From: Nadav Har'El Date: Mon, 5 Aug 2013 08:07:08 +0000 (+0300) Subject: nEPT: Fix wrong test in kvm_set_cr3 X-Git-Tag: v3.12-rc1~143^2~39 X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=b7e914501c1d92edc0c30e3bae35c14481642e55;p=kernel%2Fkernel-generic.git nEPT: Fix wrong test in kvm_set_cr3 kvm_set_cr3() attempts to check if the new cr3 is a valid guest physical address. The problem is that with nested EPT, cr3 is an *L2* physical address, not an L1 physical address as this test expects. As the comment above this test explains, it isn't necessary, and doesn't correspond to anything a real processor would do. So this patch removes it. Note that this wrong test could have also theoretically caused problems in nested NPT, not just in nested EPT. However, in practice, the problem was avoided: nested_svm_vmexit()/vmrun() do not call kvm_set_cr3 in the nested NPT case, and instead set the vmcb (and arch.cr3) directly, thus circumventing the problem. Additional potential calls to the buggy function are avoided in that we don't trap cr3 modifications when nested NPT is enabled. However, because in nested VMX we did want to use kvm_set_cr3() (as requested in Avi Kivity's review of the original nested VMX patches), we can't avoid this problem and need to fix it. Reviewed-by: Orit Wasserman Reviewed-by: Xiao Guangrong Signed-off-by: Nadav Har'El Signed-off-by: Jun Nakajima Signed-off-by: Xinhao Xu Signed-off-by: Yang Zhang Signed-off-by: Gleb Natapov Signed-off-by: Paolo Bonzini --- diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c index 05afd1a..668f19a 100644 --- a/arch/x86/kvm/x86.c +++ b/arch/x86/kvm/x86.c @@ -682,17 +682,6 @@ int kvm_set_cr3(struct kvm_vcpu *vcpu, unsigned long cr3) */ } - /* - * Does the new cr3 value map to physical memory? (Note, we - * catch an invalid cr3 even in real-mode, because it would - * cause trouble later on when we turn on paging anyway.) - * - * A real CPU would silently accept an invalid cr3 and would - * attempt to use it - with largely undefined (and often hard - * to debug) behavior on the guest side. - */ - if (unlikely(!gfn_to_memslot(vcpu->kvm, cr3 >> PAGE_SHIFT))) - return 1; vcpu->arch.cr3 = cr3; __set_bit(VCPU_EXREG_CR3, (ulong *)&vcpu->arch.regs_avail); vcpu->arch.mmu.new_cr3(vcpu);