[Bug 1001656] FreeBSD: add AF_PACKET socket familiy

[bugzilla-daemon@bugs.ecos.sourceware.org]

Please do not reply to this email, use the link below.

http://bugs.ecos.sourceware.org/show_bug.cgi?id=1001656

--- Comment #14 from Bernd Edlinger <bernd.edlinger@hotmail.de> ---
(In reply to comment #13)
> After the first quick test: Yes, it looks like the IPv6 UDP packet is now
> sent.

Good.

> However, I am wondering if it's really the right thing to stuff
> padding bytes into sockaddr_in6. I am not familiar with the whole networking
> code stuff, but I saw that in the Windows world and partly in various unix
> flavors 28 Bytes for sockaddr_in6 or even little less seem common.
> Is there
> any RFC, ISO or POSIX standard where it's said that these structures should
> have the same size?

probably not. The goal is to be compatible to other O/S where possible.
However for eCos an exact binary compatiblity is not required.
Even the defines for AF_INET and AF_INET6 may be different from what they
are on linux.

For whatever reason eCos chose to enlarge struct sockaddr from 16 bytes
(as it is on linux) to 32 bytes. I always assumed that was a smart move,
which was done to make struct sockaddr large enough to hold a sockaddr_in6,
but I never checked that, and I did not much testing with IPv6 either...

Well, struct sockaddr is the base-class of all socket types,
and therefore all other socket types should be at least as large as 32 bytes.

As a consequence sockaddr_in has a padding enlarged from 8 (as it is on linux)
enlarged to 24. But sockaddr_in6 should have done the same.

> sidenote: Why was there even a 'len'-parameter in bsd_bind() when it was not
> used up until your patch?

Definitely a BUG.

Fact is:
The code in bsd_bind does _only_ work under the assumption that
len == sizeof(sockaddr) otherwise you access beyond the end of *sa
or you only copy a part of the underlying structure.

I did not want to add more complexity here, like using dynamic allocations,
because with AF_INET/AF_INET6/AF_PACKET I can make sure that all socket
structures are exactly the same size.

On the other hand, on linux nobody sets "sa_len" but uses the len
parameter instead to tell bind() how long the structure is.
So I figured it would be better to use the len parameter to set
that member, because inside the network stack, there is no "len",
but always sa_len.

So, my intention was just to improve compatibilty to linux here,
and still have the right data in sa_len for the network stack.


Regards
Bernd.

-- 
You are receiving this mail because:
You are on the CC list for the bug.