On Mon, Sep 28, 2015 at 01:37:34PM -0300, Eduardo Habkost wrote:
> On Mon, Sep 28, 2015 at 01:38:31PM +0800, Haozhong Zhang wrote:
> > When a vcpu is created in KVM, its TSC rate is initially identical to
> > the host TSC rate. If its state is migrated to a vcpu on another
> > machine (target machine) which may uses a different host TSC rate, QEMU
> > on the target machine should notice KVM of the migrated vcpu's TSC
> > rate. In case that KVM on the target machine supports TSC scaling, guest
> > programs running on the migrated vcpu will observe the same TSC rate
> > before and after the migration.
> >
> > Signed-off-by: Haozhong Zhang <haozhong.zhang@xxxxxxxxx>
> > ---
> > kvm-all.c | 13 +++++++++++++
> > 1 file changed, 13 insertions(+)
> >
> > diff --git a/kvm-all.c b/kvm-all.c
> > index 0be4615..e8de038 100644
> > --- a/kvm-all.c
> > +++ b/kvm-all.c
> > @@ -1769,6 +1769,19 @@ void kvm_cpu_synchronize_post_reset(CPUState *cpu)
> > static void do_kvm_cpu_synchronize_post_init(void *arg)
> > {
> > CPUState *cpu = arg;
> > + CPUX86State *env = &X86_CPU(cpu)->env;
> > + int r;
> > +
> > + /*
> > + * XXX: KVM_SET_TSC_KHZ must be done before kvm_arch_put_registers().
>
> Could you explain where this requirement comes from?
>
kvm_arch_put_registers() below will setup vcpu's MSR_IA32_TSC through
KVM ioctl KVM_SET_MSRS. KVM needs to know vcpu's TSC rate so as to
correctly scale the TSC value given by QEMU, especially when vcpu's
TSC rate is different than the host TSC rate (e.g. it's migrated from
another machine w/ different host TSC rate than the current one).
> > + */
> > + r = kvm_check_extension(cpu->kvm_state, KVM_CAP_TSC_CONTROL);
> > + if (r && env->tsc_khz) {
> > + r = kvm_vcpu_ioctl(cpu, KVM_SET_TSC_KHZ, env->tsc_khz);
> > + if (r < 0) {
> > + fprintf(stderr, "KVM_SET_TSC_KHZ failed\n");
> > + }
> > + }
>
> This is duplicating the existing KVM_SET_TSC_KHZ call at
> kvm_arch_init_vcpu(). I wonder if there's a way to avoid this
> duplication. Should we set TSC KHz only at
> do_kvm_cpu_synchronize_post_init(), and remove the call from
> kvm_arch_init_vcpu()?
>
I'll check if it's safe to remove the call from kvm_arch_init_vcpu().
> Or maybe we shouldn't treat this as VM state, but as configuration, and
> let management configure the TSC frequency explicitly if the user really
> needs it to stay the same during migration.
>
> (CCing libvir-list to see if they have feedback)
>
Thanks for CC. I'll consider to add a command line option to control
the configuration of guest TSC fequency.
> --
> Eduardo
--
libvir-list mailing list
libvir-list@xxxxxxxxxx
https://www.redhat.com/mailman/listinfo/libvir-list