kernel/hung_task.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Detect Hung Task
   4  *
   5  * kernel/hung_task.c - kernel thread for detecting tasks stuck in D state
   6  *
   7  */
   8
   9 #include <linux/mm.h>
  10 #include <linux/cpu.h>
  11 #include <linux/nmi.h>
  12 #include <linux/init.h>
  13 #include <linux/delay.h>
  14 #include <linux/freezer.h>
  15 #include <linux/kthread.h>
  16 #include <linux/lockdep.h>
  17 #include <linux/export.h>
  18 #include <linux/panic_notifier.h>
  19 #include <linux/sysctl.h>
  20 #include <linux/suspend.h>
  21 #include <linux/utsname.h>
  22 #include <linux/sched/signal.h>
  23 #include <linux/sched/debug.h>
  24 #include <linux/sched/sysctl.h>
  25
  26 #include <trace/events/sched.h>
  27
  28 /*
  29  * The number of tasks checked:
  30  */
  31 int __read_mostly sysctl_hung_task_check_count = PID_MAX_LIMIT;
  32
  33 /*
  34  * Limit number of tasks checked in a batch.
  35  *
  36  * This value controls the preemptibility of khungtaskd since preemption
  37  * is disabled during the critical section. It also controls the size of
  38  * the RCU grace period. So it needs to be upper-bound.
  39  */
  40 #define HUNG_TASK_LOCK_BREAK (HZ / 10)
  41
  42 /*
  43  * Zero means infinite timeout - no checking done:
  44  */
  45 unsigned long __read_mostly sysctl_hung_task_timeout_secs = CONFIG_DEFAULT_HUNG_TASK_TIMEOUT;
  46
  47 /*
  48  * Zero (default value) means use sysctl_hung_task_timeout_secs:
  49  */
  50 unsigned long __read_mostly sysctl_hung_task_check_interval_secs;
  51
  52 int __read_mostly sysctl_hung_task_warnings = 10;
  53
  54 static int __read_mostly did_panic;
  55 static bool hung_task_show_lock;
  56 static bool hung_task_call_panic;
  57 static bool hung_task_show_all_bt;
  58
  59 static struct task_struct *watchdog_task;
  60
  61 #ifdef CONFIG_SMP
  62 /*
  63  * Should we dump all CPUs backtraces in a hung task event?
  64  * Defaults to 0, can be changed via sysctl.
  65  */
  66 static unsigned int __read_mostly sysctl_hung_task_all_cpu_backtrace;
  67 #else
  68 #define sysctl_hung_task_all_cpu_backtrace 0
  69 #endif /* CONFIG_SMP */
  70
  71 /*
  72  * Should we panic (and reboot, if panic_timeout= is set) when a
  73  * hung task is detected:
  74  */
  75 unsigned int __read_mostly sysctl_hung_task_panic =
  76                                 IS_ENABLED(CONFIG_BOOTPARAM_HUNG_TASK_PANIC);
  77
  78 static int
  79 hung_task_panic(struct notifier_block *this, unsigned long event, void *ptr)
  80 {
  81         did_panic = 1;
  82
  83         return NOTIFY_DONE;
  84 }
  85
  86 static struct notifier_block panic_block = {
  87         .notifier_call = hung_task_panic,
  88 };
  89
  90 static void check_hung_task(struct task_struct *t, unsigned long timeout)
  91 {
  92         unsigned long switch_count = t->nvcsw + t->nivcsw;
  93
  94         /*
  95          * Ensure the task is not frozen.
  96          * Also, skip vfork and any other user process that freezer should skip.
  97          */
  98         if (unlikely(t->flags & (PF_FROZEN | PF_FREEZER_SKIP)))
  99             return;
 100
 101         /*
 102          * When a freshly created task is scheduled once, changes its state to
 103          * TASK_UNINTERRUPTIBLE without having ever been switched out once, it
 104          * musn't be checked.
 105          */
 106         if (unlikely(!switch_count))
 107                 return;
 108
 109         if (switch_count != t->last_switch_count) {
 110                 t->last_switch_count = switch_count;
 111                 t->last_switch_time = jiffies;
 112                 return;
 113         }
 114         if (time_is_after_jiffies(t->last_switch_time + timeout * HZ))
 115                 return;
 116
 117         trace_sched_process_hang(t);
 118
 119         if (sysctl_hung_task_panic) {
 120                 console_verbose();
 121                 hung_task_show_lock = true;
 122                 hung_task_call_panic = true;
 123         }
 124
 125         /*
 126          * Ok, the task did not get scheduled for more than 2 minutes,
 127          * complain:
 128          */
 129         if (sysctl_hung_task_warnings) {
 130                 if (sysctl_hung_task_warnings > 0)
 131                         sysctl_hung_task_warnings--;
 132                 pr_err("INFO: task %s:%d blocked for more than %ld seconds.\n",
 133                        t->comm, t->pid, (jiffies - t->last_switch_time) / HZ);
 134                 pr_err("      %s %s %.*s\n",
 135                         print_tainted(), init_utsname()->release,
 136                         (int)strcspn(init_utsname()->version, " "),
 137                         init_utsname()->version);
 138                 pr_err("\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
 139                         " disables this message.\n");
 140                 sched_show_task(t);
 141                 hung_task_show_lock = true;
 142
 143                 if (sysctl_hung_task_all_cpu_backtrace)
 144                         hung_task_show_all_bt = true;
 145         }
 146
 147         touch_nmi_watchdog();
 148 }
 149
 150 /*
 151  * To avoid extending the RCU grace period for an unbounded amount of time,
 152  * periodically exit the critical section and enter a new one.
