linux-kernel - Re: [PATCH v3 2/4] KVM: x86: Partially allow KVM_SET_CPUID{,2} after KVM

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:   Wed, 19 Jan 2022 11:02:43 +0100
From:   Vitaly Kuznetsov <vkuznets@...hat.com>
To:     Sean Christopherson <seanjc@...gle.com>
Cc:     kvm@...r.kernel.org, Paolo Bonzini <pbonzini@...hat.com>,
        Wanpeng Li <wanpengli@...cent.com>,
        Jim Mattson <jmattson@...gle.com>,
        Igor Mammedov <imammedo@...hat.com>,
        linux-kernel@...r.kernel.org
Subject: Re: [PATCH v3 2/4] KVM: x86: Partially allow KVM_SET_CPUID{,2}
 after KVM_RUN

Sean Christopherson <seanjc@...gle.com> writes:

> On Tue, Jan 18, 2022, Vitaly Kuznetsov wrote:
>> +/* Check whether the supplied CPUID data is equal to what is already set for the vCPU. */
>> +static int kvm_cpuid_check_equal(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *e2,
>> +				 int nent)
>> +{
>> +	struct kvm_cpuid_entry2 *orig;
>> +	int i;
>> +
>> +	if (nent != vcpu->arch.cpuid_nent)
>> +		return -EINVAL;
>> +
>> +	for (i = 0; i < nent; i++) {
>> +		orig = &vcpu->arch.cpuid_entries[i];
>> +		if (e2[i].function != orig->function ||
>> +		    e2[i].index != orig->index ||
>> +		    e2[i].eax != orig->eax || e2[i].ebx != orig->ebx ||
>> +		    e2[i].ecx != orig->ecx || e2[i].edx != orig->edx)
>> +			return -EINVAL;
>
> This needs to check .flags for the above check on .index to be meaningful, and at
> that point, can't we be even more agressive and just do?
>
> 	if (memcmp(e2, vcpu->arch.cpuid_entries, nent * sizeof(e2)))
> 		return -EINVAL;
>
> 	return 0;
>

Sure, looks good to me.

>> +	}
>> +
>> +	return 0;
>> +}
>> +
>>  static void kvm_update_kvm_cpuid_base(struct kvm_vcpu *vcpu)
>>  {
>>  	u32 function;
>> @@ -313,6 +335,20 @@ static int kvm_set_cpuid(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *e2,
>>  
>>  	__kvm_update_cpuid_runtime(vcpu, e2, nent);
>> +	/*
>> +	 * KVM does not correctly handle changing guest CPUID after KVM_RUN, as
>> +	 * MAXPHYADDR, GBPAGES support, AMD reserved bit behavior, etc.. aren't
>> +	 * tracked in kvm_mmu_page_role.  As a result, KVM may miss guest page
>> +	 * faults due to reusing SPs/SPTEs. In practice no sane VMM mucks with
>> +	 * the core vCPU model on the fly. It would've been better to forbid any
>> +	 * KVM_SET_CPUID{,2} calls after KVM_RUN altogether but unfortunately
>> +	 * some VMMs (e.g. QEMU) reuse vCPU fds for CPU hotplug/unplug and do
>> +	 * KVM_SET_CPUID{,2} again. To support this legacy behavior, check
>> +	 * whether the supplied CPUID data is equal to what's already set.
>
> This is misleading/wrong. KVM_RUN isn't the only problematic ioctl(),

Well, it wasn't me who wrote the comment about KVM_RUN :-) My addition
can be improved of course.

> it's just the one that we decided to use to detect that userspace is
> being stupid.  And forbidding KVM_SET_CPUID after KVM_RUN (or even all
> problematic ioctls()) wouldn't solve problem as providing different
> CPUID configurations for vCPUs in a VM will also cause the MMU to fall
> on its face.

True, but how do we move forward? We can either let userspace do stupid
things and (potentially) create hard-to-debug problems or we try to
cover at least some use-cases with checks (like the one we introduce
here).

Different CPUID configurations for different vCPUs is actually an
interesting case. It makes me (again) think about the
allowlist/blocklist approaches: we can easily enhance the
'vcpu->arch.last_vmentry_cpu != -1' check below and start requiring
CPUIDs to [almost] match. The question then is how to change CPUID for a
multi-vCPU guest as it will become effectively forbidden. BTW, is there
a good use-case for changing CPUIDs besides testing purposes?

>
>> +	if (vcpu->arch.last_vmentry_cpu != -1)
>> +		return kvm_cpuid_check_equal(vcpu, e2, nent);
>
> And technically, checking last_vmentry_cpu doesn't forbid changing CPUID after
> KVM_RUN, it forbids changing CPUID after successfully entering the guest (or
> emulating instructions on VMX).
>
> I realize I'm being very pedantic, as a well-intended userspace is obviously not
> going to change CPUID after -EINTR or whatever.  But I do want to highlight that
> this approach is by no means bulletproof, and that what is/isn't allowed with
> respect to guest CPUID isn't necessarily associated with what is/isn't "safe".
> In other words, this check doesn't guarantee that userspace can't misuse KVM_SET_CPUID,
> and on the flip side it disallows using KVM_SET_CPUID in ways that are perfectly ok
> (if userspace is careful and deliberate).

All true but I don't see a 'bulletproof' approach here unless we start
designing new KVM API for userspace and I don't think the problem here
is a good enough justification for that. Another approach would be to
name the "don't change CPUIDs after KVM_RUN at will" comment in the code
a good enough sentinel and hope that no real world userspace actually
does such things.

-- 
Vitaly