[PATCH v10 57/59] KVM: arm64: nv: Fast-track EL1 TLBIs for VHE guests

[Marc Zyngier <maz@kernel.org>]

Due to the way ARMv8.4-NV suppresses traps when accessing EL2
system registers, we can't track when the guest changes its
HCR_EL2.TGE setting. This means we always trap EL1 TLBIs,
even if they don't affect any guest.

This obviously has a huge impact on performance, as we handle
TLBI traps as a normal exit, and a normal VHE host issues
thousands of TLBIs when booting (and quite a few when running
userspace).

A cheap way to reduce the overhead is to handle the limited
case of {E2H,TGE}=={1,1} as a guest fixup, as we already have
the right mmu configuration in place. Just execute the decoded
instruction right away and return to the guest.

Signed-off-by: Marc Zyngier <maz@kernel.org>
---
 arch/arm64/kvm/hyp/vhe/switch.c | 43 ++++++++++++++++++++++++++++++++-
 arch/arm64/kvm/hyp/vhe/tlb.c    |  6 +++--
 arch/arm64/kvm/sys_regs.c       | 12 ---------
 3 files changed, 46 insertions(+), 15 deletions(-)

diff --git a/arch/arm64/kvm/hyp/vhe/switch.c b/arch/arm64/kvm/hyp/vhe/switch.c
index a63c44bf62ce..2d305e0f145f 100644
--- a/arch/arm64/kvm/hyp/vhe/switch.c
+++ b/arch/arm64/kvm/hyp/vhe/switch.c
@@ -168,6 +168,47 @@ void deactivate_traps_vhe_put(struct kvm_vcpu *vcpu)
 	__deactivate_traps_common(vcpu);
 }
 
+static bool kvm_hyp_handle_tlbi_el1(struct kvm_vcpu *vcpu, u64 *exit_code)
+{
+	u32 instr;
+	u64 val;
+
+	/*
+	 * Ideally, we would never trap on EL1 TLB invalidations when the
+	 * guest's HCR_EL2.{E2H,TGE} == {1,1}. But "thanks" to ARMv8.4, we
+	 * don't trap writes to HCR_EL2, meaning that we can't track
+	 * changes to the virtual TGE bit. So we leave HCR_EL2.TTLB set on
+	 * the host. Oopsie...
+	 *
+	 * In order to speed-up EL1 TLBIs from the vEL2 guest when TGE is
+	 * set, try and handle these invalidation as quickly as possible,
+	 * without fully exiting (unless this needs forwarding).
+	 */
+	if (!vcpu_has_nv2(vcpu) ||
+	    !vcpu_is_el2(vcpu) ||
+	    (__vcpu_sys_reg(vcpu, HCR_EL2) & (HCR_E2H | HCR_TGE)) != (HCR_E2H | HCR_TGE))
+		return false;
+
+	instr = esr_sys64_to_sysreg(kvm_vcpu_get_esr(vcpu));
+	if (sys_reg_Op0(instr) != TLBI_Op0 ||
+	    sys_reg_Op1(instr) != TLBI_Op1_EL1)
+		return false;
+
+	val = vcpu_get_reg(vcpu, kvm_vcpu_sys_get_rt(vcpu));
+	__kvm_tlb_el1_instr(NULL, val, instr);
+	__kvm_skip_instr(vcpu);
+
+	return true;
+}
+
+static bool kvm_hyp_handle_sysreg_vhe(struct kvm_vcpu *vcpu, u64 *exit_code)
+{
+	if (kvm_hyp_handle_tlbi_el1(vcpu, exit_code))
+		return true;
+
+	return kvm_hyp_handle_sysreg(vcpu, exit_code);
+}
+
 static bool kvm_hyp_handle_eret(struct kvm_vcpu *vcpu, u64 *exit_code)
 {
 	struct kvm_cpu_context *ctxt = &vcpu->arch.ctxt;
@@ -216,7 +257,7 @@ static bool kvm_hyp_handle_eret(struct kvm_vcpu *vcpu, u64 *exit_code)
 static const exit_handler_fn hyp_exit_handlers[] = {
 	[0 ... ESR_ELx_EC_MAX]		= NULL,
 	[ESR_ELx_EC_CP15_32]		= kvm_hyp_handle_cp15_32,
-	[ESR_ELx_EC_SYS64]		= kvm_hyp_handle_sysreg,
+	[ESR_ELx_EC_SYS64]		= kvm_hyp_handle_sysreg_vhe,
 	[ESR_ELx_EC_SVE]		= kvm_hyp_handle_fpsimd,
 	[ESR_ELx_EC_FP_ASIMD]		= kvm_hyp_handle_fpsimd,
 	[ESR_ELx_EC_IABT_LOW]		= kvm_hyp_handle_iabt_low,
diff --git a/arch/arm64/kvm/hyp/vhe/tlb.c b/arch/arm64/kvm/hyp/vhe/tlb.c
index c4389db4cc22..beb162468c0b 100644
--- a/arch/arm64/kvm/hyp/vhe/tlb.c
+++ b/arch/arm64/kvm/hyp/vhe/tlb.c
@@ -201,7 +201,8 @@ void __kvm_tlb_el1_instr(struct kvm_s2_mmu *mmu, u64 val, u64 sys_encoding)
 	dsb(ishst);
 
 	/* Switch to requested VMID */
-	__tlb_switch_to_guest(mmu, &cxt);
+	if (mmu)
+		__tlb_switch_to_guest(mmu, &cxt);
 
 	/*
 	 * Execute the same instruction as the guest hypervisor did,
@@ -240,5 +241,6 @@ void __kvm_tlb_el1_instr(struct kvm_s2_mmu *mmu, u64 val, u64 sys_encoding)
 	dsb(ish);
 	isb();
 
-	__tlb_switch_to_host(&cxt);
+	if (mmu)
+		__tlb_switch_to_host(&cxt);
 }
diff --git a/arch/arm64/kvm/sys_regs.c b/arch/arm64/kvm/sys_regs.c
index 228dba82e9e5..27a29dcbfcd2 100644
--- a/arch/arm64/kvm/sys_regs.c
+++ b/arch/arm64/kvm/sys_regs.c
@@ -2987,18 +2987,6 @@ static bool handle_tlbi_el1(struct kvm_vcpu *vcpu, struct sys_reg_params *p,
 
 	WARN_ON(!vcpu_is_el2(vcpu));
 
-	if ((__vcpu_sys_reg(vcpu, HCR_EL2) & (HCR_E2H | HCR_TGE)) == (HCR_E2H | HCR_TGE)) {
-		mutex_lock(&vcpu->kvm->lock);
-		/*
-		 * ARMv8.4-NV allows the guest to change TGE behind
-		 * our back, so we always trap EL1 TLBIs from vEL2...
-		 */
-		__kvm_tlb_el1_instr(&vcpu->kvm->arch.mmu, p->regval, sys_encoding);
-		mutex_unlock(&vcpu->kvm->lock);
-
-		return true;
-	}
-
 	kvm_s2_mmu_iterate_by_vmid(vcpu->kvm, get_vmid(vttbr),
 				   &(union tlbi_info) {
 					   .va = {
-- 
2.34.1