fix: garbage in fifo, optimized async commands

10 files changed, 447 insertions, 205 deletions

diff --git a/3d_accel.h b/3d_accel.h
index 1d9fcec..06ec786 100644
--- a/3d_accel.h
+++ b/3d_accel.h
@@ -250,12 +250,14 @@ BOOL SVGA_init_hw();
 

 BOOL SVGA_valid();

 

-#define SVGA_CB_USE_CONTEXT_DEVICE 0x80000000UL

-#define SVGA_CB_SYNC               0x40000000UL

-#define SVGA_CB_FORCE_FIFO         0x20000000UL

-#define SVGA_CB_FORCE_FENCE        0x10000000UL

-#define SVGA_CB_PRESENT_ASYNC      0x08000000UL

-#define SVGA_CB_PRESENT_GPU        0x04000000UL

+#define SVGA_CB_USE_CONTEXT_DEVICE 0x80000000UL /* CB control command */

+#define SVGA_CB_SYNC               0x40000000UL /* wait to command complete */

+#define SVGA_CB_FORCE_FIFO         0x20000000UL /* insert this command to FIFO, event the CB is working */

+#define SVGA_CB_FORCE_FENCE        0x10000000UL /* force insert fence */

+#define SVGA_CB_PRESENT            0x08000000UL /* this is 'present' command: WAIT for previous 'present' and 'render' to complete */

+#define SVGA_CB_DIRTY_SURFACE      0x04000000UL /* need reread GPU SURFACE first, can combine with SVGA_CB_PRESENT */

+#define SVGA_CB_RENDER             0x02000000UL /* this is 'render' cmd, WAIT for 'present', 'update' */

+#define SVGA_CB_UPDATE             0x01000000UL /* this is 'update' cmd, updates screen on HOST, WAIT for 'update', 'present' */

 

 // SVGA_CB_FLAG_DX_CONTEXT

 

diff --git a/makefile b/makefile
index f8bb043..67836e8 100644
--- a/makefile
+++ b/makefile
@@ -12,7 +12,7 @@ OBJS += &
 

 INCS = -I$(%WATCOM)\h\win -Iddk -Ivmware

 

-VER_BUILD = 67

+VER_BUILD = 68

 

 FLAGS = -DDRV_VER_BUILD=$(VER_BUILD)

 

diff --git a/vxd_lib.c b/vxd_lib.c
index cff6666..9be9cdc 100644
--- a/vxd_lib.c
+++ b/vxd_lib.c
@@ -37,6 +37,8 @@ THE SOFTWARE.
 

 #include "code32.h"

 

+#include "vxd_strings.h"

+

 /* some CRT */

 void memset(void *dst, int c, unsigned int size)

 {

@@ -153,20 +155,49 @@ DWORD Get_VMM_Version()
 	return ver;

 }

 

+static DWORD cs_count = 0;

+

 volatile void __cdecl Begin_Critical_Section(ULONG Flags)

 {

 	static ULONG sFlags;

 	sFlags = Flags;

 	

-	_asm push ecx

-	_asm mov ecx, [sFlags]

-	VMMCall(Begin_Critical_Section);

-	_asm pop ecx	

+	if(cs_count == 0)

+	{

+		_asm push ecx

+		_asm mov ecx, [sFlags]

+		VMMCall(Begin_Critical_Section);

+		_asm pop ecx

+	}

+	cs_count++;

 }

 

-volatile  void __cdecl End_Critical_Section()

+volatile void __cdecl End_Critical_Section()

 {

-	VMMCall(End_Critical_Section);

+	if(cs_count == 0)

+	{

+		dbg_printf(dbg_cs_underflow);

+		return;

+	}

+	

+	--cs_count;

+	if(cs_count == 0)

+	{

+		VMMCall(End_Critical_Section);

+	}

+}

+

+void Cleanup_Critical_Section()

+{

+	if(cs_count > 0)

+	{

+		dbg_printf(dbg_cs_active);

+		

+		while(cs_count != 0)

+		{

+			End_Critical_Section();

+		}

+	}

 }

 

 ULONG __cdecl Create_Semaphore(ULONG TokenCount)

diff --git a/vxd_lib.h b/vxd_lib.h
index 96453e5..f385422 100644
--- a/vxd_lib.h
+++ b/vxd_lib.h
@@ -43,6 +43,7 @@ DWORD __cdecl _RegCloseKey(DWORD hKey);
 DWORD __cdecl _RegQueryValueEx(DWORD hKey, char *lpszValueName, DWORD *lpdwReserved, DWORD *lpdwType, BYTE *lpbData, DWORD *lpcbData);

 volatile void __cdecl Begin_Critical_Section(ULONG Flags);

 volatile void __cdecl End_Critical_Section();

+void Cleanup_Critical_Section();

 ULONG __cdecl Create_Semaphore(ULONG TokenCount);

 void __cdecl Destroy_Semaphore(ULONG SemHandle);

 void __cdecl Wait_Semaphore(ULONG semHandle, ULONG flags);

diff --git a/vxd_main.c b/vxd_main.c
index dfd56dd..7de327b 100644
--- a/vxd_main.c
+++ b/vxd_main.c
@@ -48,6 +48,7 @@ THE SOFTWARE.
 

