On Sun, Jan 8, 2023 at 6:43 PM Peter Geoghegan <pg@bowt.ie> wrote:
> On further reflection even v2 won't repair the page-level
> PD_ALL_VISIBLE flag in passing in this scenario. ISTM that on HEAD we
> might actually leave the all-frozen bit set in the VM, while both the
> all-visible bit and the page-level PD_ALL_VISIBLE bit remain unset.
> Again, all due to the approach we take with
> all_visible_according_to_vm, which can go stale independently of both
> the VM bit being unset and the PD_ALL_VISIBLE bit being unset (in my
> example problem scenario).
Attached is v3, which explicitly checks the need to set the
PD_ALL_VISIBLE flag at the relevant visibilitymap_set() call site. It
also has improved comments.
--
Peter Geoghegan