[libvirt] 'host-passthrough' for arm64

[Cole Robinson]

On 01/20/2014 08:19 AM, Oleg Strikov wrote:
> Hello guys,
> 
> I'm trying to come up with basic OpenStack support for arm64 node.
> I'd like to use 'libvirt_cpu_mode=host-passthrough' configuration option with
> Nova which issues <cpu mode='host-passthrough'> to libvirt xml config.
> But with this option passed libvirt crashes with 'error: unsupported
> configuration: CPU specification not supported by hypervisor'.
> This happens because the following handlers are not implemented (or
> implemented as stubs) inside src/cpu/cpu_aarch64.c:
> * AArch64Decode()
> * AArch64Update()
> * AArch64guestData()
> 
> To solve exactly this 'host-passthrough'-related issue that's enough to have
> the following set of handlers:
> 
> AArch64Decode(<...>)
> {
>     virCheckFlags(VIR_CONNECT_BASELINE_CPU_EXPAND_FEATURES, -1);
> 
>     /* I don't know any way to detect 'cortex-a57' or any other armv8 CPU for
> now */
>     /* But I don't think that we can meet anything else than cortex-a57 */
> 
>     /* We may also put 'host' there to specifically point out that
> qemu-aarch64 supports only '-cpu host' for now */
>     /* pm215 told me that the ETA for '-cpu cortex-a57' and friends is around
> 3 months from now */
> 
>     return !(VIR_STRDUP(cpu->model, "host" or "cortex-a57") == 1);
> }
> 
> static int AArch64Update(<...>)
> {
>         /* qemu-aarch64 supports only '-cpu host' for now */
> 
>         guest->match = VIR_CPU_MATCH_EXACT;
>         virCPUDefFreeModel(guest);
>         return virCPUDefCopyModel(guest, host, true);
> }
> 
> static virCPUCompareResult
> AArch64guestData(<..>)
> {
>     return  VIR_CPU_COMPARE_IDENTICAL;
> }
> 
> That's clear that these handlers provide just basic functionality
> ('host-passthrough'-only) and have to be extended in future.
> But is it something we can commit for now?
> 
> Another way to deal with this issue is to adopt some code from PPC handlers
> (including CPU model detection and best fit qemu configuration discovery).
> But this way will be blocked until:
> (1) I find any way to reliably detect CPU model on ARMv8 board (any ideas?)
> (2) pm215 implements TCG for arm64
> 
> What is the best way to choose to come up with the commitable code?
> 

IMO the above is fine, doesn't look like it should cause any compatibility
issues with a real implementation in the future, and allows launching
qemu-system-aarch64 with its preferred -cpu host flag. I'd say post the patch,
and if other folks disagree they can comment there.

Thanks,
Cole