switch_time + timeout * HZ))
212 		return;
213 
214 	/*
215 	 * This counter tracks the total number of tasks detected as hung
216 	 * since boot.
217 	 */
218 	sysctl_hung_task_detect_count++;
219 
220 	trace_sched_process_hang(t);
221 
222 	if (sysctl_hung_task_panic) {
223 		console_verbose();
224 		hung_task_show_lock = true;
225 		hung_task_call_panic = true;
226 	}
227 
228 	/*
229 	 * Ok, the task did not get scheduled for more than 2 minutes,
230 	 * complain:
231 	 */
232 	if (sysctl_hung_task_warnings || hung_task_call_panic) {
233 		if (sysctl_hung_task_warnings > 0)
234 			sysctl_hung_task_warnings--;
235 		pr_err("INFO: task %s:%d blocked for more than %ld seconds.\n",
236 		       t->comm, t->pid, (jiffies - t->last_switch_time) / HZ);
237 		pr_err("      %s %s %.*s\n",
238 			print_tainted(), init_utsname()->release,
239 			(int)strcspn(init_utsname()->version, " "),
240 			init_utsname()->version);
241 		if (t->flags & PF_POSTCOREDUMP)
242 			pr_err("      Blocked by coredump.\n");
243 		pr_err("\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
244 			" disables this message.\n");
245 		sched_show_task(t);
246 		debug_show_blocker(t);
247 		hung_task_show_lock = true;
248 
249 		if (sysctl_hung_task_all_cpu_backtrace)
250 			hung_task_show_all_bt = true;
251 		if (!sysctl_hung_task_warnings)
252 			pr_info("Future hung task reports are suppressed, see sysctl kernel.hung_task_warnings\n");
253 	}
254 
255 	touch_nmi_watchdog();
256 }
257 
258 /*
259  * To avoid extending the RCU grace period for an unbounded amount of time,
260  * periodically exit the critical section and enter a new one.
261  *
262  * For preemptible RCU it is sufficient to call rcu_read_unlock in order
263  * to exit the grace period. For classic RCU, a reschedule is required.
264  */
rcu_lock_break(struct task_struct * g,struct task_struct * t)265 static bool rcu_lock_break(struct task_struct *g, struct task_struct *t)
266 {
267 	bool can_cont;
268 
269 	get_task_struct(g);
270 	get_task_struct(t);
271 	rcu_read_unlock();
272 	cond_resched();
273 	rcu_read_lock();
274 	can_cont = pid_alive(g) && pid_alive(t);
275 	put_task_struct(t);
276 	put_task_struct(g);
277 
278 	return can_cont;
279 }
280 
281 /*
282  * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
283  * a really long time (120 seconds). If that happens, print out
284  * a warning.
285  */
check_hung_uninterruptible_tasks(unsigned long timeout)286 static void check_hung_uninterruptible_tasks(unsigned long timeout)
287 {
288 	int max_count = sysctl_hung_task_check_count;
289 	unsigned long last_break = jiffies;
290 	struct task_struct *g, *t;
291 
292 	/*
293 	 * If the system crashed already then all bets are off,
294 	 * do not report extra hung tasks:
295 	 */
296 	if (test_taint(TAINT_DIE) || did_panic)
297 		return;
298 
299 	hung_task_show_lock = false;
300 	rcu_read_lock();
301 	for_each_process_thread(g, t) {
302 		unsigned int state;
303 
304 		if (!max_count--)
305 			goto unlock;
306 		if (time_after(jiffies, last_break + HUNG_TASK_LOCK_BREAK)) {
307 			if (!rcu_lock_break(g, t))
308 				goto unlock;
309 			last_break = jiffies;
310 		}
311 		/*
312 		 * skip the TASK_KILLABLE tasks -- these can be killed
313 		 * skip the TASK_IDLE tasks -- those are genuinely idle
314 		 */
315 		state = READ_ONCE(t->__state);
316 		if ((state & TASK_UNINTERRUPTIBLE) &&
317 		    !(state & TASK_WAKEKILL) &&
318 		    !(state & TASK_NOLOAD))
319 			check_hung_task(t, timeout);
320 	}
321  unlock:
322 	rcu_read_unlock();
323 	if (hung_task_show_lock)
324 		debug_show_all_locks();
325 
326 	if (hung_task_show_all_bt) {
327 		hung_task_show_all_bt = false;
328 		trigger_all_cpu_backtrace();
329 	}
330 
331 	if (hung_task_call_panic)
332 		panic("hung_task: blocked tasks");
333 }
334 
hung_timeout_jiffies(unsigned long last_checked,unsigned long timeout)335 static long hung_timeout_jiffies(unsigned long last_checked,
336 				 unsigned long timeout)
337 {
338 	/* timeout of 0 will disable the watchdog */
339 	return timeout ? last_checked - jiffies + timeout * HZ :
340 		MAX_SCHEDULE_TIMEOUT;
341 }
342 
343 #ifdef CONFIG_SYSCTL
344 /*
345  * Process updating of timeout sysctl
346  */
proc_dohung_task_timeout_secs(const struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)347 static int proc_dohung_task_timeout_secs(const struct ctl_table *table, int write,
348 				  void *buffer,
349 				  size_t *lenp, loff_t *ppos)
350 {
351 	int ret;
352 
353 	ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
354 
