Merge branch 'akpm' (patches from Andrew)
[linux-2.6-microblaze.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_LOCK_BREAK (HZ / 10)
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_panic) {
116                 console_verbose();
117                 hung_task_show_lock = true;
118                 hung_task_call_panic = true;
119         }
120
121         /*
122          * Ok, the task did not get scheduled for more than 2 minutes,
123          * complain:
124          */
125         if (sysctl_hung_task_warnings) {
126                 if (sysctl_hung_task_warnings > 0)
127                         sysctl_hung_task_warnings--;
128                 pr_err("INFO: task %s:%d blocked for more than %ld seconds.\n",
129                         t->comm, t->pid, timeout);
130                 pr_err("      %s %s %.*s\n",
131                         print_tainted(), init_utsname()->release,
132                         (int)strcspn(init_utsname()->version, " "),
133                         init_utsname()->version);
134                 pr_err("\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
135                         " disables this message.\n");
136                 sched_show_task(t);
137                 hung_task_show_lock = true;
138         }
139
140         touch_nmi_watchdog();
141 }
142
143 /*
144  * To avoid extending the RCU grace period for an unbounded amount of time,
145  * periodically exit the critical section and enter a new one.
146  *
147  * For preemptible RCU it is sufficient to call rcu_read_unlock in order
148  * to exit the grace period. For classic RCU, a reschedule is required.
149  */
150 static bool rcu_lock_break(struct task_struct *g, struct task_struct *t)
151 {
152         bool can_cont;
153
154         get_task_struct(g);
155         get_task_struct(t);
156         rcu_read_unlock();
157         cond_resched();
158         rcu_read_lock();
159         can_cont = pid_alive(g) && pid_alive(t);
160         put_task_struct(t);
161         put_task_struct(g);
162
163         return can_cont;
164 }
165
166 /*
167  * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
168  * a really long time (120 seconds). If that happens, print out
169  * a warning.
170  */
171 static void check_hung_uninterruptible_tasks(unsigned long timeout)
172 {
173         int max_count = sysctl_hung_task_check_count;
174         unsigned long last_break = jiffies;
175         struct task_struct *g, *t;
176
177         /*
178          * If the system crashed already then all bets are off,
179          * do not report extra hung tasks:
180          */
181         if (test_taint(TAINT_DIE) || did_panic)
182                 return;
183
184         hung_task_show_lock = false;
185         rcu_read_lock();
186         for_each_process_thread(g, t) {
187                 if (!max_count--)
188                         goto unlock;
189                 if (time_after(jiffies, last_break + HUNG_TASK_LOCK_BREAK)) {
190                         if (!rcu_lock_break(g, t))
191                                 goto unlock;
192                         last_break = jiffies;
193                 }
194                 /* use "==" to skip the TASK_KILLABLE tasks waiting on NFS */
195                 if (t->state == TASK_UNINTERRUPTIBLE)
196                         check_hung_task(t, timeout);
197         }
198  unlock:
199         rcu_read_unlock();
200         if (hung_task_show_lock)
201                 debug_show_all_locks();
202         if (hung_task_call_panic) {
203                 trigger_all_cpu_backtrace();
204                 panic("hung_task: blocked tasks");
205         }
206 }
207
208 static long hung_timeout_jiffies(unsigned long last_checked,
209                                  unsigned long timeout)
210 {
211         /* timeout of 0 will disable the watchdog */
212         return timeout ? last_checked - jiffies + timeout * HZ :
213                 MAX_SCHEDULE_TIMEOUT;
214 }
215
216 /*
217  * Process updating of timeout sysctl
218  */
219 int proc_dohung_task_timeout_secs(struct ctl_table *table, int write,
220                                   void __user *buffer,
221                                   size_t *lenp, loff_t *ppos)
222 {
223         int ret;
224
225         ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
226
227         if (ret || !write)
228                 goto out;
229
230         wake_up_process(watchdog_task);
231
232  out:
233         return ret;
234 }
235
236 static atomic_t reset_hung_task = ATOMIC_INIT(0);
237
238 void reset_hung_task_detector(void)
239 {
240         atomic_set(&reset_hung_task, 1);
241 }
242 EXPORT_SYMBOL_GPL(reset_hung_task_detector);
243
244 static bool hung_detector_suspended;
245
246 static int hungtask_pm_notify(struct notifier_block *self,
247                               unsigned long action, void *hcpu)
248 {
249         switch (action) {
250         case PM_SUSPEND_PREPARE:
251         case PM_HIBERNATION_PREPARE:
252         case PM_RESTORE_PREPARE:
253                 hung_detector_suspended = true;
254                 break;
255         case PM_POST_SUSPEND:
256         case PM_POST_HIBERNATION:
257         case PM_POST_RESTORE:
258                 hung_detector_suspended = false;
259                 break;
260         default:
261                 break;
262         }
263         return NOTIFY_OK;
264 }
265
266 /*
267  * kthread which checks for tasks stuck in D state
268  */
269 static int watchdog(void *dummy)
270 {
271         unsigned long hung_last_checked = jiffies;
272
273         set_user_nice(current, 0);
274
275         for ( ; ; ) {
276                 unsigned long timeout = sysctl_hung_task_timeout_secs;
277                 unsigned long interval = sysctl_hung_task_check_interval_secs;
278                 long t;
279
280                 if (interval == 0)
281                         interval = timeout;
282                 interval = min_t(unsigned long, interval, timeout);
283                 t = hung_timeout_jiffies(hung_last_checked, interval);
284                 if (t <= 0) {
285                         if (!atomic_xchg(&reset_hung_task, 0) &&
286                             !hung_detector_suspended)
287                                 check_hung_uninterruptible_tasks(timeout);
288                         hung_last_checked = jiffies;
289                         continue;
290                 }
291                 schedule_timeout_interruptible(t);
292         }
293
294         return 0;
295 }
296
297 static int __init hung_task_init(void)
298 {
299         atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
300
301         /* Disable hung task detector on suspend */
302         pm_notifier(hungtask_pm_notify, 0);
303
304         watchdog_task = kthread_run(watchdog, NULL, "khungtaskd");
305
306         return 0;
307 }
308 subsys_initcall(hung_task_init);