Re: [Qemu-devel] [PATCH] vl.c: make sure maxcpus matches topology to prevent migration failure

[Eduardo Habkost <ehabkost@redhat.com>]

On Fri, Aug 24, 2018 at 11:13:50AM +0200, Igor Mammedov wrote:
> On Thu, 23 Aug 2018 18:32:41 +0200
> Paolo Bonzini <pbonzini@redhat.com> wrote:
> 
> > On 23/08/2018 16:51, Igor Mammedov wrote:
> > > Topology (threads*cores*sockets) must match maxcpus to be valid,
> > > otherwise we could start QEMU with invalid topology that throws
> > > a error on migration destination side, that should not be reachable:
> > > Source:
> > >   -smp 8,maxcpus=64,cores=1,threads=8,sockets=1
> > > // hotplug cpus upto maxcpus
> > > Destination:
> > >   -smp 64,maxcpus=64,cores=1,threads=8,sockets=1
> > >   qemu: cpu topology: sockets (1) * cores (1) * threads (8) < smp_cpus (64)
> This destination CLI aren't exactly correct as well since
> it should've been exactly the same -smp as on source + a bunch of -device cpufoo...
> so we can always say go fix your CLI so it won't trigger error.
>   
> > The destination should have sockets=8, shouldn't it?
> either that or cores=8 or cores=4,sockets=2 ...
>  
> > It seems to me that, at startup, you should have cpus = s*t*c and cpus
> > <= maxcpus.  Currently we check cpus <= s*t*c <= maxcpus, which doesn't
> > make much sense.
> I think that s*t*c should describe topology of whole machine
> including not yet plugged vcpus. "cpus = s*t*c" probably won't work
> for partially filled package case:
>        -smp 1,cores=1,threads=8,sockets=1
> cores/threads should reflect full package configuration
> for guest to see an expected topology.

Oh, now I remember: that's the reason we don't enforce
s*t*c == smp_cpus nor s*t*c == max_cpus.

Both "-smp 4,maxcpus=8,cores=2,threads=2,sockets=1" and
     "-smp 4,maxcpus=8,cores=2,threads=2,sockets=2"
worked since maxcpus was introduced, making the semantics of
"sockets" unclear and hard to change without breaking existing
configs.

-- 
Eduardo