My planned work on networking stack
Andre Oppermann
andre at freebsd.org
Tue Mar 2 13:19:22 PST 2004
Bruce M Simpson wrote:
>
> On Tue, Mar 02, 2004 at 07:09:02PM +0300, Gleb Smirnoff wrote:
> > I do not insist that AS pathes in kernel are good idea. If you show me an
> > other way to get AS information when constructing netflow exports in kernel,
> > I'd be thankful. I'd be also thankful if you describe how policy routing can be
> > implemented while no AS info in kernel.
> > What do other FreeBSD networking withards think?
>
> I don't see any reason why we couldn't accept, for example, a 32-bit cookie
> for abuse by a userland daemon, with pid, as it pleases (via an rtmsg
> extension and PF_ROUTE). That is generic enough to provide the tie-in
> needed with the userland RIB and the kernel FIB.
Ugh, I'm happily running my accounting in userland via BPF/PCAP and it
adds only 2-3% CPU load. The BGP information I get from MRT routing
table dumps. Pretty slick stuff. We (Claudio and me) are preparing it
for public release later this week.
>From my experience here and a performance point of view there is no need
to do netflow and related accounting stuff in the kernel at all. Userland
is much more flexible.
> ABI breakage may occur, but I would consider that the PF_ROUTE code is in need
> of an overhaul anyway (see my mail to ru@ from some months ago on -current or
> -net with code able to panic a kernel through malformed rtmsg contents).
Please don't break the current RTM5 API. We will design a nice and much
more flexible RTM6 message format later this year. It needs a good deal
of deep thought and not be rushed just for the sake of it.
--
Andre
More information about the freebsd-net
mailing list