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 static unsigned int __read_mostly sysctl_hung_task_all_cpu_backtrace;
69 #else
70 #define sysctl_hung_task_all_cpu_backtrace 0
71 #endif /* CONFIG_SMP */
72
73 /*
74 * Should we panic (and reboot, if panic_timeout= is set) when a
75 * hung task is detected:
76 */
77 unsigned int __read_mostly sysctl_hung_task_panic =
78 IS_ENABLED(CONFIG_BOOTPARAM_HUNG_TASK_PANIC);
79
80 static int
hung_task_panic(struct notifier_block * this,unsigned long event,void * ptr)81 hung_task_panic(struct notifier_block *this, unsigned long event, void *ptr)
82 {
83 did_panic = 1;
84
85 return NOTIFY_DONE;
86 }
87
88 static struct notifier_block panic_block = {
89 .notifier_call = hung_task_panic,
90 };
91
check_hung_task(struct task_struct * t,unsigned long timeout)92 static void check_hung_task(struct task_struct *t, unsigned long timeout)
93 {
94 unsigned long switch_count = t->nvcsw + t->nivcsw;
95
96 /*
97 * Ensure the task is not frozen.
98 * Also, skip vfork and any other user process that freezer should skip.
99 */
100 if (unlikely(READ_ONCE(t->__state) & TASK_FROZEN))
101 return;
102
103 /*
104 * When a freshly created task is scheduled once, changes its state to
105 * TASK_UNINTERRUPTIBLE without having ever been switched out once, it
106 * musn't be checked.
107 */
108 if (unlikely(!switch_count))
109 return;
110
111 if (switch_count != t->last_switch_count) {
112 t->last_switch_count = switch_count;
113 t->last_switch_time = jiffies;
114 return;
115 }
116 if (time_is_after_jiffies(t->last_switch_time + timeout * HZ))
117 return;
118
119 trace_sched_process_hang(t);
120
121 if (sysctl_hung_task_panic) {
122 console_verbose();
123 hung_task_show_lock = true;
124 hung_task_call_panic = true;
125 }
126
127 /*
128 * Ok, the task did not get scheduled for more than 2 minutes,
129 * complain:
130 */
131 if (sysctl_hung_task_warnings) {
132 if (sysctl_hung_task_warnings > 0)
133 sysctl_hung_task_warnings--;
134 pr_err("INFO: task %s:%d blocked for more than %ld seconds.\n",
135 t->comm, t->pid, (jiffies - t->last_switch_time) / HZ);
136 pr_err(" %s %s %.*s\n",
137 print_tainted(), init_utsname()->release,
138 (int)strcspn(init_utsname()->version, " "),
139 init_utsname()->version);
140 pr_err("\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
141 " disables this message.\n");
142 sched_show_task(t);
143 hung_task_show_lock = true;
144
145 if (sysctl_hung_task_all_cpu_backtrace)
146 hung_task_show_all_bt = true;
147 }
148
149 touch_nmi_watchdog();
150 }
151
152 /*
153 * To avoid extending the RCU grace period for an unbounded amount of time,
154 * periodically exit the critical section and enter a new one.
155 *
156 * For preemptible RCU it is sufficient to call rcu_read_unlock in order
157 * to exit the grace period. For classic RCU, a reschedule is required.
158 */
rcu_lock_break(struct task_struct * g,struct task_struct * t)159 static bool rcu_lock_break(struct task_struct *g, struct task_struct *t)
160 {
161 bool can_cont;
162
163 get_task_struct(g);
164 get_task_struct(t);
165 rcu_read_unlock();
166 cond_resched();
167 rcu_read_lock();
168 can_cont = pid_alive(g) && pid_alive(t);
169 put_task_struct(t);
170 put_task_struct(g);
171
172 return can_cont;
173 }
174
175 /*
176 * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
177 * a really long time (120 seconds). If that happens, print out
178 * a warning.
