Re: efi_enabled(EFI_PARAVIRT) use

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

On 2016年04月29日 22:53, Ard Biesheuvel wrote:
> On 29 April 2016 at 16:39, Matt Fleming <matt@codeblueprint.co.uk> wrote:
>> On Fri, 29 Apr, at 11:34:45AM, Stefano Stabellini wrote:
>>> On Fri, 29 Apr 2016, Ingo Molnar wrote:
>>>> Also, it would be nice to have all things EFI in a single tree, the conflicts are
>>>> going to be painful! There's very little reason not to carry this kind of commit:
>>>>
>>>>  arch/arm/xen/enlighten.c           |  6 +++++
>>>>  drivers/firmware/efi/arm-runtime.c | 17 +++++++++-----
>>>>  drivers/firmware/efi/efi.c         | 45 ++++++++++++++++++++++++++++++++------
>>>>  3 files changed, 56 insertions(+), 12 deletions(-)
>>>>
>>>> in the EFI tree.
>>>
>>> That's true. I'll drop this commit from xentip and let Matt pick it up
>>> or request changes as he sees fit.
>>
>> One small change I think would be sensible to make is to expand
>> EFI_PARAVIRT into a few more bits to clearly indicate the quirks on
>> Xen, and in the process, to delete EFI_PARAVIRT.
>>
Sure. How should I add this change? Rework this patch or add new one on
top of it?

>> That should address Ingo's major concern, and also make it much easier
>> to rework the code in a piecemeal fashion.
>>
>> Could somebody enumerate the things that make Xen (dom0) different on
>> arm* compared with bare metal EFI boot? The list I made for x86 was,
>>
>>   1. Has no EFI memory map
>>   2. Runtime regions do not need to be mapped
>>   3. Cannot call SetVirtualAddressMap()
>>   4. /sys/firmware/efi/fw_vendor is invisible
>>
>> The first maps to not setting EFI_MEMMAP, the second to not setting
>> EFI_RUNTIME. If we add EFI_ALREADY_VIRTUAL and EFI_FW_VENDOR_INVISIBLE
>> to efi.flags that should cover everything on x86. Does arm* require
>> anything else?
> 
> I already proposed when this patch was first under review to make the
> arm_enable_runtime_services() function bail early without error if the
> EFI_RUNTIME_SERVICES flag is already set, and the xen code could set
> that bit as well when it installs its paravirtualized alternatives. I
> don't remember exactly why that was shot down, though, but I think it
> is the only reason this code introduces references to EFI_PARAVIRT in
> the first place.
> 
Yes, in this patch we could set EFI_RUNTIME_SERVICES flag in
fdt_find_hyper_node instead of setting EFI_PARAVIRT flag, and then bail
out early in arm_enable_runtime_services() as you said. Then call
xen_efi_runtime_setup() in xen_guest_init().

While I still have a question, in this patch we use
efi_enabled(EFI_PARAVIRT) as a condition to make fdt_find_uefi_params()
and efi_get_fdt_params() execute different ways. So it needs to find a
new condition for that if we need to get rid of EFI_PARAVIRT. One I
think is that xen_initial_domain() check. Is that fine?

Thanks,
-- 
Shannon