Import -f XFree86 4.0.99.2

14 files changed, 266 insertions, 154 deletions

diff --git a/bsd/Makefile b/bsd/Makefile
index ff26c762..61cba175 100644
--- a/bsd/Makefile
+++ b/bsd/Makefile
@@ -1,5 +1,5 @@
 # $FreeBSD$
 
-SUBDIR = drm tdfx gamma
+SUBDIR = drm tdfx mga gamma
 
 .include <bsd.subdir.mk>
diff --git a/bsd/Makefile.bsd b/bsd/Makefile.bsd
index ff26c762..61cba175 100644
--- a/bsd/Makefile.bsd
+++ b/bsd/Makefile.bsd
@@ -1,5 +1,5 @@
 # $FreeBSD$
 
-SUBDIR = drm tdfx gamma
+SUBDIR = drm tdfx mga gamma
 
 .include <bsd.subdir.mk>
diff --git a/bsd/drm/auth.c b/bsd/drm/auth.c
index a2699011..37d17be5 100644
--- a/bsd/drm/auth.c
+++ b/bsd/drm/auth.c
@@ -135,12 +135,12 @@ int drm_getmagic(dev_t kdev, u_long cmd, caddr_t data,
 	if (priv->magic) {
 		auth.magic = priv->magic;
 	} else {
-		simple_lock(&lock);
 		do {
+			simple_lock(&lock);
 			if (!sequence) ++sequence; /* reserve 0 */
 			auth.magic = sequence++;
-		} while (drm_find_file(dev, auth.magic));
 		simple_unlock(&lock);
+		} while (drm_find_file(dev, auth.magic));
 		priv->magic = auth.magic;
 		drm_add_magic(dev, priv, auth.magic);
 	}
diff --git a/bsd/drm/lists.c b/bsd/drm/lists.c
index 61becaa1..b9722973 100644
--- a/bsd/drm/lists.c
+++ b/bsd/drm/lists.c
@@ -121,6 +121,7 @@ int drm_freelist_create(drm_freelist_t *bl, int count)
 	bl->low_mark  = 0;
 	bl->high_mark = 0;
 	atomic_set(&bl->wfh,   0);
+/*	bl->lock	= SPIN_LOCK_UNLOCKED; */
 	++bl->initialized;
 	return 0;
 }
@@ -159,6 +160,7 @@ int drm_freelist_put(drm_device_t *dev, drm_freelist_t *bl, drm_buf_t *buf)
 	drm_histogram_compute(dev, buf);
 #endif
 	buf->list	= DRM_LIST_FREE;
+/*
 	do {
 		old	  = (unsigned long)bl->next;
 		buf->next = (void *)old;
@@ -169,6 +171,13 @@ int drm_freelist_put(drm_device_t *dev, drm_freelist_t *bl, drm_buf_t *buf)
 			return 1;
 		}
 	} while (failed);
+*/
+
+	simple_lock(&bl->lock);
+	buf->next	= bl->next;
+	bl->next	= buf;
+	simple_unlock(&bl->lock);
+
 	atomic_inc(&bl->count);
 	if (atomic_read(&bl->count) > dma->buf_count) {
 		DRM_ERROR("%d of %d buffers free after addition of %d\n",
@@ -195,6 +204,7 @@ static drm_buf_t *drm_freelist_try(drm_freelist_t *bl)
 	if (!bl) return NULL;
 	
 				/* Get buffer */
+/*
 	do {
 		old = (unsigned int)bl->next;
 		if (!old) {
@@ -210,6 +220,16 @@ static drm_buf_t *drm_freelist_try(drm_freelist_t *bl)
 	atomic_dec(&bl->count);
 	
 	buf	  = (drm_buf_t *)old;
+*/
+	simple_lock(&bl->lock);
+	if(!bl->next){
+		simple_unlock(&bl->lock);
+		return NULL;
+	}
+	buf	= bl->next;
+	bl->next = bl->next->next;
+	simple_unlock(&bl->lock);
+	atomic_dec(&bl->count);
 	buf->next = NULL;
 	buf->list = DRM_LIST_NONE;
 	DRM_DEBUG("%d, count = %d, wfh = %d, w%d, p%d\n",
diff --git a/bsd/drm/lock.c b/bsd/drm/lock.c
index 1affcd1d..ec295b2b 100644
--- a/bsd/drm/lock.c
+++ b/bsd/drm/lock.c
@@ -135,9 +135,12 @@ static int drm_flush_queue(drm_device_t *dev, int context)
 	if (atomic_read(&q->use_count) > 1) {
 		atomic_inc(&q->block_write);
 		atomic_inc(&q->block_count);
-		error = tsleep(&q->flush_queue, PCATCH|PZERO, "drmfq", 0);
-		if (error)
-			return error;
+		for (;;) {
+			if (!DRM_BUFCOUNT(&q->waitlist)) break;
+			error = tsleep(&q->flush_queue, PCATCH|PZERO, "drmfq", 0);
+			if (error)
+				return error;
+		}
 		atomic_dec(&q->block_count);
 	}
 	atomic_dec(&q->use_count);
diff --git a/bsd/drm/memory.c b/bsd/drm/memory.c
index 81c1b1a0..f60e2128 100644
--- a/bsd/drm/memory.c
+++ b/bsd/drm/memory.c
@@ -95,7 +95,7 @@ void drm_mem_init(void)
 
 /* drm_mem_info is called whenever a process reads /dev/drm/mem. */
 
-static int _drm_mem_info SYSCTL_HANDLER_ARGS
+static int _drm_mem_info DRM_SYSCTL_HANDLER_ARGS
 {
 	drm_mem_stats_t *pt;
 	char buf[128];
@@ -127,7 +127,7 @@ static int _drm_mem_info SYSCTL_HANDLER_ARGS
 	return 0;
 }
 
-int drm_mem_info SYSCTL_HANDLER_ARGS
+int drm_mem_info DRM_SYSCTL_HANDLER_ARGS
 {
 	int ret;
 	
diff --git a/bsd/gamma/Makefile b/bsd/gamma/Makefile
index f5635910..37f3f5e7 100644
--- a/bsd/gamma/Makefile
+++ b/bsd/gamma/Makefile
@@ -3,7 +3,7 @@
 KMOD	= gamma
 SRCS    = gamma_drv.c gamma_dma.c
 SRCS   += device_if.h bus_if.h pci_if.h
-CFLAGS += ${DEBUG_FLAGS} -I..
+CFLAGS += ${DEBUG_FLAGS} -I. -I..
 KMODDEPS = drm
 
