xref: /openbmc/linux/kernel/hung_task.c (revision d3741027) 
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 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 unsigned long __read_mostly sysctl_hung_task_check_interval_secs;
51 
52 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 unsigned int __read_mostly sysctl_hung_task_panic =
76 				IS_ENABLED(CONFIG_BOOTPARAM_HUNG_TASK_PANIC);
77 
78 static int
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 
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(t->flags & (PF_FROZEN | PF_FREEZER_SKIP)))
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 		printk_prefer_direct_enter();
131 
132 		if (sysctl_hung_task_warnings > 0)
133 			sysctl_hung_task_warnings--;
134 		pr_err("INFO: task %s:%d blocked for more than %ld seconds.\n",
135 		       t->comm, t->pid, (jiffies - t->last_switch_time) / HZ);
136 		pr_err("      %s %s %.*s\n",
137 			print_tainted(), init_utsname()->release,
138 			(int)strcspn(init_utsname()->version, " "),
139 			init_utsname()->version);
140 		pr_err("\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
141 			" disables this message.\n");
142 		sched_show_task(t);
143 		hung_task_show_lock = true;
144 
145 		if (sysctl_hung_task_all_cpu_backtrace)
146 			hung_task_show_all_bt = true;
147 
148 		printk_prefer_direct_exit();
149 	}
150 
151 	touch_nmi_watchdog();
152 }
153 
154 /*
155  * To avoid extending the RCU grace period for an unbounded amount of time,
156  * periodically exit the critical section and enter a new one.
157  *
158  * For preemptible RCU it is sufficient to call rcu_read_unlock in order
159  * to exit the grace period. For classic RCU, a reschedule is required.
160  */
161 static bool rcu_lock_break(struct task_struct *g, struct task_struct *t)
162 {
163 	bool can_cont;
164 
165 	get_task_struct(g);
166 	get_task_struct(t);
167 	rcu_read_unlock();
168 	cond_resched();
169 	rcu_read_lock();
170 	can_cont = pid_alive(g) && pid_alive(t);
171 	put_task_struct(t);
172 	put_task_struct(g);
173 
174 	return can_cont;
175 }
176 
177 /*
178  * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
179  * a really long time (120 seconds). If that happens, print out
180  * a warning.
181  */
182 static void check_hung_uninterruptible_tasks(unsigned long timeout)
183 {
184 	int max_count = sysctl_hung_task_check_count;
185 	unsigned long last_break = jiffies;
186 	struct task_struct *g, *t;
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 		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 		/* use "==" to skip the TASK_KILLABLE tasks waiting on NFS */
206 		if (READ_ONCE(t->__state) == TASK_UNINTERRUPTIBLE)
207 			check_hung_task(t, timeout);
208 	}
209  unlock:
210 	rcu_read_unlock();
211 	if (hung_task_show_lock) {
212 		printk_prefer_direct_enter();
213 		debug_show_all_locks();
214 		printk_prefer_direct_exit();
215 	}
216 
217 	if (hung_task_show_all_bt) {
218 		hung_task_show_all_bt = false;
219 		printk_prefer_direct_enter();
220 		trigger_all_cpu_backtrace();
221 		printk_prefer_direct_exit();
222 	}
223 
224 	if (hung_task_call_panic)
225 		panic("hung_task: blocked tasks");
226 }
227 
228 static long hung_timeout_jiffies(unsigned long last_checked,
229 				 unsigned long timeout)
230 {
231 	/* timeout of 0 will disable the watchdog */
232 	return timeout ? last_checked - jiffies + timeout * HZ :
233 		MAX_SCHEDULE_TIMEOUT;
234 }
235 
236 #ifdef CONFIG_SYSCTL
237 /*
238  * Process updating of timeout sysctl
239  */
240 static int proc_dohung_task_timeout_secs(struct ctl_table *table, int write,
241 				  void __user *buffer,
242 				  size_t *lenp, loff_t *ppos)
243 {
244 	int ret;
245 
246 	ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
247 
248 	if (ret || !write)
249 		goto out;
250 
251 	wake_up_process(watchdog_task);
252 
253  out:
254 	return ret;
255 }
256 
257 /*
258  * This is needed for proc_doulongvec_minmax of sysctl_hung_task_timeout_secs
