On Fri, 2026-04-24 at 15:44 -0700, Jeff Davis wrote:
> Not known to be reachable after dbf217c1c7, but defend against
> similar
> issues in the future. For instance, an extension might encounter the
> problem by calling pg_newlocale_from_collation(DEFAULT_COLLATION_OID)
> from _PG_init(), and end up with a NULL pointer dereference.
>
> Backport through 17, though patch is different in 17 (also attached).
I plan to commit and backport this soon.
The patch looks good to me as it is.
Just one nit in PG 17 patch:
The provider == '\0' sentinel is correct (valid
providers are 'b'/'c'/'i'/'d', all nonzero) and the three guarded
sites (lc_collate_is_c, lc_ctype_is_c, pg_newlocale_from_collation)
cover the direct consumers symmetrically. The only thing maybe
consider is replacing the bare /* should not happen */ above the
'\0' test with something like /* provider unset; init_database_-
collation() not yet run */ -- the literal '\0' check reads as a bit
magical otherwise, but maybe you are doing that to have same
comment throughout.