Re: [PATCH net] net: Clear local_df only if crossing namespace.

[Nicolas Dichtel <nicolas.dichtel@6wind.com>]

Le 12/02/2014 18:05, Pravin Shelar a écrit :
> On Wed, Feb 12, 2014 at 1:32 AM, Nicolas Dichtel
> <nicolas.dichtel@6wind.com> wrote:
>> Le 12/02/2014 05:26, Pravin Shelar a écrit :
>>
>>> On Mon, Feb 10, 2014 at 6:11 PM, Hannes Frederic Sowa
>>> <hannes@stressinduktion.org> wrote:
>>>>
>>>> On Mon, Feb 10, 2014 at 01:00:14PM -0800, Pravin Shelar wrote:
>>>>>
>>>>> On Fri, Feb 7, 2014 at 4:58 PM, Hannes Frederic Sowa
>>>>> <hannes@stressinduktion.org> wrote:
>>>>>>
>>>>>> May I know because of wich vport, vxlan or gre, you did this change?
>>>>>>
>>>>> It affects both gre and vxlan.
>>>>
>>>>
>>>> Ok, thanks.
>>>>
>>>>>> I am feeling a bit uncomfortable handling remote and local packets that
>>>>>> differently on lower tunnel output (local_df is mostly set on locally
>>>>>> originating packets).
>>>>>
>>>>>
>>>>> For ip traffic it make sense to turn on local_df only for local
>>>>> traffic, since for remote case we can send icmp (frag-needed) back to
>>>>> source. No such thing exist for OVS tunnels. ICMP packet are not
>>>>> returned to source for the tunnels. That is why to be on safe side,
>>>>> local_df is turned on for tunnels in OVS.
>>>>
>>>>
>>>> I have a proposal:
>>>>
>>>> I don't like it that much because of the many arguments. But I currently
>>>> don't see another easy solution. Maybe we should make bool xnet an enum
>>>> and
>>>> test with bitops?
>>>>
>>>> I left the clearing of local_df in skb_scrub_packet as we need it for the
>>>> dev_forward_skb case and it should be done that in any case.
>>>>
>>>> This diff is slightly compile tested. ;)
>>>>
>>>> I can test and make proper submit if you agree.
>>>>
>>>> What do you think?
>>>>
>>>
>>> I am not sure why the caller can not just set skb->local_df before
>>> calling iptunnel_xmit() rather than passing extra arg to this
>>> function?
>>> There are not that many caller of this function.
>>
>> The benefit is that it ensures that future callers will think about this
>> point
>> ;-)
>>
>
> But that add extra test cases in fast path.
> For example OVS we can not reset skb->mark in skb_scrub_packet(). I am
> going to send patch for that too. Do you think I should add another
> argument for skb-mark clear too ?
Maybe this will be the same argument than local_df: 'bool ovs' (probably find a
better name ;-))