drm/radeon: fix CP semaphores on CIK

The CP semaphore queue on CIK has a bug that triggers if uncompleted
waits use the same address while a signal is still pending. Work around
this by using different addresses for each sync.
Signed-off-by: Christian König <christian.koenig@amd.com>
Cc: stable@vger.kernel.org
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drivers/gpu/drm/radeon/radeon.h

@@ -135,6 +135,9 @@ extern int radeon_hard_reset;
 /* R600+ */
 #define R600_RING_TYPE_UVD_INDEX	5
 
+/* number of hw syncs before falling back on blocking */
+#define RADEON_NUM_SYNCS			4
+
 /* hardcode those limit for now */
 #define RADEON_VA_IB_OFFSET			(1 << 20)
 #define RADEON_VA_RESERVED_SIZE			(8 << 20)
@@ -554,7 +557,6 @@ int radeon_mode_dumb_mmap(struct drm_file *filp,
 /*
  * Semaphores.
  */
-/* everything here is constant */
 struct radeon_semaphore {
 	struct radeon_sa_bo		*sa_bo;
 	signed				waiters;

drivers/gpu/drm/radeon/radeon_ring.c

@@ -139,7 +139,7 @@ int radeon_ib_schedule(struct radeon_device *rdev, struct radeon_ib *ib,
 	}
 
 	/* 64 dwords should be enough for fence too */
-	r = radeon_ring_lock(rdev, ring, 64 + RADEON_NUM_RINGS * 8);
+	r = radeon_ring_lock(rdev, ring, 64 + RADEON_NUM_SYNCS * 8);
 	if (r) {
 		dev_err(rdev->dev, "scheduling IB failed (%d).\n", r);
 		return r;

drivers/gpu/drm/radeon/radeon_semaphore.c

@@ -34,14 +34,15 @@
 int radeon_semaphore_create(struct radeon_device *rdev,
 			    struct radeon_semaphore **semaphore)
 {
+	uint32_t *cpu_addr;
 	int i, r;
 
 	*semaphore = kmalloc(sizeof(struct radeon_semaphore), GFP_KERNEL);
 	if (*semaphore == NULL) {
 		return -ENOMEM;
 	}
-	r = radeon_sa_bo_new(rdev, &rdev->ring_tmp_bo,
-			     &(*semaphore)->sa_bo, 8, 8, true);
+	r = radeon_sa_bo_new(rdev, &rdev->ring_tmp_bo, &(*semaphore)->sa_bo,
+			     8 * RADEON_NUM_SYNCS, 8, true);
 	if (r) {
 		kfree(*semaphore);
 		*semaphore = NULL;
@@ -49,7 +50,10 @@ int radeon_semaphore_create(struct radeon_device *rdev,
 	}
 	(*semaphore)->waiters = 0;
 	(*semaphore)->gpu_addr = radeon_sa_bo_gpu_addr((*semaphore)->sa_bo);
-	*((uint64_t*)radeon_sa_bo_cpu_addr((*semaphore)->sa_bo)) = 0;
+
+	cpu_addr = radeon_sa_bo_cpu_addr((*semaphore)->sa_bo);
+	for (i = 0; i < RADEON_NUM_SYNCS; ++i)
+		cpu_addr[i] = 0;
 
 	for (i = 0; i < RADEON_NUM_RINGS; ++i)
 		(*semaphore)->sync_to[i] = NULL;
@@ -125,6 +129,7 @@ int radeon_semaphore_sync_rings(struct radeon_device *rdev,
 				struct radeon_semaphore *semaphore,
 				int ring)
 {
+	unsigned count = 0;
 	int i, r;
 
         for (i = 0; i < RADEON_NUM_RINGS; ++i) {
@@ -140,6 +145,12 @@ int radeon_semaphore_sync_rings(struct radeon_device *rdev,
 			return -EINVAL;
 		}
 
+		if (++count > RADEON_NUM_SYNCS) {
+			/* not enough room, wait manually */
+			radeon_fence_wait_locked(fence);
+			continue;
+		}
+
 		/* allocate enough space for sync command */
 		r = radeon_ring_alloc(rdev, &rdev->ring[i], 16);
 		if (r) {
@@ -164,6 +175,8 @@ int radeon_semaphore_sync_rings(struct radeon_device *rdev,
 
 		radeon_ring_commit(rdev, &rdev->ring[i]);
 		radeon_fence_note_sync(fence, ring);
+
+		semaphore->gpu_addr += 8;
 	}
 
 	return 0;