Re: [RFC PATCH 00/13] Always build GSO/GRO functionality into the kernel

[Vlad Yasevich <vyasevic@redhat.com>]

On 11/13/2012 09:25 PM, Eric Dumazet wrote:
> On Tue, 2012-11-13 at 20:24 -0500, Vlad Yasevich wrote:
>> This patch series is a revision suggested by Eric to solve the problem where
>> a host without IPv6 support drops GSO frames from the guest.
>>
>> The problem is that GSO/GRO support is per protocol, and when said protocol
>> is not loaded or is disabled, packets attempting to go through GSO/GRO code paths
>> are dropped.  This causes retransmissions and a two orders of magnitude drop in
>> performance.
>>
>> Prior attempt to solve the problem simply enabled enough GSO/GRO functionality
>> for IPv6 protocol when IPv6 was diabled.  This did not solve the problem when
>> the protocol was not build in or was blacklisted.
>> To solve the problem, it was suggested that we separate GSO/GRO callback
>> registration from packet processing registrations.  That way
>> GSO/GRO callbacks can be built into the kernel and always be there.
>> This patch series attempts to do just that.
>> * Patches 1 and 2 split the GSO/GRO handlers from packet_type and convert
>>    to the new structure.
>> * Patches 3, 4 and 5 do the same thing for net_protocol structure.
>> * The rest of the patches try to incrementally move the IPv6 GSO/GRO
>>    code out of the module and into the static kernel build.  Some IPv6
>>    helper functions also had to move as well.
>>
>> I am currently testing the patches, but if folks could look this over
>> and send me any comments, I'd appreciate it.
>
> Seems very nice !
>
> I am just wondering if GSO/GRO is fully enabled at every step ?
>

I think so.  I ran basic tests along most of the steps and it seemed to 
be enabled.  That's why this is a 13 patch series :)  Tried to do it 
incrementally without impacting functionality.

-vlad