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 <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  */
  27 int __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_LOCK_BREAK (HZ / 10)
  37
  38 /*
  39  * Zero means infinite timeout - no checking done:
  40  */
  41 unsigned long __read_mostly sysctl_hung_task_timeout_secs = CONFIG_DEFAULT_HUNG_TASK_TIMEOUT;
  42
  43 /*
  44  * Zero (default value) means use sysctl_hung_task_timeout_secs:
  45  */
  46 unsigned long __read_mostly sysctl_hung_task_check_interval_secs;
  47
  48 int __read_mostly sysctl_hung_task_warnings = 10;
  49
  50 static int __read_mostly did_panic;
  51 static bool hung_task_show_lock;
  52 static bool hung_task_call_panic;
  53
  54 static struct task_struct *watchdog_task;
  55
  56 /*
  57  * Should we panic (and reboot, if panic_timeout= is set) when a
  58  * hung task is detected:
  59  */
  60 unsigned int __read_mostly sysctl_hung_task_panic =
  61                                 CONFIG_BOOTPARAM_HUNG_TASK_PANIC_VALUE;
  62
  63 static int __init hung_task_panic_setup(char *str)
  64 {
  65         int rc = kstrtouint(str, 0, &sysctl_hung_task_panic);
  66
  67         if (rc)
  68                 return rc;
  69         return 1;
  70 }
  71 __setup("hung_task_panic=", hung_task_panic_setup);
  72
  73 static int
  74 hung_task_panic(struct notifier_block *this, unsigned long event, void *ptr)
  75 {
  76         did_panic = 1;
  77
  78         return NOTIFY_DONE;
  79 }
  80
  81 static struct notifier_block panic_block = {
  82         .notifier_call = hung_task_panic,
  83 };
  84
  85 static void check_hung_task(struct task_struct *t, unsigned long timeout)
  86 {
  87         unsigned long switch_count = t->nvcsw + t->nivcsw;
  88
  89         /*
  90          * Ensure the task is not frozen.
  91          * Also, skip vfork and any other user process that freezer should skip.
  92          */
  93         if (unlikely(t->flags & (PF_FROZEN | PF_FREEZER_SKIP)))
  94             return;
  95
  96         /*
  97          * When a freshly created task is scheduled once, changes its state to
  98          * TASK_UNINTERRUPTIBLE without having ever been switched out once, it
  99          * musn't be checked.
 100          */
 101         if (unlikely(!switch_count))
 102                 return;
 103
 104         if (switch_count != t->last_switch_count) {
 105                 t->last_switch_count = switch_count;
 106                 t->last_switch_time = jiffies;
 107                 return;
 108         }
 109         if (time_is_after_jiffies(t->last_switch_time + timeout * HZ))
 110                 return;
 111
 112         trace_sched_process_hang(t);
 113
 114         if (sysctl_hung_task_panic) {
 115                 console_verbose();
 116                 hung_task_show_lock = true;
 117                 hung_task_call_panic = true;
 118         }
 119
 120         /*
 121          * Ok, the task did not get scheduled for more than 2 minutes,
 122          * complain:
 123          */
 124         if (sysctl_hung_task_warnings) {
 125                 if (sysctl_hung_task_warnings > 0)
 126                         sysctl_hung_task_warnings--;
 127                 pr_err("INFO: task %s:%d blocked for more than %ld seconds.\n",
 128                         t->comm, t->pid, timeout);
 129                 pr_err("      %s %s %.*s\n",
 130                         print_tainted(), init_utsname()->release,
 131                         (int)strcspn(init_utsname()->version, " "),
 132                         init_utsname()->version);
 133                 pr_err("\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
 134                         " disables this message.\n");
 135                 sched_show_task(t);
 136                 hung_task_show_lock = true;
 137         }
 138
 139         touch_nmi_watchdog();
 140 }
 141
 142 /*
 143  * To avoid extending the RCU grace period for an unbounded amount of time,
 144  * periodically exit the critical section and enter a new one.
 145  *
 146  * For preemptible RCU it is sufficient to call rcu_read_unlock in order
 147  * to exit the grace period. For classic RCU, a reschedule is required.
