thols, thanks for taking a look. I did not realize the thing about 1/4 threshold value, just that buflen/2 froze whereas 128 worked. In the original LabVIEW example I started with, it reads 1/2 of the buffer per call. I've included the exact VI in this post, along with the minimally modified original example to show the zero-timeout behavior.
I cannot look at your example since I only have up to 2017, but I already know it is a dll bug so I don't need convincing.
