Document change

[ACE_TAO.git] / TAO / docs / DIFFERENCES

@(#)DIFFERENCES 1.1 95/09/13

Note -- this documentation was distributed with the original SunSoft
IIOP reference implementation and has very little relevance to what's
in TAO.  We include it here only for "historical" reference, i.e., you
should disregard most of what's in this document.

        Doug Schmidt 

----------------------------------------

[ NOTE that the CORBA 2.0 specifications have now been published, but are
not yet sufficiently available that this document can usefully just list
any differences between that specification and this software.  As a rule,
there are no differences between this software and CORBA 2.0 in areas
where CORBA 2.0 has specified an interface. ]


This document summarizes known differences between what is implemented in
this software and currently available OMG specifications.  As a rule, these
establish compatibility with revised specifications that are currently being
prepared for publication. However, in some cases the differences resolve
problems that are currently being addressed by ORB taskforces.  When those
revised CORBA specifications are published, this document will be updated to
reflect any remaining differences from them.

The reason for those changes is that different specifications were adopted
at the same time and there were in some cases subtle differences which need
to be resolved.  Minor errata have also been fixed.  None of these changes
are substantial, with the exception of the TypeCode interpreter API which
was added to the C++ mapping.

Also, note that the goal of this software is fidelity to the IIOP, so that
issues relating (in particular) to the C++ language mapping or other OMG
specifications were judged to be of less importance at this time.


DIFFERENCES from IIOP in UNO Spec (95-3-10)
-------------------------------------------
Some changes to the original specification (94-9-32) for the IIOP protocol
were found to be necessary; most related to the subsequently adopted Interface
Repository (IFR) specification (94-11-7).  Others resulted from correction
of minor editorial errors, and experience gained as multiple teams implement
to that specification.

The bulk of those changes have been incorporated into the 95-3-10 document,
but these few have not.


CDR TYPECODE INDIRECTION

An additional constraint on the use of typecode indirection has been
identified.  Specifically, the typecode to which an indirection points
be "self-sufficient", and may not point "outside of itself" for any
further indirections.  For example, always indirections occur within
encapsulations, so it's OK for the indirection to point anywhere in
that encapsulation, or at the encoded TCKind enum value immediately
before the encapsulation's length.  For typecode encapsulations nested
inside other typecode encapsulations, the indirection may point no
further than the outermost encapsulation.

Also, when an indirection points to another typecode, the byte order
of that other typecode must be deduced from the encoded value found
there.  This is straightforward for encoded TCKind values, all of which
fit into a single byte:  if the first byte of the word is zero, then
the encoding is big-endian, else it's little-endian.

Indirecting to another indirection is not allowed, since the byte order
of the encoded offset can't consistently be deduced.


MISCELLANY

The type GIOP::Version is assumed to be identical to IIOP::Version.
As part of the editorial separation of GIOP from its Internet version
IIOP, this typedef was accidentally omitted.

This implementation of IIOP supports the OMG-IDL Primitive Data Types as
defined in Appendix A of 94-9-32, but with TCKind numbers following the new
tk_exception and tk_alias codes.  (That is, rather than using range 21-25,
these TCKind values are in the range 23-27.)

These data types are currently the topic of an RFP in process in the OMG,
and are not currently defined as part of a current OMG specification. 
Clients and servers using these extended data types are relying on CORBA
extensions that are not yet standardized; and should be prepared to change.


CORBA 2.0/CORE Differences
--------------------------
As of this writing, the new CORBA 2.0/CORE document has not yet been made
available.  This is a modification of the CORBA 1.2 document, with the
addition of (mandatory) extensions from the UNO (94-9-32) specification:
notably DSI and some new object reference operations.  This summarizes
differences between the as-yet-unpublished CORBA 2.0/CORE document and
the original text in the UNO specification.


DYNAMIC SKELETON INTERFACE

Since the new "tk_except" typecodes now include the exception ID (this is
the repository ID describing the exception's type) the "exception" operation
used to report an exception through a ServerRequest no longer includes the
exception ID.

Since the OperationDef entries for attributes are no longer found in the
interface repositories, the ServerRequest convenience operation to return
this has been removed.  Application programs (such as bridges) must compare
the operation name string with the prefixes "_get_" and "_set_" to see if
any given operation applies to an attribute.  If an operation is one of the
two for which an OMG-IDL "attribute" declaration is shorthand, then the
application must search the interface repository for that attribute.

(There are in fact three categories of operations on which an object adapter
and implementation collaborate to handle:  user defined operations, user
defined attributes, and operations defined by CORBA such as "is_a" and
"get_implementation".  Some object adapters, like the BOA, handle this last
category directly, rather than exposing it to applications.)

The "non_existent" objref operation accidentally used attribute syntax; this
has been corrected.  All operations on CORBA::Object references now use
normal operation syntax in their (pseudo) IDL definitions, and are uniformly
prefixed with underscores as part of their C++ mapping.


