VXD optimizations, fixed mouse blit bug, barriers and memory management

8 files changed, 109 insertions, 47 deletions

diff --git a/3d_accel.h b/3d_accel.h
index a263afb..3a69afe 100644
--- a/3d_accel.h
+++ b/3d_accel.h
@@ -59,6 +59,7 @@ THE SOFTWARE.
 #define OP_SVGA_QUERY_VECTOR  0x200E  /* VXD */

 #define OP_SVGA_DB_SETUP      0x200F  /* VXD */

 #define OP_SVGA_OT_SETUP      0x2010  /* VXD */

+#define OP_SVGA_FLUSHCACHE    0x2011  /* VXD */

 

 #define OP_VBE_VALID          0x3000 /* VXD, DRV */

 #define OP_VBE_SETMODE        0x3001 /* DRV */

@@ -113,6 +114,7 @@ typedef struct FBHDA
 #define FB_ACCEL_VMSVGA3D      32

 #define FB_ACCEL_VMSVGA10      64

 #define FB_MOUSE_NO_BLIT      128

+#define FB_FORCE_SOFTWARE     256

 

 /* for internal use in RING-0 by VXD only */

 BOOL FBHDA_init_hw(); 

@@ -206,6 +208,8 @@ typedef struct SVGA_DB_surface
 	DWORD height;

 	DWORD bpp;

 	DWORD gmrId; /* != 0 for GB surfaces */

+	DWORD gmrMngt; /* 1 when auto destroy MOB and region when releasing surface */

+	DWORD size; /* surface size in bytes */

 } SVGA_DB_surface_t;

 

 typedef struct SVGA_DB

@@ -284,6 +288,8 @@ typedef struct SVGA_OT_info_entry
 

 SVGA_OT_info_entry_t *SVGA_OT_setup();

 

+void SVGA_flushcache();

+

 #endif /* SVGA */

 

 /*

diff --git a/enable.c b/enable.c
index c7434d7..544bc11 100644
--- a/enable.c
+++ b/enable.c
@@ -146,12 +146,10 @@ DWORD PASCAL CreateDIBPDeviceX( LPBITMAPINFO lpInfo, LPPDEVICE lpDevice, LPVOID
 VOID WINAPI __loadds BeginAccess_VXD( LPPDEVICE lpDevice, WORD wLeft, WORD wTop, WORD wRight, WORD wBottom, WORD wFlags )

 {

 	DWORD dflags = 0;

-	if(!(wFlags & CURSOREXCLUDE))

+	if(wFlags & CURSOREXCLUDE)

 	{

-		dflags |= FBHDA_IGNORE_CURSOR;

+		FBHDA_access_begin(dflags);

 	}

-	

-	FBHDA_access_begin(dflags);

 	if(!mouse_vxd)

 	{

 		DIB_BeginAccess(lpDevice, wLeft, wTop, wRight, wBottom, wFlags);

@@ -161,16 +159,14 @@ VOID WINAPI __loadds BeginAccess_VXD( LPPDEVICE lpDevice, WORD wLeft, WORD wTop,
 VOID WINAPI __loadds EndAccess_VXD( LPPDEVICE lpDevice, WORD wFlags )

 {

 	DWORD dflags = 0;

-	if(!(wFlags & CURSOREXCLUDE))

-	{

-		dflags |= FBHDA_IGNORE_CURSOR;

-	}

-	

 	if(!mouse_vxd)

 	{

 		DIB_EndAccess(lpDevice, wFlags);

 	}

-	FBHDA_access_end(dflags);

+	if(wFlags & CURSOREXCLUDE)

+	{

+		FBHDA_access_end(dflags);

+	}

 }

 

 

diff --git a/makefile b/makefile
index 0635e2b..c02bc10 100644
--- a/makefile
+++ b/makefile
@@ -11,7 +11,7 @@ OBJS += &
 

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

 

-VER_BUILD = 32

+VER_BUILD = 34

 

 FLAGS = -DDRV_VER_BUILD=$(VER_BUILD)

 

@@ -32,13 +32,15 @@ DBGFILE =
 DBGFILE32 =

 !endif

 CFLAGS = -q -wx -s -zu -zls -6 -fp6

-CFLAGS32 = -q -wx -s -zls -6s -fp6 -mf -DVXD32

+CFLAGS32 = -q -wx -s -zls -6s -fp6 -mf -DVXD32 -fpi87 -ei -oeatxhn 

 CC = wcc

 CC32 = wcc386

 

