return 0;
 }
 
+static int tdp_mmu_split_huge_page(struct kvm *kvm, struct tdp_iter *iter,
+                                  struct kvm_mmu_page *sp, bool shared);
+
 /*
  * Handle a TDP page fault (NPT/EPT violation/misconfiguration) by installing
  * page tables and SPTEs to translate the faulting guest physical address.
                if (iter.level == fault->goal_level)
                        break;
 
-               /*
-                * If there is an SPTE mapping a large page at a higher level
-                * than the target, that SPTE must be cleared and replaced
-                * with a non-leaf SPTE.
-                */
+               /* Step down into the lower level page table if it exists. */
                if (is_shadow_present_pte(iter.old_spte) &&
-                   is_large_pte(iter.old_spte)) {
-                       if (tdp_mmu_zap_spte_atomic(vcpu->kvm, &iter))
-                               break;
+                   !is_large_pte(iter.old_spte))
+                       continue;
 
-                       /*
-                        * The iter must explicitly re-read the spte here
-                        * because the new value informs the !present
-                        * path below.
-                        */
-                       iter.old_spte = kvm_tdp_mmu_read_spte(iter.sptep);
-               }
+               /*
+                * If SPTE has been frozen by another thread, just give up and
+                * retry, avoiding unnecessary page table allocation and free.
+                */
+               if (is_removed_spte(iter.old_spte))
+                       break;
 
-               if (!is_shadow_present_pte(iter.old_spte)) {
-                       /*
-                        * If SPTE has been frozen by another thread, just
-                        * give up and retry, avoiding unnecessary page table
-                        * allocation and free.
-                        */
-                       if (is_removed_spte(iter.old_spte))
-                               break;
+               /*
+                * The SPTE is either non-present or points to a huge page that
+                * needs to be split.
+                */
+               sp = tdp_mmu_alloc_sp(vcpu);
+               tdp_mmu_init_child_sp(sp, &iter);
 
-                       sp = tdp_mmu_alloc_sp(vcpu);
-                       tdp_mmu_init_child_sp(sp, &iter);
+               sp->nx_huge_page_disallowed = fault->huge_page_disallowed;
 
-                       sp->nx_huge_page_disallowed = fault->huge_page_disallowed;
+               if (is_shadow_present_pte(iter.old_spte))
+                       ret = tdp_mmu_split_huge_page(kvm, &iter, sp, true);
+               else
+                       ret = tdp_mmu_link_sp(kvm, &iter, sp, true);
 
-                       if (tdp_mmu_link_sp(kvm, &iter, sp, true)) {
-                               tdp_mmu_free_sp(sp);
-                               break;
-                       }
+               if (ret) {
+                       tdp_mmu_free_sp(sp);
+                       break;
+               }
 
-                       if (fault->huge_page_disallowed &&
-                           fault->req_level >= iter.level) {
-                               spin_lock(&kvm->arch.tdp_mmu_pages_lock);
-                               track_possible_nx_huge_page(kvm, sp);
-                               spin_unlock(&kvm->arch.tdp_mmu_pages_lock);
-                       }
+               if (fault->huge_page_disallowed &&
+                   fault->req_level >= iter.level) {
+                       spin_lock(&kvm->arch.tdp_mmu_pages_lock);
+                       track_possible_nx_huge_page(kvm, sp);
+                       spin_unlock(&kvm->arch.tdp_mmu_pages_lock);
                }
        }
 
        return sp;
 }
 
+/* Note, the caller is responsible for initializing @sp. */
 static int tdp_mmu_split_huge_page(struct kvm *kvm, struct tdp_iter *iter,
                                   struct kvm_mmu_page *sp, bool shared)
 {
        const int level = iter->level;
        int ret, i;
 
-       tdp_mmu_init_child_sp(sp, iter);
-
        /*
         * No need for atomics when writing to sp->spt since the page table has
         * not been linked in yet and thus is not reachable from any other CPU.
                                continue;
                }
 
+               tdp_mmu_init_child_sp(sp, &iter);
+
                if (tdp_mmu_split_huge_page(kvm, &iter, sp, shared))
                        goto retry;