Re: [Qemu-devel] [PATCH v3 3/3] hw/s390x/ioinst: Fix alignment problem in struct SubchDev

[Cornelia Huck <cohuck@redhat.com>]

On Mon, 10 Dec 2018 12:27:56 +0000
Peter Maydell <peter.maydell@linaro.org> wrote:

> On Thu, 27 Sep 2018 at 09:25, Thomas Huth <thuth@redhat.com> wrote:
> >
> > struct SubchDev embeds several other structures which are marked with
> > QEMU_PACKED. This causes the compiler to not care for proper alignment
> > of these structures. When we later pass around pointers to the unaligned
> > struct members during migration, this causes problems on host architectures
> > like Sparc that can not do unaligned memory access.
> >
> > Most of the structs in ioinst.h are naturally aligned, so we can fix
> > most of the problem by removing the QEMU_PACKED statements (and use
> > QEMU_BUILD_BUG_MSG() statements instead to make sure that there is no
> > padding). However, for the struct SCHIB, we have to keep the QEMU_PACKED
> > since the compiler adds some padding here otherwise. Move this struct
> > to the beginning of struct SubchDev instead to fix the alignment problem
> > here, too.  
> 
> Unfortunately clang does not like the struct SCHIB being still
> marked packed:
> 
> /home/petmay01/linaro/qemu-from-laptop/qemu/hw/s390x/css.c:1294:25:
> warning: taking address of packed member 'pmcw' of class or structure
> 'SCHIB' may result in an unaligned pointer value
> [-Waddress-of-packed-member]
>     copy_pmcw_to_guest(&dest->pmcw, &src->pmcw);
>                         ^~~~~~~~~~
> /home/petmay01/linaro/qemu-from-laptop/qemu/hw/s390x/css.c:1294:38:
> warning: taking address of packed member 'pmcw' of class or structure
> 'SCHIB' may result in an unaligned pointer value
> [-Waddress-of-packed-member]
>     copy_pmcw_to_guest(&dest->pmcw, &src->pmcw);
>                                      ^~~~~~~~~
> /home/petmay01/linaro/qemu-from-laptop/qemu/hw/s390x/css.c:1295:25:
> warning: taking address of packed member 'scsw' of class or structure
> 'SCHIB' may result in an unaligned pointer value
> [-Waddress-of-packed-member]
>     copy_scsw_to_guest(&dest->scsw, &src->scsw);
>                         ^~~~~~~~~~
> /home/petmay01/linaro/qemu-from-laptop/qemu/hw/s390x/css.c:1295:38:
> warning: taking address of packed member 'scsw' of class or structure
> 'SCHIB' may result in an unaligned pointer value
> [-Waddress-of-packed-member]
>     copy_scsw_to_guest(&dest->scsw, &src->scsw);
>                                      ^~~~~~~~~
> /home/petmay01/linaro/qemu-from-laptop/qemu/hw/s390x/css.c:1343:27:
> warning: taking address of packed member 'pmcw' of class or structure
> 'SCHIB' may result in an unaligned pointer value
> [-Waddress-of-packed-member]
>     copy_pmcw_from_guest(&dest->pmcw, &src->pmcw);
>                           ^~~~~~~~~~
> /home/petmay01/linaro/qemu-from-laptop/qemu/hw/s390x/css.c:1343:40:
> warning: taking address of packed member 'pmcw' of class or structure
> 'SCHIB' may result in an unaligned pointer value
> [-Waddress-of-packed-member]
>     copy_pmcw_from_guest(&dest->pmcw, &src->pmcw);
>                                        ^~~~~~~~~
> /home/petmay01/linaro/qemu-from-laptop/qemu/hw/s390x/css.c:1344:27:
> warning: taking address of packed member 'scsw' of class or structure
> 'SCHIB' may result in an unaligned pointer value
> [-Waddress-of-packed-member]
>     copy_scsw_from_guest(&dest->scsw, &src->scsw);
>                           ^~~~~~~~~~
> /home/petmay01/linaro/qemu-from-laptop/qemu/hw/s390x/css.c:1344:40:
> warning: taking address of packed member 'scsw' of class or structure
> 'SCHIB' may result in an unaligned pointer value
> [-Waddress-of-packed-member]
>     copy_scsw_from_guest(&dest->scsw, &src->scsw);
>                                        ^~~~~~~~~

That's really annoying :(

> Not sure how best to address this. A couple of ideas that I had:
> 
> (1) make the 'uint64_t mba' field in the SCHIB struct into
> two uint32_t fields, adjusting all the code which needs
> to access it accordingly; then we could drop the packed
> annotation from the struct

This would mean some annoying gymnastics, but fortunately that field is
not accessed in many places.

> 
> (2) have the guts of copy_{pmcw,scsw}_{to,from}_guest() be
> macros, so we can do them inline in the copy_schib_{to,from}_guest()
> function and thus operate directly on src->pmcw.foo &c
> fields rather than ever having to take the address of any
> of the fields in src or dest

I'm not really a fan of using macros, but if it stays readable...

Not sure what the best option is here; this is why I haven't done
anything yet to fix it, as no idea was really appealing.