179 */
check_hung_uninterruptible_tasks(unsigned long timeout)180 static void check_hung_uninterruptible_tasks(unsigned long timeout)
181 {
182 int max_count = sysctl_hung_task_check_count;
183 unsigned long last_break = jiffies;
184 struct task_struct *g, *t;
185 bool need_check = true;
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 trace_android_vh_check_uninterrupt_tasks(t, timeout, &need_check);
211 if (need_check) {
212 state = READ_ONCE(t->__state);
213 if ((state & TASK_UNINTERRUPTIBLE) &&
214 !(state & TASK_WAKEKILL) &&
215 !(state & TASK_NOLOAD))
216 check_hung_task(t, timeout);
217 }
218 }
219 trace_android_vh_check_uninterrupt_tasks_done(NULL);
220 unlock:
221 rcu_read_unlock();
222 if (hung_task_show_lock)
223 debug_show_all_locks();
224
225 if (hung_task_show_all_bt) {
226 hung_task_show_all_bt = false;
227 trigger_all_cpu_backtrace();
228 }
229
230 if (hung_task_call_panic)
231 panic("hung_task: blocked tasks");
232 }
233
hung_timeout_jiffies(unsigned long last_checked,unsigned long timeout)234 static long hung_timeout_jiffies(unsigned long last_checked,
235 unsigned long timeout)
236 {
237 /* timeout of 0 will disable the watchdog */
238 return timeout ? last_checked - jiffies + timeout * HZ :
239 MAX_SCHEDULE_TIMEOUT;
240 }
241
242 #ifdef CONFIG_SYSCTL
243 /*
244 * Process updating of timeout sysctl
245 */
proc_dohung_task_timeout_secs(struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)246 static int proc_dohung_task_timeout_secs(struct ctl_table *table, int write,
247 void *buffer,
248 size_t *lenp, loff_t *ppos)
249 {
250 int ret;
251
252 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
253
254 if (ret || !write)
255 goto out;
256
257 wake_up_process(watchdog_task);
258
259 out:
260 return ret;
261 }
262
263 /*
264 * This is needed for proc_doulongvec_minmax of sysctl_hung_task_timeout_secs
265 * and hung_task_check_interval_secs
266 */
267 static const unsigned long hung_task_timeout_max = (LONG_MAX / HZ);
268 static struct ctl_table hung_task_sysctls[] = {
269 #ifdef CONFIG_SMP
270 {
271 .procname = "hung_task_all_cpu_backtrace",
272 .data = &sysctl_hung_task_all_cpu_backtrace,
273 .maxlen = sizeof(int),
274 .mode = 0644,
275 .proc_handler = proc_dointvec_minmax,
276 .extra1 = SYSCTL_ZERO,
277 .extra2 = SYSCTL_ONE,
278 },
279 #endif /* CONFIG_SMP */
280 {
281 .procname = "hung_task_panic",
282 .data = &sysctl_hung_task_panic,
283 .maxlen = sizeof(int),
284 .mode = 0644,
285 .proc_handler = proc_dointvec_minmax,
286 .extra1 = SYSCTL_ZERO,
287 .extra2 = SYSCTL_ONE,
288 },
289 {
290 .procname = "hung_task_check_count",
291 .data = &sysctl_hung_task_check_count,
292 .maxlen = sizeof(int),
293 .mode = 0644,
294 .proc_handler = proc_dointvec_minmax,
295 .extra1 = SYSCTL_ZERO,
296 },
297 {
298 .procname = "hung_task_timeout_secs",
299 .data = &sysctl_hung_task_timeout_secs,
300 .maxlen = sizeof(unsigned long),
301 .mode = 0644,
302 .proc_handler = proc_dohung_task_timeout_secs,
303 .extra2 = (void *)&hung_task_timeout_max,
304 },
305 {
306 .procname = "hung_task_check_interval_secs",
307 .data = &sysctl_hung_task_check_interval_secs,
308 .maxlen = sizeof(unsigned long),
309 .mode = 0644,
310 .proc_handler = proc_dohung_task_timeout_secs,
311 .extra2 = (void *)&hung_task_timeout_max,
312 },
313 {
314 .procname = "hung_task_warnings",
315 .data = &sysctl_hung_task_warnings,
316 .maxlen = sizeof(int),
317 .mode = 0644,
318 .proc_handler = proc_dointvec_minmax,
319 .extra1 = SYSCTL_NEG_ONE,
320 },
321 {}
322 };
323
hung_task_sysctl_init(void)324 static void __init hung_task_sysctl_init(void)
325 {
326 register_sysctl_init("kernel", hung_task_sysctls);
327 }
328 #else
329 #define hung_task_sysctl_init() do { } while (0)
330 #endif /* CONFIG_SYSCTL */
331
332
333 static atomic_t reset_hung_task = ATOMIC_INIT(0);
334
reset_hung_task_detector(void)335 void reset_hung_task_detector(void)
336 {
337 atomic_set(&reset_hung_task, 1);
338 }
339 EXPORT_SYMBOL_GPL(reset_hung_task_detector);
340
341 static bool hung_detector_suspended;
342
hungtask_pm_notify(struct notifier_block * self,unsigned long action,void * hcpu)343 static int hungtask_pm_notify(struct notifier_block *self,
344 unsigned long action, void *hcpu)
345 {
346 switch (action) {
347 case PM_SUSPEND_PREPARE:
348 case PM_HIBERNATION_PREPARE:
349 case PM_RESTORE_PREPARE:
350 hung_detector_suspended = true;
351 break;
352 case PM_POST_SUSPEND:
353 case PM_POST_HIBERNATION:
354 case PM_POST_RESTORE:
355 hung_detector_suspended = false;
356 break;
357 default:
358 break;
359 }
360 return NOTIFY_OK;
361 }
362
363 /*
364 * kthread which checks for tasks stuck in D state
365 */
watchdog(void * dummy)366 static int watchdog(void *dummy)
367 {
368 unsigned long hung_last_checked = jiffies;
369
370 set_user_nice(current, 0);
371
372 for ( ; ; ) {
373 unsigned long timeout = sysctl_hung_task_timeout_secs;
374 unsigned long interval = sysctl_hung_task_check_interval_secs;
375 long t;
376
377 if (interval == 0)
378 interval = timeout;
379 interval = min_t(unsigned long, interval, timeout);
380 t = hung_timeout_jiffies(hung_last_checked, interval);
381 if (t <= 0) {
382 if (!atomic_xchg(&reset_hung_task, 0) &&
383 !hung_detector_suspended)
384 check_hung_uninterruptible_tasks(timeout);
385 hung_last_checked = jiffies;
386 continue;
387 }
388 schedule_timeout_interruptible(t);
389 }
390
391 return 0;
392 }
393
hung_task_init(void)394 static int __init hung_task_init(void)
395 {
396 atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
397
398 /* Disable hung task detector on suspend */
399 pm_notifier(hungtask_pm_notify, 0);
400
401 watchdog_task = kthread_run(watchdog, NULL, "khungtaskd");
402 hung_task_sysctl_init();
403
404 return 0;
405 }
406 subsys_initcall(hung_task_init);
407