In the spirit of the legend Bob Marley (playing while i am sipping coffee and typing this!): AF_XDP: Could you be loved?
Jason Xing is showing a lot of love to you. Most drivers can't do zero-copy, so copy mode is the universal path. Jason will talk about recent correctness fixes (multi-buffer TX leaks, continuation descriptors, TOCTOU metadata) cleared the bugs, while fine-grained profiling unlocked close to 2× throughput.
Questions for you AF_XDP (yes, you!): Could you be locked - with finer grain? Could you be batched - without the pain? No single silver bullet! Just cycle-level perf, contention tracing, cache-line heatmaps. And the optimization love for you:
- Send without budget blues - Hot structure reshuffles (cache-line dance) - Batch xmit — one ring to rule them - Doorbell coalescing (don't ring it twice) - Skb allocation acceleration - IRQ pairs trimmed short
Don't let them fool ya — spinlock's not the ruler. Methodical, mon, methodical, one cache line at a time, thats how we get there!
Ok, enough of my spin, read more details on the talk here: https://netdevconf.info/0x1A/sessions/talk/af_xdp-copy-mode-needs-more-love....
cheers, jamal