Subject: Re: [PATCH] KVM: x86: do not go through vcpu in __get_kvmclock_ns
To: =?UTF-8?B?UmFkaW0gS3LEjW3DocWZ?= <rkrcmar@redhat.com>
References: <1478859141-25146-1-git-send-email-pbonzini@redhat.com>
 <20161114145239.GA2185@potion>
Cc: linux-kernel@vger.kernel.org, kvm@vger.kernel.org, mtosatti@redhat.com
From: Paolo Bonzini <pbonzini@redhat.com>
Message-ID: <18a835aa-407e-6102-8b41-05b90cae1784@redhat.com>
Date: Mon, 14 Nov 2016 17:17:38 +0100
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:45.0) Gecko/20100101
 Thunderbird/45.4.0
MIME-Version: 1.0
In-Reply-To: <20161114145239.GA2185@potion>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1119
Lines: 34


On 14/11/2016 15:52, Radim Krčmář wrote:
> The hunk below should return the same value in pvclock_ns and kernel_ns
> if they can be used interchangeably.  boot_ns is expected to be a bit
> delayed, because it is read late.  boot_ns shows a bounded offset from
> kernel_ns, unlike the drifting pvclock_ns.
> 
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index 83990ad3710e..30d4d3d02ac7 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -6653,6 +6653,17 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
>  		goto cancel_injection;
>  	}
>  
> +	if (vcpu->kvm->arch.use_master_clock) {
> +		s64 kernel_ns;
> +		cycle_t tsc_now, pvclock_ns, boot_ns;
> +
> +		kvm_get_time_and_clockread(&kernel_ns, &tsc_now);
> +		pvclock_ns = __pvclock_read_cycles(&vcpu->arch.hv_clock, kvm_read_l1_tsc(vcpu, tsc_now)) - vcpu->kvm->arch.kvmclock_offset;
> +		boot_ns = ktime_get_boot_ns();
> +
> +		printk("ns diff: %lld %lld\n", pvclock_ns - kernel_ns, boot_ns - kernel_ns);
> +	}
> +
>  	preempt_disable();
>  
>  	kvm_x86_ops->prepare_guest_switch(vcpu);

Ok, I'll post a v2 of this patch.

Paolo