Re: [Qemu-devel] [qemu-s390x] [PATCH v2 2/3] hw/s390x/css: Remove QEMU_PACKED from struct SenseId

[David Hildenbrand]

On 26/09/2018 10:09, Thomas Huth wrote:
> On 2018-09-26 10:07, David Hildenbrand wrote:
>> On 26/09/2018 10:04, David Hildenbrand wrote:
>>> On 26/09/2018 09:38, Thomas Huth wrote:
>>>> The uint16_t member cu_type of struct SenseId is not naturally aligned,
>>>> and since the struct is marked with QEMU_PACKED, this can lead to
>>>> unaligned memory accesses - which does not work on architectures like
>>>> Sparc. Thus remove the QEMU_PACKED here and rather copy the struct
>>>> byte by byte when we do copy_sense_id_to_guest().
>>>>
>>>> Signed-off-by: Thomas Huth <address@hidden>
>>>> ---
>>>>  hw/s390x/css.c         | 33 +++++++++++++++++----------------
>>>>  include/hw/s390x/css.h |  2 +-
>>>>  2 files changed, 18 insertions(+), 17 deletions(-)
>>>>
>>>> diff --git a/hw/s390x/css.c b/hw/s390x/css.c
>>>> index 5a9fe45..0e51b85 100644
>>>> --- a/hw/s390x/css.c
>>>> +++ b/hw/s390x/css.c
>>>> @@ -750,20 +750,20 @@ static void sch_handle_halt_func(SubchDev *sch)
>>>>  
>>>>  }
>>>>  
>>>> -static void copy_sense_id_to_guest(SenseId *dest, SenseId *src)
>>>> +static void copy_sense_id_to_guest(uint8_t *dest, SenseId *src)
>>>>  {
>>>>      int i;
>>>>  
>>>> -    dest->reserved = src->reserved;
>>>> -    dest->cu_type = cpu_to_be16(src->cu_type);
>>>> -    dest->cu_model = src->cu_model;
>>>> -    dest->dev_type = cpu_to_be16(src->dev_type);
>>>> -    dest->dev_model = src->dev_model;
>>>> -    dest->unused = src->unused;
>>>> -    for (i = 0; i < ARRAY_SIZE(dest->ciw); i++) {
>>>> -        dest->ciw[i].type = src->ciw[i].type;
>>>> -        dest->ciw[i].command = src->ciw[i].command;
>>>> -        dest->ciw[i].count = cpu_to_be16(src->ciw[i].count);
>>>> +    dest[0] = src->reserved;
>>>> +    stw_be_p(dest + 1, src->cu_type);
>>>> +    dest[3] = src->cu_model;
>>>> +    stw_be_p(dest + 4, src->dev_type);
>>>> +    dest[6] = src->dev_model;
>>>> +    dest[7] = src->unused;
>>>> +    for (i = 0; i < ARRAY_SIZE(src->ciw); i++) {
>>>> +        dest[8 + i * 4] = src->ciw[i].type;
>>>> +        dest[9 + i * 4] = src->ciw[i].command;
>>>> +        stw_be_p(dest + 10 + i * 4, src->ciw[i].count);
>>>
>>>
>>> Not really a fan of this, as we sacrifice readability due to one
>>> unaligned member. What about only converting the unaligned members (e.g.
>>> cu_type) from uint16_t to uint8_t[2] and adding a comment why this is
>>> split. Then the structure is naturally packed.
>>>
>>> We only have to fixup the places that check cu_type.
>>>
>>
>> Just realized this was basically suggested by Peter. If it would be as
>> simple as splitting VMSTATE_UINT16 into two VMSTATE_UINT8 or similar, I
>> would prefer that.
> 
> It's not that simple, it would break migration from older versions of
> QEMU due to endianness issues then.

Migration between different QEMUs (e.g. big to little) is not supported
as far as I remember. But my head always hurts when thinking about
endianness conversions, so I am pretty sure I am missing something here.

Another option would be to define a second struct with the
uint16_t->uint8_t[2] conversion as mentioned that is only used in
copy_sense_id_to_guest. The all fields except that one could be copied
without worrying about offsets in a destination buffer.

e.g. SenseIdGuest / SenseIdPacked or whatever.


-- 

Thanks,

David / dhildenb