Re: [PATCH 0/8] KVM: x86: Interrupt-based mechanism for async_pf 'page present' notifications

From: Vitaly Kuznetsov <vkuznets@redhat.com>
To: Vivek Goyal <vgoyal@redhat.com>
Cc: kvm@vger.kernel.org, x86@kernel.org,
	Paolo Bonzini <pbonzini@redhat.com>,
	Andy Lutomirski <luto@kernel.org>,
	Thomas Gleixner <tglx@linutronix.de>,
	Borislav Petkov <bp@alien8.de>, "H. Peter Anvin" <hpa@zytor.com>,
	Wanpeng Li <wanpengli@tencent.com>,
	Sean Christopherson <sean.j.christopherson@intel.com>,
	Jim Mattson <jmattson@google.com>, Gavin Shan <gshan@redhat.com>,
	Peter Zijlstra <peterz@infradead.org>,
	linux-kernel@vger.kernel.org
Subject: Re: [PATCH 0/8] KVM: x86: Interrupt-based mechanism for async_pf 'page present' notifications
Date: Thu, 14 May 2020 20:14:31 +0200	[thread overview]
Message-ID: <87lfluwfi0.fsf@vitty.brq.redhat.com> (raw)
In-Reply-To: <20200513141644.GD173965@redhat.com>

Vivek Goyal <vgoyal@redhat.com> writes:

> On Mon, May 11, 2020 at 06:47:44PM +0200, Vitaly Kuznetsov wrote:
>> Concerns were expressed around (ab)using #PF for KVM's async_pf mechanism,
>> it seems that re-using #PF exception for a PV mechanism wasn't a great
>> idea after all. The Grand Plan is to switch to using e.g. #VE for 'page
>> not present' events and normal APIC interrupts for 'page ready' events.
>> This series does the later.
>
> Hi Vitaly,
>
> How does any of this impact nested virtualization code (if any).
>
> I have tried understanding that logic, but I have to admit, I could
> never get it.
>
> arch/x86/kvm/mmu/mmu.c
>
> int kvm_handle_page_fault(struct kvm_vcpu *vcpu, u64 error_code,
>                                 u64 fault_address, char *insn, int insn_len)
> {
>         switch (vcpu->arch.apf.host_apf_reason) {
> 		case KVM_PV_REASON_PAGE_NOT_PRESENT:
> 			kvm_async_pf_task_wait(fault_address, 0);
> 		case KVM_PV_REASON_PAGE_READY:
> 			kvm_async_pf_task_wake(fault_address);
> 	}
> }
>

"[PATCH 8/8] KVM: x86: drop KVM_PV_REASON_PAGE_READY case from
kvm_handle_page_fault()" modifies this a little bit.

Basically (and if I understand this correctly) we have the following APF
related feature (bit 2 in MSR_KVM_ASYNC_PF_EN): "asynchronous page faults
are delivered to L1 as #PF vmexits.". When enabled, it allows L0 to
inject #PF when L2 guest is running. L1 will see this as '#PF vmexit'
and the code you cite will do exactly what do_async_page_fault() is
doing.

When we switch to interrupt based delivery for 'page ready' events we
don't need a special handling for them in L1 (as we don't need any
special handling for all interrupts from devices in kernel when KVM
guest is running).

I have to admit I haven't tested nested scenario yet, "what could go
wrong?" :-)

-- 
Vitaly