Can't say if this fixes *all* cases, but at least we get through
the "point" regression test now, which hyrax's last run did not.
Report: https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=hyrax&dt=2020-09-19%2021%3A27%3A23
* replaced with the real root page at the end.
*/
page = palloc0(BLCKSZ);
+ RelationOpenSmgr(state->indexrel);
smgrextend(state->indexrel->rd_smgr, MAIN_FORKNUM, GIST_ROOT_BLKNO,
page, true);
state->pages_allocated++;