Marek has been working on thread off-loading of the CS ioctl in the R600g driver in order to reduce kernel overhead. Marek expects this thread off-loading of the DRM command submission ioctl to cause up to a 15% performance increase for CPU-bound games/applications. The R600g code now also supports the new GEM_WAIT ioctl and also has greater thread-safety to prevent possible race conditions in the driver. There's also support for space checking, but the R600g driver isn't using that yet.
Two OpenGL games that are having noticeable performance improvements are Unigine Heaven and TORCS, according to Marek. His new "r600winsys2" branch is also not resulting in any new Piglit regressions. For now this new Radeon winsys is living in the Mesa branch, but it should be merged to master soon (certainly in time for Mesa 7.12).
More details can be found in Marek's mailing list announcement.