[PATCH v2 0/5] exec: unify native/compat code

[Oleg Nesterov <oleg@redhat.com>]

On 02/26, Oleg Nesterov wrote:
>
> On 02/26, Linus Torvalds wrote:
> >
> > See? The advantage of the union is that the types are correct, which
> > means that the casts are unnecessary.
>
> My point was, apart from the trivial get_arg_ptr() helper, nobody else
> uses this argv/envp, so I thought it is OK to drop the type info and
> use "void *".
>
> But as I said, I won't insist. I'll redo/resend.

Well, yes... But it turns out I didn't actually read what you proposed.

	typedef union {
		compat_uptr_t compat;
		const char __user *native;
	} conditional_user_ptr_t;

	...

	where that 'do_execve_common()' takes it's arguments as

		union conditional_user_ptr_t __user *argv,
		union conditional_user_ptr_t __user *envp

I hope you didn't really mean this...

OK, we have two kinds of pointers, the union makes sense. But I think
we do not want the 3rd kind, pointer to the union. This can't help to
avoid the casts. Yes, get_arg_ptr() can do

	&argv->native

but this still means the cast even if looks differently (and tricky).

And. How can we pass "argv" from do_execve() to do_execve_common() ?
We need another cast.

So. If you insist you prefer the pointer to the union - no need to
convince me. Just say this and I'll redo again.

This patch does:

	typedef union {
		const char __user *const __user *native;
		compat_uptr_t __user *compat;
	} conditional_user_ptr_t;

	static int do_execve_common(const char *filename,
			conditional_user_ptr_t argv,
			conditional_user_ptr_t envp,
			struct pt_regs *regs, bool compat)

get_arg_ptr() does argv.native/compat, this looks more understandable.

Do you agree?

copy_strings_kernel() still needs the cast, but this is only because
we want to add "__user" for annotation.

Oleg.