diff mbox series

[v4,66/66] KVM: arm64: nv: Fast-track EL1 TLBIs for VHE guests

Message ID 20210510165920.1913477-67-maz@kernel.org (mailing list archive)
State New, archived
Headers show
Series KVM: arm64: ARMv8.3/8.4 Nested Virtualization support | expand

Commit Message

Marc Zyngier May 10, 2021, 4:59 p.m. UTC
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 | 36 +++++++++++++++++++++++++++++++++
 arch/arm64/kvm/hyp/vhe/tlb.c    |  6 ++++--
 arch/arm64/kvm/sys_regs.c       | 25 ++++++++---------------
 3 files changed, 48 insertions(+), 19 deletions(-)
diff mbox series

Patch

diff --git a/arch/arm64/kvm/hyp/vhe/switch.c b/arch/arm64/kvm/hyp/vhe/switch.c
index b3e7d996b54d..67f8b7d89db6 100644
--- a/arch/arm64/kvm/hyp/vhe/switch.c
+++ b/arch/arm64/kvm/hyp/vhe/switch.c
@@ -166,6 +166,39 @@  void deactivate_traps_vhe_put(void)
 	__deactivate_traps_common();
 }
 
+static bool __hyp_handle_tlbi_el1(struct kvm_vcpu *vcpu)
+{
+	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 (kvm_vcpu_trap_get_class(vcpu) != ESR_ELx_EC_SYS64 ||
+	    !vcpu_mode_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 __hyp_handle_eret(struct kvm_vcpu *vcpu)
 {
 	struct kvm_cpu_context *ctxt = &vcpu->arch.ctxt;
@@ -259,6 +292,9 @@  static bool fixup_guest_exit_vhe(struct kvm_vcpu *vcpu, u64 *exit_code,
 	if (*exit_code == ARM_EXCEPTION_TRAP) {
 		if (__hyp_handle_eret(vcpu))
 			return true;
+
+		if (__hyp_handle_tlbi_el1(vcpu))
+			return true;
 	}
 
 	return fixup_guest_exit(vcpu, exit_code);
diff --git a/arch/arm64/kvm/hyp/vhe/tlb.c b/arch/arm64/kvm/hyp/vhe/tlb.c
index 001ac9f2f1b5..583255aa35ec 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 084230f6b993..6d0776737b76 100644
--- a/arch/arm64/kvm/sys_regs.c
+++ b/arch/arm64/kvm/sys_regs.c
@@ -2721,6 +2721,8 @@  static bool handle_tlbi_el1(struct kvm_vcpu *vcpu, struct sys_reg_params *p,
 			    const struct sys_reg_desc *r)
 {
 	u32 sys_encoding = sys_insn(p->Op0, p->Op1, p->CRn, p->CRm, p->Op2);
+	u64 virtual_vttbr = vcpu_read_sys_reg(vcpu, VTTBR_EL2);
+	struct kvm_s2_mmu *mmu;
 
 	/*
 	 * If we're here, this is because we've trapped on a EL1 TLBI
@@ -2739,24 +2741,13 @@  static bool handle_tlbi_el1(struct kvm_vcpu *vcpu, struct sys_reg_params *p,
 
 	mutex_lock(&vcpu->kvm->lock);
 
-	if ((__vcpu_sys_reg(vcpu, HCR_EL2) & (HCR_E2H | HCR_TGE)) != (HCR_E2H | HCR_TGE)) {
-		u64 virtual_vttbr = vcpu_read_sys_reg(vcpu, VTTBR_EL2);
-		struct kvm_s2_mmu *mmu;
-
-		mmu = lookup_s2_mmu(vcpu->kvm, virtual_vttbr, HCR_VM);
-		if (mmu)
-			__kvm_tlb_el1_instr(mmu, p->regval, sys_encoding);
+	mmu = lookup_s2_mmu(vcpu->kvm, virtual_vttbr, HCR_VM);
+	if (mmu)
+		__kvm_tlb_el1_instr(mmu, p->regval, sys_encoding);
 
-		mmu = lookup_s2_mmu(vcpu->kvm, virtual_vttbr, 0);
-		if (mmu)
-			__kvm_tlb_el1_instr(mmu, p->regval, sys_encoding);
-	} else {
-		/*
-		 * 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);
-	}
+	mmu = lookup_s2_mmu(vcpu->kvm, virtual_vttbr, 0);
+	if (mmu)
+		__kvm_tlb_el1_instr(mmu, p->regval, sys_encoding);
 
 	mutex_unlock(&vcpu->kvm->lock);