I can provoke this bug with a simple drmModeSetCursor() or drmModeMoveCursor().
The cursor images have a horizontal black stripe (not always). Position varies
on my machine. Using 3D acceleration without cursors works perfectly well
(although starting mplayer caused a deadlock on my machine).

It's also a nv50 card.

I tried bisecting it and it turns out the memory-manager rewrite caused it (as
reported on IRC). I was unable to revert the commit on top of 3.7 as it is
quite complex.