Re: [tip: perf/core] perf: Use EXPORT_SYMBOL_FOR_KVM() for the mediated APIs

[Sean Christopherson <seanjc@google.com>]

On Thu, Dec 18, 2025, Peter Zijlstra wrote:
> On Thu, Dec 18, 2025 at 09:31:09AM +0100, Peter Zijlstra wrote:
> > On Wed, Dec 17, 2025 at 12:37:57PM -0000, tip-bot2 for Peter Zijlstra wrote:
> > > diff --git a/kernel/events/core.c b/kernel/events/core.c
> > > index e6a4b1e..376fb07 100644
> > > --- a/kernel/events/core.c
> > > +++ b/kernel/events/core.c
> > > @@ -57,6 +57,7 @@
> > >  #include <linux/task_work.h>
> > >  #include <linux/percpu-rwsem.h>
> > >  #include <linux/unwind_deferred.h>
> > > +#include <linux/kvm_types.h>
> > Bah, so the !KVM architectures hate on this.
> > 
> > Sean, would something like this be acceptable?
> 
> Hmm, the other option is doing something like so:
> 
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index 376fb07d869b..014d832e8eaa 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -57,7 +57,6 @@
>  #include <linux/task_work.h>
>  #include <linux/percpu-rwsem.h>
>  #include <linux/unwind_deferred.h>
> -#include <linux/kvm_types.h>
>  
>  #include "internal.h"
>  
> @@ -6325,6 +6324,8 @@ u64 perf_event_pause(struct perf_event *event, bool reset)
>  EXPORT_SYMBOL_GPL(perf_event_pause);
>  
>  #ifdef CONFIG_PERF_GUEST_MEDIATED_PMU
> +#include <linux/kvm_types.h>

Hrm, quick and dirty, but I don't love the idea of punting on the underlying
issue, because not being able to include kvm_types.h will be a big deterrent to
using EXPORT_SYMBOL_FOR_KVM().

>  static atomic_t nr_include_guest_events __read_mostly;
>  
>  static atomic_t nr_mediated_pmu_vms __read_mostly;
> 
> > ---
> > Subject: kvm: Fix linux/kvm_types.h for !KVM architectures
> > 
> > As is, <linux/kvm_types.h> hard relies on architectures having
> > <asm/kvm_types.h> which (obviously) breaks for architectures that don't
> > have KVM support.
> > 
> > This means generic code (kernel/events/ in this case) cannot use
> > EXPORT_SYMBOL_FOR_KVM().
> > 
> > Rearrange things just so that <linux/kvm_types.h> becomes usable and
> > provides the (expected) empty stub for EXPORT_SYMBOL_FOR_KVM() for !KVM.
> > 
> > Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
> > ---
> > diff --git a/include/linux/kvm_types.h b/include/linux/kvm_types.h
> > index a568d8e6f4e8..a4cc13e41eec 100644
> > --- a/include/linux/kvm_types.h
> > +++ b/include/linux/kvm_types.h
> > @@ -6,6 +6,8 @@
> >  #include <linux/bits.h>
> >  #include <linux/export.h>
> >  #include <linux/types.h>
> > +
> > +#ifdef CONFIG_KVM
> >  #include <asm/kvm_types.h>

This won't work, because asm/kvm_types.h #defines KVM_ARCH_NR_OBJS_PER_MEMORY_CACHE,
which guards the "struct kvm_mmu_memory_cache" definition.  E.g. on x86 with
CONFIG_KVM=n, that yields errors like:

  In file included from include/linux/kvm_host.h:45,
                   from arch/x86/events/intel/core.c:17:
  arch/x86/include/asm/kvm_host.h:854:37: error: field ‘mmu_pte_list_desc_cache’ has incomplete type
    854 |         struct kvm_mmu_memory_cache mmu_pte_list_desc_cache;
        |                                     ^~~~~~~~~~~~~~~~~~~~~~~


In general, I'm hesitant to guard an include with a conditional Kconfig, precisely
because doing so has a tendency to result in wonky, config-specific build errors.

Rather than gate the check on KVM being enabled, what if we restrict the asm
include to architectures that support KVM in any capacity?  Alternatively, we
could add a HAVE_KVM, but I'd rather not add HAVE_KVM, because then we'll end up
with the same mess if architectures get clever and conditionally select HAVE_KVM
(IIRC, that's exactly what happened when HAVE_KVM was a thing in the past).

Compiled tested on all KVM architectures along with csky (and an include of
kvm_types.h in init/main.c).

--
From: Sean Christopherson <seanjc@google.com>
Date: Thu, 18 Dec 2025 15:47:59 +0000
Subject: [PATCH] KVM: Allow linux/kvm_types.h to be included on non-KVM
 architectures

Include the arch-defined asm/kvm_types.h if and only if the kernel is
being compiled for an architecture that supports KVM so that kvm_types.h
can be included in generic code without having to guard _those_ includes,
and without having to add "generic-y += kvm_types.h" for all architectures
that don't support KVM.

Assert that KVM=n if asm/kvm_types.h isn't included to provide a more
helpful error message if an arch name changes (highly unlikely) or a new
arch that supports KVM comes along.

Signed-off-by: Sean Christopherson <seanjc@google.com>
---
 include/linux/kvm_types.h | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)

diff --git a/include/linux/kvm_types.h b/include/linux/kvm_types.h
index a568d8e6f4e8..797721e298df 100644
--- a/include/linux/kvm_types.h
+++ b/include/linux/kvm_types.h
@@ -6,7 +6,23 @@
 #include <linux/bits.h>
 #include <linux/export.h>
 #include <linux/types.h>
+
+/*
+ * Include the arch-defined kvm_types.h if and only if the target architecture
+ * supports KVM, so that linux/kvm_types.h can be included in generic code
+ * without requiring _all_ architectures to add generic-y += kvm_types.h.
+ */
+#if defined(CONFIG_ARM64)	|| \
+    defined(CONFIG_LOONGARCH)	|| \
+    defined(CONFIG_MIPS)	|| \
+    defined(CONFIG_PPC)		|| \
+    defined(CONFIG_RISCV)	|| \
+    defined(CONFIG_S390)	|| \
+    defined(CONFIG_X86)
 #include <asm/kvm_types.h>
+#else
+static_assert(!IS_ENABLED(CONFIG_KVM));
+#endif
 
 #ifdef KVM_SUB_MODULES
 #define EXPORT_SYMBOL_FOR_KVM_INTERNAL(symbol) \

base-commit: 8f0b4cce4481fb22653697cced8d0d04027cb1e8
--