Re: [PATCH v2 4/4] x86: adjustments to guest handle treatment

[Julien Grall <julien@xen.org>]

Hi,

On 22/04/2020 08:56, Roger Pau Monné wrote:
> On Tue, Apr 21, 2020 at 07:44:55PM +0100, Julien Grall wrote:
>> Hi,
>>
>> On 21/04/2020 18:30, Roger Pau Monné wrote:
>>> On Tue, Apr 21, 2020 at 11:13:23AM +0200, Jan Beulich wrote:
>>>> First of all avoid excessive conversions. copy_{from,to}_guest(), for
>>>> example, work fine with all of XEN_GUEST_HANDLE{,_64,_PARAM}().
>>>
>>> I'm not sure I understand the difference between those two, as they
>>> are both placeholders for linear guest addresses?
>>>
>>> AFAICT XEN_GUEST_HANDLE should be used for guest pointers inside of an
>>> hypercall struct, while XEN_GUEST_HANDLE_PARAM is for guest pointers
>>> as hypercall arguments. But those are both just guest pointers,
>>> whether they are a parameter to the hypercall or a field in a
>>> struct, and hence could use the same type?
>>>
>>> I assume there's some reason for not doing so, and I see the comment
>>> about other arches, but again a linear guest address is just that in
>>> all arches, regardless of it's placement.
>>
>> On Arm:
>>   * XEN_GUEST_HANDLE() will always be 64-bit on both 32-bit and 64-bit
>> hypervisor.
>>   * XEN_GUEST_HANDLE_PARAM() will be 32-bit for 32-bit hypervisor. For 64-bit
>> hypervisor, it will be 64-bit.
>>
>> Per the ABI, each argument only fit a register. So you could not use
>> XEN_GUEST_HANDLE() as now an argument will be held in 2 registers on 32-bit.
>>
>> We also want the structure layout to be the same for 32-bit and 64-bit. So
>> using XEN_GUEST_HANDLE_PARAM() everywhere is not the solution as the virtual
>> address is not the same.
> 
> Right, you hide the 'padding' inside XEN_GUEST_HANDLE by making it
> have a fixed size on all bitnesses, I can see how that's not
> desirable for hypercall params though.
> 
> Iff we ever switch to an ABI that uses physical addresses instead of
> linear ones we would have to switch everything to be a 64bit integer,
> or else 32bit PAE won't work correctly. Or come up with a completely
> different ABI (ie: use a pre-allocated set of buffer pages, IIRC as
> suggested by Juergen).

I would go further here and request the arguments to always be the same 
size regardless the bitness.

> 
>>
>> We could possibly convert internally XEN_GUEST_HANDLE_PARAM() to
>> XEN_GUEST_HANDLE(), but I am not sure how this would be beneficial. We would
>> have to use 2 registers for arm 32-bit everytime.
> 
> Hm, we could maybe expand hypercall parameters to 64bit using some
> kind of translation layer between the entry point and the actual
> handler, but anyway, I get now there's a need to keep this difference.

This can be done today using guest_handle_from_param manually.

Cheers,

-- 
Julien Grall