[dpdk-dev] [RFC 0/3] Use common Linux tools to control DPDK ports

Aaron Conole aconole at redhat.com
Thu Feb 4 15:40:18 CET 2016


Hi Ferruh,

I missed your original reply to me. Sorry.

Ferruh Yigit <ferruh.yigit at intel.com> writes:
> On Tue, Jan 19, 2016 at 01:29:32PM +0200, Panu Matilainen wrote:
>> On 01/19/2016 11:59 AM, Ferruh Yigit wrote:
>>> On Mon, Jan 18, 2016 at 11:20:02AM -0500, Aaron Conole wrote:
>>>> Ferruh Yigit <ferruh.yigit at intel.com> writes:
>>>>> This work is to make DPDK ports more visible and to enable using common
>>>>> Linux tools to configure DPDK ports.
>>>>
>>>> This is a good goal. Only question - why use an additional kernel module
>>>> to do this? Is it _JUST_ for ethtool support?
>>>
>>> Kernel module used to create/destroy Linux net_devices, and module has a simple
>>> driver for that device which only handles control messages by passing them into
>>> userspace.
>>>
>>> To represent DPDK ports as Linux net_devices we need kernel support.

Why? Just create tun/tap interface, no? Then you get a queue into the
network stack, as well. Subscribe to netlink, and you can get all of the
changes that happen in the system - just look for those messages that
relate to your tun device. At least, that's what I see right away (and I
have some private patches for this, and you can take them over if you want).

I think most of the stuff you are trying to solve already exists, but I
am probably misunderstanding something (apologies for that).

>>>> I think the other stuff
>>>> can be accomplished using netlink sockets + messages, no?
>>>
>>> Netlink sockets just used to communicate kernel-space - user-space, this is not
>>> why we need a kernel module, for example this communication is implemented in
>>> original KNI as part of FIFO.
>>>
>>>> The only
>>>> trepidation I would have with something like this is the support from
>>>> major vendors - out of tree modules are not generally supportable. Might
>>>> be good to get some of the ethtool commands as netlink messages as well,
>>>> then it is supportable with no 3rd party kernel modules.
>>>
>>> Yes, there is a out of three module problem for some distros, but unfortunately
>>> we are not able to find a solution for this case without an
>>> external kernel module.
>>>
>>> This patch is still an RFC and if we receive suggested solution without a kernel
>>> module, we can work on it together.
>>
>> If it has to be in the kernel then you need to find a design that is 
>> upstreamable. Out of tree kernel modules are not a solution, they're a 
>> problem that people are working on eliminating.
>>
>
> Hi Stephen, and other Linux experts in the mail list,
>
> Can you please help finding a upstreamable solution for kernel control path?
>
> Mainly what we are looking for is userspace network driver support in
> kernel, similar to what FUSE does but a much simple version.
>
> Above KCP module basically does this, by having a network driver which
> passing requests to userspace network driver, but it is not generic
> enough.
>
> I wonder if it is possible make it more generic by extending rtnetlink support:
> 1- Add a new network driver to Linux (or update existing one like tun)
> to forward requests, get responses.
> 2- Extend rtnelink to support to attach any userspace driver to this
> device? (ip link set <device> uspace <?> ?)
>
> Does this make sense?
>
> rtnetlink already supports creating interfaces, and it provides
> kernel/user space communication,
> with "attach" support interface learns about it's peer in usersppace
> and can communicate.
>
> FUSE like communication method also can be alternative to transfer
> request and responses, but since rtnelink support exists, no need to
> create something new think.
>
> Thanks,
> ferruh


More information about the dev mailing list