 #ifdef SVGA

 # include "svga_all.h"

+# include "vxd_svga.h"

 #endif

 

 #include "version.h"

@@ -368,6 +369,8 @@ WORD __stdcall VXD_API_Proc(PCRS_32 state)
 

 	}

 	

+	Cleanup_Critical_Section();

+	

 	if(rc == 0xFFFF)

 	{

 		state->Client_EFlags |= 0x1; // set carry

@@ -492,6 +495,7 @@ DWORD __stdcall Device_IO_Control_proc(DWORD vmhandle, struct DIOCParams *params
 {

 	DWORD *inBuf  = (DWORD*)params->lpInBuffer;

 	DWORD *outBuf = (DWORD*)params->lpOutBuffer;

+	DWORD rc = 1;

 	

 //	dbg_printf(dbg_deviceiocontrol, params->dwIoControlCode);

 	

@@ -501,58 +505,74 @@ DWORD __stdcall Device_IO_Control_proc(DWORD vmhandle, struct DIOCParams *params
 		case DIOC_OPEN:

 		case DIOC_CLOSEHANDLE:

 			dbg_printf(dbg_dic_system, params->dwIoControlCode);

-			return 0;

+			rc = 0;

+			break;

 		case OP_FBHDA_SETUP:

 			outBuf[0] = (DWORD)FBHDA_setup();

-			return 0;

+			rc = 0;

+			break;

 		case OP_FBHDA_ACCESS_BEGIN:

 			FBHDA_access_begin(inBuf[0]);

-			return 0;

+			rc = 0;

+			break;

 		case OP_FBHDA_ACCESS_END:

 			FBHDA_access_end(inBuf[0]);

-			return 0;

+			rc = 0;

+			break;

 		case OP_FBHDA_ACCESS_RECT:

 			FBHDA_access_rect(inBuf[0], inBuf[1], inBuf[2], inBuf[3]);

-			return 0;

+			rc = 0;

+			break;

 		case OP_FBHDA_SWAP:

 			outBuf[0] = FBHDA_swap(inBuf[0]);

-			return 0;

+			rc = 0;

+			break;

 		case OP_FBHDA_CLEAN:

 			FBHDA_clean();

-			return 0;

+			rc = 0;

+			break;

 		case OP_FBHDA_PALETTE_SET:

 			FBHDA_palette_set(inBuf[0], inBuf[1]);

-			return 0;

+			rc = 0;

+			break;

 		case OP_FBHDA_PALETTE_GET:

 			outBuf[0] = FBHDA_palette_get(inBuf[0]);

-			return 0;

+			rc = 0;

+			break;

 #ifdef SVGA

 		case OP_SVGA_VALID:

 			outBuf[0] = SVGA_valid();

-			return 0;

+			rc = 0;

+			break;

 		case OP_SVGA_CMB_ALLOC:

 			outBuf[0] = (DWORD)SVGA_CMB_alloc();

-			return 0;

+			rc = 0;

+			break;

 		case OP_SVGA_CMB_FREE:

 			SVGA_CMB_free((DWORD*)inBuf[0]);

-			return 0;

+			rc = 0;

+			break;

 		case OP_SVGA_CMB_SUBMIT:

 			{

 				SVGA_CMB_submit_io_t *inio  = (SVGA_CMB_submit_io_t*)inBuf;

 				SVGA_CMB_status_t *status = (SVGA_CMB_status_t*)outBuf;

 				

 				SVGA_CMB_submit(inio->cmb, inio->cmb_size, status, inio->flags, inio->DXCtxId);

-				return 0;

+				rc = 0;

+				break;

 			}

 		case OP_SVGA_FENCE_GET:

 			outBuf[0] = SVGA_fence_get();

-			return 0;

+			rc = 0;

+			break;

 		case OP_SVGA_FENCE_QUERY:

 			SVGA_fence_query(&outBuf[0], &outBuf[1]);

-			return 0;

+			rc = 0;

+			break;

 		case OP_SVGA_FENCE_WAIT:

 			SVGA_fence_wait(inBuf[0]);

-			return 0;

+			rc = 0;

+			break;

 		case OP_SVGA_REGION_CREATE:

 			{

 				SVGA_region_info_t *inio  = (SVGA_region_info_t*)inBuf;

@@ -567,50 +587,65 @@ DWORD __stdcall Device_IO_Control_proc(DWORD vmhandle, struct DIOCParams *params
 					SVGA_flushcache();

 					SVGA_region_create(outio);

 				}

-				return 0;

+				rc = 0;

+				break;

 			}

 		case OP_SVGA_REGION_FREE:

 			{

 				SVGA_region_info_t *inio  = (SVGA_region_info_t*)inBuf;

 				SVGA_region_free(inio);

-				return 0;

+				rc = 0;

+				break;

 			}

 		case OP_SVGA_QUERY:

 			outBuf[0] = SVGA_query(inBuf[0], inBuf[1]);

-			return 0;

+			rc = 0;

+			break;

 		case OP_SVGA_QUERY_VECTOR:

 			SVGA_query_vector(inBuf[0], inBuf[1], inBuf[2], outBuf);

