The following commit causes rte_pktmbuf_prepend() to fail when prepending the full headroom reserved when creating the mbuf pool.

https://git.dpdk.org/dpdk-stable/commit/?h=21.11&id=d41d39bcf76900deb6aa000e929fd65734254201

Of issue is if the received frame was originally tagged (and vlan-stripped by af-socket connection) the above commit does a vlan reinsert action as part of eth_af_packet_rx(). This moves the SOF forward into the "reserved" space by four bytes. For example, if the reserved space was 128 bytes and the next stage of processing is a generic pipeline handling a variety of interface types, the generic prepend of the reserved 128-byte meta-data space will fail. This means you cannot use a common pipeline and have to have an exception now for af-packet rx handling as compared to physical NIC, tap, vhost, etc. Ideally, the initial SOF offset should be consistent relative to your design's reserved headroom, regardless of the vlan tagged status of the frame or the interface type it was received on.