1/*
2 * Detect Hung Task
3 *
4 * kernel/hung_task.c - kernel thread for detecting tasks stuck in D state
5 *
6 */
7
8#include <linux/mm.h>
9#include <linux/cpu.h>
10#include <linux/nmi.h>
11#include <linux/init.h>
12#include <linux/delay.h>
13#include <linux/freezer.h>
14#include <linux/kthread.h>
15#include <linux/lockdep.h>
16#include <linux/export.h>
17#include <linux/sysctl.h>
18#include <linux/suspend.h>
19#include <linux/utsname.h>
20#include <linux/sched/signal.h>
21#include <linux/sched/debug.h>
22#include <linux/sched/sysctl.h>
23
24#include <trace/events/sched.h>
25
26/*
27 * The number of tasks checked:
28 */
29int __read_mostly sysctl_hung_task_check_count = PID_MAX_LIMIT;
30
31/*
32 * Limit number of tasks checked in a batch.
33 *
34 * This value controls the preemptibility of khungtaskd since preemption
35 * is disabled during the critical section. It also controls the size of
36 * the RCU grace period. So it needs to be upper-bound.
37 */
38#define HUNG_TASK_LOCK_BREAK (HZ / 10)
39
40/*
41 * Zero means infinite timeout - no checking done:
42 */
43unsigned long __read_mostly sysctl_hung_task_timeout_secs = CONFIG_DEFAULT_HUNG_TASK_TIMEOUT;
44
45/*
46 * Zero (default value) means use sysctl_hung_task_timeout_secs:
47 */
48unsigned long __read_mostly sysctl_hung_task_check_interval_secs;
49
50int __read_mostly sysctl_hung_task_warnings = 10;
51
52static int __read_mostly did_panic;
53static bool hung_task_show_lock;
54static bool hung_task_call_panic;
55
56static struct task_struct *watchdog_task;
57
58/*
59 * Should we panic (and reboot, if panic_timeout= is set) when a
60 * hung task is detected:
61 */
62unsigned int __read_mostly sysctl_hung_task_panic =
63 CONFIG_BOOTPARAM_HUNG_TASK_PANIC_VALUE;
64
65static int __init hung_task_panic_setup(char *str)
66{
67 int rc = kstrtouint(str, 0, &sysctl_hung_task_panic);
68
69 if (rc)
70 return rc;
71 return 1;
72}
73__setup("hung_task_panic=", hung_task_panic_setup);
74
75static int
76hung_task_panic(struct notifier_block *this, unsigned long event, void *ptr)
77{
78 did_panic = 1;
79
80 return NOTIFY_DONE;
81}
82
83static struct notifier_block panic_block = {
84 .notifier_call = hung_task_panic,
85};
86
87static void check_hung_task(struct task_struct *t, unsigned long timeout)
88{
89 unsigned long switch_count = t->nvcsw + t->nivcsw;
90
91 /*
92 * Ensure the task is not frozen.
93 * Also, skip vfork and any other user process that freezer should skip.
94 */
95 if (unlikely(t->flags & (PF_FROZEN | PF_FREEZER_SKIP)))
96 return;
97
98 /*
99 * When a freshly created task is scheduled once, changes its state to
100 * TASK_UNINTERRUPTIBLE without having ever been switched out once, it
101 * musn't be checked.
102 */
103 if (unlikely(!switch_count))
104 return;
105
106 if (switch_count != t->last_switch_count) {
107 t->last_switch_count = switch_count;
108 t->last_switch_time = jiffies;
109 return;
110 }
111 if (time_is_after_jiffies(t->last_switch_time + timeout * HZ))
112 return;
113
114 trace_sched_process_hang(t);
115
116 if (sysctl_hung_task_panic) {
117 console_verbose();
118 hung_task_show_lock = true;
119 hung_task_call_panic = true;
120 }
121
122 /*
123 * Ok, the task did not get scheduled for more than 2 minutes,
124 * complain:
125 */
126 if (sysctl_hung_task_warnings) {
127 if (sysctl_hung_task_warnings > 0)
128 sysctl_hung_task_warnings--;
129 pr_err("INFO: task %s:%d blocked for more than %ld seconds.\n",
130 t->comm, t->pid, (jiffies - t->last_switch_time) / HZ);
131 pr_err(" %s %s %.*s\n",
132 print_tainted(), init_utsname()->release,
133 (int)strcspn(init_utsname()->version, " "),
134 init_utsname()->version);
135 pr_err("\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
136 " disables this message.\n");
137 sched_show_task(t);
138 hung_task_show_lock = true;
139 }
140
141 touch_nmi_watchdog();
142}
143
144/*
145 * To avoid extending the RCU grace period for an unbounded amount of time,
146 * periodically exit the critical section and enter a new one.
