Kernel portions of the i810 security model - verification and emission of

    state, verification of vertex buffers, implementation of swap and
    clear.

13 files changed, 1288 insertions, 272 deletions

diff --git a/linux-core/i810_dma.c b/linux-core/i810_dma.c
index c940f5cd..5c85a4dc 100644
--- a/linux-core/i810_dma.c
+++ b/linux-core/i810_dma.c
@@ -37,8 +37,9 @@
 
 #include <linux/interrupt.h>	/* For task queue support */
 
-#define I810_BUF_FREE		1
-#define I810_BUF_USED		0
+#define I810_BUF_FREE		2
+#define I810_BUF_CLIENT		1
+#define I810_BUF_HARDWARE      	0
 
 #define I810_BUF_UNMAPPED 0
 #define I810_BUF_MAPPED   1
@@ -94,7 +95,7 @@ static inline void i810_print_status_page(drm_device_t *dev)
       	DRM_DEBUG(  "hw_status: Reserved : %x\n", temp[3]);
    	DRM_DEBUG(  "hw_status: Driver Counter : %d\n", temp[5]);
    	for(i = 6; i < dma->buf_count + 6; i++) {
-	   	DRM_DEBUG(  "buffer status idx : %d used: %d\n", i - 6, temp[i]);
+	   	DRM_DEBUG( "buffer status idx : %d used: %d\n", i - 6, temp[i]);
 	}
 }
 
@@ -111,7 +112,7 @@ static drm_buf_t *i810_freelist_get(drm_device_t *dev)
 	   	drm_i810_buf_priv_t *buf_priv = buf->dev_private;
 		/* In use is already a pointer */
 	   	used = cmpxchg(buf_priv->in_use, I810_BUF_FREE, 
-			       I810_BUF_USED);
+			       I810_BUF_CLIENT);
 	   	if(used == I810_BUF_FREE) {
 			return buf;
 		}
@@ -129,8 +130,8 @@ static int i810_freelist_put(drm_device_t *dev, drm_buf_t *buf)
    	int used;
    
    	/* In use is already a pointer */
-   	used = cmpxchg(buf_priv->in_use, I810_BUF_USED, I810_BUF_FREE);
-   	if(used != I810_BUF_USED) {
+   	used = cmpxchg(buf_priv->in_use, I810_BUF_CLIENT, I810_BUF_FREE);
+   	if(used != I810_BUF_CLIENT) {
 	   	DRM_ERROR("Freeing buffer thats not in use : %d\n", buf->idx);
 	   	return -EINVAL;
 	}
@@ -218,7 +219,8 @@ static int i810_dma_get_buffer(drm_device_t *dev, drm_i810_dma_t *d,
 	retcode = i810_map_buffer(buf, filp);
 	if(retcode) {
 		i810_freelist_put(dev, buf);
-	   	DRM_DEBUG("mapbuf failed in %s retcode %d\n", __FUNCTION__, retcode);
+	   	DRM_DEBUG("mapbuf failed in %s retcode %d\n", 
+			  __FUNCTION__, retcode);
 	   	goto out_get_buf;
 	}
 	buf->pid     = current->pid;
@@ -260,7 +262,10 @@ static void i810_free_page(drm_device_t *dev, unsigned long page)
 
 static int i810_dma_cleanup(drm_device_t *dev)
 {
+	drm_device_dma_t *dma = dev->dma;
+
 	if(dev->dev_private) {
+		int i;
 	   	drm_i810_private_t *dev_priv = 
 	     		(drm_i810_private_t *) dev->dev_private;
 	   
@@ -276,6 +281,12 @@ static int i810_dma_cleanup(drm_device_t *dev)
 	   	drm_free(dev->dev_private, sizeof(drm_i810_private_t), 
 			 DRM_MEM_DRIVER);
 	   	dev->dev_private = NULL;
+
+		for (i = 0; i < dma->buf_count; i++) {
+			drm_buf_t *buf = dma->buflist[ i ];
+			drm_i810_buf_priv_t *buf_priv = buf->dev_private;
+			drm_ioremapfree(buf_priv->kernel_virtual, buf->total);
+		}
 	}
    	return 0;
 }
@@ -325,9 +336,9 @@ static int i810_freelist_init(drm_device_t *dev)
 {
       	drm_device_dma_t *dma = dev->dma;
    	drm_i810_private_t *dev_priv = (drm_i810_private_t *)dev->dev_private;
-   	u8 *hw_status = (u8 *)dev_priv->hw_status_page;
-   	int i;
    	int my_idx = 24;
+   	u32 *hw_status = (u32 *)(dev_priv->hw_status_page + my_idx);
+   	int i;
    
    	if(dma->buf_count > 1019) {
 	   	/* Not enough space in the status page for the freelist */
@@ -338,11 +349,14 @@ static int i810_freelist_init(drm_device_t *dev)
 	   	drm_buf_t *buf = dma->buflist[ i ];
 	   	drm_i810_buf_priv_t *buf_priv = buf->dev_private;
 	   
-	   	buf_priv->in_use = hw_status + my_idx;
-	   	DRM_DEBUG("buf_priv->in_use : %p\n", buf_priv->in_use);
-	   	*buf_priv->in_use = I810_BUF_FREE;
+	   	buf_priv->in_use = hw_status++;
 	   	buf_priv->my_use_idx = my_idx;
 	   	my_idx += 4;
+
+	   	*buf_priv->in_use = I810_BUF_FREE;
+
+		buf_priv->kernel_virtual = drm_ioremap(buf->bus_address, 
+						       buf->total);
 	}
 	return 0;
 }
@@ -376,9 +390,11 @@ static int i810_dma_initialize(drm_device_t *dev,
    	dev_priv->ring.Start = init->ring_start;
    	dev_priv->ring.End = init->ring_end;
    	dev_priv->ring.Size = init->ring_size;
+
    	dev_priv->ring.virtual_start = drm_ioremap(dev->agp->base + 
 						   init->ring_start, 
 						   init->ring_size);
+
    	dev_priv->ring.tail_mask = dev_priv->ring.Size - 1;
    
    	if (dev_priv->ring.virtual_start == NULL) {
@@ -387,6 +403,17 @@ static int i810_dma_initialize(drm_device_t *dev,
 			  " ring buffer\n");
 	   	return -ENOMEM;
 	}
+
+	dev_priv->w = init->w;
+	dev_priv->h = init->h;
+	dev_priv->pitch = init->pitch;
+	dev_priv->back_offset = init->back_offset;
+	dev_priv->depth_offset = init->depth_offset;
+
+	dev_priv->front_di1 = init->front_offset | init->pitch_bits;
+	dev_priv->back_di1 = init->back_offset | init->pitch_bits;
+	dev_priv->zi1 = init->depth_offset | init->pitch_bits;
+	
    
    	/* Program Hardware Status Page */
    	dev_priv->hw_status_page = i810_alloc_page(dev);
@@ -441,40 +468,274 @@ int i810_dma_init(struct inode *inode, struct file *filp,
    	return retcode;
 }
 
-static void i810_dma_dispatch_general(drm_device_t *dev, drm_buf_t *buf,
-				      int used )
+
+
+/* Most efficient way to verify state for the i810 is as it is
+ * emitted.  Non-conformant state is silently dropped.
+ *
+ * Use 'volatile' & local var tmp to force the emitted values to be
+ * identical to the verified ones.
+ */
+static void i810EmitContextVerified( drm_device_t *dev, 
+				     volatile unsigned int *code ) 
+{	
+   	drm_i810_private_t *dev_priv = dev->dev_private;
+	int i, j = 0;
+	unsigned int tmp;
+	RING_LOCALS;
+
+	BEGIN_LP_RING( I810_CTX_SETUP_SIZE );
+
+	OUT_RING( GFX_OP_COLOR_FACTOR );
+	OUT_RING( code[I810_CTXREG_CF1] );
+
+	OUT_RING( GFX_OP_STIPPLE );
+	OUT_RING( code[I810_CTXREG_ST1] );
+
+	for ( i = 4 ; i < I810_CTX_SETUP_SIZE ; i++ ) {
+		tmp = code[i];
+
+		if ((tmp & (7<<29)) == (3<<29) &&
+		    (tmp & (0x1f<<24)) < (0x1d<<24)) 
+		{
+			OUT_RING( tmp ); 
+			j++;
+		} 
+		else
+		   DRM_DEBUG("bad cmd %x\n", tmp);
+	}
+
+	if (j & 1) 
+		OUT_RING( 0 ); 
+
+	ADVANCE_LP_RING();
+}
+
+
+static void i810EmitTexVerified( drm_device_t *dev, 
+				 volatile unsigned int *code ) 
+{	
+   	drm_i810_private_t *dev_priv = dev->dev_private;
+	int i, j = 0;
+	unsigned int tmp;
+	RING_LOCALS;
+
+	BEGIN_LP_RING( I810_TEX_SETUP_SIZE );
+
+	OUT_RING( GFX_OP_MAP_INFO );
+	OUT_RING( code[I810_TEXREG_MI1] );
+	OUT_RING( code[I810_TEXREG_MI2] );
+	OUT_RING( code[I810_TEXREG_MI3] );
+
+	for ( i = 4 ; i < I810_TEX_SETUP_SIZE ; i++ ) {
+		tmp = code[i];
+
+		if ((tmp & (7<<29)) == (3<<29) &&
+		    (tmp & (0x1f<<24)) < (0x1d<<24)) 
+		{
+			OUT_RING( tmp ); 
+			j++;
+		}
+		else
+		   DRM_DEBUG("bad cmd %x\n", tmp);
+	} 
+		
+	if (j & 1) 
+		OUT_RING( 0 ); 
+
+	ADVANCE_LP_RING();
+}
+
+/* Need to do some additional checking when setting the dest buffer.
+ */
+static void i810EmitDestVerified( drm_device_t *dev, 
+				  volatile unsigned int *code ) 
+{	
+   	drm_i810_private_t *dev_priv = dev->dev_private;
+	unsigned int tmp;
+	RING_LOCALS;
+
+	BEGIN_LP_RING( I810_DEST_SETUP_SIZE + 2 );
+
+	tmp = code[I810_DESTREG_DI1];
+	if (tmp == dev_priv->front_di1 || tmp == dev_priv->back_di1) {
+		OUT_RING( CMD_OP_DESTBUFFER_INFO );
+		OUT_RING( tmp );
+	} else
+	   DRM_DEBUG("bad di1 %x (allow %x or %x)\n",
+		     tmp, dev_priv->front_di1, dev_priv->back_di1);
+
+	/* invarient:
+	 */
+	OUT_RING( CMD_OP_Z_BUFFER_INFO );
+	OUT_RING( dev_priv->zi1 );
+
+	OUT_RING( GFX_OP_DESTBUFFER_VARS );
+	OUT_RING( code[I810_DESTREG_DV1] );
+
+	OUT_RING( GFX_OP_DRAWRECT_INFO );
+	OUT_RING( code[I810_DESTREG_DR1] );
+	OUT_RING( code[I810_DESTREG_DR2] );
+	OUT_RING( code[I810_DESTREG_DR3] );
+	OUT_RING( code[I810_DESTREG_DR4] );
+	OUT_RING( 0 );
+
+	ADVANCE_LP_RING();
+}
+
+
+
+static void i810EmitState( drm_device_t *dev )
 {
-      	drm_i810_private_t *dev_priv = dev->dev_private;
-   	drm_i810_buf_priv_t *buf_priv = buf->dev_private;
-	unsigned long address = (unsigned long)buf->bus_address;
-	unsigned long start = address - dev->agp->base;
-   	RING_LOCALS;
+   	drm_i810_private_t *dev_priv = dev->dev_private;
+      	drm_i810_sarea_t *sarea_priv = dev_priv->sarea_priv;
+	unsigned int dirty = sarea_priv->dirty;
 
-   	dev_priv->counter++;
-   	DRM_DEBUG(  "dispatch counter : %ld\n", dev_priv->counter);
-   	DRM_DEBUG(  "i810_dma_dispatch\n");
-   	DRM_DEBUG(  "start : 0x%lx\n", start);
-	DRM_DEBUG(  "used : 0x%x\n", used);
-   	DRM_DEBUG(  "start + used - 4 : 0x%lx\n", start + used - 4);
-   	i810_kernel_lost_context(dev);
+	if (dirty & I810_UPLOAD_BUFFERS) {
+		i810EmitDestVerified( dev, sarea_priv->BufferState );
+		sarea_priv->dirty &= ~I810_UPLOAD_BUFFERS;
+	}
 
