Alvaro Herrera <alvherre@commandprompt.com> writes:
> In the process of looking it over again, I noticed that in an
> assert-enabled build, it's trivial to crash the server with this
> function: just pass a nonzero subobjid with an object class that doesn't
> take one. This could be fixed easily by turning the Asserts into
> elog(ERROR).
> Another problem with this function is that a lot of checks that
> currently raise errors with elog(ERROR) are now user-facing. On this
> issue one possible answer would be to do nothing; another would be to go
> over all those calls and turn them into full-fledged ereports.
Yeah, it would definitely be necessary to ensure that you couldn't cause
an Assert by passing bogus input. I wouldn't bother making the errors
into ereports though: that's adding a lot of translation burden to no
good purpose.
Please do not do this:
+/* this doesn't really need to appear in any header file */
+Datum pg_describe_object(PG_FUNCTION_ARGS);
Put the extern declaration in a header file, don't be cute.
This is useless, because getObjectDescription never returns null
(or if it does, we have a whole lot of unprotected callers to fix):
+ if (!description)
+ ereport(ERROR,
+ (errcode(ERRCODE_INVALID_OBJECT_DEFINITION),
+ errmsg("invalid object specification")));
regards, tom lane
