what workflow creates so many negative dentries?

> I'm trying to imagine why this would occur with a non-pathological program.

It seems I failed to work in a link to a previous LSFMM+BPF session on negative dentries, which I meant to do: https://lwn.net/Articles/894098/

That has some information on workloads that end up with huge numbers of negative dentries. As do some of the links from it (and links from those links :) It is a longstanding problem as can be seen in our index entry: https://lwn.net/Kernel/Index/#Dentry_cache

jake