Loading...
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 */
30int __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 */
44unsigned 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 */
49unsigned long __read_mostly sysctl_hung_task_check_interval_secs;
50
51int __read_mostly sysctl_hung_task_warnings = 10;
52
53static int __read_mostly did_panic;
54static bool hung_task_show_lock;
55static bool hung_task_call_panic;
56
57static 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 */
63unsigned int __read_mostly sysctl_hung_task_panic =
64 CONFIG_BOOTPARAM_HUNG_TASK_PANIC_VALUE;
65
66static int __init hung_task_panic_setup(char *str)
67{
68 int rc = kstrtouint(str, 0, &sysctl_hung_task_panic);
69
70 if (rc)
71 return rc;
72 return 1;
73}
74__setup("hung_task_panic=", hung_task_panic_setup);
75
76static int
77hung_task_panic(struct notifier_block *this, unsigned long event, void *ptr)
78{
79 did_panic = 1;
80
81 return NOTIFY_DONE;
82}
83
84static struct notifier_block panic_block = {
85 .notifier_call = hung_task_panic,
86};
87
88static void check_hung_task(struct task_struct *t, unsigned long timeout)
89{
90 unsigned long switch_count = t->nvcsw + t->nivcsw;
91
92 /*
93 * Ensure the task is not frozen.
94 * Also, skip vfork and any other user process that freezer should skip.
95 */
96 if (unlikely(t->flags & (PF_FROZEN | PF_FREEZER_SKIP)))
97 return;
98
99 /*
100 * When a freshly created task is scheduled once, changes its state to
101 * TASK_UNINTERRUPTIBLE without having ever been switched out once, it
102 * musn't be checked.
103 */
104 if (unlikely(!switch_count))
105 return;
106
107 if (switch_count != t->last_switch_count) {
108 t->last_switch_count = switch_count;
109 t->last_switch_time = jiffies;
110 return;
111 }
112 if (time_is_after_jiffies(t->last_switch_time + timeout * HZ))
113 return;
114
115 trace_sched_process_hang(t);
116
117 if (sysctl_hung_task_panic) {
118 console_verbose();
119 hung_task_show_lock = true;
120 hung_task_call_panic = true;
121 }
122
123 /*
124 * Ok, the task did not get scheduled for more than 2 minutes,
125 * complain:
126 */
127 if (sysctl_hung_task_warnings) {
128 if (sysctl_hung_task_warnings > 0)
129 sysctl_hung_task_warnings--;
130 pr_err("INFO: task %s:%d blocked for more than %ld seconds.\n",
131 t->comm, t->pid, (jiffies - t->last_switch_time) / HZ);
132 pr_err(" %s %s %.*s\n",
133 print_tainted(), init_utsname()->release,
134 (int)strcspn(init_utsname()->version, " "),
135 init_utsname()->version);
136 pr_err("\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
137 " disables this message.\n");
138 sched_show_task(t);
139 hung_task_show_lock = true;
140 }
141
142 touch_nmi_watchdog();
143}
144
145/*
146 * To avoid extending the RCU grace period for an unbounded amount of time,
147 * periodically exit the critical section and enter a new one.
148 *
149 * For preemptible RCU it is sufficient to call rcu_read_unlock in order
150 * to exit the grace period. For classic RCU, a reschedule is required.
151 */
152static bool rcu_lock_break(struct task_struct *g, struct task_struct *t)
153{
154 bool can_cont;
155
156 get_task_struct(g);
157 get_task_struct(t);
158 rcu_read_unlock();
159 cond_resched();
160 rcu_read_lock();
161 can_cont = pid_alive(g) && pid_alive(t);
162 put_task_struct(t);
163 put_task_struct(g);
164
165 return can_cont;
166}
167
168/*
169 * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
170 * a really long time (120 seconds). If that happens, print out
171 * a warning.
