Re: [igt-dev] [Intel-gfx] [PATCH i-g-t v2 05/17] lib/kunit: Fix illegal igt_fail() calls inside subtest body

[Janusz Krzysztofik <janusz.krzysztofik@linux.intel.com>]

On Monday, 11 September 2023 13:57:29 CEST Mauro Carvalho Chehab wrote:
> On Mon, 11 Sep 2023 11:28:32 +0200
> Janusz Krzysztofik <janusz.krzysztofik@linux.intel.com> wrote:
> 
> > Hi Mauro,
> > 
> > Thanks for review.
> > 
> > On Monday, 11 September 2023 10:52:51 CEST Mauro Carvalho Chehab wrote:
> > > On Fri,  8 Sep 2023 14:32:39 +0200
> > > Janusz Krzysztofik <janusz.krzysztofik@linux.intel.com> wrote:
> > >   
> > > > In a body of a subtest with dynamic sub-subtests, it is illegal to 
call
> > > > igt_fail() and its variants from outside of a dynamic sub-subtest 
body.
> > > > On the other hand, it is perfectly legal to call either igt_skip() and
> > > > friends or __igt_abort() or its variant from there.
> > > > 
> > > > In the current implementation of igt_kunit(), there are several places
> > > > where igt_fail() is called despite being illegal.  Moreover, it is 
called
> > > > with IGT_EXIT_ABORT as an argument with no good reason for using such
> > > > aggressive method that forces CI to trigger system reboot (in most 
cases
> > > > igt_runner can decide if abort is required).
> > > > 
> > > > Follow igt_kselftests() pattern more closely, where similar setup and
> > > > cleanup operations are performed but their potential errors are 
processed
> > > > in a more friendly way.  Move common cleanup and their corresponding 
setup
> > > > steps out of the subtest body.  Place the latter as requirements in a
> > > > preceding igt_fixture section.  Replace remaining illegal igt_fail() 
calls
> > > > with more friendly skips.  Let igt_runner decide if abort is needed.
> > > > 
> > > > Signed-off-by: Janusz Krzysztofik <janusz.krzysztofik@linux.intel.com>
> > > > ---
> > > >  lib/igt_kmod.c | 75 ++++++++++++++
+-----------------------------------
> > > >  1 file changed, 22 insertions(+), 53 deletions(-)
> > > > 
> > > > diff --git a/lib/igt_kmod.c b/lib/igt_kmod.c
> > > > index 1d1cd51170..78b8eb8f53 100644
> > > > --- a/lib/igt_kmod.c
> > > > +++ b/lib/igt_kmod.c  
> > ...
> > > > @@ -825,24 +793,21 @@ static void __igt_kunit(const char *module_name, 
const char *opts)
> > > >  		}
> > > >  	}
> > > >  
> > > > -unload:
> > > > -	igt_ktest_end(&tst);
> > > > -
> > > > -	igt_ktest_fini(&tst);
> > > > -
> > > > -	igt_skip_on_f(skip, "Skipping test, as probing KUnit module 
failed\n");
> > > > -
> > > > -	if (fail)
> > > > -		igt_fail(IGT_EXIT_ABORT);
> > > > -
> > > >  	ret = ktap_parser_stop();
> > > >  
> > > > -	if (ret != 0)
> > > > -		igt_fail(IGT_EXIT_ABORT);
> > > > +	igt_skip_on_f(ret, "KTAP parser failed\n");
> > > >  }
> > > >  
> > > >  void igt_kunit(const char *module_name, const char *name, const char 
*opts)
> > > >  {
> > > > +	struct igt_ktest tst;
> > > > +
> > > > +	if (igt_ktest_init(&tst, module_name) != 0)
> > > > +		return;  
> > > 
> > > Shouldn't it be calling igt_skip() here too?  
> > 
> > Maybe yes.  I've chosen to follow the algorithm used in igt_kselftest.  
There 
> > was an igt_skip() variant there initially but in 2017 that was converted 
to 
> > the current return only by Peter with IGT commit 9f92893b11e8 ("lib/
igt_kmod: 
> > Don't call igt_assert or igt_require without a fixture").  However, 
> > justification for dropping igt_require() instead of calling it from an 
> > igt_fixture section may not apply to kunit modules:
> > 
> > "If kmod_module_new_from_name fails, ... return normally from 
igt_kselftest, 
> > matching behaviour when the module loading is successful but it doesn't 
> > contain selftests."
> > 
> > While i915 could be built with no selftests included, a kunit module 
without 
> > any tests doesn't make sense, then silent return may be not what we need.
> 
> Yeah, selftests are handled on a different way with regards to module
> probe, so I guess we need the igt_skip there if modprobe fails.

After having a closer look at it, I think that igt_ktest_init() has nothing to 
do with actual modprobe, and it can fail only on either no memory or if 
module_name == NULL.  Anyway, I'll make the subtest skip if it fails.

> Well, you can probably simulate it by renaming a Kunit module
> and see how IGT will handle that with the current code and with
> igt_skip().

Yes, I've tired, and my results have confirmed my conclusions from code 
review.  But more important, I've found an issue in patch 15/17, "Parse KTAP 
report from the main process thread", that can cause first read() to wait 
infinitely, unless interrupted, if modprobe fails.  I've already developed a 
working fix that interrupts that read() on modprobe failure, and I'll include 
it in next version of the series.

Thanks,
Janusz

> 
> (Btw, I intend to review the other patches on this series, but need
> some time to do tests, as some changes here are not trivial)
> 
> Regards,
> Mauro
>