Re: [PATCH] KVM: MMU: Segregate mmu pages created with different cr4.pge settings

From: Marcelo Tosatti <mtosatti@redhat.com>
To: Avi Kivity <avi@redhat.com>
Cc: Alexander Graf <agraf@suse.de>,
	"kvm@vger.kernel.org" <kvm@vger.kernel.org>,
	joerg.roedel@amd.com
Subject: Re: [PATCH] KVM: MMU: Segregate mmu pages created with different cr4.pge settings
Date: Tue, 6 Jan 2009 14:43:11 -0200	[thread overview]
Message-ID: <20090106164311.GA4902@amt.cnet> (raw)
In-Reply-To: <49636AE7.4090108@redhat.com>

On Tue, Jan 06, 2009 at 04:29:59PM +0200, Avi Kivity wrote:
> Yes...
>
> Looks like kvm_unsync_page can be folded into mmu_need_write_protect  
> (after which we can drop lookup_page(), which is not a good API).  But  
> that's after we solve the current problem.
>
> Looks like the addition of a second role for non-pge mode confuses the  
> mmu.  After the second page is created, mmu_need_write_protect() will  
> return 1, but previously existing sptes can still be writable?
>
> Looks like we need to call rmap_write_protect() when the new page is  
> created.

I'm not sure about the details, but I suspect that multiple shadows
confuse NPT somehow.

Alexander can you give this a try:

diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 36aa576..2c6579e 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -363,7 +363,8 @@ void kvm_set_cr4(struct kvm_vcpu *vcpu, unsigned long cr4)
 	}
 	kvm_x86_ops->set_cr4(vcpu, cr4);
 	vcpu->arch.cr4 = cr4;
-	vcpu->arch.mmu.base_role.cr4_pge = !!(cr4 & X86_CR4_PGE);
+	if (!tdp_enabled)
+		vcpu->arch.mmu.base_role.cr4_pge = !!(cr4 & X86_CR4_PGE);
 	kvm_mmu_sync_global(vcpu);
 	kvm_mmu_reset_context(vcpu);
 }
