radeon: r345xx fixe hard lockup

This patch should fixe hard lockup and convert them in
softlockup (ie you can ssh the box but the gpu is busted
and we are waiting in loop for it to come back to reason).

1 files changed, 99 insertions, 21 deletions

diff --git a/shared-core/r300_cmdbuf.c b/shared-core/r300_cmdbuf.c
index e9f99c8f..addc075d 100644
--- a/shared-core/r300_cmdbuf.c
+++ b/shared-core/r300_cmdbuf.c
@@ -136,6 +136,18 @@ static int r300_emit_cliprects(drm_radeon_private_t *dev_priv,
 		ADVANCE_RING();
 	}
 
+	/* flus cache and wait idle clean after cliprect change */
+	BEGIN_RING(2);
+	OUT_RING(CP_PACKET0(R300_RB3D_DSTCACHE_CTLSTAT, 0));
+	OUT_RING(R300_RB3D_DC_FLUSH);
+	ADVANCE_RING();
+	BEGIN_RING(2);
+	OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
+	OUT_RING(RADEON_WAIT_3D_IDLECLEAN);
+	ADVANCE_RING();
+	/* set flush flag */
+	dev_priv->track_flush |= RADEON_FLUSH_EMITED;
+
 	return 0;
 }
 
@@ -166,13 +178,13 @@ void r300_init_reg_flags(struct drm_device *dev)
 	ADD_RANGE(0x21DC, 1);
 	ADD_RANGE(R300_VAP_UNKNOWN_221C, 1);
 	ADD_RANGE(R300_VAP_CLIP_X_0, 4);
-	ADD_RANGE(R300_VAP_PVS_WAITIDLE, 1);
+	ADD_RANGE(R300_VAP_PVS_STATE_FLUSH_REG, 1);
 	ADD_RANGE(R300_VAP_UNKNOWN_2288, 1);
 	ADD_RANGE(R300_VAP_OUTPUT_VTX_FMT_0, 2);
 	ADD_RANGE(R300_VAP_PVS_CNTL_1, 3);
 	ADD_RANGE(R300_GB_ENABLE, 1);
 	ADD_RANGE(R300_GB_MSPOS0, 5);
-	ADD_RANGE(R300_TX_CNTL, 1);
+	ADD_RANGE(R300_TX_INVALTAGS, 1);
 	ADD_RANGE(R300_TX_ENABLE, 1);
 	ADD_RANGE(0x4200, 4);
 	ADD_RANGE(0x4214, 1);
@@ -389,15 +401,28 @@ static __inline__ int r300_emit_vpu(drm_radeon_private_t *dev_priv,
 	if (sz * 16 > cmdbuf->bufsz)
 		return -EINVAL;
 
-	BEGIN_RING(5 + sz * 4);
-	/* Wait for VAP to come to senses.. */
-	/* there is no need to emit it multiple times, (only once before VAP is programmed,
-	   but this optimization is for later */
-	OUT_RING_REG(R300_VAP_PVS_WAITIDLE, 0);
+	/* VAP is very sensitive so we purge cache before we program it
+	 * and we also flush its state before & after */
+	BEGIN_RING(6);
+	OUT_RING(CP_PACKET0(R300_RB3D_DSTCACHE_CTLSTAT, 0));
+	OUT_RING(R300_RB3D_DC_FLUSH);
+	OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
+	OUT_RING(RADEON_WAIT_3D_IDLECLEAN);
+	OUT_RING(CP_PACKET0(R300_VAP_PVS_STATE_FLUSH_REG, 0));
+	OUT_RING(0);
+	ADVANCE_RING();
+	/* set flush flag */
+	dev_priv->track_flush |= RADEON_FLUSH_EMITED;
+
+	BEGIN_RING(3 + sz * 4);
 	OUT_RING_REG(R300_VAP_PVS_UPLOAD_ADDRESS, addr);
 	OUT_RING(CP_PACKET0_TABLE(R300_VAP_PVS_UPLOAD_DATA, sz * 4 - 1));
 	OUT_RING_TABLE((int *)cmdbuf->buf, sz * 4);
+	ADVANCE_RING();
 
+	BEGIN_RING(2);
+	OUT_RING(CP_PACKET0(R300_VAP_PVS_STATE_FLUSH_REG, 0));
+	OUT_RING(0);
 	ADVANCE_RING();
 
 	cmdbuf->buf += sz * 16;
@@ -425,6 +450,15 @@ static __inline__ int r300_emit_clear(drm_radeon_private_t *dev_priv,
 	OUT_RING_TABLE((int *)cmdbuf->buf, 8);
 	ADVANCE_RING();
 
+	BEGIN_RING(4);
+	OUT_RING(CP_PACKET0(R300_RB3D_DSTCACHE_CTLSTAT, 0));
+	OUT_RING(R300_RB3D_DC_FLUSH);
+	OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
+	OUT_RING(RADEON_WAIT_3D_IDLECLEAN);
+	ADVANCE_RING();
+	/* set flush flag */
+	dev_priv->track_flush |= RADEON_FLUSH_EMITED;
+
 	cmdbuf->buf += 8 * 4;
 	cmdbuf->bufsz -= 8 * 4;
 
@@ -614,13 +648,23 @@ static __inline__ int r300_emit_raw_packet3(drm_radeon_private_t *dev_priv,
 	case RADEON_CNTL_BITBLT_MULTI:
 		return r300_emit_bitblt_multi(dev_priv, cmdbuf);
 
