Re: [RFC PATCH 00/12] hw/arm/virt: Introduce cpu and cache topology support

[Andrew Jones]

On Fri, Oct 16, 2020 at 05:40:02PM +0800, Ying Fang wrote:
> 
> 
> On 10/15/2020 3:59 PM, Andrew Jones wrote:
> > On Thu, Oct 15, 2020 at 10:07:16AM +0800, Ying Fang wrote:
> > > 
> > > 
> > > On 10/14/2020 2:08 AM, Andrew Jones wrote:
> > > > On Tue, Oct 13, 2020 at 12:11:20PM +0000, Zengtao (B) wrote:
> > > > > Cc valentin
> > > > > 
> > > > > > -----Original Message-----
> > > > > > From: Qemu-devel
> > > > > > [mailto:qemu-devel-bounces+prime.zeng=hisilicon.com@nongnu.org]
> > > > > > On Behalf Of Ying Fang
> > > > > > Sent: Thursday, September 17, 2020 11:20 AM
> > > > > > To: qemu-devel@nongnu.org
> > > > > > Cc: peter.maydell@linaro.org; drjones@redhat.com; Zhanghailiang;
> > > > > > Chenzhendong (alex); shannon.zhaosl@gmail.com;
> > > > > > qemu-arm@nongnu.org; alistair.francis@wdc.com; fangying;
> > > > > > imammedo@redhat.com
> > > > > > Subject: [RFC PATCH 00/12] hw/arm/virt: Introduce cpu and cache
> > > > > > topology support
> > > > > > 
> > > > > > An accurate cpu topology may help improve the cpu scheduler's
> > > > > > decision
> > > > > > making when dealing with multi-core system. So cpu topology
> > > > > > description
> > > > > > is helpful to provide guest with the right view. Cpu cache 
> > > > > > information
> > > > > > may
> > > > > > also have slight impact on the sched domain, and even userspace
> > > > > > software
> > > > > > may check the cpu cache information to do some optimizations. Thus
> > > > > > this patch
> > > > > > series is posted to provide cpu and cache topology support for arm.
> > > > > > 
> > > > > > To make the cpu topology consistent with MPIDR, an vcpu ioctl
> > > > > 
> > > > > For aarch64, the cpu topology don't depends on the MPDIR.
> > > > > See https://patchwork.kernel.org/patch/11744387/
> > > > > 
> > > > 
> > > > The topology should not be inferred from the MPIDR Aff fields,
> > > 
> > > MPIDR is abused by ARM OEM manufactures. It is only used as a
> > > identifer for a specific cpu, not representation of the topology.
> > 
> > Right, which is why I stated topology should not be inferred from
> > it.
> > 
> > > 
> > > > but MPIDR is the CPU identifier. When describing a topology
> > > > with ACPI or DT the CPU elements in the topology description
> > > > must map to actual CPUs. MPIDR is that mapping link. KVM
> > > > currently determines what the MPIDR of a VCPU is. If KVM
> > > 
> > > KVM currently assigns MPIDR with vcpu->vcpu_id which mapped
> > > into affinity levels. See reset_mpidr in sys_regs.c
> > 
> > I know, but how KVM assigns MPIDRs today is not really important
> > to KVM userspace. KVM userspace shouldn't depend on a KVM
> > algorithm, as it could change.
> > 
> > > 
> > > > userspace is going to determine the VCPU topology, then it
> > > > also needs control over the MPIDR values, otherwise it
> > > > becomes quite messy trying to get the mapping right.
> > > If we are going to control MPIDR, shall we assign MPIDR with
> > > vcpu_id or map topology hierarchy into affinity levels or any
> > > other link schema ?
> > > 
> > 
> > We can assign them to whatever we want, as long as they're
> > unique and as long as Aff0 is assigned per the GIC requirements,
> > e.g. GICv3 requires that Aff0 be from 0 to 0xf. Also, when
> > pinning VCPUs to PCPUs we should ensure that MPIDRs with matching
> > Aff3,Aff2,Aff1 fields should actually be peers with respect to
> > the GIC.
> 
> Still not clear why vCPU's MPIDR need to match pPCPU's GIC affinity.
> Maybe I should read spec for GICv3.

Look at how IPIs are efficiently sent to "peers", where the definition
of a peer is that only Aff0 differs in its MPIDR. But, gicv3's
optimizations can only handle 16 peers. If we want pinned VCPUs to
have the same performance as PCPUS, then we should maintain this
Aff0 limit.

Thanks,
drew

> 
> > 
> > We shouldn't try to encode topology in the MPIDR in any way,
> > so we might as well simply increment a counter to assign them,
> > which could possibly be the same as the VCPU ID.
> 
> Hmm, then we can leave it as it is.
> 
> > 
> > Thanks,
> > drew
> > 
> > .
> > 
>