Re: [libvirt] domXML modeling question

Re: [libvirt] domXML modeling question

[Jim Fehlig]

Adding xen-devel to cc in case anyone there wants to comment on my latest 
proposal...

On 2/20/19 5:20 PM, Jim Fehlig wrote:
> There have been a few requests [1][2] to support Xen's max_grant_frames setting 
> in libvirt domXML, but I'm not quite sure how to model it. The documentation [3] 
> on this setting states:
> 
> Specify the maximum number of grant frames the domain is allowed to have.  This
> value controls how many pages the domain is able to grant access to for other
> domains, needed e.g. for the operation of paravirtualized devices.  The default
> is settable via xl.conf(5).

I've sent a patch to introduce an analogous default in the libvirt libxl driver

https://www.redhat.com/archives/libvir-list/2019-March/msg00123.html

> 
> It smells of a <memtune> setting, e.g. the amount of memory a domain can share, 
> but doesn't map to any of the existing settings. A new subelement <shared_limit> 
> doesn't feel right. Does anyone suggest a better way of modeling max_grant_frames?

After discussing the max_grant_frames setting a bit more with Juergen I had the 
idea to model it as IO buffer space (or DMA space) of a xenbus "controller". All 
PV devices in the guest connect to the xenbus controller and make use of the 
available I/O buffer space. Guests with more PV devices requiring more buffer 
can increase the space on the xenbus controller device.

One small wrinkle in this idea is that we currently don't model xenbus in 
libvirt. I'd need to add support for a new xenbus controller type and start 
implicitly creating it when creating guests with PV devices, similar to 
auto-creation of controllers in the qemu driver. Also, there is no existing 
controller setting for specifying buffer space. Perhaps a 'ram' attribute could 
be added, similar to specifying memory for <video> devices? E.g.

   <controller type='xenbus' ram='256'/>

Any opinion on this approach? Or other ideas for modeling this setting in libvirt?

Regards,
Jim

> 
> Another option I considered is setting the value based on number of PV devices, 
> but I think that flies in the face of libvirt's policy of not dictating policy. 
> Regardless of domain config modeling I can work on a driver-wide setting in 
> libxl.conf, similar to Xen's xl.conf(5) global.
> 
> Regards,
> Jim
> 
> [1] https://www.redhat.com/archives/libvir-list/2018-April/msg00216.html
> [2] https://www.redhat.com/archives/libvirt-users/2019-January/msg00011.html
> [3] 
> http://xenbits.xen.org/gitweb/?p=xen.git;a=blob;f=docs/man/xl.cfg.5.pod.in;h=ad81af1ed8cc983c76b5ec2c3aa02e28f042cc63;hb=HEAD#l569 


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [libvirt] domXML modeling question

[Daniel P. Berrangé]

On Mon, Mar 04, 2019 at 04:42:32PM -0700, Jim Fehlig wrote:
> Adding xen-devel to cc in case anyone there wants to comment on my latest
> proposal...
> 
> On 2/20/19 5:20 PM, Jim Fehlig wrote:
> > There have been a few requests [1][2] to support Xen's max_grant_frames
> > setting in libvirt domXML, but I'm not quite sure how to model it. The
> > documentation [3] on this setting states:
> > 
> > Specify the maximum number of grant frames the domain is allowed to have.  This
> > value controls how many pages the domain is able to grant access to for other
> > domains, needed e.g. for the operation of paravirtualized devices.  The default
> > is settable via xl.conf(5).
> 
> I've sent a patch to introduce an analogous default in the libvirt libxl driver
> 
> https://www.redhat.com/archives/libvir-list/2019-March/msg00123.html
> 
> > 
> > It smells of a <memtune> setting, e.g. the amount of memory a domain can
> > share, but doesn't map to any of the existing settings. A new subelement
> > <shared_limit> doesn't feel right. Does anyone suggest a better way of
> > modeling max_grant_frames?
> 
> After discussing the max_grant_frames setting a bit more with Juergen I had
> the idea to model it as IO buffer space (or DMA space) of a xenbus
> "controller". All PV devices in the guest connect to the xenbus controller
> and make use of the available I/O buffer space. Guests with more PV devices
> requiring more buffer can increase the space on the xenbus controller
> device.
> 
> One small wrinkle in this idea is that we currently don't model xenbus in
> libvirt. I'd need to add support for a new xenbus controller type and start
> implicitly creating it when creating guests with PV devices, similar to
> auto-creation of controllers in the qemu driver. Also, there is no existing
> controller setting for specifying buffer space. Perhaps a 'ram' attribute
> could be added, similar to specifying memory for <video> devices? E.g.
> 
>   <controller type='xenbus' ram='256'/>
> 
> Any opinion on this approach? Or other ideas for modeling this setting
> in libvirt?

Regardless of max grant frames support I think modeling xenbus as a
<controller> is a reasonable thing to want to do. I don't have a
preference on whether you call it "ram" or explicitly "maxGrantFrames"
as an attribute.


Regards,
Daniel
-- 
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel