Re: [Qemu-devel] [PATCH] pc: make sure that guest isn't able to unplug the first cpu

[Igor Mammedov <imammedo@redhat.com>]

On Tue, 28 Aug 2018 10:52:37 +1000
David Gibson <david@gibson.dropbear.id.au> wrote:

> On Mon, Aug 27, 2018 at 04:02:39PM +0200, Greg Kurz wrote:
> > On Mon, 27 Aug 2018 13:07:10 +0200
> > Igor Mammedov <imammedo@redhat.com> wrote:
> >   
> > > The first cpu unplug wasn't ever supported and corresponding
> > > monitor/qmp commands refuse to unplug it. However guest is able
> > > to issue eject request either using following command:
> > >   # echo 1 >/sys/devices/system/cpu/cpu0/firmware_node/eject  
> > 
> > I can't reproduce the issue with a pc guest and current master...
> > 
> > All I seem to get is an error in dmesg:
> > 
> > [   97.435446] processor cpu0: Offline failed.
> >   
> > > or directly writing to cpu hotplug registers, which makes
> > > qemu crash with SIGSEGV following back trace:
> > > 
> > >    kvm_flush_coalesced_mmio_buffer ()
> > >        while (ring->first != ring->last)
> > >    ...
> > >    qemu_flush_coalesced_mmio_buffer
> > >    prepare_mmio_access
> > >    flatview_read_continue
> > >    flatview_read
> > >    address_space_read_full
> > >    address_space_rw
> > >    kvm_cpu_exec(cpu!0)
> > >    qemu_kvm_cpu_thread_fn
> > > 
> > > the reason for which is that ring == KVMState::coalesced_mmio_ring
> > > happens to be a part of 1st CPU that was uplugged by guest.
> > > 
> > > Fix it by forbidding 1st cpu unplug from guest side and in addition
> > > remove CPU0._EJ0 ACPI method to make clear that unplug of the first
> > > CPU is not supported.
> > > 
> > > Signed-off-by: Igor Mammedov <imammedo@redhat.com>
> > > ---
> > > CCing spapr and s390x folks in case targets need to prevent 1st CPU unplug as well
> > >   
> > 
> > A spapr guest can _release_ the first cpu by doing something like:
> > 
> > # echo -n "/cpus/PowerPC,POWER8@0" > /sys/devices/system/cpu/release
> > 
> > But AFAIK, this doesn't unplug the cpu from a QEMU standpoint.  
> 
> Unplugging CPU 0 with device_del should be ok too.
Do you mean real unplugging (cpu0 object freed) or just remove request?