 @:
diff --git a/bsd/mga/mga_context.c b/bsd/mga/mga_context.c
index 63f1b42b..63515bab 100644
--- a/bsd/mga/mga_context.c
+++ b/bsd/mga/mga_context.c
@@ -188,7 +188,11 @@ int mga_rmctx(dev_t kdev, u_long cmd, caddr_t data, int flags, struct proc *p)
 
 	ctx = *(drm_ctx_t *) data;
 	DRM_DEBUG("%d\n", ctx.handle);
-      	if(ctx.handle != DRM_KERNEL_CONTEXT) {
+/*
+	if(ctx.handle == DRM_KERNEL_CONTEXT+1)
+		priv->remove_auth_on_close = 1;
+*/
+      	if(ctx.handle != DRM_KERNEL_CONTEXT ) {
 	   	drm_ctxbitmap_free(dev, ctx.handle);
 	}
 	
diff --git a/bsd/mga/mga_dma.c b/bsd/mga/mga_dma.c
index 7055d28d..85c29df7 100644
--- a/bsd/mga/mga_dma.c
+++ b/bsd/mga/mga_dma.c
@@ -130,7 +130,6 @@ static int mga_freelist_init(drm_device_t *dev)
 	   	item->buf = buf;
 	   	buf_priv->my_freelist = item;
 		buf_priv->discard = 0;
-		buf_priv->dispatched = 0;
 	   	dev_priv->head->next = item;
 	}
    
@@ -176,7 +175,7 @@ static __inline void mga_dma_quiescent(drm_device_t *dev)
 				  atomic_read(&dev->total_irq), 
 				  atomic_read(&dma->total_lost));
 			DRM_ERROR("lockup\n"); 
-			goto out_nolock;
+			return;
 		}
 		for (i = 0 ; i < 2000 ; i++) mga_delay();
 	}
@@ -188,15 +187,15 @@ static __inline void mga_dma_quiescent(drm_device_t *dev)
 				  atomic_read(&dev->total_irq), 
 				  atomic_read(&dma->total_lost));
 			DRM_ERROR("lockup\n"); 
-			goto out_status;
+			clear_bit(MGA_IN_DISPATCH, &dev_priv->dispatch_status);
+			return;
 		}
 		for (i = 0 ; i < 2000 ; i++) mga_delay();	  
 	}
     	sarea_priv->dirty |= MGA_DMA_FLUSH;
 
-out_status:
     	clear_bit(MGA_IN_DISPATCH, &dev_priv->dispatch_status);
-out_nolock:
+	DRM_DEBUG("exit, dispatch_status = 0x%02x\n",dev_priv->dispatch_status);
 }
 
 static void mga_reset_freelist(drm_device_t *dev)
@@ -236,20 +235,22 @@ drm_buf_t *mga_freelist_get(drm_device_t *dev)
 	   	set_bit(MGA_IN_GETBUF, &dev_priv->dispatch_status);
 	   	for (;;) {
 		   	mga_dma_schedule(dev, 0);
-		   	if(!test_bit(MGA_IN_GETBUF, 
-				     &dev_priv->dispatch_status)) 
+/*		   	if(!test_bit(MGA_IN_GETBUF, 
+				     &dev_priv->dispatch_status)) */
+			if(dev_priv->tail->age < dev_priv->last_prim_age)
 				break;
 		   	atomic_inc(&dev->total_sleeps);
 			ret = tsleep(&dev_priv->buf_queue, PZERO|PCATCH,
 				     "mgafg", 0);
-			if (ret) {
+			if (ret == EINTR) {
 				clear_bit(MGA_IN_GETBUF,
 					  &dev_priv->dispatch_status);
-				splx(s);
-			   	goto failed_getbuf;
+				break;
 			}
 		}
 		splx(s);
+		clear_bit(MGA_IN_GETBUF, &dev_priv->dispatch_status);
+		if (ret) return NULL;
 	}
    
    	if(dev_priv->tail->age < dev_priv->last_prim_age) {
@@ -263,7 +264,6 @@ drm_buf_t *mga_freelist_get(drm_device_t *dev)
 	   	return next->buf;
 	}
 
-failed_getbuf:
 	failed++;
    	return NULL;
 }
@@ -520,61 +520,55 @@ int mga_advance_primary(drm_device_t *dev)
 static __inline int mga_decide_to_fire(drm_device_t *dev)
 {
    	drm_mga_private_t *dev_priv = (drm_mga_private_t *)dev->dev_private;
-      	drm_device_dma_t  *dma	    = dev->dma;
 
    	DRM_DEBUG("%s\n", __FUNCTION__);
 
    	if(test_bit(MGA_BUF_FORCE_FIRE, &dev_priv->next_prim->buffer_status)) {
-	   	atomic_inc(&dma->total_prio);
 	   	return 1;
 	}
 
 	if (test_bit(MGA_IN_GETBUF, &dev_priv->dispatch_status) &&
 	    dev_priv->next_prim->num_dwords) {
-	   	atomic_inc(&dma->total_prio);
 	   	return 1;
 	}
 
 	if (test_bit(MGA_IN_FLUSH, &dev_priv->dispatch_status) &&
 	    dev_priv->next_prim->num_dwords) {
-	   	atomic_inc(&dma->total_prio);
 	   	return 1;
 	}
    
    	if(atomic_read(&dev_priv->pending_bufs) <= MGA_NUM_PRIM_BUFS - 1) {
 		if(test_bit(MGA_BUF_SWAP_PENDING, 
 			    &dev_priv->next_prim->buffer_status)) {
-			atomic_inc(&dma->total_dmas);
 			return 1;
 		}
 	}
 
    	if(atomic_read(&dev_priv->pending_bufs) <= MGA_NUM_PRIM_BUFS / 2) {
 		if(dev_priv->next_prim->sec_used >= MGA_DMA_BUF_NR / 8) {
-			atomic_inc(&dma->total_hit);
 			return 1;
 		}
 	}
 
    	if(atomic_read(&dev_priv->pending_bufs) >= MGA_NUM_PRIM_BUFS / 2) {
 		if(dev_priv->next_prim->sec_used >= MGA_DMA_BUF_NR / 4) {
-			atomic_inc(&dma->total_missed_free);
 			return 1;
 		}
 	}
 
-   	atomic_inc(&dma->total_tried);
    	return 0;
 }
 
 int mga_dma_schedule(drm_device_t *dev, int locked)
 {
       	drm_mga_private_t *dev_priv = (drm_mga_private_t *)dev->dev_private;
-      	drm_device_dma_t  *dma	    = dev->dma;
+	int	retval =0 ;
+
+	if (!dev_priv) return EBUSY;
 
    	if (test_and_set_bit(0, &dev->dma_flag)) {
-		atomic_inc(&dma->total_missed_dma);
-		return EBUSY;
+		retval = EBUSY;
+		goto sch_out_wakeup;
 	}
    
