[Intel-gfx] [PATCH v2] tests/gem_error_capture: Initial testcase for error state capture/dump

[Chris Wilson]

On Mon, Apr 14, 2014 at 10:01:12AM +0000, Mateo Lozano, Oscar wrote:
> > > +				&requests);
> > > +		if (req_matched == 1) {
> > > +			igt_assert(strstr(ring_name, expected_ring_name));
> > > +			igt_assert(requests == 1);
> > 
> > Bad assumption. You could still have the request from
> > gem_quiescent_gpu() which may not have been retired before the error
> > triggers.
> 
> Hmmmm... but I make a first valid gem_execbuf()+gem_sync() before the one that actually hangs. Shouldn´t that make sure that all previous requests have been retired?

No, I would not make that assumption about kernel behaviour. The only
thing that it will guarantee is that the last request with that handle
is retired. (In other words, forthcoming bug fixes already break this
assumption.)
 
> > > +
> > > +			igt_assert(getline(&line, &line_size, file) > 0);
> > > +			items = sscanf(line, "  seqno 0x%08x, emitted %ld, tail
> > 0x%08x\n",
> > > +					&seqno, &jiffies, &tail);
> > > +			igt_assert(items == 3);
> > 
> > Bad. I may change the format. s/may/will/
> 
> I still need to make sure I got a valid tail, so I need to know the format and fail if don´t understand it. Or do you want me to use a different tail, maybe the ringbuffer one?

I didn't spot where you used tail. Care to elaborate? I may be able to
suggest an alternative, or we may either code this more defensively or
make the kernel emission easier to use and hopefully more informative
for debugging.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre