Linux Audio

Check our new training course

Loading...
v4.17
 
  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/utsname.h>
 19#include <linux/sched/signal.h>
 20#include <linux/sched/debug.h>
 
 21
 22#include <trace/events/sched.h>
 23
 24/*
 25 * The number of tasks checked:
 26 */
 27int __read_mostly sysctl_hung_task_check_count = PID_MAX_LIMIT;
 28
 29/*
 30 * Limit number of tasks checked in a batch.
 31 *
 32 * This value controls the preemptibility of khungtaskd since preemption
 33 * is disabled during the critical section. It also controls the size of
 34 * the RCU grace period. So it needs to be upper-bound.
 35 */
 36#define HUNG_TASK_BATCHING 1024
 37
 38/*
 39 * Zero means infinite timeout - no checking done:
 40 */
 41unsigned long __read_mostly sysctl_hung_task_timeout_secs = CONFIG_DEFAULT_HUNG_TASK_TIMEOUT;
 42
 
 
 
 
 
 43int __read_mostly sysctl_hung_task_warnings = 10;
 44
 45static int __read_mostly did_panic;
 46static bool hung_task_show_lock;
 
 
 47
 48static struct task_struct *watchdog_task;
 49
 
 
 
 
 
 
 
 
 50/*
 51 * Should we panic (and reboot, if panic_timeout= is set) when a
 52 * hung task is detected:
 53 */
 54unsigned int __read_mostly sysctl_hung_task_panic =
 55				CONFIG_BOOTPARAM_HUNG_TASK_PANIC_VALUE;
 56
 57static int __init hung_task_panic_setup(char *str)
 58{
 59	int rc = kstrtouint(str, 0, &sysctl_hung_task_panic);
 60
 61	if (rc)
 62		return rc;
 63	return 1;
 64}
 65__setup("hung_task_panic=", hung_task_panic_setup);
 66
 67static int
 68hung_task_panic(struct notifier_block *this, unsigned long event, void *ptr)
 69{
 70	did_panic = 1;
 71
 72	return NOTIFY_DONE;
 73}
 74
 75static struct notifier_block panic_block = {
 76	.notifier_call = hung_task_panic,
 77};
 78
 79static void check_hung_task(struct task_struct *t, unsigned long timeout)
 80{
 81	unsigned long switch_count = t->nvcsw + t->nivcsw;
 82
 83	/*
 84	 * Ensure the task is not frozen.
 85	 * Also, skip vfork and any other user process that freezer should skip.
 86	 */
 87	if (unlikely(t->flags & (PF_FROZEN | PF_FREEZER_SKIP)))
 88	    return;
 89
 90	/*
 91	 * When a freshly created task is scheduled once, changes its state to
 92	 * TASK_UNINTERRUPTIBLE without having ever been switched out once, it
 93	 * musn't be checked.
 94	 */
 95	if (unlikely(!switch_count))
 96		return;
 97
 98	if (switch_count != t->last_switch_count) {
 99		t->last_switch_count = switch_count;
 
100		return;
101	}
 
 
102
103	trace_sched_process_hang(t);
104
105	if (!sysctl_hung_task_warnings && !sysctl_hung_task_panic)
106		return;
 
 
 
107
108	/*
109	 * Ok, the task did not get scheduled for more than 2 minutes,
110	 * complain:
111	 */
112	if (sysctl_hung_task_warnings) {
113		if (sysctl_hung_task_warnings > 0)
114			sysctl_hung_task_warnings--;
115		pr_err("INFO: task %s:%d blocked for more than %ld seconds.\n",
116			t->comm, t->pid, timeout);
117		pr_err("      %s %s %.*s\n",
118			print_tainted(), init_utsname()->release,
119			(int)strcspn(init_utsname()->version, " "),
120			init_utsname()->version);
121		pr_err("\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
122			" disables this message.\n");
123		sched_show_task(t);
124		hung_task_show_lock = true;
 
 
 
