xref: /linux-6.15/kernel/hung_task.c (revision b467f3ef)
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/sysctl.h>
19 #include <linux/suspend.h>
20 #include <linux/utsname.h>
21 #include <linux/sched/signal.h>
22 #include <linux/sched/debug.h>
23 #include <linux/sched/sysctl.h>
24 
25 #include <trace/events/sched.h>
26 
27 /*
28  * The number of tasks checked:
29  */
30 int __read_mostly sysctl_hung_task_check_count = PID_MAX_LIMIT;
31 
32 /*
33  * Limit number of tasks checked in a batch.
34  *
35  * This value controls the preemptibility of khungtaskd since preemption
36  * is disabled during the critical section. It also controls the size of
37  * the RCU grace period. So it needs to be upper-bound.
38  */
39 #define HUNG_TASK_LOCK_BREAK (HZ / 10)
40 
41 /*
42  * Zero means infinite timeout - no checking done:
43  */
44 unsigned long __read_mostly sysctl_hung_task_timeout_secs = CONFIG_DEFAULT_HUNG_TASK_TIMEOUT;
45 
46 /*
47  * Zero (default value) means use sysctl_hung_task_timeout_secs:
48  */
49 unsigned long __read_mostly sysctl_hung_task_check_interval_secs;
50 
51 int __read_mostly sysctl_hung_task_warnings = 10;
52 
53 static int __read_mostly did_panic;
54 static bool hung_task_show_lock;
55 static bool hung_task_call_panic;
56 
57 static struct task_struct *watchdog_task;
58 
59 /*
60  * Should we panic (and reboot, if panic_timeout= is set) when a
61  * hung task is detected:
62  */
63 unsigned int __read_mostly sysctl_hung_task_panic =
64 				CONFIG_BOOTPARAM_HUNG_TASK_PANIC_VALUE;
65 
66 static int
67 hung_task_panic(struct notifier_block *this, unsigned long event, void *ptr)
68 {
69 	did_panic = 1;
70 
71 	return NOTIFY_DONE;
72 }
73 
74 static struct notifier_block panic_block = {
75 	.notifier_call = hung_task_panic,
76 };
77 
78 static void check_hung_task(struct task_struct *t, unsigned long timeout)
79 {
80 	unsigned long switch_count = t->nvcsw + t->nivcsw;
81 
82 	/*
83 	 * Ensure the task is not frozen.
84 	 * Also, skip vfork and any other user process that freezer should skip.
85 	 */
86 	if (unlikely(t->flags & (PF_FROZEN | PF_FREEZER_SKIP)))
87 	    return;
88 
89 	/*
90 	 * When a freshly created task is scheduled once, changes its state to
91 	 * TASK_UNINTERRUPTIBLE without having ever been switched out once, it
92 	 * musn't be checked.
93 	 */
94 	if (unlikely(!switch_count))
95 		return;
96 
97 	if (switch_count != t->last_switch_count) {
98 		t->last_switch_count = switch_count;
99 		t->last_switch_time = jiffies;
100 		return;
101 	}
102 	if (time_is_after_jiffies(t->last_switch_time + timeout * HZ))
103 		return;
104 
105 	trace_sched_process_hang(t);
106 
107 	if (sysctl_hung_task_panic) {
108 		console_verbose();
109 		hung_task_show_lock = true;
110 		hung_task_call_panic = true;
111 	}
112 
113 	/*
114 	 * Ok, the task did not get scheduled for more than 2 minutes,
115 	 * complain:
116 	 */
117 	if (sysctl_hung_task_warnings) {
118 		if (sysctl_hung_task_warnings > 0)
119 			sysctl_hung_task_warnings--;
120 		pr_err("INFO: task %s:%d blocked for more than %ld seconds.\n",
121 		       t->comm, t->pid, (jiffies - t->last_switch_time) / HZ);
122 		pr_err("      %s %s %.*s\n",
123 			print_tainted(), init_utsname()->release,
124 			(int)strcspn(init_utsname()->version, " "),
125 			init_utsname()->version);
126 		pr_err("\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
127 			" disables this message.\n");
128 		sched_show_task(t);
129 		hung_task_show_lock = true;
130 	}
131 
132 	touch_nmi_watchdog();
133 }
134 
135 /*
136  * To avoid extending the RCU grace period for an unbounded amount of time,
137  * periodically exit the critical section and enter a new one.
138  *
139  * For preemptible RCU it is sufficient to call rcu_read_unlock in order
140  * to exit the grace period. For classic RCU, a reschedule is required.
