KDC worker processes project

Greg Hudson ghudson at MIT.EDU
Fri Sep 17 15:20:58 EDT 2010


On Fri, 2010-09-17 at 14:34 -0400, Ken Raeburn wrote:
> I find the disabling of automatic handling of network reconfiguration
> based on platform plus whether you're using worker processes rather
> unappealing

It's not great, but:

(1) Fundamentally, the onus is on the operating system here.  The basic
BSD sockets API makes it impossible to implement RFC 1123 section 2.3
semantics for UDP servers, and there's a limit to how much complexity
every UDP server should have to re-implement to work around that
failure--especially given that there is a well-known extension (pktinfo)
making that complexity unnecessary, and this extension is implemented on
some of the most popular Unix-like platforms.

(2) The overlap between "needs the scalability of worker processes" and
"is expected to add network interfaces the KDC cares about" seems
nonexistent.  If you're running a KDC serving a large volume of
requests, it's unlikely that you're doing so over a DHCP network
interface.

> It shouldn't be that tough to do.

It's too tough for me to do in the timeline we have available.  I won't
object if you want to do the work.

The plan you suggest is not simple to carry out.  Currently the
supervisor spends its time a wait() loop in the KDC's main.c not the
select() loop of net-server.c.





More information about the krbdev mailing list