-   	BEGIN_LP_RING(10);
-   	OUT_RING( CMD_OP_BATCH_BUFFER );
-   	OUT_RING( start | BB1_PROTECTED );
-   	OUT_RING( start + used - 4 );
-      	OUT_RING( CMD_STORE_DWORD_IDX );
-   	OUT_RING( 20 );
-   	OUT_RING( dev_priv->counter );
-   	OUT_RING( CMD_STORE_DWORD_IDX );
-   	OUT_RING( buf_priv->my_use_idx );
-   	OUT_RING( I810_BUF_FREE );   
-      	OUT_RING( CMD_REPORT_HEAD );
-      	ADVANCE_LP_RING();
-	if(buf_priv->currently_mapped == I810_BUF_MAPPED) {
-		i810_unmap_buffer(buf);
+	if (dirty & I810_UPLOAD_CTX) {
+		i810EmitContextVerified( dev, sarea_priv->ContextState );
+		sarea_priv->dirty &= ~I810_UPLOAD_CTX;
+	}
+
+	if (dirty & I810_UPLOAD_TEX0) {
+		i810EmitTexVerified( dev, sarea_priv->TexState[0] );
+		sarea_priv->dirty &= ~I810_UPLOAD_TEX0;
+	}
+
+	if (dirty & I810_UPLOAD_TEX1) {
+		i810EmitTexVerified( dev, sarea_priv->TexState[1] );
+		sarea_priv->dirty &= ~I810_UPLOAD_TEX1;
+	}
+}
+
+
+
+/* need to verify 
+ */
+static void i810_dma_dispatch_clear( drm_device_t *dev, int flags, 
+				     unsigned int clear_color,
+				     unsigned int clear_zval )
+{
+   	drm_i810_private_t *dev_priv = dev->dev_private;
+      	drm_i810_sarea_t *sarea_priv = dev_priv->sarea_priv;
+	int nbox = sarea_priv->nbox;
+	drm_clip_rect_t *pbox = sarea_priv->boxes;
+	int pitch = dev_priv->pitch;
+	int cpp = 2;
+	int i;
+	RING_LOCALS;
+
+  	i810_kernel_lost_context(dev);
+
+      	if (nbox > I810_NR_SAREA_CLIPRECTS)
+     		nbox = I810_NR_SAREA_CLIPRECTS;
+
+	for (i = 0 ; i < nbox ; i++, pbox++) {
+		unsigned int x = pbox->x1;
+		unsigned int y = pbox->y1;
+		unsigned int width = (pbox->x2 - x) * cpp;
+		unsigned int height = pbox->y2 - y;
+		unsigned int start = y * pitch + x * cpp;
+
+		if (pbox->x1 > pbox->x2 ||
+		    pbox->y1 > pbox->y2 ||
+		    pbox->x2 > dev_priv->w ||
+		    pbox->y2 > dev_priv->h)
+			continue;
+
+	   	if ( flags & I810_FRONT ) {	    
+		   	DRM_DEBUG("clear front\n");
+			BEGIN_LP_RING( 6 );	    
+			OUT_RING( BR00_BITBLT_CLIENT | 
+				  BR00_OP_COLOR_BLT | 0x3 );
+			OUT_RING( BR13_SOLID_PATTERN | (0xF0 << 16) | pitch );
+			OUT_RING( (height << 16) | width );
+			OUT_RING( start );
+			OUT_RING( clear_color );
+			OUT_RING( 0 );
+			ADVANCE_LP_RING();
+		}
+
+		if ( flags & I810_BACK ) {
+			DRM_DEBUG("clear back\n");
+			BEGIN_LP_RING( 6 );	    
+			OUT_RING( BR00_BITBLT_CLIENT | 
+				  BR00_OP_COLOR_BLT | 0x3 );
+			OUT_RING( BR13_SOLID_PATTERN | (0xF0 << 16) | pitch );
+			OUT_RING( (height << 16) | width );
+			OUT_RING( dev_priv->back_offset + start );
+			OUT_RING( clear_color );
+			OUT_RING( 0 );
+			ADVANCE_LP_RING();
+		}
+
+		if ( flags & I810_DEPTH ) {
+			DRM_DEBUG("clear depth\n");
+			BEGIN_LP_RING( 6 );	    
+			OUT_RING( BR00_BITBLT_CLIENT | 
+				  BR00_OP_COLOR_BLT | 0x3 );
+			OUT_RING( BR13_SOLID_PATTERN | (0xF0 << 16) | pitch );
+			OUT_RING( (height << 16) | width );
+			OUT_RING( dev_priv->depth_offset + start );
+			OUT_RING( clear_zval );
+			OUT_RING( 0 );
+			ADVANCE_LP_RING();
+		}
+	}
+}
+
+static void i810_dma_dispatch_swap( drm_device_t *dev )
+{
+   	drm_i810_private_t *dev_priv = dev->dev_private;
+      	drm_i810_sarea_t *sarea_priv = dev_priv->sarea_priv;
+	int nbox = sarea_priv->nbox;
+	drm_clip_rect_t *pbox = sarea_priv->boxes;
+	int pitch = dev_priv->pitch;
+	int cpp = 2;
+	int ofs = dev_priv->back_offset;
+	int i;
+	RING_LOCALS;
+
+	DRM_DEBUG("swapbuffers\n");
+
+  	i810_kernel_lost_context(dev);
+
+      	if (nbox > I810_NR_SAREA_CLIPRECTS)
+     		nbox = I810_NR_SAREA_CLIPRECTS;
+
+	for (i = 0 ; i < nbox; i++, pbox++) 
+	{
+		unsigned int w = pbox->x2 - pbox->x1;
+		unsigned int h = pbox->y2 - pbox->y1;
+		unsigned int dst = pbox->x1*cpp + pbox->y1*pitch;
+		unsigned int start = ofs + dst;
+
+		if (pbox->x1 > pbox->x2 ||
+		    pbox->y1 > pbox->y2 ||
+		    pbox->x2 > dev_priv->w ||
+		    pbox->y2 > dev_priv->h)
+			continue;
+ 
+	   	DRM_DEBUG("dispatch swap %d,%d-%d,%d!\n",
+			  pbox[i].x1, pbox[i].y1,
+			  pbox[i].x2, pbox[i].y2);
+
+		BEGIN_LP_RING( 6 );
+		OUT_RING( BR00_BITBLT_CLIENT | BR00_OP_SRC_COPY_BLT | 0x4 );
+		OUT_RING( pitch | (0xCC << 16));
+		OUT_RING( (h << 16) | (w * cpp));
+		OUT_RING( dst );
+		OUT_RING( pitch );	
+		OUT_RING( start );
+		ADVANCE_LP_RING();
 	}
 }
 
+
 static void i810_dma_dispatch_vertex(drm_device_t *dev, 
 				     drm_buf_t *buf,
 				     int discard,
@@ -487,14 +748,26 @@ static void i810_dma_dispatch_vertex(drm_device_t *dev,
    	int nbox = sarea_priv->nbox;
 	unsigned long address = (unsigned long)buf->bus_address;
 	unsigned long start = address - dev->agp->base;     
-	int i = 0;
+	int i = 0, u;
    	RING_LOCALS;
 
-   
+   	i810_kernel_lost_context(dev);
+
    	if (nbox > I810_NR_SAREA_CLIPRECTS) 
 		nbox = I810_NR_SAREA_CLIPRECTS;
-   
-   	DRM_DEBUG("dispatch vertex addr 0x%lx, used 0x%x nbox %d\n", 
+
+	if (discard) {
+		u = cmpxchg(buf_priv->in_use, I810_BUF_CLIENT, 
+			    I810_BUF_HARDWARE);
+		if(u != I810_BUF_CLIENT) {
+			DRM_DEBUG("xxxx 2\n");
+		}
+	}
+
+	if (sarea_priv->dirty)
+	   i810EmitState( dev );
+
+  	DRM_DEBUG("dispatch vertex addr 0x%lx, used 0x%x nbox %d\n", 
 		  address, used, nbox);
 
    	dev_priv->counter++;
@@ -503,16 +776,24 @@ static void i810_dma_dispatch_vertex(drm_device_t *dev,
    	DRM_DEBUG(  "start : %lx\n", start);
 	DRM_DEBUG(  "used : %d\n", used);
    	DRM_DEBUG(  "start + used - 4 : %ld\n", start + used - 4);
-   	i810_kernel_lost_context(dev);
 
 	if (used) {
+		*(u32 *)buf_priv->virtual = (GFX_OP_PRIMITIVE |
+					     sarea_priv->vertex_prim |
+					     ((used/4)-2));
+
+		if (used & 4) {
+		   *(u32 *)((u32)buf_priv->virtual + used) = 0;
+		   used += 4;
+		}
+		   
 		do {
 			if (i < nbox) {
 				BEGIN_LP_RING(4);
 				OUT_RING( GFX_OP_SCISSOR | SC_UPDATE_SCISSOR | 
 					  SC_ENABLE );
 				OUT_RING( GFX_OP_SCISSOR_INFO );
-				OUT_RING( box[i].x1 | (box[i].y1 << 16) );
+				OUT_RING( box[i].x1 | (box[i].y1<<16) );
 				OUT_RING( (box[i].x2-1) | ((box[i].y2-1)<<16) );
 				ADVANCE_LP_RING();
 			}
@@ -543,6 +824,7 @@ static void i810_dma_dispatch_vertex(drm_device_t *dev,
       	OUT_RING( CMD_REPORT_HEAD );
 	OUT_RING( 0 );
    	ADVANCE_LP_RING();
+
 	if(buf_priv->currently_mapped == I810_BUF_MAPPED) {
 		i810_unmap_buffer(buf);
 	}
@@ -711,8 +993,8 @@ static inline void i810_dma_quiescent_emit(drm_device_t *dev)
    	RING_LOCALS;
 
   	i810_kernel_lost_context(dev);
-   	BEGIN_LP_RING(4);
 
+   	BEGIN_LP_RING(4);
    	OUT_RING( INST_PARSER_CLIENT | INST_OP_FLUSH | INST_FLUSH_MAP_CACHE );
    	OUT_RING( CMD_REPORT_HEAD );
    	OUT_RING( GFX_OP_USER_INTERRUPT );
@@ -794,23 +1076,17 @@ void i810_reclaim_buffers(drm_device_t *dev, pid_t pid)
 	int		 i;
 
 	if (!dma) return;
-      	if(dev->dev_private == NULL) return;
-	if(dma->buflist == NULL) return;
+      	if (!dev->dev_private) return;
+	if (!dma->buflist) return;
+
         i810_flush_queue(dev);
 
 	for (i = 0; i < dma->buf_count; i++) {
 	   	drm_buf_t *buf = dma->buflist[ i ];
 	   	drm_i810_buf_priv_t *buf_priv = buf->dev_private;
 	   
-		/* Only buffers that need to get reclaimed ever 
-		 * get set to free 
-		 */
 		if (buf->pid == pid && buf_priv) {
-		   	cmpxchg(buf_priv->in_use, 
-				I810_BUF_USED, I810_BUF_FREE);
-			if(buf_priv->currently_mapped == I810_BUF_MAPPED) {
-				buf_priv->currently_mapped = I810_BUF_UNMAPPED;
-			}
+		   *(buf_priv->in_use) = I810_BUF_FREE;
 		}
 	}
 }
@@ -900,90 +1176,80 @@ int i810_flush_ioctl(struct inode *inode, struct file *filp,
    	return 0;
 }
 
-static int i810DmaGeneral(drm_device_t *dev, drm_i810_general_t *args)
-{
-	drm_device_dma_t *dma = dev->dma;
-   	drm_buf_t *buf = dma->buflist[ args->idx ];
-   	
-	if (!args->used) {
-	   	i810_freelist_put(dev, buf);
-	} else {  
-		i810_dma_dispatch_general( dev, buf, args->used );
-		atomic_add(args->used, &dma->total_bytes);
-		atomic_inc(&dma->total_dmas);
-	}
-	return 0; 
-}
-
-static int i810DmaVertex(drm_device_t *dev, drm_i810_vertex_t *args)
-{
-	drm_device_dma_t *dma = dev->dma;
-	drm_buf_t *buf = dma->buflist[ args->idx ];
-   	i810_dma_dispatch_vertex( dev, buf, args->discard, args->used );
-   	atomic_add(args->used, &dma->total_bytes);
-	atomic_inc(&dma->total_dmas);
-   	return 0;
-}
 
-int i810_dma_general(struct inode *inode, struct file *filp,
-		     unsigned int cmd, unsigned long arg)
+int i810_dma_vertex(struct inode *inode, struct file *filp,
+	       unsigned int cmd, unsigned long arg)
 {
 	drm_file_t *priv = filp->private_data;
 	drm_device_t *dev = priv->dev;
-	drm_i810_general_t general;
-      	drm_i810_private_t *dev_priv = (drm_i810_private_t *)dev->dev_private;
+	drm_device_dma_t *dma = dev->dma;
+   	drm_i810_private_t *dev_priv = (drm_i810_private_t *)dev->dev_private;
       	u32 *hw_status = (u32 *)dev_priv->hw_status_page;
    	drm_i810_sarea_t *sarea_priv = (drm_i810_sarea_t *) 
      					dev_priv->sarea_priv; 
+	drm_i810_vertex_t vertex;
 
-	int retcode = 0;
-	
-	copy_from_user_ret(&general, (drm_i810_general_t *)arg, sizeof(general),
+	copy_from_user_ret(&vertex, (drm_i810_vertex_t *)arg, sizeof(vertex),
 			   -EFAULT);
 
-	DRM_DEBUG("i810 dma general idx %d used %d\n",
-		  general.idx, general.used);
-   
    	if(!_DRM_LOCK_IS_HELD(dev->lock.hw_lock->lock)) {
-		DRM_ERROR("i810_dma_general called without lock held\n");
+		DRM_ERROR("i810_dma_vertex called without lock held\n");
 		return -EINVAL;
 	}
 
-	retcode = i810DmaGeneral(dev, &general);
+	DRM_DEBUG("i810 dma vertex, idx %d used %d discard %d\n",
+		  vertex.idx, vertex.used, vertex.discard);
+
+	i810_dma_dispatch_vertex( dev, 
+				  dma->buflist[ vertex.idx ], 
+				  vertex.discard, vertex.used );
+
+   	atomic_add(vertex.used, &dma->total_bytes);
+	atomic_inc(&dma->total_dmas);
 	sarea_priv->last_enqueue = dev_priv->counter-1;
    	sarea_priv->last_dispatch = (int) hw_status[5];
    
-	return retcode;
+	return 0;
 }
 
-int i810_dma_vertex(struct inode *inode, struct file *filp,
-	       unsigned int cmd, unsigned long arg)
+
+
+int i810_clear_bufs(struct inode *inode, struct file *filp,
+		   unsigned int cmd, unsigned long arg)
 {
 	drm_file_t *priv = filp->private_data;
 	drm_device_t *dev = priv->dev;
-   	drm_i810_private_t *dev_priv = (drm_i810_private_t *)dev->dev_private;
-      	u32 *hw_status = (u32 *)dev_priv->hw_status_page;
-   	drm_i810_sarea_t *sarea_priv = (drm_i810_sarea_t *) 
-     					dev_priv->sarea_priv; 
-	drm_i810_vertex_t vertex;
-	int retcode = 0;
+	drm_i810_clear_t clear;
 