 	DRM_DEBUG("%s\n", __FUNCTION__);
@@ -591,17 +585,15 @@ int mga_dma_schedule(drm_device_t *dev, int locked)
    
    	if (!locked && 
 	    !drm_lock_take(&dev->lock.hw_lock->lock, DRM_KERNEL_CONTEXT)) {
-	   	atomic_inc(&dma->total_missed_lock);
 	   	clear_bit(0, &dev->dma_flag);
 		DRM_DEBUG("Not locked\n");
-	   	return EBUSY;
+		retval = EBUSY;
+		goto sch_out_wakeup;
 	}
-   	DRM_DEBUG("I'm locked\n");
 
    	if(!test_and_set_bit(MGA_IN_DISPATCH, &dev_priv->dispatch_status)) {
 	   	/* Fire dma buffer */
 	   	if(mga_decide_to_fire(dev)) {
-		   	DRM_DEBUG("idx :%d\n", dev_priv->next_prim->idx);
 			clear_bit(MGA_BUF_FORCE_FIRE, 
 				  &dev_priv->next_prim->buffer_status);
 		   	if(dev_priv->current_prim == dev_priv->next_prim) {
@@ -613,8 +605,6 @@ int mga_dma_schedule(drm_device_t *dev, int locked)
 		} else {
 			clear_bit(MGA_IN_DISPATCH, &dev_priv->dispatch_status);
 		}
-	} else {
-		DRM_DEBUG("I can't get the dispatch lock\n");
 	}
    	
 	if (!locked) {
@@ -623,9 +613,9 @@ int mga_dma_schedule(drm_device_t *dev, int locked)
 			DRM_ERROR("\n");
 		}
 	}
-
+	clear_bit(0, &dev->dma_flag);
+sch_out_wakeup:
       	if(test_bit(MGA_IN_FLUSH, &dev_priv->dispatch_status) &&
-	   dev_priv->next_prim->num_dwords == 0 &&
 	   atomic_read(&dev_priv->pending_bufs) == 0) {
 	   	/* Everything has been processed by the hardware */
 		clear_bit(MGA_IN_FLUSH, &dev_priv->dispatch_status);
@@ -633,18 +623,10 @@ int mga_dma_schedule(drm_device_t *dev, int locked)
 	}
 
 	if(test_bit(MGA_IN_GETBUF, &dev_priv->dispatch_status) &&
-	   dev_priv->tail->age < dev_priv->last_prim_age) {
-		clear_bit(MGA_IN_GETBUF, &dev_priv->dispatch_status);
-		DRM_DEBUG("Waking up buf queue\n");
+	   dev_priv->tail->age < dev_priv->last_prim_age) 
 		wakeup(&dev_priv->buf_queue);
-	} else if (test_bit(MGA_IN_GETBUF, &dev_priv->dispatch_status)) {
-	   	DRM_DEBUG("Not waking buf_queue on %d %d\n", 
-			  atomic_read(&dev->total_irq), 
-			  dev_priv->last_prim_age);
-	}
 
-   	clear_bit(0, &dev->dma_flag);
-	return 0;
+	return retval;
 }
 
 static void mga_dma_service(void *arg)
@@ -653,7 +635,6 @@ static void mga_dma_service(void *arg)
     	drm_mga_private_t *dev_priv = (drm_mga_private_t *)dev->dev_private;
     	drm_mga_prim_buf_t *last_prim_buffer;
 
-	DRM_DEBUG("%s\n", __FUNCTION__);
     	atomic_inc(&dev->total_irq);
 	if((MGA_READ(MGAREG_STATUS) & 0x00000001) != 0x00000001) return;
       	MGA_WRITE(MGAREG_ICLEAR, 0x00000001);
@@ -663,11 +644,11 @@ static void mga_dma_service(void *arg)
 	dev_priv->sarea_priv->last_dispatch = 
 		dev_priv->last_prim_age = last_prim_buffer->prim_age;
       	clear_bit(MGA_BUF_IN_USE, &last_prim_buffer->buffer_status);
-   	wakeup(&dev_priv->wait_queue);
       	clear_bit(MGA_BUF_SWAP_PENDING, &last_prim_buffer->buffer_status);
       	clear_bit(MGA_IN_DISPATCH, &dev_priv->dispatch_status);
       	atomic_dec(&dev_priv->pending_bufs);
 	taskqueue_enqueue(taskqueue_swi, &dev->task);
+   	wakeup(&dev_priv->wait_queue);
 }
 
 static void mga_dma_task_queue(void *device, int pending)
@@ -684,6 +665,8 @@ int mga_dma_cleanup(drm_device_t *dev)
 		drm_mga_private_t *dev_priv = 
 			(drm_mga_private_t *) dev->dev_private;
       
+		if (dev->irq) mga_flush_queue(dev);
+		mga_dma_quiescent(dev);
 		if(dev_priv->ioremap) {
 			int temp = (dev_priv->warp_ucode_size + 
 				    dev_priv->primary_size + 
@@ -723,9 +706,6 @@ int mga_dma_cleanup(drm_device_t *dev)
 static int mga_dma_initialize(drm_device_t *dev, drm_mga_init_t *init) {
 	drm_mga_private_t *dev_priv;
 	drm_map_t *sarea_map = NULL;
-	int i;
-
-	DRM_DEBUG("%s\n", __FUNCTION__);
 
 	dev_priv = drm_alloc(sizeof(drm_mga_private_t), DRM_MEM_DRIVER);
 	if(dev_priv == NULL) return ENOMEM;
@@ -760,8 +740,8 @@ static int mga_dma_initialize(drm_device_t *dev, drm_mga_init_t *init) {
 
 	dev_priv->mAccess = init->mAccess;
    	dev_priv->flush_queue = 0;
-	dev_priv->buf_queue = 0;
-	dev_priv->WarpPipe = -1;
+	dev_priv->WarpPipe = 0xff000000;
+	dev_priv->vertexsize = 0;
 
    	DRM_DEBUG("chipset: %d ucode_size: %d backOffset: %x depthOffset: %x\n",
 		  dev_priv->chipset, dev_priv->warp_ucode_size, 
@@ -773,13 +753,6 @@ static int mga_dma_initialize(drm_device_t *dev, drm_mga_init_t *init) {
 	memcpy(&dev_priv->WarpIndex, &init->WarpIndex, 
 	       sizeof(drm_mga_warp_index_t) * MGA_MAX_WARP_PIPES);
 
-   	for (i = 0 ; i < MGA_MAX_WARP_PIPES ; i++) 
-		DRM_DEBUG("warp pipe %d: installed: %d phys: %lx size: %x\n",
-			  i, 
-			  dev_priv->WarpIndex[i].installed,
-			  dev_priv->WarpIndex[i].phys_addr,
-			  dev_priv->WarpIndex[i].size);
-
    	if(mga_init_primary_bufs(dev, init) != 0) {
 		DRM_ERROR("Can not initialize primary buffers\n");
 		mga_dma_cleanup(dev);
@@ -951,9 +924,7 @@ static int mga_flush_queue(drm_device_t *dev)
 
    	DRM_DEBUG("%s\n", __FUNCTION__);
 
-   	if(dev_priv == NULL) {
-	   	return 0;
-	}
+   	if(!dev_priv) return 0;
    
    	if(dev_priv->next_prim->num_dwords != 0) {
 		s = splsofttq();
@@ -1065,7 +1036,7 @@ int mga_lock(dev_t kdev, u_long cmd, caddr_t data,
 		}
 	}
    
-	DRM_DEBUG("%d %s\n", lock.context, ret ? "interrupted" : "has lock");
+	if (ret) DRM_DEBUG("%d %s\n", lock.context, ret ? "interrupted" : "has lock");
 	return ret;
 }
 		
@@ -1086,16 +1057,16 @@ int mga_flush_ioctl(dev_t kdev, u_long cmd, caddr_t data,
 	}
 
    	if(lock.flags & _DRM_LOCK_FLUSH || lock.flags & _DRM_LOCK_FLUSH_ALL) {
-		drm_mga_prim_buf_t *temp_buf =
-			dev_priv->prim_bufs[dev_priv->current_prim_idx];
+		drm_mga_prim_buf_t *temp_buf; 
+		temp_buf = dev_priv->current_prim;
 
