[openib-general] Re: IPoIB question/problem

[Roland Dreier]

    Michael> It seems that we'll need to define hard_header_cache as
    Michael> well, then. And having this appears to have performance
    Michael> implications, affecting TCP code flow in a major way (for
    Michael> better or worse).  I was thinking that using the header
    Michael> cache may be a good replacement for using the free space
    Michael> in the ha field, but this would be a major change.  What
    Michael> do you think?

You're right, I misread the neighbour.c code a little bit.  We can't
just use the update method without the whole hard header cache thing.

I looked at using that stuff originally, but it didn't seem quite
suitable for what IPoIB needed.  I can't remember why right now
unfortunately.

    Michael> A more modest approach that I was considering: keep a
    Michael> copy of the gid as part of ipoib_neigh structure, and
    Michael> make sure that the gid didnt change before posting a
    Michael> packet.  This seems to work for some of our gen1 clients.

How does adding the 16-byte memcmp to the fast path affect
performance?  It's probably OK for at least a short-term fix, because
it's better to be a little slower than to completely lose packets.

 - R.