Re: [PATCH] net-sysfs: Report link speed only when possible

[Michal Privoznik <mprivozn@redhat.com>]

On 16.06.2014 10:44, David Miller wrote:
> From: Michal Privoznik <mprivozn@redhat.com>
> Date: Mon, 16 Jun 2014 10:30:27 +0200
>
>> On 16.06.2014 10:11, David Miller wrote:
>>> From: Michal Privoznik <mprivozn@redhat.com>
>>> Date: Mon, 16 Jun 2014 09:32:35 +0200
>>>
>>>> On 13.06.2014 22:03, David Miller wrote:
>>>>> From: Michal Privoznik <mprivozn@redhat.com>
>>>>> Date: Fri, 13 Jun 2014 11:19:51 +0200
>>>>>
>>>>>> So if I were developing brand new application I could say: I'm
>>>>>> dropping all this workaround code and have it clean and require say
>>>>>> 3.16 kernel at least.
>>>>>
>>>>> Then your application wouldn't be usable on %99 of systems for a long
>>>>> long time.
>>>>>
>>>>
>>>> How come? The application is going to be usable for as long as
>>>> library/kernel APIs won't change.
>>>
>>> Because %99 of users are using a distribution kernel which is
>>> definitely
>>> going to be pre-3.16 for years.
>>>
>>
>> That's why every distribution out there has a mechanism to install
>> packages of a certain version, or those providing certain symbol,
>> whatever. Or distributions can then backport some kernel patches or
>> something. But, that's completely unrelated to the problem I'm fixing
>> here. I don't think this bikeshedding is useful for anything, sorry.
>
> You're being entirely impractical.
>
> By restricting an application to a kernel version or behavior "via
> backported patches" which doesn't even exist yet, you are foolishly
> restricting your userbase.

So? Users still have choice of not using my application. I'm okay with that.

>
> People just cope with what the current kernels support, when possible,
> and that's the right thing to do because we cannot break it on them
> exactly because people can depend upon the behavior.

Once again, we are not breaking anything. Current applications continue 
to work. I don't understand why you keep writing the opposite over and 
over again.

>
> NOBODY is checking for -EINVAL returns when reading the link speed
> sysfs file, and therefore by signalling it you will break
> applications.

That's very interesting thing to say, since even now one can experience 
EINVAL:

# cat /sys/class/net/wlan0/speed
cat: /sys/class/net/wlan0/speed: Invalid argument

How do you know for sure that NOBODY is checking -EINVAL?
For example libvirt does check EINVAL:

http://libvirt.org/git/?p=libvirt.git;a=blob;f=src/util/virnetdev.c;h=a551f9820b97aac41bbcb19c84d102c0ec3bd0aa;hb=HEAD#l1891

How can a kernel developer state that NOBODY isn't using possible kernel 
API anyway?

>
> So I will not apply a patch which adds that new behavior, sorry.

That's okay.

>
> I am not willing to discuss this further, this is fundamental and
> simple as far as I'm concerned.
>

Sure it is. That's why I'm surprised we even need to have this discussion.

Michal