Re: [TOMOYO #5 02/18] Add wrapper functions for VFShelperfunctions.

[Trond Myklebust <trond.myklebust@fys.uio.no>]

On Sat, 2007-11-17 at 13:04 +0900, Tetsuo Handa wrote:
> Hello.
> 
> Trond Myklebust wrote:
> > The problem is that you have thrown away the main tool for documenting
> > the requirement, and for enforcing correctness (i.e. function argument
> > checking by the compiler).
> I'm sorry. I wanted to know whether the below approach is possible.

"Possible" is completely irrelevant. The question you _should_ be asking
is "is it maintainable".

> > The old functions are still there, are still exported, and still take
> > the same arguments as before, but you have now added a hidden
> > requirement that I have to set last_vfsmount when I call them. If I
> > haven't read your patch, and just call one of those vfs_* functions as
> > before, without setting last_vfsmount, I break your model, and I won't
> > find out until someone reports an obscure bug at runtime.
> I don't care if some kernel module calls vfs_* without setting last_vfsmount
> because there is no reason to reject requests from kernel code.
> Setting last_vfsmount is required for requests from userland process.
> This approach makes it possible to
> (1) tell whether the vfs_* calls are from userland or kernel
> and
> (2) leave kernel modules that are not called from userland unchanged.
> This is why this patch doesn't and needn't to modify fs/*/ files.

I'm confused. How do you tell the difference between a 'userland'
request and a 'kernel' request, and why is the latter safe from a
security perspective?

Trond