Adrian Klaver <adrian.klaver@aklaver.com> writes:
> On 2/9/21 7:45 AM, hubert depesz lubaczewski wrote:
>> but I find it curious that I can set the guc using normal-ish SET, but
>> can't get it using SHOW or even select current_setting()
> Yeah, I think that part is a bug report.
After digging around in the code, I find the reason is that the entries in
pg_db_role_setting.setconfig are parsed with ParseLongOption, which quoth:
/*
* A little "long argument" simulation, although not quite GNU
* compliant. Takes a string of the form "some-option=some value" and
* returns name = "some_option" and value = "some value" in malloc'ed
* storage. Note that '-' is converted to '_' in the option name. If
* there is no '=' in the input string then value will be NULL.
*/
Sure enough,
regression=> show custom."bad-guc";
ERROR: unrecognized configuration parameter "custom.bad-guc"
regression=> show custom."bad_guc";
custom.bad_guc
----------------
1a
(1 row)
So that's where the setting went.
There's a second problem here with arbitrary GUC names, which is that
a name containing '=' isn't exactly gonna do what you want either.
There are probably other places that are not terribly careful about
funny characters in GUC names. In a quick test, I see that pg_dumpall
seems to dump the ALTER USER SET safely, but I wouldn't want to bet
that everything else copes.
I think we should probably sanitize custom GUC names at least to the
extent of forbidding '=' and '-'. Maybe we should go further and
insist they look like regular identifiers.
(Fortunately, ALTER USER SET with a custom GUC is superuser-only,
so there's no need to worry about security issues here. But we
should eliminate surprises.)
regards, tom lane
