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 static 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 EXPORT_SYMBOL_GPL(sysctl_hung_task_timeout_secs);
47
48 /*
49 * Zero (default value) means use sysctl_hung_task_timeout_secs:
50 */
51 static unsigned long __read_mostly sysctl_hung_task_check_interval_secs;
52
53 static int __read_mostly sysctl_hung_task_warnings = 10;
54
55 static int __read_mostly did_panic;
56 static bool hung_task_show_lock;
57 static bool hung_task_call_panic;
58 static bool hung_task_show_all_bt;
59
60 static struct task_struct *watchdog_task;
61
62 #ifdef CONFIG_SMP
63 /*
64 * Should we dump all CPUs backtraces in a hung task event?
65 * Defaults to 0, can be changed via sysctl.
66 */
67 static unsigned int __read_mostly sysctl_hung_task_all_cpu_backtrace;
68 #else
69 #define sysctl_hung_task_all_cpu_backtrace 0
70 #endif /* CONFIG_SMP */
71
72 /*
73 * Should we panic (and reboot, if panic_timeout= is set) when a
74 * hung task is detected:
75 */
76 static unsigned int __read_mostly sysctl_hung_task_panic =
77 IS_ENABLED(CONFIG_BOOTPARAM_HUNG_TASK_PANIC);
78
79 static int
hung_task_panic(struct notifier_block * this,unsigned long event,void * ptr)80 hung_task_panic(struct notifier_block *this, unsigned long event, void *ptr)
81 {
82 did_panic = 1;
83
84 return NOTIFY_DONE;
85 }
86
87 static struct notifier_block panic_block = {
88 .notifier_call = hung_task_panic,
89 };
90
check_hung_task(struct task_struct * t,unsigned long timeout)91 static void check_hung_task(struct task_struct *t, unsigned long timeout)
92 {
93 unsigned long switch_count = t->nvcsw + t->nivcsw;
94
95 /*
96 * Ensure the task is not frozen.
97 * Also, skip vfork and any other user process that freezer should skip.
98 */
99 if (unlikely(READ_ONCE(t->__state) & TASK_FROZEN))
100 return;
101
102 /*
103 * When a freshly created task is scheduled once, changes its state to
104 * TASK_UNINTERRUPTIBLE without having ever been switched out once, it
105 * musn't be checked.
106 */
107 if (unlikely(!switch_count))
108 return;
109
110 if (switch_count != t->last_switch_count) {
111 t->last_switch_count = switch_count;
112 t->last_switch_time = jiffies;
113 return;
114 }
115 if (time_is_after_jiffies(t->last_switch_time + timeout * HZ))
116 return;
117
118 trace_sched_process_hang(t);
119
120 if (sysctl_hung_task_panic) {
121 console_verbose();
122 hung_task_show_lock = true;
123 hung_task_call_panic = true;
124 }
125
126 /*
127 * Ok, the task did not get scheduled for more than 2 minutes,
128 * complain:
129 */
130 if (sysctl_hung_task_warnings) {
131 if (sysctl_hung_task_warnings > 0)
132 sysctl_hung_task_warnings--;
133 pr_err("INFO: task %s:%d blocked for more than %ld seconds.\n",
134 t->comm, t->pid, (jiffies - t->last_switch_time) / HZ);
135 pr_err(" %s %s %.*s\n",
136 print_tainted(), init_utsname()->release,
137 (int)strcspn(init_utsname()->version, " "),
138 init_utsname()->version);
139 pr_err("\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
140 " disables this message.\n");
141 sched_show_task(t);
142 hung_task_show_lock = true;
143
144 if (sysctl_hung_task_all_cpu_backtrace)
145 hung_task_show_all_bt = true;
146 if (!sysctl_hung_task_warnings)
147 pr_info("Future hung task reports are suppressed, see sysctl kernel.hung_task_warnings\n");
148 }
149
150 touch_nmi_watchdog();
151 }
152
153 /*
154 * To avoid extending the RCU grace period for an unbounded amount of time,
155 * periodically exit the critical section and enter a new one.
156 *
157 * For preemptible RCU it is sufficient to call rcu_read_unlock in order
158 * to exit the grace period. For classic RCU, a reschedule is required.
