drm/i915/selftests: Fixup atomic reset checking
authorChris Wilson <chris@chris-wilson.co.uk>
Wed, 26 Jun 2019 13:44:33 +0000 (14:44 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Wed, 26 Jun 2019 15:03:13 +0000 (16:03 +0100)
We require that the intel_gpu_reset() was atomic, not the whole of
i915_reset() which is guarded by a mutex. However, we do require that
i915_reset_engine() is atomic for use from within the submission tasklet.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190626134433.6318-3-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/gt/selftest_reset.c

index 64c2c8ab64ecaac0e0dd0807c6d3481901a9f9d0..641cf3aee8d5071cd2cbb4bf0052d7a0061dffa1 100644 (file)
@@ -73,11 +73,13 @@ static int igt_atomic_reset(void *arg)
        for (p = igt_atomic_phases; p->name; p++) {
                GEM_TRACE("intel_gpu_reset under %s\n", p->name);
 
-               p->critical_section_begin();
                reset_prepare(i915);
+               p->critical_section_begin();
+
                err = intel_gpu_reset(i915, ALL_ENGINES);
-               reset_finish(i915);
+
                p->critical_section_end();
+               reset_finish(i915);
 
                if (err) {
                        pr_err("intel_gpu_reset failed under %s\n", p->name);
@@ -95,12 +97,71 @@ unlock:
        return err;
 }
 
+static int igt_atomic_engine_reset(void *arg)
+{
+       struct drm_i915_private *i915 = arg;
+       const typeof(*igt_atomic_phases) *p;
+       struct intel_engine_cs *engine;
+       enum intel_engine_id id;
+       int err = 0;
+
+       /* Check that the resets are usable from atomic context */
+
+       if (!intel_has_reset_engine(i915))
+               return 0;
+
+       if (USES_GUC_SUBMISSION(i915))
+               return 0;
+
+       intel_gt_pm_get(&i915->gt);
+       igt_global_reset_lock(i915);
+
+       /* Flush any requests before we get started and check basics */
+       if (!igt_force_reset(i915))
+               goto out_unlock;
+
+       for_each_engine(engine, i915, id) {
+               tasklet_disable_nosync(&engine->execlists.tasklet);
+               intel_engine_pm_get(engine);
+
+               for (p = igt_atomic_phases; p->name; p++) {
+                       GEM_TRACE("i915_reset_engine(%s) under %s\n",
+                                 engine->name, p->name);
+
+                       p->critical_section_begin();
+                       err = i915_reset_engine(engine, NULL);
+                       p->critical_section_end();
+
+                       if (err) {
+                               pr_err("i915_reset_engine(%s) failed under %s\n",
+                                      engine->name, p->name);
+                               break;
+                       }
+               }
+
+               intel_engine_pm_put(engine);
+               tasklet_enable(&engine->execlists.tasklet);
+               if (err)
+                       break;
+       }
+
+       /* As we poke around the guts, do a full reset before continuing. */
+       igt_force_reset(i915);
+
+out_unlock:
+       igt_global_reset_unlock(i915);
+       intel_gt_pm_put(&i915->gt);
+
+       return err;
+}
+
 int intel_reset_live_selftests(struct drm_i915_private *i915)
 {
        static const struct i915_subtest tests[] = {
                SUBTEST(igt_global_reset), /* attempt to recover GPU first */
                SUBTEST(igt_wedged_reset),
                SUBTEST(igt_atomic_reset),
+               SUBTEST(igt_atomic_engine_reset),
        };
        intel_wakeref_t wakeref;
        int err = 0;