Re: OOM in libpq and infinite loop with getCopyStart()

On Thu, Mar 3, 2016 at 10:18 PM, Aleksander Alekseev
<a(dot)alekseev(at)postgrespro(dot)ru> wrote:
> I didn't checked your patch in detail yet but here is a thought I would
> like to share.
>
> In my experience usually it takes number of rewrites before patch will
> be accepted. To make sure that after every rewrite your patch still
> solves an issue you described you should probably provide a few test
> programs too. If it's possible to make these programs part of regression
> tests suite it would be just great.

Yes that's usually the case, this patch is not at its first version.

> Without such test programs I personally find it difficult to verify
> that your patch fixes something. If such examples are provided your
> patch would be much more likely to accepted. "See, this program crashes
> everything. Now we apply a patch and everything works! Cool, heh?"

The easiest way to perform tests with this patch is to take a debugger
and enforce the malloc'd pointers to NULL in the code paths. That's
wild but efficient, and that's actually for we did for the other two
libpq patches treating those OOM failures. In the case of the standby
code path what I did was to put a sleep in the WAL receiver code and
then trigger the OOM manually, leading to the errors listed upthread.
Regarding automated tests, it is largely feasible to have tests on
Linux at least by using for example LD_PRELOAD or glibc-specific
malloc hooks, though those would be platform-restricted. Here is for
example one for COPY IN/OUT, that this patch also passed (there has
been already some discussion for this patch).
http://www.postgresql.org/message-id/55FC501A.5000409@iki.fi
Applying it for recovery is doable as well by manipulating the
recovery protocol with libpq though it is aimed really at covering a
narrow case.
--
Michael