Re: DragonFly-2.1.1.346.g4d1d7 master usr.sbin/dntpd client.h defs.h dntpd.8 main.c socket.c

["Simon 'corecode' Schubert" <corecode@xxxxxxxxxxxx>]

Sepherosa Ziehau wrote:
> On Fri, Jan 9, 2009 at 9:27 PM, Simon 'corecode' Schubert
> <corecode@fs.ei.tum.de> wrote:
>> Sepherosa Ziehau wrote:
>>> On Fri, Jan 9, 2009 at 8:47 PM, Simon 'corecode' Schubert
>>> <corecode@fs.ei.tum.de> wrote:
>>>> Sepherosa Ziehau wrote:
>>>>> On Fri, Jan 9, 2009 at 8:19 PM, Simon 'corecode' Schubert
>>>>> <corecode@fs.ei.tum.de> wrote:
>>>>>> Matthias Schmidt wrote:
>>>>>>> commit 4d1d74ba3c1fe85df343e4059c80bc704eec0312
>>>>>>> Author: Matthias Schmidt <matthias@dragonflybsd.org>
>>>>>>> Date:   Fri Jan 9 12:35:25 2009 +0100
>>>>>>>
>>>>>>>  Add IPv6 support to dntpd
>>>>>>>      This commit allows dntpd to work with either IPv4 or IPv6 NTP
>>>>>>> servers.  I added
>>>>>>>  two new command line switches (-4/-6) to force dntpd to use only
>>>>>>> v4/v6
>>>>>>> addresses.
>>>>>>>  Thanks to sephe@ for some good hints.
>>>>>> Are you sure this works if the ipv6 server is unreachable?  I don't
>>>>>> think
>>>>>> that a connect(2) on a udp socket will find that out.  You'd actually
>>>>>> have
>>>>> udp_socket() only sets up a UDP socket; by calling connect(2) we could
>>>>> receive ICMP error and use write(2) like functions.  As far as I
>>>>> understand the code, it never intends to verify the reachability of
>>>>> the server.
>>>> But it has to do that now!  The DNS might return ipv6 and ipv4, but the
>>>> ipv4
>>>> one is only reachable.  So you *have* to check availability, otherwise it
>>> Do you mean the host does not have global scope ipv6 unicast addresses
>>> or the host does not have a valid default ipv6 route?  If that's what
>>> you mean, connect(2) will fail on all resolved ipv6 server addresses.
>> The host might have an ipv6 address listed in DNS, but might not listen to
>> it, or there might be a routing advertisement for ipv6 on the local network
>> but routing might be broken nevertheless, etc.  IPV6 might be
>> filtered upstream, etc.
> 
> Quickest solution, but still incorrect, is prefer ipv4 address over
> ipv6 address or use ipv4 address only by default.
> Else we try the ipv4 address if one of the ipv6 address failed
> (read(2) server response timed out or got ICMP error in the main
> loop), and try ipv6 address if ipv4 address failed.  I don't remember
> NTP has keepalive or echo like messages, so I don't think we should
> check reachability in udp_socket().

Oh, maybe we should restructure the whole thing to use multiple IPs per server, i.e. instanciate one server per IP.  Or so.  Not really easy.

cheers
  simon