-		if(temp_buf && temp_buf->num_dwords) {
 			s = splsofttq();
+		if(temp_buf && temp_buf->num_dwords) {
 			set_bit(MGA_BUF_FORCE_FIRE, &temp_buf->buffer_status);
 			mga_advance_primary(dev);
+ 		}
 			mga_dma_schedule(dev, 1);
 			splx(s);
- 		}
 	}
    	if(lock.flags & _DRM_LOCK_QUIESCENT) {
 		mga_flush_queue(dev);
diff --git a/bsd/mga/mga_drv.c b/bsd/mga/mga_drv.c
index 09937201..77d7c440 100644
--- a/bsd/mga/mga_drv.c
+++ b/bsd/mga/mga_drv.c
@@ -40,8 +40,8 @@ MODULE_DEPEND(mga, agp, 1, 1, 1);
 
 #define MGA_NAME	 "mga"
 #define MGA_DESC	 "Matrox g200/g400"
-#define MGA_DATE	 "19991213"
-#define MGA_MAJOR	 1
+#define MGA_DATE	 "20000928"
+#define MGA_MAJOR	 2
 #define MGA_MINOR	 0
 #define MGA_PATCHLEVEL	 0
 
diff --git a/bsd/mga/mga_drv.h b/bsd/mga/mga_drv.h
index 05c4d9d2..9e51a20d 100644
--- a/bsd/mga/mga_drv.h
+++ b/bsd/mga/mga_drv.h
@@ -39,8 +39,8 @@
 
 typedef struct {
 	u_int32_t buffer_status;
-   	unsigned int num_dwords;
-   	unsigned int max_dwords;
+   	int num_dwords;
+   	int max_dwords;
    	u_int32_t *current_dma_ptr;
    	u_int32_t *head;
    	u_int32_t phys_head;
@@ -50,7 +50,7 @@ typedef struct {
 } drm_mga_prim_buf_t;
 
 typedef struct _drm_mga_freelist {
-   	unsigned int age;
+   	__volatile__ unsigned int age;
    	drm_buf_t *buf;
    	struct _drm_mga_freelist *next;
    	struct _drm_mga_freelist *prev;
@@ -82,6 +82,7 @@ typedef struct _drm_mga_private {
 	int use_agp;
    	drm_mga_warp_index_t WarpIndex[MGA_MAX_G400_PIPES];
 	unsigned int WarpPipe;
+	unsigned int vertexsize;
    	atomic_t pending_bufs;
    	void *status_page;
    	unsigned long real_status_page;
@@ -191,12 +192,20 @@ typedef struct {
 		  &tmp_buf->buffer_status)) {				\
  		mga_advance_primary(dev);				\
  		mga_dma_schedule(dev, 1);				\
+		tmp_buf = dev_priv->prim_bufs[dev_priv->current_prim_idx]; \
  	} else if( tmp_buf->max_dwords - tmp_buf->num_dwords < length ||\
  	    tmp_buf->sec_used > MGA_DMA_BUF_NR/2) {			\
 		set_bit(MGA_BUF_FORCE_FIRE, &tmp_buf->buffer_status);	\
  		mga_advance_primary(dev);				\
  		mga_dma_schedule(dev, 1);				\
+		tmp_buf = dev_priv->prim_bufs[dev_priv->current_prim_idx]; \
 	}								\
+	if(MGA_VERBOSE) 						\
+		DRM_DEBUG("PRIMGETPTR in %s\n", __FUNCTION__);		\
+	dma_ptr = tmp_buf->current_dma_ptr;				\
+	num_dwords = tmp_buf->num_dwords;				\
+	phys_head = tmp_buf->phys_head;					\
+	outcount = 0;							\
 } while(0)
 
 #define PRIMGETPTR(dev_priv) do {					\
@@ -344,6 +353,73 @@ drm_mga_prim_buf_t *tmp_buf = 					\
 #define MGAREG_YTOP 				0x1c98
 #define MGAREG_ZORG 				0x1c0c
 
+/* Warp registers */
+#define MGAREG_WR0                              0x2d00
+#define MGAREG_WR1                              0x2d04
+#define MGAREG_WR2                              0x2d08
+#define MGAREG_WR3                              0x2d0c
+#define MGAREG_WR4                              0x2d10
+#define MGAREG_WR5                              0x2d14
+#define MGAREG_WR6                              0x2d18
+#define MGAREG_WR7                              0x2d1c
+#define MGAREG_WR8                              0x2d20
+#define MGAREG_WR9                              0x2d24
+#define MGAREG_WR10                             0x2d28
+#define MGAREG_WR11                             0x2d2c
+#define MGAREG_WR12                             0x2d30
+#define MGAREG_WR13                             0x2d34
+#define MGAREG_WR14                             0x2d38
+#define MGAREG_WR15                             0x2d3c
+#define MGAREG_WR16                             0x2d40
+#define MGAREG_WR17                             0x2d44
+#define MGAREG_WR18                             0x2d48
+#define MGAREG_WR19                             0x2d4c
+#define MGAREG_WR20                             0x2d50
+#define MGAREG_WR21                             0x2d54
+#define MGAREG_WR22                             0x2d58
+#define MGAREG_WR23                             0x2d5c
+#define MGAREG_WR24                             0x2d60
+#define MGAREG_WR25                             0x2d64
+#define MGAREG_WR26                             0x2d68
+#define MGAREG_WR27                             0x2d6c
+#define MGAREG_WR28                             0x2d70
+#define MGAREG_WR29                             0x2d74
+#define MGAREG_WR30                             0x2d78
+#define MGAREG_WR31                             0x2d7c
+#define MGAREG_WR32                             0x2d80
+#define MGAREG_WR33                             0x2d84
+#define MGAREG_WR34                             0x2d88
+#define MGAREG_WR35                             0x2d8c
+#define MGAREG_WR36                             0x2d90
+#define MGAREG_WR37                             0x2d94
+#define MGAREG_WR38                             0x2d98
+#define MGAREG_WR39                             0x2d9c
+#define MGAREG_WR40                             0x2da0
+#define MGAREG_WR41                             0x2da4
+#define MGAREG_WR42                             0x2da8
+#define MGAREG_WR43                             0x2dac
+#define MGAREG_WR44                             0x2db0
+#define MGAREG_WR45                             0x2db4
+#define MGAREG_WR46                             0x2db8
+#define MGAREG_WR47                             0x2dbc
+#define MGAREG_WR48                             0x2dc0
+#define MGAREG_WR49                             0x2dc4
+#define MGAREG_WR50                             0x2dc8
+#define MGAREG_WR51                             0x2dcc
+#define MGAREG_WR52                             0x2dd0
+#define MGAREG_WR53                             0x2dd4
+#define MGAREG_WR54                             0x2dd8
+#define MGAREG_WR55                             0x2ddc
+#define MGAREG_WR56                             0x2de0
+#define MGAREG_WR57                             0x2de4
+#define MGAREG_WR58                             0x2de8
+#define MGAREG_WR59                             0x2dec
+#define MGAREG_WR60                             0x2df0
+#define MGAREG_WR61                             0x2df4
+#define MGAREG_WR62                             0x2df8
+#define MGAREG_WR63                             0x2dfc
+
+
 #define PDEA_pagpxfer_enable			0x2
 