172 */
173static void check_hung_uninterruptible_tasks(unsigned long timeout)
174{
175 int max_count = sysctl_hung_task_check_count;
176 unsigned long last_break = jiffies;
177 struct task_struct *g, *t;
178
179 /*
180 * If the system crashed already then all bets are off,
181 * do not report extra hung tasks:
182 */
183 if (test_taint(TAINT_DIE) || did_panic)
184 return;
185
186 hung_task_show_lock = false;
187 rcu_read_lock();
188 for_each_process_thread(g, t) {
189 if (!max_count--)
190 goto unlock;
191 if (time_after(jiffies, last_break + HUNG_TASK_LOCK_BREAK)) {
192 if (!rcu_lock_break(g, t))
193 goto unlock;
194 last_break = jiffies;
195 }
196 /* use "==" to skip the TASK_KILLABLE tasks waiting on NFS */
197 if (t->state == TASK_UNINTERRUPTIBLE)
198 check_hung_task(t, timeout);
199 }
200 unlock:
201 rcu_read_unlock();
202 if (hung_task_show_lock)
203 debug_show_all_locks();
204 if (hung_task_call_panic) {
205 trigger_all_cpu_backtrace();
206 panic("hung_task: blocked tasks");
207 }
208}
209
210static long hung_timeout_jiffies(unsigned long last_checked,
211 unsigned long timeout)
212{
213 /* timeout of 0 will disable the watchdog */
214 return timeout ? last_checked - jiffies + timeout * HZ :
215 MAX_SCHEDULE_TIMEOUT;
216}
217
218/*
219 * Process updating of timeout sysctl
220 */
221int proc_dohung_task_timeout_secs(struct ctl_table *table, int write,
222 void __user *buffer,
223 size_t *lenp, loff_t *ppos)
224{
225 int ret;
226
227 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
228
229 if (ret || !write)
230 goto out;
231
232 wake_up_process(watchdog_task);
233
234 out:
235 return ret;
236}
237
238static atomic_t reset_hung_task = ATOMIC_INIT(0);
239
240void reset_hung_task_detector(void)
241{
242 atomic_set(&reset_hung_task, 1);
243}
244EXPORT_SYMBOL_GPL(reset_hung_task_detector);
245
246static bool hung_detector_suspended;
247
248static int hungtask_pm_notify(struct notifier_block *self,
249 unsigned long action, void *hcpu)
250{
251 switch (action) {
252 case PM_SUSPEND_PREPARE:
253 case PM_HIBERNATION_PREPARE:
254 case PM_RESTORE_PREPARE:
255 hung_detector_suspended = true;
256 break;
257 case PM_POST_SUSPEND:
258 case PM_POST_HIBERNATION:
259 case PM_POST_RESTORE:
260 hung_detector_suspended = false;
261 break;
262 default:
263 break;
264 }
265 return NOTIFY_OK;
266}
267
268/*
269 * kthread which checks for tasks stuck in D state
270 */
271static int watchdog(void *dummy)
272{
273 unsigned long hung_last_checked = jiffies;
274
275 set_user_nice(current, 0);
276
277 for ( ; ; ) {
278 unsigned long timeout = sysctl_hung_task_timeout_secs;
279 unsigned long interval = sysctl_hung_task_check_interval_secs;
280 long t;
281
282 if (interval == 0)
283 interval = timeout;
284 interval = min_t(unsigned long, interval, timeout);
285 t = hung_timeout_jiffies(hung_last_checked, interval);
286 if (t <= 0) {
287 if (!atomic_xchg(&reset_hung_task, 0) &&
288 !hung_detector_suspended)
289 check_hung_uninterruptible_tasks(timeout);
290 hung_last_checked = jiffies;
291 continue;
292 }
293 schedule_timeout_interruptible(t);
294 }
295
296 return 0;
297}
298
299static int __init hung_task_init(void)
300{
301 atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
302
303 /* Disable hung task detector on suspend */
304 pm_notifier(hungtask_pm_notify, 0);
305
306 watchdog_task = kthread_run(watchdog, NULL, "khungtaskd");
307
308 return 0;
309}
310subsys_initcall(hung_task_init);
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 */
30int __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 */
44unsigned 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 */
49unsigned long __read_mostly sysctl_hung_task_check_interval_secs;
50
51int __read_mostly sysctl_hung_task_warnings = 10;
52
53static int __read_mostly did_panic;
54static bool hung_task_show_lock;
55static bool hung_task_call_panic;
56static bool hung_task_show_all_bt;
57
58static struct task_struct *watchdog_task;
59
60#ifdef CONFIG_SMP
61/*
62 * Should we dump all CPUs backtraces in a hung task event?
63 * Defaults to 0, can be changed via sysctl.
64 */
65unsigned int __read_mostly sysctl_hung_task_all_cpu_backtrace;
66#endif /* CONFIG_SMP */
67
68/*
69 * Should we panic (and reboot, if panic_timeout= is set) when a
70 * hung task is detected:
71 */
72unsigned int __read_mostly sysctl_hung_task_panic =
73 CONFIG_BOOTPARAM_HUNG_TASK_PANIC_VALUE;
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 if (sysctl_hung_task_all_cpu_backtrace)
141 hung_task_show_all_bt = true;
142 }
143
144 touch_nmi_watchdog();
145}
146
147/*
148 * To avoid extending the RCU grace period for an unbounded amount of time,
149 * periodically exit the critical section and enter a new one.
150 *
151 * For preemptible RCU it is sufficient to call rcu_read_unlock in order
152 * to exit the grace period. For classic RCU, a reschedule is required.
