> My (completely random) guess is that there is some rare combination > of events that causes this code to fail. If so, is it feasible to > construct a test that makes this rare combination of events less rare, > so that similar future bugs are caught more quickly? Yes, I tested something similar before. What you need is create lots of PMIs with perf (running perf top should be enough) and a workload that creates lots of exits in a guest (e.g. running fio on a virtio device). This will stress test this particular path. -Andi