159 */
rcu_lock_break(struct task_struct * g,struct task_struct * t)160 static bool rcu_lock_break(struct task_struct *g, struct task_struct *t)
161 {
162 bool can_cont;
163
164 get_task_struct(g);
165 get_task_struct(t);
166 rcu_read_unlock();
167 cond_resched();
168 rcu_read_lock();
169 can_cont = pid_alive(g) && pid_alive(t);
170 put_task_struct(t);
171 put_task_struct(g);
172
173 return can_cont;
174 }
175
176 /*
177 * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
178 * a really long time (120 seconds). If that happens, print out
179 * a warning.
180 */
check_hung_uninterruptible_tasks(unsigned long timeout)181 static void check_hung_uninterruptible_tasks(unsigned long timeout)
182 {
183 int max_count = sysctl_hung_task_check_count;
184 unsigned long last_break = jiffies;
185 struct task_struct *g, *t;
186
187 /*
188 * If the system crashed already then all bets are off,
189 * do not report extra hung tasks:
190 */
191 if (test_taint(TAINT_DIE) || did_panic)
192 return;
193
194 hung_task_show_lock = false;
195 rcu_read_lock();
196 for_each_process_thread(g, t) {
197 unsigned int state;
198
199 if (!max_count--)
200 goto unlock;
201 if (time_after(jiffies, last_break + HUNG_TASK_LOCK_BREAK)) {
202 if (!rcu_lock_break(g, t))
203 goto unlock;
204 last_break = jiffies;
205 }
206 /*
207 * skip the TASK_KILLABLE tasks -- these can be killed
208 * skip the TASK_IDLE tasks -- those are genuinely idle
209 */
210 state = READ_ONCE(t->__state);
211 if ((state & TASK_UNINTERRUPTIBLE) &&
212 !(state & TASK_WAKEKILL) &&
213 !(state & TASK_NOLOAD))
214 check_hung_task(t, timeout);
215 }
216 unlock:
217 rcu_read_unlock();
218 if (hung_task_show_lock)
219 debug_show_all_locks();
220
221 if (hung_task_show_all_bt) {
222 hung_task_show_all_bt = false;
223 trigger_all_cpu_backtrace();
224 }
225
226 if (hung_task_call_panic)
227 panic("hung_task: blocked tasks");
228 }
229
hung_timeout_jiffies(unsigned long last_checked,unsigned long timeout)230 static long hung_timeout_jiffies(unsigned long last_checked,
231 unsigned long timeout)
232 {
233 /* timeout of 0 will disable the watchdog */
234 return timeout ? last_checked - jiffies + timeout * HZ :
235 MAX_SCHEDULE_TIMEOUT;
236 }
237
238 #ifdef CONFIG_SYSCTL
239 /*
240 * Process updating of timeout sysctl
241 */
proc_dohung_task_timeout_secs(struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)242 static int proc_dohung_task_timeout_secs(struct ctl_table *table, int write,
243 void *buffer,
244 size_t *lenp, loff_t *ppos)
245 {
246 int ret;
247
248 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
249
250 if (ret || !write)
251 goto out;
252
253 wake_up_process(watchdog_task);
254
255 out:
256 return ret;
257 }
258
259 /*
260 * This is needed for proc_doulongvec_minmax of sysctl_hung_task_timeout_secs
261 * and hung_task_check_interval_secs
262 */
263 static const unsigned long hung_task_timeout_max = (LONG_MAX / HZ);
264 static struct ctl_table hung_task_sysctls[] = {
265 #ifdef CONFIG_SMP
266 {
267 .procname = "hung_task_all_cpu_backtrace",
268 .data = &sysctl_hung_task_all_cpu_backtrace,
269 .maxlen = sizeof(int),
270 .mode = 0644,
271 .proc_handler = proc_dointvec_minmax,
272 .extra1 = SYSCTL_ZERO,
273 .extra2 = SYSCTL_ONE,
274 },
275 #endif /* CONFIG_SMP */
276 {
277 .procname = "hung_task_panic",
278 .data = &sysctl_hung_task_panic,