-			return 0;

+			rc = 0;

+			break;

 		case OP_SVGA_DB_SETUP:

 			outBuf[0] = (DWORD)SVGA_DB_setup();

-			return 0;

+			rc = 0;

+			break;

 		case OP_SVGA_OT_SETUP:

 			outBuf[0] = (DWORD)SVGA_OT_setup();

-			return 0;

+			rc = 0;

+			break;

 		case OP_SVGA_FLUSHCACHE:

 			SVGA_flushcache();

-			return 0;

+			rc = 0;

+			break;

 		case OP_SVGA_VXDCMD:

 			outBuf[0] = (DWORD)SVGA_vxdcmd(inBuf[0]);

-			return 0;

+			rc = 0;

+			break;

 #ifdef DBGPRINT

 		/* export some mouse function for debuging */

 		case OP_MOUSE_MOVE:

 			mouse_move(inBuf[0], inBuf[1]);

-			return 0;

+			rc = 0;

+			break;

 		case OP_MOUSE_SHOW:

 			mouse_show();

-			return 0;

+			rc = 0;

+			break;

 		case OP_MOUSE_HIDE:

 			mouse_hide();

-			return 0;

+			rc = 0;

+			break;

 #endif /*DBGPRINT */

 #endif /* SVGA */

+		default:

+			dbg_printf(dbg_dic_unknown, params->dwIoControlCode);

+			break;

 	}

-		

-	dbg_printf(dbg_dic_unknown, params->dwIoControlCode);

 	

-	return 1;

+	Cleanup_Critical_Section();

+//	dbg_printf(dbg_deviceiocontrol_leave);

+	

+	return rc;

 }

 

 void Device_Init_proc()

diff --git a/vxd_strings.h b/vxd_strings.h
index 88188dd..5db428e 100644
--- a/vxd_strings.h
+++ b/vxd_strings.h
@@ -67,7 +67,8 @@ DSTR(dbg_cmd_on, "SVGA_CMB_submit: ptr=%X, first cmd=%X, flags=%X, size=%d\n");
 DSTR(dbg_cmd_off, "SVGA_CMB_submit: end - cmd: %X\n");

 DSTR(dbg_cmd_error, "CB error: %d, first cmd %X (error at %d)\n");

 

-DSTR(dbg_deviceiocontrol, "DeviceIoControl(%x);\n");

+DSTR(dbg_deviceiocontrol, "DeviceIoControl(%x)\n");

+DSTR(dbg_deviceiocontrol_leave, "DeviceIoControl Leave\n");

 

 DSTR(dbg_gmr, "GMR: %ld at %X\n");

 DSTR(dbg_gmr_succ, "GMR ALLOC: %ld (size: %ld)\n");

@@ -150,7 +151,7 @@ DSTR(dbg_irq_install_fail, "IRQ(%d) found, but cannot be traped\n");
 

 DSTR(dbg_no_irq, "No IRQ enabled\n");

 

-DSTR(dbg_disable, "HW disable");

+DSTR(dbg_disable, "HW disable\n");

 

 DSTR(dbg_pt_build, "PT_build(%ld): BASE=%lX TYPE=%ld USER=%lp\n");

 DSTR(dbg_pt_build_2, "PT_build(%ld): pt1_entries=%ld, pt2_entries=%ld\n");

@@ -162,6 +163,13 @@ DSTR(dbg_region_2, "SVGA_region_create #2: %ld\n");
 

 DSTR(dbg_cb_error, "Error (%ld): offset %ld, error command: %ld\n");

 

+DSTR(dbg_cs_underflow, "WARNING: closing inactive CS!\n");

+DSTR(dbg_cs_active, "WARNING: CS is still active!\n");

+

+DSTR(dbg_fence_wait, "SVGA_fence_wait(%lu) from line %ld\n");

+

+DSTR(dbg_queue_check, "CB_queue_check\n");

+

 #undef DSTR

 

 #endif

diff --git a/vxd_svga.c b/vxd_svga.c
index 7b5b44e..befc82e 100644
--- a/vxd_svga.c
+++ b/vxd_svga.c
@@ -81,7 +81,7 @@ void *ctlbuf = NULL;
 /* guest frame buffer is dirty */

 BOOL ST_FB_invalid = FALSE;

 

-DWORD present_fence = 0;

+//DWORD present_fence = 0;

 /*

  * strings

  */

@@ -255,8 +255,14 @@ BOOL SVGA_fence_is_passed(DWORD fence_id)
 	return FALSE;

 }

 

+#ifndef DBGPRINT

 void SVGA_fence_wait(DWORD fence_id)

+#else

+void SVGA_fence_wait_dbg(DWORD fence_id, int line)

+#endif

 {

+//	dbg_printf(dbg_fence_wait, fence_id, line);

+	

 	for(;;)

 	{

 		if(SVGA_fence_is_passed(fence_id))

@@ -264,6 +270,7 @@ void SVGA_fence_wait(DWORD fence_id)
 			break;

 		}

 		

+#if 1

 		if(cb_support && cb_context0)

 		{

 			if(CB_queue_check(NULL))