 # Log VXD to com2

 !ifdef DBGPRINT

 CFLAGS32 += -DCOM2

+!else

+CFLAGS32 += -d0

 !endif

 

 #all : boxvmini.drv vmwsmini.drv qemumini.drv vmwsmini.vxd qemumini.vxd

diff --git a/vxd_lib.c b/vxd_lib.c
index 07fc652..5c77b58 100644
--- a/vxd_lib.c
+++ b/vxd_lib.c
@@ -101,7 +101,7 @@ DWORD Get_VMM_Version()
 	return ver;

 }

 

-void __cdecl Begin_Critical_Section(ULONG Flags)

+volatile void __cdecl Begin_Critical_Section(ULONG Flags)

 {

 	static ULONG sFlags;

 	sFlags = Flags;

@@ -112,7 +112,7 @@ void __cdecl Begin_Critical_Section(ULONG Flags)
 	_asm pop ecx	

 }

 

-void __cdecl End_Critical_Section()

+volatile  void __cdecl End_Critical_Section()

 {

 	VMMCall(End_Critical_Section);

 }

diff --git a/vxd_lib.h b/vxd_lib.h
index 4589be9..760e8f6 100644
--- a/vxd_lib.h
+++ b/vxd_lib.h
@@ -39,8 +39,8 @@ ULONG __cdecl _PhysIntoV86(ULONG PhysPage, ULONG VM, ULONG VMLinPgNum, ULONG nPa
 DWORD __cdecl _RegOpenKey(DWORD hKey, char *lpszSubKey, DWORD *lphKey);

 DWORD __cdecl _RegCloseKey(DWORD hKey);

 DWORD __cdecl _RegQueryValueEx(DWORD hKey, char *lpszValueName, DWORD *lpdwReserved, DWORD *lpdwType, BYTE *lpbData, DWORD *lpcbData);

-void __cdecl Begin_Critical_Section(ULONG Flags);

-void __cdecl End_Critical_Section();

+volatile void __cdecl Begin_Critical_Section(ULONG Flags);

+volatile void __cdecl End_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 bb2e967..499112f 100644
--- a/vxd_main.c
+++ b/vxd_main.c
@@ -550,6 +550,9 @@ DWORD __stdcall Device_IO_Control_proc(struct DIOCParams *params)
 		case OP_SVGA_OT_SETUP:

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

 			return 0;

+		case OP_SVGA_FLUSHCACHE:

+			SVGA_flushcache();

+			return 0;

 #endif /* SVGA */

 	}

 		

diff --git a/vxd_mouse_conv.h b/vxd_mouse_conv.h
index 62071fb..778e5c8 100644
--- a/vxd_mouse_conv.h
+++ b/vxd_mouse_conv.h
@@ -6,24 +6,28 @@ static BOOL calc_save(int x, int y)
 	

 	if(mouse_swap_x < 0)

 	{

-		mouse_swap_w -= x;

+		mouse_swap_w += mouse_swap_x;

 		mouse_swap_x = 0;

 	}

 	

 	if(mouse_swap_x + mouse_swap_w > hda->width)

+	{

 		mouse_swap_w = hda->width - mouse_swap_x;

+	}

 	

 	mouse_swap_h = mouse_h;

 	mouse_swap_y = y - mouse_pointy;

 	

 	if(mouse_swap_y < 0)

 	{

-		mouse_swap_h -= y;

+		mouse_swap_h += mouse_swap_y;

 		mouse_swap_y = 0;

 	}

 	

 	if(mouse_swap_y + mouse_swap_h > hda->height)

+	{

 		mouse_swap_h = hda->height - mouse_swap_y;

+	}

 	

 	if(mouse_swap_w <= 0 || mouse_swap_h <= 0)

 	{

diff --git a/vxd_svga.c b/vxd_svga.c
index faa445d..6484789 100644
--- a/vxd_svga.c
+++ b/vxd_svga.c
@@ -63,7 +63,6 @@ THE SOFTWARE.
 #pragma pack(1)

 typedef struct _cb_enable_t

 {

-//	SVGACBHeader       cbheader;

 	uint32             cmd;

 	SVGADCCmdStartStop cbstart;

 } cb_enable_t;

