On Fri, Feb 7, 2014 at 11:55 AM, Alex Wang <al...@nicira.com> wrote: > Thanks Pravin, > > >> > >> > This patch also significantly simplifies the flow miss handling >> > code and brings slight improvement to flow setup rate. >> > >> Hi Alex, >> >> I have couple of high level comments: >> >> 1. Are we trying to solve fairness here? If yes then how does it work >> for tunnel ports? If not then this is about only removing dispatcher. >> Therefore can we have a per datapath pool of pids rather than per >> vport? > > > > This is only about removing dispatcher. But we want to try addressing the > fairness issue in userspace based on this implementation. The pids per > vport implementation provides per port fairness. If we use a pids pool, > the starvation may happen in the kernel (since the upcall from different > vport > may go to the same pid). And we need some fair queueing scheme in the > kernel. I think we want to avoid such complexity in kernel. > Right, this works for physical ports. What about tunnel port? In kernel there is one vport for a tunnel type. So with this patch there is one pool of pids for all tunnel-vports for give tunnel type (for example gre). That can cause starvation among different tunnels.
> >> 2. In any case we need to handle OVS_VPORT_ATTR_UPCALL_PID attribute >> in kernel. Since it is kernel interface. >> > > I didn't follow. Do you mean we need this to pass in the nl_sock pid? > I mean vport attribute OVS_VPORT_ATTR_UPCALL_PID can not be removed. We need to handle it in kernel datapath along with new attribute related to pid pool. _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
