• 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 #undef CREATE_TRACE_POINTS
28 #include <trace/hooks/hung_task.h>
29 
30 /*
31  * The number of tasks checked:
32  */
33 int __read_mostly sysctl_hung_task_check_count = PID_MAX_LIMIT;
34 
35 /*
36  * Limit number of tasks checked in a batch.
37  *
38  * This value controls the preemptibility of khungtaskd since preemption
39  * is disabled during the critical section. It also controls the size of
40  * the RCU grace period. So it needs to be upper-bound.
41  */
42 #define HUNG_TASK_LOCK_BREAK (HZ / 10)
43 
44 /*
45  * Zero means infinite timeout - no checking done:
46  */
47 unsigned long __read_mostly sysctl_hung_task_timeout_secs = CONFIG_DEFAULT_HUNG_TASK_TIMEOUT;
48 
49 /*
50  * Zero (default value) means use sysctl_hung_task_timeout_secs:
51  */
52 unsigned long __read_mostly sysctl_hung_task_check_interval_secs;
53 
54 int __read_mostly sysctl_hung_task_warnings = 10;
55 
56 static int __read_mostly did_panic;
57 static bool hung_task_show_lock;
58 static bool hung_task_call_panic;
59 static bool hung_task_show_all_bt;
60 
61 static struct task_struct *watchdog_task;
62 
63 #ifdef CONFIG_SMP
64 /*
65  * Should we dump all CPUs backtraces in a hung task event?
66  * Defaults to 0, can be changed via sysctl.
67  */
68 unsigned int __read_mostly sysctl_hung_task_all_cpu_backtrace;
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 				CONFIG_BOOTPARAM_HUNG_TASK_PANIC_VALUE;
77 
78 static int
hung_task_panic(struct notifier_block * this,unsigned long event,void * ptr)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 
check_hung_task(struct task_struct * t,unsigned long timeout)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(frozen_or_skipped(t)))
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 		if (sysctl_hung_task_warnings > 0)
131 			sysctl_hung_task_warnings--;
132 		pr_err("INFO: task %s:%d blocked for more than %ld seconds.\n",
133 		       t->comm, t->pid, (jiffies - t->last_switch_time) / HZ);
134 		pr_err("      %s %s %.*s\n",
135 			print_tainted(), init_utsname()->release,
136 			(int)strcspn(init_utsname()->version, " "),
137 			init_utsname()->version);
138 		pr_err("\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
139 			" disables this message.\n");
140 		sched_show_task(t);
141 		hung_task_show_lock = true;
142 
143 		if (sysctl_hung_task_all_cpu_backtrace)
144 			hung_task_show_all_bt = true;
145 	}
146 
147 	touch_nmi_watchdog();
148 }
149 
150 /*
151  * To avoid extending the RCU grace period for an unbounded amount of time,
152  * periodically exit the critical section and enter a new one.
153  *
154  * For preemptible RCU it is sufficient to call rcu_read_unlock in order
155  * to exit the grace period. For classic RCU, a reschedule is required.
156  */
rcu_lock_break(struct task_struct * g,struct task_struct * t)157 static bool rcu_lock_break(struct task_struct *g, struct task_struct *t)
158 {
159 	bool can_cont;
160 
161 	get_task_struct(g);
162 	get_task_struct(t);
163 	rcu_read_unlock();
164 	cond_resched();
165 	rcu_read_lock();
166 	can_cont = pid_alive(g) && pid_alive(t);
167 	put_task_struct(t);
168 	put_task_struct(g);
169 
170 	return can_cont;
171 }
172 
173 /*
174  * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
175  * a really long time (120 seconds). If that happens, print out
176  * a warning.
