Re: [net-next.git 1/7] stmmac: review napi gro support

[Giuseppe CAVALLARO <peppe.cavallaro@st.com>]

On 4/3/2013 3:39 PM, Eric Dumazet wrote:
> On Wed, 2013-04-03 at 09:41 +0200, Giuseppe CAVALLARO wrote:
>
>> Hmm, I'm in trouble on this :-). Indeed I can understand the (fatal)
>> race and why napi_complete should be used. Sorry! So my fault and this
>> patch has to be discarded. I don't understand why I have not seen any
>> problems while running/stressing on SMP system. Have you got any idea?
>>
>
> So because you don't hit the race on your machine and your tests, we can
> remove all the needed spinlock() and various hard irq masking, and
> introduce all sort of races ?

No, no this was not my intention. If fact, I asked to discard the patch
;-) ... it is clear to me the race that can happen in that case.

I'll rework the patch to only add the GRO in the feature.

peppe

>
> Try to use a combination of two NICS, and you'll hit the bug even on UP.
>
> There is a single poll_list per cpu, and we insert new elements in this
> list under hard irq.
>
> So deleting elements _must_ be done under the protection of hard irq
> masking.
>
>
>
>
>