Is doing that manually worth it? Usually not, but for some core types (classical example is strings) or in language runtimes it can be.

Would it be awesome if this could be done automatically? Absolutely, but I understand it is a large change, and the plan is to later build upon the pattern types that are currently work in progress (and would allow you to specify custom ranged integer typed).

(is the current version of that paper, the tracking ticket insisted there's a P3125R5 and that LEWG had seen it in 2025, but it isn't listed in a mailing so it might be a mirage)

You know it's a Hana paper because it wants this to be allowed at compile time (C++ constrexpr) but joking aside this seems like a nice approach for C++ which stays agnostic about future implementation details.

If this kind of thing floats your boat, you might be interested in the non-reading variants of these as well. Mostly for things like add, max, etc but some recent architectures actually offer alternate operations to skip the read-back. The paper calls them “atomic reduction operations” https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2025/p31...

- the value is often doesn't require an update, and

- there's contention on the cache line, i.e., at least two cores frequently read or write that cache line.

But there are important details to consider:

1) The probing load must be atomic. Both the compiler and the processor in general are allowed to split non-atomic loads into two or more partial loads. Only atomic loads – even with relaxed ordering – are guaranteed to not return intermediate or mixed values from other atomic stores.

2) If the ordering on the read part of the atomic read-modify-write operation is not relaxed, the probing load must reflect this. For example, an acq-rel RMW op would require an acquire ordering on the probing read.

Well, if you target a specific architecture, then of course you can assume more guarantees than in general, portable code. And in general, a processor might distinguish between non-atomic and relaxed-atomic reads and writes – in theory.

But more important, and relevant in practice, is the behavior of the compiler. C, C++, and Rust compilers are allowed to assume that non-atomic reads aren't influenced by concurrent writes, so the compiler is allowed to split non-atomic reads into smaller reads (unlikely) or even optimize the reads away if it can prove that the memory location isn't written to by the local thread (more likely).

  atomic adds
  n = 1 ->  333e6 adds/second
  n = 2 ->  174e6
  n = 4 ->   95e6
  n = 8 ->   63e6

  atomic maxs
  n = 1 ->  161e6 maxs/second
  n = 2 ->   59e6
  n = 4 ->   39e6
  n = 8 ->   27e6

  atomic maxs with preceding check
  n = 1 ->  929e6 maxs/second
  n = 2 -> 1541e6
  n = 4 -> 3494e6
  n = 8 -> 5985e6

Also am I understanding it correctly that n is the number of threads in your example? Don't you find it suspicious that the number of operations goes up as the thread count goes up?

edit: ok, you are saying that under heavy contention the check avoids having to do the store at all. This is racy, and whether this is correct or not, would be very application specific.

edit2: I thought about this a bit, and I'm not sure i can come up with a scenario where the race matters...

edit3: ... as long as all threads are only doing atomic_max operations on the memory location, which an implementation can't assume.

What assumes that?

If your early read gives you a higher number, quitting out immediately is the same as doing the max that same nanosecond. You avoid setting a variable to the same value it already is. Doing or not doing that write shouldn't affect other atomics users, should it?

In general, I should be able to add or remove as many atomic(x=x) operations as I want without changing the result, right?

And if your early read is lower then you just do the max and having an extra read is harmless.

The only case I can think of that goes wrong is the read (and elided max) happening too early in relation to accesses to other variables, but we're assuming relaxed memory order here so that's explicitly acceptable.

A load that finds a smaller value is trickier to analyze, but i think you are just free to ignore it and just proceed with the atomic max. An underlying LL/SC loop to implement a max operation might spuriously fail anyway.

edit: here is another argument in favour: if your only atomic RMW is a cas, to implement X.atomic_max(new) you would:

  1: expected <- X 
  2: if new < expected: done
  3: else if X.cas(expected, y): done
     else goto 2 # expected implicitly refreshed

Afaik on all of x86, arm, and riscv an atomic load of a word sized datum is just a regular load.

We're talking about collating a maximum, by definition every write to that is an increase.

