Re: [RFC PATCH v3 6/9] hw/arm/virt-acpi-build: Use possible cpus in generation of MADT

[Andrew Jones <drjones@redhat.com>]

On Tue, May 18, 2021 at 12:27:59AM +0800, wangyanan (Y) wrote:
> Hi Drew,
> 
> On 2021/5/17 15:42, Andrew Jones wrote:
> > On Sun, May 16, 2021 at 06:28:57PM +0800, Yanan Wang wrote:
> > > When building ACPI tables regarding CPUs we should always build
> > > them for the number of possible CPUs, not the number of present
> > > CPUs. So we create gicc nodes in MADT for possible cpus and then
> > > ensure only the present CPUs are marked ENABLED. Furthermore, it
> > > also needed if we are going to support CPU hotplug in the future.
> > > 
> > > Co-developed-by: Andrew Jones <drjones@redhat.com>
> > > Signed-off-by: Andrew Jones <drjones@redhat.com>
> > > Co-developed-by: Ying Fang <fangying1@huawei.com>
> > > Signed-off-by: Ying Fang <fangying1@huawei.com>
> > > Co-developed-by: Yanan Wang <wangyanan55@huawei.com>
> > > Signed-off-by: Yanan Wang <wangyanan55@huawei.com>
> > > ---
> > >   hw/arm/virt-acpi-build.c | 29 +++++++++++++++++++++++++----
> > >   1 file changed, 25 insertions(+), 4 deletions(-)
> > > 
> > > diff --git a/hw/arm/virt-acpi-build.c b/hw/arm/virt-acpi-build.c
> > > index a2d8e87616..4d64aeb865 100644
> > > --- a/hw/arm/virt-acpi-build.c
> > > +++ b/hw/arm/virt-acpi-build.c
> > > @@ -481,6 +481,9 @@ build_madt(GArray *table_data, BIOSLinker *linker, VirtMachineState *vms)
> > >       const int *irqmap = vms->irqmap;
> > >       AcpiMadtGenericDistributor *gicd;
> > >       AcpiMadtGenericMsiFrame *gic_msi;
> > > +    MachineClass *mc = MACHINE_GET_CLASS(vms);
> > > +    const CPUArchIdList *possible_cpus = mc->possible_cpu_arch_ids(MACHINE(vms));
> > > +    bool pmu;
> > >       int i;
> > >       acpi_data_push(table_data, sizeof(AcpiMultipleApicTable));
> > > @@ -491,11 +494,21 @@ build_madt(GArray *table_data, BIOSLinker *linker, VirtMachineState *vms)
> > >       gicd->base_address = cpu_to_le64(memmap[VIRT_GIC_DIST].base);
> > >       gicd->version = vms->gic_version;
> > > -    for (i = 0; i < MACHINE(vms)->smp.cpus; i++) {
> > > +    for (i = 0; i < possible_cpus->len; i++) {
> > >           AcpiMadtGenericCpuInterface *gicc = acpi_data_push(table_data,
> > >                                                              sizeof(*gicc));
> > >           ARMCPU *armcpu = ARM_CPU(qemu_get_cpu(i));
> > > +        /*
> > > +         * PMU should have been either implemented for all CPUs or not,
> > > +         * so we only get information from the first CPU, which could
> > > +         * represent the others.
> > > +         */
> > > +        if (i == 0) {
> > > +            pmu = arm_feature(&armcpu->env, ARM_FEATURE_PMU);
> > > +        }
> > > +        assert(!armcpu || arm_feature(&armcpu->env, ARM_FEATURE_PMU) == pmu);
> > This doesn't belong in this patch. The commit message doesn't even mention
> > it. Also, I don't think we should do this here at all. If we want to
> > ensure that all cpus have a pmu when one does, then that should be done
> > somewhere like machvirt_init(), not in ACPI generation code which doesn't
> > even run for non-ACPI VMs.
> Sorry, I should have stated the reason of this change in the commit message.
> Actually code change here and mp_affinity part below aim to make it correct
> to create gicc entries for all possible cpus.
> 
> We only initialize and realize cpuobj for present cpus in machvirt_init,
> so that we will get null ARMCPU pointer here for the non-present cpus,
> and consequently we won't able to check from "armcpu->env" for the
> non-present cpus. The same about "armcpu->mp_affinity".
> 
> That's the reason I use PMU configuration of the first cpu to represent the
> others. I assume all cpus should have a pmu when one does here since it's
> how armcpu->env is initialized. And the assert seems not needed here.
> 
> Is there any better alternative way about this?

Move the

  if (arm_feature(&armcpu->env, ARM_FEATURE_PMU)) {
      gicc->performance_interrupt = cpu_to_le32(PPI(VIRTUAL_PMU_IRQ));
  }

into the if (possible_cpus->cpus[i].cpu != NULL) block?

> > > +
> > >           gicc->type = ACPI_APIC_GENERIC_CPU_INTERFACE;
> > >           gicc->length = sizeof(*gicc);
> > >           if (vms->gic_version == 2) {
> > > @@ -504,11 +517,19 @@ build_madt(GArray *table_data, BIOSLinker *linker, VirtMachineState *vms)
> > >               gicc->gicv_base_address = cpu_to_le64(memmap[VIRT_GIC_VCPU].base);
> > >           }
> > >           gicc->cpu_interface_number = cpu_to_le32(i);
> > > -        gicc->arm_mpidr = cpu_to_le64(armcpu->mp_affinity);
> > > +        gicc->arm_mpidr = cpu_to_le64(possible_cpus->cpus[i].arch_id);
> > Hmm, I think we may have a problem. I don't think there's any guarantee
> > that possible_cpus->cpus[i].arch_id == armcpu->mp_affinity, because
> > arch_id comes from virt_cpu_mp_affinity(), which is arm_cpu_mp_affinity,
> > but with a variable cluster size, however mp_affinity comes from
> > arm_cpu_mp_affinity with a set cluster size. Also, when KVM is used,
> > then all bets are off as to what mp_affinity is.
> Right! Arch_id is initialized by virt_cpu_mp_affinity() in machvirt and then
> mp_affinity is initialized by arch_id. Here they two have the same value.
> 
> But mp_affinity will be overridden in kvm_arch_init_vcpu() when KVM is
> enabled. Here they two won't have the same value.
> > We need to add some code that ensures arch_id == mp_affinity,
> Can we also update the arch_id at the same time when we change mp_affinity?

The proper fix is to send patches to KVM enabling userspace to control
MPIDR. Otherwise we can't be sure we don't have inconsistencies in QEMU,
since some user of possible_cpus could have made decisions or copied IDs
prior to KVM vcpu init time. Now, all that said, I think
virt_cpu_mp_affinity() should be generating the same ID as KVM does, so
maybe it doesn't matter in practice right now, but we're living with the
risk that KVM could change. For now, maybe we should just sanity check
that the KVM values match the possible_cpus values and emit warnings if
they don't?

Thanks,
drew