* Dynamically, by pre-allocating all the usb_request objects
that will be used; if the pre-allocated ones are in use,
then tx will throttle down. This behaves better under
heavy load.
* Statically, by pre-allocating fewer such requests in the
typical "no DMA queueing" case ... the best we can do is
make sure that when the next completion IRQ fires, the
controller already has a transfer ready. Having queues
deeper than two elements only helps if the CPU doesn't
need to start each transfer by hand (as with net2280).
- Diagnostics look more like other network driver diagnostics;
they use the network interface name.