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 * Total number of tasks detected as hung since boot:
35 */
36 static unsigned long __read_mostly sysctl_hung_task_detect_count;
37
38 /*
39 * Limit number of tasks checked in a batch.
40 *
41 * This value controls the preemptibility of khungtaskd since preemption
42 * is disabled during the critical section. It also controls the size of
43 * the RCU grace period. So it needs to be upper-bound.
44 */
45 #define HUNG_TASK_LOCK_BREAK (HZ / 10)
46
47 /*
48 * Zero means infinite timeout - no checking done:
49 */
50 unsigned long __read_mostly sysctl_hung_task_timeout_secs = CONFIG_DEFAULT_HUNG_TASK_TIMEOUT;
51 EXPORT_SYMBOL_GPL(sysctl_hung_task_timeout_secs);
52
53 /*
54 * Zero (default value) means use sysctl_hung_task_timeout_secs:
55 */
56 static unsigned long __read_mostly sysctl_hung_task_check_interval_secs;
57
58 static int __read_mostly sysctl_hung_task_warnings = 10;
59
60 static int __read_mostly did_panic;
61 static bool hung_task_show_lock;
62 static bool hung_task_call_panic;
63 static bool hung_task_show_all_bt;
64
65 static struct task_struct *watchdog_task;
66
67 #ifdef CONFIG_SMP
68 /*
69 * Should we dump all CPUs backtraces in a hung task event?
70 * Defaults to 0, can be changed via sysctl.
71 */
72 static unsigned int __read_mostly sysctl_hung_task_all_cpu_backtrace;
73 #else
74 #define sysctl_hung_task_all_cpu_backtrace 0
75 #endif /* CONFIG_SMP */
76
77 /*
78 * Should we panic (and reboot, if panic_timeout= is set) when a
79 * hung task is detected:
80 */
81 static unsigned int __read_mostly sysctl_hung_task_panic =
82 IS_ENABLED(CONFIG_BOOTPARAM_HUNG_TASK_PANIC);
83
84 static int
hung_task_panic(struct notifier_block * this,unsigned long event,void * ptr)85 hung_task_panic(struct notifier_block *this, unsigned long event, void *ptr)
86 {
87 did_panic = 1;
88
89 return NOTIFY_DONE;
90 }
91
92 static struct notifier_block panic_block = {
93 .notifier_call = hung_task_panic,
94 };
95
check_hung_task(struct task_struct * t,unsigned long timeout)96 static void check_hung_task(struct task_struct *t, unsigned long timeout)
97 {
98 unsigned long switch_count = t->nvcsw + t->nivcsw;
99
100 /*
101 * Ensure the task is not frozen.
102 * Also, skip vfork and any other user process that freezer should skip.
103 */
104 if (unlikely(READ_ONCE(t->__state) & TASK_FROZEN))
105 return;
106
107 /*
108 * When a freshly created task is scheduled once, changes its state to
109 * TASK_UNINTERRUPTIBLE without having ever been switched out once, it
110 * musn't be checked.
111 */
112 if (unlikely(!switch_count))
113 return;
114
115 if (switch_count != t->last_switch_count) {
116 t->last_switch_count = switch_count;
117 t->last_switch_time = jiffies;
118 return;
119 }
120 if (time_is_after_jiffies(t->last_switch_time + timeout * HZ))
121 return;
122
123 /*
124 * This counter tracks the total number of tasks detected as hung
125 * since boot.
126 */
127 sysctl_hung_task_detect_count++;
128
129 trace_sched_process_hang(t);
130
131 if (sysctl_hung_task_panic) {
132 console_verbose();
133 hung_task_show_lock = true;
134 hung_task_call_panic = true;
135 }
136
137 /*
138 * Ok, the task did not get scheduled for more than 2 minutes,
139 * complain:
140 */
141 if (sysctl_hung_task_warnings || hung_task_call_panic) {
142 if (sysctl_hung_task_warnings > 0)
143 sysctl_hung_task_warnings--;
144 pr_err("INFO: task %s:%d blocked for more than %ld seconds.\n",
145 t->comm, t->pid, (jiffies - t->last_switch_time) / HZ);
146 pr_err(" %s %s %.*s\n",
147 print_tainted(), init_utsname()->release,
148 (int)strcspn(init_utsname()->version, " "),
149 init_utsname()->version);
150 if (t->flags & PF_POSTCOREDUMP)
151 pr_err(" Blocked by coredump.\n");
152 pr_err("\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
153 " disables this message.\n");
154 sched_show_task(t);
155 hung_task_show_lock = true;
156
157 if (sysctl_hung_task_all_cpu_backtrace)
158 hung_task_show_all_bt = true;
159 if (!sysctl_hung_task_warnings)
160 pr_info("Future hung task reports are suppressed, see sysctl kernel.hung_task_warnings\n");
161 }
162
163 touch_nmi_watchdog();
164 }
165
166 /*
167 * To avoid extending the RCU grace period for an unbounded amount of time,
168 * periodically exit the critical section and enter a new one.
