On Mon, 8 Jan 2024, Vladimir Oltean wrote:
On Mon, Jan 08, 2024 at 03:23:38PM +0100, Romain Gantois wrote:
I see, the kernel docs were indeed enlightening on this point. As a side note, I've just benchmarked both the "with-inline" and "without-inline" versions. First of all, objdump seems to confirm that GCC does indeed follow this pragma in this particular case. Also, RX perfs are better with stmmac_has_ip_ethertype inlined, but TX perfs are actually consistently worse with this function inlined, which could very well be caused by cache effects.
In any case, I think it is better to remove the "inline" pragma as you said. I'll do that in v4.
Are you doing any code instrumentation, or just measuring the results and deducing what might cause them?
It might be worth looking at the perf events and seeing what function consumes the most amount of time.
CPU_CORE=0 perf record -e cycles -C $CPU_CORE sleep 10 && perf report perf record -e cache-misses -C $CPU_CORE sleep 10 && perf report
Unfortunately my hardware doesn't support these performance metrics, but I did manage to do some instrumentation with the ftrace profiler:
Same test conditions as before, 10 second iperf3 runs with unfragmented UDP packets.
no inline TX average time per call for stmmac_xmit(): 85us average time per call for stmmac_has_ip_ethertype(): 2us
no inline RX average time per call for stmmac_napi_poll_rx(): 8142us average time per call for stmmac_has_ip_ethertype(): 2us
inline TX: average time per call for stmmac_xmit(): 85us
inline RX: average time per call for stmmac_napi_poll_rx(): 8410us
It seems like this time, RX performed slightly worse with the function inline. To be honest, I'm starting to doubt the reproducibility of these tests. In any case it seems better to just remove the "inline" and let gcc do the optimizing.
Best Regards,