KVM: arm64: Drop trapping of PAuth instructions/keys

We currently insist on disabling PAuth on vcpu_load(), and get to
enable it on first guest use of an instruction or a key (ignoring
the NV case for now).

It isn't clear at all what this is trying to achieve: guests tend
to use PAuth when available, and nothing forces you to expose it
to the guest if you don't want to. This also isn't totally free:
we take a full GPR save/restore between host and guest, only to
write ten 64bit registers. The "value proposition" escapes me.

So let's forget this stuff and enable PAuth eagerly if exposed to
the guest. This results in much simpler code. Performance wise,
that's not bad either (tested on M2 Pro running a fully automated
Debian installer as the workload):

- On a non-NV guest, I can see reduction of 0.24% in the number
  of cycles (measured with perf over 10 consecutive runs)

- On a NV guest (L2), I see a 2% reduction in wall-clock time
  (measured with 'time', as M2 doesn't have a PMUv3 and NV
  doesn't support it either)

So overall, a much reduced complexity and a (small) performance
improvement.
Reviewed-by: Oliver Upton <oliver.upton@linux.dev>
Link: https://lore.kernel.org/r/20240419102935.1935571-16-maz@kernel.orgSigned-off-by: Marc Zyngier <maz@kernel.org>

arch/arm64/include/asm/kvm_emulate.h

@@ -125,11 +125,6 @@ static inline void vcpu_set_wfx_traps(struct kvm_vcpu *vcpu)
 	vcpu->arch.hcr_el2 |= HCR_TWI;
 }
 
-static inline void vcpu_ptrauth_disable(struct kvm_vcpu *vcpu)
-{
-	vcpu->arch.hcr_el2 &= ~(HCR_API | HCR_APK);
-}
-
 static inline unsigned long vcpu_get_vsesr(struct kvm_vcpu *vcpu)
 {
 	return vcpu->arch.vsesr_el2;

arch/arm64/include/asm/kvm_ptrauth.h

@@ -99,5 +99,26 @@ alternative_else_nop_endif
 .macro ptrauth_switch_to_hyp g_ctxt, h_ctxt, reg1, reg2, reg3
 .endm
 #endif /* CONFIG_ARM64_PTR_AUTH */
+
+#else  /* !__ASSEMBLY */
+
+#define __ptrauth_save_key(ctxt, key)					\
+	do {								\
+		u64 __val;                                              \
+		__val = read_sysreg_s(SYS_ ## key ## KEYLO_EL1);	\
+		ctxt_sys_reg(ctxt, key ## KEYLO_EL1) = __val;		\
+		__val = read_sysreg_s(SYS_ ## key ## KEYHI_EL1);	\
+		ctxt_sys_reg(ctxt, key ## KEYHI_EL1) = __val;		\
+	} while(0)
+
+#define ptrauth_save_keys(ctxt)						\
+	do {								\
+		__ptrauth_save_key(ctxt, APIA);				\
+		__ptrauth_save_key(ctxt, APIB);				\
+		__ptrauth_save_key(ctxt, APDA);				\
+		__ptrauth_save_key(ctxt, APDB);				\
+		__ptrauth_save_key(ctxt, APGA);				\
+	} while(0)
+
 #endif /* __ASSEMBLY__ */
 #endif /* __ASM_KVM_PTRAUTH_H */

arch/arm64/kvm/arm.c

@@ -35,10 +35,11 @@
 #include <asm/virt.h>
 #include <asm/kvm_arm.h>
 #include <asm/kvm_asm.h>
+#include <asm/kvm_emulate.h>
 #include <asm/kvm_mmu.h>
 #include <asm/kvm_nested.h>
 #include <asm/kvm_pkvm.h>
-#include <asm/kvm_emulate.h>
+#include <asm/kvm_ptrauth.h>
 #include <asm/sections.h>
 
 #include <kvm/arm_hypercalls.h>
@@ -462,6 +463,44 @@ void kvm_arch_vcpu_unblocking(struct kvm_vcpu *vcpu)
 
