Re: [Minios-devel] Single codebase for para-virtualized guests on x86 and ARM?

[Andre Przywara <andre.przywara@xxxxxxx>]

Hi,

On 05/04/18 09:49, Julien Grall wrote:
> 
> 
> On 05/04/18 09:38, Ajay Garg wrote:
>> Hi Julien.
>>
>> Thanks for the reply.
>>
>> On Thu, Apr 5, 2018 at 1:21 PM, Julien Grall <julien.grall@xxxxxxx>
>> wrote:
>>> (+ Stefano and Andre)
>>>
>>> On 04/05/2018 04:35 AM, Ajay Garg wrote:
>>>>
>>>> Hi Julien.
>>>>
>>>> Sorry for restarting this thread, but I was just wondering whether PV
>>>> on ARM might not provide the following benefits :
>>>>
>>>>      * Low overheads as compared to HVM.
>>>
>>>
>>> I don't buy this argument. The overhead is now quite high because
>>> with a PV
>>> solution as you would need to mitigate meltdown.
>>>
>>> Furthermore, this require a lot of changes in your guests and I don't
>>> see
>>> any OS community accepting such burden when virtualization extension is
>>> getting quite widespread on Arm.
>>>
>>>>      * Also support systems/procedures with no hardware-virtualization
>>>> support.
>>>
>>>
>>> There was an attempt to get PV support for Arm before the virtualization
>>> support was added ([1]). But this is dead.
>>>
>>> Do you have any platform in mind you would want to run Xen with no
>>> virtualization extension?
>>
>> Ideally, we are targetting low-end arm systems, where
>> hardware-virtualization is the one thing that we might have to
>> compromise on.
>> Beaglebones are an example.
> 
> They are low-end platform with hardware-virtualization (e.g pine64).
> So why that platform in particular? Is it because you have existing
> hardware?

As Julien said, the question is whether older platforms using an
Cortex-A8 or A9 are really worth the porting effort. Yes, the BeagleBone
is nice, but there are tons of cheap A7 and A53 based systems out there,
for instance this OrangePi Zero Plus[1] (Quad-A53) for 15 US-$. A7 based
boards go for as low as 10$.

> To be clear, I think compromising virtualization extension is a pretty
> bad idea. The burden on software will really be huge compare to the
> current solution on Xen Arm.

Yes, I fully agree. Having hardware 2-level paging for instance
simplifies a lot.
If you have such an older low-end system, then a hypervisor might not be
the best choice in the first place (regardless of virt ext or not).

Cheers,
Andre.

[1] http://www.orangepi.org/OrangePiZeroPlus

> 
> Cheers,
> 
>>
>>
>>>
>>>>
>>>>
>>>> Of course, the benefits of PV need to be really high, to justify the
>>>> efforts needed to make the required code-changes for PV-support :P
>>>
>>>
>>> HW virtualization support provides good security and performance with
>>> limited changes in the guest. Even, x86 community has been discussing to
>>> deprecate PV and move towards PVH (very similar to Arm guest).
>>>
>>> Cheers,
>>>
>>> [1] https://wiki.xenproject.org/wiki/Archived/Xen_ARM_(PV)
>>>
>>> -- 
>>> Julien Grall
>>
>>
>>
> 

_______________________________________________
Minios-devel mailing list
Minios-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/minios-devel