Re: [PATCH v5 0/3] add a test mode for SANITIZE=leak, run it in CI

On Wed, Sep 08 2021, Junio C Hamano wrote:

Ævar Arnfjörð Bjarmason  [off-list ref] writes:

quoted

We can compile git with SANITIZE=leak, and have had various efforts in
the past such as 31f9acf9ce2 (Merge branch 'ah/plugleaks', 2021-08-04)
to plug memory leaks, but have had no CI testing of it to ensure that
we don't get regressions. This series adds a GIT_TEST_* mode for
checking those regressions, and runs it in CI.

Since I submitted v2 the delta between origin/master..origin/seen
broke even t0001-init.sh when run under SANITIZE=leak, so this series
will cause test smoke on "seen".

That failure is due to a bug in es/config-based-hooks [1] and the
hn/reftable topic, i.e. these patches are legitimately catching
regressions in "seen" from day 1.

So is there a point in sending this out to the list, before sending
fixes to these broken topic and making sure they get corrected?

Because the CI does not "bisect" to tell us "ok, up to this point in
'seen', all the topics merged play well together", the overall
effect in the bigger picture is that 'seen' with this series would
cause CI to stay in failed state.

For now, I'll keep this near the tip of 'seen'.

The breakages with it are in combination with:

    ab/config-based-hooks-base
    es/config-based-hooks
    hn/reftable

You've got v4 of ab/config-based-hooks-base, the v5 is at [1], but we've
been waiting on emily to re-roll hers on top. As noted in that E-Mail
I've got a working re-roll of it as
avar-nasamuffin/config-based-hooks-restart-3 in my repo.

That'll leave hn/reftable, which given [2] I thought you were planning
to eject, and wiht the number of fixups for it / the planned re-doing of
it by Han-Wen[3] maybe it's better to do that now?

What do you think about that plan?

I.e. ejecting hn/reftable while waiting on a re-roll, and either
ejecting es/config-based-hooks while waiting, or I can submit the
avar-nasamuffin/config-based-hooks-restart-3 I've got pending Emily's
own re-roll (which may or may not be different from that).

That along with picking up the v5 of my ab/config-based-hooks-base
should make "seen" pass with SANITIZE=leak on these tests, unless
there's other just-introduced regressions. I tried re-building it a few
days ago, I haven't done that just now.

1. https://lore.kernel.org/git/cover-v5-00.36-00000000000-20210902T125110Z-avarab@gmail.com/ (local)
2. https://lore.kernel.org/git/xmqq4kaxe5dt.fsf@gitster.g/ (local)
3. https://lore.kernel.org/git/CAFQ2z_N8pUsp3cdBpybHBD-V9_1sARCZvSxr0UkMfcwCoQfCbw@mail.gmail.com/ (local)