Re: + uapi-fix-linux-sysctlh-userspace-compilation-errors.patch added to -mm tree

["Dmitry V. Levin" <ldv@altlinux.org>]

[-- Attachment #1: Type: text/plain, Size: 2170 bytes --]

On Mon, Feb 27, 2017 at 12:38:16PM +0300, Alexey Dobriyan wrote:
> On Sat, Feb 25, 2017 at 1:35 AM, Dmitry V. Levin <ldv@altlinux.org> wrote:
> > On Fri, Feb 24, 2017 at 05:35:33PM +0300, Alexey Dobriyan wrote:
> >> On Thu, Feb 23, 2017 at 02:44:31PM -0800, akpm@linux-foundation.org wrote:
> >> > /usr/include/linux/sysctl.h:38:2: error: unknown type name 'size_t'
> >> >   size_t *oldlenp;
> >> > /usr/include/linux/sysctl.h:40:2: error: unknown type name 'size_t'
> >> >   size_t newlen;
> >>
> >> > --- a/include/uapi/linux/sysctl.h~uapi-fix-linux-sysctlh-userspace-compilation-errors
> >> > +++ a/include/uapi/linux/sysctl.h
> >> > @@ -26,6 +26,10 @@
> >> >  #include <linux/types.h>
> >> >  #include <linux/compiler.h>
> >> >
> >> > +#ifndef __KERNEL__
> >> > +#include <stddef.h>                /* For size_t. */
> >> > +#endif
> >>
> >> There is __kernel_size_t for cases like this.
> >
> > No, __kernel_size_t is not for cases like this because size_t differs
> > from __kernel_size_t on x32 and mips n32.
> >
> > Fortunately, there is no sysctl syscall entry on x32, but on mips m32
> > it is there with number 6152, implemented using compat_sys_sysctl.
> > The latter operates with argument of type struct compat_sysctl_args*
> > where newlen is of type compat_size_t.
> >
> > If you change it from size_t to __kernel_size_t, you'll break mips n32.
> 
> You're gluing kernel interface to an userspace header
> which kernel doesn't control.

Are you sure it's me who does the gluing?

As no uapi header provides a definition of size_t, they are essentially
relying on size_t being provided elsewhere.

For this reason, inclusion of <stddef.h> is the most conservative fix.

> How can this fix anything?

For example,
AC_CHECK_HEADERS([<linux/sysctl.h>]) doesn't work, one has to use
AC_CHECK_HEADERS([<linux/sysctl.h>], , , [#include <stddef.h>]) instead.
The same issue arises with other autoconf checks.

I think that every uapi header file should be standalone, i.e. it must
include, directly or indirectly, every definition it uses, so that no user
should have to guess header dependencies.


-- 
ldv

[-- Attachment #2: Type: application/pgp-signature, Size: 801 bytes --]