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 */ 24int __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 */ 38unsigned long __read_mostly sysctl_hung_task_timeout_secs = CONFIG_DEFAULT_HUNG_TASK_TIMEOUT; 39 40int __read_mostly sysctl_hung_task_warnings = 10; 41 42static int __read_mostly did_panic; 43 44static 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 */ 50unsigned int __read_mostly sysctl_hung_task_panic = 51 CONFIG_BOOTPARAM_HUNG_TASK_PANIC_VALUE; 52 53static 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 63static int 64hung_task_panic(struct notifier_block *this, unsigned long event, void *ptr) 65{ 66 did_panic = 1; 67 68 return NOTIFY_DONE; 69} 70 71static struct notifier_block panic_block = { 72 .notifier_call = hung_task_panic, 73}; 74 75static 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) 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 pr_err("INFO: task %s:%d blocked for more than %ld seconds.\n", 112 t->comm, t->pid, timeout); 113 pr_err(" %s %s %.*s\n", 114 print_tainted(), init_utsname()->release, 115 (int)strcspn(init_utsname()->version, " "), 116 init_utsname()->version); 117 pr_err("\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\"" 118 " disables this message.\n"); 119 sched_show_task(t); 120 debug_show_held_locks(t); 121 122 touch_nmi_watchdog(); 123 124 if (sysctl_hung_task_panic) { 125 trigger_all_cpu_backtrace(); 126 panic("hung_task: blocked tasks"); 127 } 128} 129 130/* 131 * To avoid extending the RCU grace period for an unbounded amount of time, 132 * periodically exit the critical section and enter a new one. 133 * 134 * For preemptible RCU it is sufficient to call rcu_read_unlock in order 135 * to exit the grace period. For classic RCU, a reschedule is required. 136 */ 137static bool rcu_lock_break(struct task_struct *g, struct task_struct *t) 138{ 139 bool can_cont; 140 141 get_task_struct(g); 142 get_task_struct(t); 143 rcu_read_unlock(); 144 cond_resched(); 145 rcu_read_lock(); 146 can_cont = pid_alive(g) && pid_alive(t); 147 put_task_struct(t); 148 put_task_struct(g); 149 150 return can_cont; 151} 152 153/* 154 * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for 155 * a really long time (120 seconds). If that happens, print out 156 * a warning. 157 */ 158static void check_hung_uninterruptible_tasks(unsigned long timeout) 159{ 160 int max_count = sysctl_hung_task_check_count; 161 int batch_count = HUNG_TASK_BATCHING; 162 struct task_struct *g, *t; 163 164 /* 165 * If the system crashed already then all bets are off, 166 * do not report extra hung tasks: 167 */ 168 if (test_taint(TAINT_DIE) || did_panic) 169 return; 170 171 rcu_read_lock(); 172 for_each_process_thread(g, t) { 173 if (!max_count--) 174 goto unlock; 175 if (!--batch_count) { 176 batch_count = HUNG_TASK_BATCHING; 177 if (!rcu_lock_break(g, t)) 178 goto unlock; 179 } 180 /* use "==" to skip the TASK_KILLABLE tasks waiting on NFS */ 181 if (t->state == TASK_UNINTERRUPTIBLE) 182 check_hung_task(t, timeout); 183 } 184 unlock: 185 rcu_read_unlock(); 186} 187 188static unsigned long timeout_jiffies(unsigned long timeout) 189{ 190 /* timeout of 0 will disable the watchdog */ 191 return timeout ? timeout * HZ : MAX_SCHEDULE_TIMEOUT; 192} 193 194/* 195 * Process updating of timeout sysctl 196 */ 197int proc_dohung_task_timeout_secs(struct ctl_table *table, int write, 198 void __user *buffer, 199 size_t *lenp, loff_t *ppos) 200{ 201 int ret; 202 203 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos); 204 205 if (ret || !write) 206 goto out; 207 208 wake_up_process(watchdog_task); 209 210 out: 211 return ret; 212} 213 214static atomic_t reset_hung_task = ATOMIC_INIT(0); 215 216void reset_hung_task_detector(void) 217{ 218 atomic_set(&reset_hung_task, 1); 219} 220EXPORT_SYMBOL_GPL(reset_hung_task_detector); 221 222/* 223 * kthread which checks for tasks stuck in D state 224 */ 225static int watchdog(void *dummy) 226{ 227 set_user_nice(current, 0); 228 229 for ( ; ; ) { 230 unsigned long timeout = sysctl_hung_task_timeout_secs; 231 232 while (schedule_timeout_interruptible(timeout_jiffies(timeout))) 233 timeout = sysctl_hung_task_timeout_secs; 234 235 if (atomic_xchg(&reset_hung_task, 0)) 236 continue; 237 238 check_hung_uninterruptible_tasks(timeout); 239 } 240 241 return 0; 242} 243 244static int __init hung_task_init(void) 245{ 246 atomic_notifier_chain_register(&panic_notifier_list, &panic_block); 247 watchdog_task = kthread_run(watchdog, NULL, "khungtaskd"); 248 249 return 0; 250} 251subsys_initcall(hung_task_init); 252