259  * and hung_task_check_interval_secs
260  */
261 static const unsigned long hung_task_timeout_max = (LONG_MAX / HZ);
262 static struct ctl_table hung_task_sysctls[] = {
263 #ifdef CONFIG_SMP
264 	{
265 		.procname	= "hung_task_all_cpu_backtrace",
266 		.data		= &sysctl_hung_task_all_cpu_backtrace,
267 		.maxlen		= sizeof(int),
268 		.mode		= 0644,
269 		.proc_handler	= proc_dointvec_minmax,
270 		.extra1		= SYSCTL_ZERO,
271 		.extra2		= SYSCTL_ONE,
272 	},
273 #endif /* CONFIG_SMP */
274 	{
275 		.procname	= "hung_task_panic",
276 		.data		= &sysctl_hung_task_panic,
277 		.maxlen		= sizeof(int),
278 		.mode		= 0644,
279 		.proc_handler	= proc_dointvec_minmax,
280 		.extra1		= SYSCTL_ZERO,
281 		.extra2		= SYSCTL_ONE,
282 	},
283 	{
284 		.procname	= "hung_task_check_count",
285 		.data		= &sysctl_hung_task_check_count,
286 		.maxlen		= sizeof(int),
287 		.mode		= 0644,
288 		.proc_handler	= proc_dointvec_minmax,
289 		.extra1		= SYSCTL_ZERO,
290 	},
291 	{
292 		.procname	= "hung_task_timeout_secs",
293 		.data		= &sysctl_hung_task_timeout_secs,
294 		.maxlen		= sizeof(unsigned long),
295 		.mode		= 0644,
296 		.proc_handler	= proc_dohung_task_timeout_secs,
297 		.extra2		= (void *)&hung_task_timeout_max,
298 	},
299 	{
300 		.procname	= "hung_task_check_interval_secs",
301 		.data		= &sysctl_hung_task_check_interval_secs,
302 		.maxlen		= sizeof(unsigned long),
303 		.mode		= 0644,
304 		.proc_handler	= proc_dohung_task_timeout_secs,
305 		.extra2		= (void *)&hung_task_timeout_max,
306 	},
307 	{
308 		.procname	= "hung_task_warnings",
309 		.data		= &sysctl_hung_task_warnings,
310 		.maxlen		= sizeof(int),
311 		.mode		= 0644,
312 		.proc_handler	= proc_dointvec_minmax,
313 		.extra1		= SYSCTL_NEG_ONE,
314 	},
315 	{}
316 };
317 
318 static void __init hung_task_sysctl_init(void)
319 {
320 	register_sysctl_init("kernel", hung_task_sysctls);
321 }
322 #else
323 #define hung_task_sysctl_init() do { } while (0)
324 #endif /* CONFIG_SYSCTL */
325 
326 
327 static atomic_t reset_hung_task = ATOMIC_INIT(0);
328 
329 void reset_hung_task_detector(void)
330 {
331 	atomic_set(&reset_hung_task, 1);
332 }
333 EXPORT_SYMBOL_GPL(reset_hung_task_detector);
334 
335 static bool hung_detector_suspended;
336 
337 static int hungtask_pm_notify(struct notifier_block *self,
338 			      unsigned long action, void *hcpu)
339 {
340 	switch (action) {
341 	case PM_SUSPEND_PREPARE:
342 	case PM_HIBERNATION_PREPARE:
343 	case PM_RESTORE_PREPARE:
344 		hung_detector_suspended = true;
345 		break;
346 	case PM_POST_SUSPEND:
347 	case PM_POST_HIBERNATION:
348 	case PM_POST_RESTORE:
349 		hung_detector_suspended = false;
350 		break;
351 	default:
352 		break;
353 	}
354 	return NOTIFY_OK;
355 }
356 
357 /*
358  * kthread which checks for tasks stuck in D state
359  */
360 static int watchdog(void *dummy)
361 {
362 	unsigned long hung_last_checked = jiffies;
363 
364 	set_user_nice(current, 0);
365 
366 	for ( ; ; ) {
367 		unsigned long timeout = sysctl_hung_task_timeout_secs;
368 		unsigned long interval = sysctl_hung_task_check_interval_secs;
369 		long t;
370 
371 		if (interval == 0)
372 			interval = timeout;
373 		interval = min_t(unsigned long, interval, timeout);
374 		t = hung_timeout_jiffies(hung_last_checked, interval);
375 		if (t <= 0) {
376 			if (!atomic_xchg(&reset_hung_task, 0) &&
377 			    !hung_detector_suspended)
378 				check_hung_uninterruptible_tasks(timeout);
379 			hung_last_checked = jiffies;
380 			continue;
381 		}
382 		schedule_timeout_interruptible(t);
383 	}
384 
385 	return 0;
386 }
387 
388 static int __init hung_task_init(void)
389 {
390 	atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
391 
392 	/* Disable hung task detector on suspend */
393 	pm_notifier(hungtask_pm_notify, 0);
394 
395 	watchdog_task = kthread_run(watchdog, NULL, "khungtaskd");
396 	hung_task_sysctl_init();
397 
398 	return 0;
399 }
400 subsys_initcall(hung_task_init);
401