On Wed, Mar 01, 2017 at 11:28:33AM -0700, Jason Gunthorpe wrote: > On Wed, Mar 01, 2017 at 04:02:08PM +0200, Erez Shitrit wrote: > > However, New features are exisiting in common network interfaces that > > are very hard to implement in IPoIB interfaces while it uses the RDMA > > layer, examples include TSS and RSS, tunneling offloads, and XDP. > > Rather than continuously porting IP network interface developments into > > the RDMA stack, we propose adding an abstract network data-path interfaces > > to RDMA devices. > > I should also say that we have discussed this for a while and I am > broadly in favor of this approach of delegating IPoIB skb transmit > directly to the low level driver. I agree. > > There are too many optimizations that do not fit sanely into verbs, > and the entire overhead of building a verbs wr and posting it then > translating that to a driver wr is unnecessary layering. Yes, Niranjana has more details but the idea with the new rdma_netdev is that the type defines what skbs the driver can expect and the low level driver does the correct optimizations if the rdma_netdev is used. I assume none of this applies to Connected Mode? > > Longer term, it might be quite nice if the part verbs specific part of > ipoib was cleanly split out into the 'default' implementation of your > get_ipoib_accel_ops. This seems reasonable, Ira > > Jason > -- > To unsubscribe from this list: send the line "unsubscribe linux-rdma" in > the body of a message to majordomo@xxxxxxxxxxxxxxx > More majordomo info at http://vger.kernel.org/majordomo-info.html -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