I believe this is a bit trickier than that, you would also need at least some kind of atomic barrier to preserve the ordering semantics of the successful update case.

Anyway, much as I have cases where it matters I guess what I’m saying is I think you’re right that vfork is rarely actually necessary, especially since you’d probably have a much easier time getting a faster and still deterministic spawn if it ever actually becomes a bottleneck for something you care about.

You can always accomplish that sort of thing by using a helper program that ultimately execs the desired one -- just prefix it and its arguments to the intended argv.

Yes, and typically for the same reasons as in JVMs etc: `vfork()` is just faster.

In the Linux VFS for example, there are optimized functions for reading and writing, but if those are not implemented, a fallback to unoptimized functions is done at the call sites. Both sets are function pointers and you only need to implement one if I recall correctly.

I think the "is/is not" question is not so clear. If you think of "is" as a whether there's a homomorphism, then it makes sense to say that it is OOP, but it can qualify as being something else too, ie. it's not an exclusionary relationaship.

If you think of the differences as being OOP implies contracts with the compiler and data abstraction does not (beyond a simple structure saying where the members are in memory), it becomes easier to see the two as different things.

You can implement OOP in C if you do vtables for inheritance hierarchies manually, among other things, but that is different than what Linux does.

It's a function belonging to an object, to which is dynamically dispatched with something I would call a vtable. To me that sounds like a classic example of OOP.

Data abstraction is a core of OOP.

This pattern can be used to implement inheritance, when it isn't here that doesn't mean its not OOP.

> Object-oriented programming (OOP) is a programming paradigm based on the object – a software entity that encapsulates data and function(s). An OOP computer program consists of objects that interact with one another. A programming language that provides OOP features is classified as an OOP language [...]

You don't disagree, that this kernel pattern is about data abstraction. You probably don't disagree, that the kernel uses functions. The kernel uses "objects" (FS implementations) that follow a defined set of functions, sometimes called "class" (vtables/wtables/however you like to call them). Therefore I conclude what the kernel does here is a prime example of OOP.

I know ADT and OOP are different concepts, in another answer I wrote what I think the base differences are. But they are related, and in the definitions I am familiar with, ADTs are a base concept for OOP. And OOP can be an implementation for ADTs.

If you don't think the implementations I provide are enough to apply to the kernel, can maybe provide your own definition according to which we can evaluate this, because I feel like we are beating around the bush. But please not something that says this can only be happening in an OOP language, to these I would plainly disagree with, because OOP is a paradigm and not a property of a language.

Sorry, but where did you got this definition from? I've always thought OOP as a way of organizing your data and your code, sometimes supported by language-specific constructs, but not necessarily.

Can you organize your data into lists, trees, and hashmaps even if your language does not have those as native types? So you can think in a OO way even if the language has no notion of objects, methods, etc.

It is from experience with object oriented languages (mainly C++ and Java). Technically, you can do everything manually, but that involves shoehorning things into the OO paradigm that do not naturally fit, like the article author did when he claimed struct file_operations was a vtable when it has ->check_flags(), which would be equivalent to a static member function in C++. That is never in a vtable.

If Al Viro were trying to restrict himself to object oriented programming, he would need to remove function pointers to what are effectively the equivalent of static member functions in C++ to turn it into a proper vtable, and handle accesses to that function through the “class”, rather than the “object”.

Of course, since he is not doing object oriented programming, placing pointers to what would be virtual member functions and static member functions into the same structure is fine. There will never be a use case where you want to inherit from a filesystem implementation’s struct file_operations, so there is no need for the decoupling that object oriented programming forces.

> I've always thought OOP as a way of organizing your data and your code, sometimes supported by language-specific constructs, but not necessarily.

It certainly can be, but it is not the only way.

> Can you organize your data into lists, trees, and hashmaps even if your language does not have those as native types?

This is an odd question. First, exactly what is a native type? If you mean primitive types, then yes. Even C++ does that. If you mean standard library compound types, again, yes. The C++ STL started as a third party library at SGI before becoming part of the C++ standard. If you mean types that you can define, then probably not without a bunch of pain, as then we are going back to the dark days of manually remembering offsets as people had to do in assembly language, although it is technically possible to do in both C and C++.