 #define WIA_wmode_suspend			0x0
diff --git a/bsd/mga/mga_state.c b/bsd/mga/mga_state.c
index 73029bbe..fb365631 100644
--- a/bsd/mga/mga_state.c
+++ b/bsd/mga/mga_state.c
@@ -37,6 +37,20 @@
 typedef u_int16_t u16;
 typedef u_int32_t u32;
 
+#define MGAEMITCLIP_SIZE        10
+#define MGAEMITCTX_SIZE         20
+#define MGAG200EMITTEX_SIZE     20
+#define MGAG400EMITTEX0_SIZE    30
+#define MGAG400EMITTEX1_SIZE    25
+#define MGAG400EMITPIPE_SIZE    50
+#define MGAG200EMITPIPE_SIZE    15
+
+#define MAX_STATE_SIZE ((MGAEMITCLIP_SIZE * MGA_NR_SAREA_CLIPRECTS) + \
+                        MGAEMITCTX_SIZE + MGAG400EMITTEX0_SIZE + \
+                        MGAG400EMITTEX1_SIZE + MGAG400EMITPIPE_SIZE)
+
+
+
 static void mgaEmitClipRect(drm_mga_private_t * dev_priv,
 			    drm_clip_rect_t * box)
 {
@@ -49,6 +63,7 @@ static void mgaEmitClipRect(drm_mga_private_t * dev_priv,
 	PRIMGETPTR(dev_priv);
 
 	/* Force reset of dwgctl (eliminates clip disable) */
+	if (dev_priv->chipset == MGA_CARD_TYPE_G400) {
 #if 0
 	PRIMOUTREG(MGAREG_DMAPAD, 0);
 	PRIMOUTREG(MGAREG_DWGSYNC, 0);
@@ -60,11 +75,11 @@ static void mgaEmitClipRect(drm_mga_private_t * dev_priv,
 	PRIMOUTREG(MGAREG_DWGCTL, regs[MGA_CTXREG_DWGCTL]);
 	PRIMOUTREG(MGAREG_LEN + MGAREG_MGA_EXEC, 0x80000000);
 #endif
-
+	}
 	PRIMOUTREG(MGAREG_DMAPAD, 0);
 	PRIMOUTREG(MGAREG_CXBNDRY, ((box->x2) << 16) | (box->x1));
-	PRIMOUTREG(MGAREG_YTOP, box->y1 * dev_priv->stride / 2);
-	PRIMOUTREG(MGAREG_YBOT, box->y2 * dev_priv->stride / 2);
+	PRIMOUTREG(MGAREG_YTOP, box->y1 * dev_priv->stride / dev_priv->cpp);
+	PRIMOUTREG(MGAREG_YBOT, box->y2 * dev_priv->stride / dev_priv->cpp);
 
 	PRIMADVANCE(dev_priv);
 }
@@ -76,7 +91,7 @@ static void mgaEmitContext(drm_mga_private_t * dev_priv)
 	PRIMLOCALS;
 	DRM_DEBUG("%s\n", __FUNCTION__);
 
-	/* This takes a max of 15 dwords */
+	/* This takes a max of 20 dwords */
 	PRIMGETPTR(dev_priv);
 
 	PRIMOUTREG(MGAREG_DSTORG, regs[MGA_CTXREG_DSTORG]);
@@ -94,6 +109,12 @@ static void mgaEmitContext(drm_mga_private_t * dev_priv)
 		PRIMOUTREG(MGAREG_TDUALSTAGE0, regs[MGA_CTXREG_TDUAL0]);
 		PRIMOUTREG(MGAREG_TDUALSTAGE1, regs[MGA_CTXREG_TDUAL1]);
 		PRIMOUTREG(MGAREG_FCOL, regs[MGA_CTXREG_FCOL]);
+
+                PRIMOUTREG(MGAREG_STENCIL, regs[MGA_CTXREG_STENCIL]);
+                PRIMOUTREG(MGAREG_STENCILCTL, regs[MGA_CTXREG_STENCILCTL]);
+                PRIMOUTREG(MGAREG_DMAPAD, 0);
+                PRIMOUTREG(MGAREG_DMAPAD, 0);
+
 	} else {
 		PRIMOUTREG(MGAREG_FCOL, regs[MGA_CTXREG_FCOL]);
 		PRIMOUTREG(MGAREG_DMAPAD, 0);
@@ -128,9 +149,9 @@ static void mgaG200EmitTex(drm_mga_private_t * dev_priv)
 	PRIMOUTREG(MGAREG_TEXORG4, regs[MGA_TEXREG_ORG4]);
 	PRIMOUTREG(MGAREG_TEXWIDTH, regs[MGA_TEXREG_WIDTH]);
 	PRIMOUTREG(MGAREG_TEXHEIGHT, regs[MGA_TEXREG_HEIGHT]);
-	PRIMOUTREG(0x2d00 + 24 * 4, regs[MGA_TEXREG_WIDTH]);
+	PRIMOUTREG(MGAREG_WR24, regs[MGA_TEXREG_WIDTH]);
 
-	PRIMOUTREG(0x2d00 + 34 * 4, regs[MGA_TEXREG_HEIGHT]);
+	PRIMOUTREG(MGAREG_WR34, regs[MGA_TEXREG_HEIGHT]);
 	PRIMOUTREG(MGAREG_TEXTRANS, 0xffff);
 	PRIMOUTREG(MGAREG_TEXTRANSHIGH, 0xffff);
 	PRIMOUTREG(MGAREG_DMAPAD, 0);
@@ -138,11 +159,12 @@ static void mgaG200EmitTex(drm_mga_private_t * dev_priv)
 	PRIMADVANCE(dev_priv);
 }
 
+#define TMC_dualtex_enable              0x80
+
 static void mgaG400EmitTex0(drm_mga_private_t * dev_priv)
 {
 	drm_mga_sarea_t *sarea_priv = dev_priv->sarea_priv;
 	unsigned int *regs = sarea_priv->TexState[0];
-	int multitex = sarea_priv->WarpPipe & MGA_T2;
 	PRIMLOCALS;
 	DRM_DEBUG("%s\n", __FUNCTION__);
 
@@ -163,22 +185,21 @@ static void mgaG400EmitTex0(drm_mga_private_t * dev_priv)
 	PRIMOUTREG(MGAREG_TEXORG4, regs[MGA_TEXREG_ORG4]);
 	PRIMOUTREG(MGAREG_TEXWIDTH, regs[MGA_TEXREG_WIDTH]);
 	PRIMOUTREG(MGAREG_TEXHEIGHT, regs[MGA_TEXREG_HEIGHT]);
-	PRIMOUTREG(0x2d00 + 49 * 4, 0);
+        PRIMOUTREG(MGAREG_WR49, 0);
+
+        PRIMOUTREG(MGAREG_WR57, 0);
+        PRIMOUTREG(MGAREG_WR53, 0);
+        PRIMOUTREG(MGAREG_WR61, 0);
+        PRIMOUTREG(MGAREG_WR52, 0x40);
 
