Re: [Qemu-devel] [PATCH 1/3] qapi: Report support for -device cpu hotplug in query-machines

[Peter Krempa]

On Fri, Jun 24, 2016 at 16:56:21 +1000, David Gibson wrote:
> On Fri, 24 Jun 2016 07:41:11 +0200
> Peter Krempa <address@hidden> wrote:
> 
> > On Fri, Jun 24, 2016 at 14:56:51 +1000, David Gibson wrote:
> > 
> > [...]
> > 
> > > > You are correct - query-commands says whether 'query-hotpluggable-cpus'
> > > > exists as a command.  But that is insufficient.  See my review, or the
> > > > v2 patch, where the above poor wording was corrected to say what was
> > > > really meant: knowing whether query-hotpluggable-cpus exists is
> > > > insufficient to tell you whether a given cpu type can be hotplugged.  So
> > > > adding one more piece of witness (for every type of cpu supported, we
> > > > also advertise if it is hotpluggable) is enough for libvirt to
> > > > efficiently take advantage of the new query-hotpluggable-cpus command.  
> > > 
> > > Ah, right.  Or to put it another way, the availability of
> > > query-hotpluggable-cpus is global across qemu, whereas actually being
> > > able to use it for hotplug is per machine type.
> > > 
> > > Would it be possible to do this instead by attempting to invoke
> > > query-hopluggable-cpus and seeing if it returns any information?  
> > 
> > It is not strictly necessary for us to have this in the context of
> > usability. If the user requests using the new hotplug feature we will
> > try it unconditionally and call query-hotpluggable-cpus before even
> > starting guest execution. A failure to query the state will then result
> > in termination of the VM.
> 
> Oh.. I wasn't expecting the feature would be enabled at user request -
> I thought libvirt would just use it if available.

Hmm, I think that will be possible to use the feature all the time after
all. I'll need to add multiple states for that though to track it in
the XML so that we can re-create it on migration the right way.

> > It is necessary though to report the availability of the feature to the
> > user via our domain capabilities API which some higher layer management
> > apps use to make decisions.
> 
> Right... what advantage does adding the machine flag have over
> attempting the query-hotpluggable-cpus?

Mostly the ability for oVirt or open stack being able to query whether
it's available for given machine type prior to even attempting to launch
the VM. Otherwise the'll be left to either guessing or re-implementing
the support matrix.

This means that the ability to query it with machine types is not
strictly necessary for implementing the feature in libvirt but a
very-nice-to-have thing to add to our capabilities queries.

> > This would also be necessary if we wanted to switch by default to the
> > new approach, but that's not really possible as libvirt tries to
> > guarantee that a config valid on certain version will be still valid
> > even when it was migrated to a newer version and then back.
> 
> Sorry, I've lost track of what the "This" is that would be necessary.

Never mind on this point. If I design the XML a bit differently it will
be possible to use this this feature if present all the time. Some cpus
just won't be available for unplug.

> > My current plan is to start qemu with -smp cpus=1,... and then call
> > query-hotpluggable-cpus and then hotplug all of them until the requested
> > configuration is satisfied. This approach is necessary so that we can
> > query for the model and topology info so that we don't need to
> > re-implement all the numbering and naming logic from qemu.
> 
> Um.. why?  What's the problem with just staring with -smp cpus=whatever
> and then using query-hotpluggable-cpus?
> 
> > Additionally this will require us to mark one CPU as non-hotpluggable as
> > -smp cpus=0,maxcpus=10 is basically translated to -smp
> > cpus=10,maxcpus=10.
> 
> The latter is true, but I'm not clear why it implies the former.

It necessarily doesn't imply it. I originally wanted to have most cpus
unpluggable but thinking again it is not really necessary I just need to
adapt the interface design.

Let's see how it turns out.

Peter