Re: [Xen-devel] Re: [RFC PATCH 03/35] Add Xen interface header files

[Anthony Liguori <aliguori@us.ibm.com>]

Hollis Blanchard wrote:
> On Tue, 2006-05-09 at 16:15 +0100, Christoph Hellwig wrote:
>   
>>> +#ifdef __XEN__
>>> +#define __DEFINE_GUEST_HANDLE(name, type) \
>>> +    typedef struct { type *p; } __guest_handle_ ## name
>>> +#else
>>> +#define __DEFINE_GUEST_HANDLE(name, type) \
>>> +    typedef type * __guest_handle_ ## name
>>> +#endif
>>>       
>> please get rid of all these stupid typedefs 
>>     
>
> These typedefs are a new hack to work around a basic interface problem:
> instead of explicitly-sized types, Xen uses longs and pointers in its
> interface. On PowerPC in particular, where we need a 32-bit userland
> communicating with a 64-bit hypervisor, those types don't work.
>
> However, the maintainers are reluctant to switch the interface to use
> explicitly-sized types because it would break binary compatibility.
> These ugly "HANDLE" macros allow PowerPC to do what we need without
> affecting binary compatibility on x86.
>   

Is this strictly true though?  The ABI for Power and x86 are not 
necessarily dependent on each other.  One could just as easily define a 
typedef like:

#if defined(__ppc__)
typedef uint64_t guest_handle_t;
#elif defined(__x86__)
typedef unsigned long guest_handle_t;
#endif

I thought the use of GUEST_HANDLE was to maintain type safety.  It 
certainly helps the issue you point out but it's not strictly necessary.

IMHO, this trick makes the code pretty ugly.  I'd rather see it 
disappear in favor of something more akin to the above.

Regards,

Anthony Liguori