Thanks for the great responses! I've cobbled them together to follow
a train of thought, interspersing them below:
> > 1) Alternative Implementations
> > ------------------------------
> > Invariably, when I point out the above issue to developers, they
> > react with surprise. It is not common knowledge that Ada95 lacks this
> > guarantee, and the bugs it can manifest are elusive, since dynamic
> > reconfiguration is typically a massive operation (akin to startup) with
> > a combinatorial explosion that consequently sees little testing.
> I don't think I understand exactly what you want to do. There seem
> to be several solutions with minimal impact on existing code. It seems
> unlikely the Ada rules will change any time soon, especially since
> they appear to have a very reasonable and understandable justification.
My thrust in this dialogue is an Ada advocacy issue. I'm not seeking to
clarify a language wart or suggest workarounds, but rather to expose the
perspective of Ada developers: it's a fact of life that the latter are
never totally conversant with the ARM. Also, I'm not as focused on the
possible implementations (below), per se, as on how the language and
the public literature -mislead- those developers into subtle trouble,
as evidenced by the real-world implementations that I see. So, for an
"alternative implementation", I'm looking for an approach that is not
driven -merely- by avoidance of this language limitation (assuming
the developers are even aware of it), but rather hoping for one that
seems as "obvious and intuitive" to those developers as the flawed
> Do you ever have a case of multiple tasks registering the same callback
> routine? If so, how does one unregister its callback as opposed to
> its sibling's callback?
A good question! I haven't seen any code that specifically worries
about that -- a real issue nonetheless. The more likely paranoia is
that during a complex series of reconfigurations something will fall
down the crack, and a given subscriber will re-subscribe having not
unsubscribed first, thus ending up on the callback list more than
once. This paranoia gives rise to code that compares the subscribing
callback to those already on the list, in order to avoid this situation.
> > a) single 'Access: The "proper" thing to do is a technical tweak, as
> > described at <http://www.adaic.org/docs/95style/html/sec_7/7-3-2.html>
> > Define a single 'Access constant along with each callback subprogram
> > declaration, and use that constant for all references. This isn't a
> > bad solution, but it presumes the subscribers know how the publisher
> > is implemented (or will evolve). Maybe such a rule should be added to
> > program coding standards -- as a cliche.
> I don't follow this objection. All the subscriber needs to know is the
> profile of the callback; what else can change that affects the
> declaration of the single 'access constant?
The issue is in expecting developers to -know- they should use the
single 'Access constant. Typically, the publisher side of the code is
developed separately (by a different group or a third party) and all
that the subscriber developers have available is the publisher's spec --
as is appropriate per Ada's contract model. The latter folks should not
need to know how the "unsubscribe" operation has been implemented (or
will evolve), and in particular should not be expected to -compensate-
for its (possible) misuse of an Ada construct. This is the -only-
reason that solution (a) is needed at all: it is nothing but pure and
unexpected compensation that impacts -every- subscriber.
> The best solution to your problem is a coding standard requiring a
> single 'Access for each subprogram in your program. To be effective,
> this would have to be checked with a tool. As a practical matter, you
> won't see any wrappers with library-level subprograms that aren't in
> generic units, so the pattern usually works. (Which may be the most
> dangerous thing in this case.)
This should probably be done regardless, and may indeed be the "best
solution". It would help by raising awareness of this arcane language
limitation, and by addressing it as a cliche: ie, one should code
this way all the time, whether the workaround is needed or not. Thus,
each callback subprogram declaration would be paired with a (single)
'Access constant, imitating the cliche of declaring a class-wide access
type for each tagged type.
> > b) use 'Address: The System.Address of the callback could be used
> > instead of its 'Access value. This would require a system service for
> > performing the call (usually provided by the OS), but it defeats the
> > type checking that 'Access enjoys. And besides, it's not clear that
> > 'Address has any more guarantee for matching than 'Access does.
No one commented on this, but I would infer from Randy's lengthy
explanation (thanks!) that 'Address suffers from the same limitations
as 'Access (and if not, it should!) and hence is not an alternative.
> > c) subscriber id: This approach assigns a unique identifier to each
> > subscriber and uses that instead of the 'Access value for matching
> > in the callback list. This adds complexity and maintenance concerns,
> > and resists the introduction of a new subscriber into the system
> > -- although I've heard it argued that that is a -good- thing: only
> > "authorized" subscribers will be able to subscribe, by virtue of having
> > been assigned an id.
> How about the "laundry ticket" method
> My_Registration := Register(Callback'access);
> followed later by
> You could even have the laundry ticket My_Registration be a controlled
> type so it could be guaranteed to Unregister itself when it goes away.
This is version of solution (c). Basically, the issue is that developers
have a two-fold expectation of a subprogram pointer, namely, to support
indirect invocation and also subprogram identification. Since Ada does
not guarantee the latter, solutions of this kind decouple the two.
> This is indeed a better design. It is possible to have a fast
> implementation to look for the ID (AVL tree for example). More over
> it is then possible to have a "readable" log file. For each use of
> the interface you can log which client is calling.
Ideally, it should be requirements such as those that would drive the
developers to an approach like this one -- not mere avoidance of a
language wart. But lacking requirements beyond those in my original
email, what justifies this approach over the flawed one (other than
the flaw itself)? Do we see such an example in the public literature?
