C++ Patterns: The Badge
awesomekling.github.io
awesomekling.github.io
This badge thing is clever, but seems like yet more accidental/unnecessary complexity on top of the already-probably-unnecessary private/protected/friend - as requirements change you now have this extra layer of access-control cruft to reason about, that you probably wouldn't even miss if it didn't exist.
Not to get too ranty, and no disrespect to Andreas, but the C++ culture seems to thrive in unnecessary / accidental complexity, deep template hacks and the like that add friction to refactoring, obfuscate underlying logic so you miss opportunities to simplify, and turn straightforward programming problems into brain-bending template puzzles.
OOP also; I've seen so many crazy class architectures that could just be a handful of plain-old-functions, complex virutal "type families" that could just be a TypeOfThingThisIs enum with some if statements in the respective functions, etc.
I don't know the Serenity codebase and use-case very well, and perhaps this kind of thing has a place in large libraries, but honestly even then I'd lean towards just making everything simple & public and use a prefix when needed, or describe appropriate usage in documentation. Seems to work in the python world (minus a few outlier codebases that go OOP-crazy).
It also doesn't work very well with code completion - implementation methods get mixed up with surface API - and usability by third parties suffers.
If you're on a team of one, and you own the project for its whole lifetime, then you're fine.
But in the case of say an open source library, how likely are clients going to update their code with each new release of the library? Typical case is to do it from time to time, especially with compiled code / shared libs, and expect a few incompatibilities to sort out and fix when you do.
In a fast evolving large team project, seems it depends on how well the boundaries of the subsystems are worked out. Misusing supposed-to-be-private members is one minor possibility amidst many challenges.
The main downside of Badge<T> is that it comes with a performance penalty and ABI change for something that should be statically deducible (empty struct is one byte in C++).
You can have a look here: https://godbolt.org/z/DmeYL-
The one byte rule does nothing here, because its value is undefined and its address not observable.
edit: if you find a case where an empty struct parameter causes code to be emitted, let's consider filing bugs.
also, tons of "generic" numerical code with templates everywhere, where the only two possible instantiations are "float" and "double". Some people do really like to add useless complexity everywhere.
You never really need to do that. In the rare cases you need to, once every five years or so, it is trivial to change your code.
Personally, I think that bashing/penalizing ignorance (and I mean the word ignorance literally, as being uninformed about a topic, no offense here whatsoever) is one of the lowest forms of abuse. It says much more about the downvoter than it does about the downvoted. And this not even considering the fact that searching online for "m_" isn't going to help you answer your own question.
If everybody just downvoted the questions they deem unworthy of an answer because "you are supposed to know this", there would be no questions answered.
Sorry for the OT.
Because complaining about coding conventions on a discussion about a design pattern is noise and adds nothing to the discussion.
And judging by other comments to the question, I wasn't alone, luckily.
For example, if we develop some framework to be used at multiple sites in the company, we can be 100% sure that at some point someone will assume that anything marked 'public' in the API is fair game to use (which IMO is a reasonable perspective). If they start using internal API's or depend on class internals that are only public just because the language did not provide us with a reasonable way to hide them, at some point we will change them and working systems could break elsewhere. In a similar fashion, misuse of private API's by violating preconditions that are completely opaque to users of your (intended) public API can lead to the worst kinds of bugs and unpredictable runtime failures.
Software development always seems so simple and pragmatic if you don't need to care about other people you have no control over.
... in your setup. All shops are different. To each their own. There's no private/protected scaffolding in the Linux kernel to give the most obvious example.
I don't know anything about kernel programming, but I assume that there's some very, very strict conventions about how the various parts of the Linux kernel interact, and nothing gets accepted for merging unless you follow them. This would be similar to imposing scope/visibility constraints at the language level, but considering C doesn't have anything for that, an extremely strict development process is the best you can do.
It's a user/m_ prefix vs compiler/private. Im not sure if former is better. It's a strong trade off imho.
In my experience relying on people reading code and agreeing on imlicit contracts does not scale beyond 5 people, but maybe I've been mistreated by life.
And that's only ever possible if you're the only "dude", and your codebase is like <10k loc
The only mechanism for safe-guarding access to private parts is using privately defined structs that are passed around in a form of opaque pointers. Some project use it (PGPphone did, the original SSH did, etc.), some don't, e.g. BSD and Linux kernels.
It's really more of a matter of developers exercising some basic thought when using other people's code. And it just happens so that there's more of them outside of the C++ group than in it.
If a class is a black box, when you put your hand inside it, it becomes the "pain box" from the Bene Gesserit.
