On Mon, 2017-01-16 at 17:04 -0700, Andrew Collins wrote: > The fq_codel packet scheduler always regenerates the skb flow hash. Is there > any reason > to do this other than the recent hash perturbation additions? > > I ask because I have a case where an incoming set of TCP flows is > encapsulated in a > single ipsec tunnel, which is then classified on egress into a single flow by > fq_codel > resulting in poor behavior. > > Reusing the skb hash set in receive results in much better behavior, as the > value is > determined pre-encapsulation and flows end up being distributed more > naturally across > codel queues.
> > Is opportunistically using the pre-existing skb hash (if available) > problematic? I guess this would be fine. We never exported the 'perturbation' or allowed to change/set it. The only 'risk' would be having a buggy device setting a wrong L4 skb->hash, or buggy protocol messing with skb->hash for a given flow. > Is there a better way to manage flow separation in routed+encapsulated > traffic? Encapsulated traffic is fine, since flow dissector skips encap header(s) up to the L4 header. Problem is about encrypted traffic, since presumably this L4 header is opaque for flow dissector ;)