169 *
170 * For preemptible RCU it is sufficient to call rcu_read_unlock in order
171 * to exit the grace period. For classic RCU, a reschedule is required.
172 */
rcu_lock_break(struct task_struct * g,struct task_struct * t)173 static bool rcu_lock_break(struct task_struct *g, struct task_struct *t)
174 {
175 bool can_cont;
176
177 get_task_struct(g);
178 get_task_struct(t);
179 rcu_read_unlock();
180 cond_resched();
181 rcu_read_lock();
182 can_cont = pid_alive(g) && pid_alive(t);
183 put_task_struct(t);
184 put_task_struct(g);
185
186 return can_cont;
187 }
188
189 /*
190 * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
191 * a really long time (120 seconds). If that happens, print out
192 * a warning.
193 */
check_hung_uninterruptible_tasks(unsigned long timeout)194 static void check_hung_uninterruptible_tasks(unsigned long timeout)
195 {
196 int max_count = sysctl_hung_task_check_count;
197 unsigned long last_break = jiffies;
198 struct task_struct *g, *t;
199
200 /*
201 * If the system crashed already then all bets are off,
202 * do not report extra hung tasks:
203 */
204 if (test_taint(TAINT_DIE) || did_panic)
205 return;
206
207 hung_task_show_lock = false;
208 rcu_read_lock();
209 for_each_process_thread(g, t) {
210 unsigned int state;
211
212 if (!max_count--)
213 goto unlock;
214 if (time_after(jiffies, last_break + HUNG_TASK_LOCK_BREAK)) {
215 if (!rcu_lock_break(g, t))
216 goto unlock;
217 last_break = jiffies;
218 }
219 /*
220 * skip the TASK_KILLABLE tasks -- these can be killed
221 * skip the TASK_IDLE tasks -- those are genuinely idle
222 */
223 state = READ_ONCE(t->__state);
224 if ((state & TASK_UNINTERRUPTIBLE) &&
225 !(state & TASK_WAKEKILL) &&
226 !(state & TASK_NOLOAD))
227 check_hung_task(t, timeout);
228 }
229 unlock:
230 rcu_read_unlock();
231 if (hung_task_show_lock)
232 debug_show_all_locks();
233
234 if (hung_task_show_all_bt) {
235 hung_task_show_all_bt = false;
236 trigger_all_cpu_backtrace();
237 }
238
239 if (hung_task_call_panic)
240 panic("hung_task: blocked tasks");
241 }
242
hung_timeout_jiffies(unsigned long last_checked,unsigned long timeout)243 static long hung_timeout_jiffies(unsigned long last_checked,
244 unsigned long timeout)
245 {
246 /* timeout of 0 will disable the watchdog */
247 return timeout ? last_checked - jiffies + timeout * HZ :
248 MAX_SCHEDULE_TIMEOUT;
249 }
250
251 #ifdef CONFIG_SYSCTL
252 /*
253 * Process updating of timeout sysctl
254 */
proc_dohung_task_timeout_secs(const struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)255 static int proc_dohung_task_timeout_secs(const struct ctl_table *table, int write,
256 void *buffer,
257 size_t *lenp, loff_t *ppos)
258 {
259 int ret;
260
261 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
262
263 if (ret || !write)
264 goto out;
265
266 wake_up_process(watchdog_task);
267
268 out:
269 return ret;
270 }
271
272 /*
273 * This is needed for proc_doulongvec_minmax of sysctl_hung_task_timeout_secs
274 * and hung_task_check_interval_secs
275 */
276 static const unsigned long hung_task_timeout_max = (LONG_MAX / HZ);
277 static const struct ctl_table hung_task_sysctls[] = {
278 #ifdef CONFIG_SMP
279 {
280 .procname = "hung_task_all_cpu_backtrace",
281 .data = &sysctl_hung_task_all_cpu_backtrace,
282 .maxlen = sizeof(int),
283 .mode = 0644,
284 .proc_handler = proc_dointvec_minmax,
285 .extra1 = SYSCTL_ZERO,
286 .extra2 = SYSCTL_ONE,
287 },
288 #endif /* CONFIG_SMP */
289 {
290 .procname = "hung_task_panic",
291 .data = &sysctl_hung_task_panic,