@@ -273,6 +280,7 @@ void SVGA_fence_wait(DWORD fence_id)
 				return;

 			}

 		}

+#endif

 		SVGA_Sync();

 	}

 }

@@ -1071,12 +1079,7 @@ void FBHDA_access_rect(DWORD left, DWORD top, DWORD right, DWORD bottom)
 	if(fb_lock_cnt++ == 0)

 	{

 		BOOL readback = FALSE;

-				

-		if(present_fence != 0)

-		{

-			SVGA_fence_wait(present_fence);

-//			readback = TRUE;

-		}

+		SVGA_CMB_wait_update();

 		

 		if(ST_FB_invalid)

 		{

@@ -1173,7 +1176,7 @@ void FBHDA_access_end(DWORD flags)
 	  	stupdate->rect.w = w;

 	  	stupdate->rect.h = h;

 

-			submit_cmdbuf(cmd_offset, SVGA_CB_PRESENT_ASYNC, 0);

+			submit_cmdbuf(cmd_offset, SVGA_CB_UPDATE, 0);

 		}

 	  else if(hda->bpp == 32)

 	  {

@@ -1188,7 +1191,7 @@ void FBHDA_access_end(DWORD flags)
 	  	cmd_update->width  = hda->width;

 	  	cmd_update->height = hda->height;

 	  	

-			submit_cmdbuf(cmd_offset, SVGA_CB_PRESENT_ASYNC, 0);

+			submit_cmdbuf(cmd_offset, SVGA_CB_UPDATE, 0);

 	  }

 	} // fb_lock_cnt == 0

 	

diff --git a/vxd_svga.h b/vxd_svga.h
index 299089a..9e4eadf 100644
--- a/vxd_svga.h
+++ b/vxd_svga.h
@@ -25,6 +25,14 @@ BOOL SVGA_fence_is_passed(DWORD fence_id);
 DWORD SVGA_fence_passed();

 DWORD SVGA_GetDevCap(DWORD search_id);

 

+#ifdef DBGPRINT

+void SVGA_fence_wait_dbg(DWORD fence_id, int line);

+

+#define SVGA_fence_wait(_fence) SVGA_fence_wait_dbg(_fence, __LINE__)

+

+#endif

+

+

 /* screen target */

 extern BOOL  st_used;

 extern DWORD st_flags;

@@ -63,5 +71,6 @@ void SVGA_CMB_free(DWORD *cmb);
 void SVGA_CB_start();

 void SVGA_CB_stop();

 void SVGA_CB_restart();

+void SVGA_CMB_wait_update();

 

 #endif /* __VXD_SVGA_H__INCLUDED__ */

diff --git a/vxd_svga_cb.c b/vxd_svga_cb.c
index 27d41fa..17df34f 100644
--- a/vxd_svga_cb.c
+++ b/vxd_svga_cb.c
@@ -44,6 +44,10 @@ THE SOFTWARE.
  * types

  */

 

+#define CBQ_PRESENT 0x01

+#define CBQ_RENDER  0x02

+#define CBQ_UPDATE  0x04

+

 #pragma pack(push)

 #pragma pack(1)

 typedef struct _cb_enable_t

@@ -75,7 +79,7 @@ extern BOOL gb_support;
 extern BOOL cb_support;

 extern BOOL cb_context0;

 

-extern DWORD present_fence;

+//extern DWORD present_fence;

 extern BOOL ST_FB_invalid;

 

 BOOL CB_queue_check(SVGACBHeader *tracked);

@@ -90,7 +94,6 @@ static uint64 cb_next_id = {0, 0};
 /*

  * Macros

  */

-

 #define WAIT_FOR_CB(_cb, _forcesync) \

 	do{ \

 		/*dbg_printf(dbg_wait_cb, __LINE__);*/ \

@@ -264,14 +267,36 @@ inline BOOL CB_queue_check_inline(SVGACBHeader *tracked)
 

 BOOL CB_queue_check(SVGACBHeader *tracked)

 {

+//	dbg_printf(dbg_queue_check);

 	return CB_queue_check_inline(tracked);

 }

 

-void CB_queue_insert(SVGACBHeader *cb)

+static BOOL CB_queue_is_flags_set(DWORD flags)

+{

+	cb_queue_t *item = cb_queue_info.first;

+	

+	if(flags == 0) return FALSE;

+	

+	while(item != NULL)

+	{

+		if((item->flags & flags) != 0)

+		{

+			return TRUE;

+		}

+		

+		item = item->next;

+	}

+	

+	return FALSE;

+}

+

+void CB_queue_insert(SVGACBHeader *cb, DWORD flags)

 {

 	cb_queue_t *item = (cb_queue_t*)(cb-1);

 	item->next = NULL;

-	

+	item->flags = flags;

+	item->data_size = cb->length;

+

 	if(cb_queue_info.last != NULL)

 	{

 		cb_queue_info.last->next = item;

@@ -300,14 +325,112 @@ void CB_queue_erase()
 	cb_queue_info.last  = NULL;

 }

 

+static DWORD flags_to_cbq(DWORD cb_flags)

+{

+	DWORD r = 0;

+	

+	if((cb_flags & SVGA_CB_PRESENT) != 0)

