Re: [RFC PATCH] KVM: x86: Allow Qemu/KVM to use PVH entry point

From: Boris Ostrovsky
Date: Wed Nov 29 2017 - 09:26:24 EST


On 11/29/2017 09:18 AM, Roger Pau Monné wrote:
> On Wed, Nov 29, 2017 at 03:11:12PM +0100, Juergen Gross wrote:
>> On 29/11/17 15:03, Boris Ostrovsky wrote:
>>> On 11/29/2017 03:50 AM, Roger Pau Monné wrote:
>>>> On Wed, Nov 29, 2017 at 09:21:59AM +0100, Juergen Gross wrote:
>>>>> On 28/11/17 20:34, Maran Wilson wrote:
>>>>>> For certain applications it is desirable to rapidly boot a KVM virtual
>>>>>> machine. In cases where legacy hardware and software support within the
>>>>>> guest is not needed, Qemu should be able to boot directly into the
>>>>>> uncompressed Linux kernel binary without the need to run firmware.
>>>>>>
>>>>>> There already exists an ABI to allow this for Xen PVH guests and the ABI is
>>>>>> supported by Linux and FreeBSD:
>>>>>>
>>>>>> https://xenbits.xen.org/docs/unstable/misc/hvmlite.html
>>>> I would also add a link to:
>>>>
>>>> http://xenbits.xen.org/docs/unstable/hypercall/x86_64/include,public,arch-x86,hvm,start_info.h.html#Struct_hvm_start_info
>>>>
>>>>>> This PoC patch enables Qemu to use that same entry point for booting KVM
>>>>>> guests.
>>>>>>
>>>>>> Even though the code is still PoC quality, I'm sending this as an RFC now
>>>>>> since there are a number of different ways the specific implementation
>>>>>> details can be handled. I chose a shared code path for Xen and KVM guests
>>>>>> but could just as easily create a separate code path that is advertised by
>>>>>> a different ELF note for KVM. There also seems to be some flexibility in
>>>>>> how the e820 table data is passed and how (or if) it should be identified
>>>>>> as e820 data. As a starting point, I've chosen the options that seem to
>>>>>> result in the smallest patch with minimal to no changes required of the
>>>>>> x86/HVM direct boot ABI.
>>>>> I like the idea.
>>>>>
>>>>> I'd rather split up the different hypervisor types early and use a
>>>>> common set of service functions instead of special casing xen_guest
>>>>> everywhere. This would make it much easier to support the KVM PVH
>>>>> boot without the need to configure the kernel with CONFIG_XEN.
>>>>>
>>>>> Another option would be to use the same boot path as with grub: set
>>>>> the boot params in zeropage and start at startup_32.
>>>> I think I prefer this approach since AFAICT it should allow for
>>>> greater code share with the common boot path.
>>> zeropage is x86/Linux-specific so we'd need some sort of firmware (like
>>> grub) between a hypervisor and Linux to convert hvm_start_info to
>>> bootparams.
>> qemu?

I think KVM folks didn't want to do this. I can't find the thread but I
believe it was somewhere during Clear Containers discussion. Paolo?


> But then it won't be using the PVH entry point, and would just use the
> native one?
>
> My understanding was that the PVH shim inside of Linux will prepare a
> zero-page when booted using the PVH entry point, and then jump into
> the native boot path.

Right, but that's not what Juergen's second option is. IIUIC with that
option Linux starts with zeropage already prepared. No shim in the kernel.

-boris