-	copy_from_user_ret(&vertex, (drm_i810_vertex_t *)arg, sizeof(vertex),
+   	copy_from_user_ret(&clear, (drm_i810_clear_t *)arg, sizeof(clear), 
 			   -EFAULT);
+   
    	if(!_DRM_LOCK_IS_HELD(dev->lock.hw_lock->lock)) {
-		DRM_ERROR("i810_dma_vertex called without lock held\n");
+		DRM_ERROR("i810_clear_bufs called without lock held\n");
 		return -EINVAL;
 	}
 
-	DRM_DEBUG("i810 dma vertex, idx %d used %d discard %d\n",
-		  vertex.idx, vertex.used, vertex.discard);
+	i810_dma_dispatch_clear( dev, clear.flags, 
+				 clear.clear_color, 
+				 clear.clear_depth );
+   	return 0;
+}
 
-	retcode = i810DmaVertex(dev, &vertex);
-	sarea_priv->last_enqueue = dev_priv->counter-1;
-   	sarea_priv->last_dispatch = (int) hw_status[5];
+int i810_swap_bufs(struct inode *inode, struct file *filp,
+		  unsigned int cmd, unsigned long arg)
+{
+	drm_file_t *priv = filp->private_data;
+	drm_device_t *dev = priv->dev;
    
-	return retcode;
+	DRM_DEBUG("i810_swap_bufs\n");
 
+   	if(!_DRM_LOCK_IS_HELD(dev->lock.hw_lock->lock)) {
+		DRM_ERROR("i810_swap_buf called without lock held\n");
+		return -EINVAL;
+	}
+
+	i810_dma_dispatch_swap( dev );
+   	return 0;
 }
 
 int i810_getage(struct inode *inode, struct file *filp, unsigned int cmd,
diff --git a/linux-core/i810_drm.h b/linux-core/i810_drm.h
index 7e435879..4c8e09f6 100644
--- a/linux-core/i810_drm.h
+++ b/linux-core/i810_drm.h
@@ -5,35 +5,112 @@
  * if you change them, you must change the defines in the Xserver.
  */
 
-/* Might one day want to support the client-side ringbuffer code again.
- */
 #ifndef _I810_DEFINES_
 #define _I810_DEFINES_
 
-#define I810_USE_BATCH			1
 #define I810_DMA_BUF_ORDER		12
 #define I810_DMA_BUF_SZ 		(1<<I810_DMA_BUF_ORDER)
 #define I810_DMA_BUF_NR 		256
-#define I810_NR_SAREA_CLIPRECTS 	2
+#define I810_NR_SAREA_CLIPRECTS 	8
 
 /* Each region is a minimum of 64k, and there are at most 64 of them.
  */
-
 #define I810_NR_TEX_REGIONS 64
 #define I810_LOG_MIN_TEX_REGION_SIZE 16
 #endif
 
+#define I810_UPLOAD_TEX0IMAGE  0x1 /* handled clientside */
+#define I810_UPLOAD_TEX1IMAGE  0x2 /* handled clientside */
+#define I810_UPLOAD_CTX        0x4
+#define I810_UPLOAD_BUFFERS    0x8
+#define I810_UPLOAD_TEX0       0x10
+#define I810_UPLOAD_TEX1       0x20
+#define I810_UPLOAD_CLIPRECTS  0x40
+
+
+/* Indices into buf.Setup where various bits of state are mirrored per
+ * context and per buffer.  These can be fired at the card as a unit,
+ * or in a piecewise fashion as required.
+ */
+
+/* Destbuffer state 
+ *    - backbuffer linear offset and pitch -- invarient in the current dri
+ *    - zbuffer linear offset and pitch -- also invarient
+ *    - drawing origin in back and depth buffers.
+ *
+ * Keep the depth/back buffer state here to acommodate private buffers
+ * in the future.
+ */
+#define I810_DESTREG_DI0  0	/* CMD_OP_DESTBUFFER_INFO (2 dwords) */
+#define I810_DESTREG_DI1  1
+#define I810_DESTREG_DV0  2	/* GFX_OP_DESTBUFFER_VARS (2 dwords) */
+#define I810_DESTREG_DV1  3
+#define I810_DESTREG_DR0  4	/* GFX_OP_DRAWRECT_INFO (4 dwords) */
+#define I810_DESTREG_DR1  5
+#define I810_DESTREG_DR2  6
+#define I810_DESTREG_DR3  7
+#define I810_DESTREG_DR4  8
+#define I810_DEST_SETUP_SIZE 10
+
+/* Context state
+ */
+#define I810_CTXREG_CF0   0	/* GFX_OP_COLOR_FACTOR */
+#define I810_CTXREG_CF1   1	
+#define I810_CTXREG_ST0   2     /* GFX_OP_STIPPLE */
+#define I810_CTXREG_ST1   3
+#define I810_CTXREG_VF    4	/* GFX_OP_VERTEX_FMT */
+#define I810_CTXREG_MT    5	/* GFX_OP_MAP_TEXELS */
+#define I810_CTXREG_MC0   6	/* GFX_OP_MAP_COLOR_STAGES - stage 0 */
+#define I810_CTXREG_MC1   7     /* GFX_OP_MAP_COLOR_STAGES - stage 1 */
+#define I810_CTXREG_MC2   8	/* GFX_OP_MAP_COLOR_STAGES - stage 2 */
+#define I810_CTXREG_MA0   9	/* GFX_OP_MAP_ALPHA_STAGES - stage 0 */
+#define I810_CTXREG_MA1   10	/* GFX_OP_MAP_ALPHA_STAGES - stage 1 */
+#define I810_CTXREG_MA2   11	/* GFX_OP_MAP_ALPHA_STAGES - stage 2 */
+#define I810_CTXREG_SDM   12	/* GFX_OP_SRC_DEST_MONO */
+#define I810_CTXREG_FOG   13	/* GFX_OP_FOG_COLOR */
+#define I810_CTXREG_B1    14	/* GFX_OP_BOOL_1 */
+#define I810_CTXREG_B2    15	/* GFX_OP_BOOL_2 */
+#define I810_CTXREG_LCS   16	/* GFX_OP_LINEWIDTH_CULL_SHADE_MODE */
+#define I810_CTXREG_PV    17	/* GFX_OP_PV_RULE -- Invarient! */
+#define I810_CTXREG_ZA    18	/* GFX_OP_ZBIAS_ALPHAFUNC */
+#define I810_CTXREG_AA    19	/* GFX_OP_ANTIALIAS */
+#define I810_CTX_SETUP_SIZE 20 
+
+/* Texture state (per tex unit)
+ */
+#define I810_TEXREG_MI0  0	/* GFX_OP_MAP_INFO (4 dwords) */
+#define I810_TEXREG_MI1  1	
+#define I810_TEXREG_MI2  2	
+#define I810_TEXREG_MI3  3	
+#define I810_TEXREG_MF   4	/* GFX_OP_MAP_FILTER */
+#define I810_TEXREG_MLC  5	/* GFX_OP_MAP_LOD_CTL */
+#define I810_TEXREG_MLL  6	/* GFX_OP_MAP_LOD_LIMITS */
+#define I810_TEXREG_MCS  7	/* GFX_OP_MAP_COORD_SETS ??? */
+#define I810_TEX_SETUP_SIZE 8
+
+#define I810_FRONT   0x1
+#define I810_BACK    0x2
+#define I810_DEPTH   0x4
+
+
 typedef struct _drm_i810_init {
-   	enum {
-	   	I810_INIT_DMA = 0x01,
-	   	I810_CLEANUP_DMA = 0x02
+	enum {
+		I810_INIT_DMA = 0x01,
+		I810_CLEANUP_DMA = 0x02
 	} func;
-   	int ring_map_idx;
-      	int buffer_map_idx;
+	int ring_map_idx;
+	int buffer_map_idx;
 	int sarea_priv_offset;
-   	unsigned long ring_start;
-   	unsigned long ring_end;
-   	unsigned long ring_size;
+	unsigned int ring_start;
+	unsigned int ring_end;
+	unsigned int ring_size;
+	unsigned int front_offset;
+	unsigned int back_offset;
+	unsigned int depth_offset;
+	unsigned int w;
+	unsigned int h;
+	unsigned int pitch;
+	unsigned int pitch_bits; 
 } drm_i810_init_t;
 
 /* Warning: If you change the SAREA structure you must change the Xserver
@@ -46,6 +123,11 @@ typedef struct _drm_i810_tex_region {
 } drm_i810_tex_region_t;
 
 typedef struct _drm_i810_sarea {
+   	unsigned int ContextState[I810_CTX_SETUP_SIZE];
+   	unsigned int BufferState[I810_DEST_SETUP_SIZE];
+   	unsigned int TexState[2][I810_TEX_SETUP_SIZE];
+   	unsigned int dirty;
+
 	unsigned int nbox;
 	drm_clip_rect_t boxes[I810_NR_SAREA_CLIPRECTS];
 
@@ -72,12 +154,18 @@ typedef struct _drm_i810_sarea {
 	int last_dispatch;	/* age of the most recently dispatched buffer */
 	int last_quiescent;     /*  */
 	int ctxOwner;		/* last context to upload state */
+
+	int vertex_prim;
+
 } drm_i810_sarea_t;
 
-typedef struct _drm_i810_general {
-   	int idx;
-	int used;
-} drm_i810_general_t;
+typedef struct _drm_i810_clear {
+	int clear_color;
+	int clear_depth;
+	int flags;
+} drm_i810_clear_t;
+
+
 
 /* These may be placeholders if we have more cliprects than
  * I810_NR_SAREA_CLIPRECTS.  In that case, the client sets discard to
diff --git a/linux-core/i810_drv.c b/linux-core/i810_drv.c
index 03585bdf..75b402da 100644
--- a/linux-core/i810_drv.c
+++ b/linux-core/i810_drv.c
@@ -107,10 +107,11 @@ static drm_ioctl_desc_t	      i810_ioctls[] = {
 
    	[DRM_IOCTL_NR(DRM_IOCTL_I810_INIT)]   = { i810_dma_init,   1, 1 },
    	[DRM_IOCTL_NR(DRM_IOCTL_I810_VERTEX)] = { i810_dma_vertex, 1, 0 },
-   	[DRM_IOCTL_NR(DRM_IOCTL_I810_DMA)]    = { i810_dma_general,1, 0 },
+   	[DRM_IOCTL_NR(DRM_IOCTL_I810_CLEAR)]  = { i810_clear_bufs, 1, 0 },
       	[DRM_IOCTL_NR(DRM_IOCTL_I810_FLUSH)]  = { i810_flush_ioctl,1, 0 },
    	[DRM_IOCTL_NR(DRM_IOCTL_I810_GETAGE)] = { i810_getage,     1, 0 },
 	[DRM_IOCTL_NR(DRM_IOCTL_I810_GETBUF)] = { i810_getbuf,     1, 0 },
+   	[DRM_IOCTL_NR(DRM_IOCTL_I810_SWAP)]   = { i810_swap_bufs,  1, 0 },
 };
 
 #define I810_IOCTL_COUNT DRM_ARRAY_SIZE(i810_ioctls)
diff --git a/linux-core/i810_drv.h b/linux-core/i810_drv.h
index e775d1b6..671489d3 100644
--- a/linux-core/i810_drv.h
+++ b/linux-core/i810_drv.h
@@ -55,6 +55,15 @@ typedef struct drm_i810_private {
 
    	atomic_t flush_done;
    	wait_queue_head_t flush_queue;	/* Processes waiting until flush    */
+
+
+	
+	u32 front_di1, back_di1, zi1;
+	
+	int back_offset;
+	int depth_offset;
+	int w, h;
+	int pitch;
 } drm_i810_private_t;
 
 				/* i810_drv.c */
@@ -128,11 +137,9 @@ typedef struct drm_i810_buf_priv {
    	int my_use_idx;
 	int currently_mapped;
 	void *virtual;
+	void *kernel_virtual;
 } drm_i810_buf_priv_t;
 
-#define I810_DMA_GENERAL 0
-#define I810_DMA_VERTEX  1
-#define I810_DMA_DISCARD 2	/* not used */
 
 #define I810_VERBOSE 0
 
@@ -140,9 +147,11 @@ typedef struct drm_i810_buf_priv {
 int i810_dma_vertex(struct inode *inode, struct file *filp,
 		    unsigned int cmd, unsigned long arg);
 
-int i810_dma_general(struct inode *inode, struct file *filp,
-		     unsigned int cmd, unsigned long arg);
+int i810_swap_bufs(struct inode *inode, struct file *filp,
+		   unsigned int cmd, unsigned long arg);
 
+int i810_clear_bufs(struct inode *inode, struct file *filp,
+		    unsigned int cmd, unsigned long arg);
 
 #define GFX_OP_USER_INTERRUPT 		((0<<29)|(2<<23))
 #define GFX_OP_BREAKPOINT_INTERRUPT	((0<<29)|(1<<23))
@@ -196,5 +205,22 @@ int i810_dma_general(struct inode *inode, struct file *filp,
 #define SCI_YMAX_MASK      (0xffff<<16)
 #define SCI_XMAX_MASK      (0xffff<<0)
 