141  */
142 static bool rcu_lock_break(struct task_struct *g, struct task_struct *t)
143 {
144 	bool can_cont;
145 
146 	get_task_struct(g);
147 	get_task_struct(t);
148 	rcu_read_unlock();
149 	cond_resched();
150 	rcu_read_lock();
151 	can_cont = pid_alive(g) && pid_alive(t);
152 	put_task_struct(t);
153 	put_task_struct(g);
154 
155 	return can_cont;
156 }
157 
158 /*
159  * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
160  * a really long time (120 seconds). If that happens, print out
161  * a warning.
162  */
163 static void check_hung_uninterruptible_tasks(unsigned long timeout)
164 {
165 	int max_count = sysctl_hung_task_check_count;
166 	unsigned long last_break = jiffies;
167 	struct task_struct *g, *t;
168 
169 	/*
170 	 * If the system crashed already then all bets are off,
171 	 * do not report extra hung tasks:
172 	 */
173 	if (test_taint(TAINT_DIE) || did_panic)
174 		return;
175 
176 	hung_task_show_lock = false;
177 	rcu_read_lock();
178 	for_each_process_thread(g, t) {
179 		if (!max_count--)
180 			goto unlock;
181 		if (time_after(jiffies, last_break + HUNG_TASK_LOCK_BREAK)) {
182 			if (!rcu_lock_break(g, t))
183 				goto unlock;
184 			last_break = jiffies;
185 		}
186 		/* use "==" to skip the TASK_KILLABLE tasks waiting on NFS */
187 		if (t->state == TASK_UNINTERRUPTIBLE)
188 			check_hung_task(t, timeout);
189 	}
190  unlock:
191 	rcu_read_unlock();
192 	if (hung_task_show_lock)
193 		debug_show_all_locks();
194 	if (hung_task_call_panic) {
195 		trigger_all_cpu_backtrace();
196 		panic("hung_task: blocked tasks");
197 	}
198 }
199 
200 static long hung_timeout_jiffies(unsigned long last_checked,
201 				 unsigned long timeout)
202 {
203 	/* timeout of 0 will disable the watchdog */
204 	return timeout ? last_checked - jiffies + timeout * HZ :
205 		MAX_SCHEDULE_TIMEOUT;
206 }
207 
208 /*
209  * Process updating of timeout sysctl
210  */
211 int proc_dohung_task_timeout_secs(struct ctl_table *table, int write,
212 				  void __user *buffer,
213 				  size_t *lenp, loff_t *ppos)
214 {
215 	int ret;
216 
217 	ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
218 
219 	if (ret || !write)
220 		goto out;
221 
222 	wake_up_process(watchdog_task);
223 
224  out:
225 	return ret;
226 }
227 
228 static atomic_t reset_hung_task = ATOMIC_INIT(0);
229 
230 void reset_hung_task_detector(void)
231 {
232 	atomic_set(&reset_hung_task, 1);
233 }
234 EXPORT_SYMBOL_GPL(reset_hung_task_detector);
235 
236 static bool hung_detector_suspended;
237 
238 static int hungtask_pm_notify(struct notifier_block *self,
239 			      unsigned long action, void *hcpu)
240 {
241 	switch (action) {
242 	case PM_SUSPEND_PREPARE:
243 	case PM_HIBERNATION_PREPARE:
244 	case PM_RESTORE_PREPARE:
245 		hung_detector_suspended = true;
246 		break;
247 	case PM_POST_SUSPEND:
248 	case PM_POST_HIBERNATION:
249 	case PM_POST_RESTORE:
250 		hung_detector_suspended = false;
251 		break;
252 	default:
253 		break;
254 	}
255 	return NOTIFY_OK;
256 }
257 
258 /*
259  * kthread which checks for tasks stuck in D state
260  */
261 static int watchdog(void *dummy)
262 {
263 	unsigned long hung_last_checked = jiffies;
264 
265 	set_user_nice(current, 0);
266 
267 	for ( ; ; ) {
268 		unsigned long timeout = sysctl_hung_task_timeout_secs;
269 		unsigned long interval = sysctl_hung_task_check_interval_secs;
270 		long t;
271 
272 		if (interval == 0)
273 			interval = timeout;
274 		interval = min_t(unsigned long, interval, timeout);
275 		t = hung_timeout_jiffies(hung_last_checked, interval);
276 		if (t <= 0) {
277 			if (!atomic_xchg(&reset_hung_task, 0) &&
278 			    !hung_detector_suspended)
279 				check_hung_uninterruptible_tasks(timeout);
280 			hung_last_checked = jiffies;
281 			continue;
282 		}
283 		schedule_timeout_interruptible(t);
284 	}
285 
286 	return 0;
287 }
288 
289 static int __init hung_task_init(void)
290 {
291 	atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
292 
293 	/* Disable hung task detector on suspend */
294 	pm_notifier(hungtask_pm_notify, 0);
295 
296 	watchdog_task = kthread_run(watchdog, NULL, "khungtaskd");
297 
298 	return 0;
299 }
300 subsys_initcall(hung_task_init);
301