Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Wed, 20 Jul 2022 17:38:58 +0000
From:   Sean Christopherson <seanjc@google.com>
To:     Chao Gao <chao.gao@intel.com>
Cc:     Kechen Lu <kechenl@nvidia.com>, kvm@vger.kernel.org,
        pbonzini@redhat.com, vkuznets@redhat.com, somduttar@nvidia.com,
        linux-kernel@vger.kernel.org
Subject: Re: [RFC PATCH v3 5/7] KVM: x86: add vCPU scoped toggling for
 disabled exits
Message-ID: <Ytg9shoNq9XfTiHS@google.com>
References: <20220615011622.136646-1-kechenl@nvidia.com>
 <20220615011622.136646-6-kechenl@nvidia.com>
 <20220615024311.GA7808@gao-cwp>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20220615024311.GA7808@gao-cwp>
Precedence: bulk

On Wed, Jun 15, 2022, Chao Gao wrote:
> >@@ -5980,6 +5987,8 @@ int kvm_vm_ioctl_irq_line(struct kvm *kvm, struct kvm_irq_level *irq_event,
> > int kvm_vm_ioctl_enable_cap(struct kvm *kvm,
> > 			    struct kvm_enable_cap *cap)
> > {
> >+	struct kvm_vcpu *vcpu;
> >+	unsigned long i;
> > 	int r;
> > 
> > 	if (cap->flags)
> >@@ -6036,14 +6045,17 @@ int kvm_vm_ioctl_enable_cap(struct kvm *kvm,
> > 			break;
> > 
> > 		mutex_lock(&kvm->lock);
> >-		if (kvm->created_vcpus)
> >-			goto disable_exits_unlock;
> >+		if (kvm->created_vcpus) {
> >+			kvm_for_each_vcpu(i, vcpu, kvm) {
> >+				kvm_ioctl_disable_exits(vcpu->arch, cap->args[0]);
> >+				static_call(kvm_x86_update_disabled_exits)(vcpu);
> 
> IMO, this won't work on Intel platforms.

It's not safe on AMD either because at best the behavior is non-deterministic if
the vCPU is already running in the guest, and at worst could cause explosions,
e.g. if hardware doesn't like software modifying in-use VMCB state.

> Because, to manipulate a vCPU's VMCS, vcpu_load() should be invoked in
> advance to load the VMCS.  Alternatively, you can add a request KVM_REQ_XXX
> and defer updating VMCS to the next vCPU entry.

Definitely use a request, doing vcpu_load() from a KVM-scoped ioctl() would be
a mess as KVM would need to acquire the per-vCPU lock for every vCPU.