Re: Detached address check

["Robert Swindells" <rjs%fdy2.co.uk@localhost>]

On Tue, November 21, 2017 9:26 am, Roy Marples wrote:
> On 18/11/2017 13:37, Robert Swindells wrote:
>> On Fri, November 17, 2017 4:57 pm, Robert Swindells wrote:
>>> I feel that the changes made to check whether a sending address is
>>> valid
>>> are too strict.
>>>
>>> The check means that an attempt to connect from the address of an
>>> interface that is marked detached to the same address will fail even
>>> though there is still a route for this address to the loopback device.
>>>
>>> The change was introduced in:
>>>
>>> <http://mail-index.netbsd.org/source-changes/2016/09/15/msg077726.html>
>>
>> To clarify, I expect to be able to do:
>>
>> # ifconfig wm0 inet 192.168.0.1 netmask 0xffffff00 up
>> # ping 192.168.0.1
>>
>> and for it to work every time, not just when wm0 has detected a
>> carrier. In the traditional usage, the packets don't go through wm0
>> anyway they go through lo0.
>
> Aside from ping, what else do you really want?

I want all network programs to work.

>>
>> The RFC 4862 referenced in the commit is for IPv6, I don't think that
>> IPv4 behaviour should change too.
>>
>> I'm not even convinced that I want this behaviour for IPv6, I would
>> prefer that addresses set using ifconfig(8) should stay valid until
>> I change them.
>
> What would you actually do with them though?
> You have localhost - 127.0.0.1 which works just fine.

I shouldn't need to use localhost to connect to the same machine, that
isn't the way that UNIX networking has traditionally worked, there
wouldn't be any need for the route from each interface address to lo0
if we adopt your new programming model.

I feel that setting DETACHED/TENTATIVE based on link status should
be removed from IPv4 altogether. I have commented it out in my tree
and my machine is working again.

For IPv6, I would enable it with a sysctl whose value is only set to
true when ipv6mode="autohost".

If this change makes it into NetBSD-8 then I think you are going to
spend a lot of time answering user questions.

Robert Swindells