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