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;
Date:   Fri, 18 Jun 2021 20:53:48 +0000
From:   Sean Christopherson <seanjc@google.com>
To:     Maxim Levitsky <mlevitsk@redhat.com>
Cc:     kvm@vger.kernel.org, Paolo Bonzini <pbonzini@redhat.com>,
        open list <linux-kernel@vger.kernel.org>,
        "open list:DOCUMENTATION" <linux-doc@vger.kernel.org>,
        Wanpeng Li <wanpengli@tencent.com>,
        Ingo Molnar <mingo@redhat.com>,
        Vitaly Kuznetsov <vkuznets@redhat.com>,
        Thomas Gleixner <tglx@linutronix.de>,
        Borislav Petkov <bp@alien8.de>, Joerg Roedel <joro@8bytes.org>,
        "maintainer:X86 ARCHITECTURE (32-BIT AND 64-BIT)" <x86@kernel.org>,
        Jim Mattson <jmattson@google.com>,
        Jonathan Corbet <corbet@lwn.net>,
        "H. Peter Anvin" <hpa@zytor.com>
Subject: Re: [PATCH v3 8/8] KVM: x86: avoid loading PDPTRs after migration
 when possible
Message-ID: <YM0H3Hvs8/3+twnc@google.com>
References: <20210607090203.133058-1-mlevitsk@redhat.com>
 <20210607090203.133058-9-mlevitsk@redhat.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20210607090203.133058-9-mlevitsk@redhat.com>
Precedence: bulk

> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index 11260e83518f..eadfc9caf500 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -815,6 +815,8 @@ int load_pdptrs(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu, unsigned long cr3)
>  
>  	memcpy(mmu->pdptrs, pdpte, sizeof(mmu->pdptrs));
>  	kvm_register_mark_dirty(vcpu, VCPU_EXREG_PDPTR);
> +	vcpu->arch.pdptrs_restored_oob = false;
> +
>  out:
>  
>  	return ret;
> @@ -10113,6 +10115,7 @@ static int __set_sregs2(struct kvm_vcpu *vcpu, struct kvm_sregs2 *sregs2)
>  
>  		kvm_register_mark_dirty(vcpu, VCPU_EXREG_PDPTR);
>  		mmu_reset_needed = 1;
> +		vcpu->arch.pdptrs_restored_oob = true;

Setting pdptrs_restored_oob[*] here and _only_ clearing it on successful
load_pdptrs() is not robust.  Potential problems once the flag is set:

  1.  Userspace calls KVM_SET_SREGS{,2} without valid PDPTRs.  Flag is now stale.
  2.  kvm_check_nested_events() VM-Exits to L1 before the flag is processed.
      Flag is now stale.

(2) might not be problematic in practice since the "normal" load_pdptrs()
should reset the flag on the next VM-Enter, but it's really, really hard to tell.
E.g. what if an SMI causes an exit and _that_ non-VM-Enter reload of L2 state
is the first to trip the flag?  The bool is essentially an extension of
KVM_REQ_GET_NESTED_STATE_PAGES, I think it makes sense to clear the flag whenever
KVM_REQ_GET_NESTED_STATE_PAGES is cleared.

Another thing that's not obvious is the required ordering between KVM_SET_SREGS2
and KVM_SET_NESTED_STATE.  AFAICT it's not documented, but that may be PEBKAC on
my end.  E.g. what happens if walk_mmu == &root_mmu (L1 active in targte KVM)
when SET_SREGS2 is called, and _then_ KVM_SET_NESTED_STATE is called?

[*] pdptrs_from_userspace in Paolo's tree.