147 *
148 * For preemptible RCU it is sufficient to call rcu_read_unlock in order
149 * to exit the grace period. For classic RCU, a reschedule is required.
150 */
151static bool rcu_lock_break(struct task_struct *g, struct task_struct *t)
152{
153 bool can_cont;
154
155 get_task_struct(g);
156 get_task_struct(t);
157 rcu_read_unlock();
158 cond_resched();
159 rcu_read_lock();
160 can_cont = pid_alive(g) && pid_alive(t);
161 put_task_struct(t);
162 put_task_struct(g);
163
164 return can_cont;
165}
166
167/*
168 * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
169 * a really long time (120 seconds). If that happens, print out
170 * a warning.
171 */
172static void check_hung_uninterruptible_tasks(unsigned long timeout)
173{
174 int max_count = sysctl_hung_task_check_count;
175 unsigned long last_break = jiffies;
176 struct task_struct *g, *t;
177
178 /*
179 * If the system crashed already then all bets are off,
180 * do not report extra hung tasks:
181 */
182 if (test_taint(TAINT_DIE) || did_panic)
183 return;
184
185 hung_task_show_lock = false;
186 rcu_read_lock();
187 for_each_process_thread(g, t) {
188 if (!max_count--)
189 goto unlock;
190 if (time_after(jiffies, last_break + HUNG_TASK_LOCK_BREAK)) {
191 if (!rcu_lock_break(g, t))
192 goto unlock;
193 last_break = jiffies;
194 }
195 /* use "==" to skip the TASK_KILLABLE tasks waiting on NFS */
196 if (t->state == TASK_UNINTERRUPTIBLE)
197 check_hung_task(t, timeout);
198 }
199 unlock:
200 rcu_read_unlock();
201 if (hung_task_show_lock)
202 debug_show_all_locks();
203 if (hung_task_call_panic) {
204 trigger_all_cpu_backtrace();
205 panic("hung_task: blocked tasks");
206 }
207}
208
209static long hung_timeout_jiffies(unsigned long last_checked,
210 unsigned long timeout)
211{
212 /* timeout of 0 will disable the watchdog */
213 return timeout ? last_checked - jiffies + timeout * HZ :
214 MAX_SCHEDULE_TIMEOUT;
215}
216
217/*
218 * Process updating of timeout sysctl
219 */
220int proc_dohung_task_timeout_secs(struct ctl_table *table, int write,
221 void __user *buffer,
222 size_t *lenp, loff_t *ppos)
223{
224 int ret;
225
226 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
227
228 if (ret || !write)
229 goto out;
230
231 wake_up_process(watchdog_task);
232
233 out:
234 return ret;
235}
236
237static atomic_t reset_hung_task = ATOMIC_INIT(0);
238
239void reset_hung_task_detector(void)
240{
241 atomic_set(&reset_hung_task, 1);
242}
243EXPORT_SYMBOL_GPL(reset_hung_task_detector);
244
245static bool hung_detector_suspended;
246
247static int hungtask_pm_notify(struct notifier_block *self,
248 unsigned long action, void *hcpu)
249{
250 switch (action) {
251 case PM_SUSPEND_PREPARE:
252 case PM_HIBERNATION_PREPARE:
253 case PM_RESTORE_PREPARE:
254 hung_detector_suspended = true;
255 break;
256 case PM_POST_SUSPEND:
257 case PM_POST_HIBERNATION:
258 case PM_POST_RESTORE:
259 hung_detector_suspended = false;
260 break;
261 default:
262 break;
263 }
264 return NOTIFY_OK;
265}
266
267/*
268 * kthread which checks for tasks stuck in D state
269 */
270static int watchdog(void *dummy)
271{
272 unsigned long hung_last_checked = jiffies;
273
274 set_user_nice(current, 0);
275
276 for ( ; ; ) {
277 unsigned long timeout = sysctl_hung_task_timeout_secs;
278 unsigned long interval = sysctl_hung_task_check_interval_secs;
279 long t;
280
281 if (interval == 0)
282 interval = timeout;
283 interval = min_t(unsigned long, interval, timeout);
284 t = hung_timeout_jiffies(hung_last_checked, interval);
285 if (t <= 0) {
286 if (!atomic_xchg(&reset_hung_task, 0) &&
287 !hung_detector_suspended)
288 check_hung_uninterruptible_tasks(timeout);
289 hung_last_checked = jiffies;
290 continue;
291 }
292 schedule_timeout_interruptible(t);
293 }
294
295 return 0;
296}
297
298static int __init hung_task_init(void)
299{
300 atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
301
302 /* Disable hung task detector on suspend */
303 pm_notifier(hungtask_pm_notify, 0);
304
305 watchdog_task = kthread_run(watchdog, NULL, "khungtaskd");
306
307 return 0;
308}
309subsys_initcall(hung_task_init);
310