Re: [PATCH] support: Kill process group for test failure

["Carlos O'Donell" <carlos at redhat dot com>]

On Thu, Feb 20, 2020 at 9:53 AM Carlos O'Donell <codonell@redhat.com> wrote:
>
> On 2/20/20 9:34 AM, Adhemerval Zanella wrote:
> > Some testcases that create multiple subprocesses might abort or exit
> > prior waiting for their children.  In such case, support_test_main
> > does not try to kill the spawned test process group (as in the
> > test timeout case).
> >
> > On example that we are observing in internal tests is when
> > malloc/tst-mallocfork2 fails to fork in the signal handling (due
> > either maximum number of process or other non expected failure).
> >
> > This patch kill the process group in the case of failed execution,
> > similar on how it is done on timeout.
>
> LGTM.
>
> Reviewed-by: Carlos O'Donell <carlos@redhat.com>

I'm withdrawing my reviewed-by here, since there is a race.

Florian highlighted that the children are all going to be reparented
to init and that therefore we can't catch them anymore.

The only plausible solution here is to use the controlling terminal to
kill the orphan children.

> > Checked on x86_64-linux-gnu.
> > ---
> >  support/support_test_main.c | 7 +++++++
> >  1 file changed, 7 insertions(+)
> >
> > diff --git a/support/support_test_main.c b/support/support_test_main.c
> > index e3f0bf15f2..ac9f710fb7 100644
> > --- a/support/support_test_main.c
> > +++ b/support/support_test_main.c
> > @@ -459,6 +459,9 @@ support_test_main (int argc, char **argv, const struct test_config *config)
> >    /* Process terminated normaly without timeout etc.  */
> >    if (WIFEXITED (status))
> >      {
> > +      /* It is expected that a successful test execution handles all its
> > +      children.  */
>
> OK.
>
> > +
> >        if (config->expected_status == 0)
> >          {
> >            if (config->expected_signal == 0)
> > @@ -486,6 +489,10 @@ support_test_main (int argc, char **argv, const struct test_config *config)
> >    /* Process was killed by timer or other signal.  */
> >    else
> >      {
> > +      /* Kill the whole process group if test process aborts or exits prior
> > +      waiting for them.  */
> > +      kill (-test_pid, SIGKILL);
>
> OK. Send negation of process group id to kill the whole process group. The
> process group id is the same as the id of the process that created the group
> so test_pid is the right value.
>
> Notes:
> - Should we be using test_pgid to make this clear?
> - Should this cleanup be refactored a bit to avoid duplication from
>   signal_handler() and support_test_main() e.g. kill_process_group ()
>   which runs kill looks for errors prints diagnostic etc.
>
> > +
> >        if (config->expected_signal == 0)
> >          {
> >            printf ("Didn't expect signal from child: got `%s'\n",
> >
>
>
> --
> Cheers,
> Carlos.