| 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/suspend.h> |
| 19 | #include <linux/utsname.h> |
| 20 | #include <linux/sched/signal.h> |
| 21 | #include <linux/sched/debug.h> |
| 22 | |
| 23 | #include <trace/events/sched.h> |
| 24 | |
| 25 | /* |
| 26 | * The number of tasks checked: |
| 27 | */ |
| 28 | int __read_mostly sysctl_hung_task_check_count = PID_MAX_LIMIT; |
| 29 | |
| 30 | /* |
| 31 | * Limit number of tasks checked in a batch. |
| 32 | * |
| 33 | * This value controls the preemptibility of khungtaskd since preemption |
| 34 | * is disabled during the critical section. It also controls the size of |
| 35 | * the RCU grace period. So it needs to be upper-bound. |
| 36 | */ |
| 37 | #define HUNG_TASK_LOCK_BREAK (HZ / 10) |
| 38 | |
| 39 | /* |
| 40 | * Zero means infinite timeout - no checking done: |
| 41 | */ |
| 42 | unsigned long __read_mostly sysctl_hung_task_timeout_secs = CONFIG_DEFAULT_HUNG_TASK_TIMEOUT; |
| 43 | |
| 44 | int __read_mostly sysctl_hung_task_warnings = 10; |
| 45 | |
| 46 | static int __read_mostly did_panic; |
| 47 | static bool hung_task_show_lock; |
| 48 | static bool hung_task_call_panic; |
| 49 | |
| 50 | static struct task_struct *watchdog_task; |
| 51 | |
| 52 | /* |
| 53 | * Should we panic (and reboot, if panic_timeout= is set) when a |
| 54 | * hung task is detected: |
| 55 | */ |
| 56 | unsigned int __read_mostly sysctl_hung_task_panic = |
| 57 | CONFIG_BOOTPARAM_HUNG_TASK_PANIC_VALUE; |
| 58 | |
| 59 | static int __init hung_task_panic_setup(char *str) |
| 60 | { |
| 61 | int rc = kstrtouint(str, 0, &sysctl_hung_task_panic); |
| 62 | |
| 63 | if (rc) |
| 64 | return rc; |
| 65 | return 1; |
| 66 | } |
| 67 | __setup("hung_task_panic=", hung_task_panic_setup); |
| 68 | |
| 69 | static int |
| 70 | hung_task_panic(struct notifier_block *this, unsigned long event, void *ptr) |
| 71 | { |
| 72 | did_panic = 1; |
| 73 | |
| 74 | return NOTIFY_DONE; |
| 75 | } |
| 76 | |
| 77 | static struct notifier_block panic_block = { |
| 78 | .notifier_call = hung_task_panic, |
| 79 | }; |
| 80 | |
| 81 | static void check_hung_task(struct task_struct *t, unsigned long timeout) |
| 82 | { |
| 83 | unsigned long switch_count = t->nvcsw + t->nivcsw; |
| 84 | |
| 85 | /* |
| 86 | * Ensure the task is not frozen. |
| 87 | * Also, skip vfork and any other user process that freezer should skip. |
| 88 | */ |
| 89 | if (unlikely(t->flags & (PF_FROZEN | PF_FREEZER_SKIP))) |
| 90 | return; |
| 91 | |
| 92 | /* |
| 93 | * When a freshly created task is scheduled once, changes its state to |
| 94 | * TASK_UNINTERRUPTIBLE without having ever been switched out once, it |
| 95 | * musn't be checked. |
| 96 | */ |
| 97 | if (unlikely(!switch_count)) |
| 98 | return; |
| 99 | |
| 100 | if (switch_count != t->last_switch_count) { |
| 101 | t->last_switch_count = switch_count; |
| 102 | return; |
| 103 | } |
| 104 | |
| 105 | trace_sched_process_hang(t); |
| 106 | |
| 107 | if (sysctl_hung_task_panic) { |
| 108 | console_verbose(); |
