Re: [dpdk-dev] [PATCH v2 10/10] kni: add API to set link status on kernel interface

[Jason Wang <jasowang@redhat.com>]

On 2018年09月12日 07:14, Stephen Hemminger wrote:
> On Tue, 11 Sep 2018 19:07:47 -0300
> Dan Gora <dg@adax.com> wrote:
>
>> On Tue, Sep 11, 2018 at 6:52 PM, Stephen Hemminger
>> <stephen@networkplumber.org> wrote:
>>> The carrier state has no meaning when device is down, at least for physical
>>> devices. Because often the PHY is powered off when the device is marked down.
>> The thing that caught my attention is that when you mark a kernel
>> ethernet device 'down', you get a message that the link is down in the
>> syslog.
>>
>> snappy:root:bash 2645 => ip link set down dev eth0
>> Sep 11 18:32:48 snappy kernel: e1000e: eth0 NIC Link is Down
>>
>> With this method, that's not possible because you cannot change the
>> link state from the callback from kni_net_release.
>>
>> The carrier state doesn't have any meaning from a data transfer point
>> of view, but it's often useful for being able to diagnose connectivity
>> issues (is my cable plugged in or not).
>>
>> I'm still not really clear what the objection really is to the ioctl
>> method.  Is it just the number of changes?  That the kernel driver has
>> to change as well?  Just that there is another way to do it?
>>
>> thanks
>> dan
> I want to see KNI as part of the standard Linux kernel at some future date.
> Having KNI as an out of tree driver means it is doomed to chasing tail lights
> for the Linux kernel ABI instability and also problems with Linux distributions.

Why not use vhost_net instead? KNI duplicates its function.

Thanks

>
> One of the barriers to entry for Linux drivers is introducing new ioctl's.
> Ioctl's have issues with being device specific and also 32/64 compatiablity.
> If KNI has ioctl's it makes it harder to get merged some day.
>
> I freely admit that this is forcing KNI to respond to something that is not
> there yet, so if it is too hard, then doing it with ioctl is going to be
> necessary.