+	{

+		r |= CBQ_PRESENT;

+	}

+	

+	if((cb_flags & SVGA_CB_RENDER) != 0)

+	{

+		r |= CBQ_RENDER;

+	}

+	

+	if((cb_flags & SVGA_CB_UPDATE) != 0)

+	{

+		r |= SVGA_CB_UPDATE;

+	}

+	

+	return r;

+}

+

+static DWORD flags_to_cbq_check(DWORD cb_flags)

+{

+	DWORD r = 0;

+

+	if((cb_flags & SVGA_CB_PRESENT) != 0)

+	{

+		r |= CBQ_PRESENT | CBQ_RENDER;

+	}

+	

+	if((cb_flags & SVGA_CB_RENDER) != 0)

+	{

+		r |= CBQ_RENDER | SVGA_CB_UPDATE;

+	}

+	

+	if((cb_flags & SVGA_CB_UPDATE) != 0)

+	{

+		r |= CBQ_RENDER | SVGA_CB_UPDATE;

+	}

+	

+	return r;

+}

+

+static uint32 fence_present = 0;

+static uint32 fence_render  = 0;

+static uint32 fence_update  = 0;

+

+static void flags_fence_check(DWORD cb_flags)

+{

+	DWORD to_check = flags_to_cbq_check(cb_flags);

+	

+	if((to_check & CBQ_PRESENT) != 0)

+	{

+		if(fence_present)

+		{

+			SVGA_fence_wait(fence_present);

+			fence_present = 0;

+		}

+	}

+	

+	if((to_check & CBQ_RENDER) != 0)

+	{

+		if(fence_render)

+		{

+			SVGA_fence_wait(fence_render);

+			fence_render = 0;

+		}

+	}

+	

+	if((to_check & CBQ_UPDATE) != 0)

+	{

+		if(fence_update)

+		{

+			SVGA_fence_wait(fence_update);

+			fence_update = 0;

+		}

+	}

+}

+

+static void flags_fence_insert(DWORD cb_flags, uint32 fence)

+{

+	if((cb_flags & SVGA_CB_PRESENT) != 0)

+	{

+		fence_present = fence;

+	}

+	

+	if((cb_flags & SVGA_CB_RENDER) != 0)

+	{

+		fence_render = fence;

+	}

+	

+	if((cb_flags & SVGA_CB_UPDATE) != 0)

+	{

+		fence_update = fence;

+	}

+}

+

+#define flags_fifo_fence_need(_flags) (((_flags) & (SVGA_CB_SYNC | SVGA_CB_FORCE_FENCE | SVGA_CB_PRESENT | SVGA_CB_RENDER | SVGA_CB_UPDATE)) != 0)

+#define flags_cb_fence_need(_flags) (((_flags) & (SVGA_CB_FORCE_FENCE)) != 0)

 

 static void SVGA_CMB_submit_critical(DWORD FBPTR cmb, DWORD cmb_size, SVGA_CMB_status_t FBPTR status, DWORD flags, DWORD DXCtxId)

 {

 	DWORD fence = 0;

 	SVGACBHeader *cb = ((SVGACBHeader *)cmb)-1;

 	

-	CB_queue_check_inline(NULL);

-	

 	if(status)

 	{

 		cb->status = SVGA_CB_STATUS_NONE;

@@ -320,43 +443,28 @@ static void SVGA_CMB_submit_critical(DWORD FBPTR cmb, DWORD cmb_size, SVGA_CMB_s
 //	debug_cmdbuf_trace(cmb, cmb_size, SVGA_3D_CMD_BLIT_SURFACE_TO_SCREEN);

 //		debug_draw(cmb, cmb_size);

 #endif

-

-	if(flags & SVGA_CB_PRESENT_ASYNC)

-	{

-		if(present_fence != 0)

-		{

-			SVGA_fence_wait(present_fence);

-		}

-		

-		flags |= SVGA_CB_FORCE_FENCE;

-	}

-	else if(present_fence)

-	{

-		if(SVGA_fence_is_passed(present_fence))

-		{

-			present_fence = 0;

-		}

-	}

 	

-	if(flags & SVGA_CB_PRESENT_GPU)

+	if(flags & SVGA_CB_DIRTY_SURFACE)

 	{

 		ST_FB_invalid = TRUE;

 	}

-	

+

 	if(  /* CB are supported and enabled */

-		cb_support &&

-		(cb_context0 || (flags & SVGA_CB_USE_CONTEXT_DEVICE)) &&

+		cb_support && cb_context0 &&

 		(flags & SVGA_CB_FORCE_FIFO) == 0

 	)

 	{

+		/***

+		 *

+		 * COMMAND BUFFER procesing

+		 *

+		 ***/

 		DWORD cbhwctxid = SVGA_CB_CONTEXT_0;

+		DWORD cbq_check = flags_to_cbq_check(flags);

 		

-		if(flags & SVGA_CB_USE_CONTEXT_DEVICE)

-		{

-			cbhwctxid = SVGA_CB_CONTEXT_DEVICE;

-		}

+		//if(flags & SVGA_CB_USE_CONTEXT_DEVICE)	cbhwctxid = SVGA_CB_CONTEXT_DEVICE;

