Preemptive threads with growable stacks. There was some discussion around getting segmented stacks into the kernel, but I'm not sure that's the best approach. There might have to be some novel work done in making contiguous stacks work in a shared address space.