Re: [PATCH] y2038: fix socket.h header inclusion

[Arnd Bergmann <arnd@arndb.de>]

On Sun, Mar 17, 2019 at 7:20 PM Deepa Dinamani <deepa.kernel@gmail.com> wrote:
> On Fri, Mar 15, 2019 at 2:20 PM Florian Weimer <fw@deneb.enyo.de> wrote:
> > > On Thu, Mar 14, 2019 at 7:41 PM Florian Weimer <fw@deneb.enyo.de> wrote:
> > >> > diff --git a/arch/alpha/include/uapi/asm/socket.h
> > >> > b/arch/alpha/include/uapi/asm/socket.h
> > >> > index 0d0fddb7e738..976e89b116e5 100644
> > >> > --- a/arch/alpha/include/uapi/asm/socket.h
> > >> > +++ b/arch/alpha/include/uapi/asm/socket.h
> > >> > @@ -2,8 +2,8 @@
> > >> >  #ifndef _UAPI_ASM_SOCKET_H
> > >> >  #define _UAPI_ASM_SOCKET_H
> > >> >
> > >> > +#include <linux/posix_types.h>
> > >> >  #include <asm/sockios.h>
> > >> > -#include <asm/bitsperlong.h>
> > >>
> > >> This breaks POSIX conformance in glibc because the
> > >> <linux/posix_types.h> header is not namespace clean.  It contains the
> > >> identifiers fds_bits and val:
> > >>
> > >>         unsigned long fds_bits[__FD_SETSIZE / (8 * sizeof(long))];
> > >>
> > >>         int     val[2];
> > >
> > > What is problematic about the struct members here? I had thought that
> > > only the struct names have to be in a namespace to be usable here,
> > > but not the members.
> >
> > According POSIX, a user can do this:
> >
> > #define fds_bits 1024
> >
> > before including the <sys/socket.h> header file.  Similarly for val.
> >
> > Since glibc pulls in <asm/socket.h> indirectly, the result is a parse
> > error, even though the programmer did nothing wrong (fds_bits is not
> > an identifier used by POSIX, nor is it in the implementation
> > namespace, ans <sys/socket.h> is a POSIX header).

Ok, I see. Thanks for the explanation!

> > > We could use asm/posix_types.h instead of linux/posix_types.h,
> > > would that address your concern?
> >
> > It should fix the fds_bits case, I think.  But
> > <asm-generic/posix_types.h> still uses val, so that part of the issue
> > remains.
>
> Would moving kernel namespace types(__kernel prefix) to a different
> header file(kernel_types.h?) and then including this from
> linux/posix_types.h.
> And, for socket.h just including kernel_types.h make sense?

I fear we have considered linux/posix_types.h to be something that
can be included anywhere for a long time, so it may be better to
ensure that this is actually the case, and avoid the problem with those
two structures but leave the rest untouched.

I think we can move  __kernel_fsid_t into include/uapi/asm-generic/statfs.h,
which is the only thing that needs it anyway. We have two definitions of
it today, the non-generic one being for mips32, but incidentally there was
a patch the other day to remove that and use the generic one instead.

With that done, we can change asm/socket.h to just use asm/posix_types.h.

I would still prefer to solve the problem for linux/posix_types.h as well,
but I'm not sure even how __kernel_fd_set  is used today in
user space, if at all. Commit 8ded2bbc1845 ("posix_types.h: Cleanup
stale __NFDBITS and related definitions") removed most of the fd_set
definition after a long discussion [1], and since then it has been
basically impossible to use 'struct fd_set'  from the kernel in a
meaningful way without including the libc headers or duplicating
them.

Should we just remove __kernel_fd_set from the exported headers and
define the internal fd_set directly in include/linux/types.h? (Adding the
folks from the old thread to Cc).

      Arnd

[1] https://lore.kernel.org/lkml/20120724181209.GA10534@zod.bos.redhat.com/t/