On Thu, Apr 19, 2012 at 5:13 PM, Simon Horman <ho...@verge.net.au> wrote: > On Thu, Apr 19, 2012 at 01:29:08PM -0700, Jesse Gross wrote: >> On Wed, Apr 18, 2012 at 9:50 PM, Simon Horman <ho...@verge.net.au> wrote: >> > It seems to me that some changes are needed to the tunneling core code, >> > in particular handle_offloads(), to allow GSO skbs to be passed >> > unsegmented to STT. Perhaps a new mutable->flags flag is in order? >> >> I agree that there definitely needs to be additional information given >> to the tunneling core to enable it to decide whether or not to emulate >> offloads. However, I think the information belongs in tnl_ops instead >> of mutable->flags since it is shared by all tunnels of a given type. > > Thanks, I will look into that. > > At this point I am thinking of a single flag to indicate that > the protocol handler is able to handle GSO skbs. I'll see about preparing > a patch to implement that. But perhaps a single flag is insufficient > for some cases?
I think it needs to be a little more sophisticated that that. In the cases that where you currently drop the packet due to being incompatible with STT (such as an extra level of vlan tag, L4 offset too large, etc.) we should do software segmentation to avoid the need to offload over STT. Also not all types of GSO types are supported over STT (such as FCoE or future ones that don't exist today). Probably a function to evaluate a packet for offloading is more appropriate. _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
