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 int __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         int rc = kstrtouint(str, 0, &sysctl_hung_task_panic);
  56
  57         if (rc)
  58                 return rc;
  59         return 1;
  60 }
  61 __setup("hung_task_panic=", hung_task_panic_setup);
  62
  63 static int
  64 hung_task_panic(struct notifier_block *this, unsigned long event, void *ptr)
  65 {
  66         did_panic = 1;
  67
  68         return NOTIFY_DONE;
  69 }
  70
  71 static struct notifier_block panic_block = {
  72         .notifier_call = hung_task_panic,
  73 };
  74
  75 static void check_hung_task(struct task_struct *t, unsigned long timeout)
  76 {
  77         unsigned long switch_count = t->nvcsw + t->nivcsw;
  78
  79         /*
  80          * Ensure the task is not frozen.
  81          * Also, skip vfork and any other user process that freezer should skip.
  82          */
  83         if (unlikely(t->flags & (PF_FROZEN | PF_FREEZER_SKIP)))
  84             return;
  85
  86         /*
  87          * When a freshly created task is scheduled once, changes its state to
  88          * TASK_UNINTERRUPTIBLE without having ever been switched out once, it
  89          * musn't be checked.
  90          */
  91         if (unlikely(!switch_count))
  92                 return;
  93
  94         if (switch_count != t->last_switch_count) {
  95                 t->last_switch_count = switch_count;
  96                 return;
  97         }
  98
  99         trace_sched_process_hang(t);
 100
 101         if (!sysctl_hung_task_warnings && !sysctl_hung_task_panic)
 102                 return;
 103
 104         if (sysctl_hung_task_warnings > 0)
 105                 sysctl_hung_task_warnings--;
 106
 107         /*
 108          * Ok, the task did not get scheduled for more than 2 minutes,
 109          * complain:
 110          */
 111         if (sysctl_hung_task_warnings) {
 112                 pr_err("INFO: task %s:%d blocked for more than %ld seconds.\n",
 113                         t->comm, t->pid, timeout);
 114                 pr_err("      %s %s %.*s\n",
 115                         print_tainted(), init_utsname()->release,
 116                         (int)strcspn(init_utsname()->version, " "),
 117                         init_utsname()->version);
 118                 pr_err("\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
 119                         " disables this message.\n");
 120                 sched_show_task(t);
 121                 debug_show_all_locks();
 122         }
 123
 124         touch_nmi_watchdog();
 125
 126         if (sysctl_hung_task_panic) {
 127                 trigger_all_cpu_backtrace();
 128                 panic("hung_task: blocked tasks");
 129         }
 130 }
 131
 132 /*
 133  * To avoid extending the RCU grace period for an unbounded amount of time,
 134  * periodically exit the critical section and enter a new one.
 135  *
 136  * For preemptible RCU it is sufficient to call rcu_read_unlock in order
 137  * to exit the grace period. For classic RCU, a reschedule is required.
 138  */
 139 static bool rcu_lock_break(struct task_struct *g, struct task_struct *t)
 140 {
 141         bool can_cont;
 142
 143         get_task_struct(g);
 144         get_task_struct(t);
 145         rcu_read_unlock();
 146         cond_resched();
 147         rcu_read_lock();
 148         can_cont = pid_alive(g) && pid_alive(t);
 149         put_task_struct(t);
 150         put_task_struct(g);
 151
 152         return can_cont;
 153 }
 154
 155 /*
 156  * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
 157  * a really long time (120 seconds). If that happens, print out
 158  * a warning.
 159  */
 160 static void check_hung_uninterruptible_tasks(unsigned long timeout)
 161 {
 162         int max_count = sysctl_hung_task_check_count;
 163         int batch_count = HUNG_TASK_BATCHING;
 164         struct task_struct *g, *t;
 165
 166         /*
 167          * If the system crashed already then all bets are off,
 168          * do not report extra hung tasks:
 169          */
 170         if (test_taint(TAINT_DIE) || did_panic)
 171                 return;
 172
 173         rcu_read_lock();
 174         for_each_process_thread(g, t) {
 175                 if (!max_count--)
 176                         goto unlock;
 177                 if (!--batch_count) {
 178                         batch_count = HUNG_TASK_BATCHING;
 179                         if (!rcu_lock_break(g, t))
 180                                 goto unlock;
 181                 }
 182                 /* use "==" to skip the TASK_KILLABLE tasks waiting on NFS */
 183                 if (t->state == TASK_UNINTERRUPTIBLE)
 184                         check_hung_task(t, timeout);
 185         }
 186  unlock:
 187         rcu_read_unlock();
 188 }
 189
 190 static long hung_timeout_jiffies(unsigned long last_checked,
 191                                  unsigned long timeout)
 192 {
 193         /* timeout of 0 will disable the watchdog */
 194         return timeout ? last_checked - jiffies + timeout * HZ :
 195                 MAX_SCHEDULE_TIMEOUT;
 196 }
 197
 198 /*
 199  * Process updating of timeout sysctl
 200  */
 201 int proc_dohung_task_timeout_secs(struct ctl_table *table, int write,
 202                                   void __user *buffer,
 203                                   size_t *lenp, loff_t *ppos)
 204 {
 205         int ret;
 206
 207         ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
 208
 209         if (ret || !write)
 210                 goto out;
 211
 212         wake_up_process(watchdog_task);
 213
 214  out:
 215         return ret;
 216 }
 217
 218 static atomic_t reset_hung_task = ATOMIC_INIT(0);
 219
 220 void reset_hung_task_detector(void)
 221 {
 222         atomic_set(&reset_hung_task, 1);
 223 }
 224 EXPORT_SYMBOL_GPL(reset_hung_task_detector);
 225
 226 /*
 227  * kthread which checks for tasks stuck in D state
 228  */
 229 static int watchdog(void *dummy)
 230 {
 231         unsigned long hung_last_checked = jiffies;
 232
 233         set_user_nice(current, 0);
 234
 235         for ( ; ; ) {
 236                 unsigned long timeout = sysctl_hung_task_timeout_secs;
 237                 long t = hung_timeout_jiffies(hung_last_checked, timeout);
 238
 239                 if (t <= 0) {
 240                         if (!atomic_xchg(&reset_hung_task, 0))
 241                                 check_hung_uninterruptible_tasks(timeout);
 242                         hung_last_checked = jiffies;
 243                         continue;
 244                 }
 245                 schedule_timeout_interruptible(t);
 246         }
 247
 248         return 0;
 249 }
 250
 251 static int __init hung_task_init(void)
 252 {
 253         atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
 254         watchdog_task = kthread_run(watchdog, NULL, "khungtaskd");
 255
 256         return 0;
 257 }
 258 subsys_initcall(hung_task_init);