Date:   Tue, 14 Mar 2023 17:47:43 -0700
In-Reply-To: <9f8c89bf44a8eb68407369d26956c24082562cd2.camel@intel.com>
Mime-Version: 1.0
References: <20230310214232.806108-1-seanjc@google.com> <20230310214232.806108-6-seanjc@google.com>
 <ce2330db94b05605a0649a3da0595211c5bd71dd.camel@intel.com>
 <ZA9avcHRPoIqZP/n@google.com> <9f8c89bf44a8eb68407369d26956c24082562cd2.camel@intel.com>
Message-ID: <ZBEVc0/vD5tEj29e@google.com>
Subject: Re: [PATCH v2 05/18] x86/reboot: KVM: Disable SVM during reboot via
 virt/KVM reboot callback
From:   Sean Christopherson <seanjc@google.com>
To:     Kai Huang <kai.huang@intel.com>
Cc:     Chao Gao <chao.gao@intel.com>,
        "dave.hansen@linux.intel.com" <dave.hansen@linux.intel.com>,
        "bp@alien8.de" <bp@alien8.de>, "x86@kernel.org" <x86@kernel.org>,
        "mingo@redhat.com" <mingo@redhat.com>,
        "tglx@linutronix.de" <tglx@linutronix.de>,
        "andrew.cooper3@citrix.com" <andrew.cooper3@citrix.com>,
        "pbonzini@redhat.com" <pbonzini@redhat.com>,
        "kvm@vger.kernel.org" <kvm@vger.kernel.org>,
        "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>
Content-Type: text/plain; charset="iso-8859-1"
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

On Tue, Mar 14, 2023, Huang, Kai wrote:
> On Mon, 2023-03-13 at 10:18 -0700, Sean Christopherson wrote:
> > On Mon, Mar 13, 2023, Huang, Kai wrote:
> > > On Fri, 2023-03-10 at 13:42 -0800, Sean Christopherson wrote:
> > > > Use the virt callback to disable SVM (and set GIF=3D1) during an em=
ergency
> > > > instead of blindly attempting to disable SVM.=EF=BF=BD Like the VMX=
 case, if KVM
> > > > (or an out-of-tree hypervisor) isn't loaded/active, SVM can't be in=
 use.
> > > >=20
> > > > Signed-off-by: Sean Christopherson <seanjc@google.com>
> > >=20
> > > [...]
> > >=20
> > > > -#if IS_ENABLED(CONFIG_KVM_INTEL)
> > > > +#if IS_ENABLED(CONFIG_KVM_INTEL) || IS_ENABLED(CONFIG_KVM_AMD)
> > > > =EF=BF=BD/* RCU-protected callback to disable virtualization prior =
to reboot. */
> > > > =EF=BF=BDstatic cpu_emergency_virt_cb __rcu *cpu_emergency_virt_cal=
lback;
> > > > =EF=BF=BD
> > > > @@ -821,7 +821,7 @@ int crashing_cpu =3D -1;
> > > > =EF=BF=BD */
> > > > =EF=BF=BDvoid cpu_emergency_disable_virtualization(void)
> > > > =EF=BF=BD{
> > > > -#if IS_ENABLED(CONFIG_KVM_INTEL)
> > > > +#if IS_ENABLED(CONFIG_KVM_INTEL) || IS_ENABLED(CONFIG_KVM_AMD)
> > > > =EF=BF=BD	cpu_emergency_virt_cb *callback;
> > > > =EF=BF=BD
> > > > =EF=BF=BD	rcu_read_lock();
> > > > @@ -830,8 +830,6 @@ void cpu_emergency_disable_virtualization(void)
> > > > =EF=BF=BD		callback();
> > > > =EF=BF=BD	rcu_read_unlock();
> > > > =EF=BF=BD#endif
> > > > -	/* KVM_AMD doesn't yet utilize the common callback. */
> > > > -	cpu_emergency_svm_disable();
> > > > =EF=BF=BD}
> > >=20
> > > Shouldn't the callback be always present since you want to consider '=
out-of-
> > > tree' hypervisor case?
> >=20
> > No?  The kernel doesn't provide any guarantees for out-of-tree code.  I=
 don't have
> > a super strong preference, though I do like the effective documentation=
 the checks
> > provide.  Buy more importantly, my understanding is that the x86 mainta=
iners want
> > to limit the exposure for these types of interfaces, e.g. `git grep IS_=
ENABLED\(CONFIG_KVM`
> > for a variety of hooks that are enabled iff KVM is enabled in the kerne=
l config.
>=20
> How about doing the embracing the code to do the emergency virt callback =
as the
> last step?

I like that idea, it also makes a few other patches a bit cleaner.

> I like the "cleanup" work in this series regardless whether we should gua=
rd the
> emergency virt callback with CONFIG_KVM_INTEL || CONFIG_KVM_AMD.  If we d=
o the
> actual "cleanup" work first, and put the CONFIG_KVM_INTEL || CONFIG_KVM_A=
MD as
> the last step, it is also easier to review I guess, because it's kinda a
> separate logic independent to the actual "cleanup" work.
>=20
> Also, personally I don't particularly like the middle state in patch 04:
>=20
>  void cpu_emergency_disable_virtualization(void)
>  {
>  #if IS_ENABLED(CONFIG_KVM_INTEL)
> -	cpu_crash_vmclear_loaded_vmcss();
> -#endif
> +	cpu_emergency_virt_cb *callback;
> =20
> -	cpu_emergency_vmxoff();
> +	rcu_read_lock();
> +	callback =3D rcu_dereference(cpu_emergency_virt_callback);
> +	if (callback)
> +		callback();
> +	rcu_read_unlock();
> +#endif
> +	/* KVM_AMD doesn't yet utilize the common callback. */
>  	cpu_emergency_svm_disable();
>  }
>=20
> Which eventually got fixed up in patch 05:
>=20
>  void cpu_emergency_disable_virtualization(void)
>  {
> -#if IS_ENABLED(CONFIG_KVM_INTEL)
> +#if IS_ENABLED(CONFIG_KVM_INTEL) || IS_ENABLED(CONFIG_KVM_AMD)
>  	cpu_emergency_virt_cb *callback;
> =20
>  	rcu_read_lock();
> @@ -830,8 +830,6 @@ void cpu_emergency_disable_virtualization(void)
>  		callback();
>  	rcu_read_unlock();
>  #endif
> -	/* KVM_AMD doesn't yet utilize the common callback. */
> -	cpu_emergency_svm_disable();
>  }
> =20
> Could we just merge the two patches together?=20

I'd prefer not to squash the two.  I agree it's ugly, but I dislike convert=
ing
VMX and SVM at the same time.  I'm not totally opposed to moving everything=
 in
one fell swoop, but my preference is to keep them separate.