r368649 fixed a regression in r362031 that was MFC-ed to stable/12 as
a part of r362572. That commit reordered IPI send and local TLB flush in
TLB invalidations.
Without this fix we've been seeing problems with stale memory content
where changes done under a mutex were not immediately observed by
another thread after taking the same mutex. Those inconsistenices were
correlated to copy-on-write faults for pages contaning the data.
The change needed some adaptations as I elected to skip two significant
intermediate changes: