Re: [PATCH net-next 0/3] net: Switch tag HW extraction/insertion

[Florian Fainelli <f.fainelli@gmail.com>]

On 30/07/15 23:08, Eric Dumazet wrote:
> On Thu, 2015-07-30 at 18:51 -0700, Florian Fainelli wrote:
>> On 30/07/15 15:51, David Miller wrote:
>>> From: David Miller <davem@davemloft.net>
>>> Date: Thu, 30 Jul 2015 14:19:35 -0700 (PDT)
>>>
>>>> This looks fine, series applied, thanks.
>>>
>>> I think your control block is too large, you'll need to rework this
>>> somehow.
>>
>> So napi_gro_cb really is 48 bytes on 64-bits architectures (had not
>> realized it was so big).
>>
>> What would you recommend to do here considering that this driver is
>> currently used on 32-bits platforms, but I see no reason why someone
>> would no want to use this feature on a 64-bit platform, yet we are
>> competing with napi_gro_cb, and adding a new skbuff member is pretty
>> much a no-no? Would it be acceptable to have a new member which is ifdef
>> CONFIG_NET_DSA_TAG_BRCM?
>>
>> FWIW, this does provide a small 2-3% throughput increase for RX.
> 
> Which layer will read this tag after GRO processing ?

DSA would read this tag, but in general any ethertype hook using
packet_type would be in the same boat.

> 
> It seems you simply can use skb->cb[] like other layers : At offset 0.
> 
> BTW brcm_tag_rcv() does not even use GRO, as it calls
> netif_receive_skb()

That is right here, we will come from the RX processing of a driver that
might use napi_gro_receive, as it turns out the sequence looks like this
here:

bcm_sysport_desc_rx()
	-> extract tag
	eth_type_trans() sets skb->protocol to ETH_P_XDSA
	napi_gro_receive(), sets skb->cb to napi_gro_cb
		-> walk the list of packet_type find the one for ETH_P_XDSA
		-> brcm_tag_rcv()

BTW, there is no build time assertion that napi_gro_cb is not exceeding
skb->cb right now, even though they both have the same size on 64-bits,
should we have one?
-- 
Florian