+#define GFX_OP_COLOR_FACTOR      ((0x3<<29)|(0x1d<<24)|(0x1<<16)|0x0)
+#define GFX_OP_STIPPLE           ((0x3<<29)|(0x1d<<24)|(0x83<<16))
+#define GFX_OP_MAP_INFO          ((0x3<<29)|(0x1d<<24)|0x2)
+#define GFX_OP_DESTBUFFER_VARS   ((0x3<<29)|(0x1d<<24)|(0x85<<16)|0x0)
+#define GFX_OP_DRAWRECT_INFO     ((0x3<<29)|(0x1d<<24)|(0x80<<16)|(0x3))
+#define GFX_OP_PRIMITIVE         ((0x3<<29)|(0x1f<<24))
+
+#define CMD_OP_Z_BUFFER_INFO     ((0x0<<29)|(0x16<<23))
+#define CMD_OP_DESTBUFFER_INFO   ((0x0<<29)|(0x15<<23))
+
+#define BR00_BITBLT_CLIENT   0x40000000
+#define BR00_OP_COLOR_BLT    0x10000000
+#define BR00_OP_SRC_COPY_BLT 0x10C00000
+#define BR13_SOLID_PATTERN   0x80000000
+
+
+
 #endif
 
diff --git a/linux/Makefile.linux b/linux/Makefile.linux
index 1db7a81f..1a34ad0c 100644
--- a/linux/Makefile.linux
+++ b/linux/Makefile.linux
@@ -190,3 +190,253 @@ $(PROGOBJS): $(PROGHEADERS)
 
 clean:
 	rm -f *.o *.a *.po *~ core $(PROGS)
+# DO NOT DELETE
+
+agpsupport.o: drmP.h
+auth.o: drmP.h
+bufs.o: drmP.h /usr/include/linux/un.h
+context.o: drmP.h
+ctxbitmap.o: drmP.h
+dma.o: drmP.h /usr/include/linux/interrupt.h /usr/include/linux/kernel.h
+dma.o: /usr/include/asm/bitops.h /usr/include/asm/atomic.h
+dma.o: /usr/include/asm/hardirq.h /usr/include/linux/threads.h
+dma.o: /usr/include/asm/softirq.h
+drawable.o: drmP.h
+drm.o: /usr/include/asm/ioctl.h mga_drm.h i810_drm.h
+drmstat.o: /usr/include/stdio.h /usr/include/features.h
+drmstat.o: /usr/include/sys/cdefs.h /usr/include/gnu/stubs.h
+drmstat.o: /usr/lib/gcc-lib/i386-redhat-linux/egcs-2.91.66/include/stddef.h
+drmstat.o: /usr/lib/gcc-lib/i386-redhat-linux/egcs-2.91.66/include/stdarg.h
+drmstat.o: /usr/include/bits/types.h /usr/include/libio.h
+drmstat.o: /usr/include/_G_config.h /usr/include/bits/stdio_lim.h
+drmstat.o: /usr/include/stdlib.h /usr/include/sys/types.h /usr/include/time.h
+drmstat.o: /usr/include/endian.h /usr/include/bits/endian.h
+drmstat.o: /usr/include/sys/select.h /usr/include/bits/select.h
+drmstat.o: /usr/include/bits/sigset.h /usr/include/sys/sysmacros.h
+drmstat.o: /usr/include/alloca.h /usr/include/unistd.h
+drmstat.o: /usr/include/bits/posix_opt.h /usr/include/bits/confname.h
+drmstat.o: /usr/include/getopt.h /usr/include/sys/time.h
+drmstat.o: /usr/include/bits/time.h /usr/include/sys/mman.h
+drmstat.o: /usr/include/bits/mman.h /usr/include/strings.h
+drmstat.o: /usr/include/errno.h /usr/include/bits/errno.h
+drmstat.o: /usr/include/linux/errno.h /usr/include/asm/errno.h
+drmstat.o: /usr/include/signal.h /usr/include/bits/signum.h
+drmstat.o: /usr/include/bits/siginfo.h /usr/include/bits/sigaction.h
+drmstat.o: /usr/include/bits/sigcontext.h /usr/include/asm/sigcontext.h
+drmstat.o: /usr/include/bits/sigstack.h /usr/include/fcntl.h
+drmstat.o: /usr/include/bits/fcntl.h xf86drm.h
+fops.o: drmP.h /usr/include/linux/poll.h /usr/include/asm/poll.h
+gamma_dma.o: drmP.h gamma_drv.h /usr/include/linux/interrupt.h
+gamma_dma.o: /usr/include/linux/kernel.h /usr/include/asm/bitops.h
+gamma_dma.o: /usr/include/asm/atomic.h /usr/include/asm/hardirq.h
+gamma_dma.o: /usr/include/linux/threads.h /usr/include/asm/softirq.h
+gamma_drv.o: drmP.h gamma_drv.h
+i810_bufs.o: drmP.h i810_drv.h /usr/include/linux/un.h
+i810_context.o: /usr/include/linux/sched.h /usr/include/asm/param.h
+i810_context.o: /usr/include/linux/binfmts.h /usr/include/linux/ptrace.h
+i810_context.o: /usr/include/asm/ptrace.h /usr/include/linux/capability.h
+i810_context.o: /usr/include/linux/types.h /usr/include/linux/config.h
+i810_context.o: /usr/include/linux/autoconf.h
+i810_context.o: /usr/include/linux/posix_types.h /usr/include/linux/stddef.h
+i810_context.o: /usr/include/asm/posix_types.h /usr/include/asm/types.h
+i810_context.o: /usr/include/linux/fs.h /usr/include/linux/linkage.h
+i810_context.o: /usr/include/linux/limits.h /usr/include/linux/wait.h
+i810_context.o: /usr/include/linux/vfs.h /usr/include/asm/statfs.h
+i810_context.o: /usr/include/linux/net.h /usr/include/linux/socket.h
+i810_context.o: /usr/include/asm/socket.h /usr/include/asm/sockios.h
+i810_context.o: /usr/include/linux/sockios.h /usr/include/linux/uio.h
+i810_context.o: /usr/include/linux/kdev_t.h /usr/include/linux/ioctl.h
+i810_context.o: /usr/include/asm/ioctl.h /usr/include/linux/list.h
+i810_context.o: /usr/include/linux/dcache.h /usr/include/linux/stat.h
+i810_context.o: /usr/include/linux/cache.h /usr/include/asm/cache.h
+i810_context.o: /usr/include/asm/atomic.h /usr/include/asm/bitops.h
+i810_context.o: /usr/include/linux/personality.h /usr/include/linux/threads.h
+i810_context.o: /usr/include/linux/kernel.h /usr/include/linux/times.h
+i810_context.o: /usr/include/linux/timex.h /usr/include/asm/timex.h
+i810_context.o: /usr/include/asm/msr.h /usr/include/asm/system.h
+i810_context.o: /usr/include/asm/segment.h /usr/include/linux/bitops.h
+i810_context.o: /usr/include/asm/semaphore.h /usr/include/asm/rwlock.h
+i810_context.o: /usr/include/asm/page.h /usr/include/linux/smp.h
+i810_context.o: /usr/include/linux/tty.h /usr/include/linux/sem.h
+i810_context.o: /usr/include/linux/ipc.h /usr/include/asm/ipcbuf.h
+i810_context.o: /usr/include/asm/sembuf.h /usr/include/linux/signal.h
+i810_context.o: /usr/include/asm/signal.h /usr/include/asm/siginfo.h
+i810_context.o: /usr/include/linux/securebits.h /usr/include/linux/time.h
+i810_context.o: /usr/include/linux/param.h /usr/include/linux/resource.h
+i810_context.o: /usr/include/asm/resource.h /usr/include/linux/timer.h
+i810_context.o: /usr/include/asm/processor.h /usr/include/asm/vm86.h
+i810_context.o: /usr/include/asm/math_emu.h /usr/include/asm/sigcontext.h
+i810_context.o: drmP.h i810_drv.h
+i810_dma.o: drmP.h i810_drv.h /usr/include/linux/interrupt.h
+i810_dma.o: /usr/include/linux/kernel.h /usr/include/asm/bitops.h
+i810_dma.o: /usr/include/asm/atomic.h /usr/include/asm/hardirq.h
+i810_dma.o: /usr/include/linux/threads.h /usr/include/asm/softirq.h
+i810_drv.o: drmP.h i810_drv.h
+i810_state.o: drmP.h i810_drv.h drm.h /usr/include/asm/ioctl.h mga_drm.h
+i810_state.o: i810_drm.h
+init.o: drmP.h
+ioctl.o: drmP.h
+lists.o: drmP.h
+lock.o: drmP.h
+memory.o: drmP.h
+mga_bufs.o: drmP.h mga_drv.h /usr/include/linux/un.h
+mga_context.o: /usr/include/linux/sched.h /usr/include/asm/param.h
+mga_context.o: /usr/include/linux/binfmts.h /usr/include/linux/ptrace.h
+mga_context.o: /usr/include/asm/ptrace.h /usr/include/linux/capability.h
+mga_context.o: /usr/include/linux/types.h /usr/include/linux/config.h
+mga_context.o: /usr/include/linux/autoconf.h /usr/include/linux/posix_types.h
+mga_context.o: /usr/include/linux/stddef.h /usr/include/asm/posix_types.h
+mga_context.o: /usr/include/asm/types.h /usr/include/linux/fs.h
+mga_context.o: /usr/include/linux/linkage.h /usr/include/linux/limits.h
+mga_context.o: /usr/include/linux/wait.h /usr/include/linux/vfs.h
+mga_context.o: /usr/include/asm/statfs.h /usr/include/linux/net.h
+mga_context.o: /usr/include/linux/socket.h /usr/include/asm/socket.h
+mga_context.o: /usr/include/asm/sockios.h /usr/include/linux/sockios.h
+mga_context.o: /usr/include/linux/uio.h /usr/include/linux/kdev_t.h
+mga_context.o: /usr/include/linux/ioctl.h /usr/include/asm/ioctl.h
+mga_context.o: /usr/include/linux/list.h /usr/include/linux/dcache.h
+mga_context.o: /usr/include/linux/stat.h /usr/include/linux/cache.h
+mga_context.o: /usr/include/asm/cache.h /usr/include/asm/atomic.h
+mga_context.o: /usr/include/asm/bitops.h /usr/include/linux/personality.h
+mga_context.o: /usr/include/linux/threads.h /usr/include/linux/kernel.h
+mga_context.o: /usr/include/linux/times.h /usr/include/linux/timex.h
+mga_context.o: /usr/include/asm/timex.h /usr/include/asm/msr.h
+mga_context.o: /usr/include/asm/system.h /usr/include/asm/segment.h
+mga_context.o: /usr/include/linux/bitops.h /usr/include/asm/semaphore.h
+mga_context.o: /usr/include/asm/rwlock.h /usr/include/asm/page.h
+mga_context.o: /usr/include/linux/smp.h /usr/include/linux/tty.h
+mga_context.o: /usr/include/linux/sem.h /usr/include/linux/ipc.h
+mga_context.o: /usr/include/asm/ipcbuf.h /usr/include/asm/sembuf.h
+mga_context.o: /usr/include/linux/signal.h /usr/include/asm/signal.h
+mga_context.o: /usr/include/asm/siginfo.h /usr/include/linux/securebits.h
+mga_context.o: /usr/include/linux/time.h /usr/include/linux/param.h
+mga_context.o: /usr/include/linux/resource.h /usr/include/asm/resource.h
+mga_context.o: /usr/include/linux/timer.h /usr/include/asm/processor.h
+mga_context.o: /usr/include/asm/vm86.h /usr/include/asm/math_emu.h
+mga_context.o: /usr/include/asm/sigcontext.h drmP.h mga_drv.h
+mga_dma.o: drmP.h mga_drv.h /usr/include/linux/interrupt.h
+mga_dma.o: /usr/include/linux/kernel.h /usr/include/asm/bitops.h
+mga_dma.o: /usr/include/asm/atomic.h /usr/include/asm/hardirq.h
+mga_dma.o: /usr/include/linux/threads.h /usr/include/asm/softirq.h
+mga_drv.o: drmP.h mga_drv.h
+mga_state.o: drmP.h mga_drv.h drm.h /usr/include/asm/ioctl.h mga_drm.h
+mga_state.o: i810_drm.h
+picker.o: /usr/include/linux/autoconf.h /usr/include/linux/version.h
+proc.o: drmP.h
+sigio.o: /usr/include/unistd.h /usr/include/features.h
+sigio.o: /usr/include/sys/cdefs.h /usr/include/gnu/stubs.h
+sigio.o: /usr/include/bits/posix_opt.h /usr/include/bits/types.h
+sigio.o: /usr/lib/gcc-lib/i386-redhat-linux/egcs-2.91.66/include/stddef.h
+sigio.o: /usr/include/bits/confname.h /usr/include/getopt.h
+sigio.o: /usr/include/signal.h /usr/include/bits/sigset.h
+sigio.o: /usr/include/bits/signum.h /usr/include/time.h
+sigio.o: /usr/include/bits/siginfo.h /usr/include/bits/sigaction.h
+sigio.o: /usr/include/bits/sigcontext.h /usr/include/asm/sigcontext.h
+sigio.o: /usr/include/bits/sigstack.h /usr/include/fcntl.h
+sigio.o: /usr/include/bits/fcntl.h /usr/include/sys/types.h
+sigio.o: /usr/include/endian.h /usr/include/bits/endian.h
+sigio.o: /usr/include/sys/select.h /usr/include/bits/select.h
+sigio.o: /usr/include/sys/sysmacros.h /usr/include/sys/time.h
+sigio.o: /usr/include/bits/time.h /usr/include/errno.h
+sigio.o: /usr/include/bits/errno.h /usr/include/linux/errno.h
+sigio.o: /usr/include/asm/errno.h /usr/include/stdio.h
+sigio.o: /usr/lib/gcc-lib/i386-redhat-linux/egcs-2.91.66/include/stdarg.h
+sigio.o: /usr/include/libio.h /usr/include/_G_config.h
+sigio.o: /usr/include/bits/stdio_lim.h /usr/include/string.h
+tdfx_context.o: /usr/include/linux/sched.h /usr/include/asm/param.h
+tdfx_context.o: /usr/include/linux/binfmts.h /usr/include/linux/ptrace.h
+tdfx_context.o: /usr/include/asm/ptrace.h /usr/include/linux/capability.h
+tdfx_context.o: /usr/include/linux/types.h /usr/include/linux/config.h
+tdfx_context.o: /usr/include/linux/autoconf.h
+tdfx_context.o: /usr/include/linux/posix_types.h /usr/include/linux/stddef.h
+tdfx_context.o: /usr/include/asm/posix_types.h /usr/include/asm/types.h
+tdfx_context.o: /usr/include/linux/fs.h /usr/include/linux/linkage.h
+tdfx_context.o: /usr/include/linux/limits.h /usr/include/linux/wait.h
+tdfx_context.o: /usr/include/linux/vfs.h /usr/include/asm/statfs.h
+tdfx_context.o: /usr/include/linux/net.h /usr/include/linux/socket.h
+tdfx_context.o: /usr/include/asm/socket.h /usr/include/asm/sockios.h
+tdfx_context.o: /usr/include/linux/sockios.h /usr/include/linux/uio.h
+tdfx_context.o: /usr/include/linux/kdev_t.h /usr/include/linux/ioctl.h
+tdfx_context.o: /usr/include/asm/ioctl.h /usr/include/linux/list.h
+tdfx_context.o: /usr/include/linux/dcache.h /usr/include/linux/stat.h
+tdfx_context.o: /usr/include/linux/cache.h /usr/include/asm/cache.h
+tdfx_context.o: /usr/include/asm/atomic.h /usr/include/asm/bitops.h
+tdfx_context.o: /usr/include/linux/personality.h /usr/include/linux/threads.h
+tdfx_context.o: /usr/include/linux/kernel.h /usr/include/linux/times.h
+tdfx_context.o: /usr/include/linux/timex.h /usr/include/asm/timex.h
+tdfx_context.o: /usr/include/asm/msr.h /usr/include/asm/system.h
+tdfx_context.o: /usr/include/asm/segment.h /usr/include/linux/bitops.h
+tdfx_context.o: /usr/include/asm/semaphore.h /usr/include/asm/rwlock.h
+tdfx_context.o: /usr/include/asm/page.h /usr/include/linux/smp.h
+tdfx_context.o: /usr/include/linux/tty.h /usr/include/linux/sem.h
+tdfx_context.o: /usr/include/linux/ipc.h /usr/include/asm/ipcbuf.h
+tdfx_context.o: /usr/include/asm/sembuf.h /usr/include/linux/signal.h
+tdfx_context.o: /usr/include/asm/signal.h /usr/include/asm/siginfo.h
+tdfx_context.o: /usr/include/linux/securebits.h /usr/include/linux/time.h
+tdfx_context.o: /usr/include/linux/param.h /usr/include/linux/resource.h
+tdfx_context.o: /usr/include/asm/resource.h /usr/include/linux/timer.h
+tdfx_context.o: /usr/include/asm/processor.h /usr/include/asm/vm86.h
+tdfx_context.o: /usr/include/asm/math_emu.h /usr/include/asm/sigcontext.h
+tdfx_context.o: drmP.h tdfx_drv.h
+tdfx_drv.o: drmP.h tdfx_drv.h
+vm.o: drmP.h
+xf86_OSproc.o: /usr/include/X11/Xfuncproto.h
+xf86drm.o: /usr/include/stdio.h /usr/include/features.h
+xf86drm.o: /usr/include/sys/cdefs.h /usr/include/gnu/stubs.h
+xf86drm.o: /usr/lib/gcc-lib/i386-redhat-linux/egcs-2.91.66/include/stddef.h
+xf86drm.o: /usr/lib/gcc-lib/i386-redhat-linux/egcs-2.91.66/include/stdarg.h
+xf86drm.o: /usr/include/bits/types.h /usr/include/libio.h
+xf86drm.o: /usr/include/_G_config.h /usr/include/bits/stdio_lim.h
+xf86drm.o: /usr/include/stdlib.h /usr/include/sys/types.h /usr/include/time.h
+xf86drm.o: /usr/include/endian.h /usr/include/bits/endian.h
+xf86drm.o: /usr/include/sys/select.h /usr/include/bits/select.h
+xf86drm.o: /usr/include/bits/sigset.h /usr/include/sys/sysmacros.h
+xf86drm.o: /usr/include/alloca.h /usr/include/unistd.h
+xf86drm.o: /usr/include/bits/posix_opt.h /usr/include/bits/confname.h
+xf86drm.o: /usr/include/getopt.h /usr/include/string.h /usr/include/ctype.h
+xf86drm.o: /usr/include/fcntl.h /usr/include/bits/fcntl.h
+xf86drm.o: /usr/include/errno.h /usr/include/bits/errno.h
+xf86drm.o: /usr/include/linux/errno.h /usr/include/asm/errno.h
+xf86drm.o: /usr/include/signal.h /usr/include/bits/signum.h
+xf86drm.o: /usr/include/bits/siginfo.h /usr/include/bits/sigaction.h
+xf86drm.o: /usr/include/bits/sigcontext.h /usr/include/asm/sigcontext.h
+xf86drm.o: /usr/include/bits/sigstack.h /usr/include/sys/stat.h
+xf86drm.o: /usr/include/bits/stat.h /usr/include/sys/ioctl.h
+xf86drm.o: /usr/include/bits/ioctls.h /usr/include/asm/ioctls.h
+xf86drm.o: /usr/include/asm/ioctl.h /usr/include/bits/ioctl-types.h
+xf86drm.o: /usr/include/sys/ttydefaults.h /usr/include/sys/mman.h
+xf86drm.o: /usr/include/bits/mman.h /usr/include/sys/time.h
+xf86drm.o: /usr/include/bits/time.h xf86drm.h drm.h mga_drm.h i810_drm.h
+xf86drmHash.o: xf86drm.h /usr/include/stdio.h /usr/include/features.h
+xf86drmHash.o: /usr/include/sys/cdefs.h /usr/include/gnu/stubs.h
+xf86drmHash.o: /usr/lib/gcc-lib/i386-redhat-linux/egcs-2.91.66/include/stddef.h
+xf86drmHash.o: /usr/lib/gcc-lib/i386-redhat-linux/egcs-2.91.66/include/stdarg.h
+xf86drmHash.o: /usr/include/bits/types.h /usr/include/libio.h
+xf86drmHash.o: /usr/include/_G_config.h /usr/include/bits/stdio_lim.h
+xf86drmHash.o: /usr/include/stdlib.h /usr/include/sys/types.h
+xf86drmHash.o: /usr/include/time.h /usr/include/endian.h
+xf86drmHash.o: /usr/include/bits/endian.h /usr/include/sys/select.h
+xf86drmHash.o: /usr/include/bits/select.h /usr/include/bits/sigset.h
+xf86drmHash.o: /usr/include/sys/sysmacros.h /usr/include/alloca.h
+xf86drmRandom.o: xf86drm.h /usr/include/stdio.h /usr/include/features.h
+xf86drmRandom.o: /usr/include/sys/cdefs.h /usr/include/gnu/stubs.h
+xf86drmRandom.o: /usr/lib/gcc-lib/i386-redhat-linux/egcs-2.91.66/include/stddef.h
+xf86drmRandom.o: /usr/lib/gcc-lib/i386-redhat-linux/egcs-2.91.66/include/stdarg.h
+xf86drmRandom.o: /usr/include/bits/types.h /usr/include/libio.h
+xf86drmRandom.o: /usr/include/_G_config.h /usr/include/bits/stdio_lim.h
+xf86drmRandom.o: /usr/include/stdlib.h /usr/include/sys/types.h
+xf86drmRandom.o: /usr/include/time.h /usr/include/endian.h
+xf86drmRandom.o: /usr/include/bits/endian.h /usr/include/sys/select.h
+xf86drmRandom.o: /usr/include/bits/select.h /usr/include/bits/sigset.h
+xf86drmRandom.o: /usr/include/sys/sysmacros.h /usr/include/alloca.h
+xf86drmSL.o: xf86drm.h /usr/include/stdio.h /usr/include/features.h
+xf86drmSL.o: /usr/include/sys/cdefs.h /usr/include/gnu/stubs.h
+xf86drmSL.o: /usr/lib/gcc-lib/i386-redhat-linux/egcs-2.91.66/include/stddef.h
+xf86drmSL.o: /usr/lib/gcc-lib/i386-redhat-linux/egcs-2.91.66/include/stdarg.h
+xf86drmSL.o: /usr/include/bits/types.h /usr/include/libio.h
+xf86drmSL.o: /usr/include/_G_config.h /usr/include/bits/stdio_lim.h
+xf86drmSL.o: /usr/include/stdlib.h /usr/include/sys/types.h
+xf86drmSL.o: /usr/include/time.h /usr/include/endian.h
+xf86drmSL.o: /usr/include/bits/endian.h /usr/include/sys/select.h
+xf86drmSL.o: /usr/include/bits/select.h /usr/include/bits/sigset.h
+xf86drmSL.o: /usr/include/sys/sysmacros.h /usr/include/alloca.h
diff --git a/linux/drm.h b/linux/drm.h
index 86ab0410..d196d912 100644
--- a/linux/drm.h
+++ b/linux/drm.h
@@ -345,8 +345,9 @@ typedef struct drm_agp_info {
 /* I810 specific ioctls */
 #define DRM_IOCTL_I810_INIT    DRM_IOW( 0x40, drm_i810_init_t)
 #define DRM_IOCTL_I810_VERTEX  DRM_IOW( 0x41, drm_i810_vertex_t)
-#define DRM_IOCTL_I810_DMA     DRM_IOW( 0x42, drm_i810_general_t)
+#define DRM_IOCTL_I810_CLEAR   DRM_IOW( 0x42, drm_i810_clear_t)
 #define DRM_IOCTL_I810_FLUSH   DRM_IO ( 0x43)
 #define DRM_IOCTL_I810_GETAGE  DRM_IO ( 0x44)
 #define DRM_IOCTL_I810_GETBUF  DRM_IOW( 0x45, drm_i810_dma_t)
+#define DRM_IOCTL_I810_SWAP    DRM_IO ( 0x46)
 #endif
diff --git a/linux/i810_dma.c b/linux/i810_dma.c
index c940f5cd..5c85a4dc 100644
--- a/linux/i810_dma.c
+++ b/linux/i810_dma.c
@@ -37,8 +37,9 @@
 
