On Tue, Jul 21, 2020 at 09:48:15PM +0200, Uwe Kleine-König wrote:
> Hello Johan,
> 
> On Tue, Jul 14, 2020 at 09:13:55AM +0200, Johan Hovold wrote:
> > On Tue, Jul 07, 2020 at 06:59:58PM +0200, Uwe Kleine-König wrote:
> > > +	while (firstrun ||
> > > +	       icount.rx != trigger_data->rx ||
> > > +	       icount.tx != trigger_data->tx) {
> > > +
> > > +		led_set_brightness(trigger_data->led_cdev, LED_ON);
> > > +
> > > +		msleep(100);
> > > +
> > > +		led_set_brightness(trigger_data->led_cdev, LED_OFF);
> > > +
> > > +		trigger_data->rx = icount.rx;
> > > +		trigger_data->tx = icount.tx;
> > > +		firstrun = false;
> > > +
> > > +		ret = tty_get_icount(trigger_data->tty, &icount);
> > > +		if (ret)
> > > +			return;
> > > +	}
> > 
> > Haven't looked at the latest proposal in detail, but this looks broken
> > as you can potentially loop indefinitely in a worker thread, and with no
> > way to stop the trigger (delayed work).
> 
> I don't think that potentially looping indefinitely is a problem, but
> indeed it should drop the lock during each iteration. Will think about
> how to adapt.

You musn't queue work that can run for long on the global shared
workqueue as it affects flushing:

	* system_wq is the one used by schedule[_delayed]_work[_on]().
	* Multi-CPU multi-threaded.  There are users which expect relatively
	* short queue flush time.  Don't queue works which can run for too
	* long.

Work that potentially run indefinitely is an absolute no-no. :)

Johan