tjcw / bpf-examples Goto Github PK

At the moment, I have receive socket and send socket using separate umems. That causes a requirement for a 'memcpy' in user space between where the packet is received to, and where the packet can be sent from. This issue replaces the separate umems wiith a shared umem and eliminates this memcpy.

It should also be possible to reinject the packet to the same queue from where it was redirected to user space. This should cause the packet to go to the transmitting interface, then be redirected again to the veth where it was initially redirected to user space, then be sent to the bridge and on its way. Pakcet time-to-live should be decremented by 1 at each hop. and the packet should reach its destination before the time-to-live gets to zero.

At the moment, packets are reinjected to the destination node's transmit queue. This is OK when there are only 2 namespaces involved, but becomes cumbersome as the number of namespaces increases as would be the case in a Kubernetes system. This issue will change to using the source node's transmit queue, from where linux kerne; routing weill get the packet to the right destination. The packet will not be directed to user space a second time because the map is updated before the packet arrives again.