Re: [Qemu-devel] [RFC PATCH v2 0/9] Core based CPU hotplug for PowerPC sPAPR

[David Gibson]

On Mon, Mar 21, 2016 at 11:43:34AM +0100, Igor Mammedov wrote:
> On Mon, 21 Mar 2016 14:57:53 +1100
> David Gibson <address@hidden> wrote:
> 
> > On Fri, Mar 18, 2016 at 08:59:32AM +0530, Bharata B Rao wrote:
> > > On Thu, Mar 17, 2016 at 09:03:43PM +1100, David Gibson wrote:  
> > > > On Wed, Mar 16, 2016 at 04:48:50PM +0100, Igor Mammedov wrote:  
> > > > > On Wed, 16 Mar 2016 09:18:03 +0530
> > > > > Bharata B Rao <address@hidden> wrote:
> > > > >   
> > > > > > On Mon, Mar 14, 2016 at 10:47:28AM +0100, Igor Mammedov wrote:  
> > > > > > > On Fri, 11 Mar 2016 10:24:29 +0530
> > > > > > > Bharata B Rao <address@hidden> wrote:
> > > > > > >     
> > > > > > > > Hi,
> > > > > > > > 
> > > > > > > > This is the next version of "Core based CPU hotplug for PowerPC 
> > > > > > > > sPAPR" that
> > > > > > > > was posted at
> > > > > > > > https://lists.gnu.org/archive/html/qemu-ppc/2016-03/msg00081.html
> > > > > > > > 
> > > > > > > > device_add semantics
> > > > > > > > --------------------
> > > > > > > > For -smp 16,sockets=1,cores=2,threads=8,maxcpus=32
> > > > > > > > (qemu) device_add 
> > > > > > > > spapr-cpu-core,id=core2,core=16,cpu_model=host[,threads=8]    
> > > > > > > do you plan to allow user to hotplug different cpu_models?
> > > > > > > If not it would be better to hide cpu_model from user
> > > > > > > and set it from machine pre_plug handler.    
> > > > > > 
> > > > > > In my earlier implementations I derived cpu model from -cpu and 
> > > > > > threads from
> > > > > > -smp,threads= commandline options and never exposed them to 
> > > > > > device_add
> > > > > > command.
> > > > > > 
> > > > > > Though we don't support heterogenous systems (different cpu models 
> > > > > > and/or
> > > > > > threads) now, it was felt that it should be easy enough to support 
> > > > > > such
> > > > > > systems if required in future, that's how cpu_model and threads 
> > > > > > became
> > > > > > options for device_add.
> > > > > > 
> > > > > > One of the things that David felt was missing from my earlier QMP 
> > > > > > query
> > > > > > command (and which is true in your QMP query implementation also) 
> > > > > > is that
> > > > > > we aren't exporting cpu_model at all, at least for not-yet-plugged 
> > > > > > cores.
> > > > > > So should we include that or let management figure that out since it
> > > > > > would already know about the CPU model.  
> > > > > 1.
> > > > > so since you are not planning supporting heterogeneous setup yet,
> > > > > I'd suggest to refrain from making user to provide cpu_model at
> > > > > device_add time. Instead make machine code to set it for cores it
> > > > > creates before core.realize() (yet another use for pre_plug()).
> > > > > 
> > > > > That way mgmt doesn't have to figure out what cpu_model to set at
> > > > > device_add time and doesn't have find out what property to use for 
> > > > > it.  
> > > > 
> > > > Yes.. of course you could also do the same thing for nr_threads, so
> > > > I'm wondering whether there's a good argument to keep one in
> > > > pre_plug() and one in query-hotpluggable-cpus.  
> > > 
> > > Right, so what should be the way forward ? Should we keep cpu_model= and
> > > threads= options with device_add or just threads=  or neither ?  
> > 
> > I'm inclined to keep them both in device_add - I like the idea of
> > having an example on day 0 of advertising extra properties (beyond
> > nr_threads and location) to set from query-hotpluggable-cpus.
> > 
> > But, I'd probably change my mind if Igor or someone has a stronger
> > opinion.
> I don't have a strong opinion on this, but you have to keep in mind
> that one you make it ABI you probably would have to maintain it forever.
> 
> So far 'threads' and 'cpu_model' look like a constant values,
> fixed at start-up time for every core.
> Taking in account that user is not supposed to change them during
> hotplug time and that they are the same for every core,
> I'd go for conservative route and hide them in pre_plug() for now.
> You always can expose them later if needed.

Hm, yes, I see your point.  Hiding them in pre-plug does also make
life more convenient for someone (not libvirt) manually experimenting
with this - less to type on their device-add command.

> > If we advertise cpu_model, however, it should probably be changed to
> > cpu thread class name, since IIUC that's an existing advertised part
> > of the QOM interface, but cpu_model isn't.
> I still think that spapr-cpu-core should be an abstract type
> with a concrete set of derived cores types per each thread type.
> But this question is not related to hotplug, but rather to
> start-up of QEMU from scratch with -device and supported types
> discovery. So I'd postpone question for later and that's yet another
> reason why I'd like to hide cpu_model from user for now.

Ah.. yes, I think you're probably right, and we should have derived
types.  It's a little bit awkward for spapr, but we can probably do
some macro magic to make it not too bad.

-- 
David Gibson                    | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au  | minimalist, thank you.  NOT _the_ _other_
                                | _way_ _around_!
http://www.ozlabs.org/~dgibson

signature.asc
Description: PGP signature