Re: [Xen-devel] Libvir: a simple C virtualization control library

[Daniel Veillard <veillard@xxxxxxxxxx>]

On Sun, Dec 18, 2005 at 01:41:03PM -0600, Anthony Liguori wrote:
> Hi Daniel,

  Hi Anthony,

> I think the goal of having a library that allows applications to 
> interact with Xen is a great idea.

  thanks !

> I suggest though that instead of taking the current approach of 
> reimplementing a hypercall library, you consider just using Xend's HTTP 
> interface.
> 
> xm/xend communicate via an web services API that's based on 
> S-Expressions.  Xend can listen on a domain socket and/or a TCP socket 
> for incoming connections so both transports should be supported.
> 
> You can access this interface by doing an HTTP request with the 
> Content-Type set to application/sxp.  The protocol uses the URL to 
> identify the command and options and returns s-expressions.  You can get 
> a feeling for the supported commands by enabling the TCP server and 
> going to http://localhost:8000/xend/domain/

  Okay, I will check this specific API too, thanks :-) !

  One can take a technical look, there is pros and cons for both approaches,
for example querying another process though localhost TCP + HTTP + Python
marshalling + python interpreter and back to get state informations for
a domain when this could be a simple hypervisor call in pure C, well
the HTTP RPC approach does not sound fantastic. When creating a new domain,
I agree that the performance aspect is neglectible. From a technical view
again using directly the hypercalls introduce an ABI dependancy, which the
Xen authors may or may not want to garantee at this point.

  Then there is non-technical viewpoint, I am personally (I don't represent
my employer here) uneasy with a framework purely GPLed in user land
especially as a lot of the work is being done in user space. To me GPL is fine
for completely self-contained work with standard open APIs like a kernel can
be, the competition of ideas and implementations (which is the core factor
why OSS/free software end up being the best) happens either inside the project
or between projects reimplementing the standard APIs (e.g. Linux/BSD/Solaris).
But having a project GPLed without standard APIs hence creating an unicity of
the implementation by the nature of the licence I don't feel great about.
I understand the motivation to drive people to collabrate on a single
implementation when the project is bootstrapping, but using the licence as
a mechanism for this sounds wrong to me, Xen 3.0 and upcoming integration
in Linux main kernel show the project is getting ready to become mainstream,
and now is time to define the standard APIs and make sure the licence don't
constrain the adoption.

  Not having contributed yet to Xen, my viewpoint can be seen as presumptuous,
I am sorry for this, I recognize the huge amount of work which has been done
and is still being done, but ideally libvir should not exist because the 
problems of a standard, easy to reuse API for Xen should not exist. I don't
think a GPL library is an answer, Xen and virtualization in general will
be used by a variety of tools, and I want to see an healthy competition
of project bringing tools to the users, and not just corporate, cluster users.
I hope virtualization will reach the masses, and will integrate seamlessly
in the user's desktop, that's why I think Xen is fantastic, and why I think
we should not bet on a single collection of tools, getting there will require
open APIs and competing open-source projects building those users tools.

> This would allow you to have an LGPL control library yet not rewrite all 
> of Xend.  It also requires minimal dependencies since I believe libxml2 
> already has an HTTP client (though you'd have to add a domain socket 
> transport).

   libvir doesn't require libxml2 :-) the dependancy may be added later
if interfaces using XML as input get exposed though. There is also all
the security and reliability aspects of opening up an HTTP server with
side effect (even if limited to localhost) especially when it's is exposing
a full programming environment.

> An interesting thing to do too would be to attempt to add support to 
> Xend for another mime type (like text/xml) and perhaps even support 
> XML-RPC.  This would be really excellent long term as it could eliminate 
> a ton of code in Xend and xm by just reusing the python XML-RPC support.

  Using XML would force to tighten up some grey area, like for example
what is the encoding of a domain name string in a /etc/xen config file
(it's python so all bets are off at the moment). But this would add one more
layer of processing and dependancies. For something like creating a domain
or querying/changing long lived settings that would be just fine, but for
anything like monitoring this could be just too much.

   Thanks for the suggestion again, yes I will look at this too, this may
be the best technical solution in the current situation :-)

Daniel

-- 
Daniel Veillard      | Red Hat http://redhat.com/
veillard@xxxxxxxxxx  | libxml GNOME XML XSLT toolkit  http://xmlsoft.org/
http://veillard.com/ | Rpmfind RPM search engine http://rpmfind.net/

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel