Bug#964902: libio-socket-ip-perl: AI_ADDRCONFIG breaks many test suites on IPv6-only hosts

[Dominic Hargreaves]

On Sat, Jul 11, 2020 at 11:14:23PM +0300, Niko Tyni wrote:
> Package: libio-socket-ip-perl
> Version: 0.39-2
> Severity: serious
> Tags: patch ipv6
> 
> This is a follow-up for #962047 "fails to build on IPv6-only buildds".
> 
> The background here is that a while ago new official build daemons were
> added that only have IPv6 connectivity, and this exposed a new class
> of build failures, mainly due to getaddrinfo(3) behaviour with the
> AI_ADDRCONFIG flag.
> 
> Quoting Julien Cristau there:
> 
> > FWIW, it seems IO::Socket::IP passes AI_ADDRCONFIG to getaddrinfo, which
> > then doesn't return ipv4 addresses because the host doesn't have
> > non-local ones, even though the address we're trying to resolve is
> > "127.0.0.1".
> > 
> > Passing either GetAddrInfoFlags => AI_NUMERICHOST or GetAddrInfoFlags =>
> > 0 to both IO::Socket::IP->new calls in the test file lets it pass, by
> > turning off the smarts in getaddrinfo.
> 
> We fixed the test suite of libio-socket-ip-perl itself in 0.39-2 by
> making it pass AI_NUMERICHOST in all the tests. However, it turns out
> that the problem is much more widespread than this.
> 
> I've been testing systematically for these issues by test rebuilding
> the whole archive, and found
> 
> 1) 138 packages that failed to build on a host with an IPv6 address
>    on a "real" interface, and IPv4 + IPv6 on lo, but no internet
>    connection
> 
> 2) 8 of those 138 also failed to build on a host with both an IPv6 address
>    and an IPv4 address on a "real interface", still without a real internet
>    connection (these are regular "needs internet" bugs, not IPv6 specific)
> 
> 3) only 42 of the remaining 130 packages fail to build with the same setup
>    as 1), after applying the patch discussed below to IO::Socket::IP.
> 
> So it looks like the AI_ADDRCONFIG default of IO-Socket-IP is causing
> 88 potential build failures on the new IPv6-only buildds. Examples of
> those can be seen on for instance
> 
>  https://buildd.debian.org/status/logs.php?pkg=libmojolicious-perl
> 
>  https://buildd.debian.org/status/logs.php?pkg=libtest-redisserver-perl
> 
>  https://buildd.debian.org/status/logs.php?pkg=libwww-mechanize-perl
> 
> The attached patch changes IO-Socket-IP to pass AI_NUMERICHOST to
> getaddrinfo(3) when the address looks like an IPv4 numeric address (rather
> than a hostname), and special cases a PeerHost value of "localhost"
> not to use AI_ADDRCONFIG.
> 
> I have tried quite a few variants, and this was the best I could do
> without removing AI_ADDRCONFIG (which is a documented feature) altogether.
> (TODO: The patch does not currently update the documentation about the
> new behaviour.)
> 
> It would be good to get something like this upstream of course, and it
> will also need to go in src:perl which has a copy of IO::Socket::IP.
> 
> I've tested that this makes libio-socket-perl pass its own test suite
> without the modifications in 0.39-2, so this supersedes the current test
> suite patch.
> 
> Build logs can currently (but not guaranteed indefinitely) be found at
> 
>  http://perl.debian.net/rebuild-logs/sid-v6only/
> 
>  http://perl.debian.net/rebuild-logs/experimental-v6only/
> 
> where the latter had a patched libio-socket-ip-perl package pre-installed
> in the chroot.

[snip list of affected packages]

As a heads up, there was some discussion on debian-devel about how this
is something which should be fixed in libc. This is the bug about the
issue from this perspective:

https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=952740

There doesn't seem any prospect of this being fixed in the near future,
so I'm planning to apply Niko's patch to libio-socket-ip-perl (and perl,
probably starting with experimental) to solve the immediate issues of
failing tests.

Dominic.