diff --git a/shared-core/radeon_cp.c b/shared-core/radeon_cp.c
index 13f5ba4e..b0d5b27a 100644
--- a/shared-core/radeon_cp.c
+++ b/shared-core/radeon_cp.c
@@ -644,6 +644,24 @@ static void radeon_cp_init_ring_buffer( drm_device_t *dev,
 
 	RADEON_WRITE( RADEON_SCRATCH_UMSK, 0x7 );
 
+	/* Writeback doesn't seem to work everywhere, test it first */
+	DRM_WRITE32( &dev_priv->scratch[1], 0 );
+	RADEON_WRITE( RADEON_SCRATCH_REG1, 0xdeadbeef );
+
+	for ( tmp = 0 ; tmp < dev_priv->usec_timeout ; tmp++ ) {
+		if ( DRM_READ32( &dev_priv->scratch[1] ) == 0xdeadbeef )
+			break;
+		DRM_UDELAY( 1 );
+	}
+
+	if ( tmp < dev_priv->usec_timeout ) {
+		dev_priv->writeback_works = TRUE;
+		DRM_DEBUG( "writeback test succeeded, tmp=%d\n", tmp );
+	} else {
+		dev_priv->writeback_works = FALSE;
+		DRM_DEBUG( "writeback test failed\n" );
+	}
+
 	dev_priv->sarea_priv->last_frame = dev_priv->scratch[0] = 0;
 	RADEON_WRITE( RADEON_LAST_FRAME_REG,
 		      dev_priv->sarea_priv->last_frame );
@@ -1168,7 +1186,7 @@ drm_buf_t *radeon_freelist_get( drm_device_t *dev )
 	start = dev_priv->last_buf;
 
 	for ( t = 0 ; t < dev_priv->usec_timeout ; t++ ) {
-		u32 done_age = DRM_READ32(&dev_priv->scratch[1]);
+		u32 done_age = GET_SCRATCH( 1 );
 		DRM_DEBUG("done_age = %d\n",done_age);
 		for ( i = start ; i < dma->buf_count ; i++ ) {
 			buf = dma->buflist[i];
diff --git a/shared-core/radeon_drv.h b/shared-core/radeon_drv.h
index ae66f12e..15c0d4dd 100644
--- a/shared-core/radeon_drv.h
+++ b/shared-core/radeon_drv.h
@@ -76,6 +76,7 @@ typedef struct drm_radeon_private {
    	drm_radeon_freelist_t *tail;
 	int last_buf;
 	volatile u32 *scratch;
+	int writeback_works;
 
 	int usec_timeout;
 	int is_pci;
@@ -220,6 +221,10 @@ extern int radeon_cp_flip( DRM_IOCTL_ARGS );
 #define RADEON_SCRATCH_UMSK		0x0770
 #define RADEON_SCRATCH_ADDR		0x0774
 
+#define GET_SCRATCH( x )	(dev_priv->writeback_works			\
+				? DRM_READ32( &dev_priv->scratch[(x)] )		\
+				: RADEON_READ( RADEON_SCRATCH_REG0 + 4*(x) ) )
+
 #define RADEON_HOST_PATH_CNTL		0x0130
 #	define RADEON_HDP_SOFT_RESET		(1 << 26)
 #	define RADEON_HDP_WC_TIMEOUT_MASK	(7 << 28)
diff --git a/shared-core/radeon_state.c b/shared-core/radeon_state.c
index 9e04cbea..fb821055 100644
--- a/shared-core/radeon_state.c
+++ b/shared-core/radeon_state.c
@@ -1855,13 +1855,13 @@ int radeon_cp_getparam( DRM_IOCTL_ARGS )
 		value = dev_priv->agp_buffers_offset;
 		break;
 	case RADEON_PARAM_LAST_FRAME:
-		value = DRM_READ32(&dev_priv->scratch[0]);
+		value = GET_SCRATCH( 0 );
 		break;
 	case RADEON_PARAM_LAST_DISPATCH:
-		value = DRM_READ32(&dev_priv->scratch[1]);
+		value = GET_SCRATCH( 1 );
 		break;
 	case RADEON_PARAM_LAST_CLEAR:
-		value = DRM_READ32(&dev_priv->scratch[2]);
+		value = GET_SCRATCH( 2 );
 		break;
 	default:
 		return DRM_ERR(EINVAL);
diff --git a/shared/radeon_cp.c b/shared/radeon_cp.c
index 13f5ba4e..b0d5b27a 100644
--- a/shared/radeon_cp.c
+++ b/shared/radeon_cp.c
@@ -644,6 +644,24 @@ static void radeon_cp_init_ring_buffer( drm_device_t *dev,
 
