drm/i915: Define explicit wedged on init reset state

We're currently using scratch presence as a way of identifying that we
entered wedged state at driver initialization time.
Let's use a separate flag rather than rely on scratch.
Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: https://patchwork.freedesktop.org/patch/msgid/20190926133142.2838-1-chris@chris-wilson.co.uk

drivers/gpu/drm/i915/gt/intel_reset.c

@@ -811,7 +811,8 @@ static bool __intel_gt_unset_wedged(struct intel_gt *gt)
 	if (!test_bit(I915_WEDGED, &gt->reset.flags))
 		return true;
 
-	if (!gt->scratch) /* Never full initialised, recovery impossible */
+	/* Never fully initialised, recovery impossible */
+	if (test_bit(I915_WEDGED_ON_INIT, &gt->reset.flags))
 		return false;
 
 	GEM_TRACE("start\n");
@@ -1279,6 +1280,14 @@ int intel_gt_terminally_wedged(struct intel_gt *gt)
 	return intel_gt_is_wedged(gt) ? -EIO : 0;
 }
 
+void intel_gt_set_wedged_on_init(struct intel_gt *gt)
+{
+	BUILD_BUG_ON(I915_RESET_ENGINE + I915_NUM_ENGINES >
+		     I915_WEDGED_ON_INIT);
+	intel_gt_set_wedged(gt);
+	set_bit(I915_WEDGED_ON_INIT, &gt->reset.flags);
+}
+
 void intel_gt_init_reset(struct intel_gt *gt)
 {
 	init_waitqueue_head(&gt->reset.queue);

drivers/gpu/drm/i915/gt/intel_reset.h

@@ -45,6 +45,12 @@ void intel_gt_set_wedged(struct intel_gt *gt);
 bool intel_gt_unset_wedged(struct intel_gt *gt);
 int intel_gt_terminally_wedged(struct intel_gt *gt);
 
+/*
+ * There's no unset_wedged_on_init paired with this one.
+ * Once we're wedged on init, there's no going back.
+ */
+void intel_gt_set_wedged_on_init(struct intel_gt *gt);
+
 int __intel_gt_reset(struct intel_gt *gt, intel_engine_mask_t engine_mask);
 
 int intel_reset_guc(struct intel_gt *gt);
@@ -68,6 +74,9 @@ void __intel_fini_wedge(struct intel_wedge_me *w);
 
 static inline bool __intel_reset_failed(const struct intel_reset *reset)
 {
+	GEM_BUG_ON(test_bit(I915_WEDGED_ON_INIT, &reset->flags) ?
+		   !test_bit(I915_WEDGED, &reset->flags) : false);
+
 	return unlikely(test_bit(I915_WEDGED, &reset->flags));
 }
 

drivers/gpu/drm/i915/gt/intel_reset_types.h

@@ -29,11 +29,17 @@ struct intel_reset {
 	 * we set the #I915_WEDGED bit. Prior to command submission, e.g.
 	 * i915_request_alloc(), this bit is checked and the sequence
 	 * aborted (with -EIO reported to userspace) if set.
+	 *
+	 * #I915_WEDGED_ON_INIT - If we fail to initialize the GPU we can no
+	 * longer use the GPU - similar to #I915_WEDGED bit. The difference in
+	 * in the way we're handling "forced" unwedged (e.g. through debugfs),
+	 * which is not allowed in case we failed to initialize.
 	 */
 	unsigned long flags;
 #define I915_RESET_BACKOFF	0
 #define I915_RESET_MODESET	1
 #define I915_RESET_ENGINE	2
+#define I915_WEDGED_ON_INIT	(BITS_PER_LONG - 2)
 #define I915_WEDGED		(BITS_PER_LONG - 1)
 
 	struct mutex mutex; /* serialises wedging/unwedging */

drivers/gpu/drm/i915/i915_gem.c

@@ -1411,7 +1411,7 @@ int i915_gem_init(struct drm_i915_private *dev_priv)
 err_gt:
 	mutex_unlock(&dev_priv->drm.struct_mutex);
 
-	intel_gt_set_wedged(&dev_priv->gt);
+	intel_gt_set_wedged_on_init(&dev_priv->gt);
 	i915_gem_suspend(dev_priv);
 	i915_gem_suspend_late(dev_priv);