[ovs-dev] [PATCH net-next v6] net: openvswitch: IPv6: Add IPv6 extension header support

[Nicolas Dichtel]

Le 30/09/2021 à 18:11, Cpp Code a écrit :
> On Wed, Sep 29, 2021 at 6:19 AM Jakub Kicinski <kuba at kernel.org> wrote:
>>
>> On Wed, 29 Sep 2021 08:19:05 +0200 Nicolas Dichtel wrote:
>>>> /* Insert a kernel only KEY_ATTR */
>>>> #define OVS_KEY_ATTR_TUNNEL_INFO    __OVS_KEY_ATTR_MAX
>>>> #undef OVS_KEY_ATTR_MAX
>>>> #define OVS_KEY_ATTR_MAX            __OVS_KEY_ATTR_MAX
>>> Following the other thread [1], this will break if a new app runs over an old
>>> kernel.
>>
>> Good point.
>>
>>> Why not simply expose this attribute to userspace and throw an error if a
>>> userspace app uses it?
>>
>> Does it matter if it's exposed or not? Either way the parsing policy
>> for attrs coming from user space should have a reject for the value.
>> (I say that not having looked at the code, so maybe I shouldn't...)
> 
> To remove some confusion, there are some architectural nuances if we
> want to extend code without large refactor.
> The ovs_key_attr is defined only in kernel side. Userspace side is
> generated from this file. As well the code can be built without kernel
> modules.
> The code inside OVS repository and net-next is not identical, but I
> try to keep some consistency.
I didn't get why OVS_KEY_ATTR_TUNNEL_INFO cannot be exposed to userspace.

> 
> JFYI This is the file responsible for generating userspace part:
> https://github.com/openvswitch/ovs/blob/master/build-aux/extract-odp-netlink-h
> This is the how corresponding file for ovs_key_attr looks inside OVS:
> https://github.com/openvswitch/ovs/blob/master/datapath/linux/compat/include/linux/openvswitch.h
> one can see there are more values than in net-next version.
There are still some '#ifdef __KERNEL__'. The standard 'make headers_install'
filters them. Why not using this standard mechanism?

In this file, there are two attributes (OVS_KEY_ATTR_PACKET_TYPE and
OVS_KEY_ATTR_ND_EXTENSIONS) that doesn't exist in the kernel.
This will also breaks if an old app runs over a new kernel. I don't see how it
is possible to keep the compat between {old|new} {kernel|app}.


Regards,
Nicolas