 }
 
+static void vcpu_set_pauth_traps(struct kvm_vcpu *vcpu)
+{
+	if (vcpu_has_ptrauth(vcpu)) {
+		/*
+		 * Either we're running running an L2 guest, and the API/APK
+		 * bits come from L1's HCR_EL2, or API/APK are both set.
+		 */
+		if (unlikely(vcpu_has_nv(vcpu) && !is_hyp_ctxt(vcpu))) {
+			u64 val;
+
+			val = __vcpu_sys_reg(vcpu, HCR_EL2);
+			val &= (HCR_API | HCR_APK);
+			vcpu->arch.hcr_el2 &= ~(HCR_API | HCR_APK);
+			vcpu->arch.hcr_el2 |= val;
+		} else {
+			vcpu->arch.hcr_el2 |= (HCR_API | HCR_APK);
+		}
+
+		/*
+		 * Save the host keys if there is any chance for the guest
+		 * to use pauth, as the entry code will reload the guest
+		 * keys in that case.
+		 * Protected mode is the exception to that rule, as the
+		 * entry into the EL2 code eagerly switch back and forth
+		 * between host and hyp keys (and kvm_hyp_ctxt is out of
+		 * reach anyway).
+		 */
+		if (is_protected_kvm_enabled())
+			return;
+
+		if (vcpu->arch.hcr_el2 & (HCR_API | HCR_APK)) {
+			struct kvm_cpu_context *ctxt;
+			ctxt = this_cpu_ptr_hyp_sym(kvm_hyp_ctxt);
+			ptrauth_save_keys(ctxt);
+		}
+	}
+}
+
 void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
 {
 	struct kvm_s2_mmu *mmu;
@@ -500,8 +539,8 @@ void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
 	else
 		vcpu_set_wfx_traps(vcpu);
 
-	if (vcpu_has_ptrauth(vcpu))
-		vcpu_ptrauth_disable(vcpu);
+	vcpu_set_pauth_traps(vcpu);
+
 	kvm_arch_vcpu_load_debug_state_flags(vcpu);
 
 	if (!cpumask_test_cpu(cpu, vcpu->kvm->arch.supported_cpus))

arch/arm64/kvm/handle_exit.c

@@ -217,14 +217,12 @@ static int handle_sve(struct kvm_vcpu *vcpu)
  * Two possibilities to handle a trapping ptrauth instruction:
  *
  * - Guest usage of a ptrauth instruction (which the guest EL1 did not
- *   turn into a NOP). If we get here, it is that we didn't fixup
- *   ptrauth on exit, and all that we can do is give the guest an
- *   UNDEF (as the guest isn't supposed to use ptrauth without being
- *   told it could).
+ *   turn into a NOP). If we get here, it is because we didn't enable
+ *   ptrauth for the guest. This results in an UNDEF, as it isn't
+ *   supposed to use ptrauth without being told it could.
  *
  * - Running an L2 NV guest while L1 has left HCR_EL2.API==0, and for
- *   which we reinject the exception into L1. API==1 is handled as a
- *   fixup so the only way to get here is when API==0.
+ *   which we reinject the exception into L1.
  *
  * Anything else is an emulation bug (hence the WARN_ON + UNDEF).
  */

arch/arm64/kvm/hyp/include/hyp/switch.h

@@ -27,6 +27,7 @@
 #include <asm/kvm_hyp.h>
 #include <asm/kvm_mmu.h>
 #include <asm/kvm_nested.h>
+#include <asm/kvm_ptrauth.h>
 #include <asm/fpsimd.h>
 #include <asm/debug-monitors.h>
 #include <asm/processor.h>
@@ -447,82 +448,6 @@ static inline bool handle_tx2_tvm(struct kvm_vcpu *vcpu)
 	return true;
 }
 