@@ -84,7 +83,7 @@ static BOOL cb_context0 = FALSE;
 

 static BOOL SVGA_is_valid = FALSE;

 

-static SVGACBHeader *last_cb = NULL;

+static volatile SVGACBHeader *last_cb = NULL;

 

 /* temp for reading registry */

 static union

@@ -139,10 +138,8 @@ static void SVGA_Sync()
 	SVGA_WriteReg(SVGA_REG_SYNC, 1);

 }

 

-static void SVGA_Flush_CB()

+static void SVGA_Flush_CB_critical()

 {

-	Begin_Critical_Section(0);

-	

 	/* wait for actual CB */

 	if(last_cb != NULL)

 	{

@@ -154,7 +151,12 @@ static void SVGA_Flush_CB()
 	

 	/* drain FIFO */

 	SVGA_Flush();

-	

+}

+

+static void SVGA_Flush_CB()

+{

+	Begin_Critical_Section(0);

+	SVGA_Flush_CB_critical();

 	End_Critical_Section();

 }

 

@@ -386,8 +388,7 @@ DWORD *SVGA_CMB_alloc()
 

 void SVGA_CMB_free(DWORD *cmb)

 {

-	SVGACBHeader *cb = (SVGACBHeader *)cmb;

-	--cb;

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

 	

 	Begin_Critical_Section(0);

 	

@@ -475,6 +476,7 @@ void SVGA_CMB_submit(DWORD FBPTR cmb, DWORD cmb_size, SVGA_CMB_status_t FBPTR st
 			cb->id.hi  = cb_next_id.hi;

 			cb->length = cmb_size;

 			

+#if 0

 			/* wait for last CB to complete */

 			if(last_cb && cb != last_cb)

 			{

@@ -483,12 +485,11 @@ void SVGA_CMB_submit(DWORD FBPTR cmb, DWORD cmb_size, SVGA_CMB_status_t FBPTR st
 					SVGA_Sync();

 				}

 			}

+#endif

 			

-			//Begin_Critical_Section(0);

 			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);

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

 			SVGA_Sync();

-			//End_Critical_Section();

 			

 			last_cb = cb;

 	

@@ -543,8 +544,6 @@ void SVGA_CMB_submit(DWORD FBPTR cmb, DWORD cmb_size, SVGA_CMB_status_t FBPTR st
 		

 		dwords += 2;

 		

-		//Begin_Critical_Section(0);

-		

 		nextCmd = gSVGA.fifoMem[SVGA_FIFO_NEXT_CMD];

 		max     = gSVGA.fifoMem[SVGA_FIFO_MAX];

 		min     = gSVGA.fifoMem[SVGA_FIFO_MIN];

@@ -564,8 +563,6 @@ void SVGA_CMB_submit(DWORD FBPTR cmb, DWORD cmb_size, SVGA_CMB_status_t FBPTR st
 			dwords--;

 		}

 		

-		//End_Critical_Section();

-		

 		if(flags & SVGA_CB_SYNC)

 		{

 			SVGA_fence_wait(fence);

@@ -967,21 +964,17 @@ BOOL SVGA_region_create(SVGA_region_info_t *rinfo)
 	

 	spare_region_used:

 	

-	End_Critical_Section();

-	

 	// JH: no need here

 	//SVGA_Flush_CB();

 	

 	if(rinfo->region_id <= SVGA_ReadReg(SVGA_REG_GMR_MAX_IDS))

 	{

 		/* register GMR */

-		Begin_Critical_Section(0);

 		SVGA_WriteReg(SVGA_REG_GMR_ID, rinfo->region_id);

 		SVGA_WriteReg(SVGA_REG_GMR_DESCRIPTOR, rinfo->region_ppn);

 		SVGA_Sync(); // notify register change

-		End_Critical_Section();

 		

-		SVGA_Flush_CB();

+		SVGA_Flush_CB_critical();

 	}

 	else

 	{

@@ -989,6 +982,8 @@ BOOL SVGA_region_create(SVGA_region_info_t *rinfo)
 		rinfo->region_ppn = 0;

 	}

 	

+	End_Critical_Section();

+	

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

 	

 	return TRUE;

@@ -1002,14 +997,13 @@ void SVGA_region_free(SVGA_region_info_t *rinfo)
 {

 	BYTE *free_ptr = (BYTE*)rinfo->address;

