OK, problem solved. Should have known it would be my code. The netdev is our own and on a previous update from upstream I missed the fact that the packet hash has to be initialised in the rxq_recv routine. All good now.
Tony On Tue, Sep 22, 2015 at 5:25 PM, Tony van der Peet <tony.vanderp...@gmail.com> wrote: > OK, hold that thought, I think I just worked it out for myself! Looks > like there is a calculation in emc_processing, which implies that the > hashes are already calculated? It's too late at the end of a long day. > Still something not quite right though, so I'll keep looking - any > further advice appreciated. > > Tony > > On Tue, Sep 22, 2015 at 5:18 PM, Tony van der Peet > <tony.vanderp...@gmail.com> wrote: >> In dpif-netdev.c, line 3264 (master branch), appears this comment (the >> routine is fast_path_processing). >> >> /* Key length is needed in all the cases, hash computed on demand. */ >> >> I have inspected this code quite a bit, and can't find where the >> hashes are calculated. In particular, this appears to cause a memory >> leak in associated emc code (flows with an extra reference on them are >> not being deleted). This is because the hash is not calculated when >> emc_insert is called. >> >> Could someone please point out where the hashes are calculated, or >> confirm that there might be an issue here? Or that my understanding is >> incorrect and we don't need those hash values? >> >> If there is a fix required, I have a reliable test that shows up this >> issue, so will be happy to test it out. >> >> Thanks >> Tony _______________________________________________ discuss mailing list discuss@openvswitch.org http://openvswitch.org/mailman/listinfo/discuss
