On Thu, Nov 14, 2013 at 03:28:25PM -0800, Joe Stringer wrote: > Previously, we tracked status changes for ofports on a per-device basis. > Each time in the main thread's loop, we would loop through all ofports > and manually check whether the status has changed for corresponding > devices. > > This patch shifts change_seq above the netdevice layer, with one atomic > variable tracking status change for all ports. In the average case where > ports are not constantly going up or down, this allows us to check > change_seq once per loop and not poll any ports. In the worst case, > execution is expected to be similar to how it is currently. > > As change_seq is no longer tracked per-device, it doesn't make sense to > cache this status in each ofport struct. As such, we shift this into the > ofproto struct. > > In a test environment of 5000 internal ports and 50 tunnel ports with > bfd, this reduces CPU usage from about 45% to about 35%. > > Signed-off-by: Joe Stringer <joestrin...@nicira.com>
I think we could delete netdev_dummy_changed() and netdev_vport_changed() entirely and just call netdev_notify() from the caller. I think that every netdev construct function now calls netdev_notify(). If so, I guess we could call it centrally from netdev_open(). I'd expect that bond_wait() should use netdev_seq_wait() instead of a direct comparison followed by poll_immediate_wake(). In the new comment here: + * although implementations should try to avoid this. netdev.c contains more + * information about this notification interface. */ I'd be more specific about where to look, naming the related functions. Otherwise this looks good, thanks. _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