| 109 | hung_task_show_lock = true; |
| 110 | hung_task_call_panic = true; |
| 111 | } |
| 112 | |
| 113 | /* |
| 114 | * Ok, the task did not get scheduled for more than 2 minutes, |
| 115 | * complain: |
| 116 | */ |
| 117 | if (sysctl_hung_task_warnings) { |
| 118 | if (sysctl_hung_task_warnings > 0) |
| 119 | sysctl_hung_task_warnings--; |
| 120 | pr_err("INFO: task %s:%d blocked for more than %ld seconds.\n", |
| 121 | t->comm, t->pid, timeout); |
| 122 | pr_err(" %s %s %.*s\n", |
| 123 | print_tainted(), init_utsname()->release, |
| 124 | (int)strcspn(init_utsname()->version, " "), |
| 125 | init_utsname()->version); |
| 126 | pr_err("\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\"" |
| 127 | " disables this message.\n"); |
| 128 | sched_show_task(t); |
| 129 | hung_task_show_lock = true; |
| 130 | } |
| 131 | |
| 132 | touch_nmi_watchdog(); |
| 133 | } |
| 134 | |
| 135 | /* |
| 136 | * To avoid extending the RCU grace period for an unbounded amount of time, |
| 137 | * periodically exit the critical section and enter a new one. |
| 138 | * |
| 139 | * For preemptible RCU it is sufficient to call rcu_read_unlock in order |
| 140 | * to exit the grace period. For classic RCU, a reschedule is required. |
| 141 | */ |
| 142 | static bool rcu_lock_break(struct task_struct *g, struct task_struct *t) |
| 143 | { |
| 144 | bool can_cont; |
| 145 | |
| 146 | get_task_struct(g); |
| 147 | get_task_struct(t); |
| 148 | rcu_read_unlock(); |
| 149 | cond_resched(); |
| 150 | rcu_read_lock(); |
| 151 | can_cont = pid_alive(g) && pid_alive(t); |
| 152 | put_task_struct(t); |
| 153 | put_task_struct(g); |
| 154 | |
| 155 | return can_cont; |
| 156 | } |
| 157 | |
| 158 | /* |
| 159 | * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for |
| 160 | * a really long time (120 seconds). If that happens, print out |
| 161 | * a warning. |
| 162 | */ |
| 163 | static void check_hung_uninterruptible_tasks(unsigned long timeout) |
| 164 | { |
| 165 | int max_count = sysctl_hung_task_check_count; |
| 166 | unsigned long last_break = jiffies; |
| 167 | struct task_struct *g, *t; |
| 168 | |
| 169 | /* |
| 170 | * If the system crashed already then all bets are off, |
| 171 | * do not report extra hung tasks: |
| 172 | */ |
| 173 | if (test_taint(TAINT_DIE) || did_panic) |
| 174 | return; |
| 175 | |
| 176 | hung_task_show_lock = false; |
| 177 | rcu_read_lock(); |
| 178 | for_each_process_thread(g, t) { |
| 179 | if (!max_count--) |
| 180 | goto unlock; |
| 181 | if (time_after(jiffies, last_break + HUNG_TASK_LOCK_BREAK)) { |
| 182 | if (!rcu_lock_break(g, t)) |
| 183 | goto unlock; |
| 184 | last_break = jiffies; |
| 185 | } |
| 186 | /* use "==" to skip the TASK_KILLABLE tasks waiting on NFS */ |
| 187 | if (t->state == TASK_UNINTERRUPTIBLE) |
| 188 | check_hung_task(t, timeout); |
| 189 | } |
| 190 | unlock: |
| 191 | rcu_read_unlock(); |
| 192 | if (hung_task_show_lock) |
| 193 | debug_show_all_locks(); |
| 194 | if (hung_task_call_panic) { |