 		

-		if(flags & SVGA_CB_FORCE_FENCE)

+		if(flags_cb_fence_need(flags))

 		{

 			DWORD dwords = cmb_size/sizeof(DWORD);

 			fence = SVGA_fence_get();

@@ -365,6 +473,12 @@ static void SVGA_CMB_submit_critical(DWORD FBPTR cmb, DWORD cmb_size, SVGA_CMB_s
 			cmb_size += sizeof(DWORD)*2;

 		}

 		

+		/* wait */

+		do

+		{

+			CB_queue_check_inline(NULL);

+		} while(CB_queue_is_flags_set(cbq_check));

+		

 		if(cmb_size == 0)

 		{

 			cb->status = SVGA_PROC_COMPLETED;

@@ -377,10 +491,10 @@ static void SVGA_CMB_submit_critical(DWORD FBPTR cmb, DWORD cmb_size, SVGA_CMB_s
 		}

 		else

 		{

-			cb->status = SVGA_CB_STATUS_NONE;

+			cb->status      = SVGA_CB_STATUS_NONE;

 			cb->errorOffset = 0;

-			cb->offset = 0; /* VMware modified this, needs to be clear */

-			cb->flags  = SVGA_CB_FLAG_NO_IRQ;

+			cb->offset      = 0; /* VMware modified this, needs to be clear */

+			cb->flags       = SVGA_CB_FLAG_NO_IRQ;

 			

 			if(flags & SVGA_CB_FLAG_DX_CONTEXT)

 			{

@@ -392,11 +506,11 @@ static void SVGA_CMB_submit_critical(DWORD FBPTR cmb, DWORD cmb_size, SVGA_CMB_s
 				cb->dxContext = 0;

 			}

 			

-			cb->id.low = cb_next_id.low;

-			cb->id.hi  = cb_next_id.hi;

-			cb->length = cmb_size;

+			cb->id.low      = cb_next_id.low;

+			cb->id.hi       = cb_next_id.hi;

+			cb->length      = cmb_size;

 			

-			CB_queue_insert(cb);			

+			CB_queue_insert(cb, flags_to_cbq(flags));			

 			

 			SVGA_WriteReg(SVGA_REG_COMMAND_HIGH, 0); // high part of 64-bit memory address...

 			SVGA_WriteReg(SVGA_REG_COMMAND_LOW, (cb->ptr.pa.low - sizeof(SVGACBHeader)) | cbhwctxid);

@@ -417,10 +531,10 @@ static void SVGA_CMB_submit_critical(DWORD FBPTR cmb, DWORD cmb_size, SVGA_CMB_s
 						fence = SVGA_fence_passed();

 					}

 				}

-				else

+				/*else

 				{

 					dbg_printf(dbg_cb_suc);

-				}

+				}*/

 				

 				if(status)

 				{

@@ -440,45 +554,30 @@ static void SVGA_CMB_submit_critical(DWORD FBPTR cmb, DWORD cmb_size, SVGA_CMB_s
 			}

 		}	

 	}

-	else /* copy to FIFO */

+	else

 	{

+		/***

+		 *

+		 * FIFO procesing

+		 *

+		 ***/

 		DWORD *ptr = cmb;

 		DWORD dwords = cmb_size/sizeof(DWORD);

 		DWORD nextCmd, max, min;

 		

 		/* insert fence CMD */

-		fence = SVGA_fence_get();

-		ptr[dwords]   = SVGA_CMD_FENCE;

-		ptr[dwords+1] = fence;

-		

-		dwords += 2;

-		

-		//SVGA_Flush();

-		

-		nextCmd = gSVGA.fifoMem[SVGA_FIFO_NEXT_CMD];

-		max     = gSVGA.fifoMem[SVGA_FIFO_MAX];

-		min     = gSVGA.fifoMem[SVGA_FIFO_MIN];

-		

-		/* copy to fifo */

-		while(dwords > 0)

+		if(flags_fifo_fence_need(flags))

 		{

-			gSVGA.fifoMem[nextCmd/sizeof(DWORD)] = *ptr;

-			ptr++;

-			

-			nextCmd += sizeof(uint32);

-			if (nextCmd >= max)

-			{

-				nextCmd = min;

-			}

-			gSVGA.fifoMem[SVGA_FIFO_NEXT_CMD] = nextCmd;

-			dwords--;

+			fence = SVGA_fence_get();

+			ptr[dwords++] = SVGA_CMD_FENCE;

+			ptr[dwords++] = fence;

 		}

 		

-		//SVGA_Sync();

+		flags_fence_check(flags);

 		

-		if(flags & SVGA_CB_SYNC)

+		if(dwords == 0)

 		{

-			SVGA_fence_wait(fence);

+			cb->status = SVGA_PROC_COMPLETED;

 			if(status)

 			{

 				status->sStatus = SVGA_PROC_COMPLETED;

@@ -488,15 +587,49 @@ static void SVGA_CMB_submit_critical(DWORD FBPTR cmb, DWORD cmb_size, SVGA_CMB_s
 		}

 		else

 		{

-			if(status)

