diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/include/asm/kvm_host.h
index ac16f96c878d..108966a9db12 100644
--- a/arch/arm64/include/asm/kvm_host.h
+++ b/arch/arm64/include/asm/kvm_host.h
@@ -411,6 +411,9 @@ struct kvm_arch {
 	/* Masks for VNCR-backed and general EL2 sysregs */
 	struct kvm_sysreg_masks	*sysreg_masks;
 
+	/* Count the number of VNCR_EL2 TLBs */
+	atomic_t vncr_tlb_count;
+
 	/*
 	 * For an untrusted host VM, 'pkvm.handle' is used to lookup
 	 * the associated pKVM instance in the hypervisor.
diff --git a/arch/arm64/kvm/hyp/vhe/switch.c b/arch/arm64/kvm/hyp/vhe/switch.c
index c09b1d411c58..fa776b18a32c 100644
--- a/arch/arm64/kvm/hyp/vhe/switch.c
+++ b/arch/arm64/kvm/hyp/vhe/switch.c
@@ -424,10 +424,15 @@ static bool kvm_hyp_handle_tlbi_el2(struct kvm_vcpu *vcpu, u64 *exit_code)
 		return false;
 
 	/*
-	 * If we have to check for any VNCR mapping being invalidated,
-	 * go back to the slow path for further processing.
+	 * If we have to check for any VNCR TLB being invalidated, go back
+	 * to the slow path for further processing.
+	 *
+	 * The synchronisation betweem TLBI and walk is provided by the
+	 * speculative increment of the TLB counter on walk, and the
+	 * invalidation counter. Yes, this is fiddly.
 	 */
-	if (vcpu_el2_e2h_is_set(vcpu) && vcpu_el2_tge_is_set(vcpu))
+	if (vcpu_el2_e2h_is_set(vcpu) && vcpu_el2_tge_is_set(vcpu) &&
+	    atomic_read(&vcpu->kvm->arch.vncr_tlb_count))
 		return false;
 
 	__kvm_skip_instr(vcpu);
diff --git a/arch/arm64/kvm/nested.c b/arch/arm64/kvm/nested.c
index 550c9bd3dbe7..20af94197a8a 100644
--- a/arch/arm64/kvm/nested.c
+++ b/arch/arm64/kvm/nested.c
@@ -48,6 +48,7 @@ void kvm_init_nested(struct kvm *kvm)
 {
 	kvm->arch.nested_mmus = NULL;
 	kvm->arch.nested_mmus_size = 0;
+	atomic_set(&kvm->arch.vncr_tlb_count, 0);
 }
 
 static int init_nested_s2_mmu(struct kvm *kvm, struct kvm_s2_mmu *mmu)
@@ -1016,10 +1017,12 @@ u16 get_asid_by_regime(struct kvm_vcpu *vcpu, enum trans_regime regime)
 	return asid;
 }
 
-static void invalidate_vncr(struct vncr_tlb *vt)
+static void invalidate_vncr(struct kvm *kvm, struct vncr_tlb *vt)
 {
+	BUG_ON(!vt->valid);
 	vt->valid = false;
 	unmap_l1_vncr(vt);
+	atomic_dec(&kvm->arch.vncr_tlb_count);
 }
 
 static bool vncr_tlb_intersects(struct vncr_tlb *vt, u64 addr,
@@ -1066,7 +1069,7 @@ static void kvm_invalidate_vncr_ipa(struct kvm *kvm, u64 start, u64 end)
 	 */
 	kvm_for_each_vncr_tlb(i, vcpu, vt, kvm)
 		if (vncr_tlb_intersects(vt, vt->wr.pa, start, end - start))
-			invalidate_vncr(vt);
+			invalidate_vncr(kvm, vt);
 }
 
 struct s1e2_tlbi_scope {
@@ -1123,7 +1126,7 @@ static void invalidate_vncr_va(struct kvm *kvm,
 			break;
 		}
 
-		invalidate_vncr(vt);
+		invalidate_vncr(kvm, vt);
 	}
 }
 
@@ -1359,13 +1362,20 @@ void kvm_arch_flush_shadow_all(struct kvm *kvm)
  *   intersects with the TLBI request, invalidate it, and unmap the page
  *   from the fixmap. Because we need to look at all the vcpu-private TLBs,
  *   this requires some wide-ranging locking to ensure that nothing races
- *   against it. This may require some refcounting to avoid the search when
- *   no such TLB is present.
+ *   against it. This requires some refcounting to avoid the search when
+ *   no such TLB is present (see below).
  *
  * - On MMU notifiers, we must invalidate our TLB in a similar way, but
  *   looking at the IPA instead. The funny part is that there may not be a
  *   stage-2 mapping for this page if L1 hasn't accessed it using LD/ST
  *   instructions.
+ *
+ * - vncr_tlb_count tracks the number of valid VNCR TLBs VM-wide. This isn't
+ *   the number of *mapped* L1 VNCR pages, which is likely be a subset (and
+ *   by definition, a TLBI handled from L1 runs with the canonical VNCR
+ *   page, not the L1's). The innermost trap handling code checks this to
+ *   find out whether to return to the guest ASAP (no L1 TLBs) or to visit
+ *   this part of the world for some extra invalidation work.
  */
 
 int kvm_vcpu_allocate_vncr_tlb(struct kvm_vcpu *vcpu)
@@ -1420,7 +1430,8 @@ static int kvm_translate_vncr(struct kvm_vcpu *vcpu, bool *is_gmem)
 	 */
 	scoped_guard(write_lock, &vcpu->kvm->mmu_lock) {
 		this_cpu_reset_vncr_fixmap(vcpu);
-		vt->valid = false;
+		if (vt->valid)
+			invalidate_vncr(vcpu->kvm, vt);
 
 		vt->wi = (struct s1_walk_info) {
 			.regime	= TR_EL20,
@@ -1545,7 +1556,20 @@ int kvm_handle_vncr_abort(struct kvm_vcpu *vcpu)
 		return -EIO;
 	}
 
+	/*
+	 * Speculatively increment the TLB count to make sure concurrent
+	 * TLBIs will take the slow path, and will interact with the retry
+	 * mechanism. Drop it again on error.
+	 */
+	atomic_inc(&vcpu->kvm->arch.vncr_tlb_count);
+	smp_mb__after_atomic();
+
 	ret = kvm_translate_vncr(vcpu, &is_gmem);
+	if (ret) {
+		smp_mb__before_atomic();
+		atomic_dec(&vcpu->kvm->arch.vncr_tlb_count);
+	}
+
 	switch (ret) {
 	case -EAGAIN:
 		/* Let's try again... */