Doing development on udlfb on Linux 2.6.32, crossed paths with these two quirks.
The usb_skeleton.c sample driver appears to run afoul of this problem, resulting in a warn_on during bulk transfer completion, if done the way the sample recommends. Because completion routines can be called with irqs disabled, usb_buffer_free on dma-ready buffers can’t be called from them — which pushed the udlfb implementation to diverge from the skeleton model and pre-allocate a small set of urbs, and free only on driver disconnect.
Calls to usb_bulk_msg (simple synchronous bulk out) appeared to be silently failing for largish transfers (generally 16-32K). Nasty to track down, since no errors are returned. Caused removal all use of usb_bulk_msg from udlfb, except for transactions that are known to be small.