diff --git a/dlls/gdi32/bitblt.c b/dlls/gdi32/bitblt.c
index aa1ba0dda1e..a1b4b6fa393 100644
--- a/dlls/gdi32/bitblt.c
+++ b/dlls/gdi32/bitblt.c
@@ -567,7 +567,7 @@ BOOL WINAPI NtGdiPatBlt( HDC hdc, INT left, INT top, INT width, INT height, DWOR
 
 
 /***********************************************************************
- *           BitBlt    (GDI32.@)
+ *           NtGdiBitBlt    (win32u.@)
  */
 BOOL WINAPI NtGdiBitBlt( HDC hdc_dst, INT x_dst, INT y_dst, INT width, INT height,
                          HDC hdc_src, INT x_src, INT y_src, DWORD rop, DWORD bk_color, FLONG fl )
@@ -638,12 +638,11 @@ BOOL WINAPI NtGdiStretchBlt( HDC hdcDst, INT xDst, INT yDst, INT widthDst, INT h
 #define BKGND_ROP3(ROP4)	(ROP3Table[((ROP4)>>24) & 0xFF])
 
 /***********************************************************************
- *           MaskBlt [GDI32.@]
+ *           NtGdiMaskBlt    (win32u.@)
  */
-BOOL WINAPI MaskBlt(HDC hdcDest, INT nXDest, INT nYDest,
-                        INT nWidth, INT nHeight, HDC hdcSrc,
-			INT nXSrc, INT nYSrc, HBITMAP hbmMask,
-			INT xMask, INT yMask, DWORD dwRop)
+BOOL WINAPI NtGdiMaskBlt( HDC hdcDest, INT nXDest, INT nYDest, INT nWidth, INT nHeight,
+                          HDC hdcSrc, INT nXSrc, INT nYSrc, HBITMAP hbmMask,
+                          INT xMask, INT yMask, DWORD dwRop, DWORD bk_color )
 {
     HBITMAP hBitmap1, hOldBitmap1, hBitmap2, hOldBitmap2;
     HDC hDC1, hDC2;
@@ -782,9 +781,10 @@ BOOL WINAPI MaskBlt(HDC hdcDest, INT nXDest, INT nYDest,
     };
 
     if (!hbmMask)
-	return BitBlt(hdcDest, nXDest, nYDest, nWidth, nHeight, hdcSrc, nXSrc, nYSrc, FRGND_ROP3(dwRop));
+        return NtGdiBitBlt( hdcDest, nXDest, nYDest, nWidth, nHeight, hdcSrc,
+                            nXSrc, nYSrc, FRGND_ROP3(dwRop), bk_color, 0 );
 
-    hbrMask = CreatePatternBrush(hbmMask);
+    hbrMask = NtGdiCreatePatternBrushInternal( hbmMask, FALSE, FALSE );
     hbrDst = NtGdiSelectBrush( hdcDest, get_stock_object(NULL_BRUSH) );
 
     /* make bitmap */
@@ -793,9 +793,9 @@ BOOL WINAPI MaskBlt(HDC hdcDest, INT nXDest, INT nYDest,
     hOldBitmap1 = NtGdiSelectBitmap(hDC1, hBitmap1);
 
     /* draw using bkgnd rop */
-    BitBlt(hDC1, 0, 0, nWidth, nHeight, hdcDest, nXDest, nYDest, SRCCOPY);
+    NtGdiBitBlt( hDC1, 0, 0, nWidth, nHeight, hdcDest, nXDest, nYDest, SRCCOPY, 0, 0 );
     hbrTmp = NtGdiSelectBrush(hDC1, hbrDst);
-    BitBlt(hDC1, 0, 0, nWidth, nHeight, hdcSrc, nXSrc, nYSrc, BKGND_ROP3(dwRop));
+    NtGdiBitBlt( hDC1, 0, 0, nWidth, nHeight, hdcSrc, nXSrc, nYSrc, BKGND_ROP3(dwRop), 0, 0 );
     NtGdiSelectBrush(hDC1, hbrTmp);
 
     /* make bitmap */
@@ -804,18 +804,19 @@ BOOL WINAPI MaskBlt(HDC hdcDest, INT nXDest, INT nYDest,
     hOldBitmap2 = NtGdiSelectBitmap(hDC2, hBitmap2);
 
