Re: [mingo@elte.hu: [git pull] headers_check fixes]

[Linus Torvalds <torvalds@linux-foundation.org>]

On Wed, 28 Jan 2009, Ingo Molnar wrote:
> 
> It would be much better if we exported _much_ less and reduced our 
> cross-section to user-space. Also, the include/linux/Kbuild rules are all 
> but transparent: it would also be nice if whatever we exported was be 
> visible straight in the header itself, to make it obvious to people who 
> modify/extend those files that those definitions are going to be exported 
> to user-space.
> 
> Some __user_export tag on structures perhaps? I have no good ideas here - 
> #ifdefs are ugly and tags obscure the purity of the code.

I agree, and we probably could do so with some sparse extension (just make 
the rule be that in order to make a __user pointer, the base type needs to 
have been tagged with __user_visible). So we _could_ do something like 
that, and have a sane checking model where sparse would warn if it sees a 
user pointer of something that wasn't marked as being a user-visible type.

But I doubt it's really worth it. A lot of the usage cases for users end 
up being about constants etc that we really can't check sanely and 
automatically. And some historical user-space usage is literally just 
about user space finding the kernel headers really helpful (ie using the 
list.h headers for user-space lists, or the inline asms for atomic stuff 
for threaded apps that don't rely on ptrace locking).

		Linus