| 195 | trigger_all_cpu_backtrace(); |
| 196 | panic("hung_task: blocked tasks"); |
| 197 | } |
| 198 | } |
| 199 | |
| 200 | static long hung_timeout_jiffies(unsigned long last_checked, |
| 201 | unsigned long timeout) |
| 202 | { |
| 203 | /* timeout of 0 will disable the watchdog */ |
| 204 | return timeout ? last_checked - jiffies + timeout * HZ : |
| 205 | MAX_SCHEDULE_TIMEOUT; |
| 206 | } |
| 207 | |
| 208 | /* |
| 209 | * Process updating of timeout sysctl |
| 210 | */ |
| 211 | int proc_dohung_task_timeout_secs(struct ctl_table *table, int write, |
| 212 | void __user *buffer, |
| 213 | size_t *lenp, loff_t *ppos) |
| 214 | { |
| 215 | int ret; |
| 216 | |
| 217 | ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos); |
| 218 | |
| 219 | if (ret || !write) |
| 220 | goto out; |
| 221 | |
| 222 | wake_up_process(watchdog_task); |
| 223 | |
| 224 | out: |
| 225 | return ret; |
| 226 | } |
| 227 | |
| 228 | static atomic_t reset_hung_task = ATOMIC_INIT(0); |
| 229 | |
| 230 | void reset_hung_task_detector(void) |
| 231 | { |
| 232 | atomic_set(&reset_hung_task, 1); |
| 233 | } |
| 234 | EXPORT_SYMBOL_GPL(reset_hung_task_detector); |
| 235 | |
| 236 | static bool hung_detector_suspended; |
| 237 | |
| 238 | static int hungtask_pm_notify(struct notifier_block *self, |
| 239 | unsigned long action, void *hcpu) |
| 240 | { |
| 241 | switch (action) { |
| 242 | case PM_SUSPEND_PREPARE: |
| 243 | case PM_HIBERNATION_PREPARE: |
| 244 | case PM_RESTORE_PREPARE: |
| 245 | hung_detector_suspended = true; |
| 246 | break; |
| 247 | case PM_POST_SUSPEND: |
| 248 | case PM_POST_HIBERNATION: |
| 249 | case PM_POST_RESTORE: |
| 250 | hung_detector_suspended = false; |
| 251 | break; |
| 252 | default: |
| 253 | break; |
| 254 | } |
| 255 | return NOTIFY_OK; |
| 256 | } |
| 257 | |
| 258 | /* |
| 259 | * kthread which checks for tasks stuck in D state |
| 260 | */ |
| 261 | static int watchdog(void *dummy) |
| 262 | { |
| 263 | unsigned long hung_last_checked = jiffies; |
| 264 | |
| 265 | set_user_nice(current, 0); |
| 266 | |
| 267 | for ( ; ; ) { |
| 268 | unsigned long timeout = sysctl_hung_task_timeout_secs; |
| 269 | long t = hung_timeout_jiffies(hung_last_checked, timeout); |
| 270 | |
| 271 | if (t <= 0) { |
| 272 | if (!atomic_xchg(&reset_hung_task, 0) && |
| 273 | !hung_detector_suspended) |
| 274 | check_hung_uninterruptible_tasks(timeout); |
| 275 | hung_last_checked = jiffies; |
| 276 | continue; |
| 277 | } |
| 278 | schedule_timeout_interruptible(t); |
| 279 | } |
| 280 | |
| 281 | return 0; |
| 282 | } |
| 283 | |
| 284 | static int __init hung_task_init(void) |
| 285 | { |
| 286 | atomic_notifier_chain_register(&panic_notifier_list, &panic_block); |
| 287 | |
| 288 | /* Disable hung task detector on suspend */ |
| 289 | pm_notifier(hungtask_pm_notify, 0); |
| 290 | |
| 291 | watchdog_task = kthread_run(watchdog, NULL, "khungtaskd"); |
| 292 | |
| 293 | return 0; |
| 294 | } |
| 295 | subsys_initcall(hung_task_init); |