How to avoid zombie kernel threads?

[Jarno Paananen]

Hi,

I'm coding a driver (can be found from
http://hardsid.sourceforge.net/ is someone is actually interested)
that uses a kernel thread to do the actual work asynchronously from
rest of the world. The thread is created when opening a character
device and exits when the device is closed.

This works fine otherwise but if the user mode process opens and
closes the device multiple times during its lifetime I get N-1
zombie kernel threads where N is the number of opens.

The code goes like this:

in device open:

        DECLARE_MUTEX_LOCKED(sem);
        int rmmod = 0; 

        rmmod = 0;
        notify = &sem;
        kernel_thread(hsid_thread, (void *)sid_data, 0);
        down(&sem);
        notify = NULL;
 

in device close:

        notify = &sem;
        rmmod = 1;
        up(&todoSem); // just to wake the thread to do something
        down(&sem);
        notify = NULL;
        rmmod = 0;

and the thread itself does:

[daemonize() etc.]

    /* Notify the parent */
    if(notify != NULL)
        up(notify);

    for(;;)
    {
        if (rmmod || signal_pending(current))
            break;

        /* We sit here waiting for something to do */
        down_interruptible(&todoSem);

        if (rmmod || signal_pending(current))
            break;

        [actual work]
    }

    if(notify != NULL)
        up(notify);

    return 0;

I think this worked fine in earlier 2.4 versions (not sure though),
but I'm now seeing this in both 2.4.18-pre7 and 2.5.2-dj6, UP and SMP.

Thanks,

// Jarno