Hi,

-----Original Message-----

On 11/18/21 5:18 AM, Heikki Krogerus wrote:

quoted

Hi,

On Thu, Nov 18, 2021 at 05:23:52PM +0800, Xu Yang wrote:

quoted

@@ -6428,6 +6432,9 @@ void tcpm_unregister_port(struct tcpm_port

*port)

quoted

quoted

  {
     int i;

You need to take the port lock here, no?

         mutex_lock(&port->lock);

quoted

+    kthread_destroy_worker(port->wq);
+    port->wq = NULL;

         mutex_unlock(&port->lock);

I don't think the timer code runs under the lock, so that won't help.
But, yes, the code is inherently racy. At the very least, port->wq
would have to be set to NULL first, but even that would have to be
synchronized. I wonder how other drivers handle that situation.

Guenter

Actually, the race is unlikely to happen if the hrtimer's expire time is a
little bigger. But it's better to avoid this situation from happening. So it
may be necessary to use a flag to indicate to the hrtimer the port is going
to be unregistered and it's no need to queue work anymore.

The patch maybe like following:

struct tcpm_port  *tcpm_register_port(...)
{
	...
	port->registered = true;
	return;
}

static enum hrtimer_restart  state_machine_timer_handler(timer)
{
	if (port->registered)
		kthread_queue_work(port->wq, work);
	return HRTIMER_NORESTART;
}

void tcpm_unregister_port(port)
{
	port->registered = false;
	kthread_destroy_worker(port->wq);

	hrtimer_cancel(&timer);
	...
}

What about this idea? Or I will submit it as patch V2.

Xu Yang