     /* draw using foregnd rop */
-    BitBlt(hDC2, 0, 0, nWidth, nHeight, hdcDest, nXDest, nYDest, SRCCOPY);
+    NtGdiBitBlt( hDC2, 0, 0, nWidth, nHeight, hdcDest, nXDest, nYDest, SRCCOPY, 0, 0 );
     hbrTmp = NtGdiSelectBrush(hDC2, hbrDst);
-    BitBlt(hDC2, 0, 0, nWidth, nHeight, hdcSrc, nXSrc, nYSrc, FRGND_ROP3(dwRop));
+    NtGdiBitBlt( hDC2, 0, 0, nWidth, nHeight, hdcSrc, nXSrc, nYSrc, FRGND_ROP3(dwRop), 0, 0 );
 
     /* combine both using the mask as a pattern brush */
     NtGdiSelectBrush(hDC2, hbrMask);
     SetBrushOrgEx(hDC2, -xMask, -yMask, NULL);
-    BitBlt(hDC2, 0, 0, nWidth, nHeight, hDC1, 0, 0, 0xac0744 ); /* (D & P) | (S & ~P) */ 
+    /* (D & P) | (S & ~P) */
+    NtGdiBitBlt(hDC2, 0, 0, nWidth, nHeight, hDC1, 0, 0, 0xac0744, 0, 0 );
     NtGdiSelectBrush(hDC2, hbrTmp);
 
     /* blit to dst */
-    BitBlt(hdcDest, nXDest, nYDest, nWidth, nHeight, hDC2, 0, 0, SRCCOPY);
+    NtGdiBitBlt( hdcDest, nXDest, nYDest, nWidth, nHeight, hDC2, 0, 0, SRCCOPY, bk_color, 0 );
 
     /* restore all objects */
     NtGdiSelectBrush(hdcDest, hbrDst);
diff --git a/dlls/gdi32/emfdc.c b/dlls/gdi32/emfdc.c
index 84069bdba3a..66c87ef9412 100644
--- a/dlls/gdi32/emfdc.c
+++ b/dlls/gdi32/emfdc.c
@@ -1388,6 +1388,102 @@ BOOL EMFDC_StretchBlt( DC_ATTR *dc_attr, INT x_dst, INT y_dst, INT width_dst, IN
                               height_src, rop, EMR_STRETCHBLT );
 }
 