355 	if (ret || !write)
356 		goto out;
357 
358 	wake_up_process(watchdog_task);
359 
360  out:
361 	return ret;
362 }
363 
364 /*
365  * This is needed for proc_doulongvec_minmax of sysctl_hung_task_timeout_secs
366  * and hung_task_check_interval_secs
367  */
368 static const unsigned long hung_task_timeout_max = (LONG_MAX / HZ);
369 static const struct ctl_table hung_task_sysctls[] = {
370 #ifdef CONFIG_SMP
371 	{
372 		.procname	= "hung_task_all_cpu_backtrace",
373 		.data		= &sysctl_hung_task_all_cpu_backtrace,
374 		.maxlen		= sizeof(int),
375 		.mode		= 0644,
376 		.proc_handler	= proc_dointvec_minmax,
377 		.extra1		= SYSCTL_ZERO,
378 		.extra2		= SYSCTL_ONE,
379 	},
380 #endif /* CONFIG_SMP */
381 	{
382 		.procname	= "hung_task_panic",
383 		.data		= &sysctl_hung_task_panic,
384 		.maxlen		= sizeof(int),
385 		.mode		= 0644,
386 		.proc_handler	= proc_dointvec_minmax,
387 		.extra1		= SYSCTL_ZERO,
388 		.extra2		= SYSCTL_ONE,
389 	},
390 	{
391 		.procname	= "hung_task_check_count",
392 		.data		= &sysctl_hung_task_check_count,
393 		.maxlen		= sizeof(int),
394 		.mode		= 0644,
395 		.proc_handler	= proc_dointvec_minmax,
396 		.extra1		= SYSCTL_ZERO,
397 	},
398 	{
399 		.procname	= "hung_task_timeout_secs",
400 		.data		= &sysctl_hung_task_timeout_secs,
401 		.maxlen		= sizeof(unsigned long),
402 		.mode		= 0644,
403 		.proc_handler	= proc_dohung_task_timeout_secs,
404 		.extra2		= (void *)&hung_task_timeout_max,
405 	},
406 	{
407 		.procname	= "hung_task_check_interval_secs",
408 		.data		= &sysctl_hung_task_check_interval_secs,
409 		.maxlen		= sizeof(unsigned long),
410 		.mode		= 0644,
411 		.proc_handler	= proc_dohung_task_timeout_secs,
412 		.extra2		= (void *)&hung_task_timeout_max,
413 	},
414 	{
415 		.procname	= "hung_task_warnings",
416 		.data		= &sysctl_hung_task_warnings,
417 		.maxlen		= sizeof(int),
418 		.mode		= 0644,
419 		.proc_handler	= proc_dointvec_minmax,
420 		.extra1		= SYSCTL_NEG_ONE,
421 	},
422 	{
423 		.procname	= "hung_task_detect_count",
424 		.data		= &sysctl_hung_task_detect_count,
425 		.maxlen		= sizeof(unsigned long),
426 		.mode		= 0444,
427 		.proc_handler	= proc_doulongvec_minmax,
428 	},
429 };
430 
hung_task_sysctl_init(void)431 static void __init hung_task_sysctl_init(void)
432 {
433 	register_sysctl_init("kernel", hung_task_sysctls);
434 }
435 #else
436 #define hung_task_sysctl_init() do { } while (0)
437 #endif /* CONFIG_SYSCTL */
438 
439 
440 static atomic_t reset_hung_task = ATOMIC_INIT(0);
441 
reset_hung_task_detector(void)442 void reset_hung_task_detector(void)
443 {
444 	atomic_set(&reset_hung_task, 1);
445 }
446 EXPORT_SYMBOL_GPL(reset_hung_task_detector);
447 
448 static bool hung_detector_suspended;
449 
hungtask_pm_notify(struct notifier_block * self,unsigned long action,void * hcpu)450 static int hungtask_pm_notify(struct notifier_block *self,
451 			      unsigned long action, void *hcpu)
452 {
453 	switch (action) {
454 	case PM_SUSPEND_PREPARE:
455 	case PM_HIBERNATION_PREPARE:
456 	case PM_RESTORE_PREPARE:
457 		hung_detector_suspended = true;
458 		break;
459 	case PM_POST_SUSPEND:
460 	case PM_POST_HIBERNATION:
461 	case PM_POST_RESTORE:
462 		hung_detector_suspended = false;
463 		break;
464 	default:
465 		break;
466 	}
467 	return NOTIFY_OK;
468 }
469 
470 /*
471  * kthread which checks for tasks stuck in D state
472  */
watchdog(void * dummy)473 static int watchdog(void *dummy)
474 {
475 	unsigned long hung_last_checked = jiffies;
476 
477 	set_user_nice(current, 0);
478 
479 	for ( ; ; ) {
480 		unsigned long timeout = sysctl_hung_task_timeout_secs;
481 		unsigned long interval = sysctl_hung_task_check_interval_secs;
482 		long t;
483 
484 		if (interval == 0)
485 			interval = timeout;
486 		interval = min_t(unsigned long, interval, timeout);
487 		t = hung_timeout_jiffies(hung_last_checked, interval);
488 		if (t <= 0) {
489 			if (!atomic_xchg(&reset_hung_task, 0) &&
490 			    !hung_detector_suspended)
491 				check_hung_uninterruptible_tasks(timeout);
492 			hung_last_checked = jiffies;
493 			continue;
494 		}
495 		schedule_timeout_interruptible(t);
496 	}
497 
498 	return 0;
499 }
500 
hung_task_init(void)501 static int __init hung_task_init(void)
502 {
503 	atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
504 
505 	/* Disable hung task detector on suspend */
506 	pm_notifier(hungtask_pm_notify, 0);
507 
508 	watchdog_task = kthread_run(watchdog, NULL, "khungtaskd");
509 	hung_task_sysctl_init();
510 
511 	return 0;
512 }
513 subsys_initcall(hung_task_init);
514