Re: CONFIG_* symbols in UAPI headers?

[David Howells]

Arnd Bergmann <arnd@xxxxxxxx> wrote:

> > I just stumbled over the MAP_UNINITIALIZED defintion, initially
> > added by:
> >
> > commit ea637639591def87a54cea811cbac796980cb30d
> > Author: Jie Zhang <jie.zhang@xxxxxxxxxx>
> > Date:   Mon Dec 14 18:00:02 2009 -0800
> >
> >     nommu: fix malloc performance by adding uninitialized flag
> >
> > The defintion depends on CONFIG_MMAP_ALLOW_UNINITIALIZED, which
> > will never be set by userspace.  How is this supposed to work?
> >
> > Shoudn't we define the symbol unconditionally and just turn it
> > into a no-op in the implementation?

Yes.

> Right, good catch. That should work. It can probably be done
> by adding another check before the conditional, like:
> 
>        /* clear anonymous mappings that don't ask for uninitialized data */
>        if (!vma->vm_file &&
>            !(IS_ENABLED(CONFIG_MMAP_ALLOW_UNINITIALIZED) &&
>              (flags & MAP_UNINITIALIZED))
>                memset((void *)region->vm_start, 0,
>                       region->vm_end - region->vm_start);

Sounds good.

> > There are a few similar issues, like struct elf_prstatus having
> > a different layout depending on CONFIG_BINFMT_ELF_FDPIC, or
> > MAX_SHARED_LIBS defending on CONFIG_BINFMT_SHARED_FLAT.

Because the kernel code uses that header and that struct too, so you'd break
compilation of binfmt_elf_fdpic.c.  There is a way round it - and that's to
copy the struct into the non-UAPI backing header and delete the conditional
section from the UAPI one.  You'd have to stop the non-UAPI header from
#including the UAPI header, though, and you'd have to hope that no one is
trying to set it in userspace (gdb doesn't).

David