Received-SPF: pass (google.com: domain of linux-kernel+bounces-182513-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:45d1:ec00::1 as permitted sender) client-ip=2604:1380:45d1:ec00::1;
Date: Fri, 17 May 2024 11:17:55 -0700
In-Reply-To: <2450ce49-2230-45a2-bc0d-b21071f2cce6@redhat.com>
Precedence: bulk
Mime-Version: 1.0
References: <20240507154459.3950778-1-pbonzini@redhat.com> <20240507154459.3950778-8-pbonzini@redhat.com>
 <ZkVHh49Hn8gB3_9o@google.com> <7c0bbec7-fa5c-4f55-9c08-ca0e94e68f7c@redhat.com>
 <ZkeH8agqiHzay5r9@google.com> <2450ce49-2230-45a2-bc0d-b21071f2cce6@redhat.com>
Message-ID: <ZkefU_PhjvnaEE7Q@google.com>
Subject: Re: [PATCH 7/7] KVM: VMX: Introduce test mode related to EPT
 violation VE
From: Sean Christopherson <seanjc@google.com>
To: Paolo Bonzini <pbonzini@redhat.com>
Cc: linux-kernel@vger.kernel.org, kvm@vger.kernel.org, 
	Isaku Yamahata <isaku.yamahata@intel.com>
Content-Type: text/plain; charset="us-ascii"

On Fri, May 17, 2024, Paolo Bonzini wrote:
> On 5/17/24 18:38, Sean Christopherson wrote:
> > > > I've hit this three times now when running KVM-Unit-Tests (I'm pretty sure it's
> > > > the EPT test, unsurprisingly).  And unless I screwed up my testing, I verified it
> > > > still fires with Isaku's fix[*], though I'm suddenly having problems repro'ing.
> > > > 
> > > > I'll update tomorrow as to whether I botched my testing of Isaku's fix, or if
> > > > there's another bug lurking.
> > > > 
> > > > https://lore.kernel.org/all/20240515173209.GD168153@ls.amr.corp.intel.com
> > > I cannot reproduce it on a Skylake (Xeon Gold 5120), with or without Isaku's
> > > fix, with either ./runtests.sh or your reproducer line.
> > > 
> > > However I can reproduce it only if eptad=0 and with the following line:
> > > 
> > > ./x86/run x86/vmx.flat -smp 1 -cpu max,host-phys-bits,+vmx -m 2560 \
> > >    -append 'ept_access_test_not_present ept_access_test_read_only'
> > 
> > FWIW, I tried that on RPL, still no failure.
> 
> Ok, so it does look like a CPU issue.  Even with the fixes you identified, I
> don't see any other solution than adding scary text in Kconfig, defaulting
> it to "n", and adding an also-very-scary pr_err_once("...") the first time
> VMPTRLD is executed with CONFIG_KVM_INTEL_PROVE_VE.

I don't think we need to make it super scary, at least not yet.  KVM just needs
to not kill the VM, which thanks to the BUSY flag is trivial: just resume the guest.
Then the failure is "just" a WARN, which won't be anywhere near as problematic for
KVM developers.  I doubt syzbot will hit this, purely because syzbot runs almost
exclusively in VMs, i.e. won't have #VE support.

If we don't have a resolution by rc6 or so, then maybe consider doing something
more drastic?

I agree that it should be off by default though.  And the help text should be
more clear that this intended only for developers and testing environments.

I have a handful of patches, including one to not kill the VM.  I'll try to post
them later today, mostly just need to write changelogs.

diff --git a/arch/x86/kvm/Kconfig b/arch/x86/kvm/Kconfig
index 75082c4a9ac4..5c22186671e9 100644
--- a/arch/x86/kvm/Kconfig
+++ b/arch/x86/kvm/Kconfig
@@ -98,15 +98,15 @@ config KVM_INTEL
 
 config KVM_INTEL_PROVE_VE
         bool "Check that guests do not receive #VE exceptions"
-        default KVM_PROVE_MMU || DEBUG_KERNEL
-        depends on KVM_INTEL
+        depends on KVM_INTEL && KVM_PROVE_MMU
         help
-
           Checks that KVM's page table management code will not incorrectly
           let guests receive a virtualization exception.  Virtualization
           exceptions will be trapped by the hypervisor rather than injected
           in the guest.
 
+          This should never be enabled in a production environment.
+
           If unsure, say N.
 
 config X86_SGX_KVM