125	}
126
127	touch_nmi_watchdog();
128
129	if (sysctl_hung_task_panic) {
130		if (hung_task_show_lock)
131			debug_show_all_locks();
132		trigger_all_cpu_backtrace();
133		panic("hung_task: blocked tasks");
134	}
135}
136
137/*
138 * To avoid extending the RCU grace period for an unbounded amount of time,
139 * periodically exit the critical section and enter a new one.
140 *
141 * For preemptible RCU it is sufficient to call rcu_read_unlock in order
142 * to exit the grace period. For classic RCU, a reschedule is required.
143 */
144static bool rcu_lock_break(struct task_struct *g, struct task_struct *t)
145{
146	bool can_cont;
147
148	get_task_struct(g);
149	get_task_struct(t);
150	rcu_read_unlock();
151	cond_resched();
152	rcu_read_lock();
153	can_cont = pid_alive(g) && pid_alive(t);
154	put_task_struct(t);
155	put_task_struct(g);
156
157	return can_cont;
158}
159
160/*
161 * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
162 * a really long time (120 seconds). If that happens, print out
163 * a warning.
164 */
165static void check_hung_uninterruptible_tasks(unsigned long timeout)
166{
167	int max_count = sysctl_hung_task_check_count;
168	int batch_count = HUNG_TASK_BATCHING;
169	struct task_struct *g, *t;
170
171	/*
172	 * If the system crashed already then all bets are off,
173	 * do not report extra hung tasks:
174	 */
175	if (test_taint(TAINT_DIE) || did_panic)
176		return;
177
178	hung_task_show_lock = false;
179	rcu_read_lock();
180	for_each_process_thread(g, t) {
181		if (!max_count--)
182			goto unlock;
183		if (!--batch_count) {
184			batch_count = HUNG_TASK_BATCHING;
185			if (!rcu_lock_break(g, t))
186				goto unlock;
 
187		}
188		/* use "==" to skip the TASK_KILLABLE tasks waiting on NFS */
189		if (t->state == TASK_UNINTERRUPTIBLE)
190			check_hung_task(t, timeout);
191	}
192 unlock:
193	rcu_read_unlock();
194	if (hung_task_show_lock)
195		debug_show_all_locks();
 
 
 
 
 
 
 
 
196}
197
198static long hung_timeout_jiffies(unsigned long last_checked,
199				 unsigned long timeout)
200{
201	/* timeout of 0 will disable the watchdog */
202	return timeout ? last_checked - jiffies + timeout * HZ :
203		MAX_SCHEDULE_TIMEOUT;
204}
205
206/*
207 * Process updating of timeout sysctl
208 */
209int proc_dohung_task_timeout_secs(struct ctl_table *table, int write,
210				  void __user *buffer,
211				  size_t *lenp, loff_t *ppos)
212{
213	int ret;
214
215	ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
216
217	if (ret || !write)
218		goto out;
219
220	wake_up_process(watchdog_task);
221
222 out:
223	return ret;
224}
225
226static atomic_t reset_hung_task = ATOMIC_INIT(0);
227
228void reset_hung_task_detector(void)
229{
230	atomic_set(&reset_hung_task, 1);
231}
232EXPORT_SYMBOL_GPL(reset_hung_task_detector);
233
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
234/*
235 * kthread which checks for tasks stuck in D state
236 */
237static int watchdog(void *dummy)
238{
239	unsigned long hung_last_checked = jiffies;
240
241	set_user_nice(current, 0);
242
243	for ( ; ; ) {
244		unsigned long timeout = sysctl_hung_task_timeout_secs;
245		long t = hung_timeout_jiffies(hung_last_checked, timeout);
 
246
 
 
 
 
247		if (t <= 0) {
248			if (!atomic_xchg(&reset_hung_task, 0))
 
249				check_hung_uninterruptible_tasks(timeout);
250			hung_last_checked = jiffies;
251			continue;
252		}
253		schedule_timeout_interruptible(t);
254	}
255
256	return 0;
257}
258
259static int __init hung_task_init(void)
260{
261	atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
 
 
 