 #include <linux/interrupt.h>	/* For task queue support */
 
-#define I810_BUF_FREE		1
-#define I810_BUF_USED		0
+#define I810_BUF_FREE		2
+#define I810_BUF_CLIENT		1
+#define I810_BUF_HARDWARE      	0
 
 #define I810_BUF_UNMAPPED 0
 #define I810_BUF_MAPPED   1
@@ -94,7 +95,7 @@ static inline void i810_print_status_page(drm_device_t *dev)
       	DRM_DEBUG(  "hw_status: Reserved : %x\n", temp[3]);
    	DRM_DEBUG(  "hw_status: Driver Counter : %d\n", temp[5]);
    	for(i = 6; i < dma->buf_count + 6; i++) {
-	   	DRM_DEBUG(  "buffer status idx : %d used: %d\n", i - 6, temp[i]);
+	   	DRM_DEBUG( "buffer status idx : %d used: %d\n", i - 6, temp[i]);
 	}
 }
 
@@ -111,7 +112,7 @@ static drm_buf_t *i810_freelist_get(drm_device_t *dev)
 	   	drm_i810_buf_priv_t *buf_priv = buf->dev_private;
 		/* In use is already a pointer */
 	   	used = cmpxchg(buf_priv->in_use, I810_BUF_FREE, 
-			       I810_BUF_USED);
+			       I810_BUF_CLIENT);
 	   	if(used == I810_BUF_FREE) {
 			return buf;
 		}
@@ -129,8 +130,8 @@ static int i810_freelist_put(drm_device_t *dev, drm_buf_t *buf)
    	int used;
    
    	/* In use is already a pointer */
-   	used = cmpxchg(buf_priv->in_use, I810_BUF_USED, I810_BUF_FREE);
-   	if(used != I810_BUF_USED) {
+   	used = cmpxchg(buf_priv->in_use, I810_BUF_CLIENT, I810_BUF_FREE);
+   	if(used != I810_BUF_CLIENT) {
 	   	DRM_ERROR("Freeing buffer thats not in use : %d\n", buf->idx);
 	   	return -EINVAL;
 	}
@@ -218,7 +219,8 @@ static int i810_dma_get_buffer(drm_device_t *dev, drm_i810_dma_t *d,
 	retcode = i810_map_buffer(buf, filp);
 	if(retcode) {
 		i810_freelist_put(dev, buf);
-	   	DRM_DEBUG("mapbuf failed in %s retcode %d\n", __FUNCTION__, retcode);
+	   	DRM_DEBUG("mapbuf failed in %s retcode %d\n", 
+			  __FUNCTION__, retcode);
 	   	goto out_get_buf;
 	}
 	buf->pid     = current->pid;
@@ -260,7 +262,10 @@ static void i810_free_page(drm_device_t *dev, unsigned long page)
 
 static int i810_dma_cleanup(drm_device_t *dev)
 {
+	drm_device_dma_t *dma = dev->dma;
+
 	if(dev->dev_private) {
+		int i;
 	   	drm_i810_private_t *dev_priv = 
 	     		(drm_i810_private_t *) dev->dev_private;
 	   
@@ -276,6 +281,12 @@ static int i810_dma_cleanup(drm_device_t *dev)
 	   	drm_free(dev->dev_private, sizeof(drm_i810_private_t), 
 			 DRM_MEM_DRIVER);
 	   	dev->dev_private = NULL;
+
+		for (i = 0; i < dma->buf_count; i++) {
+			drm_buf_t *buf = dma->buflist[ i ];
+			drm_i810_buf_priv_t *buf_priv = buf->dev_private;
+			drm_ioremapfree(buf_priv->kernel_virtual, buf->total);
+		}
 	}
    	return 0;
 }
@@ -325,9 +336,9 @@ static int i810_freelist_init(drm_device_t *dev)
 {
       	drm_device_dma_t *dma = dev->dma;
    	drm_i810_private_t *dev_priv = (drm_i810_private_t *)dev->dev_private;
-   	u8 *hw_status = (u8 *)dev_priv->hw_status_page;
-   	int i;
    	int my_idx = 24;
+   	u32 *hw_status = (u32 *)(dev_priv->hw_status_page + my_idx);
+   	int i;
    
    	if(dma->buf_count > 1019) {
 	   	/* Not enough space in the status page for the freelist */
@@ -338,11 +349,14 @@ static int i810_freelist_init(drm_device_t *dev)
 	   	drm_buf_t *buf = dma->buflist[ i ];
 	   	drm_i810_buf_priv_t *buf_priv = buf->dev_private;
 	   
