drm/radeon/kms: IB locking dumps out a lockdep ordering issue

We sometimes lock IB then the ring and sometimes the ring then
the IB. This is mostly due to the IB locking not being well defined
about what data in the structs it actually locks. Define what I
believe is the correct behaviour and gets rid of the lock dep ordering
warning.
Signed-off-by: Dave Airlie <airlied@redhat.com>

drivers/gpu/drm/radeon/r600_blit_kms.c

@@ -538,8 +538,8 @@ int r600_vb_ib_get(struct radeon_device *rdev)
 
 void r600_vb_ib_put(struct radeon_device *rdev)
 {
-	mutex_lock(&rdev->ib_pool.mutex);
 	radeon_fence_emit(rdev, rdev->r600_blit.vb_ib->fence);
+	mutex_lock(&rdev->ib_pool.mutex);
 	list_add_tail(&rdev->r600_blit.vb_ib->list, &rdev->ib_pool.scheduled_ibs);
 	mutex_unlock(&rdev->ib_pool.mutex);
 	radeon_ib_free(rdev, &rdev->r600_blit.vb_ib);

drivers/gpu/drm/radeon/radeon.h

@@ -402,6 +402,10 @@ struct radeon_ib {
 	uint32_t		length_dw;
 };
 
+/*
+ * locking -
+ * mutex protects scheduled_ibs, ready, alloc_bm
+ */
 struct radeon_ib_pool {
 	struct mutex		mutex;
 	struct radeon_object	*robj;

drivers/gpu/drm/radeon/radeon_ring.c

@@ -56,10 +56,12 @@ int radeon_ib_get(struct radeon_device *rdev, struct radeon_ib **ib)
 		set_bit(i, rdev->ib_pool.alloc_bm);
 		rdev->ib_pool.ibs[i].length_dw = 0;
 		*ib = &rdev->ib_pool.ibs[i];
+		mutex_unlock(&rdev->ib_pool.mutex);
 		goto out;
 	}
 	if (list_empty(&rdev->ib_pool.scheduled_ibs)) {
 		/* we go do nothings here */
+		mutex_unlock(&rdev->ib_pool.mutex);
 		DRM_ERROR("all IB allocated none scheduled.\n");
 		r = -EINVAL;
 		goto out;
@@ -69,10 +71,13 @@ int radeon_ib_get(struct radeon_device *rdev, struct radeon_ib **ib)
 			 struct radeon_ib, list);
 	if (nib->fence == NULL) {
 		/* we go do nothings here */
+		mutex_unlock(&rdev->ib_pool.mutex);
 		DRM_ERROR("IB %lu scheduled without a fence.\n", nib->idx);
 		r = -EINVAL;
 		goto out;
 	}
+	mutex_unlock(&rdev->ib_pool.mutex);
+
 	r = radeon_fence_wait(nib->fence, false);
 	if (r) {
 		DRM_ERROR("radeon: IB(%lu:0x%016lX:%u)\n", nib->idx,
@@ -81,12 +86,17 @@ int radeon_ib_get(struct radeon_device *rdev, struct radeon_ib **ib)
 		goto out;
 	}
 	radeon_fence_unref(&nib->fence);
+
 	nib->length_dw = 0;
+
+	/* scheduled list is accessed here */
+	mutex_lock(&rdev->ib_pool.mutex);
 	list_del(&nib->list);
 	INIT_LIST_HEAD(&nib->list);
+	mutex_unlock(&rdev->ib_pool.mutex);
+
 	*ib = nib;
 out:
-	mutex_unlock(&rdev->ib_pool.mutex);
 	if (r) {
 		radeon_fence_unref(&fence);
 	} else {
@@ -110,9 +120,10 @@ void radeon_ib_free(struct radeon_device *rdev, struct radeon_ib **ib)
 		return;
 	}
 	list_del(&tmp->list);
-	if (tmp->fence) {
+	INIT_LIST_HEAD(&tmp->list);
+	if (tmp->fence)
 		radeon_fence_unref(&tmp->fence);
-	}
+
 	tmp->length_dw = 0;
 	clear_bit(tmp->idx, rdev->ib_pool.alloc_bm);
 	mutex_unlock(&rdev->ib_pool.mutex);
@@ -122,25 +133,24 @@ int radeon_ib_schedule(struct radeon_device *rdev, struct radeon_ib *ib)
 {
 	int r = 0;
 
-	mutex_lock(&rdev->ib_pool.mutex);
 	if (!ib->length_dw || !rdev->cp.ready) {
 		/* TODO: Nothings in the ib we should report. */
-		mutex_unlock(&rdev->ib_pool.mutex);
 		DRM_ERROR("radeon: couldn't schedule IB(%lu).\n", ib->idx);
 		return -EINVAL;
 	}
+
 	/* 64 dwords should be enough for fence too */
 	r = radeon_ring_lock(rdev, 64);
 	if (r) {
 		DRM_ERROR("radeon: scheduling IB failled (%d).\n", r);
-		mutex_unlock(&rdev->ib_pool.mutex);
 		return r;
 	}
 	radeon_ring_ib_execute(rdev, ib);
 	radeon_fence_emit(rdev, ib->fence);
-	radeon_ring_unlock_commit(rdev);
+	mutex_lock(&rdev->ib_pool.mutex);
 	list_add_tail(&ib->list, &rdev->ib_pool.scheduled_ibs);
 	mutex_unlock(&rdev->ib_pool.mutex);
+	radeon_ring_unlock_commit(rdev);
 	return 0;
 }