New IP header cleanup branch available for testing
My leaf repo now has a branch called 'iphdr'. Here's the git remote setup:
[remote "dillon"] url = git://leaf.dragonflybsd.org/~dillon/dragonfly.git fetch = +refs/heads/*:refs/remotes/dillon/*
$ git fetch dillon $ git branch iphdr dillon/iphdr $ git checkout iphdr
I have done basic testing with ipfw, pf, ipv4 fragmentation, ipv4 TCP, a little bridging, if_tap, and ipv4 UDP.
Tons of other things haven't been tested yet. ipv6, bpf filters, carp, more extensive ipfw2 and pf, and so forth. Donn't bother with IPSEC, it has other issues which will have to be solved in the next stage.
Only do testing if you have some experience with the interfaces in question.
This branch contains the network byte ordering and adjustment fixes for ip_len, ip_off, and a large chunk of the protosw cleanup. I am going to leave the branch alone except for bug fixes.
I am now starting work on another branch that will contain a rewrite of the whole ether/ip demux and dispatch mechanic. All protocols will become per-cpu and there will be just one protocol support thread for each cpu. NETISRs will go away and be replaced with a more direct netmsg queueing operation. The toeplitz hash will be integrated into the demux mechanic (probably integrated with ip_lengthcheck()), so the cpu switch will occur even before a protocol is handed the packet. Fast-forwarding will still occur in the demux and be able to bypass netmsg queueing.
Hopefully when all is said and done the combination of the stage 1 fixes now available and this stage 2 work will give us a huge degree of flexibility with regards to managing packet mbufs, including an ability to trivially requeue them (something IPSEC and other tunnel implementations need to be able to do).
This second stage will probably take a week at least.