Re: [Qemu-devel] [PATCH] target-i386: enable kvm_pv_unhalt by default

[Eduardo Habkost]

On Wed, Sep 13, 2017 at 04:39:58PM +0200, Alexander Graf wrote:
> Commit f010bc643a (target-i386: add feature kvm_pv_unhalt) introduced the
> kvm_pv_unhalt feature but didn't enable it by default.
> 
> Without kvm_pv_unhalt we see a measurable degradation in scheduling
> performance, so enabling it by default does make sense IMHO. This patch
> just flips it to default to on by default.
> 
>   [With kvm_pv_unhalt disabled]
>   $ perf bench sched messaging -l 10000
>     Total time: 8.573 [sec]
> 
>   [With kvm_pv_unhalt enabled]
>   $ perf bench sched messaging -l 10000
>     Total time: 4.416 [sec]
> 
> Signed-off-by: Alexander Graf <address@hidden>
> 
> ---
> 
> Let's ask everyone who was involved back then whether this is a feature
> that is good to enable by default. My measurements imply so, but who knows...
> 
> Also, I'd kindly like to ask for guidance on how to make this reasonably
> well backwards compatible. I assume we only want to flip the default in newer
> machine models? If so, how?

There are two things that need to be taken care of:

* Guest ABI compatibility: if running an older machine-type, the
  feature needs to remain disabled.
* machine-type runnability: libvirt doesn't expect an existing VM
  to become not runnable on the same host if changing only the
  machine-type.  This can happen if the new machine-type enables
  a feature that's not supported by the host kernel.

Guest ABI compatibility is relatively easy to handle: see how we
deal with kvm-pv-eoi on pc-1.2.  Moving the machine-type-specific
kvm-defaults to PCMachineClass (instead of introducing new
pc_compat_*() functions) would be interesting, though; I will
take a look and try to implement that.

machine-type runnability is trickier: it's a problem in theory if
people can be running old kernels that didn't support the
feature, but may be not a problem in practice if the feature is
so old that everybody running a recent QEMU is also running a
recent enough kernel.  In either case, the problem still needs to
be addressed somehow; just documenting the minimum kernel version
QEMU needs would probably be enough.

For reference, kvm_pv_unhalt was introduced in Linux v3.12-rc1.

> ---
>  target/i386/cpu.c | 1 +
>  1 file changed, 1 insertion(+)
> 
> diff --git a/target/i386/cpu.c b/target/i386/cpu.c
> index 69676e13e1..c58f4ab24f 100644
> --- a/target/i386/cpu.c
> +++ b/target/i386/cpu.c
> @@ -1581,6 +1581,7 @@ static PropValue kvm_default_props[] = {
>      { "kvm-asyncpf", "on" },
>      { "kvm-steal-time", "on" },
>      { "kvm-pv-eoi", "on" },
> +    { "kvm-pv-unhalt", "on" },
>      { "kvmclock-stable-bit", "on" },
>      { "x2apic", "on" },
>      { "acpi", "off" },
> -- 
> 2.12.3
> 

-- 
Eduardo