> > d) sockets: A variation of this is that every possible subscriber is
> > allocated a "socket" and the socket table replaces the callback list:
> > no search (involving a match) is needed, since each subscriber has its
> > own socket. The table is statically maximal, the equivalent of every
> > subscriber having been added at system startup. Dynamic reconfiguration
> > is achieved via a boolean flag in each socket, indicating whether a
> > given subscriber participates in the current configuration (ie, whether
> > or not the event handler should call that subscriber's callback). This
> > may be efficient but can consume a great deal of memory, much of it
> > wasted on non-configured subscribers.
No comments on this one. It avoids the identification issue altogether.
> > e) dispatching: This is a more sophisticated approach, wherein the
> > callbacks are mediated via tagged type dispatching rather than by
> > 'Access subprogram values.
> Every publish-subscribe interface I have ever seen in OO languages
> used polymorphism. Function pointers are used in languages that
> don't directly support polymorphism. IMHO, the only thing access to
> subprograms is good for is interfacing to C (or other low-level) code;
> dispatching handles this type of stuff more cleanly. For instance,
> you can carry data in the dispatching object, which is quite handy in
> most instances.
Same question as above: what requirements justify this approach over
the flawed one (other than the flaw itself)? Said another way: assuming
the developers are ignorant of the language limitation, what line of
reasoning would (seredipitously) lead them to this solution rather
than to the unexpectedly flawed one? BTW, for the implementations I
have seen, the callback profile includes an input parameter for the
same data that the dispatching object would provide.
> But the *really* best solution is an OOP-based solution. This
> is extensible and type-safe and doesn't require unsafe compares and
> doesn't even need (visible) access types. Admittedly, there is a bit
> more work up front, but I think access types should avoided unless there
> is really dynamic allocation (which can't happen with subprograms in
"Extensibility" is a requirement that could argue for this approach.
However in my discussions with developers, it is very difficult to pin
down just what constitutes extensibility: it seems to be a matter of
taste... and I've heard arguments that the flawed solution does meet
extensibility requirements. But it's the "bit more work up front",
lacking some other motivation, that leads developers to -not- choose
this solution. In this dialogue here, I'm hoping for an "obvious and
intuitive" reason that developers who are ignorant of the language
limitation would choose an unflawed solution.
> Of course this is the design to use with Ada95, as Corey said, it is
> the implementation used in OO languages, it works just fine and does
> not suffer the 'Access issue you have raised.
I'd sure like to see it in the Ada literature! Lacking that, is there
a web site that provides a good example?
> > ...and there are certainly other designs.
> In this example, the event handler can provide an entry on which tasks
> may block, being released when the event occurs. Each "subscriber"
> can provide a waiting task that waits on the event, then invokes
> whatever action the specific "subscriber" requires. A "subscriber"
> can "unsubscribe" by having the waiting task terminate the entry
> call prematurely.
I've heard this approach discussed, but have never actually seen it. In
embedded real-time systems, the timing requirements tend to override
the concurrency ones, and rendezvous is avoided in favor of callback
invocation. I neglected to list this approach in my original email.
> If you have piles of existing code containing
> you could make the Register procedure look at the code at location
> Callback'access and, if it's a short wrapper, simulate it to find the
> real routine. You would need some information on just what sort of
> code your compiler generates, of course.
Wow, I never would have thought of that one! At the very least, this
approach is arcane and depends on both the compiler and its version.
Not "obvious and intuitive": I'd be glad if the 'Access comparison
issue itself were more widely known, let alone this kind of knowledge.
> > 2) Language Revision
> > --------------------
> > This issue naturally leads to the question: can the next language
> > revision for Ada provide the missing guarantee, assuring a match
> > in comparisons of subprogram access values that designate the same
> > subprogram?
> Perhaps we could make pragma Export (Ada, ...) do what you want?
> (pragma Import has additional semantics for convention Ada).
Ideally, a language revision would simply hard-guarantee subprogram
'Access comparisons -- without making them optional. As in solution
(a), a pragma approach would require subscriber developers to -know-
the publisher's need for applying the pragma in every subscriber.
> IMHO, the only thing access to
> subprograms is good for is interfacing to C (or other low-level) code;
Well, eliminating subprogram 'Access comparisons altogether would
be a step backwards. It's a desirable feature to keep -- if it were
> Certainly, we could get rid of the rule by disallowing the use of
> 'Access on subprograms that come from generic instances, but that
> is a far worse wart on the language than the current rule. We want
> subprograms in generics to be as 'normal' as possible.
And -partially- eliminating subprogram 'Access comparisons would be a
(smaller) step backwards. What about a compiler implementation that
included a value (address or otherwise) with every wrapper, whose
sole purpose is to provide the "identification" part of the two-fold
developer expectation of 'Access?
> If you are using GNAT compiler then you may consider the Code_Address
> attribute instead of Access (see Implementation Defined Attributes in
> GNAT Reference Manual).
Since I don't have access to the Reference Manual: does 'Code_Address
meet the two-fold developer expectation, namely: provide the desired
comparison guarantee, as well as invocation of the subprogram thus
designated? If so, could it be standardized? At the very least, it
is an existence proof that the language limitation -could- be removed.
C. Daniel Cooper ==========v=================v=======================v
Adv Computing Technologist | processes | All opinions are mine |
206-655-3519 | + architectures | and may not represent |
[log in to unmask] | = systems | those of my employer. |