FYI:
Kris Kennaway and I have been working on diagnosing and addressing network
stack contention and scheduling issues for high performance DNS serving with
nsd and bind9. One of the changes in the pipeline is converting from using a
mutex to protect the pcbinfo and inpcb data structures to an rwlock, so that
UDP, which doesn't require exclusion for either of those structures in a
number of key cases, doesn't have to experience as high contention. TCP would
continue to use them only exclusively [for now]. This is a heads up that this
change will likely end up in HEAD in the next month or two, but we're still
working through some issues. This may well be something that can be MFC'd as
rwlocks are intentionally fairly layout compatible with mutexes, and these
locks and the data structures they protect are not considered public
structures from the perspective of most kernel modules (although there are
potentially issues with firewalls that reach up the stack to look at th
connection table).
One of those issues is that we need to demonstrate to ourselves that exclusive
access contention is managed as well with rwlocks as with sleep mutexes, as
these locks would continue to be fairly highly contended in TCP. The other
issue is that rwlocks don't support full priority propagation for reader
access, although Jeff Roberson has recently improved fairness to writers with
many readers.
Robert N M Watson
Computer Laboratory
University of Cambridge
___
freebsd-net@freebsd.org mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-net
To unsubscribe, send any mail to [EMAIL PROTECTED]
