Mandeep Singh Baines | e162b39 | 2009-01-15 11:08:40 -0800 | [diff] [blame] | 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/module.h> |
| 17 | #include <linux/sysctl.h> |
| 18 | |
| 19 | /* |
Mandeep Singh Baines | ce9dbe2 | 2009-02-04 20:35:48 -0800 | [diff] [blame] | 20 | * The number of tasks checked: |
Mandeep Singh Baines | e162b39 | 2009-01-15 11:08:40 -0800 | [diff] [blame] | 21 | */ |
Mandeep Singh Baines | ce9dbe2 | 2009-02-04 20:35:48 -0800 | [diff] [blame] | 22 | unsigned long __read_mostly sysctl_hung_task_check_count = PID_MAX_LIMIT; |
| 23 | |
| 24 | /* |
| 25 | * Limit number of tasks checked in a batch. |
| 26 | * |
| 27 | * This value controls the preemptibility of khungtaskd since preemption |
| 28 | * is disabled during the critical section. It also controls the size of |
| 29 | * the RCU grace period. So it needs to be upper-bound. |
| 30 | */ |
| 31 | #define HUNG_TASK_BATCHING 1024 |
Mandeep Singh Baines | e162b39 | 2009-01-15 11:08:40 -0800 | [diff] [blame] | 32 | |
| 33 | /* |
| 34 | * Zero means infinite timeout - no checking done: |
| 35 | */ |
| 36 | unsigned long __read_mostly sysctl_hung_task_timeout_secs = 120; |
| 37 | static unsigned long __read_mostly hung_task_poll_jiffies; |
| 38 | |
| 39 | unsigned long __read_mostly sysctl_hung_task_warnings = 10; |
| 40 | |
| 41 | static int __read_mostly did_panic; |
| 42 | |
| 43 | static struct task_struct *watchdog_task; |
| 44 | |
| 45 | /* |
| 46 | * Should we panic (and reboot, if panic_timeout= is set) when a |
| 47 | * hung task is detected: |
| 48 | */ |
| 49 | unsigned int __read_mostly sysctl_hung_task_panic = |
| 50 | CONFIG_BOOTPARAM_HUNG_TASK_PANIC_VALUE; |
| 51 | |
| 52 | static int __init hung_task_panic_setup(char *str) |
| 53 | { |
| 54 | sysctl_hung_task_panic = simple_strtoul(str, NULL, 0); |
| 55 | |
| 56 | return 1; |
| 57 | } |
| 58 | __setup("hung_task_panic=", hung_task_panic_setup); |
| 59 | |
| 60 | static int |
| 61 | hung_task_panic(struct notifier_block *this, unsigned long event, void *ptr) |
| 62 | { |
| 63 | did_panic = 1; |
| 64 | |
| 65 | return NOTIFY_DONE; |
| 66 | } |
| 67 | |
| 68 | static struct notifier_block panic_block = { |
| 69 | .notifier_call = hung_task_panic, |
| 70 | }; |
| 71 | |
| 72 | /* |
| 73 | * Returns seconds, approximately. We don't need nanosecond |
| 74 | * resolution, and we don't need to waste time with a big divide when |
| 75 | * 2^30ns == 1.074s. |
| 76 | */ |
| 77 | static unsigned long get_timestamp(void) |
| 78 | { |
| 79 | int this_cpu = raw_smp_processor_id(); |
| 80 | |
| 81 | return cpu_clock(this_cpu) >> 30LL; /* 2^30 ~= 10^9 */ |
| 82 | } |
| 83 | |
Mandeep Singh Baines | 603a148 | 2009-01-17 10:31:48 -0800 | [diff] [blame] | 84 | static void check_hung_task(struct task_struct *t, unsigned long now, |
| 85 | unsigned long timeout) |
Mandeep Singh Baines | e162b39 | 2009-01-15 11:08:40 -0800 | [diff] [blame] | 86 | { |
| 87 | unsigned long switch_count = t->nvcsw + t->nivcsw; |
| 88 | |
| 89 | if (t->flags & PF_FROZEN) |
| 90 | return; |
| 91 | |
| 92 | if (switch_count != t->last_switch_count || !t->last_switch_timestamp) { |
| 93 | t->last_switch_count = switch_count; |
| 94 | t->last_switch_timestamp = now; |