+			nextCmd = gSVGA.fifoMem[SVGA_FIFO_NEXT_CMD];

+			max     = gSVGA.fifoMem[SVGA_FIFO_MAX];

+			min     = gSVGA.fifoMem[SVGA_FIFO_MIN];

+			

+			/* copy to fifo */

+			while(dwords > 0)

 			{

-				status->sStatus = SVGA_PROC_FENCE;

-				status->qStatus = NULL;

-				status->fifo_fence_used = fence;

+				gSVGA.fifoMem[nextCmd/sizeof(DWORD)] = *ptr;

+				ptr++;

+				

+				nextCmd += sizeof(uint32);

+				if (nextCmd >= max)

+				{

+					nextCmd = min;

+				}

+				gSVGA.fifoMem[SVGA_FIFO_NEXT_CMD] = nextCmd;

+				dwords--;

 			}

-		}

-		

-		cb->status = SVGA_PROC_COMPLETED;

+			

+			if(flags & SVGA_CB_SYNC)

+			{

+				SVGA_fence_wait(fence);

+				if(status)

+				{

+					status->sStatus = SVGA_PROC_COMPLETED;

+					status->qStatus = NULL;

+					status->fifo_fence_used = 0;

+				}

+			}

+			else

+			{

+				flags_fence_insert(flags, fence);

+				

+				if(status)

+				{

+					status->sStatus = SVGA_PROC_FENCE;

+					status->qStatus = NULL;

+					status->fifo_fence_used = fence;

+				}

+			}

+			

+			cb->status = SVGA_PROC_COMPLETED;

+		} // size > 0

 	} /* FIFO */

 	

 	if(status)

@@ -504,11 +637,6 @@ static void SVGA_CMB_submit_critical(DWORD FBPTR cmb, DWORD cmb_size, SVGA_CMB_s
 		status->fifo_fence_last = SVGA_fence_passed();

 	}

 	

-	if(flags & SVGA_CB_PRESENT_ASYNC)

-	{

-		present_fence = fence;

-	}

-	

 	//dbg_printf(dbg_cmd_off, cmb[0]);

 }

 

@@ -627,7 +755,6 @@ void SVGA_CB_stop()
 		SVGA_Sync();

 		

 		CB_queue_erase();

-		present_fence = 0; // remove waiting fence

 		

 		dbg_printf(dbg_cb_stop_status, status);

 	}

@@ -643,3 +770,18 @@ void SVGA_CB_restart()
 	

 	SVGA_CB_start();

 }

+

+void SVGA_CMB_wait_update()

+{

+	if(cb_support && cb_context0)

+	{

+		do

+		{

+			CB_queue_check_inline(NULL);

+		} while(CB_queue_is_flags_set(CBQ_UPDATE | CBQ_UPDATE));

+	}

+	else

+	{

+		flags_fence_check(SVGA_CB_UPDATE);

+	}

+}

diff --git a/vxd_svga_mem.c b/vxd_svga_mem.c
index fb1d87b..16c3820 100644
--- a/vxd_svga_mem.c
+++ b/vxd_svga_mem.c
@@ -714,94 +714,104 @@ BOOL SVGA_region_create(SVGA_region_info_t *rinfo)
 		ULONG blk_pages_raw = 0;

 		

 		/* allocate user block */

-		maddr = _PageAllocate(nPages+pt_pages, pa_type, pa_vm, pa_align, 0x0, 0x100000, NULL, pa_flags);

-

-		if(!maddr)

+		if(!rinfo->mobonly)

 		{

-			End_Critical_Section();

-			return FALSE;

+			maddr = _PageAllocate(nPages+pt_pages, pa_type, pa_vm, pa_align, 0x0, 0x100000, NULL, pa_flags);

 		}

-

-		laddr = maddr + pt_pages*P_SIZE;

-

-		/* determine how many physical continuous blocks we have */

-		base_ppn = getPPN(laddr);

-		base_cnt = 1;

-		blocks   = 1;

-		for(pgi = 1; pgi < nPages; pgi++)

+		else

 		{

-			taddr = laddr + pgi*P_SIZE;

-			tppn = getPPN(taddr);

-			

-			if(tppn != base_ppn + base_cnt)

-			{

-				base_ppn = tppn;

-				base_cnt = 1;

-				blocks++;

-			}

-			else

-			{

-				base_cnt++;

-			}

+			maddr = _PageAllocate(nPages+pt_pages, pa_type, pa_vm, 0, 0x0, 0x100000, NULL, PAGEFIXED);

 		}

 

-		// number of pages to store regions information

-		blk_pages_raw = (blocks + MDONPAGE - 1)/MDONPAGE;

-

-		// number of pages including pages-cross descriptors

-		blk_pages = (blocks + blk_pages_raw + MDONPAGE - 1)/MDONPAGE;

-		

-		dbg_printf(dbg_region_1, blocks+blk_pages_raw, SVGA_ReadReg(SVGA_REG_GMR_MAX_DESCRIPTOR_LENGTH));

-

-		/* allocate memory for GMR descriptor */

-		pgblk = _PageAllocate(blk_pages, pa_type, pa_vm, pa_align, 0x0, 0x100000, NULL, pa_flags);

