[igt-dev] [PATCH i-g-t] i915/gem_exec_parallel: Add skip_on_simulation to whole binary

[Chris Wilson]

Quoting Katarzyna Dec (2019-02-08 13:32:38)
> On Fri, Feb 08, 2019 at 01:27:51PM +0000, Chris Wilson wrote:
> > Quoting Katarzyna Dec (2019-02-08 13:23:59)
> > > On Fri, Feb 08, 2019 at 01:14:26PM +0000, Chris Wilson wrote:
> > > > Quoting Katarzyna Dec (2019-02-08 13:10:37)
> > > > > On simulation environment these tests were failing. During debug
> > > > > we decided that whole binary should be skipped, because there is
> > > > > no additional coverage in simulation.
> > > > 
> > > > How did they fail? There are some paths here that are rarely touched
> > > > elsewhere.
> > > > -Chris
> > > I like you challenges Chris :)
> > > 
> > > gem_exec_parallel@*context* were failing with error:
> > > 
> > > (gem_exec_parallel:1245) ioctl_wrappers-CRITICAL: Test assertion failure
> > > function gem_execbuf, file ../lib/ioctl_wrappers.c:605:
> > > (gem_exec_parallel:1245) ioctl_wrappers-CRITICAL: Failed assertion:
> > > __gem_execbuf(fd, execbuf) == 0
> > > (gem_exec_parallel:1245) ioctl_wrappers-CRITICAL: error: -28 != 0
> > 
> > Should not happen.
> But happends and is reproducible on some CI runs.

Well fix it. It is not a valid error in this case. ENOSPC is only valid
in the case the user submitted more than could be fitted into the GGTT
(or had an overlap in their softpin). In this instance, it means that
one _context_ impacted another client, and that is simply not allowed.

> > > and @*fds* with:
> > > 
> > > (gem_exec_parallel:1315) igt_debugfs-CRITICAL: Last errno: 24, Too many open
> > > files
> > 
> > That is boring, but more indicative of system setup failure. We should
> > be configuring the system to allow ourselves, as root, to use as many fd
> > as can fit in memory.
> > 
> > I'd recommend a pass through the basic stress test at least. It won't
> > take a lot to exercise some unusual conditions.
> > -Chris
> 
> What if we skip all *context* and all *fds* and leave rest as it is?
> Will it be resonable in this shape?

I would rewrite it such that the threads ran for a certain period of
time rather than iterations; set basic to something small (say 2s) and
left everything else to run for 150s (tends to be the common value I've
used for random stress tests). Then skip everything but the basic-all
test.
-Chris