i915/i915_pm_rps: Add fence-order and engine-order subtests

This series adds two tests that check if the iteration order of fences has an impact on applying waitboost. These regression tests make sure that the heuristic works as expected, no matter the order in which we wait on fences. For more background, see [1] and [2].

In addition to this, the patchset also includes a simple fix for pm_rps_exit_handler, added in preparation for multi-gt support and per-test sysfs files initialization.

Following the review comments, the initial patch was split up into two. So, this series is actually v5 of the "i915/i915_pm_rps: Check impact of fence ordering on waitboosting" patch. The previous versions can be seen at [3].

v5:
- Added missing close() calls on sysfs handle
- Extracted pm_rps_exit_handler change into a dedicated patch
v4:
- Added test descriptions with igt_describe()
v3:
- Removed unnecessary calls to gem_write
v2:
- Introduced batch_create() with arbitration points to prevent hangs
- Added engine-order subtest
- Added reporting of GPU frequency
- Changed runtimes and switched prints from us to ms

[1] - https://lore.kernel.org/intel-gfx/[email protected]/
[2] - https://gitlab.freedesktop.org/drm/intel/-/issues/6284
[3] - https://patchwork.freedesktop.org/patch/493954

Chris Wilson (2):
i915/i915_pm_rps: Check impact of fence ordering on waitboosting i915/i915_pm_rps: Write values in pm_rps_exit_handler only for min-max
configs

 tests/i915/i915_pm_rps.c | 265 ++++++++++++++++++++++++++++++++++++++-
 1 file changed, 259 insertions(+), 6 deletions(-)

--
2.25.1