 153  *
 154  * For preemptible RCU it is sufficient to call rcu_read_unlock in order
 155  * to exit the grace period. For classic RCU, a reschedule is required.
 156  */
 157 static bool rcu_lock_break(struct task_struct *g, struct task_struct *t)
 158 {
 159         bool can_cont;
 160
 161         get_task_struct(g);
 162         get_task_struct(t);
 163         rcu_read_unlock();
 164         cond_resched();
 165         rcu_read_lock();
 166         can_cont = pid_alive(g) && pid_alive(t);
 167         put_task_struct(t);
 168         put_task_struct(g);
 169
 170         return can_cont;
 171 }
 172
 173 /*
 174  * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
 175  * a really long time (120 seconds). If that happens, print out
 176  * a warning.
 177  */
 178 static void check_hung_uninterruptible_tasks(unsigned long timeout)
 179 {
 180         int max_count = sysctl_hung_task_check_count;
 181         unsigned long last_break = jiffies;
 182         struct task_struct *g, *t;
 183
 184         /*
 185          * If the system crashed already then all bets are off,
 186          * do not report extra hung tasks:
 187          */
 188         if (test_taint(TAINT_DIE) || did_panic)
 189                 return;
 190
 191         hung_task_show_lock = false;
 192         rcu_read_lock();
 193         for_each_process_thread(g, t) {
 194                 if (!max_count--)
 195                         goto unlock;
 196                 if (time_after(jiffies, last_break + HUNG_TASK_LOCK_BREAK)) {
 197                         if (!rcu_lock_break(g, t))
 198                                 goto unlock;
 199                         last_break = jiffies;
 200                 }
 201                 /* use "==" to skip the TASK_KILLABLE tasks waiting on NFS */
 202                 if (READ_ONCE(t->__state) == TASK_UNINTERRUPTIBLE)
 203                         check_hung_task(t, timeout);
 204         }
 205  unlock:
 206         rcu_read_unlock();
 207         if (hung_task_show_lock)
 208                 debug_show_all_locks();
 209
 210         if (hung_task_show_all_bt) {
 211                 hung_task_show_all_bt = false;
 212                 trigger_all_cpu_backtrace();
 213         }
 214
 215         if (hung_task_call_panic)
 216                 panic("hung_task: blocked tasks");
 217 }
 218
 219 static long hung_timeout_jiffies(unsigned long last_checked,
 220                                  unsigned long timeout)
 221 {
 222         /* timeout of 0 will disable the watchdog */
 223         return timeout ? last_checked - jiffies + timeout * HZ :
 224                 MAX_SCHEDULE_TIMEOUT;
 225 }
 226
 227 #ifdef CONFIG_SYSCTL
 228 /*
 229  * Process updating of timeout sysctl
 230  */
 231 static int proc_dohung_task_timeout_secs(struct ctl_table *table, int write,
 232                                   void __user *buffer,
 233                                   size_t *lenp, loff_t *ppos)
 234 {
 235         int ret;
 236
 237         ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
 238
 239         if (ret || !write)
 240                 goto out;
 241
 242         wake_up_process(watchdog_task);
 243
 244  out:
 245         return ret;
 246 }
 247
 248 /*
 249  * This is needed for proc_doulongvec_minmax of sysctl_hung_task_timeout_secs
 250  * and hung_task_check_interval_secs
 251  */
 252 static const unsigned long hung_task_timeout_max = (LONG_MAX / HZ);
 253 static struct ctl_table hung_task_sysctls[] = {
 254 #ifdef CONFIG_SMP
 255         {
 256                 .procname       = "hung_task_all_cpu_backtrace",
 257                 .data           = &sysctl_hung_task_all_cpu_backtrace,
 258                 .maxlen         = sizeof(int),
 259                 .mode           = 0644,
 260                 .proc_handler   = proc_dointvec_minmax,
 261                 .extra1         = SYSCTL_ZERO,
 262                 .extra2         = SYSCTL_ONE,
 263         },
 264 #endif /* CONFIG_SMP */
 265         {
 266                 .procname       = "hung_task_panic",
 267                 .data           = &sysctl_hung_task_panic,
 268                 .maxlen         = sizeof(int),
 269                 .mode           = 0644,
 270                 .proc_handler   = proc_dointvec_minmax,
 271                 .extra1         = SYSCTL_ZERO,
 272                 .extra2         = SYSCTL_ONE,
