Received-SPF: pass (google.com: domain of linux-kernel+bounces-212287-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:4601:e00::3 as permitted sender) client-ip=2604:1380:4601:e00::3;
Date: Wed, 12 Jun 2024 14:41:27 -0700
In-Reply-To: <20240207172646.3981-9-xin3.li@intel.com>
Precedence: bulk
Mime-Version: 1.0
References: <20240207172646.3981-1-xin3.li@intel.com> <20240207172646.3981-9-xin3.li@intel.com>
Message-ID: <ZmoWB_XtA0wR2K4Q@google.com>
Subject: Re: [PATCH v2 08/25] KVM: VMX: Initialize VMCS FRED fields
From: Sean Christopherson <seanjc@google.com>
To: Xin Li <xin3.li@intel.com>
Cc: linux-kernel@vger.kernel.org, kvm@vger.kernel.org, 
	linux-doc@vger.kernel.org, linux-kselftest@vger.kernel.org, 
	pbonzini@redhat.com, corbet@lwn.net, tglx@linutronix.de, mingo@redhat.com, 
	bp@alien8.de, dave.hansen@linux.intel.com, x86@kernel.org, hpa@zytor.com, 
	shuah@kernel.org, vkuznets@redhat.com, peterz@infradead.org, 
	ravi.v.shankar@intel.com, xin@zytor.com
Content-Type: text/plain; charset="us-ascii"

On Wed, Feb 07, 2024, Xin Li wrote:
> diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
> index d58ed2d3d379..b7b772183ee4 100644
> --- a/arch/x86/kvm/vmx/vmx.c
> +++ b/arch/x86/kvm/vmx/vmx.c
> @@ -1470,6 +1470,18 @@ void vmx_vcpu_load_vmcs(struct kvm_vcpu *vcpu, int cpu,
>  				    (unsigned long)(cpu_entry_stack(cpu) + 1));
>  		}
>  
> +#ifdef CONFIG_X86_64

Don't bother, practically no one cares about 32-bit KVM these days, and I highly
don't anyone that runs 32-bit KVM cares about the code footprint to this degree.

> +		/* Per-CPU FRED MSRs */
> +		if (kvm_cpu_cap_has(X86_FEATURE_FRED)) {
> +			vmcs_write64(HOST_IA32_FRED_RSP1, read_msr(MSR_IA32_FRED_RSP1));
> +			vmcs_write64(HOST_IA32_FRED_RSP2, read_msr(MSR_IA32_FRED_RSP2));
> +			vmcs_write64(HOST_IA32_FRED_RSP3, read_msr(MSR_IA32_FRED_RSP3));
> +			vmcs_write64(HOST_IA32_FRED_SSP1, read_msr(MSR_IA32_FRED_SSP1));
> +			vmcs_write64(HOST_IA32_FRED_SSP2, read_msr(MSR_IA32_FRED_SSP2));
> +			vmcs_write64(HOST_IA32_FRED_SSP3, read_msr(MSR_IA32_FRED_SSP3));

That's a lot of RDMSRs to eat on every task migration.  How hard would it be to
add a per-CPU cache for each of these?  Or is there a pre-existing way to get at
the info that's faster than RDMSR?

> +		}
> +#endif
> +
>  		vmx->loaded_vmcs->cpu = cpu;
>  	}
>  }
> @@ -4321,6 +4333,15 @@ void vmx_set_constant_host_state(struct vcpu_vmx *vmx)
>  	 */
>  	vmcs_write16(HOST_DS_SELECTOR, 0);
>  	vmcs_write16(HOST_ES_SELECTOR, 0);
> +
> +	/*
> +	 * FRED MSRs are per-cpu, however FRED CONFIG and STKLVLS MSRs
> +	 * are the same on all CPUs, thus they are initialized here.

Eh, just trim this to:

	/* FRED CONFIG and STKLVLS are the same on all CPUs. */

> +	 */
> +	if (kvm_cpu_cap_has(X86_FEATURE_FRED)) {
> +		vmcs_write64(HOST_IA32_FRED_CONFIG, read_msr(MSR_IA32_FRED_CONFIG));
> +		vmcs_write64(HOST_IA32_FRED_STKLVLS, read_msr(MSR_IA32_FRED_STKLVLS));
> +	}
>  #else
>  	vmcs_write16(HOST_DS_SELECTOR, __KERNEL_DS);  /* 22.2.4 */
>  	vmcs_write16(HOST_ES_SELECTOR, __KERNEL_DS);  /* 22.2.4 */
> @@ -4865,6 +4886,19 @@ static void __vmx_vcpu_reset(struct kvm_vcpu *vcpu)
>  	 */
>  	vmx->pi_desc.nv = POSTED_INTR_VECTOR;
>  	vmx->pi_desc.sn = 1;
> +
> +#ifdef CONFIG_X86_64
> +	if (kvm_cpu_cap_has(X86_FEATURE_FRED)) {
> +		vmcs_write64(GUEST_IA32_FRED_CONFIG, 0);
> +		vmcs_write64(GUEST_IA32_FRED_RSP1, 0);
> +		vmcs_write64(GUEST_IA32_FRED_RSP2, 0);
> +		vmcs_write64(GUEST_IA32_FRED_RSP3, 0);
> +		vmcs_write64(GUEST_IA32_FRED_STKLVLS, 0);
> +		vmcs_write64(GUEST_IA32_FRED_SSP1, 0);
> +		vmcs_write64(GUEST_IA32_FRED_SSP2, 0);
> +		vmcs_write64(GUEST_IA32_FRED_SSP3, 0);
> +	}

Somewhat of a moot point, but this belongs in init_vmcs(), not __vmx_vcpu_reset().