Re: [vmw_vmci RFC 00/11] VMCI for Linux

From: Dor Laor
Date: Thu Jun 14 2012 - 07:52:13 EST

On 06/06/2012 08:06 AM, Greg KH wrote:
On Tue, Jun 05, 2012 at 12:02:51AM -0700, Dmitry Torokhov wrote:
Hi Greg,

On Mon, Jun 04, 2012 at 03:57:57PM -0700, Greg KH wrote:
On Fri, Jun 01, 2012 at 08:33:02AM -0700, Andy King wrote:
Greg,

Thanks so much for the comments and apologies for the delayed response.

Don't we have something like this already for KVM and maybe Xen?
virtio? Can't you use that code instead of a new block of code that
is only used by vmware users? It has virtual pci devices which
should give you what you want/need here, right?

If not, why doesn't that work for you? Would it be easier to just
extend it?

The VMCI virtual device for which this driver is intended has been
around a lot longer than this submission might suggest. The virtual
hardware was released in a product before Rusty sent his RFC and
quite a bit before it made it to mainline; there was, regrettably,
no virtio then.

As such, it was designed to be its own transport, and it's something
that is now very much fixed at the hardware level (enhancements
not withstanding), and which we have to support all the way back.

What "hardware" are you refering to here?

The virtual hardware that is currently shipping and has been shipping
for a few years.


In addition to that, our hypervisor endpoints are written using
the existing device backend; virtio doesn't currently make a lot of
sense for them, and would require a lot of additional work.

All of this is unfortunate. While I agree that virtio is certainly
the right approach, and we need to avoid this proliferation, I think
at this point we'd really like to try and upstream this in its current
form. There's certainly the possibility going forwards that we could
add a glue layer, such that other clients could use virtio if they're
willing to write their own hypervisor endpoints.

Does that sound reasonable?

Not really, why should we take an interface that is tied to something
that you are saying isn't something we should be using?

That is not what Andy said. If virtio was available when we started
shipping VMCI then we certainly could have used that, but since it
wasn't there we invented something else.

Ok, that makes sense.

Don't you also
have control over the hypervisor side of things in order to properly
design this type of thing?

We do not have a time machine to go back and change products that we
already shipped to the customers. It is probably the same story as with
Hyper-V's vmbus which is not virtio.

Besides, virtio is not available on non-Linux guests with we have to
support as well, and than affected the design decisions in hypervisor
layer that have been made several years ago.

Indeed there is no time machine and I definitely agree you won't be able to port VMCI to virtio-serial w/o major breakage for existing users.

Nevertheless, as I wrote on my previous feedback [1], you can make the VMCI socket generic in a way that virtio-serial users would be able to enjoy. This way VMW users will get out of the box experience w/o any change while KVM users will get socket abstraction layer which carry some benefit over the virtio-serial ports.

Dor

[1] https://lkml.org/lkml/2012/5/16/126


Ok, thanks for clearing that up, I was confused here.

greg k-h
_______________________________________________
Virtualization mailing list
Virtualization@xxxxxxxxxxxxxxxxxxxxxxxxxx
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/