-	   	buf_priv->in_use = hw_status + my_idx;
-	   	DRM_DEBUG("buf_priv->in_use : %p\n", buf_priv->in_use);
-	   	*buf_priv->in_use = I810_BUF_FREE;
+	   	buf_priv->in_use = hw_status++;
 	   	buf_priv->my_use_idx = my_idx;
 	   	my_idx += 4;
+
+	   	*buf_priv->in_use = I810_BUF_FREE;
+
+		buf_priv->kernel_virtual = drm_ioremap(buf->bus_address, 
+						       buf->total);
 	}
 	return 0;
 }
@@ -376,9 +390,11 @@ static int i810_dma_initialize(drm_device_t *dev,
    	dev_priv->ring.Start = init->ring_start;
    	dev_priv->ring.End = init->ring_end;
    	dev_priv->ring.Size = init->ring_size;
+
    	dev_priv->ring.virtual_start = drm_ioremap(dev->agp->base + 
 						   init->ring_start, 
 						   init->ring_size);
+
    	dev_priv->ring.tail_mask = dev_priv->ring.Size - 1;
    
    	if (dev_priv->ring.virtual_start == NULL) {
@@ -387,6 +403,17 @@ static int i810_dma_initialize(drm_device_t *dev,
 			  " ring buffer\n");
 	   	return -ENOMEM;
 	}
+
+	dev_priv->w = init->w;
+	dev_priv->h = init->h;
+	dev_priv->pitch = init->pitch;
+	dev_priv->back_offset = init->back_offset;
+	dev_priv->depth_offset = init->depth_offset;
+
+	dev_priv->front_di1 = init->front_offset | init->pitch_bits;
+	dev_priv->back_di1 = init->back_offset | init->pitch_bits;
+	dev_priv->zi1 = init->depth_offset | init->pitch_bits;
+	
    
    	/* Program Hardware Status Page */
    	dev_priv->hw_status_page = i810_alloc_page(dev);
@@ -441,40 +468,274 @@ int i810_dma_init(struct inode *inode, struct file *filp,
    	return retcode;
 }
 
-static void i810_dma_dispatch_general(drm_device_t *dev, drm_buf_t *buf,
-				      int used )
+
+
+/* Most efficient way to verify state for the i810 is as it is
+ * emitted.  Non-conformant state is silently dropped.
+ *
+ * Use 'volatile' & local var tmp to force the emitted values to be
+ * identical to the verified ones.
+ */
+static void i810EmitContextVerified( drm_device_t *dev, 
+				     volatile unsigned int *code ) 
+{	
+   	drm_i810_private_t *dev_priv = dev->dev_private;
+	int i, j = 0;
+	unsigned int tmp;
+	RING_LOCALS;
+
+	BEGIN_LP_RING( I810_CTX_SETUP_SIZE );
+
+	OUT_RING( GFX_OP_COLOR_FACTOR );
+	OUT_RING( code[I810_CTXREG_CF1] );
+
+	OUT_RING( GFX_OP_STIPPLE );
+	OUT_RING( code[I810_CTXREG_ST1] );
+
+	for ( i = 4 ; i < I810_CTX_SETUP_SIZE ; i++ ) {
+		tmp = code[i];
+
+		if ((tmp & (7<<29)) == (3<<29) &&
+		    (tmp & (0x1f<<24)) < (0x1d<<24)) 
+		{
+			OUT_RING( tmp ); 
+			j++;
+		} 
+		else
+		   DRM_DEBUG("bad cmd %x\n", tmp);
+	}
+
+	if (j & 1) 
+		OUT_RING( 0 ); 
+
+	ADVANCE_LP_RING();
+}
+
+
+static void i810EmitTexVerified( drm_device_t *dev, 
+				 volatile unsigned int *code ) 
+{	
+   	drm_i810_private_t *dev_priv = dev->dev_private;
+	int i, j = 0;
+	unsigned int tmp;
+	RING_LOCALS;
+
+	BEGIN_LP_RING( I810_TEX_SETUP_SIZE );
+
+	OUT_RING( GFX_OP_MAP_INFO );
+	OUT_RING( code[I810_TEXREG_MI1] );
+	OUT_RING( code[I810_TEXREG_MI2] );
+	OUT_RING( code[I810_TEXREG_MI3] );
+
+	for ( i = 4 ; i < I810_TEX_SETUP_SIZE ; i++ ) {
+		tmp = code[i];
+
+		if ((tmp & (7<<29)) == (3<<29) &&
+		    (tmp & (0x1f<<24)) < (0x1d<<24)) 
+		{
+			OUT_RING( tmp ); 
+			j++;
+		}
+		else
+		   DRM_DEBUG("bad cmd %x\n", tmp);
+	} 
+		
+	if (j & 1) 
+		OUT_RING( 0 ); 
+
+	ADVANCE_LP_RING();
+}
+
+/* Need to do some additional checking when setting the dest buffer.
+ */
+static void i810EmitDestVerified( drm_device_t *dev, 
+				  volatile unsigned int *code ) 
+{	
+   	drm_i810_private_t *dev_priv = dev->dev_private;
+	unsigned int tmp;
+	RING_LOCALS;
+
+	BEGIN_LP_RING( I810_DEST_SETUP_SIZE + 2 );
+
+	tmp = code[I810_DESTREG_DI1];
+	if (tmp == dev_priv->front_di1 || tmp == dev_priv->back_di1) {
+		OUT_RING( CMD_OP_DESTBUFFER_INFO );
+		OUT_RING( tmp );
+	} else
+	   DRM_DEBUG("bad di1 %x (allow %x or %x)\n",
+		     tmp, dev_priv->front_di1, dev_priv->back_di1);
+
+	/* invarient:
+	 */
+	OUT_RING( CMD_OP_Z_BUFFER_INFO );
+	OUT_RING( dev_priv->zi1 );
+
+	OUT_RING( GFX_OP_DESTBUFFER_VARS );
+	OUT_RING( code[I810_DESTREG_DV1] );
+
+	OUT_RING( GFX_OP_DRAWRECT_INFO );
+	OUT_RING( code[I810_DESTREG_DR1] );
+	OUT_RING( code[I810_DESTREG_DR2] );
+	OUT_RING( code[I810_DESTREG_DR3] );
+	OUT_RING( code[I810_DESTREG_DR4] );
+	OUT_RING( 0 );
+
+	ADVANCE_LP_RING();
+}
+
+
+
+static void i810EmitState( drm_device_t *dev )
 {
-      	drm_i810_private_t *dev_priv = dev->dev_private;
-   	drm_i810_buf_priv_t *buf_priv = buf->dev_private;
-	unsigned long address = (unsigned long)buf->bus_address;
-	unsigned long start = address - dev->agp->base;
-   	RING_LOCALS;
+   	drm_i810_private_t *dev_priv = dev->dev_private;
+      	drm_i810_sarea_t *sarea_priv = dev_priv->sarea_priv;
+	unsigned int dirty = sarea_priv->dirty;
 
-   	dev_priv->counter++;
-   	DRM_DEBUG(  "dispatch counter : %ld\n", dev_priv->counter);
-   	DRM_DEBUG(  "i810_dma_dispatch\n");
-   	DRM_DEBUG(  "start : 0x%lx\n", start);
-	DRM_DEBUG(  "used : 0x%x\n", used);
-   	DRM_DEBUG(  "start + used - 4 : 0x%lx\n", start + used - 4);
-   	i810_kernel_lost_context(dev);
+	if (dirty & I810_UPLOAD_BUFFERS) {
+		i810EmitDestVerified( dev, sarea_priv->BufferState );
+		sarea_priv->dirty &= ~I810_UPLOAD_BUFFERS;
+	}
 
-   	BEGIN_LP_RING(10);
-   	OUT_RING( CMD_OP_BATCH_BUFFER );
-   	OUT_RING( start | BB1_PROTECTED );
-   	OUT_RING( start + used - 4 );
-      	OUT_RING( CMD_STORE_DWORD_IDX );
-   	OUT_RING( 20 );
-   	OUT_RING( dev_priv->counter );
-   	OUT_RING( CMD_STORE_DWORD_IDX );
-   	OUT_RING( buf_priv->my_use_idx );
-   	OUT_RING( I810_BUF_FREE );   
-      	OUT_RING( CMD_REPORT_HEAD );
-      	ADVANCE_LP_RING();
-	if(buf_priv->currently_mapped == I810_BUF_MAPPED) {
-		i810_unmap_buffer(buf);
+	if (dirty & I810_UPLOAD_CTX) {
+		i810EmitContextVerified( dev, sarea_priv->ContextState );
+		sarea_priv->dirty &= ~I810_UPLOAD_CTX;
+	}
+
+	if (dirty & I810_UPLOAD_TEX0) {
+		i810EmitTexVerified( dev, sarea_priv->TexState[0] );
+		sarea_priv->dirty &= ~I810_UPLOAD_TEX0;
+	}
+
+	if (dirty & I810_UPLOAD_TEX1) {
+		i810EmitTexVerified( dev, sarea_priv->TexState[1] );
+		sarea_priv->dirty &= ~I810_UPLOAD_TEX1;
+	}
+}
+
+
+
+/* need to verify 
+ */
+static void i810_dma_dispatch_clear( drm_device_t *dev, int flags, 
+				     unsigned int clear_color,
+				     unsigned int clear_zval )
+{
+   	drm_i810_private_t *dev_priv = dev->dev_private;
+      	drm_i810_sarea_t *sarea_priv = dev_priv->sarea_priv;
+	int nbox = sarea_priv->nbox;
+	drm_clip_rect_t *pbox = sarea_priv->boxes;
+	int pitch = dev_priv->pitch;
+	int cpp = 2;
+	int i;
+	RING_LOCALS;
+
+  	i810_kernel_lost_context(dev);
+
+      	if (nbox > I810_NR_SAREA_CLIPRECTS)
+     		nbox = I810_NR_SAREA_CLIPRECTS;
+
+	for (i = 0 ; i < nbox ; i++, pbox++) {
+		unsigned int x = pbox->x1;
+		unsigned int y = pbox->y1;
+		unsigned int width = (pbox->x2 - x) * cpp;
+		unsigned int height = pbox->y2 - y;
+		unsigned int start = y * pitch + x * cpp;
+
+		if (pbox->x1 > pbox->x2 ||
+		    pbox->y1 > pbox->y2 ||
+		    pbox->x2 > dev_priv->w ||
+		    pbox->y2 > dev_priv->h)
+			continue;
+
+	   	if ( flags & I810_FRONT ) {	    
+		   	DRM_DEBUG("clear front\n");
+			BEGIN_LP_RING( 6 );	    
+			OUT_RING( BR00_BITBLT_CLIENT | 
+				  BR00_OP_COLOR_BLT | 0x3 );
+			OUT_RING( BR13_SOLID_PATTERN | (0xF0 << 16) | pitch );
+			OUT_RING( (height << 16) | width );
+			OUT_RING( start );
+			OUT_RING( clear_color );
+			OUT_RING( 0 );
+			ADVANCE_LP_RING();
+		}
+
+		if ( flags & I810_BACK ) {
+			DRM_DEBUG("clear back\n");
+			BEGIN_LP_RING( 6 );	    
+			OUT_RING( BR00_BITBLT_CLIENT | 
+				  BR00_OP_COLOR_BLT | 0x3 );
+			OUT_RING( BR13_SOLID_PATTERN | (0xF0 << 16) | pitch );
+			OUT_RING( (height << 16) | width );
+			OUT_RING( dev_priv->back_offset + start );
+			OUT_RING( clear_color );
+			OUT_RING( 0 );
+			ADVANCE_LP_RING();
+		}
+
+		if ( flags & I810_DEPTH ) {
+			DRM_DEBUG("clear depth\n");
+			BEGIN_LP_RING( 6 );	    
+			OUT_RING( BR00_BITBLT_CLIENT | 
+				  BR00_OP_COLOR_BLT | 0x3 );
+			OUT_RING( BR13_SOLID_PATTERN | (0xF0 << 16) | pitch );
+			OUT_RING( (height << 16) | width );
+			OUT_RING( dev_priv->depth_offset + start );
+			OUT_RING( clear_zval );
+			OUT_RING( 0 );
+			ADVANCE_LP_RING();
+		}
+	}
+}
+
+static void i810_dma_dispatch_swap( drm_device_t *dev )
+{
+   	drm_i810_private_t *dev_priv = dev->dev_private;
+      	drm_i810_sarea_t *sarea_priv = dev_priv->sarea_priv;
+	int nbox = sarea_priv->nbox;
+	drm_clip_rect_t *pbox = sarea_priv->boxes;
+	int pitch = dev_priv->pitch;
+	int cpp = 2;
+	int ofs = dev_priv->back_offset;
+	int i;
+	RING_LOCALS;
+
+	DRM_DEBUG("swapbuffers\n");
+
+  	i810_kernel_lost_context(dev);
+
+      	if (nbox > I810_NR_SAREA_CLIPRECTS)
+     		nbox = I810_NR_SAREA_CLIPRECTS;
+
+	for (i = 0 ; i < nbox; i++, pbox++) 
+	{
+		unsigned int w = pbox->x2 - pbox->x1;
+		unsigned int h = pbox->y2 - pbox->y1;
+		unsigned int dst = pbox->x1*cpp + pbox->y1*pitch;
+		unsigned int start = ofs + dst;
+
+		if (pbox->x1 > pbox->x2 ||
+		    pbox->y1 > pbox->y2 ||
+		    pbox->x2 > dev_priv->w ||
+		    pbox->y2 > dev_priv->h)
+			continue;
+ 
+	   	DRM_DEBUG("dispatch swap %d,%d-%d,%d!\n",
+			  pbox[i].x1, pbox[i].y1,
+			  pbox[i].x2, pbox[i].y2);
+
+		BEGIN_LP_RING( 6 );
+		OUT_RING( BR00_BITBLT_CLIENT | BR00_OP_SRC_COPY_BLT | 0x4 );
+		OUT_RING( pitch | (0xCC << 16));
+		OUT_RING( (h << 16) | (w * cpp));
+		OUT_RING( dst );
+		OUT_RING( pitch );	
+		OUT_RING( start );
+		ADVANCE_LP_RING();
 	}
 }
 
