Some further observations...

There are other calls to uvm_pagedeactivate() which may be affected by
making it not call pmap_clear_reference().  It seems pretty clear that
just removing the pmap_clear_reference() there is going to cause them
to deactivate the page and then immediately reactivate it because the
bit is still set.  So, it may be necessary to adjust those callers to
avoid doing an extra pmap_clear_reference() in the page scanner.

It's worth noting that those code paths are *only* used to handle
various msync() and madvise() parameters, but that doesn't really make
them less of a problem.  B-)