Hi,
On 2025-02-21 15:58:45 +0700, Zharkov Roman wrote:
> We running TAP tests which operates with a lot of instances. And some of
> these tests randomly fail due "address already in use". It turned out that
> the meson does not set environment variable MESON_BUILD_ROOT when running
> the test() function [0].
I don't think meson ever set MESON_BUILD_ROOT during tests, so afaict the
portlock logic just just never worked with meson?
> As a result each test uses its own "portlock" directory [1]. The small TAP
> test 001_portlock_env_test.pl shows the test environment variables. As we
> can see MESON_BUILD_ROOT is really undefined.
I'm rather baffled this hasn't caused more problems...
I have recently seen a CI failure that looked like it likely was caused by
this. It was in a back branch, and I thought that was the explanation, due to
the portlock logic not yet existing. But the portlock logic is old enough, so
it likely was this.
But I would have expected many more errors.
> Can we explicitly set the MESON_BUILD_ROOT environment variable when running
> a test? With included patch for the src/tools/testwrap file, each instance
> gets an unique TCP port.
I don't like that, feels like a "namespace violation" to set a MESON_*
variable, why not set the top_builddir or a dedicated variable?
And I don't think it should be done in testwrap, but to test_env in the
top-level meson.build.
Something like the attached?
Greetings,
Andres Freund
