Re: [Qemu-devel] [PATCH v3 09/19] pc: delay setting number of boot CPUs to machine_done time

[Igor Mammedov]

On Tue, 12 Jul 2016 00:29:08 -0300
Eduardo Habkost <address@hidden> wrote:

> On Wed, Jul 06, 2016 at 08:20:45AM +0200, Igor Mammedov wrote:
> > currently present CPUs counter in CMOS only contains
> > smp_cpus (i.e. initial CPUs specified with -smp X) and
> > doesn't account for CPUs created with -device.
> > If VM is started with additional CPUs added with
> >  -device, it will hang in BIOS waiting for condition
> >    smp_cpus == counted_cpus
> > forever as counted_cpus will include -device CPUs as well
> > and be more than smp_cpus.
> > 
> > make present CPUs counter in CMOS to count all CPUs
> > (initial and coldplugged with -device) by delaying
> > it to machine done time when it possible to count
> > CPUs added with -device.  
> 
> Do you plan to fix the remaining code using smp_cpus? e.g.:
perhaps after Drew's -smp refactoring or maybe during it,
it looks like good candidate for that series,
I'll work with Drew on that.

> 
> 1) x86_cpu_realizefn():
>     if (cpu->env.features[FEAT_1_EDX] & CPUID_APIC || smp_cpus > 1) {
>             x86_cpu_apic_create(cpu, &local_err);
>     [...]
>   (Maybe we should simply make realizefn fail if we try to create a second CPU
>   using -device or device_add without CPUID_APIC)
wouldn't that break some setups that doing it but still able
to boot?

> 
> 2) the smp_cpus checks at hw/i386/kvmvapic.c.
> 
> > 
> > Signed-off-by: Igor Mammedov <address@hidden>
> > ---
> >  hw/i386/pc.c | 12 +++++++++---
> >  1 file changed, 9 insertions(+), 3 deletions(-)
> > 
> > diff --git a/hw/i386/pc.c b/hw/i386/pc.c
> > index 6691825..3206572 100644
> > --- a/hw/i386/pc.c
> > +++ b/hw/i386/pc.c
> > @@ -471,9 +471,6 @@ void pc_cmos_init(PCMachineState *pcms,
> >      rtc_set_memory(s, 0x5c, val >> 8);
> >      rtc_set_memory(s, 0x5d, val >> 16);
> >  
> > -    /* set the number of CPU */
> > -    rtc_set_memory(s, 0x5f, smp_cpus - 1);
> > -
> >      object_property_add_link(OBJECT(pcms), "rtc_state",
> >                               TYPE_ISA_DEVICE,
> >                               (Object **)&pcms->rtc,
> > @@ -1157,10 +1154,19 @@ void pc_cpus_init(PCMachineState *pcms)
> >  static
> >  void pc_machine_done(Notifier *notifier, void *data)
> >  {
> > +    int i, boot_cpus = 0;
> >      PCMachineState *pcms = container_of(notifier,
> >                                          PCMachineState, machine_done);
> >      PCIBus *bus = pcms->bus;
> >  
> > +    for (i = 0; i < pcms->possible_cpus->len; i++) {
> > +        if (pcms->possible_cpus->cpus[i].cpu) {
> > +            boot_cpus++;
> > +        }
> > +    }  
> 
> Any specific reason you chose to check possible_cpus instead of the
> arch-independent CPU list from exec.c? I believe other architectures will be
> interested in a generic way to count online CPUs instead of using smp_cpus.
When others would need to do it we can switch to arch-independent CPU list but
for now I'd prefer to keep using possible_cpus throughout pc.c for consistency
reasons (so it would be the single point of bookkeeping CPUs on PC),
and I most likely will do the same for ARM probably reusing some of PC code.

 
> > +    /* set the number of CPUs */
> > +    rtc_set_memory(pcms->rtc, 0x5f, boot_cpus - 1);
> > +
> >      if (bus) {
> >          int extra_hosts = 0;
> >  
> > -- 
> > 2.7.0
> >   
>