-		if(!pgblk)

+		if(!maddr)

 		{

-			_PageFree((PVOID)laddr, 0);

 			End_Critical_Section();

 			return FALSE;

 		}

 

-		desc = (SVGAGuestMemDescriptor*)pgblk;

-		memset(desc, 0, blk_pages*P_SIZE);

-

-		blocks         = 1;

-		desc->ppn      = getPPN(laddr);

-		desc->numPages = 1;

-		/* fill GMR physical structure */

-		for(pgi = 1; pgi < nPages; pgi++)

+		laddr = maddr + pt_pages*P_SIZE;

+		

+		if(!rinfo->mobonly)

 		{

-			taddr = laddr + pgi*P_SIZE;

-			tppn = getPPN(taddr);

-					

-			if(tppn == desc->ppn + desc->numPages)

+			/* determine how many physical continuous blocks we have */

+			base_ppn = getPPN(laddr);

+			base_cnt = 1;

+			blocks   = 1;

+			for(pgi = 1; pgi < nPages; pgi++)

 			{

-				desc->numPages++;

+				taddr = laddr + pgi*P_SIZE;

+				tppn = getPPN(taddr);

+				

+				if(tppn != base_ppn + base_cnt)

+				{

+					base_ppn = tppn;

+					base_cnt = 1;

+					blocks++;

+				}

+				else

+				{

+					base_cnt++;

+				}

 			}

-			else

+	

+			// number of pages to store regions information

+			blk_pages_raw = (blocks + MDONPAGE - 1)/MDONPAGE;

+	

+			// number of pages including pages-cross descriptors

+			blk_pages = (blocks + blk_pages_raw + MDONPAGE - 1)/MDONPAGE;

+			

+			dbg_printf(dbg_region_1, blocks+blk_pages_raw, SVGA_ReadReg(SVGA_REG_GMR_MAX_DESCRIPTOR_LENGTH));

+	

+			/* allocate memory for GMR descriptor */

+			pgblk = _PageAllocate(blk_pages, pa_type, pa_vm, pa_align, 0x0, 0x100000, NULL, pa_flags);

+			if(!pgblk)

+			{

+				_PageFree((PVOID)laddr, 0);

+				End_Critical_Section();

+				return FALSE;

+			}

+	

+			desc = (SVGAGuestMemDescriptor*)pgblk;

+			memset(desc, 0, blk_pages*P_SIZE);

+	

+			blocks         = 1;

+			desc->ppn      = getPPN(laddr);

+			desc->numPages = 1;

+			/* fill GMR physical structure */

+			for(pgi = 1; pgi < nPages; pgi++)

 			{

-				/* the next descriptor is on next page page */

-				if(((blocks+1) % MDONPAGE) == 0)

+				taddr = laddr + pgi*P_SIZE;

+				tppn = getPPN(taddr);

+						

+				if(tppn == desc->ppn + desc->numPages)

 				{

+					desc->numPages++;

+				}

+				else

+				{

+					/* the next descriptor is on next page page */

+					if(((blocks+1) % MDONPAGE) == 0)

+					{

+						desc++;

+						desc->numPages = 0;

+						desc->ppn = getPPN((DWORD)(desc+1));

+						blocks++;

+					}

+	

 					desc++;

-					desc->numPages = 0;

-					desc->ppn = getPPN((DWORD)(desc+1));

+					desc->ppn = tppn;

+					desc->numPages = 1;

 					blocks++;

 				}

-

-				desc++;

-				desc->ppn = tppn;

-				desc->numPages = 1;

-				blocks++;

 			}

+	

+			desc++;

+			desc->ppn = 0;

+			desc->numPages = 0;

+			

+			dbg_printf(dbg_region_2, blocks);

 		}

 

-		desc++;

-		desc->ppn = 0;

-		desc->numPages = 0;

-		

-		dbg_printf(dbg_region_2, blocks);

-

 		rinfo->mob_address    = (void*)maddr;

 		rinfo->mob_ppn        = 0;

 		rinfo->mob_pt_depth   = 0;

@@ -837,7 +847,7 @@ BOOL SVGA_region_create(SVGA_region_info_t *rinfo)
 			SVGA_WriteReg(SVGA_REG_GMR_DESCRIPTOR, rinfo->region_ppn);

 			SVGA_Sync(); // notify register change

 			

-			SVGA_Flush_CB_critical();

+			//SVGA_Flush_CB_critical();

 		}

 		else

 		{

@@ -859,6 +869,7 @@ BOOL SVGA_region_create(SVGA_region_info_t *rinfo)
 		mob->sizeInBytes = rinfo->size;

 

 		submit_cmdbuf(cmdoff, SVGA_CB_SYNC, 0);

+		SVGA_Sync();

 		

   	rinfo->is_mob = 1;

 	}

@@ -871,7 +882,7 @@ BOOL SVGA_region_create(SVGA_region_info_t *rinfo)
 	

 	End_Critical_Section();

 	

-	//dbg_printf(dbg_gmr_succ, rinfo->region_id, rinfo->size);

+	dbg_printf(dbg_gmr_succ, rinfo->region_id, rinfo->size);

 	

 	return TRUE;

 }