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