| 95 | return; |
| 96 | } |
Mandeep Singh Baines | 603a148 | 2009-01-17 10:31:48 -0800 | [diff] [blame] | 97 | if ((long)(now - t->last_switch_timestamp) < timeout) |
Mandeep Singh Baines | e162b39 | 2009-01-15 11:08:40 -0800 | [diff] [blame] | 98 | return; |
| 99 | if (!sysctl_hung_task_warnings) |
| 100 | return; |
| 101 | sysctl_hung_task_warnings--; |
| 102 | |
| 103 | /* |
| 104 | * Ok, the task did not get scheduled for more than 2 minutes, |
| 105 | * complain: |
| 106 | */ |
| 107 | printk(KERN_ERR "INFO: task %s:%d blocked for more than " |
Mandeep Singh Baines | 603a148 | 2009-01-17 10:31:48 -0800 | [diff] [blame] | 108 | "%ld seconds.\n", t->comm, t->pid, timeout); |
Mandeep Singh Baines | e162b39 | 2009-01-15 11:08:40 -0800 | [diff] [blame] | 109 | printk(KERN_ERR "\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\"" |
| 110 | " disables this message.\n"); |
| 111 | sched_show_task(t); |
| 112 | __debug_show_held_locks(t); |
| 113 | |
| 114 | t->last_switch_timestamp = now; |
| 115 | touch_nmi_watchdog(); |
| 116 | |
| 117 | if (sysctl_hung_task_panic) |
| 118 | panic("hung_task: blocked tasks"); |
| 119 | } |
| 120 | |
| 121 | /* |
Mandeep Singh Baines | ce9dbe2 | 2009-02-04 20:35:48 -0800 | [diff] [blame] | 122 | * To avoid extending the RCU grace period for an unbounded amount of time, |
| 123 | * periodically exit the critical section and enter a new one. |
| 124 | * |
| 125 | * For preemptible RCU it is sufficient to call rcu_read_unlock in order |
| 126 | * exit the grace period. For classic RCU, a reschedule is required. |
| 127 | */ |
| 128 | static void rcu_lock_break(struct task_struct *g, struct task_struct *t) |
| 129 | { |
| 130 | get_task_struct(g); |
| 131 | get_task_struct(t); |
| 132 | rcu_read_unlock(); |
| 133 | cond_resched(); |
| 134 | rcu_read_lock(); |
| 135 | put_task_struct(t); |
| 136 | put_task_struct(g); |
| 137 | } |
| 138 | |
| 139 | /* |
Mandeep Singh Baines | e162b39 | 2009-01-15 11:08:40 -0800 | [diff] [blame] | 140 | * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for |
| 141 | * a really long time (120 seconds). If that happens, print out |
| 142 | * a warning. |
| 143 | */ |
Mandeep Singh Baines | 603a148 | 2009-01-17 10:31:48 -0800 | [diff] [blame] | 144 | static void check_hung_uninterruptible_tasks(unsigned long timeout) |
Mandeep Singh Baines | e162b39 | 2009-01-15 11:08:40 -0800 | [diff] [blame] | 145 | { |
| 146 | int max_count = sysctl_hung_task_check_count; |
Mandeep Singh Baines | ce9dbe2 | 2009-02-04 20:35:48 -0800 | [diff] [blame] | 147 | int batch_count = HUNG_TASK_BATCHING; |
Mandeep Singh Baines | e162b39 | 2009-01-15 11:08:40 -0800 | [diff] [blame] | 148 | unsigned long now = get_timestamp(); |
| 149 | struct task_struct *g, *t; |
| 150 | |
| 151 | /* |
| 152 | * If the system crashed already then all bets are off, |
| 153 | * do not report extra hung tasks: |
| 154 | */ |
| 155 | if (test_taint(TAINT_DIE) || did_panic) |
| 156 | return; |
| 157 | |
Mandeep Singh Baines | 94be52d | 2009-02-05 09:56:08 -0800 | [diff] [blame^] | 158 | rcu_read_lock(); |
Mandeep Singh Baines | e162b39 | 2009-01-15 11:08:40 -0800 | [diff] [blame] | 159 | do_each_thread(g, t) { |
| 160 | if (!--max_count) |
| 161 | goto unlock; |
