[igt-dev] [PATCH i-g-t v2 5/5] i915/gem_ctx_isolation.c - If initialization fails, exit

[Dale B Stimson]

On 2020-02-13 10:29:55, Petri Latvala wrote:
> On Wed, Feb 12, 2020 at 05:28:40PM -0800, Dale B Stimson wrote:
> > At the start of igt_main, failure of the initial tests for successful
> > initialization transfer control to the end of an igt_fixture block.
> > From there, execution of the main per-engine loop is attempted.
> > Instead, the test should be caused to exit.
> > 
> > If initialization fails, exit.
> > 
> > Signed-off-by: Dale B Stimson <dale.b.stimson at intel.com>
> > ---
> >  tests/i915/gem_ctx_isolation.c | 15 +++++++++++++++
> >  1 file changed, 15 insertions(+)
> > 
> > diff --git a/tests/i915/gem_ctx_isolation.c b/tests/i915/gem_ctx_isolation.c
> > index 07ffbb84a..b11158dab 100644
> > --- a/tests/i915/gem_ctx_isolation.c
> > +++ b/tests/i915/gem_ctx_isolation.c
> > @@ -898,10 +898,13 @@ igt_main
> >  	int fd = -1;
> >  	struct eng_mmio_base_table_s *mbp = NULL;
> >  	uint32_t mmio_base = 0;
> > +	/* igt_fixture block is skipped if --list-subtests, so start with true. */
> > +	bool init_successful = true;
> >  
> >  	igt_fixture {
> >  		int gen;
> >  
> > +		init_successful = false;
> >  		fd = drm_open_driver(DRIVER_INTEL);
> >  		igt_require_gem(fd);
> >  		igt_require(gem_has_contexts(fd));
> > @@ -916,8 +919,20 @@ igt_main
> >  		igt_skip_on(gen > LAST_KNOWN_GEN);
> >  
> >  		mbp = gem_engine_mmio_base_info_get(fd);
> > +		init_successful = true;
> >  	}
> >  
> > +	if (!init_successful) {
> > +		igt_exit_early();
> > +	}
> > +
> 
> NAK. All this dancing around the infrastructure just makes changing
> the infrastructure later be awkward and produce weird errors.
> 
> If something in the fixture failed, with this code you never enter the
> subtest, making the test result 'notrun' instead of the correct 'skip'
> or 'fail'.
> 
> What is the problem this is trying to solve? Incorrect engine list
> used? If you have a subtest per static engine, all CI does is execute
> per static engine. Converting this test to use dynamic subtests for
> engines is the way forward.
> 
> 
> -- 
> Petri Latvala

NAK understood and accepted.

I will address this in a different way, targeting the underlying issues
instead of the symptom.  Please see my patch (just sent to ML):
  lib/i915/gem_engine_topology.c - intel_get_current_engine invalid result

To answer to your question about what this was trying to solve:

Briefly, and specifically addressing gem_ctx_isolation:

As-is observed behavior when open (or debugfs open) fails: per-engine loop
executes forever:
    Subtest vecs0-nonpriv: FAIL
    Subtest vecs0-nonpriv-switch: FAIL
    Subtest vecs0-clean: FAIL
    Subtest vecs0-dirty-create: FAIL
    Subtest vecs0-dirty-switch: FAIL
    Subtest vecs0-none: FAIL
    Subtest vecs0-S3: FAIL
    Subtest vecs0-S4: FAIL
    Subtest vecs0-reset: FAIL
    And repeat, ad infinitum for vecs0