279 .maxlen = sizeof(int),
280 .mode = 0644,
281 .proc_handler = proc_dointvec_minmax,
282 .extra1 = SYSCTL_ZERO,
283 .extra2 = SYSCTL_ONE,
284 },
285 {
286 .procname = "hung_task_check_count",
287 .data = &sysctl_hung_task_check_count,
288 .maxlen = sizeof(int),
289 .mode = 0644,
290 .proc_handler = proc_dointvec_minmax,
291 .extra1 = SYSCTL_ZERO,
292 },
293 {
294 .procname = "hung_task_timeout_secs",
295 .data = &sysctl_hung_task_timeout_secs,
296 .maxlen = sizeof(unsigned long),
297 .mode = 0644,
298 .proc_handler = proc_dohung_task_timeout_secs,
299 .extra2 = (void *)&hung_task_timeout_max,
300 },
301 {
302 .procname = "hung_task_check_interval_secs",
303 .data = &sysctl_hung_task_check_interval_secs,
304 .maxlen = sizeof(unsigned long),
305 .mode = 0644,
306 .proc_handler = proc_dohung_task_timeout_secs,
307 .extra2 = (void *)&hung_task_timeout_max,
308 },
309 {
310 .procname = "hung_task_warnings",
311 .data = &sysctl_hung_task_warnings,
312 .maxlen = sizeof(int),
313 .mode = 0644,
314 .proc_handler = proc_dointvec_minmax,
315 .extra1 = SYSCTL_NEG_ONE,
316 },
317 {}
318 };
319
hung_task_sysctl_init(void)320 static void __init hung_task_sysctl_init(void)
321 {
322 register_sysctl_init("kernel", hung_task_sysctls);
323 }
324 #else
325 #define hung_task_sysctl_init() do { } while (0)
326 #endif /* CONFIG_SYSCTL */
327
328
329 static atomic_t reset_hung_task = ATOMIC_INIT(0);
330
reset_hung_task_detector(void)331 void reset_hung_task_detector(void)
332 {
333 atomic_set(&reset_hung_task, 1);
334 }
335 EXPORT_SYMBOL_GPL(reset_hung_task_detector);
336
337 static bool hung_detector_suspended;
338
hungtask_pm_notify(struct notifier_block * self,unsigned long action,void * hcpu)339 static int hungtask_pm_notify(struct notifier_block *self,
340 unsigned long action, void *hcpu)
341 {
342 switch (action) {
343 case PM_SUSPEND_PREPARE:
344 case PM_HIBERNATION_PREPARE:
345 case PM_RESTORE_PREPARE:
346 hung_detector_suspended = true;
347 break;
348 case PM_POST_SUSPEND:
349 case PM_POST_HIBERNATION:
350 case PM_POST_RESTORE:
351 hung_detector_suspended = false;
352 break;
353 default:
354 break;
355 }
356 return NOTIFY_OK;
357 }
358
359 /*
360 * kthread which checks for tasks stuck in D state
361 */
watchdog(void * dummy)362 static int watchdog(void *dummy)
363 {
364 unsigned long hung_last_checked = jiffies;
365
366 set_user_nice(current, 0);
367
368 for ( ; ; ) {
369 unsigned long timeout = sysctl_hung_task_timeout_secs;
370 unsigned long interval = sysctl_hung_task_check_interval_secs;
371 long t;
372
373 if (interval == 0)
374 interval = timeout;
375 interval = min_t(unsigned long, interval, timeout);
376 t = hung_timeout_jiffies(hung_last_checked, interval);
377 if (t <= 0) {
378 if (!atomic_xchg(&reset_hung_task, 0) &&
379 !hung_detector_suspended)
380 check_hung_uninterruptible_tasks(timeout);
381 hung_last_checked = jiffies;
382 continue;
383 }
384 schedule_timeout_interruptible(t);
385 }
386
387 return 0;
388 }
389
hung_task_init(void)390 static int __init hung_task_init(void)
391 {
392 atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
393
394 /* Disable hung task detector on suspend */
395 pm_notifier(hungtask_pm_notify, 0);
396
397 watchdog_task = kthread_run(watchdog, NULL, "khungtaskd");
398 hung_task_sysctl_init();
399
400 return 0;
401 }
402 subsys_initcall(hung_task_init);
403