-	PRIMOUTREG(0x2d00 + 57 * 4, 0);
-	PRIMOUTREG(0x2d00 + 53 * 4, 0);
-	PRIMOUTREG(0x2d00 + 61 * 4, 0);
+        PRIMOUTREG(MGAREG_WR60, 0x40);
+        PRIMOUTREG(MGAREG_WR54, regs[MGA_TEXREG_WIDTH] | 0x40);
+        PRIMOUTREG(MGAREG_WR62, regs[MGA_TEXREG_HEIGHT] | 0x40);
 	PRIMOUTREG(MGAREG_DMAPAD, 0);
 
-	if (!multitex) {
-		PRIMOUTREG(0x2d00 + 52 * 4, 0x40);
-		PRIMOUTREG(0x2d00 + 60 * 4, 0x40);
 		PRIMOUTREG(MGAREG_DMAPAD, 0);
 		PRIMOUTREG(MGAREG_DMAPAD, 0);
-	}
 
-	PRIMOUTREG(0x2d00 + 54 * 4, regs[MGA_TEXREG_WIDTH] | 0x40);
-	PRIMOUTREG(0x2d00 + 62 * 4, regs[MGA_TEXREG_HEIGHT] | 0x40);
 	PRIMOUTREG(MGAREG_TEXTRANS, 0xffff);
 	PRIMOUTREG(MGAREG_TEXTRANSHIGH, 0xffff);
 
@@ -212,14 +233,15 @@ static void mgaG400EmitTex1(drm_mga_private_t * dev_priv)
 	PRIMOUTREG(MGAREG_TEXORG4, regs[MGA_TEXREG_ORG4]);
 	PRIMOUTREG(MGAREG_TEXWIDTH, regs[MGA_TEXREG_WIDTH]);
 	PRIMOUTREG(MGAREG_TEXHEIGHT, regs[MGA_TEXREG_HEIGHT]);
-	PRIMOUTREG(0x2d00 + 49 * 4, 0);
+        PRIMOUTREG(MGAREG_WR49, 0);
+
+        PRIMOUTREG(MGAREG_WR57, 0);
+        PRIMOUTREG(MGAREG_WR53, 0);
+        PRIMOUTREG(MGAREG_WR61, 0);
+        PRIMOUTREG(MGAREG_WR52, regs[MGA_TEXREG_WIDTH] | 0x40);
 
-	PRIMOUTREG(0x2d00 + 57 * 4, 0);
-	PRIMOUTREG(0x2d00 + 53 * 4, 0);
-	PRIMOUTREG(0x2d00 + 61 * 4, 0);
-	PRIMOUTREG(0x2d00 + 52 * 4, regs[MGA_TEXREG_WIDTH] | 0x40);
+        PRIMOUTREG(MGAREG_WR60, regs[MGA_TEXREG_HEIGHT] | 0x40);
 
-	PRIMOUTREG(0x2d00 + 60 * 4, regs[MGA_TEXREG_HEIGHT] | 0x40);
 	PRIMOUTREG(MGAREG_TEXTRANS, 0xffff);
 	PRIMOUTREG(MGAREG_TEXTRANSHIGH, 0xffff);
 	PRIMOUTREG(MGAREG_TEXCTL2, regs[MGA_TEXREG_CTL2] | 0x00008000);
@@ -227,12 +249,17 @@ static void mgaG400EmitTex1(drm_mga_private_t * dev_priv)
 	PRIMADVANCE(dev_priv);
 }
 
+#define MAGIC_FPARAM_HEX_VALUE 0x46480000
+/* This is the hex value of 12800.0f which is a magic value we must
+ * set in wr56.
+ */
+
+
 #define EMIT_PIPE 50
 static void mgaG400EmitPipe(drm_mga_private_t * dev_priv)
 {
 	drm_mga_sarea_t *sarea_priv = dev_priv->sarea_priv;
 	unsigned int pipe = sarea_priv->WarpPipe;
-	float fParam = 12800.0f;
 	PRIMLOCALS;
 	DRM_DEBUG("%s\n", __FUNCTION__);
 
@@ -266,14 +293,14 @@ static void mgaG400EmitPipe(drm_mga_private_t * dev_priv)
 			PRIMOUTREG(MGAREG_DWGCTL, MGA_FLUSH_CMD);
 
 			PRIMOUTREG(MGAREG_LEN + MGAREG_MGA_EXEC, 1);
-			PRIMOUTREG(MGAREG_DMAPAD, 0);
 			PRIMOUTREG(MGAREG_DWGSYNC, 0x7000);
-			PRIMOUTREG(MGAREG_DMAPAD, 0);
-
-			PRIMOUTREG(MGAREG_TEXCTL2, 0 | 0x00008000);
+			PRIMOUTREG(MGAREG_TEXCTL2, 0x00008000);
 			PRIMOUTREG(MGAREG_LEN + MGAREG_MGA_EXEC, 0);
+
 			PRIMOUTREG(MGAREG_TEXCTL2, 0x80 | 0x00008000);
 			PRIMOUTREG(MGAREG_LEN + MGAREG_MGA_EXEC, 0);
+			PRIMOUTREG(MGAREG_TEXCTL2, 0x00008000);
+			PRIMOUTREG(MGAREG_DMAPAD, 0);
 		}
 
 		PRIMOUTREG(MGAREG_WVRTXSZ, 0x00001807);
@@ -289,18 +316,20 @@ static void mgaG400EmitPipe(drm_mga_private_t * dev_priv)
 
 	PRIMOUTREG(MGAREG_WFLAG, 0);
 	PRIMOUTREG(MGAREG_WFLAG1, 0);
-	PRIMOUTREG(0x2d00 + 56 * 4, *((u32 *) (&fParam)));
+	PRIMOUTREG(MGAREG_WR56, MAGIC_FPARAM_HEX_VALUE);
 	PRIMOUTREG(MGAREG_DMAPAD, 0);
 
-	PRIMOUTREG(0x2d00 + 49 * 4, 0);	/* Tex stage 0 */
-	PRIMOUTREG(0x2d00 + 57 * 4, 0);	/* Tex stage 0 */
-	PRIMOUTREG(0x2d00 + 53 * 4, 0);	/* Tex stage 1 */
-	PRIMOUTREG(0x2d00 + 61 * 4, 0);	/* Tex stage 1 */
+        PRIMOUTREG(MGAREG_WR49, 0);     /* Tex stage 0 */
+        PRIMOUTREG(MGAREG_WR57, 0);     /* Tex stage 0 */
+        PRIMOUTREG(MGAREG_WR53, 0);     /* Tex stage 1 */
+        PRIMOUTREG(MGAREG_WR61, 0);     /* Tex stage 1 */
+
+
+        PRIMOUTREG(MGAREG_WR54, 0x40);  /* Tex stage 0 : w */
+        PRIMOUTREG(MGAREG_WR62, 0x40);  /* Tex stage 0 : h */
+        PRIMOUTREG(MGAREG_WR52, 0x40);  /* Tex stage 1 : w */
+        PRIMOUTREG(MGAREG_WR60, 0x40);  /* Tex stage 1 : h */
 
