Re: [PATCH 0/4 RESEND] exec: unify compat/non-compat code

[Oleg Nesterov <oleg@redhat.com>]

On 02/25, Linus Torvalds wrote:
>
> On Fri, Feb 25, 2011 at 9:52 AM, Oleg Nesterov <oleg@redhat.com> wrote:
> >> On 12/01, Milton Miller wrote:
> >> >
> >> > > +#ifdef CONFIG_COMPAT
> >> > > +int compat_do_execve(char * filename,
> >> > > + compat_uptr_t __user *argv,
> >> > > + compat_uptr_t __user *envp,
> >> > > + struct pt_regs * regs)
> >> > > +{
> >> > > + return do_execve_common(filename,
> >> > > +                         (void __user*)argv, (void __user*)envp,
> >> >
> >> > Shouldn't these be compat_ptr(argv)?  (makes a difference on s390)
>
> Indeed. The "compat_uptr_t __user *argv" is wrong, and it should be just
>
>     compat_uptr_t argv;
>
> and then every time you turn it into a pointer, it should use
> "compat_ptr(argv)".

Oh, perhaps, and I was thinking about this too. But this is another
issue, no? Or I misunderstood.

First of all, I agree that perhaps it makes sense to change the
signature of compat_do_execve()

	-	compat_do_execve(compat_uptr_t __user *argv)
	+	compat_do_execve(compat_uptr_t argv)

but this has nothing to do with this series. We can do this before
or after ("after" seems simpler").

>  - user passes "compat_uptr_t"

Yes,

>  - the kernel can turn that into "compat_uptr_t __user *" by doing
>
>        compat_uptr_t __user *pptr;
>        pptr = compat_ptr(argv);

Yes! and the kernel already does this before it calls compat_do_execve(),
iow compat_do_execve() gets the result of compat_ptr(compat_ptr_from_user).

>  - the kernel needs to fetch the individual entries with
>
>        compat_uptr_t cuptr = get_user(pptr);
>
>  - the kernel can then turn _those_ into the actual pointers to the string with
>
>        const char __user *str = compat_ptr(cuptr);

Yes, and this is exactly what get_arg_ptr(compat => true) does.

> > So, once again, this should not (and can not) be compat_ptr(argv) afaics.
>
> It can be, and probably should.

Only if we change the signature of compat_do_execve(). With the current
code yet another compat_ptr() is not needed and it is simply wrong, this
is what I meant when I replied to Milton.

> But the low-level s390 wrapper
> function may have done one of the levels already. It probably
> shouldn't, and we _should_ do the "compat_ptr()" thing a the generic C
> level.

Agreed, but currently this compat_ptr() thing belongs to the caller.

IOW. Lets look at the current code. arch/ calls
compat_do_execve(compat_uptr_t __user *argv)->compat_count(argv) which
does get_user(argv) without any conversion, because argv was already
converted or arch/ is buggy.

Both do_execve() and compat_do_execve() accept the valid pointer
which does not need any conversion. But this pointer points to different
things, either to "char*" of "compat_uptr_t".

However, please see my reply to 2-3/5, I agree that this is confusing
and can be cleanuped.

Or do you think I missed something else?

Oleg.