No disrespect to you, QuadrupleA, do whatever makes you the most productive, but I wouldn't want to work together with you in this way :)
Anyway language rants aside, your project looks very cool - I remember seeing it posted here a while back.
Over a long enough time, every API will be called in every possible way. If an API misbehaves, some code somewhere will begin to rely on the specific kind of misbehavior. This is just the reality of publishing API's, so you deal with it and fix bugs as they are discovered.
If you make your entire class definition public in a large project, you will eventually have code poking at the internals of your class. I've seen this countless times, it's some kind of inevitable natural thing.
Developer scalability:
You know everything about how your code works, but I don't. A well-defined API with clear access rules gives me a fair chance to avoid using your class in ways you didn't intend.
In small-enough teams, or with small-enough projects, it's possible for everyone to understand (mostly) everything, but there's some magic number of concurrent developers where that privilege ends.
This is a recipe for disaster as soon as inheritance is in the picture.
This is the sort of accidental complexity that access specifiers eliminated, but somehow you managed to reintroduce by trying to reinvent the wheel.
The 'm_' prefix only means "private member variable, don'access it" to you and you alone, and anyone who will stumble on your code will not be aware or obligated to follow your personal style. That would not be the case.if you simply used a basic feature of C++ that's been available since the time of "C with classes".
To me, you're the epitome of the main source of C++'s problems: programmers who failed to learn the basics of a language and instead decided to develop convoluted hacks to circumvent proper use of the language because they believe their new discovery is clever although in reality they only managed to needlessly add complexity to a code base while adding brittle and bug-prone code.
What else? Do away with memory deallocation because there's nothing of the sort in Python and your C++ code is packed with segfaults?
And by the way, the pimpl pattern is used to provide libraries a stable ABI in spite of what change might be done to the implementation. It is not a fad or a clever-looking trick.
In my IDE in c++ I can quickly refactor m_ to non-m to make it "semantically public" without having to edit header files, change struct packing order / ABI, etc. And most of the time I don't even bother with m_. It's one less thing to think about. I honestly don't think access specifiers are worth wasting mental cpu cycles on in most projects.
Besides, in c++ people can always cast your intricately-protected object to a void pointer and start manipulating it nefarioisly anyway. Just document and communicate, and write simple straightforward code.
I used pimpl at the time for faster recompiles when changing "private" stuff. Better build tools got me around that problem personally, but I'm sure it has it's uses.
Python does not have access modifiers. C++ does. It makes absolutely no sense to write Python code in C++ while entirely oblivious to basic, age-old C++ features just because your background lies somewhere else and you failed or refused to learn even the basics.
And you know what actually works instead of just "working fine"? Having the compiler throw an error if a careless programmer tries to access private variables.
> In my IDE in c++ I can quickly refactor m_ to non-m to make it "semantically public" without having to edit header files, change struct packing order / ABI, etc.
Don't you understand that's completely irrelevant? Just set the private member as private and it does not matter at all which naming convention you follow. Don't you understand that having to do nothing is better than expecting team members to be prescient about your convoluted and absurd naming convention used to avoid best practices?
> Besides, in c++ people can always cast your intricately-protected object to a void pointer and start manipulating it nefarioisly anyway.
Don't you understand that't entirely irrelevant? You're trying to argue that it's ok to ignore best practices such as using member access specifyers, which actually get the compiler to validate the code, if you stick with naming conventions, but arguing that it's technically possible to circumvent compiler checks under some specific circumstances is an absurd statement. Think about it: if you feel that enforcing a coding style is enough to enforce private member access then don't you agree that the same code review that is supposed to enforce your personal style can quite easily catch your hypothetical casts?
I understand them well, and I've experimented with their use in my own projects, and decided they're a minor complexity I don't need to bother with. And that decision has served me well.
Dogmatic adherence to "best practices" is silly, even when people can agree on what they are. Learn them, yes, evaluate them carefully, but ultimately choose what works best for your constraints. Including what your team's conventions are.
The programmers I've seen taking most about "best practices" generally don't think very critically about their approaches and go with dogmatic answers, often tying up their codebases in ridiculous complexity. E.g. the "best practice" now to make a website is a React.js SPA, and now every site out there takes 20 seconds to load with 10 little spinner icons while all the HTTP requests go out to their microservices on docker clusters, and everything requires a hundred-library-deep build stack.
That’s just the cost of progress. Gotta break a few eggs to make an omelet. No pain, no gain. There are always casualties in war. Get with the times, grandpa.