DIFFERENCES from IDL C++ Mapping (94-9-14)
------------------------------------------
There are four basic ways in which the IIOP framework does not comply
with the OMG's IDL C++ mapping specification, beyond the use of the CORBA2
module/namespace rather than the CORBA module/namespace.  (CORBA2 is used
so that this software can be linked with existing ORBs, which should be
using the CORBA module/namespace already.)

These ways are beyond the portability-derived use of the "class" construct
(instead of the C++ namespace construct), and the use of CORBA2::Environment
(rather than C++ exceptions), both of which are accomodated by the mapping:

    (1) This framework does not attempt to be complete.
    
        Completeness is not required to implement the Internet IOP,
        so time was not spent in providing a complete mapping.

    (2) In some cases the API provided is not the one in the mapping.
        For example CORBA2::Environment is not always passed in all
        operations.  This was done to promote ease of implementation
        and coding.

    (3) Implementation details are generally exposed.

        This isn't so much a noncompliance issue (it's not specified
        that such details must be hidden) as an issue of allowing
        noncompliant applications.  The effort required to actively
        prevent use of implementation details by higher level code
        would be nontrivial, and could slow down IIOP code, so no
        time was applied to hiding such details.

        (However, note that when the software was modified to use COM,
        standard C++ techniques were used to hide most such details.)

    (4) The C++ mapping needs completion in some areas.

        Notably, 94-9-14 does not satisfy requirements to provide
        support for for all data types in an "Any" or exception without
        needing any precompiled data type support routines.  The
        TypeCode interpreter addresses this issue.

        Some additional operations were not adopted at the time 94-9-14
        was defined; both 94-9-32 and 94-11-7 added new ORB interfaces
        which are not found in the initial C++ mapping document.

With respect to this fourth issue, descriptions of the API extensions
used is provided later in this file.  The additional CORBA operations
are not described since their mapping is straightforward if it isn't given
in those documents.


ACCESS TO ALL DATA IN "ANY" AND EXCEPTIONS

The "void *" value in an "Any", and any "CORBA2::Exception *" value, may
always be passed to the TypeCode interpreter (see next).  This allows access
to all data held within an "Any" and an Exception.  The "void *" value in an
Any with an exception TypeCode is of type CORBA2::Exception *".  Due to
the way CORBA is specified, it is not possible to tell whether an "Any"
holding an exception denotes a user or standard (system-defined) exception
except by exhaustive comparison against exception IDs.

Also, _all_ legal OMG-IDL data types may be held in an "Any", and may be
manipulated using the TypeCode interpreter.  This satisfies a language
mapping requirement that DII (and DSI) be able to manipulate all OMG-IDL
data types without requiring compiler generated support for them.


TYPECODE INTERPRETER

In any given C/C++ execution environment there is a binary standard for how
data is represented.  While that standard will differ between execution
environments, there are a very limited number of ways in which those
representations differ.  Those differences are encapsulated only in the
TypeCode interpreter, so that portable applications using the dynamic
typing infrastructure in the ORB can completely ignore how it's done in
any particular environment.

The APIs in this software distribution have been suggested for adoption as
part of the OMG specification suite.  The two APIs are nonstatic member
functions in the C++ mapping's TypeCode pseudo-object, and are augmented
by a single new enumeration type.

    unsigned long       size() 

        This returns the size of an instance of the type that is
        described by the TypeCode.
        
        For example, when invoked on the typecode constant CORBA::_tc_Short,
        the value returned is sizeof(CORBA::Short); and when invoked on the
        typecode for a structure, it is the size of that structure (including
        any internal and tail padding needed).  When invoked on a sequence
        typecode, it returns a value that does not include the size for
        any embedded buffer.
    
    enum traverse_status {TRAVERSE_STOP, TRAVERSE_CONTINUE };
        
        This is a data type used in the traverse() member function.
        It allows data type traversal to be terminated early for
        non-exceptional conditions, and eliminates the confusion
        that some similar APIs have created when they use a single
        boolean value (does TRUE mean to stop, or to continue?).

    traverse_status     traverse (
                            const void  *value1;
                            const void  *value2
                            traverse_status     visit (
                                    TypeCode_ptr        tc,
                                    const void          *visit_value1,
                                    const void          *visit_value2,
                                    void                *visit_context
                                ),
                            void        *context
                        );

        (In the current language mapping, CORBA2::Environment references
        are passed as the final parameter to the 'traverse' and 'visit'
        routines for use when reporting exceptions.)

        The pointers "value1" and "value2" point to instances of the
        data type described by the typecode (or are null pointers).

        For each constituent of that data type (e.g. structure member)
        the visit() routine is called once.   The constituent's type
        is described by "tc"; "visit_value1" points to the constituent
        of that type in "value1" (assuming the traverse routine was
        not passed a null pointer) and similarly for "visit_value2".
        The "visit_context" parameter is the context parameter passed
        to the traverse() routine, and can point to whatever data is
        needed by the visit() routine.

