DMARC-Filter: OpenDMARC Filter v1.3.2 mx1.redhat.com 4BAA05AFC9
From: Bandan Das <bsd@redhat.com>
To: Radim =?utf-8?B?S3LEjW3DocWZ?= <rkrcmar@redhat.com>
Cc: David Hildenbrand <david@redhat.com>, kvm@vger.kernel.org,
        pbonzini@redhat.com, linux-kernel@vger.kernel.org
Subject: Re: [PATCH v4 3/3] KVM: nVMX: Emulate EPTP switching for the L1 hypervisor
References: <20170710204936.4001-1-bsd@redhat.com>
        <20170710204936.4001-4-bsd@redhat.com>
        <2d50ebc4-9328-ce08-b55b-6a331ee13cc3@redhat.com>
        <jpgy3ruam9i.fsf@linux.bootlegged.copy>
        <jpg1spmsugn.fsf@linux.bootlegged.copy> <20170711193235.GE3326@potion>
        <jpgd1964usz.fsf@linux.bootlegged.copy>
        <20170711202118.GC28875@potion>
        <jpgzica3e83.fsf@linux.bootlegged.copy> <20170711204521.GF3326@potion>
Date: Tue, 11 Jul 2017 17:08:48 -0400
In-Reply-To: <20170711204521.GF3326@potion> ("Radim \=\?utf-8\?B\?S3LEjW3DocWZ\?\=
 \=\?utf-8\?B\?Iidz\?\= message of "Tue,
        11 Jul 2017 22:45:21 +0200")
Message-ID: <jpgeftmwujj.fsf@linux.bootlegged.copy>
User-Agent: Gnus/5.13 (Gnus v5.13) Emacs/25.2 (gnu/linux)
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: 8BIT
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 4825
Lines: 114

Radim Krčmář <rkrcmar@redhat.com> writes:

> 2017-07-11 16:34-0400, Bandan Das:
>> Radim Krčmář <rkrcmar@redhat.com> writes:
>> 
>> > 2017-07-11 15:50-0400, Bandan Das:
>> >> Radim Krčmář <rkrcmar@redhat.com> writes:
>> >> > 2017-07-11 14:24-0400, Bandan Das:
>> >> >> Bandan Das <bsd@redhat.com> writes:
>> >> >> > If there's a triple fault, I think it's a good idea to inject it
>> >> >> > back. Basically, there's no need to take care of damage control
>> >> >> > that L1 is intentionally doing.
>> >> >> >
>> >> >> >>> +			goto fail;
>> >> >> >>> +		kvm_mmu_unload(vcpu);
>> >> >> >>> +		vmcs12->ept_pointer = address;
>> >> >> >>> +		kvm_mmu_reload(vcpu);
>> >> >> >>
>> >> >> >> I was thinking about something like this:
>> >> >> >>
>> >> >> >> kvm_mmu_unload(vcpu);
>> >> >> >> old = vmcs12->ept_pointer;
>> >> >> >> vmcs12->ept_pointer = address;
>> >> >> >> if (kvm_mmu_reload(vcpu)) {
>> >> >> >> 	/* pointer invalid, restore previous state */
>> >> >> >> 	kvm_clear_request(KVM_REQ_TRIPLE_FAULT, vcpu);
>> >> >> >> 	vmcs12->ept_pointer = old;
>> >> >> >> 	kvm_mmu_reload(vcpu);
>> >> >> >> 	goto fail;
>> >> >> >> }
>> >> >> >>
>> >> >> >> The you can inherit the checks from mmu_check_root().
>> >> >> 
>> >> >> Actually, thinking about this a bit more, I agree with you. Any fault
>> >> >> with a vmfunc operation should end with a vmfunc vmexit, so this
>> >> >> is a good thing to have. Thank you for this idea! :)
>> >> >
>> >> > SDM says
>> >> >
>> >> >   IF tent_EPTP is not a valid EPTP value (would cause VM entry to fail
>> >> >   if in EPTP) THEN VMexit;
>> >> 
>> >> This section here:
>> >> As noted in Section 25.5.5.2, an execution of the
>> >> EPTP-switching VM function that causes a VM exit (as specified
>> >> above), uses the basic exit reason 59, indicating “VMFUNC”.
>> >> The length of the VMFUNC instruction is saved into the
>> >> VM-exit instruction-length field. No additional VM-exit
>> >> information is provided.
>> >> 
>> >> Although, it adds (as specified above), from testing, any vmexit that
>> >> happens as a result of the execution of the vmfunc instruction always
>> >> has exit reason 59.
>> >> 
>> >> IMO, the case David pointed out comes under "as a result of the
>> >> execution of the vmfunc instruction", so I would prefer exiting
>> >> with reason 59.
>> >
>> > Right, the exit reason is 59 for reasons that trigger a VM exit
>> > (i.e. invalid EPTP value, the four below), but kvm_mmu_reload() checks
>> > unrelated stuff.
>> >
>> > If the EPTP value is correct, then the switch should succeed.
>> > If the EPTP is correct, but bogus, then the guest should get
>> > EPT_MISCONFIG VM exit on its first access (when reading the
>> > instruction).  Source: I added
>> 
>> My point is that we are using kvm_mmu_reload() to emulate eptp
>> switching. If that emulation of vmfunc fails, it should exit with reason
>> 59.
>
> Yeah, we just disagree on what is a vmfunc failure.
>
>> >   vmcs_write64(EPT_POINTER, vmcs_read64(EPT_POINTER) | (1ULL << 40));
>> >
>> > shortly before a VMLAUNCH on L0. :)
>> 
>> What happens if this ept pointer is actually in the eptp list and the guest
>> switches to it using vmfunc ? I think it will exit with reason 59.
>
> I think otherwise, because it doesn't cause a VM entry failure on
> bare-metal (and SDM says that we get a VM exit only if there would be a
> VM entry failure).
> I expect the vmfunc to succeed and to get a EPT_MISCONFIG right after.
> (Experiment pending :])
>
>> > I think that we might be emulating this case incorrectly and throwing
>> > triple faults when it should be VM exits in vcpu_run().
>> 
>> No, I agree with not throwing a triple fault. We should clear it out.
>> But we should emulate a vmfunc vmexit back to L1 when kvm_mmu_load fails.
>
> Here we disagree.  I think that it's a bug do the VM exit, so we can

Why do you think it's a bug ? The eptp switching function really didn't
succeed as far as our emulation goes when kvm_mmu_reload() fails.
And as such, the generic vmexit failure event should be a vmfunc vmexit.
We cannot strictly follow the spec here, the spec doesn't even mention a way
to emulate eptp switching. If setting up the switching succeeded and the
new root pointer is invalid or whatever, I really don't care what happens
next but this is not the case. We fail to get a new root pointer and without
that, we can't even make a switch!

> just keep the original bug -- we want to eventually fix it and it's no
> worse till then.

Anyway, can you please confirm again what is the behavior that you
are expecting if kvm_mmu_reload fails ? This would be a rarely used
branch and I am actually fine diverging from what I think is right if
I can get the reviewers to agree on a common thing.

(Thanks for giving this a closer look, Radim. I really appreciate it.)

Bandan