Hi Denis, Marcel,
I'll respond just once, since you both commented on pretty much the same topics.
> > I notice that your version of the patches are not forming
> the IPv6 link-local address and configuring it on the network
> interface. That's ok, as long as connman takes care of it,
> but it does mean that Ethernet interfaces, which always have
> a link-local address, will autoconfigure immediately while
> point-to-point interfaces will only autoconfigure when
> connman sets the link-local address on them. We talked about
> this with Marcel and at the time concluded that it would make
> more sense to keep things consistent by having oFono
> configure the link-local address on the point-to-point
> interfaces. I had this implemented in my later patch sets.
[Denis]
> Samuel felt strongly that this should be done by ConnMan.
> There's still
> some questions left around this area. Namely whether we
> should prevent
> auto-configuration or not, etc.
Disabling IPv6 stateless address would be a bold move indeed, since it is
declared mandatory in both IETF and 3GPP standards. Please see [RFC4294] and
[3GPP 23.060]. The section "Dynamic IPv6 Address Allocation" in 23.060 is very
clear on how IPv6 address allocation in 3GPPP networks is done. See also the
section about IPv6 prefix delegation (relevant to IPv6 tethering) and
applicable parts of 24.008.
> > Having separate lists for IPv4 and IPv6 DNS servers made
> sense in my first patch set, because a dual context could be
> emulated with separate IPv4 and IPv6 contexts and those DNS
> servers might have been behind different network interfaces.
> However, now this just creates additional complexity for the
> drivers. A dual context will get a list of DNS server
> addresses, which may contain IPv4 addresses, IPv6 addresses
> or both. Now the driver has to sort them into two separate
> lists for IPv4 and IPv6. Note that you can make A and AAAA
> queries to any server, so there is no particular reason to
> separate the lists based on address family.
[Denis]
> We waffled on this one, but it seemed better to keep it symmetric with
> the way ConnMan API is setup for IPv6. For AT modems that support
> dual-stack, the IPv4 and IPv6 information is returned
> separately anyway,
> so for the majority of the drivers separating the details is
> more efficient.
Sounds like connman has some unnecessary complexity here as well. A single DNS
server list really should be sufficient. Not a big deal, though, as long as
connman combines the lists at some point and gives the whole thing to the DNS
resolver.
> > void ofono_gprs_context_set_ipv6_prefix_length(struct
> ofono_gprs_context *gc,
> > unsigned
> char length);
> > void ofono_gprs_context_set_ipv6_gateway(struct
> ofono_gprs_context *gc,
> > const char
> *gateway);
> >
> > I'm not sure these are really needed, which is why I
> dropped these from subsequent patches. This information is
> not received from the cellular network as part of context
> activation signalling. On-link prefixes, routes and default
> gateways are received as part of the standard IPv6 stateless
> address autoconfiguration when the interfaces is brought up.
> The only reason to have these would if a specific modem with
> a virtual ethernet interface deviates from the standard
> address configuration practises for some reasons.
>
[Denis]
> At least the gateway is reported separately for IPv4 and
> IPv6. Refer to
> 27.007 Section 10.1.23 "PDP Context Read Dynamic Parameters".
[Marcel]
> The AT commands are pretty clear in that they give us actually a
> netmask.
True at first glance. However, please note the following:
1) The netmask and gateway values returned here are invented by the modem. They
do not come from the network, because they are not carried anywhere in the
signalling messages related to PDP context activation. Stateless IPv6 address
autoconfiguration (potentially followed by DHCPv6) is the only way to get the
real information from the network.
2) AT+CGPADDR and AT+CGCONTRDP were introduced in 3GPP rel8 and they are both
optional. I.e. pre-rel8 equipment does not have them and rel8+ equipment is not
required to have them. In particular, they are not required for a modem to be
able to support IPv6, so you cannot rely on these commands being implemented in
the modem.
3) IPv6 stateless address autoconfiguration is mandatory in 3GPP standards, as
well as in IETF standards. There's no two ways about it.
Granted, AT+CGPADDR and AT+GCCONTRDP may be useful for supporting some quirky
modems, as already seen with IPv4 and virtual ethernet interfaces. Lacking a
concrete example for such a modem I don't see much point in adding these
methods yet, though.
[Denis]
> For IPv6 we decided to go with prefix length to keep symmetry with
> connman, even though 27.007 reports a netmask. Drivers will need to
> convert between IPv6 netmask and prefix length, so this might
> have to be
> addressed in the future.
Actually, the latest versions of 27.007 introduce the AT+CGPIAF command, which
allows the use of proper textual representations for IPv6 addresses and
prefixes. Looks like it is finally dawning on the AT spec writers that IPv6 is
actually very different from IPv4. ;)
[Denis]
> However, you might be right and that 27.007 is not a good
> reflection of
> reality. This has happened before ;)
It is not. 27.007 is just an interface specification for modem vendords. It
does not specify the behaviour and protocols between a UE and the network. You
really need to look at 23.060 and 24.008. It will save you some hassle in the
future.
[Marcel]
> Our success rate with testing IPv6 on a real network with current
> hardware was rather limited.
That is obvious.
Regards,
MikaL
_______________________________________________
ofono mailing list
[email protected]
http://lists.ofono.org/listinfo/ofono