drm/radeon: some r420s have a CP race with the DMA engine.

This patch makes sure the CP doesn't DMA do VRAM while 2D
is active by inserting a CP resync token.

todo: port to kms.
Signed-off-by: Dave Airlie <airlied@redhat.com>

drivers/gpu/drm/radeon/radeon_cp.c

@@ -616,6 +616,18 @@ static void radeon_do_cp_start(drm_radeon_private_t * dev_priv)
 
 	dev_priv->cp_running = 1;
 
+	/* on r420, any DMA from CP to system memory while 2D is active
+	 * can cause a hang.  workaround is to queue a CP RESYNC token
+	 */
+	if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R420) {
+		BEGIN_RING(3);
+		OUT_RING(CP_PACKET0(R300_CP_RESYNC_ADDR, 1));
+		OUT_RING(5); /* scratch reg 5 */
+		OUT_RING(0xdeadbeef);
+		ADVANCE_RING();
+		COMMIT_RING();
+	}
+
 	BEGIN_RING(8);
 	/* isync can only be written through cp on r5xx write it here */
 	OUT_RING(CP_PACKET0(RADEON_ISYNC_CNTL, 0));
@@ -653,8 +665,19 @@ static void radeon_do_cp_reset(drm_radeon_private_t * dev_priv)
  */
 static void radeon_do_cp_stop(drm_radeon_private_t * dev_priv)
 {
+	RING_LOCALS;
 	DRM_DEBUG("\n");
 
+	/* finish the pending CP_RESYNC token */
+	if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R420) {
+		BEGIN_RING(2);
+		OUT_RING(CP_PACKET0(R300_RB3D_DSTCACHE_CTLSTAT, 0));
+		OUT_RING(R300_RB3D_DC_FINISH);
+		ADVANCE_RING();
+		COMMIT_RING();
+		radeon_do_wait_for_idle(dev_priv);
+	}
+
 	RADEON_WRITE(RADEON_CP_CSQ_CNTL, RADEON_CSQ_PRIDIS_INDDIS);
 
 	dev_priv->cp_running = 0;

drivers/gpu/drm/radeon/radeon_drv.h

@@ -1099,6 +1099,9 @@ extern u32 radeon_get_scratch(drm_radeon_private_t *dev_priv, int index);
 #	define RADEON_CSQ_PRIBM_INDBM		(4 << 28)
 #	define RADEON_CSQ_PRIPIO_INDPIO		(15 << 28)
 
+#define R300_CP_RESYNC_ADDR		0x0778
+#define R300_CP_RESYNC_DATA		0x077c
+
 #define RADEON_AIC_CNTL			0x01d0
 #	define RADEON_PCIGART_TRANSLATE_EN	(1 << 0)
 #	define RS400_MSI_REARM	                (1 << 3)