What you are asking seems to be exactly what data abstraction is, which involves making an interface that separates use and implementation, allowing different data structures to be used to organize data using the same interface. As per Wikipedia:

> For example, one could define an abstract data type called lookup table which uniquely associates keys with values, and in which values may be retrieved by specifying their corresponding keys. Such a lookup table may be implemented in various ways: as a hash table, a binary search tree, or even a simple linear list of (key:value) pairs. As far as client code is concerned, the abstract properties of the type are the same in each case.

https://en.wikipedia.org/wiki/Abstraction_(computer_science)...

Getting back to doing data structures without object oriented programming, this is often done in C using a structure definition and the CPP (C PreProcessor) via intrusive data structures. Those break encapsulation, but are great for performance since they can coalesce memory allocations and reduce pointer indirections for objects indexed by multiple structures. They also are extremely beneficial for debugging, since you can see all data structures indexing the object. Here are some of the more common examples:

https://github.com/openbsd/src/blob/master/sys/sys/queue.h

https://github.com/openbsd/src/blob/master/sys/sys/tree.h

sys/queue.h is actually part of the POSIX standard, while sys/tree.h never achieved standardization. You will find a number of libraries that implement trees like libuutil on Solaris/Illumos, glib on GNU, sys/tree.h on BSD, and others. The implementations are portable to other platforms, so you can pick the one you want and use it.

As for “hash maps” or hash tables, those tend to be more purpose built in practice to fit the data from what I have seen. However, generic implementations exist:

https://stackoverflow.com/questions/6118539/why-are-there-no...

That said, anyone using hash tables at scale should pay very close attention to how their hash function distributes keys to ensure it is as close to uniformly random as possible, or you are going to have a bad time. Most other applications would be fine using binary search trees. It probably is not a good idea to use hash tables with user controlled keys from a security perspective, since then a guy named Bob can pick keys that cause collisions to slow everything down in a DoS attack. An upgrade from binary search trees that does not risk issues from hash function collisions would be B-trees.

By the way, B-trees are containers and cannot function as intrusive data structures, so you give up some convenience when debugging if you use B-Trees.

You don't need classes for OOP. C++ not putting methods that logically operate on an object, but don't need a pointer to it, into the automatically created vtable, is an optimization and an implementation detail. I don't know why you think that putting this function into a vtable precludes OOP.

Wait, how does inheritance work when the method is not in the vtable?

As for how inheritance works when the member function is not in the vtable, that depends on what kind of member function it is. All C++ functions are given a mangled name that is stuffed into C’s infrastructure for linking symbols. For static member functions, inheritance is irrelevant since they are tied to the class. Calls to static member functions go directly to the mangled function with no indirections, just as if it had been a global function. For non-static virtual member functions, you use the vtable pointer to find it. For non-virtual member functions, the call goes straight to the function as if a global function had been called (and the this pointer is still passed, even if the function does not use it), since the compiler knows the type and thus can tell the linker to have calls there go to the function through the appropriately mangled name. It is just like calling a global function.

Yes. Since we are not in C++ we can choose to get rid of this useless pointer.

> Removing it on member functions that do not use it would pose a problem if another class inherited from this class and overrode the function definition with one that did use it.

That problem has nothing to do with the this pointer specifically. When you change the method signature of an inherited method you always have this problem. This simply means, that the superclass prescribes limits to subclasses, which is why it's possible to use a subclass inplace of a superclass.

> Maybe in very special cases whole program optimization could safely remove the this pointer, but it is questionable whether any compiler author

Yes, that's why its not done in C++, but we can do it, if we handroll it.

> it would break ABI stability

It does not if it has always been like this.

> For static member functions, inheritance is irrelevant since they are tied to the class. Calls to static member functions go directly to the mangled function with no indirections

In other words, ->check_flags() can't be implemented as a static member functions in C++. It would simply have a this pointer, that it just wouldn't use, since C++ has no way to express non-static member functions, that just don't take a this pointer.

