Date:   Fri, 10 Mar 2023 21:29:36 +0000
In-Reply-To: <20230310161354.1889b539@gandalf.local.home>
Mime-Version: 1.0
References: <cover.1678474914.git.jpoimboe@kernel.org> <432e4844ba65840af4a24f5e3f561aead867f6e7.1678474914.git.jpoimboe@kernel.org>
 <ZAucD8gHx8Xp8Dlb@google.com> <20230310161354.1889b539@gandalf.local.home>
Message-ID: <ZAuhQNS8mjRt1bOG@google.com>
Subject: Re: [RFC][PATCH 5/5] x86/kvm: Simplify static call handling
From:   Sean Christopherson <seanjc@google.com>
To:     Steven Rostedt <rostedt@goodmis.org>
Cc:     Josh Poimboeuf <jpoimboe@kernel.org>, x86@kernel.org,
        linux-kernel@vger.kernel.org,
        Peter Zijlstra <peterz@infradead.org>,
        Mark Rutland <mark.rutland@arm.com>,
        Jason Baron <jbaron@akamai.com>,
        Ard Biesheuvel <ardb@kernel.org>,
        Christophe Leroy <christophe.leroy@csgroup.eu>,
        Paolo Bonzini <pbonzini@redhat.com>
Content-Type: text/plain; charset="us-ascii"
Precedence: bulk

On Fri, Mar 10, 2023, Steven Rostedt wrote:
> On Fri, 10 Mar 2023 13:07:27 -0800
> Sean Christopherson <seanjc@google.com> wrote:
> 
> > "KVM: x86:" please, "x86/kvm" is for guest-side changes.
> > 
> > On Fri, Mar 10, 2023, Josh Poimboeuf wrote:
> > > diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
> > > index 1dfba499d3e5..612531e1c478 100644
> > > --- a/arch/x86/include/asm/kvm_host.h
> > > +++ b/arch/x86/include/asm/kvm_host.h
> > > @@ -1789,8 +1789,6 @@ extern struct kvm_x86_ops kvm_x86_ops;
> > >  
> > >  #define KVM_X86_OP(func) \
> > >  	DECLARE_STATIC_CALL(kvm_x86_##func, *(((struct kvm_x86_ops *)0)->func));
> > > -#define KVM_X86_OP_OPTIONAL KVM_X86_OP
> > > -#define KVM_X86_OP_OPTIONAL_RET0 KVM_X86_OP
> > >  #include <asm/kvm-x86-ops.h>
> > >  
> > >  int kvm_x86_vendor_init(struct kvm_x86_init_ops *ops);
> > > diff --git a/arch/x86/kvm/pmu.c b/arch/x86/kvm/pmu.c
> > > index 6accb46295a3..5f7f860c5f17 100644
> > > --- a/arch/x86/kvm/pmu.c
> > > +++ b/arch/x86/kvm/pmu.c
> > > @@ -77,20 +77,15 @@ static struct kvm_pmu_ops kvm_pmu_ops __read_mostly;
> > >  #define KVM_X86_PMU_OP(func)					     \
> > >  	DEFINE_STATIC_CALL_NULL(kvm_x86_pmu_##func,			     \
> > >  				*(((struct kvm_pmu_ops *)0)->func));
> > > -#define KVM_X86_PMU_OP_OPTIONAL KVM_X86_PMU_OP
> > >  #include <asm/kvm-x86-pmu-ops.h>
> > >  
> > >  void kvm_pmu_ops_update(const struct kvm_pmu_ops *pmu_ops)
> > >  {
> > >  	memcpy(&kvm_pmu_ops, pmu_ops, sizeof(kvm_pmu_ops));
> > >  
> > > -#define __KVM_X86_PMU_OP(func) \
> > > -	static_call_update(kvm_x86_pmu_##func, kvm_pmu_ops.func);
> > >  #define KVM_X86_PMU_OP(func) \
> > > -	WARN_ON(!kvm_pmu_ops.func); __KVM_X86_PMU_OP(func)  
> > 
> > I would much prefer to keep KVM mostly as-is, specifically so that we don't lose
> > this WARN_ON() that guards against a vendor module neglecting to implement a
> > mandatory callback.  This effectively gives KVM "full" protection against consuming
> > an unexpectedly-NULL function pointer.
> 
> As in my reply to patch 0/5, I suggested that static_call_update(NULL)
> would trigger a WARN_ON() always. Then this could be cleaned up and still
> get that warning.

I don't think that provides the functionality KVM wants/needs.  KVM only disallows
NULL updates for select mandatory hooks.  For optional hooks, KVM needs to support
NULL updates in some capacity to handle the scenario where a vendor module is
reloaded with different settings, e.g. loading kvm_intel with enable_apicv=0 after
running with enable_apicv=1.

WARN_ON() a static_call_update(..., NULL) should be ok, but I believe KVM would
still need/want macro shenanigans, e.g.

#define __KVM_X86_OP(func) \
	static_call_update(kvm_x86_##func,
			   kvm_x86_ops.func ? kvm_x86_ops.func : STATIC_CALL_NOP);
#define KVM_X86_OP(func) \
	WARN_ON(!kvm_x86_ops.func); __KVM_X86_OP(func)
#define KVM_X86_OP_OPTIONAL __KVM_X86_OP
#define KVM_X86_OP_OPTIONAL_RET0(func) __KVM_X86_OP