khungtaskd 检测处于TASK_UNINTERRUPTIBLE状态的进程

在调试进程时dmesg碰到如下信息

======================
== Upstream sync On ==
======================
eth3: link up, full duplex, speed 1 Gbps
INFO: task PonMgr:510 blocked for more than 120 seconds.
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
PonMgr          D c03e08cc     0   510    509 0x00000000
Backtrace:
[<c03e0514>] (__schedule+0x0/0x494) from [<c03e0ad0>] (schedule+0x84/0x8c)
[<c03e0a4c>] (schedule+0x0/0x8c) from [<c03dee44>] (schedule_timeout+0x20/0x1e4)
[<c03dee24>] (schedule_timeout+0x0/0x1e4) from [<c03dfe98>] (__down+0x80/0xb4)
r7:00000000 r6:c7b67bc0 r5:7fffffff r4:bf038044
[<c03dfe18>] (__down+0x0/0xb4) from [<c012e190>] (down+0x34/0x48)
r8:00000000 r7:00000034 r6:be1ffa40 r5:a0000013 r4:bf038044
[<c012e15c>] (down+0x0/0x48) from [<bf0229a8>] (gpon_evt_read+0x1c/0xc8 [gpon_drv])
r5:be1ffa40 r4:00000034
[<bf02298c>] (gpon_evt_read+0x0/0xc8 [gpon_drv]) from [<c0195e80>] (vfs_read+0xb8/0x134)
r7:c5c05f70 r6:be1ffa40 r5:c6b84bc0 r4:00000034
[<c0195dc8>] (vfs_read+0x0/0x134) from [<c0195f40>] (sys_read+0x44/0x70)
r8:00000000 r7:00000003 r6:00000034 r5:be1ffa40 r4:c6b84bc0
[<c0195efc>] (sys_read+0x0/0x70) from [<c000d7a0>] (ret_fast_syscall+0x0/0x30)
r9:c5c04000 r8:c000d948 r6:00000034 r5:0000000c r4:be1ffa40

开始认为是代码存在问题，导致Backtrace；后搜索代码发现这是内核的一个保护机制；出现这个是因为编写的read函数一直在阻塞，没有接收到有效信息；

kernel/hung_task.c

内核线程启动如下：

static int __init hung_task_init(void)
{
     atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
     watchdog_task = kthread_run(watchdog, NULL, "khungtaskd");

     return 0;
}
module_init(hung_task_init);

/*
* kthread which checks for tasks stuck in D state
*/
static int watchdog(void *dummy)
{
     set_user_nice(current, 0);

     for ( ; ; ) {
          unsigned long timeout = sysctl_hung_task_timeout_secs;

          while (schedule_timeout_interruptible(timeout_jiffies(timeout)))
               timeout = sysctl_hung_task_timeout_secs;

          check_hung_uninterruptible_tasks(timeout);
     }

     return 0;
}

.config文件中配置sysctl_hung_task_timeout_secs为120，那么这个机制没隔2分钟检测一次；依次检测每一个进程控制块；当进程处于TASK_UNINTERRUPTIBLE状态时，调用check_hung_task；

     do_each_thread(g, t) {
          if (!max_count--)
               goto unlock;
          if (!--batch_count) {
               batch_count = HUNG_TASK_BATCHING;
               if (!rcu_lock_break(g, t))
                    goto unlock;
          }
          /* use "==" to skip the TASK_KILLABLE tasks waiting on NFS */
          if (t->state == TASK_UNINTERRUPTIBLE)
               check_hung_task(t, timeout);
     } while_each_thread(g, t);

接下来判断该进程是否试图进行过调度：

static void check_hung_task(struct task_struct *t, unsigned long timeout)

     unsigned long switch_count = t->nvcsw + t->nivcsw;
     /*
     * Ensure the task is not frozen.
     * Also, skip vfork and any other user process that freezer should skip.
     */
     if (unlikely(t->flags & (PF_FROZEN | PF_FREEZER_SKIP)))
         return;
     /*
     * When a freshly created task is scheduled once, changes its state to
     * TASK_UNINTERRUPTIBLE without having ever been switched out once, it
     * musn't be checked.
     */
     if (unlikely(!switch_count))
          return;
     if (switch_count != t->last_switch_count) {
          t->last_switch_count = switch_count;
          return;
     }
     if (!sysctl_hung_task_warnings)
          return;
     sysctl_hung_task_warnings--;

首先搞清nvcsw，nivcsw两个变量的含义：

nvcsw: Context voluntary switch counter 

nivcsw:  Context involuntary switch counter 

很明显这是上下文切换 自愿/非自愿的次数统计；

last_switch_count的意义是：hung task detection， 也就是只供khungtaskd检测TASK_UNINTERRUPTIBLE进程使用；

t->nvcsw + t->nivcsw==last_switch_count  就表示这两分钟内该进程没有被调度过，正是要找的进程，

这里last_switch_count==10，也就是 该机制最多只打印10次信息有效，10次后的检测结果就看不到了；

接下来打印堆栈信息

     printk(KERN_ERR "INFO: task %s:%d blocked for more than "
               "%ld seconds.\n", t->comm, t->pid, timeout);
     printk(KERN_ERR "\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
               " disables this message.\n");
     sched_show_task(t);
     debug_show_held_locks(t);

     touch_nmi_watchdog();

     if (sysctl_hung_task_panic)
          panic("hung_task: blocked tasks");

sysctl_hung_task_panic=CONFIG_BOOTPARAM_HUNG_TASK_PANIC_VALUE;

.config中没有定义CONFIG_BOOTPARAM_HUNG_TASK_PANIC_VALUE，  所以内核不会panic；