This seems unnecessarily personal. See https://news.ycombinator.com/newsguidelines.html for the, um, best practices around these parts, which you are not following here.
I've spent most of my adult life working on large C++ codebases with public API's (most notably Qt and WebKit), which has led me to cultivate a defensive programming mindset. Someone mentioned Hyrum's law, which is 100% accurate in my experience.
I should be honest and admit that Badge<T> is not yet diligently applied everywhere in the Serenity codebase. I wanted to "feel it out" for a while first, but now I've decided I like it, so I'm applying it more and more going forward.
Oh and while I have your attention, I recently posted a May 2019 update video[1] on the Serenity OS project if you're curious how things are going :)
I'm actually curious why something like this has never been included as a language feature, it seems trivial to come up with a syntax for fine-grained access control, that is fully backwards-compatible and a true zero-cost abstraction.
Most obvious first shot:
class A
{
private:
void badgedFunction() const friend Device, SomeOtherClass;
};
I've only just decided that I enjoy using this pattern so my internal standards committee hadn't progressed to the proposal stage yet. :)
Your syntax makes perfect sense to me, and I think it would make a great addition to the language. I wonder what the real working group would say. (I tried quickly googling for something that might sound like an existing proposal but came up empty.)
Alternatively, there are other ways to leverage friendship and access control in C++. Here's one option:
template <typename Registry>
class DeviceManager;
class Device {
template<typename Registry> friend class DeviceManager;
int y;
};
class VFS {
friend struct DeviceManager<VFS>;
int x;
};
template <typename Registry>
struct DeviceManager {
static void register_device (Registry& registry, Device& device) {
registry.x = 1; // accessing privates
device.y = 2; // accessing privates
}
};
int main() {
VFS vfs;
Device dev;
DeviceManager<VFS>::register_device (vfs, dev);
}
This is also flexible. Remove 'static' and you get yourself a stateful mediator/manager. Add a variadic template register_devices_s_ member to Devicemanager and you've got yourself a convenience function for registering multiple devices (of heterogeneous types) without bloating the VFS API and, without runtime cost, and without losing type safety (by e.g. casting down to Device& from USBDevice&).
This doesn't seem so different from:
class A { private: friend void foo(A &, B &); int x; };
class B { private: friend void foo(A &, B &); int y; };
void foo(A &a, B &b)
{
a.x = 1;
b.y = 2;
}
int main() {
A a; B b;
foo(a, b);
}
> there is no restriction on who calls DeviceManager<VFS>::register_device
Well, you can make it private inside DeviceManager and give the class its own friends :-))
Another benefit of my solution is that adding things to DeviceManager<VFS> doesn't require changing the VFS.h or Device.h headers at all. Anything that links these 2 classes is an entirely separate concern and very much in line with thinking in terms of models or concepts rather than types. What you want at the end of the day is an absolute bare minimum of member functions in any class.
To be honest, whatever the solution, the key is that having the Device class self-register to the VFS in its constructor (as in Andreas's blog post) is, imho, an anti-pattern and is what is really leading to this Badge mess.
I'm not sure if it would work, cppreference has this to say about default arguments:
> The names used in the default arguments are looked up, checked for accessibility, and bound at the point of declaration, but are executed at the point of the function call
https://en.cppreference.com/w/cpp/language/default_arguments
Unfortunately it seems like default arguments can't use private constructors. Both GCC and Clang immediately fail at the declaration. Not fair, C++!
/home/andreas/src/serenity/Kernel/FileSystem/VirtualFileSystem.h:81:52: error: 'Badge<T>::Badge() [with T = Device]' is private within this context
Take a look: https://godbolt.org/z/DmeYL-
It's worth noting that it does apply to the latest gcc and Clang but not to all the compilers listed there. For an example of a difference, the latest djggp (7.2.0) inserts an extra push in the call when the empty argument is present.
AFAICT the C++ standard itself does not specify the particulars of calling convention, which is a function of the platform, compiler and flags/attributes.
There are an astonishing number of uses for default arguments that never appear at the call site.
My fav constructs an object that is just storage for a value that needs to live a little bit longer than the call itself.
I usually avoid 'friend' in these situations and just write
//private:
But I might would call the class Friend<T> or something.
A nice feature I've noticed in Elixir is that giving a function a "@doc false" annotation will actually prevent IDE/REPL autocomplete from completing the function. (It's still there to call if you type it yourself.) Maybe that's something C++ tooling could copy.
