This is a sample application for AF_XDP sockets. It was originally in the Linux source code tree under samples/bpf and called xdpsock. The main purpose of it is to demonstrate the various capabilities of AF_XDP. It is not meant to be a minimal example of what you need to do to get things up and running.
The application contains three micro-benchmarks:
rxdrop: Receives a packet then immediately drops it without touching the packet data. (-r command line option)
txpush: Sends pregenerated packets as fast as possible. Never touches packet data during the send phase. (-t command line option)
l2fwd: A simple mac swap benchmark that receives a packet, swaps the mac addresses in the Ethernet header, then sends it out. This benchmark touches the first cache line of the packets.
Here is the simplest command line for launching the rxdrop benchmark in this example:
sudo taskset -c <app core> ./xdpsock -i <interface> -q <queue_id> -rWhen running in the default softirq mode, it is good to have the application and the driver executing on different cores. So, make app_core != queue_id, in a standard Linux config, otherwise performance will suffer.
Some interesting capabilities that are demonstrated by xdpsock:
Busy-poll mode. In this mode both the application and the driver can be run efficiently on the same core. The kernel driver is explicitly invoked by the application by calling either recvmsg() or sendto(). Invoke this by setting the -B option. The -b option can be used to set the batch size that the driver will use. For example:
sudo taskset -c 2 ./xdpsock -i <interface> -q 2 -l -N -B -b 256Note that you have to enable busy_poll explicitly for the netdev. Here is one example:
echo 2 | sudo tee /sys/class/net/<interface>/napi_defer_hard_irqs
echo 200000 | sudo tee /sys/class/net/<interface>/gro_flush_timeoutXDP_SHARED_UMEM mode, where a single umem is shared between multiple sockets. Note that this example demonstrates how to share the umem between sockets bound to the same netdev and queue id. This can be enabled with the -M option. For an example on how to share a umem between sockets with different netdevs and/or queue ids, please take a look at the AF_XDP-forwarding example.
Various statistics can be dumped with the -a and -x options. The -I option can also be used to dump statistics about interrupts, but requires the user to supply the interrupt vector name that can be found in /proc/interrupts.
Unaligned mode in which buffers can be placed at any place in the umem. For the default aligned mode, recieved buffers will be placed at a standard offset (256 bytes) from the start of the chunk, and a packet cannot cross a chunk boundary. In unaligned mode, there are no chunks. If you put an address into the fill ring, the packet will be received at exactly that address. In aligned mode, it will be put at the standard offset in the chunk that the fill address points to.
Unaligned mode requires huge pages to be feasible. Please see the example code.