 148  */
 149 static bool rcu_lock_break(struct task_struct *g, struct task_struct *t)
 150 {
 151         bool can_cont;
 152
 153         get_task_struct(g);
 154         get_task_struct(t);
 155         rcu_read_unlock();
 156         cond_resched();
 157         rcu_read_lock();
 158         can_cont = pid_alive(g) && pid_alive(t);
 159         put_task_struct(t);
 160         put_task_struct(g);
 161
 162         return can_cont;
 163 }
 164
 165 /*
 166  * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
 167  * a really long time (120 seconds). If that happens, print out
 168  * a warning.
 169  */
 170 static void check_hung_uninterruptible_tasks(unsigned long timeout)
 171 {
 172         int max_count = sysctl_hung_task_check_count;
 173         unsigned long last_break = jiffies;
 174         struct task_struct *g, *t;
 175
 176         /*
 177          * If the system crashed already then all bets are off,
 178          * do not report extra hung tasks:
 179          */
 180         if (test_taint(TAINT_DIE) || did_panic)
 181                 return;
 182
 183         hung_task_show_lock = false;
 184         rcu_read_lock();
 185         for_each_process_thread(g, t) {
 186                 if (!max_count--)
 187                         goto unlock;
 188                 if (time_after(jiffies, last_break + HUNG_TASK_LOCK_BREAK)) {
 189                         if (!rcu_lock_break(g, t))
 190                                 goto unlock;
 191                         last_break = jiffies;
 192                 }
 193                 /* use "==" to skip the TASK_KILLABLE tasks waiting on NFS */
 194                 if (t->state == TASK_UNINTERRUPTIBLE)
 195                         check_hung_task(t, timeout);
 196         }
 197  unlock:
 198         rcu_read_unlock();
 199         if (hung_task_show_lock)
 200                 debug_show_all_locks();
 201         if (hung_task_call_panic) {
 202                 trigger_all_cpu_backtrace();
 203                 panic("hung_task: blocked tasks");
 204         }
 205 }
 206
 207 static long hung_timeout_jiffies(unsigned long last_checked,
 208                                  unsigned long timeout)
 209 {
 210         /* timeout of 0 will disable the watchdog */
 211         return timeout ? last_checked - jiffies + timeout * HZ :
 212                 MAX_SCHEDULE_TIMEOUT;
 213 }
 214
 215 /*
 216  * Process updating of timeout sysctl
 217  */
 218 int proc_dohung_task_timeout_secs(struct ctl_table *table, int write,
 219                                   void __user *buffer,
 220                                   size_t *lenp, loff_t *ppos)
 221 {
 222         int ret;
 223
 224         ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
 225
 226         if (ret || !write)
 227                 goto out;
 228
 229         wake_up_process(watchdog_task);
 230
 231  out:
 232         return ret;
 233 }
 234
 235 static atomic_t reset_hung_task = ATOMIC_INIT(0);
 236
 237 void reset_hung_task_detector(void)
 238 {
 239         atomic_set(&reset_hung_task, 1);
 240 }
 241 EXPORT_SYMBOL_GPL(reset_hung_task_detector);
 242
 243 /*
 244  * kthread which checks for tasks stuck in D state
 245  */
 246 static int watchdog(void *dummy)
 247 {
 248         unsigned long hung_last_checked = jiffies;
 249
 250         set_user_nice(current, 0);
 251
 252         for ( ; ; ) {
 253                 unsigned long timeout = sysctl_hung_task_timeout_secs;
 254                 unsigned long interval = sysctl_hung_task_check_interval_secs;
 255                 long t;
 256
 257                 if (interval == 0)
 258                         interval = timeout;
 259                 interval = min_t(unsigned long, interval, timeout);
 260                 t = hung_timeout_jiffies(hung_last_checked, interval);
 261                 if (t <= 0) {
 262                         if (!atomic_xchg(&reset_hung_task, 0))
 263                                 check_hung_uninterruptible_tasks(timeout);
 264                         hung_last_checked = jiffies;
 265                         continue;
 266                 }
 267                 schedule_timeout_interruptible(t);
 268         }
 269
 270         return 0;
 271 }
 272
 273 static int __init hung_task_init(void)
 274 {
 275         atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
 276         watchdog_task = kthread_run(watchdog, NULL, "khungtaskd");
 277
 278         return 0;
 279 }
 280 subsys_initcall(hung_task_init);