Any visible interface, public or private (via friend), will eventually be used by other programmers for other than its intended use case simply because it is there and accessible. For the maintainer of an interface with a single intended purpose, what should be a clean, tidy modification within the scope of its intended use ends up breaking all kinds of unrelated code that the maintainer wasn't even aware of that is using the interface in ways the designer didn't anticipate. It is a common form of architectural rot. Reducing the incidence of this in large code bases is usually enforced by soft means, like sternly worded comments, vigorous policing, etc but that only goes so far. Ideally, an access policy would be in the code itself and enforced by the compiler.
The public/private labels don't address this. If you make things public, the world has the ability to do unseemly things with your interface. If you friend a bunch of classes so they can access private methods, those classes have the ability to do unseemly things with your private implementation details. Both cases embrittle the architecture. What you really want is fine-grained ACLs to class members to document both intent and dependencies.
This method of restricting usage of methods to specific classes at compile time is clever and provides good documentation of design intent, I haven't seen that pattern before. I also would guess that the compiler elides the Badge argument since it is never used.
> With a sufficient number of users of an API, it does not matter what you promise in the contract: all observable behaviors of your system will be depended on by somebody.
Essentially, a bugfix that prevents CPU overheating breaks someone's workflow involving a temperature sensor (jokingly).
typedef struct foo_struct *foo;
foo foo_make();
void foo_destroy(foo);
void foo_frobnicate(foo, int how);
The programmers who use this cannot just edit a header file to gain access to something; the don't have a header file which tells them anything. They could reverse engineer the data structure, but to thwart that, you could scramble the order of the structure members with each new release of the component, so their hack wouldn't be backward or forward compatible.
I think you could also solve this specific problem with an internal header file in c++ too (by moving register/deregister to it's own class) but this solution solves other problems like binary compatibility.
It's also zero width, so no temporaries, etc.
A few bytes of stack is the least of my concerns on some file system device registration function that is called five times when the system boots.
If we imagine a software organization "going to town" with this badge approach so that there are badges all over the code base, it's not hard to imagine how it would be a nuisance.
They're supposed to trigger a reconsideration of the API when a new developer runs into a restriction, but more often in my experience the new developer will just add a new public interface.
Flipping a "private:" to "public:" has no effect on the binary compatibility.
That's not the case on Windows. The access qualifier is part of the mangled name. https://en.wikiversity.org/wiki/Visual_C%2B%2B_name_mangling...
Of course you can get around if you really, really want to. But I think this subthread is starting to split hairs.
--
C++'s private inheritance can be used here to avoid some of the gymnastics, wherein, that private aspect of the class can be passed out into the appropriate party at creation time or otherwise.
To illustrate: https://gist.github.com/arithma/0d96aec5c07e2d8dc24cfc6cd31f...
Note: I have never used this in production as far as I remember.
template<typename T>
Badge<T> FakeBadge() {
struct Stub {}
return reinterpret_cast<Badge<T>>(Stub());
} template<typename T>
class Badge {
friend T;
Badge() {}
};
template<typename T>
Badge<T> FakeBadge() {
struct Stub {};
return reinterpret_cast<Badge<T>>(Stub());
};
class Device {
public:
static Badge<Device> getBadge() { return Badge<Device>(); }
};
void needDeviceBadge(Badge<Device> badge)
{
}
int main()
{
needDeviceBadge(Device::getBadge()); // no error
needDeviceBadge(FakeBadge<Device>()); // error!
return 0;
}
badge.cc: In instantiation of ‘Badge<T> FakeBadge() [with T = Device]’:
badge.cc:25:38: required from here
badge.cc:10:10: error: invalid cast from type ‘FakeBadge() [with T = Device]::Stub’ to type ‘Badge<Device>’
return reinterpret_cast<Badge<T>>(Stub());
^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
That doesn't even have to do with friendship/protection. If we move that line into the Device class, the reinterpret_cast still doesn't compile.
The rule should be that if the default constructor of T is not accessible in the scope, then any cast to T * or T & requires a diagnostic. Or something like that.
If I can allocate a suitably aligned buffer of sizeof (T) zeros, and then treat that as a T, I've effectively constructed a T.
It would be better to do eg "return Device::private:get_badge()", but C++ defectively doesn't support that.
Edit: Actually, from reading some of the other comments in the thread, the standard apparently specifies that empty structs are 1 byte in size, so technically they would both be the same size, so casting between them should be fine.
template<typename T>
Badge<T> FakeBadge() {
static struct Stub {} stub;
return *(Badge<T>*)&stub;
};The improvement that Badge<T> makes is the <T> part; you don't need to look up the PassKey declaration to see who can construct PassKeys, it's encoded right there in the type instead.
