Virtual PMD devices are not freed up

[sangjinhan]

According to @tootoonchian, in the deinit() function of the PMD driver should do something more than just calling rte_eth_dev_stop(). I suspect rte_eth_dev_stop() is enough for physical PMDs, but for virtual PMDs (e.g., bonding) we need explicit resource deallocation.

[tootoonchian]

Two reasons I do this (i.e., rte_eth_dev_close() + rte_eth_dev_detach() after rte_eth_dev_stop()):

1. As you said, with virtual PMDs stopping the port is not enough to attach to it again later.
2. I can safely change the physical port's driver after detaching if the driver has hotplug support.

We can guard the rte_eth_dev_detach() call with a check similar to lib/librte_ethdev/rte_ethdev.c > rte_eth_dev_is_detachable() (which is not public), or just call it and have its internal check fail in which case it logs an error and does nothing.

[sangjinhan]

Now the port is closed & detached if and only if it was dynamically attached (hot plugged).