[openib-general] [PATCH] IB/ipoib: NAPI

[Roland Dreier]

    Shirley> I forgot to mention these NAPI parameters should be
    Shirley> tunable for different device drivers, like dev->weight,
    Shirley> or set up in lower driver.

    Michael> So we need something like poll_weight in struct
    Michael> ib_device, to give a hint on how expensive an interrupt
    Michael> is versus poll?  Seems to make sense, and actually might
    Michael> be useful for other ULPs.  Roland, what do you think?

How could a low-level driver possibly know the cost of an interrupt vs
polling a CQ?  It depends on the particular CPU/cache/chipset details
of the system and it might not even be the same from one PCI slot to
another.

If this value makes a real difference in practice, we can make it
tunable but I would like to see some hard benchmarks that show it
making a big difference one way or another.  But we have too many
knobs as it is so I'm inclined to just pick a value that works OK.

 - R.