Re: hardwired VMI crap

[Jeremy Fitzhardinge <jeremy@goop.org>]

Ingo Molnar wrote:
>  - /One/ _intelligent_ higher-level virtualization API/ABI. Xen's API is 
>    quite advanced on this front.

At last!  Some love!

The Xen approach has always been to prefer high-level interfaces over
lower-level ones, so that guests can meaningfully participate in their
own virtualization.  There are some necessarily low-level things, but
conceptually simple things like "create a new vcpu" should have simple
interfaces.  There's no point in going to the effort of emulating a
whole pile of real hardware if Xen can present an interface which is a
close match to an existing high-level interface within the operating system.

>  This would be shared by /all/ 
>    hypervisors and could occasionally be reused by other hardware 
>    platforms as well. It can be morphed via small wrappers into some
>    'hypervisor personality' kind of hypervisor backends, but no
>    fundamental transformation happens.
>   

Because of both general conceptual cleanliness and Xen's requirements,
the pv_ops interface has tended towards fairly high-level interfaces
where possible and useful.

VMI's design tends towards being a closer match to some approximation of
the real underlying hardware, which I suppose is a reflection of its
origins as an extension of a fully virtualizing hypervisor.  I don't
have any particular problem with that, and I think its a perfectly
reasonable approach if that's the path you want to take.

What this means, however, is that the existing arch/i386 code needs to
be refactored so that VMI can reuse it in a sensible way, so that it can
implement the high-level operations in terms of the existing building
blocks that running on bare hardware has to use anyway.

Things like genirq and genapic should help with that in principle. 
genirq certainly cleaned up Xen's interface to the interrupt subsystem. 
I haven't yet looked at genapic, but from what Zach and Chris Wright
have said, I get the impression that it isn't yet up to meeting our
requirements.

The result of this tension has been a general desire to keep pv_ops a
high-level interface, but for pragmatic reasons it has grown a few
low-level operations (like the apic read and write interfaces) which
better suit VMI's needs.  I would like to see these go away as they get
replaced with high-level interfaces. 

I think this is the natural course we've been following anyway.  Xen has
the advantage of starting from a relatively clean slate here, but
reusing existing entrenched pieces cleanly in new ways always takes
careful thought and hard work.

> what we do _NOT_ want is some mixture of 'simplified' and 'hardwired' 
> native hardware access mixed with hypercalls that somehow ends up 
> creating a Frankenstein mixture of 'virtual silicon', which is specified 
> nowhere else but in VMWare's proprietary hypervisor source code that we 
> have no way to fix and no way to even see!

No, but I'm not prejudiced against virtual hardware.  If we have a piece
of code that thinks its talking to an apic, then I think its OK to use
that code whether its a real apic or a virtual one, _so long as its
being used in a way that's consistent with its intended interface_.  I
have to admit I have not looked at apics - real or virtual - in any
detail, so I won't claim to really understand the details of the
existing arch/i386 code or what VMI's trying to do, but it does seem to
me that it could all be much cleaner.

And clean is good, we all love clean - and so, agreement!

    J