-	PRIMOUTREG(0x2d00 + 54 * 4, 0x40);	/* Tex stage 0 : w */
-	PRIMOUTREG(0x2d00 + 62 * 4, 0x40);	/* Tex stage 0 : h */
-	PRIMOUTREG(0x2d00 + 52 * 4, 0x40);	/* Tex stage 1 : w */
-	PRIMOUTREG(0x2d00 + 60 * 4, 0x40);	/* Tex stage 1 : h */
 
 	/* Dma pading required due to hw bug */
 	PRIMOUTREG(MGAREG_DMAPAD, 0xffffffff);
@@ -326,12 +355,12 @@ static void mgaG200EmitPipe(drm_mga_private_t * dev_priv)
 	PRIMOUTREG(MGAREG_WIADDR, WIA_wmode_suspend);
 	PRIMOUTREG(MGAREG_WVRTXSZ, 7);
 	PRIMOUTREG(MGAREG_WFLAG, 0);
-	PRIMOUTREG(0x2d00 + 24 * 4, 0);	/* tex w/h */
+	PRIMOUTREG(MGAREG_WR24, 0);     /* tex w/h */
 
-	PRIMOUTREG(0x2d00 + 25 * 4, 0x100);
-	PRIMOUTREG(0x2d00 + 34 * 4, 0);	/* tex w/h */
-	PRIMOUTREG(0x2d00 + 42 * 4, 0xFFFF);
-	PRIMOUTREG(0x2d00 + 60 * 4, 0xFFFF);
+        PRIMOUTREG(MGAREG_WR25, 0x100);
+        PRIMOUTREG(MGAREG_WR34, 0);     /* tex w/h */
+        PRIMOUTREG(MGAREG_WR42, 0xFFFF);
+        PRIMOUTREG(MGAREG_WR60, 0xFFFF);
 
 	/* Dma pading required due to hw bug */
 	PRIMOUTREG(MGAREG_DMAPAD, 0xffffffff);
@@ -495,7 +524,6 @@ static void mga_dma_dispatch_tex_blit(drm_device_t * dev,
 	y2 = length / 64;
 
 	PRIM_OVERFLOW(dev, dev_priv, 30);
-	PRIMGETPTR(dev_priv);
 
 	PRIMOUTREG(MGAREG_DSTORG, destOrg);
 	PRIMOUTREG(MGAREG_MACCESS, 0x00000000);
@@ -513,10 +541,11 @@ static void mga_dma_dispatch_tex_blit(drm_device_t * dev,
 	PRIMOUTREG(MGAREG_FXBNDRY, (63 << 16));
 	PRIMOUTREG(MGAREG_YDSTLEN + MGAREG_MGA_EXEC, y2);
 
+	PRIMOUTREG(MGAREG_DMAPAD, 0);
 	PRIMOUTREG(MGAREG_SRCORG, 0);
 	PRIMOUTREG(MGAREG_PITCH, dev_priv->stride / dev_priv->cpp);
-	PRIMOUTREG(MGAREG_DMAPAD, 0);
-	PRIMOUTREG(MGAREG_DMAPAD, 0);
+	PRIMOUTREG(MGAREG_DWGSYNC, 0x7000);
+
 	PRIMADVANCE(dev_priv);
 }
 
@@ -529,7 +558,6 @@ static void mga_dma_dispatch_vertex(drm_device_t * dev, drm_buf_t * buf)
 	int length = buf->used;
 	int use_agp = PDEA_pagpxfer_enable;
 	int i = 0;
-	int primary_needed;
 	PRIMLOCALS;
 	DRM_DEBUG("%s\n", __FUNCTION__);
 
@@ -545,11 +573,16 @@ static void mga_dma_dispatch_vertex(drm_device_t * dev, drm_buf_t * buf)
 		 * these numbers (Overestimating this doesn't hurt).  
 		 */
 		buf_priv->dispatched = 1;
-		primary_needed = (50 + 15 + 15 + 30 + 25 +
-				  10 + 15 * MGA_NR_SAREA_CLIPRECTS);
-		PRIM_OVERFLOW(dev, dev_priv, primary_needed);
+		PRIM_OVERFLOW(dev, dev_priv,
+			      (MAX_STATE_SIZE + (5 * MGA_NR_SAREA_CLIPRECTS)));
 		mgaEmitState(dev_priv);
 
+#if 0
+                length = dev_priv->vertexsize * 3 * 4;
+#endif
+
+
+
 		do {
 			if (i < sarea_priv->nbox) {
 				DRM_DEBUG("idx %d Emit box %d/%d:"
@@ -597,7 +630,6 @@ static void mga_dma_dispatch_indices(drm_device_t * dev,
 	unsigned int address = (unsigned int) buf->bus_address;
 	int use_agp = PDEA_pagpxfer_enable;
 	int i = 0;
-	int primary_needed;
 	PRIMLOCALS;
 	DRM_DEBUG("%s\n", __FUNCTION__);
 
@@ -611,9 +643,9 @@ static void mga_dma_dispatch_indices(drm_device_t * dev,
 		 * these numbers (Overestimating this doesn't hurt).  
 		 */
 		buf_priv->dispatched = 1;
-		primary_needed = (50 + 15 + 15 + 30 + 25 +
-				  10 + 15 * MGA_NR_SAREA_CLIPRECTS);
-		PRIM_OVERFLOW(dev, dev_priv, primary_needed);
+                PRIM_OVERFLOW(dev, dev_priv,
+                              (MAX_STATE_SIZE + (5 * MGA_NR_SAREA_CLIPRECTS)));
+
 		mgaEmitState(dev_priv);
 
 		do {
@@ -639,6 +671,7 @@ static void mga_dma_dispatch_indices(drm_device_t * dev,
 				    SETADD_mode_vertlist));
 			PRIMOUTREG(MGAREG_SETUPEND,
 				   ((address + end) | use_agp));
+				  
 			PRIMADVANCE(dev_priv);
 		} while (++i < sarea_priv->nbox);
 	}
@@ -653,7 +686,10 @@ static void mga_dma_dispatch_indices(drm_device_t * dev,
 
 static void mga_dma_dispatch_clear(drm_device_t * dev, int flags,
 				   unsigned int clear_color,
-				   unsigned int clear_zval)
+                                   unsigned int clear_zval,
+                                   unsigned int clear_colormask,
+                                   unsigned int clear_depthmask)
+
 {
 	drm_mga_private_t *dev_priv = dev->dev_private;
 	drm_mga_sarea_t *sarea_priv = dev_priv->sarea_priv;
@@ -662,7 +698,6 @@ static void mga_dma_dispatch_clear(drm_device_t * dev, int flags,
 	drm_clip_rect_t *pbox = sarea_priv->boxes;
 	unsigned int cmd;
 	int i;
-	int primary_needed;
 	PRIMLOCALS;
 	DRM_DEBUG("%s\n", __FUNCTION__);
 
@@ -671,11 +706,7 @@ static void mga_dma_dispatch_clear(drm_device_t * dev, int flags,
 	else
 		cmd = MGA_CLEAR_CMD | DC_atype_rstr;
 
