ANDROID: KVM: arm64: Disallow kvm_pgtable_stage2_unmap on host S2

The TLBI logic is unsafe for the host stage-2 table because not all
valid ptes are reference counted. Hence we could discard a child table
with only a __kvm_tlb_flush_vmid_ipa which is insufficient to
invalidate all mappings still reachable from the child table.

Bug: 278749606
Bug: 360102354
Change-Id: I0080b357743b335a4915327456f7ffc1dc2b7069
Signed-off-by: Keir Fraser <keirf@google.com>

arch/arm64/kvm/hyp/pgtable.c

diff --git a/arch/arm64/kvm/hyp/pgtable.c b/arch/arm64/kvm/hyp/pgtable.c
index b11bceb..460263d4 100644
--- a/arch/arm64/kvm/hyp/pgtable.c
+++ b/arch/arm64/kvm/hyp/pgtable.c
@@ -1157,6 +1157,18 @@ int kvm_pgtable_stage2_unmap(struct kvm_pgtable *pgt, u64 addr, u64 size)
 		.flags	= KVM_PGTABLE_WALK_LEAF | KVM_PGTABLE_WALK_TABLE_POST,
 	};
 
+	/*
+	 * stage2_unmap_walker's TLBI logic is unsafe for the pKVM host stage-2
+	 * table because a child table may have a refcount of 1 while still
+	 * containing valid mappings. The use of __kvm_tlb_flush_vmid_ipa in
+	 * stage2_unmap_clear_pte is then insufficient to invalidate all leaf
+	 * mappings reachable from the child table. All other stage-2 tables
+	 * hold a reference for every non-zero PTE, and are thus guaranteed to
+	 * be completely empty when refcount is 1.
+	 */
+	if (WARN_ON(pgt->flags & KVM_PGTABLE_S2_IDMAP))
+		return -EINVAL;
+
 	ret = kvm_pgtable_walk(pgt, addr, size, &walker);
 	if (stage2_unmap_defer_tlb_flush(pgt))
 		/* Perform the deferred TLB invalidations */