+
 static void i810_dma_dispatch_vertex(drm_device_t *dev, 
 				     drm_buf_t *buf,
 				     int discard,
@@ -487,14 +748,26 @@ static void i810_dma_dispatch_vertex(drm_device_t *dev,
    	int nbox = sarea_priv->nbox;
 	unsigned long address = (unsigned long)buf->bus_address;
 	unsigned long start = address - dev->agp->base;     
-	int i = 0;
+	int i = 0, u;
    	RING_LOCALS;
 
-   
+   	i810_kernel_lost_context(dev);
+
    	if (nbox > I810_NR_SAREA_CLIPRECTS) 
 		nbox = I810_NR_SAREA_CLIPRECTS;
-   
-   	DRM_DEBUG("dispatch vertex addr 0x%lx, used 0x%x nbox %d\n", 
+
+	if (discard) {
+		u = cmpxchg(buf_priv->in_use, I810_BUF_CLIENT, 
+			    I810_BUF_HARDWARE);
+		if(u != I810_BUF_CLIENT) {
+			DRM_DEBUG("xxxx 2\n");
+		}
+	}
+
+	if (sarea_priv->dirty)
+	   i810EmitState( dev );
+
+  	DRM_DEBUG("dispatch vertex addr 0x%lx, used 0x%x nbox %d\n", 
 		  address, used, nbox);
 
    	dev_priv->counter++;
@@ -503,16 +776,24 @@ static void i810_dma_dispatch_vertex(drm_device_t *dev,
    	DRM_DEBUG(  "start : %lx\n", start);
 	DRM_DEBUG(  "used : %d\n", used);
    	DRM_DEBUG(  "start + used - 4 : %ld\n", start + used - 4);
-   	i810_kernel_lost_context(dev);
 
 	if (used) {
+		*(u32 *)buf_priv->virtual = (GFX_OP_PRIMITIVE |
+					     sarea_priv->vertex_prim |
+					     ((used/4)-2));
+
+		if (used & 4) {
+		   *(u32 *)((u32)buf_priv->virtual + used) = 0;
+		   used += 4;
+		}
+		   
 		do {
 			if (i < nbox) {
 				BEGIN_LP_RING(4);
 				OUT_RING( GFX_OP_SCISSOR | SC_UPDATE_SCISSOR | 
 					  SC_ENABLE );
 				OUT_RING( GFX_OP_SCISSOR_INFO );
-				OUT_RING( box[i].x1 | (box[i].y1 << 16) );
+				OUT_RING( box[i].x1 | (box[i].y1<<16) );
 				OUT_RING( (box[i].x2-1) | ((box[i].y2-1)<<16) );
 				ADVANCE_LP_RING();
 			}
@@ -543,6 +824,7 @@ static void i810_dma_dispatch_vertex(drm_device_t *dev,
       	OUT_RING( CMD_REPORT_HEAD );
 	OUT_RING( 0 );
    	ADVANCE_LP_RING();
+
 	if(buf_priv->currently_mapped == I810_BUF_MAPPED) {
 		i810_unmap_buffer(buf);
 	}
@@ -711,8 +993,8 @@ static inline void i810_dma_quiescent_emit(drm_device_t *dev)
    	RING_LOCALS;
 
   	i810_kernel_lost_context(dev);
-   	BEGIN_LP_RING(4);
 
+   	BEGIN_LP_RING(4);
    	OUT_RING( INST_PARSER_CLIENT | INST_OP_FLUSH | INST_FLUSH_MAP_CACHE );
    	OUT_RING( CMD_REPORT_HEAD );
    	OUT_RING( GFX_OP_USER_INTERRUPT );
@@ -794,23 +1076,17 @@ void i810_reclaim_buffers(drm_device_t *dev, pid_t pid)
 	int		 i;
 
 	if (!dma) return;
-      	if(dev->dev_private == NULL) return;
-	if(dma->buflist == NULL) return;
+      	if (!dev->dev_private) return;
+	if (!dma->buflist) return;
+
         i810_flush_queue(dev);
 
 	for (i = 0; i < dma->buf_count; i++) {
 	   	drm_buf_t *buf = dma->buflist[ i ];
 	   	drm_i810_buf_priv_t *buf_priv = buf->dev_private;
 	   
-		/* Only buffers that need to get reclaimed ever 
-		 * get set to free 
-		 */
 		if (buf->pid == pid && buf_priv) {
-		   	cmpxchg(buf_priv->in_use, 
-				I810_BUF_USED, I810_BUF_FREE);
-			if(buf_priv->currently_mapped == I810_BUF_MAPPED) {
-				buf_priv->currently_mapped = I810_BUF_UNMAPPED;
-			}
+		   *(buf_priv->in_use) = I810_BUF_FREE;
 		}
 	}
 }
@@ -900,90 +1176,80 @@ int i810_flush_ioctl(struct inode *inode, struct file *filp,
    	return 0;
 }
 
-static int i810DmaGeneral(drm_device_t *dev, drm_i810_general_t *args)
-{
-	drm_device_dma_t *dma = dev->dma;
-   	drm_buf_t *buf = dma->buflist[ args->idx ];
-   	
-	if (!args->used) {
-	   	i810_freelist_put(dev, buf);
-	} else {  
-		i810_dma_dispatch_general( dev, buf, args->used );
-		atomic_add(args->used, &dma->total_bytes);
-		atomic_inc(&dma->total_dmas);
-	}
-	return 0; 
-}
-
-static int i810DmaVertex(drm_device_t *dev, drm_i810_vertex_t *args)
-{
-	drm_device_dma_t *dma = dev->dma;
-	drm_buf_t *buf = dma->buflist[ args->idx ];
-   	i810_dma_dispatch_vertex( dev, buf, args->discard, args->used );
-   	atomic_add(args->used, &dma->total_bytes);
-	atomic_inc(&dma->total_dmas);
-   	return 0;
-}
 
-int i810_dma_general(struct inode *inode, struct file *filp,
-		     unsigned int cmd, unsigned long arg)
+int i810_dma_vertex(struct inode *inode, struct file *filp,
+	       unsigned int cmd, unsigned long arg)
 {
 	drm_file_t *priv = filp->private_data;
 	drm_device_t *dev = priv->dev;
-	drm_i810_general_t general;
-      	drm_i810_private_t *dev_priv = (drm_i810_private_t *)dev->dev_private;
+	drm_device_dma_t *dma = dev->dma;
+   	drm_i810_private_t *dev_priv = (drm_i810_private_t *)dev->dev_private;
       	u32 *hw_status = (u32 *)dev_priv->hw_status_page;
    	drm_i810_sarea_t *sarea_priv = (drm_i810_sarea_t *) 
      					dev_priv->sarea_priv; 
+	drm_i810_vertex_t vertex;
 
-	int retcode = 0;
-	
-	copy_from_user_ret(&general, (drm_i810_general_t *)arg, sizeof(general),
+	copy_from_user_ret(&vertex, (drm_i810_vertex_t *)arg, sizeof(vertex),
 			   -EFAULT);
 
-	DRM_DEBUG("i810 dma general idx %d used %d\n",
-		  general.idx, general.used);
-   
    	if(!_DRM_LOCK_IS_HELD(dev->lock.hw_lock->lock)) {
-		DRM_ERROR("i810_dma_general called without lock held\n");
+		DRM_ERROR("i810_dma_vertex called without lock held\n");
 		return -EINVAL;
 	}
 
-	retcode = i810DmaGeneral(dev, &general);
+	DRM_DEBUG("i810 dma vertex, idx %d used %d discard %d\n",
+		  vertex.idx, vertex.used, vertex.discard);
+
+	i810_dma_dispatch_vertex( dev, 
+				  dma->buflist[ vertex.idx ], 
+				  vertex.discard, vertex.used );
+
+   	atomic_add(vertex.used, &dma->total_bytes);
+	atomic_inc(&dma->total_dmas);
 	sarea_priv->last_enqueue = dev_priv->counter-1;
    	sarea_priv->last_dispatch = (int) hw_status[5];
    
-	return retcode;
+	return 0;
 }
 
-int i810_dma_vertex(struct inode *inode, struct file *filp,
-	       unsigned int cmd, unsigned long arg)
+
+
+int i810_clear_bufs(struct inode *inode, struct file *filp,
+		   unsigned int cmd, unsigned long arg)
 {
 	drm_file_t *priv = filp->private_data;
 	drm_device_t *dev = priv->dev;
-   	drm_i810_private_t *dev_priv = (drm_i810_private_t *)dev->dev_private;
-      	u32 *hw_status = (u32 *)dev_priv->hw_status_page;
-   	drm_i810_sarea_t *sarea_priv = (drm_i810_sarea_t *) 
-     					dev_priv->sarea_priv; 
-	drm_i810_vertex_t vertex;
-	int retcode = 0;
+	drm_i810_clear_t clear;
 
-	copy_from_user_ret(&vertex, (drm_i810_vertex_t *)arg, sizeof(vertex),
+   	copy_from_user_ret(&clear, (drm_i810_clear_t *)arg, sizeof(clear), 
 			   -EFAULT);
+   
    	if(!_DRM_LOCK_IS_HELD(dev->lock.hw_lock->lock)) {
-		DRM_ERROR("i810_dma_vertex called without lock held\n");
+		DRM_ERROR("i810_clear_bufs called without lock held\n");
 		return -EINVAL;
 	}
 
-	DRM_DEBUG("i810 dma vertex, idx %d used %d discard %d\n",
-		  vertex.idx, vertex.used, vertex.discard);
+	i810_dma_dispatch_clear( dev, clear.flags, 
+				 clear.clear_color, 
+				 clear.clear_depth );
+   	return 0;
+}
 
-	retcode = i810DmaVertex(dev, &vertex);
-	sarea_priv->last_enqueue = dev_priv->counter-1;
-   	sarea_priv->last_dispatch = (int) hw_status[5];
+int i810_swap_bufs(struct inode *inode, struct file *filp,
+		  unsigned int cmd, unsigned long arg)
+{
+	drm_file_t *priv = filp->private_data;
+	drm_device_t *dev = priv->dev;
    
-	return retcode;
+	DRM_DEBUG("i810_swap_bufs\n");
 
+   	if(!_DRM_LOCK_IS_HELD(dev->lock.hw_lock->lock)) {
+		DRM_ERROR("i810_swap_buf called without lock held\n");
+		return -EINVAL;
+	}
+
+	i810_dma_dispatch_swap( dev );
+   	return 0;
 }
 
 int i810_getage(struct inode *inode, struct file *filp, unsigned int cmd,
diff --git a/linux/i810_drm.h b/linux/i810_drm.h
index 7e435879..4c8e09f6 100644
--- a/linux/i810_drm.h
+++ b/linux/i810_drm.h
@@ -5,35 +5,112 @@
  * if you change them, you must change the defines in the Xserver.
  */
 
-/* Might one day want to support the client-side ringbuffer code again.
- */
 #ifndef _I810_DEFINES_
 #define _I810_DEFINES_
 
-#define I810_USE_BATCH			1
 #define I810_DMA_BUF_ORDER		12
 #define I810_DMA_BUF_SZ 		(1<<I810_DMA_BUF_ORDER)
 #define I810_DMA_BUF_NR 		256
-#define I810_NR_SAREA_CLIPRECTS 	2
+#define I810_NR_SAREA_CLIPRECTS 	8
 
 /* Each region is a minimum of 64k, and there are at most 64 of them.
  */
-
 #define I810_NR_TEX_REGIONS 64
 #define I810_LOG_MIN_TEX_REGION_SIZE 16
 #endif
 
