Re: [Qemu-devel] [RFC 4/8] qdev: link based hotplug

["Michael S. Tsirkin" <mst@redhat.com>]

On Mon, Mar 24, 2014 at 01:20:46PM +0100, Andreas Färber wrote:
> Am 21.03.2014 11:35, schrieb Igor Mammedov:
> > BTW not related to hotplug but why I used link<>s:
> > 
> > I've added link<>s as an attempt to visualize Andreas' idea to use them for
> 
> Anthony's :)
> 
> > hotplug and mgmt. It has it's own benefits if we try to provide more or
> > less uniform QOM interface view for management. What I have in mind is that
> > we could have tree like this:
> >  /machine/node[...]/dimm[...]
> >                    /cpu[...]/core[...]/thread[...]
> > 
> > where leaves are link<>s which are prebuilt at startup and set when device
> > is added. It provides an easy to enumerate interface for mgmt and also
> > gives us a quite informative path that encodes topology and later
> > we could use it instead of custom properties. For example:
> > 
> >   device_add x86cpu,path=/machine/node[1]/cpu[0]/core[3]/thread[2]
> > vs
> >   device_add x86cpu,apic-id=[who knows how it's calculated]
> 
> This still collides with what Anthony and me have been saying about CPU
> hot-add: It should not happen on thread level. cpu-add covers the
> current approach, but device_add should add a full socket and definitely
> not in that /machine/node[n]/... path. You or someone replied on Paolo's
> NUMA thread that this was only as an internal convenience for lookups,
> now you're pushing this as user ABI. NACK to the latter.
> 
> I was hoping that we could let device_add auto-discover free link<>
> slots like it does for buses today;

That's friendly for HMP but a huge maintainance pain.
In the end hardly anyone uses HMP for hotplug so I
don't think there's a lot of value in auto-discovery,
make management as explicit as possible.

> having two places to link the same
> object would complicate that, so we need to be careful in designing our
> link<>s: Having /machine/node[0]/cpu[0] be a link<> would mean no second
> link<> directly on /machine/cpu[0], i.e. the NUMA node acting as a
> sub-container of the board; on the other hand having a link<> to the
> thread CPUState from some NUMA-specific path outside of the composition
> model would still be possible due to the different link<> type but
> having cpu[0] be a link to the actual socket object rules out using the
> same name for a NUMA-only container object as proposed.
> 
> > or
> >   device_add dimm,path=/machine/node[0]/dimm[5]
> > vs
> >   device_add dimm,node=0,slot=5
> > 
> > i.e. being added device could decode all needed information from above
> > provided path instead of creating a bunch of custom properties.
> 
> Hm, the advantage of having properties there would be better error
> checking in terms of restricting paths. I'd be open to having both.
> 
> I do wonder in both cases if we should use paths relative to /machine to
> avoid them cluttering the command line?
> 
> Regards,
> Andreas
> 
> -- 
> SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg, Germany
> GF: Jeff Hawn, Jennifer Guild, Felix Imendörffer; HRB 16746 AG Nürnberg