Flush the cairo surface before we try to write to it.

skia/ext/bitmap_platform_device_linux.cc

Index: skia/ext/bitmap_platform_device_linux.cc
diff --git a/skia/ext/bitmap_platform_device_linux.cc b/skia/ext/bitmap_platform_device_linux.cc
index 60feb453d56a3865ff96f96588752ae4db15441f..034d6bdac749ee75eab405d6386a9b22b9f8f893 100644
--- a/skia/ext/bitmap_platform_device_linux.cc
+++ b/skia/ext/bitmap_platform_device_linux.cc
@@ -153,9 +153,12 @@ bool BitmapPlatformDevice::IsVectorial() {
 cairo_t* BitmapPlatformDevice::beginPlatformPaint() {
   data_->LoadConfig();
   cairo_t* cairo = data_->bitmap_context();
-  // Tell Cairo that we've (probably) modified its pixel buffer without
-  // its knowledge.
-  cairo_surface_mark_dirty(cairo_get_target(cairo));
+  cairo_surface_t* surface = cairo_get_target(cairo);
+  // Tell cairo to flush anything it has pending.
+  cairo_surface_flush(surface);

[Evan Martin, 2010/12/14 22:27:11]

Doesn't this belong in endPlatformPaint()?
It seems from the docs that you're supposed to call it before using non-cairo
APIs to work on the surface... I guess that's what the mark_dirty is for,
though, huh.

[vandebo (ex-Chrome), 2010/12/14 22:47:53]

I read the Docs the other way around.  flush needs to be called before you
modify the surface directly (cairo may have things it hasn't finalized to the
surface).  Then, if you modify the surface, you need to call mark_dirty before
making further cairo calls so that cairo knows to invalidate any cached state it
has.  Asserts in new versions of mark_dirty reinforce this interpretation.

Linux doesn't currently have an endPlatformPaint.  I assume this is ok and
currently works because mark_dirty just flushes cairo caches and doesn't
recompute anything.

+  // Tell Cairo that we (probably) modified (actually, will modify) its pixel
+  // buffer directly.
+  cairo_surface_mark_dirty(surface);
   return cairo;
 }