D3D9 Interoperability issue driver lock-down when invoking cuD3D9UnmapResources()

I mapped a D3D9 A8R8B8G8 render-target into CUDA memory space, access it, and when I unmap the RT the whole system locked-down for 5-seconds.

The same programme runs perfectly on CUDA2.0, according to my test this issue only appears when:

  1. CUDA2.1(with its bundled driver) and 2.2 release(with 185.85)
  2. RT is format of A8R8G8B8 (formats like G16R16 do not trigger this issue)

Is any one encountered the same problem??