Re: Runtime services support for Xen on ARM

[Shannon Zhao <shannon.zhao@linaro.org>]

On 2015/11/10 20:36, Ian Campbell wrote:
> On Tue, 2015-11-10 at 12:26 +0000, Stefano Stabellini wrote:
>> CC'ing xen-devel and Jan
>>
>> On Tue, 10 Nov 2015, Shannon Zhao wrote:
>>> Hi Stefano,
>>>
>>> I'm working on adding Runtime services support at Xen side. Most of
>>> work
>>> is adding the ARM part in xen/common/efi/runtime.c.
>>>
>>> There is one problem which block me. That is how to implement
>>> efi_rs_enter() and efi_rs_leave() for ARM, since I think current
>>> implementation is x86 specific and won't work on ARM. Also the
>>> rtc_lock.
>>>
>>> Could you give some suggestion? Thanks!
>>
>> efi_rs_enter() and efi_rs_leave() look very PV x86 specific. It is
>> possible that we don't actually need to do anything at all on ARM, aside
>> from refactoring the code. Jan?
>
> I think these functions derive from the USE_SET_VIRTUAL_ADDRESS_MAP option.
> If that is set we call efi_rs->SetVirtualAddressMap to tell the f/w about
> our virtual address layout and can then call rs directly with no faff.
>
> Unfortunately SetVirtualAddressMap is incompatible with kexec (I think you
> can only call it once), so we have this USE_SET_VIRTUAL_ADDRESS_MAP option.
>
> Without USE_SET_VIRTUAL_ADDRESS_MAP we need to switch to a set of page
> tables which are "OK" for calling the runtime services. I'm not sure what
> "OK" means here -- but I suspect it means "1:1" in some part.
>
> So sadly I think we do _eventually_ need to support this mode (for kexec),
> which would mean quite a bit of refactoring (since the relevant code
> in xen/common/efi/boot.c has some x86-isms).
>
> But right now, since we do not yet support kexec,  I think we could get the
> ball rolling wrt RS support by setting USE_SET_VIRTUAL_ADDRESS_MAP on ARM
> and dodging the need to make efi_rs_enter() work right now. (IOW make it
> the problem of whomever wants kexec support...)
>
Yeah, we could set USE_SET_VIRTUAL_ADDRESS_MAP at this moment, but as 
you said, it needs to support kexec and make efi_rs_enter() work 
eventually. So if anyone know this well, please share me your ideas. 
Thanks very much.

> Unless there are issues other than kexec with SetVirtualAddressMap which
> I'm not thinking of now?
>
>> Similarly we don't have an rtc_lock spinlock today and it is not
>> actually needed, but for the sake of keeping the common code common, I
>> would just introduce rtc_lock on ARM too.
>
> Agreed.
>

-- 
Shannon