Re: future plans for libxl?

[Andre Przywara <andre.przywara@amd.com>]

Vincent Hanquez wrote:
> On 15/04/10 08:21, Andre Przywara wrote:
>> I stumbled upon missing 'info' support, so I implemented a basic version
>> of it. A few questions about it:
>> 1.) Is there a fixed interface for libxenlight? I see libxc interfaces
>> duplicated (like {xc,libxl_get}_physinfo), is that thought to provide a
>> more stable interface than xc does (which changed quite a bit lastly).
>> What about extending this interface? For complete xl info I need more
>> field of the physinfo sysctl to be provided by libxl_get_physinfo, is it
>> OK to extend the struct libxl_physinfo?
> 
> yes, i've duplicated them for this exact purpose. it could work 
> hand-in-hand with what i'm describing below.
> 
> for extending usually, you would just add a new field at the end of the
> structure.
> 
> it's probably ok to extends the physinfo structure. however we tried
> to separate some things that would be not very useful from a toolstack
> point of view, compared to actually carrying all sorts of info of
> questionable usefulness.
OK, what is missing currently is nr_nodes, {hw,virt}_caps and the whole 
bunch of (currently in flux) NUMA information. I think the whole NUMA 
info should be moved into either a topology or a NUMA structure 
(together with libxl_get_... functions). But I'd like to see at least 
nr_nodes and the caps within the physinfo structure. One can debate 
about the nr_nodes fields, though.

What about xc_version(), by the way? This provides a lot of information 
currently shown in xm info, I didn't found any interface for that in 
libxenlight. Is it OK to implement it? Or shall I call xc_version() 
directly from xl.c?

>> 2.) I see a simple check of LIBXL_VERSION when doing libxl_ctx_init().
>> Is that meant to be that hard or is a compatibility scheme planned
>> (like: support apps using on older interface and somehow emulate it?)
> 
> it takes lots of resources to do a compat scheme. but eventually it 
> could all be done with the LIBXL_VERSION check. if you upgrade
> the library and something change, the previously compiled client
> would still use a LIBXL_VERSION < compared to the current one in the 
> library.
> 
> at this point, libxl could provide a mechanism to provide old -> new 
> structure mechanism to provide automatic compatibilty with old 
> userspace. note that it wouldn't cover function prototype change, but 
> only structure change, and this is an extremly painful process
> where you need to keep your old structure around, and have a 
> (potentially) big fast switch case in function that need this compat layer.
QEMU implements something in it's migration interface to support 
different version of the emulated devices. It's easier there though 
because they read each member explicitly from the stream and don't need 
to fiddle around with mapping structs.

> it's much much simpler to just let client and library to be the exact 
> same version, which is what the ctx_init function is doing at this point 
> compared to the elaborate scheme of compat.
I agree that the whole effort is probably not worth the gains. 
Especially since we have only one library one could update this together 
with the tools. If we came to some kind of stable interface, one could 
use the standard UNIX library versioning scheme.
The only question is if we can keep pace with the hypervisor 
development. If it introduces new features, hypercalls, extended 
structures do we really want to break compatibility or do we refrain 
from implementing new features?
I suppose that at least in the -unstable tree we don't care about 
compatibility and only keep a stable interface in the -testing trees, 
the libxl version could then be named after the stable hypervisor version.

Regards,
Andre.

-- 
Andre Przywara
AMD-Operating System Research Center (OSRC), Dresden, Germany
Tel: +49 351 448-3567-12