Re: [PATCH] [NET]: Fix Ooops of napi net_rx_action.

[David Miller <davem@davemloft.net>]

From: "Joonwoo Park" <joonwpark81@gmail.com>
Date: Tue, 11 Dec 2007 18:13:34 +0900

Joonwoo-ssi annyoung haseyo,

> [NET]: Fix Ooops of napi net_rx_action.
> Before doing list_move_tail napi poll_list, it should be ensured
> 
> Signed-off-by: Joonwoo Park <joonwpark81@gmail.com>
> ---
> diff --git a/net/core/dev.c b/net/core/dev.c
> index 86d6261..74bd5ab 100644
> --- a/net/core/dev.c
> +++ b/net/core/dev.c
> @@ -2207,7 +2207,8 @@ static void net_rx_action(struct softirq_action *h)
>  		 * still "owns" the NAPI instance and therefore can
>  		 * move the instance around on the list at-will.
>  		 */
> -		if (unlikely(work == weight))
> +		if (unlikely((test_bit(NAPI_STATE_SCHED, &n->state))
> +				&& (work == weight)))
>  			list_move_tail(&n->poll_list, list);
>  
>  		netpoll_poll_unlock(have);

How can the NAPI_STATE_SCHED bit be cleared externally yet we take
this list_move_tail() code path?

If NAPI_STATE_SCHED is cleared, work will be zero which will never be
equal to 'weight', and this we'll never attempt the list_move_tail().

If something clears NAPI_STATE_SCHED meanwhile, we have a serious race
and your patch is an incomplete bandaid.  For example, if it can
happen, then a case like:

		if (test_bit(NAPI_STATE_SCHED, &n->state))
	... something clears NAPI_STATE_SCHED right now ...
			work = n->poll(n, weight);

can crash too.