153 */
154static bool rcu_lock_break(struct task_struct *g, struct task_struct *t)
155{
156 bool can_cont;
157
158 get_task_struct(g);
159 get_task_struct(t);
160 rcu_read_unlock();
161 cond_resched();
162 rcu_read_lock();
163 can_cont = pid_alive(g) && pid_alive(t);
164 put_task_struct(t);
165 put_task_struct(g);
166
167 return can_cont;
168}
169
170/*
171 * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
172 * a really long time (120 seconds). If that happens, print out
173 * a warning.
174 */
175static void check_hung_uninterruptible_tasks(unsigned long timeout)
176{
177 int max_count = sysctl_hung_task_check_count;
178 unsigned long last_break = jiffies;
179 struct task_struct *g, *t;
180
181 /*
182 * If the system crashed already then all bets are off,
183 * do not report extra hung tasks:
184 */
185 if (test_taint(TAINT_DIE) || did_panic)
186 return;
187
188 hung_task_show_lock = false;
189 rcu_read_lock();
190 for_each_process_thread(g, t) {
191 if (!max_count--)
192 goto unlock;
193 if (time_after(jiffies, last_break + HUNG_TASK_LOCK_BREAK)) {
194 if (!rcu_lock_break(g, t))
195 goto unlock;
196 last_break = jiffies;
197 }
198 /* use "==" to skip the TASK_KILLABLE tasks waiting on NFS */
199 if (t->state == TASK_UNINTERRUPTIBLE)
200 check_hung_task(t, timeout);
201 }
202 unlock:
203 rcu_read_unlock();
204 if (hung_task_show_lock)
205 debug_show_all_locks();
206
207 if (hung_task_show_all_bt) {
208 hung_task_show_all_bt = false;
209 trigger_all_cpu_backtrace();
210 }
211
212 if (hung_task_call_panic)
213 panic("hung_task: blocked tasks");
214}
215
216static long hung_timeout_jiffies(unsigned long last_checked,
217 unsigned long timeout)
218{
219 /* timeout of 0 will disable the watchdog */
220 return timeout ? last_checked - jiffies + timeout * HZ :
221 MAX_SCHEDULE_TIMEOUT;
222}
223
224/*
225 * Process updating of timeout sysctl
226 */
227int proc_dohung_task_timeout_secs(struct ctl_table *table, int write,
228 void __user *buffer,
229 size_t *lenp, loff_t *ppos)
230{
231 int ret;
232
233 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
234
235 if (ret || !write)
236 goto out;
237
238 wake_up_process(watchdog_task);
239
240 out:
241 return ret;
242}
243
244static atomic_t reset_hung_task = ATOMIC_INIT(0);
245
246void reset_hung_task_detector(void)
247{
248 atomic_set(&reset_hung_task, 1);
249}
250EXPORT_SYMBOL_GPL(reset_hung_task_detector);
251
252static bool hung_detector_suspended;
253
254static int hungtask_pm_notify(struct notifier_block *self,
255 unsigned long action, void *hcpu)
256{
257 switch (action) {
258 case PM_SUSPEND_PREPARE:
259 case PM_HIBERNATION_PREPARE:
260 case PM_RESTORE_PREPARE:
261 hung_detector_suspended = true;
262 break;
263 case PM_POST_SUSPEND:
264 case PM_POST_HIBERNATION:
265 case PM_POST_RESTORE:
266 hung_detector_suspended = false;
267 break;
268 default:
269 break;
270 }
271 return NOTIFY_OK;
272}
273
274/*
275 * kthread which checks for tasks stuck in D state
276 */
277static int watchdog(void *dummy)
278{
279 unsigned long hung_last_checked = jiffies;
280
281 set_user_nice(current, 0);
282
283 for ( ; ; ) {
284 unsigned long timeout = sysctl_hung_task_timeout_secs;
285 unsigned long interval = sysctl_hung_task_check_interval_secs;
286 long t;
287
288 if (interval == 0)
289 interval = timeout;
290 interval = min_t(unsigned long, interval, timeout);
291 t = hung_timeout_jiffies(hung_last_checked, interval);
292 if (t <= 0) {
293 if (!atomic_xchg(&reset_hung_task, 0) &&
294 !hung_detector_suspended)
295 check_hung_uninterruptible_tasks(timeout);
296 hung_last_checked = jiffies;
297 continue;
298 }
299 schedule_timeout_interruptible(t);
300 }
301
302 return 0;
303}
304
305static int __init hung_task_init(void)
306{
307 atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
308
309 /* Disable hung task detector on suspend */
310 pm_notifier(hungtask_pm_notify, 0);
311
312 watchdog_task = kthread_run(watchdog, NULL, "khungtaskd");
313
314 return 0;
315}
316subsys_initcall(hung_task_init);