Break out RSS init path into sys/net; inet4 code in sys/netinet/; inet6 code in sys/netinet6/

[adrian]

Summary

The RSS code currently lives in sys/netinet/ and supports both IPv4 and IPv6.
But as part of the general network stack integration, stuff outside of netinet/
(eg, netisr) will need some information about the RSS setup and CPU mapping
so it can make decisions about CPU / RSS bucket allocations for netisr, CPU
pinning, later on it'll query NUMA topology and setup netisr accordingly, etc.

So:

• the stuff to do with RSS configuration now lives in sys/net/rss_config.[ch];
• the sysctl space is still net.inet.netisr for now, but that may change before 11-RELEASE is cut;
• sys/netinet/in_rss.[ch] has IPv4 specific code in it;
• sys/netinet6/in6_rss.[ch] has IPv6 specific code in it;
• the IPv6 RSS code now uses pointers instead of copying the src/dst IPv6 addresses around.

[bz]

I can't really review this in Phabricator. Is there any chance we could reduce some noise from that change, e.g., making the IPv6 faddr/laddr pointers doesn't really belong in here. If we can brea such stuff out, reviewing the rest is a lot easier.

sys/net/rss_config.c

32

This probably just wants to become $FreeBSD$

sys/net/rss_config.h

107

Aeh, while I can see the point, the function will not allow you to abstract the KPI/KBI away either this way (static inline ...). Just saying.

sys/netinet/in_pcbgroup.c

50

This is misordered

sys/netinet/ip_output.c

76

misordered

sys/netinet6/in6_pcbgroup.c

43–44

Outside #ifdef INET6

[adrian]

How about I do a separate smaller commit first to turn the ipv6 address fields into pointers; and -then- do the code shuffle?

[jfv]

Changes in my code are trivial, and fine by me.

[adrian]

Re-do the RSS code split; the IPv6 API change is already in -HEAD.

Delay putting the RSS hash configuration in a global uint32_t;
I'll do that in a separate pass after this work is in the tree.

[adrian]

Update to latest -HEAD - make RSS on ixl(4) compile.

[adrian]

All done; it's in -HEAD.