[Olsr-dev] mtu problem with multicast ipv6 packets

[John Hay]

On Mon, Oct 20, 2008 at 11:53:34AM +0200, Bernd Petrovitsch wrote:
> On Sun, 2008-10-19 at 23:22 +0200, John Hay wrote:
> > On Sun, Oct 19, 2008 at 10:51:02PM +0200, Bernd Petrovitsch wrote:
> [...]
> > > On Son, 2008-10-19 at 21:30 +0200, John Hay wrote:
> > > [...]
> > > > So here is my patch to clip the mtu in the IPv6 case. I have hardcoded
> > > > OLSR_IPV6_MTU, but at least on the BSDs it is defined in netinet/ip6.h as
> > > > IPV6_MMTU. It did not look as if ip6.h is pulled in at the moment, so I did
> > > > not know if we should start now. :-)
> > > 
> > > Why not?
> > 
> > Well looking through interfaces.h and defs.h, it seemed to hardcode
> > some of the network related values, so I was not sure which way to
> 
> Hmm, these were there before my time. Actually I never thought about
> them before.
> 
> > go. Also I wasn't sure if it was a standard define in a standard
> > place. I found quite a few BSD references, but not much else.
> > 
> > > Is there any reason not to use a #define in a system header and hardcode
> > > magic values in some application?
> > > 
> > > Hmm, on my Fedora-9, I can find
> > > ----  snip  ----
> > > /* The latest drafts declared increase in minimal mtu up to 1280. */
> > > 
> > > #define IPV6_MIN_MTU	1280
> > > ----  snip  ----
> > > in <linux/ipv6.h> which seems to be the equivalent value.
> > 
> > Ok, so linux use a different define in a different file, OSX will
> > probably be like the BSDs, so that leaves Windows... What else is
> > there to worry about?
> 
> Nothing - if no one complains, it works or no one uses it;-)
> 
> I tend to judge compile time breakage as quite harmless as they should
> be easy to fix.
> I just did `find /usr/include -name '*.h' | xargs fgrep -w 1280` - after
> IPV6_MMTU didn't show up - and the above turned up.
> 
> As a general strategy (nice word;-):
> I would take already #define'd constants instead of adding the n+1
> #define for them.
> That adds the problem of "different names for the same thing in
> different OSs". Personally I would use the more descriptive/usual/easier
> to understand in the source and add a compatibility #define for the
> others.
> So for the current case: Use IPV6_MIN_MTU everywhere and add a
> 	#define IPV6_MIN_MTU IPV6_MMTU
> under a appropriate #ifdef[1].

That sounds ok.

> If the (olsrd-)local interpretation is some what different (e.g. that
> value is the actual always used MTU and not the minimum possible):
> Use a new name and have #if ... #else ... #endif

I think that the idea is that for things like multicast, you should
fall back to the minimum that is guaranteed. In the IPv6 world, that
is 1280 bytes. For normal packets, path mtu discovery will take care
of mtu mismatches.

> I see no problem if there is one (or more) header file which hides such
> compatibility issues (and I actually wonder why people can't agree on
> the same #define'd name for the same thing).
> 
> For the current case:
> But reading the patch once again (and believing that I understand;-):
> Perhaps renaming OLSR_DEFAULT_MTU to OLSR_IPV4_DEFAULT_MTU (or
> IPV4_DEFAULT_MTU[0]) and adding OLSR_IPV6_DEFAULT_MTU is the most clear
> way to go.
> Or did I miss something?

Or OLSR_IPV6_MULTICAST_MTU, but DEFAULT is keeping it closer to the
ipv4 equivalent. :-)

> 
> 	Bernd
> 
> [0]: this olsr_ prefix feels somewhat religious.

That is probably just to make sure that there is no collision in the
name space.

> [1]: IPV6_MIN_MTU is IMHO clearer than IPV6_MMTU.

I agree with this.

John
-- 
John Hay -- (spam-protected) / (spam-protected)