-static inline bool esr_is_ptrauth_trap(u64 esr)
-{
-	switch (esr_sys64_to_sysreg(esr)) {
-	case SYS_APIAKEYLO_EL1:
-	case SYS_APIAKEYHI_EL1:
-	case SYS_APIBKEYLO_EL1:
-	case SYS_APIBKEYHI_EL1:
-	case SYS_APDAKEYLO_EL1:
-	case SYS_APDAKEYHI_EL1:
-	case SYS_APDBKEYLO_EL1:
-	case SYS_APDBKEYHI_EL1:
-	case SYS_APGAKEYLO_EL1:
-	case SYS_APGAKEYHI_EL1:
-		return true;
-	}
-
-	return false;
-}
-
-#define __ptrauth_save_key(ctxt, key)					\
-	do {								\
-	u64 __val;                                                      \
-	__val = read_sysreg_s(SYS_ ## key ## KEYLO_EL1);                \
-	ctxt_sys_reg(ctxt, key ## KEYLO_EL1) = __val;                   \
-	__val = read_sysreg_s(SYS_ ## key ## KEYHI_EL1);                \
-	ctxt_sys_reg(ctxt, key ## KEYHI_EL1) = __val;                   \
-} while(0)
-
-DECLARE_PER_CPU(struct kvm_cpu_context, kvm_hyp_ctxt);
-
-static bool kvm_hyp_handle_ptrauth(struct kvm_vcpu *vcpu, u64 *exit_code)
-{
-	struct kvm_cpu_context *ctxt;
-	u64 enable = 0;
-
-	if (!vcpu_has_ptrauth(vcpu))
-		return false;
-
-	/*
-	 * NV requires us to handle API and APK independently, just in
-	 * case the hypervisor is totally nuts. Please barf >here<.
-	 */
-	if (vcpu_has_nv(vcpu) && !is_hyp_ctxt(vcpu)) {
-		switch (ESR_ELx_EC(kvm_vcpu_get_esr(vcpu))) {
-		case ESR_ELx_EC_PAC:
-			if (!(__vcpu_sys_reg(vcpu, HCR_EL2) & HCR_API))
-				return false;
-
-			enable |= HCR_API;
-			break;
-
-		case ESR_ELx_EC_SYS64:
-			if (!(__vcpu_sys_reg(vcpu, HCR_EL2) & HCR_APK))
-				return false;
-
-			enable |= HCR_APK;
-			break;
-		}
-	} else {
-		enable = HCR_API | HCR_APK;
-	}
-
-	ctxt = this_cpu_ptr(&kvm_hyp_ctxt);
-	__ptrauth_save_key(ctxt, APIA);
-	__ptrauth_save_key(ctxt, APIB);
-	__ptrauth_save_key(ctxt, APDA);
-	__ptrauth_save_key(ctxt, APDB);
-	__ptrauth_save_key(ctxt, APGA);
-
-
-	vcpu->arch.hcr_el2 |= enable;
-	sysreg_clear_set(hcr_el2, 0, enable);
-
-	return true;
-}
-
 static bool kvm_hyp_handle_cntpct(struct kvm_vcpu *vcpu)
 {
 	struct arch_timer_context *ctxt;
@@ -610,9 +535,6 @@ static bool kvm_hyp_handle_sysreg(struct kvm_vcpu *vcpu, u64 *exit_code)
 	    __vgic_v3_perform_cpuif_access(vcpu) == 1)
 		return true;
 
-	if (esr_is_ptrauth_trap(kvm_vcpu_get_esr(vcpu)))
-		return kvm_hyp_handle_ptrauth(vcpu, exit_code);
-
 	if (kvm_hyp_handle_cntpct(vcpu))
 		return true;
 

arch/arm64/kvm/hyp/nvhe/switch.c

@@ -191,7 +191,6 @@ static const exit_handler_fn hyp_exit_handlers[] = {
 	[ESR_ELx_EC_IABT_LOW]		= kvm_hyp_handle_iabt_low,
 	[ESR_ELx_EC_DABT_LOW]		= kvm_hyp_handle_dabt_low,
 	[ESR_ELx_EC_WATCHPT_LOW]	= kvm_hyp_handle_watchpt_low,
-	[ESR_ELx_EC_PAC]		= kvm_hyp_handle_ptrauth,
 	[ESR_ELx_EC_MOPS]		= kvm_hyp_handle_mops,
 };
 
@@ -203,7 +202,6 @@ static const exit_handler_fn pvm_exit_handlers[] = {
 	[ESR_ELx_EC_IABT_LOW]		= kvm_hyp_handle_iabt_low,
 	[ESR_ELx_EC_DABT_LOW]		= kvm_hyp_handle_dabt_low,
 	[ESR_ELx_EC_WATCHPT_LOW]	= kvm_hyp_handle_watchpt_low,
-	[ESR_ELx_EC_PAC]		= kvm_hyp_handle_ptrauth,
 	[ESR_ELx_EC_MOPS]		= kvm_hyp_handle_mops,
 };
 

arch/arm64/kvm/hyp/vhe/switch.c

@@ -41,9 +41,8 @@ DEFINE_PER_CPU(unsigned long, kvm_hyp_vector);
  * - TGE: we want the guest to use EL1, which is incompatible with
  *   this bit being set
  *
- * - API/APK: for hysterical raisins, we enable PAuth lazily, which
- *   means that the guest's bits cannot be directly applied (we really
- *   want to see the traps). Revisit this at some point.
+ * - API/APK: they are already accounted for by vcpu_load(), and can
+ *   only take effect across a load/put cycle (such as ERET)
  */
 #define NV_HCR_GUEST_EXCLUDE	(HCR_TGE | HCR_API | HCR_APK)
 
@@ -268,7 +267,6 @@ static const exit_handler_fn hyp_exit_handlers[] = {
 	[ESR_ELx_EC_IABT_LOW]		= kvm_hyp_handle_iabt_low,
 	[ESR_ELx_EC_DABT_LOW]		= kvm_hyp_handle_dabt_low,
 	[ESR_ELx_EC_WATCHPT_LOW]	= kvm_hyp_handle_watchpt_low,
-	[ESR_ELx_EC_PAC]		= kvm_hyp_handle_ptrauth,
 	[ESR_ELx_EC_ERET]		= kvm_hyp_handle_eret,
 	[ESR_ELx_EC_MOPS]		= kvm_hyp_handle_mops,
 };