On Mon, Apr 23, 2018 at 07:15:58PM +0200, Paolo Bonzini wrote:
On 22/04/2018 02:53, Anthoine Bourgeois wrote:Since the commit "8003c9ae204e: add APIC Timer periodic/oneshot mode VMX preemption timer support", a Windows 10 guest has some erratic timer spikes after few hours. As the uptime of the VM grows the spikes are larger. Here the results on a 150000 times 1ms timer without any load: Before 8003c9ae204e | After 8003c9ae204e Max 1834us | 86000us Mean 1100us | 1021us Deviation 59us | 149us Here the results on a 150000 times 1ms timer with a cpu-z stress test: Before 8003c9ae204e | After 8003c9ae204e Max 32000us | 140000us Mean 1006us | 1997us Deviation 140us | 11095us The current patch partially revert the previous commit by removing the target timer expectation to go back to the straight hrtimer calls. The APIC Timer periodic/oneshot mode support is kept because it is necessary on the new Windows Spring update. v2: Check if the tsc deadline is already expired. Thank you Mika. Cc: Mika Penttilä <mika.penttila@xxxxxxxxxxxx Signed-off-by: Anthoine Bourgeois <anthoine.bourgeois@xxxxxxxxxxxxxxx> --- arch/x86/kvm/lapic.c | 57 +++++++++++++++++++++++++--------------------------- arch/x86/kvm/lapic.h | 1 - 2 files changed, 27 insertions(+), 31 deletions(-) diff --git a/arch/x86/kvm/lapic.c b/arch/x86/kvm/lapic.c index 70dcb5548022..8b5c2a69a3b6 100644 --- a/arch/x86/kvm/lapic.c +++ b/arch/x86/kvm/lapic.c @@ -1173,7 +1173,7 @@ static void apic_send_ipi(struct kvm_lapic *apic) static u32 apic_get_tmcct(struct kvm_lapic *apic) { - ktime_t remaining, now; + ktime_t remaining; s64 ns; u32 tmcct; @@ -1184,8 +1184,7 @@ static u32 apic_get_tmcct(struct kvm_lapic *apic) apic->lapic_timer.period == 0) return 0; - now = ktime_get(); - remaining = ktime_sub(apic->lapic_timer.target_expiration, now); + remaining = hrtimer_get_remaining(&apic->lapic_timer.timer);I'm confused, how can this work when the preemption timer is in use (vcpu->arch.apic->lapic_timer.hv_timer_in_use is true)?
I don't really know, this hunk is only a revert that works for me. I'm still seeking what is the root cause. My guest is the target_expiration variable is mis compute sometimes. What I see is the spikes are linear over time at the rate of 1ms more every 1 minutes 30 seconds. Anthoine
