Re: Races in net_rx_action vs netpoll?

[David Miller <davem@davemloft.net>]

From: Olaf Kirch <okir@lst.de>
Date: Wed, 4 Jul 2007 14:16:32 +0200

Another locking bug in netpoll, why am I not surprised? :-/

Thanks for reporting this Olaf.

> I think the only real fix for this is to restrict who is allowed
> to remove the interface from the poll_list. Only net_rx_action
> should be allowed to do so. A possible patch is given below
> (beware, it's untested so far)

I'm happy to entertain this kind of solution, but we really
need to first have an interface to change multiple bits
at a time in one atomic operation, because by itself this
patch doubles the number of atomices we do when starting
a NAPI poll.

>  static inline int __netif_rx_schedule_prep(struct net_device *dev)
>  {
> +	/* The driver may have decided that there's no more work
> +	 * to be done - but now another interrupt arrives, and
> +	 * we changed our mind. */
> +	smp_mb__before_clear_bit();
> +	clear_bit(__LINK_STATE_RX_COMPLETE, &dev->state);
> +
>  	return !test_and_set_bit(__LINK_STATE_RX_SCHED, &dev->state);
>  }
>  

Because of that change.