Re: [PATCH net-next 0/6] A step closer to RFC 6458 compliancy

[Michael Tuexen <Michael.Tuexen@lurchi.franken.de>]

On 18 Jun 2014, at 15:25, Vlad Yasevich <vyasevich@gmail.com> wrote:

> On 06/18/2014 09:16 AM, David Laight wrote:
>> From: Michael Tuexen [ 
>>> On 18 Jun 2014, at 10:42, David Laight <David.Laight@ACULAB.COM> wrote:
>>> 
>>>> From: Vlad Yasevich
>>>>> On 06/17/2014 11:36 AM, David Laight wrote:
>>>>>> From: Of Geir Ola Vaagland
>>>>>>> These patches are part of my master thesis project. I have been searching for discrepancies
>>> between
>>>>>>> the socket API specificiation in RFC 6458 and the current Linux SCTP implementation. The
>>> following
>>>>>>> patches are my humble attempts at getting somewhat closer to compliancy.
>>>>>> 
>>>>>> I've just been reading RFC 6458 - HTF did it get past the editors and
>>>>>> then published in its current form?
>>>>>> Lots of the structures have implied padding.
>>>> ...
>>>>> I've argued the padding issue, but the editor stance is that it's implementation
>>>>> dependent.
>>>> 
>>>> It wouldn't be as bad if the RFC said that the structure contained the
>>>> fields that followed (as is typical of the posix definitions),
>>>> but instead it gives a definition of the structure.
>> 
>>> That would have been a possibility, but it was never suggested.
>>> As far as I know, C does not guarantee the memory layout for structs,
>>> except for the sequence of the components. So a compiler might add
>>> some padding at any place. When implementing this, you need to take
>>> care of this (and your job might be simpler, since you might only
>>> work with a specific set of compilers).
>>> In FreeBSD we also added some padding to some structures since they
>>> "evolved" during the lifetime of of the internet draft and we wanted
>>> to preserve some compatibility.
>>> I agree, that one must take care of the implied padding and I will double
>>> check how this is handled in FreeBSD. Not sure...
>> 
>> You need to add explicit named pad fields in order to zero them.
>> (since you don't really want a memset())
>> That is against my reading of the RFC.
>> 
>> What does FreeBSD do about the 'sockaddr_storage'?
>> I'd have thought it had the same rules as NetBSD - where (IIRC) it should never
>> be instantiated, but only exists as a pointer type for function parameters.
>> 
> 
> I don't remember any such rules when sockaddr_storage was defined.  Can you
> point to any document stating such rules?
> It is definitely useful as a container object at times.
I agree. The description would be OK for struct sockaddr. I have never instantiated
a variable of that type. The only use (I know of) of struct sockaddr_storage
is as a container object.

Best regards
Michael
> 
> -vlad
> 
>