Paul Brook wrote:
> On Monday 04 August 2008, Gerd Hoffmann wrote:
>> Blue Swirl wrote:
>>> On 8/4/08, Gerd Hoffmann <kraxel@xxxxxxxxxx> wrote:
>>>> +/* i386 protocol version */
>>>> +#pragma pack(push, 4)
>>> What's wrong with __attribute__(__aligned__)?
>> That one is tricky because we must be binary compatible with xen. And
>> the block driver protocol has a small design flaw: It has a 64bit value
>> which is not aligned on a 64bit border. This leads to different struct
>> layouts on i386 and x86_64 due to different alignment rules. Unless you
>> force them into something else, like we do in that header file to deal
>> with it. Which is needed to run 32bit guests on 64bit hosts.
>
> qemu should be host independent. You need to use packed structures with 
> appropriate padding.
I think the current code is correct. Ok, the verbose version. The
struct in question looks like this:
 struct blkif_request {
 uint8_t operation;
 uint8_t nr_segments;
 blkif_vdev_t handle;
 uint64_t id;
 [ ... ]
 };
The critical element is id. blkif_vdev_t is uint16_t, which means id is
32bit aligned but not 64bit aligned.
On i386 64bit values get 32bit alignments by default.
On x86_64 64bit values get 64bit alignments by default.
Thus x86_64 has a 32bit padding hole just before id. i386 hasn't.
So we define a i386 and a x86_64 version of the struct. The i386
version is wrapped into a pragma like this ...
 #pragma pack(push, 4)
 struct blkif_x86_32_request {
 [ ... ]
 }
 #pragma pack(pop)
... to enforce i386 alignment rules everythere.
The x86_64 version looks like this ...
 struct blkif_x86_64_request {
 [ ... ]
 uint64_t __attribute__((__aligned__(8))) id;
 [ ... ]
 };
... to make sure the id element is 64bit aligned everythere.
That gives us the correct struct layouts for both i386 and x86_64 builds.
cheers,
 Gerd
-- 
http://kraxel.fedorapeople.org/xenner/
_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel