The win10 buildbot appears to have this failure consistently (100%) on the 3.10 branch.  The first such failure appears to be yesterday (https://buildbot.python.org/all/#/builders/600/builds/79), following commit da8097aaf5a55c23f5b5ddbeffc2d90d06e00d93 - GC changes for mmap.

I'm not yet having much luck reproducing manually (which is perplexing given the consistent failures during the automatic tests).  I did see one test run where 8 other tests failed instead, but haven't been able to repeat.  The 3.x branch also seems to be doing fine on the buildbot while including the same patch that was cherry picked for 3.10.

So this could be a false lead, but it might be another place to review.  Or for anyone who can at least sporadically get a failure (I'm still trying), something to try reverting.