        Members are traversed in first-to-last order, as defined in the
        IDL specification for the data type.  So for example, the
        visit routine for a structure could print out each element
        on a IO stream passed through the context parameter, perhaps in
        a neatly formatted form intended for human consumption.

        The visit() function may choose to recursively traverse() each
        element.  So for example if "tc->kind()" in a visit routine
        returned CORBA::tk_struct, the contents of that struct would
        be ignored unless the visit routine invoked traverse() using
        the typecode and values passed to it.

        If the visit() routine returns TRAVERSE_CONTINUE, succeeding
        constituents of the data type are visited in turn; if it
        returns TRAVERSE_STOP, the traverse() routine returns that
        value to its caller.  Of course, if a visit() routine calls
        traverse() itself, it may choose to ignore TRAVERSE_STOP.

The traverse() routine is a powerful tool.  It is used in the IIOP code
itself in several ways; look at such use for tutorial details.  You could
implement data value comparison and "debug" data dumping as simple exercises
in the use of these APIs.

The "marshaling interpreter" (marshal.cc) uses it to encode and decode
values according to the IIOP protocol specification.  The "Any" code
(any.cc) uses traverse() both to free embedded pointers, and to make "deep"
copies of any data structure given its TypeCode.

Only that "deep copy" visit routine uses "value2" as anything other than a
null pointer; it allows efficient construction of "deep copies" without
needing extra space for temporary values or coroutine stacks.  (A general
purpose two-value comparison could also use "value2".)  Most uses of the API
only manipulate a single data value at a time; no realistic need has yet
been seen for manipulating more than two data values at once.

With respect to the OMG C and C++ mappings, it is clear that this style API
must be provided for the C mapping, but some people have noted that a
"purer" object oriented style API could also be provided in C++.  That style
would use another internal visit routine, performing the requisite "switch"
over the fixed number of TCKind values, and then make a virtual function
call to an instance of a C++ class whose private state was the "context" and
whose member functions matched the arms of the switch.


MEMORY ALLOCATION

In order to dynamically manipulate instances of arbitrary data types,
applications need to be able to allocate and free memory.  The OMG C++
mapping only says how to do this for data types which have static C++
interfaces defined, which is clearly inadequate for using constructed
types with the DII/DSI/Any family of interfaces.

This infrastructure requires the standard "malloc" and "free" primitives to
be used, and cast to the appropriate type.  Data allocated using malloc
will normally be sized according to TypeCode::size(), and then be stored
inside an Any.  When the Any is deleted, if the ORB deallocates the memory
it always uses "free" (including for any nested pointers, and correctly
handling cases such as arrays and sequences).

Note that to support implementations where C and C++ language bindings
share the same ORB infrastructure, this solution is inadequate.  This is
because the C binding's CORBA_free() interface would have no way to
determine the type of the data being freed.  Instead, typed allocation
APIs will need to be used even when using the dynamically typed CORBA
subsystem ... a TypeCode::malloc() routine would suffice, if it returned
memory that was internally tagged with that TypeCode.  In such a case,
the CORBA_free() routine could use that TypeCode to "deep free" data as
required, and C++ "new" and "delete" operators would need to know about
the internal tagging for all those data types.  Such tagged allocation
would need to be used for all data that was to be freed by the ORB.

(Having the C mapping require use of "typed free" routines, instead of the
CORBA_free interface, is sufficient to eliminate this problem.)


PASSING EXCEPTIONS THROUGH THE DII

The C++ mapping's Dynamic Invocation Interface (DII) has key omissions
in that it doesn't say how to access to user-defined exception values,
and implicitly requires ORBs either to pass "excess" data on the wire
or else to consult an interface repository to deal with exceptions or
(deprecated) "context" strings.

This software changes the DII specification in two ways to support the
requirement for DII users to be able to see user-defined exceptions, yet
not to violate the OMG-IDL type model by reporting illegal exceptions:

    * The Object::_create_request() member function, through which DII
      users provide all the information included in an IDL operation's
      signature, has an additional parameter.

      That parameter is a sequence of exception typecodes, describing
      the user-defined exceptions that the operation can return.  (The
      standard exceptions defined in CORBA may always be returned.)
      When any other exception is returned, the client's ORB reports that
      the object's type definition has been violated by the server's ORB.

    * The TypeCode interpreter (see above) may be used to examine all
      exception values reported through a CORBA::Environment.  A new
      Exception::id() operation may be used to determine which typecode
      should be used.

Note that a number of portability and specification problems have been
identified in the current C++ mapping for DII, e.g. for details of memory
management.  Later versions of this code may attempt to closely comply
with an improved mapping for DII, to the extent that the interpretation
used here differs from that more unambiguous specification.

Since there is no efficient way to distinguish typecodes for user defined
exceptions from ones for system-defined ("standard") exceptions (comparing
the exception ID against all the system defined ones is inefficient :-)
a new enum type "ExceptionType" is defined.  (An analogous type is already
defined for the C language mapping.)  This is used to report exceptions
through DII and DSI.