Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Subject: Re: [PATCH v2] KVM: x86/pmu: Support full width counting
To:     Like Xu <like.xu@linux.intel.com>
Cc:     Sean Christopherson <sean.j.christopherson@intel.com>,
        Vitaly Kuznetsov <vkuznets@redhat.com>,
        Wanpeng Li <wanpengli@tencent.com>,
        Jim Mattson <jmattson@google.com>,
        Joerg Roedel <joro@8bytes.org>, kvm@vger.kernel.org,
        linux-kernel@vger.kernel.org
References: <20200507021452.174646-1-like.xu@linux.intel.com>
From:   Paolo Bonzini <pbonzini@redhat.com>
Message-ID: <3fb56700-7f0b-59e1-527a-f8eb601185b1@redhat.com>
Date:   Thu, 7 May 2020 09:57:08 +0200
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:68.0) Gecko/20100101
 Thunderbird/68.6.0
MIME-Version: 1.0
In-Reply-To: <20200507021452.174646-1-like.xu@linux.intel.com>
Content-Type: text/plain; charset=windows-1252
Content-Language: en-US
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On 07/05/20 04:14, Like Xu wrote:
> +static inline u64 vmx_get_perf_capabilities(void)
> +{
> +	u64 perf_cap = 0;
> +
> +	if (boot_cpu_has(X86_FEATURE_PDCM))
> +		rdmsrl(MSR_IA32_PERF_CAPABILITIES, perf_cap);
> +
> +	/* Currently, KVM only supports Full-Width Writes. */
> +	perf_cap &= PMU_CAP_FW_WRITES;
> +
> +	return perf_cap;
> +}
> +

Since counters are virtualized, it seems to me that you can support
PMU_CAP_FW_WRITES unconditionally, even if the host lacks it.  So just
return PMU_CAP_FW_WRITES from this function.

> +	case MSR_IA32_PERF_CAPABILITIES:
> +		return 1; /* RO MSR */
>  	default:

You need to allow writes from the host if (data &
~vmx_get_perf_capabilities()) == 0.

> -			if (!msr_info->host_initiated)
> +		if ((pmc = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0)) ||
> +			(pmc = get_gp_pmc(pmu, msr, MSR_IA32_PMC0))) {
> +			if (data & ~pmu->counter_bitmask[KVM_PMC_GP])
> +				return 1;
> +			if (!fw_writes_is_enabled(pmu))
>  				data = (s64)(s32)data;


You are dropping the test on msr_info->host_initiated here, you should
keep it otherwise you allow full-width write to MSR_IA32_PERFCTR0 as
well.  So:

#define MSR_PMC_FULL_WIDTH_BIT      (MSR_IA32_PMC0 - MSR_IA32_PERFCTR0)

	if (!msr_info->host_initiated && !(msr & MSR_PMC_FULL_WIDTH_BIT))
		data = (s64)(s32)data;

> +	case MSR_IA32_PERF_CAPABILITIES:
> +		if (!nested)
> +			return 1;
> +		msr->data = vmx_get_perf_capabilities();
> +		return 0;

The !nested check is wrong.

> 
> +++ b/arch/x86/kvm/x86.c
> @@ -1220,6 +1220,13 @@ static const u32 msrs_to_save_all[] = {
>  	MSR_ARCH_PERFMON_EVENTSEL0 + 12, MSR_ARCH_PERFMON_EVENTSEL0 + 13,
>  	MSR_ARCH_PERFMON_EVENTSEL0 + 14, MSR_ARCH_PERFMON_EVENTSEL0 + 15,
>  	MSR_ARCH_PERFMON_EVENTSEL0 + 16, MSR_ARCH_PERFMON_EVENTSEL0 + 17,
> +
> +	MSR_IA32_PMC0, MSR_IA32_PMC0 + 1, MSR_IA32_PMC0 + 2,
> +	MSR_IA32_PMC0 + 3, MSR_IA32_PMC0 + 4, MSR_IA32_PMC0 + 5,
> +	MSR_IA32_PMC0 + 6, MSR_IA32_PMC0 + 7, MSR_IA32_PMC0 + 8,
> +	MSR_IA32_PMC0 + 9, MSR_IA32_PMC0 + 10, MSR_IA32_PMC0 + 11,
> +	MSR_IA32_PMC0 + 12, MSR_IA32_PMC0 + 13, MSR_IA32_PMC0 + 14,
> +	MSR_IA32_PMC0 + 15, MSR_IA32_PMC0 + 16, MSR_IA32_PMC0 + 17,
>  };

This is not needed because the full-width content is already accessible
from the host via MSR_IA32_PERFCTRn.

Given the bugs, it is clear that you should also modify the pmu.c
testcase for kvm-unit-tests to cover full-width writes (and especially
the non-full-width write behavior of MSR_IA32_PERFCTRn).  Even before
the QEMU side is begin worked on, you can test it with "-cpu
host,migratable=off".

Thanks,

Paolo