[Qemu-devel] Re: [PATCH v3 07/10] qemu-kvm: Cleanup/fix TSC and PV clock writeback

[Marcelo Tosatti]

On Thu, Feb 25, 2010 at 09:48:47AM +0100, Jan Kiszka wrote:
> Marcelo Tosatti wrote:
> > On Thu, Feb 25, 2010 at 12:58:26AM +0100, Jan Kiszka wrote:
> >> Marcelo Tosatti wrote:
> >>> On Thu, Feb 25, 2010 at 12:45:55AM +0100, Jan Kiszka wrote:
> >>>> Marcelo Tosatti wrote:
> >>>>> On Wed, Feb 24, 2010 at 03:17:55PM +0100, Jan Kiszka wrote:
> >>>>>> Drop kvm_load_tsc in favor of level-dependent writeback in
> >>>>>> kvm_arch_load_regs. KVM's PV clock MSRs fall in the same category and
> >>>>>> should therefore only be written back on full sync.
> >>>>>>
> >>>>>> Signed-off-by: Jan Kiszka <address@hidden>
> >>>>>> ---
> >>>>>>  qemu-kvm-x86.c        |   19 +++++--------------
> >>>>>>  qemu-kvm.h            |    4 ----
> >>>>>>  target-i386/machine.c |    5 -----
> >>>>>>  3 files changed, 5 insertions(+), 23 deletions(-)
> >>>>>>
> >>>>>> diff --git a/qemu-kvm-x86.c b/qemu-kvm-x86.c
> >>>>>> index 840c1c9..84fd7fa 100644
> >>>>>> --- a/qemu-kvm-x86.c
> >>>>>> +++ b/qemu-kvm-x86.c
> >>>>>> @@ -965,8 +965,11 @@ void kvm_arch_load_regs(CPUState *env, int level)
> >>>>>>          set_msr_entry(&msrs[n++], MSR_LSTAR  ,           env->lstar);
> >>>>>>      }
> >>>>>>  #endif
> >>>>>> -    set_msr_entry(&msrs[n++], MSR_KVM_SYSTEM_TIME,  
> >>>>>> env->system_time_msr);
> >>>>>> -    set_msr_entry(&msrs[n++], MSR_KVM_WALL_CLOCK,  
> >>>>>> env->wall_clock_msr);
> >>>>>> +    if (level == KVM_PUT_FULL_STATE) {
> >>>>>> +        set_msr_entry(&msrs[n++], MSR_IA32_TSC, env->tsc);
> >>>>>> +        set_msr_entry(&msrs[n++], MSR_KVM_SYSTEM_TIME, 
> >>>>>> env->system_time_msr);
> >>>>>> +        set_msr_entry(&msrs[n++], MSR_KVM_WALL_CLOCK, 
> >>>>>> env->wall_clock_msr);
> >>>>>> +    }
> >>>>> As things stand today, the TSC should only be written on migration. See
> >>>>> 53f658b3c33616a4997ee254311b335e59063289 in the kernel.
> >>>> Migration and power-up - that's what this patch ensures (=>
> >>>> KVM_PUT_FULL_STATE). Or where do you see any problem?
> >>>>
> >>>> Jan
> >>>>
> >>> The problem is it should not write on power up (the kernel attempts
> >>> to synchronize the TSCs in that case, see the commit).
> >>>
> >> OK, need to read this more carefully.
> >>
> >> I do not yet understand the difference from user space POV: it tries to
> >> transfer the identical TSC values to all currently stopped VCPU threads.
> > 
> > guest tsc = host tsc + offset
> > 
> > So at the time you set_msr(TSC), the guest visible TSC starts ticking. 
> > For SMP guests, this does not happen exactly at the same time for all
> > vcpus.
> 
> Ouch.
> 
> > 
> >> That should not be different if we are booting a fresh VM or loading a
> >> complete state of a migrated image. If it does, it looks like a KVM
> >> kernel deficit on first glance.
> > 
> > Yes it is a deficit. After migration TSCs of SMP guests go out of sync.
> > Zachary is working on that.
> > 
> 
> OK, so we need a workaround, ideally without reintroducing hooks. Is
> this one acceptable?
> 
> 
> diff --git a/qemu-kvm-x86.c b/qemu-kvm-x86.c
> index 84fd7fa..285c05a 100644
> --- a/qemu-kvm-x86.c
> +++ b/qemu-kvm-x86.c
> @@ -966,7 +966,15 @@ void kvm_arch_load_regs(CPUState *env, int level)
>      }
>  #endif
>      if (level == KVM_PUT_FULL_STATE) {
> -        set_msr_entry(&msrs[n++], MSR_IA32_TSC, env->tsc);
> +        /*
> +         * KVM is yet unable to synchronize TSC values of multiple VCPUs on
> +         * writeback. Until this is fixed, we only write the offset to SMP
> +         * guests after migration, desynchronizing the VCPUs, but avoiding
> +         * huge jump-backs that would occur without any writeback at all.
> +         */
> +        if (smp_cpus == 1 || env->tsc != 0) {
> +            set_msr_entry(&msrs[n++], MSR_IA32_TSC, env->tsc);
> +        }
>          set_msr_entry(&msrs[n++], MSR_KVM_SYSTEM_TIME, env->system_time_msr);
>          set_msr_entry(&msrs[n++], MSR_KVM_WALL_CLOCK, env->wall_clock_msr);
>      }

Well, migration of the SMP TSCs is not precise, but it is better than
zeroing the TSCs on migration. So something like a new state (migration
only), or a hack that mimicks that behaviour, is needed :(