[AFS3-std] Comments on draft-keiser-afs3-xdr-union-06

[Michael Meffie]

On Mon, 7 Apr 2014 14:47:58 -0400
Benjamin Kaduk <kaduk@MIT.EDU> wrote:

> On Sat, 5 Apr 2014, Simon Wilkinson wrote:
> 
> > Hi,
> >
> > As an extension to RFC4506 this document seems fine. One nit,
> > however is that our current rxgen doesn't implement the 4506 syntax
> > for unions.
> >
> > A RFC4506 style union definition is:
> >
> > union switch (DESC-TYPE DESC-NAME) {
> >    ...
> > } NAME;
> 
> If I am reading it correctly (e.g., section 4.18), this is the RFC4506 way 
> to declare and name a single instance of a union, at the location where it 
> appears...
> 
> > However, AFS's rxgen takes a union defintion in the form
> >
> > union NAME switch (DESC-TYPE DESC-NAME) {
> >    ...
> > };
> 
> ... whereas this would be the RFC 4506 syntax for typdef-ing the name 
> 'NAME' to refer to the union type defined here.
> 
> It looks like the OpenAFS codebase's .xg files do not use enum, only union
> and struct, but only in the "implicit typedef" form.  We do not always use 
> the typedef'd type name when putting structs in other structs, which 
> perhaps adds some confusion ("struct AFSFid netFid" and "AFSFid netFid" 
> would be equivalent as members of some other struct).
> 
> > For an extended union, this document defines
> >
> > ext-union switch (DESC-TYPE DESC-NAME) {
> >    ...
> > } NAME;
> 
> This draft prefixes the example as "Extensible discriminated unions are 
> defined in RPC-L as follows:", which could be interpreted to mean the 
> "declare-and-name-a-single-instance-where-it-appears" case, and the 
> implicit typedef case is, well, implicit.
> 
> > So we're consistent with RFC4506, but inconsistent with the syntax used 
> > in deployed AFS protocol definitions. I don't know to what extent this 
> > is an issue, but it would seem difficult if we're moving towards a style 
> > of XDR that can't be compiled by the current protocol compilers.
> 
> I'm not actually very familiar with either the OpenAFS rxgen or RC 4506 
> XDR, but it seems like maybe it is just the case that OpenAFS rxgen does 
> not handle unions in the non-implicit-typedef case?
> 
> Perhaps I am just confused.

I think you are correct. I had to read RFC4506 section 6.3 a couple
of times to be sure.

Unions are defined twice in RFC4506 section 6.3.

First there is the declaration variation (with the identifier last)

      declaration:
           type-specifier identifier
           | (etc...)

      type-specifier:
           "typedef" declaration ";"
           | (etc...)

      union-type-spec:
          "union" union-body

and then there is the definition variation (with the identifier in the
middle):

      type-def:
           "union" identifier union-body ";"
           | (etc..)

      definition:
           type-def
           | (etc...)


I see only one case of unions being used in the OpenAFS xg files, which
is the definition form ("union" identifier union-body).

    union ktc_tokenUnion switch (afs_int32 at_type) {
        case AFSTOKEN_UNION_KAD:
    	token_rxkad at_kad;
    };

Testing and reading the OpenAFS rxgen, it indeed only supports the
definition form (that is, OpenAFS rxgen expects the identifier to
follow the "union" token). 


-- 
Michael Meffie <mmeffie@sinenomine.net>