+BOOL EMFDC_MaskBlt( DC_ATTR *dc_attr, INT x_dst, INT y_dst, INT width_dst, INT height_dst,
+                    HDC hdc_src, INT x_src, INT y_src, HBITMAP mask,
+                    INT x_mask, INT y_mask, DWORD rop )
+{
+    unsigned char mask_info_buffer[FIELD_OFFSET(BITMAPINFO, bmiColors[256])];
+    BITMAPINFO *mask_bits_info = (BITMAPINFO *)mask_info_buffer;
+    struct emf *emf = dc_attr->emf;
+    BITMAPINFO mask_info = {{ sizeof( mask_info.bmiHeader ) }};
+    BITMAPINFO src_info = {{ sizeof( src_info.bmiHeader ) }};
+    HBITMAP bitmap, blit_bitmap = NULL, mask_bitmap = NULL;
+    UINT bmi_size, size, mask_info_size = 0;
+    EMRMASKBLT *emr = NULL;
+    BITMAPINFO *bmi;
+    HDC blit_dc, mask_dc = NULL;
+    BOOL ret = FALSE;
+
+    if (!rop_uses_src( rop ))
+        return EMFDC_PatBlt( dc_attr, x_dst, y_dst, width_dst, height_dst, rop );
+
+    if (!(bitmap = GetCurrentObject( hdc_src, OBJ_BITMAP ))) return FALSE;
+    blit_dc = hdc_src;
+    blit_bitmap = bitmap;
+    if (!(bmi_size = get_bitmap_info( &blit_dc, &blit_bitmap, &src_info ))) return FALSE;
+
+    if (mask)
+    {
+        mask_dc = hdc_src;
+        mask_bitmap = mask;
+        if (!(mask_info_size = get_bitmap_info( &mask_dc, &mask_bitmap, &mask_info ))) goto err;
+        if (mask_info.bmiHeader.biBitCount == 1)
+            mask_info_size = sizeof(BITMAPINFOHEADER); /* don't include colors */
+    }
+    else mask_info.bmiHeader.biSizeImage = 0;
+
+    size = sizeof(*emr) + bmi_size + src_info.bmiHeader.biSizeImage +
+        mask_info_size + mask_info.bmiHeader.biSizeImage;
+
+    if (!(emr = HeapAlloc(GetProcessHeap(), 0, size))) goto err;
+
+    emr->emr.iType = EMR_MASKBLT;
+    emr->emr.nSize = size;
+    emr->rclBounds.left = x_dst;
+    emr->rclBounds.top = y_dst;
+    emr->rclBounds.right = x_dst + width_dst - 1;
+    emr->rclBounds.bottom = y_dst + height_dst - 1;
+    emr->xDest = x_dst;
+    emr->yDest = y_dst;
+    emr->cxDest = width_dst;
+    emr->cyDest = height_dst;
+    emr->dwRop = rop;
+    emr->xSrc = x_src;
+    emr->ySrc = y_src;
+    NtGdiGetTransform( hdc_src, 0x204, &emr->xformSrc );
+    emr->crBkColorSrc = GetBkColor( hdc_src );
+    emr->iUsageSrc = DIB_RGB_COLORS;
+    emr->offBmiSrc = sizeof(*emr);
+    emr->cbBmiSrc = bmi_size;
+    emr->offBitsSrc = emr->offBmiSrc + bmi_size;
+    emr->cbBitsSrc = src_info.bmiHeader.biSizeImage;
+    emr->xMask = x_mask;
+    emr->yMask = y_mask;
+    emr->iUsageMask = DIB_PAL_MONO;
+    emr->offBmiMask = mask_info_size ? emr->offBitsSrc + emr->cbBitsSrc : 0;
+    emr->cbBmiMask = mask_info_size;
+    emr->offBitsMask = emr->offBmiMask + emr->cbBmiMask;
+    emr->cbBitsMask = mask_info.bmiHeader.biSizeImage;
+
+    bmi = (BITMAPINFO *)((char *)emr + emr->offBmiSrc);
+    bmi->bmiHeader = src_info.bmiHeader;
+    ret = GetDIBits( blit_dc, blit_bitmap, 0, src_info.bmiHeader.biHeight,
+                     (char *)emr + emr->offBitsSrc, bmi, DIB_RGB_COLORS );
+    if (!ret) goto err;
+
+    if (mask_info_size)
+    {
+        mask_bits_info->bmiHeader = mask_info.bmiHeader;
+        ret = GetDIBits( blit_dc, mask_bitmap, 0, mask_info.bmiHeader.biHeight,
+                         (char *)emr + emr->offBitsMask, mask_bits_info, DIB_RGB_COLORS );
+        if (ret) memcpy( (char *)emr + emr->offBmiMask, mask_bits_info, mask_info_size );
+    }
+
+    if (ret)
+    {
+        ret = emfdc_record( emf, (EMR *)emr );
+        if (ret) emfdc_update_bounds( emf, &emr->rclBounds );
+    }
+
+err:
+    HeapFree( GetProcessHeap(), 0, emr );
+    if (mask_bitmap != mask) DeleteObject( mask_bitmap );
+    if (mask_dc != hdc_src) DeleteObject( mask_dc );
+    if (blit_bitmap != bitmap) DeleteObject( blit_bitmap );
+    if (blit_dc != hdc_src) DeleteDC( blit_dc );
+    return ret;
+}
+
 BOOL EMFDC_StretchDIBits( DC_ATTR *dc_attr, INT x_dst, INT y_dst, INT width_dst, INT height_dst,
                           INT x_src, INT y_src, INT width_src, INT height_src, const void *bits,
                           const BITMAPINFO *info, UINT usage, DWORD rop )
diff --git a/dlls/gdi32/gdi_private.h b/dlls/gdi32/gdi_private.h
index 2dddf361d5b..2f0a6b4c0d7 100644
--- a/dlls/gdi32/gdi_private.h
+++ b/dlls/gdi32/gdi_private.h
@@ -203,6 +203,9 @@ extern BOOL EMFDC_IntersectClipRect( DC_ATTR *dc_attr, INT left, INT top, INT ri
                                      INT bottom ) DECLSPEC_HIDDEN;
 extern BOOL EMFDC_InvertRgn( DC_ATTR *dc_attr, HRGN hrgn ) DECLSPEC_HIDDEN;
 extern BOOL EMFDC_LineTo( DC_ATTR *dc_attr, INT x, INT y ) DECLSPEC_HIDDEN;
+extern BOOL EMFDC_MaskBlt( DC_ATTR *dc_attr, INT x_dst, INT y_dst, INT width_dst, INT height_dst,
+                           HDC hdc_src, INT x_src, INT y_src, HBITMAP mask,
+                           INT x_mask, INT y_mask, DWORD rop ) DECLSPEC_HIDDEN;
 extern BOOL EMFDC_ModifyWorldTransform( DC_ATTR *dc_attr, const XFORM *xform,
                                         DWORD mode ) DECLSPEC_HIDDEN;
 extern BOOL EMFDC_MoveTo( DC_ATTR *dc_attr, INT x, INT y ) DECLSPEC_HIDDEN;
diff --git a/dlls/gdi32/gdidc.c b/dlls/gdi32/gdidc.c
index d5645595984..25ef85f2db5 100644
--- a/dlls/gdi32/gdidc.c
+++ b/dlls/gdi32/gdidc.c
@@ -1499,6 +1499,23 @@ BOOL WINAPI StretchBlt( HDC hdc, INT x_dst, INT y_dst, INT width_dst, INT height
                             height_src, rop, 0 /* FIXME */ );
 }
 
