Re: Utility of PV on non-obsolete hardware

["Mathew, Cherry G." <cherry%zyx.in@localhost>]

On 4 April 2019 4:53:12 PM GMT+05:30, Greg Troxel <gdt%lexort.com@localhost> wrote:
>"Mathew, Cherry G." <cherry%zyx.in@localhost> writes:
>
>> On 4 April 2019 3:06:18 PM GMT+05:30, tg%gmplib.org@localhost wrote:
>
>>>Seriously, plain PV is perhaps only useful on machines without
>EPT/NPT.
>>>Our test systems with Intel Conroe (2006), Intel Penryn (2007), and
>AND
>>>K8 (2003) are the only ones which benefit from PV.
>>
>> That's a good point - I'm surprised about the I/o performance though
>- were you using virtio ?
>>
>>  I'm about 70% of the way enabling PVHVM for domU - we have PoC but
>needs a bit more debugging etc. Also looking at non PV Dom0 solutions.
>
>Is this slowness of PV expected and understood?  Is it about the cost
>of page
>table update hypercalls?  Could it be a difference in using multiple
>cores vs not?

There's several issues with PV - on 64bit CPUs, the kernel shares the same privilege level as userland. This means that a whole bunch of privilege related buffers need to be flushed on every kernel/user boundary transition.

Furthermore the address spaces are completely isolated as opposed to native where the kernel is mapped into userland ASs but gated by hw  privilege mechanisms. On top of this, PV needs hypercalls to update paging related state, which are hw  facilitated on HVM. You'd see a marginal speed up on 32 domU because it doesn't share kernel AS with userland.

So yeah this is known. The only advantage with PV is that you can have virtualisation on CPUs that don't natively support it.

I'm hoping NetBSD will have full HVM container support for domU and dom0 RSN.

Cherry 

-- 
Sent from my phone. Please excuse my brevity.