+#define I810_UPLOAD_TEX0IMAGE  0x1 /* handled clientside */
+#define I810_UPLOAD_TEX1IMAGE  0x2 /* handled clientside */
+#define I810_UPLOAD_CTX        0x4
+#define I810_UPLOAD_BUFFERS    0x8
+#define I810_UPLOAD_TEX0       0x10
+#define I810_UPLOAD_TEX1       0x20
+#define I810_UPLOAD_CLIPRECTS  0x40
+
+
+/* Indices into buf.Setup where various bits of state are mirrored per
+ * context and per buffer.  These can be fired at the card as a unit,
+ * or in a piecewise fashion as required.
+ */
+
+/* Destbuffer state 
+ *    - backbuffer linear offset and pitch -- invarient in the current dri
+ *    - zbuffer linear offset and pitch -- also invarient
+ *    - drawing origin in back and depth buffers.
+ *
+ * Keep the depth/back buffer state here to acommodate private buffers
+ * in the future.
+ */
+#define I810_DESTREG_DI0  0	/* CMD_OP_DESTBUFFER_INFO (2 dwords) */
+#define I810_DESTREG_DI1  1
+#define I810_DESTREG_DV0  2	/* GFX_OP_DESTBUFFER_VARS (2 dwords) */
+#define I810_DESTREG_DV1  3
+#define I810_DESTREG_DR0  4	/* GFX_OP_DRAWRECT_INFO (4 dwords) */
+#define I810_DESTREG_DR1  5
+#define I810_DESTREG_DR2  6
+#define I810_DESTREG_DR3  7
+#define I810_DESTREG_DR4  8
+#define I810_DEST_SETUP_SIZE 10
+
+/* Context state
+ */
+#define I810_CTXREG_CF0   0	/* GFX_OP_COLOR_FACTOR */
+#define I810_CTXREG_CF1   1	
+#define I810_CTXREG_ST0   2     /* GFX_OP_STIPPLE */
+#define I810_CTXREG_ST1   3
+#define I810_CTXREG_VF    4	/* GFX_OP_VERTEX_FMT */
+#define I810_CTXREG_MT    5	/* GFX_OP_MAP_TEXELS */
+#define I810_CTXREG_MC0   6	/* GFX_OP_MAP_COLOR_STAGES - stage 0 */
+#define I810_CTXREG_MC1   7     /* GFX_OP_MAP_COLOR_STAGES - stage 1 */
+#define I810_CTXREG_MC2   8	/* GFX_OP_MAP_COLOR_STAGES - stage 2 */
+#define I810_CTXREG_MA0   9	/* GFX_OP_MAP_ALPHA_STAGES - stage 0 */
+#define I810_CTXREG_MA1   10	/* GFX_OP_MAP_ALPHA_STAGES - stage 1 */
+#define I810_CTXREG_MA2   11	/* GFX_OP_MAP_ALPHA_STAGES - stage 2 */
+#define I810_CTXREG_SDM   12	/* GFX_OP_SRC_DEST_MONO */
+#define I810_CTXREG_FOG   13	/* GFX_OP_FOG_COLOR */
+#define I810_CTXREG_B1    14	/* GFX_OP_BOOL_1 */
+#define I810_CTXREG_B2    15	/* GFX_OP_BOOL_2 */
+#define I810_CTXREG_LCS   16	/* GFX_OP_LINEWIDTH_CULL_SHADE_MODE */
+#define I810_CTXREG_PV    17	/* GFX_OP_PV_RULE -- Invarient! */
+#define I810_CTXREG_ZA    18	/* GFX_OP_ZBIAS_ALPHAFUNC */
+#define I810_CTXREG_AA    19	/* GFX_OP_ANTIALIAS */
+#define I810_CTX_SETUP_SIZE 20 
+
+/* Texture state (per tex unit)
+ */
+#define I810_TEXREG_MI0  0	/* GFX_OP_MAP_INFO (4 dwords) */
+#define I810_TEXREG_MI1  1	
+#define I810_TEXREG_MI2  2	
+#define I810_TEXREG_MI3  3	
+#define I810_TEXREG_MF   4	/* GFX_OP_MAP_FILTER */
+#define I810_TEXREG_MLC  5	/* GFX_OP_MAP_LOD_CTL */
+#define I810_TEXREG_MLL  6	/* GFX_OP_MAP_LOD_LIMITS */
+#define I810_TEXREG_MCS  7	/* GFX_OP_MAP_COORD_SETS ??? */
+#define I810_TEX_SETUP_SIZE 8
+
+#define I810_FRONT   0x1
+#define I810_BACK    0x2
+#define I810_DEPTH   0x4
+
+
 typedef struct _drm_i810_init {
-   	enum {
-	   	I810_INIT_DMA = 0x01,
-	   	I810_CLEANUP_DMA = 0x02
+	enum {
+		I810_INIT_DMA = 0x01,
+		I810_CLEANUP_DMA = 0x02
 	} func;
-   	int ring_map_idx;
-      	int buffer_map_idx;
+	int ring_map_idx;
+	int buffer_map_idx;
 	int sarea_priv_offset;
-   	unsigned long ring_start;
-   	unsigned long ring_end;
-   	unsigned long ring_size;
+	unsigned int ring_start;
+	unsigned int ring_end;
+	unsigned int ring_size;
+	unsigned int front_offset;
+	unsigned int back_offset;
+	unsigned int depth_offset;
+	unsigned int w;
+	unsigned int h;
+	unsigned int pitch;
+	unsigned int pitch_bits; 
 } drm_i810_init_t;
 
 /* Warning: If you change the SAREA structure you must change the Xserver
@@ -46,6 +123,11 @@ typedef struct _drm_i810_tex_region {
 } drm_i810_tex_region_t;
 
 typedef struct _drm_i810_sarea {
+   	unsigned int ContextState[I810_CTX_SETUP_SIZE];
+   	unsigned int BufferState[I810_DEST_SETUP_SIZE];
+   	unsigned int TexState[2][I810_TEX_SETUP_SIZE];
+   	unsigned int dirty;
+
 	unsigned int nbox;
 	drm_clip_rect_t boxes[I810_NR_SAREA_CLIPRECTS];
 
@@ -72,12 +154,18 @@ typedef struct _drm_i810_sarea {
 	int last_dispatch;	/* age of the most recently dispatched buffer */
 	int last_quiescent;     /*  */
 	int ctxOwner;		/* last context to upload state */
+
+	int vertex_prim;
+
 } drm_i810_sarea_t;
 
-typedef struct _drm_i810_general {
-   	int idx;
-	int used;
-} drm_i810_general_t;
+typedef struct _drm_i810_clear {
+	int clear_color;
+	int clear_depth;
+	int flags;
+} drm_i810_clear_t;
+
+
 
 /* These may be placeholders if we have more cliprects than
  * I810_NR_SAREA_CLIPRECTS.  In that case, the client sets discard to
diff --git a/linux/i810_drv.c b/linux/i810_drv.c
index 03585bdf..75b402da 100644
--- a/linux/i810_drv.c
+++ b/linux/i810_drv.c
@@ -107,10 +107,11 @@ static drm_ioctl_desc_t	      i810_ioctls[] = {
 
    	[DRM_IOCTL_NR(DRM_IOCTL_I810_INIT)]   = { i810_dma_init,   1, 1 },
    	[DRM_IOCTL_NR(DRM_IOCTL_I810_VERTEX)] = { i810_dma_vertex, 1, 0 },
-   	[DRM_IOCTL_NR(DRM_IOCTL_I810_DMA)]    = { i810_dma_general,1, 0 },
+   	[DRM_IOCTL_NR(DRM_IOCTL_I810_CLEAR)]  = { i810_clear_bufs, 1, 0 },
       	[DRM_IOCTL_NR(DRM_IOCTL_I810_FLUSH)]  = { i810_flush_ioctl,1, 0 },
    	[DRM_IOCTL_NR(DRM_IOCTL_I810_GETAGE)] = { i810_getage,     1, 0 },
 	[DRM_IOCTL_NR(DRM_IOCTL_I810_GETBUF)] = { i810_getbuf,     1, 0 },
+   	[DRM_IOCTL_NR(DRM_IOCTL_I810_SWAP)]   = { i810_swap_bufs,  1, 0 },
 };
 
 #define I810_IOCTL_COUNT DRM_ARRAY_SIZE(i810_ioctls)
diff --git a/linux/i810_drv.h b/linux/i810_drv.h
index e775d1b6..671489d3 100644
--- a/linux/i810_drv.h
+++ b/linux/i810_drv.h
@@ -55,6 +55,15 @@ typedef struct drm_i810_private {
 
    	atomic_t flush_done;
    	wait_queue_head_t flush_queue;	/* Processes waiting until flush    */
+
+
+	
+	u32 front_di1, back_di1, zi1;
+	
+	int back_offset;
+	int depth_offset;
+	int w, h;
+	int pitch;
 } drm_i810_private_t;
 
 				/* i810_drv.c */
@@ -128,11 +137,9 @@ typedef struct drm_i810_buf_priv {
    	int my_use_idx;
 	int currently_mapped;
 	void *virtual;
+	void *kernel_virtual;
 } drm_i810_buf_priv_t;
 
-#define I810_DMA_GENERAL 0
-#define I810_DMA_VERTEX  1
-#define I810_DMA_DISCARD 2	/* not used */
 
 #define I810_VERBOSE 0
 
@@ -140,9 +147,11 @@ typedef struct drm_i810_buf_priv {
 int i810_dma_vertex(struct inode *inode, struct file *filp,
 		    unsigned int cmd, unsigned long arg);
 
-int i810_dma_general(struct inode *inode, struct file *filp,
-		     unsigned int cmd, unsigned long arg);
+int i810_swap_bufs(struct inode *inode, struct file *filp,
+		   unsigned int cmd, unsigned long arg);
 
+int i810_clear_bufs(struct inode *inode, struct file *filp,
+		    unsigned int cmd, unsigned long arg);
 
 #define GFX_OP_USER_INTERRUPT 		((0<<29)|(2<<23))
 #define GFX_OP_BREAKPOINT_INTERRUPT	((0<<29)|(1<<23))
@@ -196,5 +205,22 @@ int i810_dma_general(struct inode *inode, struct file *filp,
 #define SCI_YMAX_MASK      (0xffff<<16)
 #define SCI_XMAX_MASK      (0xffff<<0)
 
+#define GFX_OP_COLOR_FACTOR      ((0x3<<29)|(0x1d<<24)|(0x1<<16)|0x0)
+#define GFX_OP_STIPPLE           ((0x3<<29)|(0x1d<<24)|(0x83<<16))
+#define GFX_OP_MAP_INFO          ((0x3<<29)|(0x1d<<24)|0x2)
+#define GFX_OP_DESTBUFFER_VARS   ((0x3<<29)|(0x1d<<24)|(0x85<<16)|0x0)
+#define GFX_OP_DRAWRECT_INFO     ((0x3<<29)|(0x1d<<24)|(0x80<<16)|(0x3))
+#define GFX_OP_PRIMITIVE         ((0x3<<29)|(0x1f<<24))
+
+#define CMD_OP_Z_BUFFER_INFO     ((0x0<<29)|(0x16<<23))
+#define CMD_OP_DESTBUFFER_INFO   ((0x0<<29)|(0x15<<23))
+
+#define BR00_BITBLT_CLIENT   0x40000000
+#define BR00_OP_COLOR_BLT    0x10000000
+#define BR00_OP_SRC_COPY_BLT 0x10C00000
+#define BR13_SOLID_PATTERN   0x80000000
+
+
+
 #endif
 
diff --git a/linux/mga_drm.h b/linux/mga_drm.h
index 12a858e7..eb8c434a 100644
--- a/linux/mga_drm.h
+++ b/linux/mga_drm.h
@@ -37,6 +37,7 @@
  */
 #ifndef _MGA_DEFINES_
 #define _MGA_DEFINES_
+
 #define MGA_F  0x1		/* fog */
 #define MGA_A  0x2		/* alpha */
 #define MGA_S  0x4		/* specular */
@@ -61,11 +62,11 @@
 
 #define MGA_MAX_G400_PIPES 16
 #define MGA_MAX_G200_PIPES  8	/* no multitex */
-
 #define MGA_MAX_WARP_PIPES MGA_MAX_G400_PIPES
 
 #define MGA_CARD_TYPE_G200 1
 #define MGA_CARD_TYPE_G400 2
+
 #define MGA_FRONT   0x1
 #define MGA_BACK    0x2
 #define MGA_DEPTH   0x4
@@ -110,8 +111,8 @@
 #define MGA_UPLOAD_TEX0       0x2
 #define MGA_UPLOAD_TEX1       0x4
 #define MGA_UPLOAD_PIPE       0x8
-#define MGA_UPLOAD_TEX0IMAGE  0x10
-#define MGA_UPLOAD_TEX1IMAGE  0x20
+#define MGA_UPLOAD_TEX0IMAGE  0x10 /* handled client-side */
+#define MGA_UPLOAD_TEX1IMAGE  0x20 /* handled client-side */
 #define MGA_UPLOAD_2D 	      0x40
 #define MGA_WAIT_AGE          0x80 /* handled client-side */
 #define MGA_UPLOAD_CLIPRECTS  0x100 /* handled client-side */
diff --git a/shared-core/drm.h b/shared-core/drm.h
index 86ab0410..d196d912 100644
--- a/shared-core/drm.h
+++ b/shared-core/drm.h
@@ -345,8 +345,9 @@ typedef struct drm_agp_info {
 /* I810 specific ioctls */
 #define DRM_IOCTL_I810_INIT    DRM_IOW( 0x40, drm_i810_init_t)
 #define DRM_IOCTL_I810_VERTEX  DRM_IOW( 0x41, drm_i810_vertex_t)
-#define DRM_IOCTL_I810_DMA     DRM_IOW( 0x42, drm_i810_general_t)
+#define DRM_IOCTL_I810_CLEAR   DRM_IOW( 0x42, drm_i810_clear_t)
 #define DRM_IOCTL_I810_FLUSH   DRM_IO ( 0x43)
 #define DRM_IOCTL_I810_GETAGE  DRM_IO ( 0x44)
 #define DRM_IOCTL_I810_GETBUF  DRM_IOW( 0x45, drm_i810_dma_t)
+#define DRM_IOCTL_I810_SWAP    DRM_IO ( 0x46)
 #endif
diff --git a/shared/drm.h b/shared/drm.h
index 86ab0410..d196d912 100644
--- a/shared/drm.h
+++ b/shared/drm.h
@@ -345,8 +345,9 @@ typedef struct drm_agp_info {
 /* I810 specific ioctls */
 #define DRM_IOCTL_I810_INIT    DRM_IOW( 0x40, drm_i810_init_t)
 #define DRM_IOCTL_I810_VERTEX  DRM_IOW( 0x41, drm_i810_vertex_t)
-#define DRM_IOCTL_I810_DMA     DRM_IOW( 0x42, drm_i810_general_t)
+#define DRM_IOCTL_I810_CLEAR   DRM_IOW( 0x42, drm_i810_clear_t)
 #define DRM_IOCTL_I810_FLUSH   DRM_IO ( 0x43)
 #define DRM_IOCTL_I810_GETAGE  DRM_IO ( 0x44)
 #define DRM_IOCTL_I810_GETBUF  DRM_IOW( 0x45, drm_i810_dma_t)
+#define DRM_IOCTL_I810_SWAP    DRM_IO ( 0x46)
 #endif