> thus can tell the linker to have calls there go to the function

In our case the linker can only resolve the call to the appropriate vtable, since the type isn't known until runtime.

If you were trying to implement OOP in the kernel in C and implemented a vtable, you cannot get rid of the this pointer in vtable entries since a child class might want to use it in the overrode definition. It is one of the same reasons why you cannot remove it in C++. The entire point of a vtable is to enable inheritance. If OOP really were being done, an out of tree module could make a class that inherits from this one without needing any code changes and use the this pointer, but you cannot do that if you drop the this pointer. I already explained this.

You already prescribe with the chosen arguments in the superclass on which things the child implementation can depend. Why not also do this with the first argument?

Also why do you care exactly about the order of arguments? The nature of the function doesn't change, it's entirely arbitrary and orthogonal to the paradigm the function implements. Another example is the implementation of the equality operator between objects. In languages with syntactic sugar you typically have (self, other), but if its the true equality operator then the order doesn't matter.

Isn't this exactly how most (every?) OOP language implements it? You would say a C++ virtual method isn't OOP?

The entire point of OOP is to make contracts with the compiler that forcibly tie certain things together that are not tied together with data abstraction. Member functions are subject to inheritance and polymorphism. Member function pointers are not. Changing the type of your class will never magically change the contents of a member function pointer, but it will change the constants of a non-virtual member function. A member function will have a this pointer to refer to the class. A member function pointer does not unless you explicitly add one (named something other than this in C++).

You claim when the compiler does this, it's OOP, but when I do it, it's not?

Of course you could implement the same in C++ and then it can't be the same as the vtable introduced by the compiler, so you would just end up with to vtables, you own and the one introduced by the compiler.

As I said, these are NOT vtables. The fact that you and some others keep thinking of them as vtables misleads you into thinking that this can be done using the object oriented tools of C++. It cannot without major hacks and the result would be slower, harder to read and only something that a bureaucrat could like.

> As I said, these are NOT vtables

Ok, you just define vtables differently then me. To me a vtable is a table of virtual functions that are used to implement polymorphic behaviour of objects. This applies to their usage in the kernel and the article. Feel free to introduce a new term for this. If your only distinction is whether these are created by a compiler, this is just a distinction I don't care about.

As for the one distinction you recognize and think is invalid, that distinction is given by the definition you found. You refuse to obey the definition you yourself quoted to settle matter elsewhere in the thread.

It's not the only distinction I recognize, its the distinction you think matters here and that seams to be the basis for our disagreement.

This is the term (vtable, VMT) I got told in lectures to describe this pattern, you have yet not pointed me to a different term that you would recognize to be this, so in lack of a better term I will continue to use this.

As to why I think this distinction does not matter here, is because I perceive the compiler to be a tool that generates code which is controlled by the programmer. Thus the programmer in both cases creates codes with the same paradigm, they only differ in the tools used. We generally don't name things differently depending on which tools are used in the creation, except if they are created with a different intention.

Not completely true either, it eventually supported most of the normal 3d primitives but gaming performance was never a priority because there were few developers and they weren't employed by AMD/ATI -- which also meant that some cards would only reach full feature support after their EOL, sadly.

The amdgpu also driver benefits from a lot of the groundwork that has been done since. The radeon driver is older than kernel features like KMS (kernel modesetting) and GEM (graphics execution manager), and the LLVM-based shader compiler in mesa (userspace). I'd say that the radeon driver was actually the proving ground for many of these features, because it was the most capable open source 3d driver: The Intel 845/915 hardware barely supported 3d operations, and the only 3d-capable open source driver for Nvidia was the reverse-engineered nouveau driver.

Luckily, many people working on the amdgpu driver are actually on AMD's payroll these days.

I forgot that name "fglrx", probably a mental self-defense mechanism. Those were some bad times, trying to get different display outputs to work at the same time, guessing and testing values in xorg.conf, so on. There was some community utility someone wrote to try and help with installation, reinstallation, configuration and reconfiguration, but the name eludes me now.

I would edit my post to correct it, but it seems the edit window has passed.