Merge branch 'reject-ptr-scalar-mix'
[sfrench/cifs-2.6.git] / kernel / hung_task.c
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/suspend.h>
19 #include <linux/utsname.h>
20 #include <linux/sched/signal.h>
21 #include <linux/sched/debug.h>
22
23 #include <trace/events/sched.h>
24
25 /*
26  * The number of tasks checked:
27  */
28 int __read_mostly sysctl_hung_task_check_count = PID_MAX_LIMIT;
29
30 /*
31  * Limit number of tasks checked in a batch.
32  *
33  * This value controls the preemptibility of khungtaskd since preemption
34  * is disabled during the critical section. It also controls the size of
35  * the RCU grace period. So it needs to be upper-bound.
36  */
37 #define HUNG_TASK_BATCHING 1024
38
39 /*
40  * Zero means infinite timeout - no checking done:
41  */
42 unsigned long __read_mostly sysctl_hung_task_timeout_secs = CONFIG_DEFAULT_HUNG_TASK_TIMEOUT;
43
44 /*
45  * Zero (default value) means use sysctl_hung_task_timeout_secs:
46  */
47 unsigned long __read_mostly sysctl_hung_task_check_interval_secs;
48
49 int __read_mostly sysctl_hung_task_warnings = 10;
50
51 static int __read_mostly did_panic;
52 static bool hung_task_show_lock;
53 static bool hung_task_call_panic;
54
55 static struct task_struct *watchdog_task;
56
57 /*
58  * Should we panic (and reboot, if panic_timeout= is set) when a
59  * hung task is detected:
60  */
61 unsigned int __read_mostly sysctl_hung_task_panic =
62                                 CONFIG_BOOTPARAM_HUNG_TASK_PANIC_VALUE;
63
64 static int __init hung_task_panic_setup(char *str)
65 {
66         int rc = kstrtouint(str, 0, &sysctl_hung_task_panic);
67
68         if (rc)
69                 return rc;
70         return 1;
71 }
72 __setup("hung_task_panic=", hung_task_panic_setup);
73
74 static int
75 hung_task_panic(struct notifier_block *this, unsigned long event, void *ptr)
76 {
77         did_panic = 1;
78
79         return NOTIFY_DONE;
80 }
81
82 static struct notifier_block panic_block = {
83         .notifier_call = hung_task_panic,
84 };
85
86 static void check_hung_task(struct task_struct *t, unsigned long timeout)
87 {
88         unsigned long switch_count = t->nvcsw + t->nivcsw;
89
90         /*
91          * Ensure the task is not frozen.
92          * Also, skip vfork and any other user process that freezer should skip.
93          */
94         if (unlikely(t->flags & (PF_FROZEN | PF_FREEZER_SKIP)))
95             return;
96
97         /*
98          * When a freshly created task is scheduled once, changes its state to
99          * TASK_UNINTERRUPTIBLE without having ever been switched out once, it
100          * musn't be checked.
101          */
102         if (unlikely(!switch_count))
103                 return;
104
105         if (switch_count != t->last_switch_count) {
106                 t->last_switch_count = switch_count;
107                 t->last_switch_time = jiffies;
108                 return;
109         }
110         if (time_is_after_jiffies(t->last_switch_time + timeout * HZ))
111                 return;
112
113         trace_sched_process_hang(t);
114
115         if (!sysctl_hung_task_warnings && !sysctl_hung_task_panic)
116                 return;
117
118         /*
119          * Ok, the task did not get scheduled for more than 2 minutes,
120          * complain:
121          */
122         if (sysctl_hung_task_warnings) {
123                 if (sysctl_hung_task_warnings > 0)
124                         sysctl_hung_task_warnings--;
125                 pr_err("INFO: task %s:%d blocked for more than %ld seconds.\n",
126                         t->comm, t->pid, timeout);
127                 pr_err("      %s %s %.*s\n",
128                         print_tainted(), init_utsname()->release,
129                         (int)strcspn(init_utsname()->version, " "),
130                         init_utsname()->version);
131                 pr_err("\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
132                         " disables this message.\n");
133                 sched_show_task(t);
134                 hung_task_show_lock = true;
135         }
136
137         touch_nmi_watchdog();
138
139         if (sysctl_hung_task_panic) {
140                 hung_task_show_lock = true;
141                 hung_task_call_panic = true;
142         }
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  */
152 static 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  */
173 static void check_hung_uninterruptible_tasks(unsigned long timeout)
174 {
175         int max_count = sysctl_hung_task_check_count;
176         int batch_count = HUNG_TASK_BATCHING;
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 (!--batch_count) {
192                         batch_count = HUNG_TASK_BATCHING;
193                         if (!rcu_lock_break(g, t))
194                                 goto unlock;
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
210 static 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  */
221 int 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
238 static atomic_t reset_hung_task = ATOMIC_INIT(0);
239
240 void reset_hung_task_detector(void)
241 {
242         atomic_set(&reset_hung_task, 1);
243 }
244 EXPORT_SYMBOL_GPL(reset_hung_task_detector);
245
246 static bool hung_detector_suspended;
247
248 static 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  */
271 static 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
299 static 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 }
310 subsys_initcall(hung_task_init);