Re: [Qemu-devel] [PATCH V3 05/10] net/net.c: Add vnet header length to SocketReadState

[Zhang Chen <zhangchen.fnst@cn.fujitsu.com>]

On 05/09/2017 01:36 PM, Jason Wang wrote:
>
>
> On 2017年05月09日 12:03, Zhang Chen wrote:
>>
>>
>> On 05/09/2017 10:40 AM, Jason Wang wrote:
>>>
>>>
>>> On 2017年05月08日 11:47, Zhang Chen wrote:
>>>>
>>>>
>>>> On 05/03/2017 06:42 PM, Jason Wang wrote:
>>>>>
>>>>>
>>>>> On 2017年05月03日 11:43, Zhang Chen wrote:
>>>>>>
>>>>>>
>>>>>> On 05/02/2017 12:53 PM, Jason Wang wrote:
>>>>>>>
>>>>>>>
>>>>>>> On 2017年04月28日 17:47, Zhang Chen wrote:
>>>>>>>> Address Jason Wang's comments add vnet header length to 
>>>>>>>> SocketReadState.
>>>>>>>
>>>>>>> Instead of saying this, you can add "Suggested-by: Jason Wang 
>>>>>>> <jasowang@redhat.com>" above your sign-off.
>>>>>>
>>>>>> OK.
>>>>>>
>>>>>>>
>>>>>>>> So we change net_fill_rstate() to read
>>>>>>>> struct  {int size; int vnet_hdr_len; const uint8_t buf[];}.
>>>>>>>
>>>>>>> This makes me thinking about the backward compatibility. I think 
>>>>>>> we'd better try to keep it as much as possible. E.g how about 
>>>>>>> pack vnet_hdr_len into higher bits of size?
>>>>>>>
>>>>>>
>>>>>> Do you means split uint32_t size to uint16_t size and uint16_t 
>>>>>> vnet_hdr_len ?
>>>>>> If yes, we also can't keep compatibility with old version.
>>>>>> Old code send a uint32_t size, we read it as uint16_t size is 
>>>>>> always wrong.
>>>>>>
>>>>>> Thanks
>>>>>> Zhang Chen
>>>>>
>>>>> Consider it's unlikely to send or receive packet >= 64K, we can 
>>>>> reuse higher bits (e.g highest 8 bits). Then we can still read 
>>>>> uint32_t and then check its highest 8 bits. If it was zero, we 
>>>>> know vnet header is zero, if not it could be read as vnet header 
>>>>> length.
>>>>
>>>> I got your point, but in this way, packet size must < 64K, if size 
>>>> >=64K we still can't maintain the backward compatibility,
>>>> For the packet sender that didn't know the potential packet size 
>>>> limit, I think we should make code explicitly declared like
>>>> currently code. Otherwise we will make other people confused and 
>>>> make code difficult to maintain.
>>>>
>>>> Thanks
>>>> Zhang Chen
>>>>
>>>
>>> Yes, this is an possible issue in the future. Rethink about this, 
>>> what if we just compare vnet header? Is there any issue of doing 
>>> this? (Sorry, I remember this is a reason, but forget now).
>>
>> Yes, we can compare all packet with vnet header, the compare 
>> performance is very low. but we can't parse raw packet to 
>> tcp/udp/icmp packet, because we didn't know the vnet_hdr_len as the 
>> offset.
>>
>> Thanks
>> Zhang Chen
>
> Aha, so I think it's time to use the new format but:
>
> - probably need a new option to enable this support for filter

Do you means we should add the new option for every filter object like that:
Now:
-object filter-mirror,id=m0,netdev=hn0,queue=tx,outdev=mirror0
Feature:
-object 
filter-mirror,id=m0,netdev=hn0,queue=tx,outdev=mirror0,has_vnet_hdr_len=on

And colo-compare also need this option to get the vnet_hdr_len.


> - let's don't touch socket backend, and make it still can work with 
> filters whose vnet_hder is off

OK, Maybe we can add a new variable in net_fill_rstate().

Thanks
Zhang Chen

>
> Thanks
>
>>
>>>
>>> Thanks
>>>
>>>
>>> .
>>>
>>
>
>
>
> .
>

-- 
Thanks
Zhang Chen