On 2026-07-31 12:33, Chris Hofstaedtler wrote:
> On Sat, Jul 18, 2026 at 08:45:25PM -0700, H. Peter Anvin wrote:
>> On 2026-07-18 20:03, Chris Hofstaedtler wrote:
>>> On Tue, Feb 14, 2017 at 09:32:43PM +0100, Uwe Kleine-König wrote:
>>>> On 02/14/2017 08:17 PM, H. Peter Anvin wrote:
>>>>> Okay, let me chime in here.
>>>>
>>>> That's great, thanks.
>>>>
>>>>> AI_ADDRCONFIG seems to be the Wrong Thing[TM].
>>>>> AI_PASSIVE seems to be the Right Thing[TM].
>>>>
>>>> And you even seem to agree with me, that's still greater :-)
>>>
>>> Was this patch ever applied? I see set_sock_addr still uses
>>> hints.ai_flags = AI_CANONNAME | AI_ADDRCONFIG;
>>>
>>> Maybe another solution got applied? I'd love to close the bug on the
>>> Debian side.
>>
>> Would it be possible to see if b2c36a0b5ab71ff4abb1e321ba05064ea333177b
>> solves
>> the problem?
>
> So I don't know what Uwe had in mind back when suggesting the GAI
> changes. However when trying commit
> b2c36a0b5ab71ff4abb1e321ba05064ea333177b now, per its description,
> it should be possible to pick any address and tftpd should then
> listen on it?
>
> I think these situations should basically be the same:
> * as described in the bug, an IP address is not assigned *yet* to
> the interface
> * my test scenario, I picked an IP address that I did not assign to
> any interface
>
> In my test scenario, tftpd with commit
> b2c36a0b5ab71ff4abb1e321ba05064ea333177b applied still does not
> start, instead it gives:
> in.tftpd: cannot bind to local IPv4 socket: Cannot assign requested address
>
> Now I should probably test this on a system that has *no* IPv4
> address at all, to see if that makes a difference. But that's a bit
> harder right now.
>
> From my understanding of how other programs work, to really support
> binding to an address that is not yet assigned to an interface,
> bind() also needs setsockopt IP_FREEBIND first.
>
> Thanks for working on this!
>
I think you might be right. It's a bit dangerous, but for the specific case of
tftpd it is probably the right thing to do.
Another thing that I would like to do is to support systemd socket activation,
mainly because it naturally solves a lot of the problems with needing to drop
privilege. It is just that *any* changes to this ancient code base is really
painful.
-hpa
