Re: [PATCH v2 0/2] Lock and Pointer guards

[Linus Torvalds]

On Thu, Jun 8, 2023 at 1:53 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> Or perhaps use the smart-pointer concept applied to our classes like:
>
> #define smart_ptr(name, var) \
>         __INSTANTIATE_VAR(name, var)

So this is the only situation where I think "ptr" makes sense (your
"fat pointer" argument is nonsensical - sure, you can treat anything
as a pointer if you're brave enough, but that just means that
"pointer" becomes a meaningless word).

However, I think that for "smart pointers", we don't need any of this
complexity at all, and we don't need that ugly syntax.

> Then we can write:
>
> DEFINE_CLASS(kfree, void *, kfree(THIS), p, void *p)
>
>         smart_ptr(kfree, mem) = kzalloc_node(...);
>         if (!mem)
>                 return -ENOMEM;

No, the above is broken, and would result in us using "void *" in
situations where we really *really* don't want that.

For automatic freeing, you want something that can handle different
types properly, and without having to constantly declare the types
somewhere else before use.

And no, you do *not* want that "kfree(THIS)" kind of interface,
because you want the compiler to inline the freeing function wrapper,
and notice _statically_ when somebody zeroed the variable and not even
call "kfree()", because otherwise you'd have a pointless call to
kfree(NULL) in the success path too.

So for convenient automatic pointer freeing, you want an interface
much more akin to

        struct whatever *ptr __automatic_kfree = kmalloc(...);

which is much more legible, doesn't have any type mis-use issues, and
is also just trivially dealt with by a

  static inline void automatic_kfree_wrapper(void *pp)
  { void *p = *(void **)pp; if (p) kfree(p); }
  #define __automatic_kfree \
        __attribute__((__cleanup__(automatic_kfree_wrapper)))
  #define no_free_ptr(p) \
        ({ __auto_type __ptr = (p); (p) = NULL; __ptr; })

which I just tested generates the sane code even for the "set the ptr
to NULL and return success" case.

The above allows you to trivially do things like

        struct whatever *p __automatic_kfree = kmalloc(..);

        if (!do_something(p))
                return -ENOENT;

        return no_free_ptr(p);

and it JustWorks(tm).

And yes, it needs a couple of different versions of that
"__automatic_kfree()" wrapper for the different freeing cases (kvfree,
rcu_free, whatever), but those are all trivial one-liners.

And no, I didn't think too much about those names. "__automatic_kfree"
is too damn long to type, but you hated "auto". And "no_free_ptr()" is
not wonderful name either. But I tried to make the naming at least be
obvious, if not wonderful.

               Linus