 	

-	SVGA_Flush_CB();

-	

 	Begin_Critical_Section(0);

+	SVGA_Sync();

+	SVGA_Flush_CB_critical();

+	

 	SVGA_WriteReg(SVGA_REG_GMR_ID, rinfo->region_id);

 	SVGA_WriteReg(SVGA_REG_GMR_DESCRIPTOR, 0);

 	SVGA_Sync(); // notify register change

-		

-	//SVGA_Flush_CB();

 	

 	/* save 2 large regions */

 	if(rinfo->size >= SPARE_REGION_LARGE_SIZE)

@@ -1073,8 +1067,53 @@ void SVGA_region_free(SVGA_region_info_t *rinfo)
 	rinfo->mob_address    = NULL;

 	rinfo->region_ppn     = 0;

 	rinfo->mob_ppn        = 0;

+	

+}

+

+/**

+ * Destroy saved regions and free memory

+ *

+ **/

+void SVGA_flushcache()

+{

+	int i;

+	

+	Begin_Critical_Section(0);

+	

+	for(i = 0; i < SPARE_REGIONS; i++)

+	{

+		if(spare_regions[i].used != FALSE)

+		{

+			SVGA_region_info_t *rinfo = &spare_regions[i].region;

+			BYTE *free_ptr = (BYTE*)rinfo->address;

+			

+			spare_regions[i].used = FALSE;

+			

+			if(rinfo->region_address != NULL)

+			{

+				dbg_printf(dbg_pagefree, rinfo->region_address);

+				_PageFree((PVOID)rinfo->region_address, 0);

+			}

+			else

+			{

+				free_ptr -= P_SIZE;

+			}

+			

+			if(rinfo->mob_address != NULL)

+			{

+				dbg_printf(dbg_pagefree, rinfo->mob_address);

+				_PageFree((PVOID)rinfo->mob_address, 0);

+			}

+			

+			dbg_printf(dbg_pagefree, free_ptr);

+			_PageFree((PVOID)free_ptr, 0);

+		}

+	}

+

+	End_Critical_Section();

 }

 

+

 /**

  * GPU10: start context0

  *

@@ -1285,14 +1324,22 @@ BOOL SVGA_setmode(DWORD w, DWORD h, DWORD bpp)
 		return FALSE;

 	}

 	

+	/* free chached regions */

+	SVGA_flushcache();

+	

+	FBHDA_access_begin(0);

+	

+	Begin_Critical_Section(0);

+	

 	/* Make sure, that we drain full FIFO */

-	SVGA_Flush_CB(); 

+	SVGA_Sync();

+	SVGA_Flush_CB_critical(); 

       

 	/* stop command buffer context 0 */

 	SVGA_CB_stop();

             

 	SVGA_SetModeLegacy(w, h, bpp); /* setup by legacy registry */

-	SVGA_Flush_CB(); /* make sure, that is really set */

+	SVGA_Flush_CB_critical(); /* make sure, that is really set */

 	

 	has3D = SVGA3D_Init();

       

@@ -1300,9 +1347,9 @@ BOOL SVGA_setmode(DWORD w, DWORD h, DWORD bpp)
 	if(SVGA_hasAccelScreen())

 	{

 		SVGA_defineScreen(w, h, bpp);

-		SVGA_Flush_CB();

+		SVGA_Flush_CB_critical();

 	}

-      

+	

 	/* start command buffer context 0 */

 	SVGA_CB_start();

      

@@ -1326,7 +1373,7 @@ BOOL SVGA_setmode(DWORD w, DWORD h, DWORD bpp)
       

 	SVGA_WriteReg(SVGA_REG_ENABLE, TRUE);

 	SVGA_Sync();

-	SVGA_Flush_CB();

+	SVGA_Flush_CB_critical();

 	

 	hda->width   = SVGA_ReadReg(SVGA_REG_WIDTH);

 	hda->height  = SVGA_ReadReg(SVGA_REG_HEIGHT);

@@ -1345,6 +1392,10 @@ BOOL SVGA_setmode(DWORD w, DWORD h, DWORD bpp)
 	{

 		hda->flags &= ~((DWORD)FB_ACCEL_VMSVGA3D);

 	}

+	

+	End_Critical_Section();

+	

+	FBHDA_access_end(0);

 

   return TRUE;

 }