+/***********************************************************************
+ *           MaskBlt [GDI32.@]
+ */
+BOOL WINAPI MaskBlt( HDC hdc, INT x_dst, INT y_dst, INT width_dst, INT height_dst,
+                     HDC hdc_src, INT x_src, INT y_src, HBITMAP mask,
+                     INT x_mask, INT y_mask, DWORD rop )
+{
+    DC_ATTR *dc_attr;
+
+    if (!(dc_attr = get_dc_attr( hdc ))) return FALSE;
+    if (dc_attr->emf && !EMFDC_MaskBlt( dc_attr, x_dst, y_dst, width_dst, height_dst,
+                                        hdc_src, x_src, y_src, mask, x_mask, y_mask, rop ))
+        return FALSE;
+    return NtGdiMaskBlt( hdc, x_dst, y_dst, width_dst, height_dst, hdc_src, x_src, y_src,
+                         mask, x_mask, y_mask, rop, 0 /* FIXME */ );
+}
+
 /******************************************************************************
  *           GdiAlphaBlend   (GDI32.@)
  */
diff --git a/include/ntgdi.h b/include/ntgdi.h
index 9e07268a307..cb7b9639e20 100644
--- a/include/ntgdi.h
+++ b/include/ntgdi.h
@@ -344,6 +344,9 @@ BOOL     WINAPI NtGdiFontIsLinked( HDC hdc );
 INT      WINAPI NtGdiIntersectClipRect( HDC hdc, INT left, INT top, INT right, INT bottom );
 BOOL     WINAPI NtGdiInvertRgn( HDC hdc, HRGN hrgn );
 BOOL     WINAPI NtGdiLineTo( HDC hdc, INT x, INT y );
+BOOL     WINAPI NtGdiMaskBlt( HDC hdc, INT x_dst, INT y_dst, INT width_dst, INT height_dst,
+                              HDC hdc_src, INT x_src, INT y_src, HBITMAP mask,
+                              INT x_mask, INT y_mask, DWORD rop, DWORD bk_color );
 BOOL     WINAPI NtGdiModifyWorldTransform( HDC hdc, const XFORM *xform, DWORD mode );
 BOOL     WINAPI NtGdiMoveTo( HDC hdc, INT x, INT y, POINT *pt );
 INT      WINAPI NtGdiOffsetClipRgn( HDC hdc, INT x, INT y );