On Wed, May 21, 2014 at 03:52:39PM -0700, Ben Pfaff wrote: > On Thu, May 22, 2014 at 10:25:04AM +1200, Joe Stringer wrote: > > So just to make sure I'm understanding correctly, > > > > 1) revalidator calls dpif_flow_dump_next_may_destroy_keys(), which > > indicates that it will not destroy the buffer. > > 2) revalidator calls dpif_flow_dump_next(), which dumps a flow without > > actions. The most likely case is that this is the last flow in the batch. > > 3) dpif_flow_dump_next() attempts to retrieve the actions, but fails. > > 4) dpif_flow_dump_next() continues in the loop, and fetches a new batch via > > nl_dump_next(), breaking the guarantee that may_destroy_keys() gave. > > > > Then, the proposed solution is to eliminate flow deletion batching. > > Yes, that's correct. > > > I have a couple of thoughts:- > > > > Could we fix it instead by changing the > > dpif_linux_flow_dump_next_may_destroy_keys() to copy the ofpbuf, seek ahead > > and check if actions exist so that it provides the proper guarantee? > > Hmm. That is an approach that I did not consider. It would be > better, so I'll see what I can do.
I sent out v2 that adopts this idea. _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
