Re: pgsql: Move tablespace path re-creation from the makefiles to pg_regres

On Wed, Mar 24, 2021 at 07:56:59PM -0700, Noah Misch wrote:
> That would entail forbidding various shell metacharacters in @testtablespace@.
> One could avoid imposing such a restriction by adding a mkdir() function to
> regress.c and writing it like:
>
> CREATE FUNCTION mkdir(text)
>    RETURNS void AS '@libdir@/regress@DLSUFFIX@', 'regress_mkdir'
>    LANGUAGE C STRICT\;
>   REVOKE ALL ON FUNCTION mkdir FROM public;
> SELECT mkdir('@testtablespace@');
> CREATE TABLESPACE regress_tblspace LOCATION '@testtablespace@';

Sounds simple.

>> I doubt that all the Windows environments would be able to get their
>> hands on that.
>
>> And I am not sure either that this would work when it
>> comes to the CI case, again on Windows.
>
> How might a CI provider break that?

I am wondering about potential issues when it comes to create the
base tablespace path from the Postgres backend, while HEAD does it
while relying on a restricted token obtained when starting
pg_regress.

I have compiled a simple patch that uses a SQL function for the base
tablespace directory creation, that I have tested on Linux and MSVC.
To get some coverage with the CF bot, I am adding a CF entry with this
thread.

I am still not sure if people would prefer this approach over what's
on HEAD.  So if there are any opinions, please feel free.
--
Michael