292 .maxlen = sizeof(int),
293 .mode = 0644,
294 .proc_handler = proc_dointvec_minmax,
295 .extra1 = SYSCTL_ZERO,
296 .extra2 = SYSCTL_ONE,
297 },
298 {
299 .procname = "hung_task_check_count",
300 .data = &sysctl_hung_task_check_count,
301 .maxlen = sizeof(int),
302 .mode = 0644,
303 .proc_handler = proc_dointvec_minmax,
304 .extra1 = SYSCTL_ZERO,
305 },
306 {
307 .procname = "hung_task_timeout_secs",
308 .data = &sysctl_hung_task_timeout_secs,
309 .maxlen = sizeof(unsigned long),
310 .mode = 0644,
311 .proc_handler = proc_dohung_task_timeout_secs,
312 .extra2 = (void *)&hung_task_timeout_max,
313 },
314 {
315 .procname = "hung_task_check_interval_secs",
316 .data = &sysctl_hung_task_check_interval_secs,
317 .maxlen = sizeof(unsigned long),
318 .mode = 0644,
319 .proc_handler = proc_dohung_task_timeout_secs,
320 .extra2 = (void *)&hung_task_timeout_max,
321 },
322 {
323 .procname = "hung_task_warnings",
324 .data = &sysctl_hung_task_warnings,
325 .maxlen = sizeof(int),
326 .mode = 0644,
327 .proc_handler = proc_dointvec_minmax,
328 .extra1 = SYSCTL_NEG_ONE,
329 },
330 {
331 .procname = "hung_task_detect_count",
332 .data = &sysctl_hung_task_detect_count,
333 .maxlen = sizeof(unsigned long),
334 .mode = 0444,
335 .proc_handler = proc_doulongvec_minmax,
336 },
337 };
338
hung_task_sysctl_init(void)339 static void __init hung_task_sysctl_init(void)
340 {
341 register_sysctl_init("kernel", hung_task_sysctls);
342 }
343 #else
344 #define hung_task_sysctl_init() do { } while (0)
345 #endif /* CONFIG_SYSCTL */
346
347
348 static atomic_t reset_hung_task = ATOMIC_INIT(0);
349
reset_hung_task_detector(void)350 void reset_hung_task_detector(void)
351 {
352 atomic_set(&reset_hung_task, 1);
353 }
354 EXPORT_SYMBOL_GPL(reset_hung_task_detector);
355
356 static bool hung_detector_suspended;
357
hungtask_pm_notify(struct notifier_block * self,unsigned long action,void * hcpu)358 static int hungtask_pm_notify(struct notifier_block *self,
359 unsigned long action, void *hcpu)
360 {
361 switch (action) {
362 case PM_SUSPEND_PREPARE:
363 case PM_HIBERNATION_PREPARE:
364 case PM_RESTORE_PREPARE:
365 hung_detector_suspended = true;
366 break;
367 case PM_POST_SUSPEND:
368 case PM_POST_HIBERNATION:
369 case PM_POST_RESTORE:
370 hung_detector_suspended = false;
371 break;
372 default:
373 break;
374 }
375 return NOTIFY_OK;
376 }
377
378 /*
379 * kthread which checks for tasks stuck in D state
380 */
watchdog(void * dummy)381 static int watchdog(void *dummy)
382 {
383 unsigned long hung_last_checked = jiffies;
384
385 set_user_nice(current, 0);
386
387 for ( ; ; ) {
388 unsigned long timeout = sysctl_hung_task_timeout_secs;
389 unsigned long interval = sysctl_hung_task_check_interval_secs;
390 long t;
391
392 if (interval == 0)
393 interval = timeout;
394 interval = min_t(unsigned long, interval, timeout);
395 t = hung_timeout_jiffies(hung_last_checked, interval);
396 if (t <= 0) {
397 if (!atomic_xchg(&reset_hung_task, 0) &&
398 !hung_detector_suspended)
399 check_hung_uninterruptible_tasks(timeout);
400 hung_last_checked = jiffies;
401 continue;
402 }
403 schedule_timeout_interruptible(t);
404 }
405
406 return 0;
407 }
408
hung_task_init(void)409 static int __init hung_task_init(void)
410 {
411 atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
412
413 /* Disable hung task detector on suspend */
414 pm_notifier(hungtask_pm_notify, 0);
415
416 watchdog_task = kthread_run(watchdog, NULL, "khungtaskd");
417 hung_task_sysctl_init();
418
419 return 0;
420 }
421 subsys_initcall(hung_task_init);
422