> On Thu, May 29, 2014 at 1:21 AM, YAMAMOTO Takashi > <yamam...@valinux.co.jp> wrote: >>>> what would the new definition of DP_HASH action look like? >>>> >>> flow_hash_5tuple >> >> DP_HASH was defined as "datapath specific hash of the packet". >> my question is what it will be after your proposed change. > > It will be the same. > > User space computed hash values are not going to match with those > computed in the datapath. If the application require the value to match, > one way to implement it is to use a helper function to ask datapath to > compute it. > > DP_HASH is currently only used by bonding. User space computed hash may > lead to selecting a different port than the datapath port. The packet will > still be delivered, although may be delivered out of order, but this is > an inherent problem when datapath and slow path are in use for the > same flow. On the other hand, and most likely, if a flow is handled > entirely in slow path or in datapath, this implementation should work fine. > > In case we have future applications that require slow path > execution of DP_HASH to produce the same values produced by the > datapath, my proposed patch will not work. > > Does this make sense? What do you think?
i agree that it isn't a problem for bond usage. my concern is a lack of semantics description of the action. YAMAMOTO Takashi > > > >> >> YAMAMOTO Takashi _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
