On 04/01/19 17:30, Sean Christopherson wrote:

quoted

+
+		if (kvm_available_flush_tlb_with_range()
+		    && (tmp_spte_ret & SET_SPTE_NEED_REMOTE_TLB_FLUSH)) {
+			struct kvm_mmu_page *leaf_sp = page_header(sp->spt[i]
+					& PT64_BASE_ADDR_MASK);
+			list_add(&leaf_sp->flush_link, &flush_list);
+		}
+
+		set_spte_ret |= tmp_spte_ret;
+
 	}
 
 	if (set_spte_ret & SET_SPTE_NEED_REMOTE_TLB_FLUSH)
-		kvm_flush_remote_tlbs(vcpu->kvm);
+		kvm_flush_remote_tlbs_with_list(vcpu->kvm, &flush_list);

This is a bit confusing and potentially fragile.  It's not obvious that
kvm_flush_remote_tlbs_with_list() is guaranteed to call
kvm_flush_remote_tlbs() when kvm_available_flush_tlb_with_range() is
false, and you're relying on the kvm_flush_remote_tlbs_with_list() call
chain to never optimize away the empty list case.  Rechecking
kvm_available_flush_tlb_with_range() isn't expensive.

Alternatively, do not check it during the loop: always build the
flush_list, and always call kvm_flush_remote_tlbs_with_list.  The
function can then check whether the list is empty, and the OR of
tmp_spte_ret on every iteration goes away.

Paolo