[PATCH] Ensure blitter quiescience before reading pixels from the framebuffer

[Bernardo Innocenti]

On my laptop, I sometimes see trapezoids and glyphs painted in
the wrong color.  It happens mostly in Gtk widgets, but I could
reproduce the problem also with a small cairo testcase.

I can't test this fix right now because I'm travelling and I
can't afford to download and build all the xserver prerequisites,
but the race I'm fixing here would perfectly justify the symptoms
I see.

Unfortunately, calling XSync() at this point will stall the GPU
and has therefore a huge impact on performance.  But I can't
think of another way to protect access.  If we're always going
to accesse these 1x1 repeated pixmaps with the CPU, a better
fix would be not uploading them in the first place.

diff --git a/hw/xfree86/xaa/xaaPict.c b/hw/xfree86/xaa/xaaPict.c
index 55f0f6e..52852df 100644
--- a/hw/xfree86/xaa/xaaPict.c
+++ b/hw/xfree86/xaa/xaaPict.c
@@ -237,6 +237,7 @@ XAADoComposite (
 
 	if((pSrc->pDrawable->width == 1) && (pSrc->pDrawable->height == 1)) {
 	   CARD16 red, green, blue, alpha;
+           XAASync();
            CARD32 pixel =
                 *((CARD32*)(((PixmapPtr)(pSrc->pDrawable))->devPrivate.ptr));
 
@@ -594,8 +595,9 @@ XAADoGlyphs (CARD8         op,
        !(infoRec->WriteBitmapFlags & NO_TRANSPARENCY))
     {
 	CARD16 red, green, blue, alpha;
-	CARD32 pixel =
-                *((CARD32*)(((PixmapPtr)(pSrc->pDrawable))->devPrivate.ptr));
+	CARD32 pixel;
+	XaaSync();
+	pixel = *((CARD32*)(((PixmapPtr)(pSrc->pDrawable))->devPrivate.ptr));
 	CARD32 *bits, *pntr, *pnt;
 	int x, y, i, n, left, top, right, bottom, width, height, pitch;
 	int L, T, R, B, X, Y, h, w, dwords, row, column, nbox;
-- 
1.5.2.3


-- 
   // Bernardo Innocenti
 \X/  http://www.codewiz.org/