-	case RADEON_CP_INDX_BUFFER:	/* DRAW_INDX_2 without INDX_BUFFER seems to lock up the gpu */
+	case RADEON_CP_INDX_BUFFER:
+		/* whenever we send vertex we clear flush & purge */
+		dev_priv->track_flush ^= (RADEON_FLUSH_EMITED |
+					  RADEON_PURGE_EMITED);
+		/* DRAW_INDX_2 without INDX_BUFFER seems to lock up the gpu */
 		return r300_emit_indx_buffer(dev_priv, cmdbuf);
-	case RADEON_CP_3D_DRAW_IMMD_2:	/* triggers drawing using in-packet vertex data */
-	case RADEON_CP_3D_DRAW_VBUF_2:	/* triggers drawing of vertex buffers setup elsewhere */
-	case RADEON_CP_3D_DRAW_INDX_2:	/* triggers drawing using indices to vertex buffer */
+	case RADEON_CP_3D_DRAW_IMMD_2:
+		/* triggers drawing using in-packet vertex data */
+	case RADEON_CP_3D_DRAW_VBUF_2:
+		/* triggers drawing of vertex buffers setup elsewhere */
+	case RADEON_CP_3D_DRAW_INDX_2:
+		/* triggers drawing using indices to vertex buffer */
 	case RADEON_WAIT_FOR_IDLE:
 	case RADEON_CP_NOP:
+		/* whenever we send vertex we clear flush & purge */
+		dev_priv->track_flush ^= (RADEON_FLUSH_EMITED |
+					  RADEON_PURGE_EMITED);
 		/* these packets are safe */
 		break;
 	default:
@@ -715,16 +759,47 @@ static __inline__ int r300_emit_packet3(drm_radeon_private_t *dev_priv,
 static __inline__ void r300_pacify(drm_radeon_private_t *dev_priv)
 {
 	RING_LOCALS;
+	
+	if ((dev_priv->track_flush & RADEON_PURGE_EMITED)) {
+		/* purge already emited without vertex in btw don't purge
+		 * again or lockup will likely happen */
+		return;
+	}
 
-	BEGIN_RING(6);
-	OUT_RING(CP_PACKET0(R300_RB3D_DSTCACHE_CTLSTAT, 0));
-	OUT_RING(R300_RB3D_DSTCACHE_UNKNOWN_0A);
+	/* flush & purge zbuffer */
+	BEGIN_RING(2);
 	OUT_RING(CP_PACKET0(R300_ZB_ZCACHE_CTLSTAT, 0));
-	OUT_RING(R300_ZB_ZCACHE_CTLSTAT_ZC_FLUSH_FLUSH_AND_FREE|
-		 R300_ZB_ZCACHE_CTLSTAT_ZC_FREE_FREE);
-	OUT_RING(CP_PACKET3(RADEON_CP_NOP, 0));
-	OUT_RING(0x0);
+	OUT_RING(R300_ZC_FLUSH | R300_ZC_FREE);
+	ADVANCE_RING();
+	/* flush & purge 3d */
+	BEGIN_RING(2);
+	OUT_RING(CP_PACKET0(R300_RB3D_DSTCACHE_CTLSTAT, 0));
+	OUT_RING(R300_RB3D_DC_FLUSH | R300_RB3D_DC_FREE);
+	ADVANCE_RING();
+	/* flush & purge texture */
+	BEGIN_RING(2);
+	OUT_RING(CP_PACKET0(R300_TX_INVALTAGS, 0));
+	OUT_RING(0);
+	ADVANCE_RING();
+	/* FIXME: is this one really needed ? */
+	BEGIN_RING(2);
+	OUT_RING(CP_PACKET0(R300_RB3D_AARESOLVE_CTL, 0));
+	OUT_RING(0);
 	ADVANCE_RING();
+	BEGIN_RING(2);
+	OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
+	OUT_RING(RADEON_WAIT_3D_IDLECLEAN);
+	ADVANCE_RING();
+	/* flush & purge 2d through E2 as RB2D will trigger lockup */
+	BEGIN_RING(4);
+	OUT_RING(CP_PACKET0(R300_DSTCACHE_CTLSTAT, 0));
+	OUT_RING(R300_RB2D_DC_FLUSH | R300_RB2D_DC_FREE);
+	OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
+	OUT_RING(RADEON_WAIT_2D_IDLECLEAN |
+		 RADEON_WAIT_HOST_IDLECLEAN);
+	ADVANCE_RING();
+	/* set flush & purge flags */
+	dev_priv->track_flush |= RADEON_FLUSH_EMITED | RADEON_PURGE_EMITED;
 }
 
 /**
@@ -902,12 +977,15 @@ int r300_do_cp_cmdbuf(struct drm_device *dev,
 	struct drm_buf *buf = NULL;
 	int emit_dispatch_age = 0;
 	int ret = 0;
+	RING_LOCALS;
 
 	DRM_DEBUG("\n");
 
-	/* See the comment above r300_emit_begin3d for why this call must be here,
-	 * and what the cleanup gotos are for. */
-	r300_pacify(dev_priv);
+	/* start by a wait, should be necessary */
+	BEGIN_RING(2);
+	OUT_RING(CP_PACKET0(RADEON_WAIT_UNTIL, 0));
+	OUT_RING(RADEON_WAIT_3D_IDLE | RADEON_WAIT_2D_IDLE);
+	ADVANCE_RING();
 
 	if (cmdbuf->nbox <= R300_SIMULTANEOUS_CLIPRECTS) {
 		ret = r300_emit_cliprects(dev_priv, cmdbuf, 0);