 	RADEON_WRITE( RADEON_SCRATCH_UMSK, 0x7 );
 
+	/* Writeback doesn't seem to work everywhere, test it first */
+	DRM_WRITE32( &dev_priv->scratch[1], 0 );
+	RADEON_WRITE( RADEON_SCRATCH_REG1, 0xdeadbeef );
+
+	for ( tmp = 0 ; tmp < dev_priv->usec_timeout ; tmp++ ) {
+		if ( DRM_READ32( &dev_priv->scratch[1] ) == 0xdeadbeef )
+			break;
+		DRM_UDELAY( 1 );
+	}
+
+	if ( tmp < dev_priv->usec_timeout ) {
+		dev_priv->writeback_works = TRUE;
+		DRM_DEBUG( "writeback test succeeded, tmp=%d\n", tmp );
+	} else {
+		dev_priv->writeback_works = FALSE;
+		DRM_DEBUG( "writeback test failed\n" );
+	}
+
 	dev_priv->sarea_priv->last_frame = dev_priv->scratch[0] = 0;
 	RADEON_WRITE( RADEON_LAST_FRAME_REG,
 		      dev_priv->sarea_priv->last_frame );
@@ -1168,7 +1186,7 @@ drm_buf_t *radeon_freelist_get( drm_device_t *dev )
 	start = dev_priv->last_buf;
 
 	for ( t = 0 ; t < dev_priv->usec_timeout ; t++ ) {
-		u32 done_age = DRM_READ32(&dev_priv->scratch[1]);
+		u32 done_age = GET_SCRATCH( 1 );
 		DRM_DEBUG("done_age = %d\n",done_age);
 		for ( i = start ; i < dma->buf_count ; i++ ) {
 			buf = dma->buflist[i];
diff --git a/shared/radeon_drv.h b/shared/radeon_drv.h
index ae66f12e..15c0d4dd 100644
--- a/shared/radeon_drv.h
+++ b/shared/radeon_drv.h
@@ -76,6 +76,7 @@ typedef struct drm_radeon_private {
    	drm_radeon_freelist_t *tail;
 	int last_buf;
 	volatile u32 *scratch;
+	int writeback_works;
 
 	int usec_timeout;
 	int is_pci;
@@ -220,6 +221,10 @@ extern int radeon_cp_flip( DRM_IOCTL_ARGS );
 #define RADEON_SCRATCH_UMSK		0x0770
 #define RADEON_SCRATCH_ADDR		0x0774
 
+#define GET_SCRATCH( x )	(dev_priv->writeback_works			\
+				? DRM_READ32( &dev_priv->scratch[(x)] )		\
+				: RADEON_READ( RADEON_SCRATCH_REG0 + 4*(x) ) )
+
 #define RADEON_HOST_PATH_CNTL		0x0130
 #	define RADEON_HDP_SOFT_RESET		(1 << 26)
 #	define RADEON_HDP_WC_TIMEOUT_MASK	(7 << 28)
diff --git a/shared/radeon_state.c b/shared/radeon_state.c
index 9e04cbea..fb821055 100644
--- a/shared/radeon_state.c
+++ b/shared/radeon_state.c
@@ -1855,13 +1855,13 @@ int radeon_cp_getparam( DRM_IOCTL_ARGS )
 		value = dev_priv->agp_buffers_offset;
 		break;
 	case RADEON_PARAM_LAST_FRAME:
-		value = DRM_READ32(&dev_priv->scratch[0]);
+		value = GET_SCRATCH( 0 );
 		break;
 	case RADEON_PARAM_LAST_DISPATCH:
-		value = DRM_READ32(&dev_priv->scratch[1]);
+		value = GET_SCRATCH( 1 );
 		break;
 	case RADEON_PARAM_LAST_CLEAR:
-		value = DRM_READ32(&dev_priv->scratch[2]);
+		value = GET_SCRATCH( 2 );
 		break;
 	default:
 		return DRM_ERR(EINVAL);