 273         },
 274         {
 275                 .procname       = "hung_task_check_count",
 276                 .data           = &sysctl_hung_task_check_count,
 277                 .maxlen         = sizeof(int),
 278                 .mode           = 0644,
 279                 .proc_handler   = proc_dointvec_minmax,
 280                 .extra1         = SYSCTL_ZERO,
 281         },
 282         {
 283                 .procname       = "hung_task_timeout_secs",
 284                 .data           = &sysctl_hung_task_timeout_secs,
 285                 .maxlen         = sizeof(unsigned long),
 286                 .mode           = 0644,
 287                 .proc_handler   = proc_dohung_task_timeout_secs,
 288                 .extra2         = (void *)&hung_task_timeout_max,
 289         },
 290         {
 291                 .procname       = "hung_task_check_interval_secs",
 292                 .data           = &sysctl_hung_task_check_interval_secs,
 293                 .maxlen         = sizeof(unsigned long),
 294                 .mode           = 0644,
 295                 .proc_handler   = proc_dohung_task_timeout_secs,
 296                 .extra2         = (void *)&hung_task_timeout_max,
 297         },
 298         {
 299                 .procname       = "hung_task_warnings",
 300                 .data           = &sysctl_hung_task_warnings,
 301                 .maxlen         = sizeof(int),
 302                 .mode           = 0644,
 303                 .proc_handler   = proc_dointvec_minmax,
 304                 .extra1         = SYSCTL_NEG_ONE,
 305         },
 306         {}
 307 };
 308
 309 static void __init hung_task_sysctl_init(void)
 310 {
 311         register_sysctl_init("kernel", hung_task_sysctls);
 312 }
 313 #else
 314 #define hung_task_sysctl_init() do { } while (0)
 315 #endif /* CONFIG_SYSCTL */
 316
 317
 318 static atomic_t reset_hung_task = ATOMIC_INIT(0);
 319
 320 void reset_hung_task_detector(void)
 321 {
 322         atomic_set(&reset_hung_task, 1);
 323 }
 324 EXPORT_SYMBOL_GPL(reset_hung_task_detector);
 325
 326 static bool hung_detector_suspended;
 327
 328 static int hungtask_pm_notify(struct notifier_block *self,
 329                               unsigned long action, void *hcpu)
 330 {
 331         switch (action) {
 332         case PM_SUSPEND_PREPARE:
 333         case PM_HIBERNATION_PREPARE:
 334         case PM_RESTORE_PREPARE:
 335                 hung_detector_suspended = true;
 336                 break;
 337         case PM_POST_SUSPEND:
 338         case PM_POST_HIBERNATION:
 339         case PM_POST_RESTORE:
 340                 hung_detector_suspended = false;
 341                 break;
 342         default:
 343                 break;
 344         }
 345         return NOTIFY_OK;
 346 }
 347
 348 /*
 349  * kthread which checks for tasks stuck in D state
 350  */
 351 static int watchdog(void *dummy)
 352 {
 353         unsigned long hung_last_checked = jiffies;
 354
 355         set_user_nice(current, 0);
 356
 357         for ( ; ; ) {
 358                 unsigned long timeout = sysctl_hung_task_timeout_secs;
 359                 unsigned long interval = sysctl_hung_task_check_interval_secs;
 360                 long t;
 361
 362                 if (interval == 0)
 363                         interval = timeout;
 364                 interval = min_t(unsigned long, interval, timeout);
 365                 t = hung_timeout_jiffies(hung_last_checked, interval);
 366                 if (t <= 0) {
 367                         if (!atomic_xchg(&reset_hung_task, 0) &&
 368                             !hung_detector_suspended)
 369                                 check_hung_uninterruptible_tasks(timeout);
 370                         hung_last_checked = jiffies;
 371                         continue;
 372                 }
 373                 schedule_timeout_interruptible(t);
 374         }
 375
 376         return 0;
 377 }
 378
 379 static int __init hung_task_init(void)
 380 {
 381         atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
 382
 383         /* Disable hung task detector on suspend */
 384         pm_notifier(hungtask_pm_notify, 0);
 385
 386         watchdog_task = kthread_run(watchdog, NULL, "khungtaskd");
 387         hung_task_sysctl_init();
 388
 389         return 0;
 390 }
 391 subsys_initcall(hung_task_init);