 
262	watchdog_task = kthread_run(watchdog, NULL, "khungtaskd");
263
264	return 0;
265}
266subsys_initcall(hung_task_init);
v5.14.15
  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 */
 31int __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 */
 45unsigned long __read_mostly sysctl_hung_task_timeout_secs = CONFIG_DEFAULT_HUNG_TASK_TIMEOUT;
 46
 47/*
 48 * Zero (default value) means use sysctl_hung_task_timeout_secs:
 49 */
 50unsigned long __read_mostly sysctl_hung_task_check_interval_secs;
 51
 52int __read_mostly sysctl_hung_task_warnings = 10;
 53
 54static int __read_mostly did_panic;
 55static bool hung_task_show_lock;
 56static bool hung_task_call_panic;
 57static bool hung_task_show_all_bt;
 58
 59static struct task_struct *watchdog_task;
 60
 61#ifdef CONFIG_SMP
 62/*
 63 * Should we dump all CPUs backtraces in a hung task event?
 64 * Defaults to 0, can be changed via sysctl.
 65 */
 66unsigned int __read_mostly sysctl_hung_task_all_cpu_backtrace;
 67#endif /* CONFIG_SMP */
 68
 69/*
 70 * Should we panic (and reboot, if panic_timeout= is set) when a
 71 * hung task is detected:
 72 */
 73unsigned int __read_mostly sysctl_hung_task_panic =
 74				CONFIG_BOOTPARAM_HUNG_TASK_PANIC_VALUE;
 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		if (sysctl_hung_task_all_cpu_backtrace)
142			hung_task_show_all_bt = true;
143	}
144
145	touch_nmi_watchdog();
 
 
 
 
 
 
 
146}
147
148/*
149 * To avoid extending the RCU grace period for an unbounded amount of time,
150 * periodically exit the critical section and enter a new one.
151 *
152 * For preemptible RCU it is sufficient to call rcu_read_unlock in order
153 * to exit the grace period. For classic RCU, a reschedule is required.
154 */
155static bool rcu_lock_break(struct task_struct *g, struct task_struct *t)
156{
157	bool can_cont;
158
159	get_task_struct(g);
160	get_task_struct(t);
161	rcu_read_unlock();
162	cond_resched();
163	rcu_read_lock();
164	can_cont = pid_alive(g) && pid_alive(t);
165	put_task_struct(t);
166	put_task_struct(g);
167
168	return can_cont;
169}
170
171/*
172 * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
173 * a really long time (120 seconds). If that happens, print out
174 * a warning.
175 */
176static void check_hung_uninterruptible_tasks(unsigned long timeout)
177{
178	int max_count = sysctl_hung_task_check_count;
179	unsigned long last_break = jiffies;
180	struct task_struct *g, *t;
181
182	/*
183	 * If the system crashed already then all bets are off,
184	 * do not report extra hung tasks:
185	 */
186	if (test_taint(TAINT_DIE) || did_panic)
187		return;
188
189	hung_task_show_lock = false;
190	rcu_read_lock();
191	for_each_process_thread(g, t) {
192		if (!max_count--)
193			goto unlock;
194		if (time_after(jiffies, last_break + HUNG_TASK_LOCK_BREAK)) {
 
195			if (!rcu_lock_break(g, t))
196				goto unlock;
197			last_break = jiffies;
198		}
199		/* use "==" to skip the TASK_KILLABLE tasks waiting on NFS */
200		if (READ_ONCE(t->__state) == TASK_UNINTERRUPTIBLE)
201			check_hung_task(t, timeout);
202	}
203 unlock:
204	rcu_read_unlock();
205	if (hung_task_show_lock)
206		debug_show_all_locks();
207
208	if (hung_task_show_all_bt) {
209		hung_task_show_all_bt = false;
210		trigger_all_cpu_backtrace();
211	}
212
213	if (hung_task_call_panic)
214		panic("hung_task: blocked tasks");
215}
216
217static long hung_timeout_jiffies(unsigned long last_checked,
218				 unsigned long timeout)
219{
220	/* timeout of 0 will disable the watchdog */
221	return timeout ? last_checked - jiffies + timeout * HZ :
222		MAX_SCHEDULE_TIMEOUT;
223}
224
225/*
226 * Process updating of timeout sysctl
227 */
228int proc_dohung_task_timeout_secs(struct ctl_table *table, int write,
229				  void *buffer, 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);