-	primary_needed = nbox * 70;
-	if (primary_needed == 0)
-		primary_needed = 70;
-	PRIM_OVERFLOW(dev, dev_priv, primary_needed);
-	PRIMGETPTR(dev_priv);
+	PRIM_OVERFLOW(dev, dev_priv, 35 * MGA_NR_SAREA_CLIPRECTS);
 
 	for (i = 0; i < nbox; i++) {
 		unsigned int height = pbox[i].y2 - pbox[i].y1;
@@ -687,7 +718,7 @@ static void mga_dma_dispatch_clear(drm_device_t * dev, int flags,
 		if (flags & MGA_FRONT) {
 			DRM_DEBUG("clear front\n");
 			PRIMOUTREG(MGAREG_DMAPAD, 0);
-			PRIMOUTREG(MGAREG_DMAPAD, 0);
+			PRIMOUTREG(MGAREG_PLNWT, clear_colormask);
 			PRIMOUTREG(MGAREG_YDSTLEN,
 				   (pbox[i].y1 << 16) | height);
 			PRIMOUTREG(MGAREG_FXBNDRY,
@@ -702,7 +733,7 @@ static void mga_dma_dispatch_clear(drm_device_t * dev, int flags,
 		if (flags & MGA_BACK) {
 			DRM_DEBUG("clear back\n");
 			PRIMOUTREG(MGAREG_DMAPAD, 0);
-			PRIMOUTREG(MGAREG_DMAPAD, 0);
+			PRIMOUTREG(MGAREG_PLNWT, clear_colormask);
 			PRIMOUTREG(MGAREG_YDSTLEN,
 				   (pbox[i].y1 << 16) | height);
 			PRIMOUTREG(MGAREG_FXBNDRY,
@@ -717,7 +748,7 @@ static void mga_dma_dispatch_clear(drm_device_t * dev, int flags,
 		if (flags & MGA_DEPTH) {
 			DRM_DEBUG("clear depth\n");
 			PRIMOUTREG(MGAREG_DMAPAD, 0);
-			PRIMOUTREG(MGAREG_DMAPAD, 0);
+			PRIMOUTREG(MGAREG_PLNWT, clear_depthmask);
 			PRIMOUTREG(MGAREG_YDSTLEN,
 				   (pbox[i].y1 << 16) | height);
 			PRIMOUTREG(MGAREG_FXBNDRY,
@@ -746,31 +777,32 @@ static void mga_dma_dispatch_swap(drm_device_t * dev)
 	int nbox = sarea_priv->nbox;
 	drm_clip_rect_t *pbox = sarea_priv->boxes;
 	int i;
-	int primary_needed;
+	int pixel_stride = dev_priv->stride / dev_priv->cpp;
+
 	PRIMLOCALS;
 	DRM_DEBUG("%s\n", __FUNCTION__);
 
-	primary_needed = nbox * 5;
-	primary_needed += 60;
-	PRIM_OVERFLOW(dev, dev_priv, primary_needed);
-	PRIMGETPTR(dev_priv);
+        PRIM_OVERFLOW(dev, dev_priv, (MGA_NR_SAREA_CLIPRECTS * 5) + 20);
+
+        PRIMOUTREG(MGAREG_DMAPAD, 0);
+        PRIMOUTREG(MGAREG_DMAPAD, 0);
+        PRIMOUTREG(MGAREG_DWGSYNC, 0x7100);
+        PRIMOUTREG(MGAREG_DWGSYNC, 0x7000);
 
 	PRIMOUTREG(MGAREG_DSTORG, dev_priv->frontOffset);
 	PRIMOUTREG(MGAREG_MACCESS, dev_priv->mAccess);
 	PRIMOUTREG(MGAREG_SRCORG, dev_priv->backOffset);
-	PRIMOUTREG(MGAREG_AR5, dev_priv->stride / 2);
+        PRIMOUTREG(MGAREG_AR5, pixel_stride);
 
 	PRIMOUTREG(MGAREG_DMAPAD, 0);
 	PRIMOUTREG(MGAREG_DMAPAD, 0);
 	PRIMOUTREG(MGAREG_DMAPAD, 0);
 	PRIMOUTREG(MGAREG_DWGCTL, MGA_COPY_CMD);
 
+
 	for (i = 0; i < nbox; i++) {
 		unsigned int h = pbox[i].y2 - pbox[i].y1;
-		unsigned int start = pbox[i].y1 * dev_priv->stride / 2;
-
-		DRM_DEBUG("dispatch swap %d,%d-%d,%d!\n",
-			  pbox[i].x1, pbox[i].y1, pbox[i].x2, pbox[i].y2);
+		unsigned int start = pbox[i].y1 * pixel_stride;
 
 		PRIMOUTREG(MGAREG_AR0, start + pbox[i].x2 - 1);
 		PRIMOUTREG(MGAREG_AR3, start + pbox[i].x1);
@@ -814,7 +846,10 @@ int mga_clear_bufs(dev_t kdev, u_long cmd, caddr_t data,
 	 */
 	dev_priv->sarea_priv->dirty |= MGA_UPLOAD_CTX;
 	mga_dma_dispatch_clear(dev, clear.flags,
-			       clear.clear_color, clear.clear_depth);
+				clear.clear_color,
+                               clear.clear_depth,
+                               clear.clear_color_mask,
+                               clear.clear_depth_mask);
 	PRIMUPDATE(dev_priv);
 	mga_flush_write_combine();
 	s = splsofttq();
diff --git a/bsd/tdfx/Makefile b/bsd/tdfx/Makefile
index 471a5fbf..4362a5ba 100644
--- a/bsd/tdfx/Makefile
+++ b/bsd/tdfx/Makefile
@@ -3,7 +3,7 @@
 KMOD	= tdfx
 SRCS    = tdfx_drv.c tdfx_context.c
 SRCS   += device_if.h bus_if.h pci_if.h
-CFLAGS += ${DEBUG_FLAGS} -I..
+CFLAGS += ${DEBUG_FLAGS} -I. -I..
 KMODDEPS = drm
 
 @:
diff --git a/bsd/tdfx/tdfx_drv.c b/bsd/tdfx/tdfx_drv.c
index 91e7d560..202df7c3 100644
--- a/bsd/tdfx/tdfx_drv.c
+++ b/bsd/tdfx/tdfx_drv.c
@@ -42,7 +42,7 @@ MODULE_DEPEND(tdfx, agp, 1, 1, 1);
 
 #define TDFX_NAME	 "tdfx"
 #define TDFX_DESC	 "tdfx"
-#define TDFX_DATE	 "19991009"
+#define TDFX_DATE	 "20000928"
 #define TDFX_MAJOR	 1
 #define TDFX_MINOR	 0
 #define TDFX_PATCHLEVEL  0
@@ -64,6 +64,9 @@ static int tdfx_probe(device_t dev)
 	case 0x0005121a:
 		s = "3Dfx Voodoo 3 graphics accelerator";
 		break;
+	case 0x0009121a:
+		s = "3Dfx Voodoo 5 graphics accelerator";
+		break;
 	}
 
 	if (s) {