On 12/29/2022 3:41 PM, Zhang, Chen wrote:
-----Original Message-----
From: Yang, Weijiang <weijiang.yang@xxxxxxxxx>
Sent: Thursday, December 29, 2022 3:03 PM
To: Zhang, Chen <chen.zhang@xxxxxxxxx>
Cc: Gao, Chao <chao.gao@xxxxxxxxx>; Pawan Gupta
<pawan.kumar.gupta@xxxxxxxxxxxxxxx>; Paolo Bonzini
<pbonzini@xxxxxxxxxx>; Christopherson,, Sean <seanjc@xxxxxxxxxx>; H.
Peter Anvin <hpa@xxxxxxxxx>; Dave Hansen <dave.hansen@xxxxxxxxxxxxxxx>;
Borislav Petkov <bp@xxxxxxxxx>; Ingo Molnar <mingo@xxxxxxxxxx>; Thomas
Gleixner <tglx@xxxxxxxxxxxxx>; x86@xxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx;
kvm@xxxxxxxxxxxxxxx
Subject: Re: [RFC PATCH 6/9] kvm/x86: Add ARCH_CAP_VIRTUAL_ENUM for
guest MSR_IA32_ARCH_CAPABILITIES
On 12/29/2022 10:58 AM, Zhang, Chen wrote:
[...]
\
+ ARCH_CAP_VIRTUAL_ENUM)
static u64 kvm_get_arch_capabilities(void)
{
@@ -1607,6 +1611,13 @@ static u64 kvm_get_arch_capabilities(void)
*/
data |= ARCH_CAP_PSCHANGE_MC_NO;
+ /*
+ * Virtual MSRs can allow guests to notify VMM whether or not
+ * they are using specific software mitigation, allowing a VMM
+ * to enable there hardware control only where necessary.
+ */
+ data |= ARCH_CAP_VIRTUAL_ENUM;
IMO, this is: data &= ARCH_CAP_VIRTUAL_ENUM; because it requires
platform support.
Intel defined the virtual MSRs for software mitigations for all platforms.
KVM should be unconditionally opened it for the software mitigation in
migration pools.
For example migration from the old platform to the new platform.
Please check the Software Mitigations in Migration Pools section in
documents:
https://www.intel.com/content/www/us/en/developer/articles/technical/s
oftware-security-guidance/technical-documentation/branch-history-injec
tion.html
If this series running on old platforms, how VMM can set specific vmcs fields,
e.g., “virtualize IA32_SPEC_CTRL” VM-execution control, to mitigate guest
issues?
Enable the virtual MSRs does not means to enable the “virtualize IA32_SPEC_CTRL”.
KVM will check "cpu_has_virt_spec_ctrl()" before set specific VMCS.
What I'm concerned is, if the feature is not supported on some
platforms, unconditionally
exposing to guest would make it wrongly "think" the feature is
supported on HW, but actually
WMM can do nothing to mitigate issues. Read through the patch set, looks
like guest cannot
even get any errors during feature configuration in this case. So I
suggest add more
sanity checks. Did I miss something?
Thanks
Chen
+
/*
* If we're doing cache flushes (either "always" or "cond")
* we will do one whenever the guest does a vmlaunch/vmresume.
@@ -1657,6 +1668,9 @@ static int kvm_get_msr_feature(struct
kvm_msr_entry *msr)
case MSR_IA32_UCODE_REV:
rdmsrl_safe(msr->index, &msr->data);
break;
+ case MSR_VIRTUAL_ENUMERATION:
+ msr->data = VIRT_ENUM_MITIGATION_CTRL_SUPPORT;
Need to check bit 63 of host MSR_ARCH_CAPABILITIES before expose the
feature.
Refer to the above comments.
Thanks
Chen
+ break;
default:
return static_call(kvm_x86_get_msr_feature)(msr);
}