177  */
check_hung_uninterruptible_tasks(unsigned long timeout)178 static void check_hung_uninterruptible_tasks(unsigned long timeout)
179 {
180 	int max_count = sysctl_hung_task_check_count;
181 	unsigned long last_break = jiffies;
182 	struct task_struct *g, *t;
183 	bool need_check = true;
184 
185 	/*
186 	 * If the system crashed already then all bets are off,
187 	 * do not report extra hung tasks:
188 	 */
189 	if (test_taint(TAINT_DIE) || did_panic)
190 		return;
191 
192 	hung_task_show_lock = false;
193 	rcu_read_lock();
194 	for_each_process_thread(g, t) {
195 		if (!max_count--)
196 			goto unlock;
197 		if (time_after(jiffies, last_break + HUNG_TASK_LOCK_BREAK)) {
198 			if (!rcu_lock_break(g, t))
199 				goto unlock;
200 			last_break = jiffies;
201 		}
202 		trace_android_vh_check_uninterruptible_tasks(t, timeout, &need_check);
203 		if (need_check)
204 			/* use "==" to skip the TASK_KILLABLE tasks waiting on NFS */
205 			if (READ_ONCE(t->__state) == TASK_UNINTERRUPTIBLE)
206 				check_hung_task(t, timeout);
207 	}
208 	trace_android_vh_check_uninterruptible_tasks_dn(NULL);
209  unlock:
210 	rcu_read_unlock();
211 	if (hung_task_show_lock)
212 		debug_show_all_locks();
213 
214 	if (hung_task_show_all_bt) {
215 		hung_task_show_all_bt = false;
216 		trigger_all_cpu_backtrace();
217 	}
218 
219 	if (hung_task_call_panic)
220 		panic("hung_task: blocked tasks");
221 }
222 
hung_timeout_jiffies(unsigned long last_checked,unsigned long timeout)223 static long hung_timeout_jiffies(unsigned long last_checked,
224 				 unsigned long timeout)
225 {
226 	/* timeout of 0 will disable the watchdog */
227 	return timeout ? last_checked - jiffies + timeout * HZ :
228 		MAX_SCHEDULE_TIMEOUT;
229 }
230 
231 /*
232  * Process updating of timeout sysctl
233  */
proc_dohung_task_timeout_secs(struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)234 int proc_dohung_task_timeout_secs(struct ctl_table *table, int write,
235 				  void *buffer, size_t *lenp, loff_t *ppos)
236 {
237 	int ret;
238 
239 	ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
240 
241 	if (ret || !write)
242 		goto out;
243 
244 	wake_up_process(watchdog_task);
245 
246  out:
247 	return ret;
248 }
249 
250 static atomic_t reset_hung_task = ATOMIC_INIT(0);
251 
reset_hung_task_detector(void)252 void reset_hung_task_detector(void)
253 {
254 	atomic_set(&reset_hung_task, 1);
255 }
256 EXPORT_SYMBOL_GPL(reset_hung_task_detector);
257 
258 static bool hung_detector_suspended;
259 
hungtask_pm_notify(struct notifier_block * self,unsigned long action,void * hcpu)260 static int hungtask_pm_notify(struct notifier_block *self,
261 			      unsigned long action, void *hcpu)
262 {
263 	switch (action) {
264 	case PM_SUSPEND_PREPARE:
265 	case PM_HIBERNATION_PREPARE:
266 	case PM_RESTORE_PREPARE:
267 		hung_detector_suspended = true;
268 		break;
269 	case PM_POST_SUSPEND:
270 	case PM_POST_HIBERNATION:
271 	case PM_POST_RESTORE:
272 		hung_detector_suspended = false;
273 		break;
274 	default:
275 		break;
276 	}
277 	return NOTIFY_OK;
278 }
279 
280 /*
281  * kthread which checks for tasks stuck in D state
282  */
watchdog(void * dummy)283 static int watchdog(void *dummy)
284 {
285 	unsigned long hung_last_checked = jiffies;
286 
287 	set_user_nice(current, 0);
288 
289 	for ( ; ; ) {
290 		unsigned long timeout = sysctl_hung_task_timeout_secs;
291 		unsigned long interval = sysctl_hung_task_check_interval_secs;
292 		long t;
293 
294 		if (interval == 0)
295 			interval = timeout;
296 		interval = min_t(unsigned long, interval, timeout);
297 		t = hung_timeout_jiffies(hung_last_checked, interval);
298 		if (t <= 0) {
299 			if (!atomic_xchg(&reset_hung_task, 0) &&
300 			    !hung_detector_suspended)
301 				check_hung_uninterruptible_tasks(timeout);
302 			hung_last_checked = jiffies;
303 			continue;
304 		}
305 		schedule_timeout_interruptible(t);
306 	}
307 
308 	return 0;
309 }
310 
hung_task_init(void)311 static int __init hung_task_init(void)
312 {
313 	atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
314 
315 	/* Disable hung task detector on suspend */
316 	pm_notifier(hungtask_pm_notify, 0);
317 
318 	watchdog_task = kthread_run(watchdog, NULL, "khungtaskd");
319 
320 	return 0;
321 }
322 subsys_initcall(hung_task_init);
323