I do like how PassKey makes the copy ctor private. I think we can improve on that and delete both the copy and move ctors in Badge.
As a programmer, I would feel frustrated to have to figure this out instead of spending time on something a user would actually care about.
Arguably the Badge pattern as presented in the article is just an honor system that provides zero security, since it's trivial to forge a fake Badge: https://repl.it/repls/BountifulQuerulousCron
https://repl.it/repls/HatefulMadeupMicrostation
Of course, type systems are orthogonal to any real security, at least in most languages, and certainly all languages with no-holds-barred interface to machine code.
In this case, once you know what a Badge<T> is, it’s obvious what it communicates about a function that takes one.
In Java or C# (and of course in C++), you could have a token system where the token can only be provided by Device :
VFS: void register(DeviceToken):
DeviceToken: empty interface (marker)
Device: private DeviceToken provideDeviceToken();
This hides both VFS and Device classes from accessing each others' private members.
You can also constrain how gets to provide DeviceToken by adding sufficient comments in code.
The misfeature is this: a class A can only declare a specific member function of class B as a friend, if that B member function is public!
Example:
The idea here is that Device has a static member function called Device::registrationHelper. That specific function (not the entire Device class) is declared a friend to VFS, and so that function can call VFS::registerDevice.
[Edit: this doesn't quite match the Badge solution, though. If this worked, it would have the problem that the registrationHelper has access to all of VFS, which is almost as bad as the entire Device class being a friend of VFS. That is one of the problems which motivate Badge. Nice try, though! The C++ restriction is probably worth fixing anyway. What the C++ friendship mechanism really needs here is to be able to say that a specific member function in B is allowed to access a specific member of A.]
class Device;
class VFS;
class Device {
public: // we want this private!
static void registrationHelper(VFS &v, Device &d);
private:
void registerWith(VFS &vfs) { registrationHelper(vfs, *this); }
};
class VFS {
private:
void registerDevice(const Device &);
friend void Device::registrationHelper(VFS &, Device &);
};
void Device::registrationHelper(VFS &v, Device &d)
{
v.registerDevice(d);
}
If we make Device::registrationHelper private to prevent this, then the "friend void Device::registrationHelper" declaration in VFS fails.
This is an oversight; the privacy of the Device::registrationHelper identifier means that the VFS class is not even allowed to mention its name for the sake of declaring it a friend.
That should be fixed in C++. Declaring a function a friend shouldn't require it to be accessible; we are not lifting its address or calling it; we are just saying that it can call us. Allowing a function to call us is not a form of access to that function, yet is being prevented by access specifiers.
To me this is a violation of the Interface Segregation Principle. I think the author went the right way to fix it, but not far enough.
He added a parameter to segregate the device registration operations from other operations. It works, but the segregation is artificial : you still need a Device object (which seems to be used by a lot of classes, another smell, god object ?)
Instead of asking for an entire Device object, register_device should ask for a different type, specific to that operation, like « NewDevice », « UnregisteredDevice », etc. This object would depend on very specific information only available at the device creation. Now, if another developer tried to call register_device, he would need to create a weird object that he never heard of, with parameters he cannot even provide.
Thoughts ?
edit: what this means is that even if the struct had data members, you couldn't brace initialize it like an aggregate type, and to achieve something syntactically similar, you would need a suitable constructor.
A glaring security hole. Any old hacker can forge or clone a data structure. This "badge" (AKA token) has to be explicitly unpredictably replay-proof generated and hard to forge, and also automatically verified.
If people would only obey simple rules, C++ would be a memory safe language. Alas, they don't.
class Device;
class VFS {
public:
class {
friend Device;
void operator()(int y) {/* do stuff with y here*/}} register_device;
};
class Device {
public:
void foo(VFS fs) {
fs.register_device(1);
}
};
int main(int argc, char *argv[]) {
// fails:
// VFS().register_device(2);
// works:
Device().foo(VFS());
return 0;
}
my c++ skills are a bit rusty, because most of the time python works just fine.
1. Aesthetics. (This is down to personal taste of course, but I much prefer how Badge gets the job done without needing multiple lines of code at every declaration.)
2. What if I need to put the function implementation out-of-line?
No warning:
void VFS::register_device(Badge<Device>, Device& device) { ... }
void VFS::register_device(Badge<Device> badge, Device& device) { ... }> Now, nobody except Device should ever be calling VFS::register_device() or VFS::unregister_device(),
Why?
Either have VFS call the constructor because non-registered devices are banned and RAII is good, or don't create arbitrary restrictions to hamstring how the rest of the system manages Devices.