[PATCH v3 15/15] KVM: x86: Disallow guest CPUID lookups when IRQs are disabled

From: Sean Christopherson
Date: Tue Aug 15 2023 - 16:38:56 EST


Now that KVM has a framework for caching guest CPUID feature flags, add
a "rule" that IRQs must be enabled when doing guest CPUID lookups, and
enforce the rule via a lockdep assertion. CPUID lookups are slow, and
within KVM, IRQs are only ever disabled in hot paths, e.g. the core run
loop, fast page fault handling, etc. I.e. querying guest CPUID with IRQs
disabled, especially in the run loop, should be avoided.

Signed-off-by: Sean Christopherson <seanjc@xxxxxxxxxx>
---
arch/x86/kvm/cpuid.c | 13 +++++++++++++
1 file changed, 13 insertions(+)

diff --git a/arch/x86/kvm/cpuid.c b/arch/x86/kvm/cpuid.c
index 67e9f79fe059..e961e9a05847 100644
--- a/arch/x86/kvm/cpuid.c
+++ b/arch/x86/kvm/cpuid.c
@@ -11,6 +11,7 @@
#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt

#include <linux/kvm_host.h>
+#include "linux/lockdep.h"
#include <linux/export.h>
#include <linux/vmalloc.h>
#include <linux/uaccess.h>
@@ -84,6 +85,18 @@ static inline struct kvm_cpuid_entry2 *cpuid_entry2_find(
struct kvm_cpuid_entry2 *e;
int i;

+ /*
+ * KVM has a semi-arbitrary rule that querying the guest's CPUID model
+ * with IRQs disabled is disallowed. The CPUID model can legitimately
+ * have over one hundred entries, i.e. the lookup is slow, and IRQs are
+ * typically disabled in KVM only when KVM is in a performance critical
+ * path, e.g. the core VM-Enter/VM-Exit run loop. Nothing will break
+ * if this rule is violated, this assertion is purely to flag potential
+ * performance issues. If this fires, consider moving the lookup out
+ * of the hotpath, e.g. by caching information during CPUID updates.
+ */
+ lockdep_assert_irqs_enabled();
+
for (i = 0; i < nent; i++) {
e = &entries[i];

--
2.41.0.694.ge786442a9b-goog