Mandeep Singh Baines | ce9dbe2 | 2009-02-04 20:35:48 -0800 | [diff] [blame] | 162 | if (!--batch_count) { |
| 163 | batch_count = HUNG_TASK_BATCHING; |
| 164 | rcu_lock_break(g, t); |
| 165 | /* Exit if t or g was unhashed during refresh. */ |
| 166 | if (t->state == TASK_DEAD || g->state == TASK_DEAD) |
| 167 | goto unlock; |
| 168 | } |
Mandeep Singh Baines | e162b39 | 2009-01-15 11:08:40 -0800 | [diff] [blame] | 169 | /* use "==" to skip the TASK_KILLABLE tasks waiting on NFS */ |
| 170 | if (t->state == TASK_UNINTERRUPTIBLE) |
Mandeep Singh Baines | 603a148 | 2009-01-17 10:31:48 -0800 | [diff] [blame] | 171 | check_hung_task(t, now, timeout); |
Mandeep Singh Baines | e162b39 | 2009-01-15 11:08:40 -0800 | [diff] [blame] | 172 | } while_each_thread(g, t); |
| 173 | unlock: |
Mandeep Singh Baines | 94be52d | 2009-02-05 09:56:08 -0800 | [diff] [blame^] | 174 | rcu_read_unlock(); |
Mandeep Singh Baines | e162b39 | 2009-01-15 11:08:40 -0800 | [diff] [blame] | 175 | } |
| 176 | |
| 177 | static void update_poll_jiffies(void) |
| 178 | { |
| 179 | /* timeout of 0 will disable the watchdog */ |
| 180 | if (sysctl_hung_task_timeout_secs == 0) |
| 181 | hung_task_poll_jiffies = MAX_SCHEDULE_TIMEOUT; |
| 182 | else |
| 183 | hung_task_poll_jiffies = sysctl_hung_task_timeout_secs * HZ / 2; |
| 184 | } |
| 185 | |
| 186 | /* |
| 187 | * Process updating of timeout sysctl |
| 188 | */ |
| 189 | int proc_dohung_task_timeout_secs(struct ctl_table *table, int write, |
| 190 | struct file *filp, void __user *buffer, |
| 191 | size_t *lenp, loff_t *ppos) |
| 192 | { |
| 193 | int ret; |
| 194 | |
| 195 | ret = proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos); |
| 196 | |
| 197 | if (ret || !write) |
| 198 | goto out; |
| 199 | |
| 200 | update_poll_jiffies(); |
| 201 | |
| 202 | wake_up_process(watchdog_task); |
| 203 | |
| 204 | out: |
| 205 | return ret; |
| 206 | } |
| 207 | |
| 208 | /* |
| 209 | * kthread which checks for tasks stuck in D state |
| 210 | */ |
| 211 | static int watchdog(void *dummy) |
| 212 | { |
| 213 | set_user_nice(current, 0); |
| 214 | update_poll_jiffies(); |
| 215 | |
| 216 | for ( ; ; ) { |
Mandeep Singh Baines | 603a148 | 2009-01-17 10:31:48 -0800 | [diff] [blame] | 217 | unsigned long timeout; |
| 218 | |
Mandeep Singh Baines | e162b39 | 2009-01-15 11:08:40 -0800 | [diff] [blame] | 219 | while (schedule_timeout_interruptible(hung_task_poll_jiffies)); |
Mandeep Singh Baines | 603a148 | 2009-01-17 10:31:48 -0800 | [diff] [blame] | 220 | |
| 221 | /* |
| 222 | * Need to cache timeout here to avoid timeout being set |
| 223 | * to 0 via sysctl while inside check_hung_*_tasks(). |
| 224 | */ |
| 225 | timeout = sysctl_hung_task_timeout_secs; |
| 226 | if (timeout) |
| 227 | check_hung_uninterruptible_tasks(timeout); |
Mandeep Singh Baines | e162b39 | 2009-01-15 11:08:40 -0800 | [diff] [blame] | 228 | } |
| 229 | |
| 230 | return 0; |
| 231 | } |
| 232 | |
| 233 | static int __init hung_task_init(void) |
| 234 | { |
| 235 | atomic_notifier_chain_register(&panic_notifier_list, &panic_block); |
| 236 | watchdog_task = kthread_run(watchdog, NULL, "khungtaskd"); |
| 237 | |
| 238 | return 0; |
| 239 | } |
| 240 | |
| 241 | module_init(hung_task_init); |