Re: [tip:timers/core] [timers] 7ee9887703: netperf.Throughput_Mbps -1.2% regression

[Frederic Weisbecker]

On Tue, Mar 05, 2024 at 10:17:43AM +0800, Oliver Sang wrote:
>      20.97 ±  3%      -9.7       11.23 ± 16%      -9.6       11.40 ±  4%  perf-profile.calltrace.cycles-pp.__napi_poll.net_rx_action.__do_softirq.irq_exit_rcu.sysvec_apic_timer_interrupt

And yes, fewer time spent in IRQ-tail softirq processing.

>       1.70 ± 13%      +5.6        7.28 ± 14%      +5.8        7.54 ± 15%  perf-profile.calltrace.cycles-pp.__napi_poll.net_rx_action.__do_softirq.run_ksoftirqd.smpboot_thread_fn

And more time spent in ksoftirqd softirq processing.

This can match the increase in involuntary context switches: IRQ-tail softirq
processing takes too much time, possibly due to long lasting remote timer
expiring, ksoftirqd is then scheduled, preempting netperf (through
might_resched()/cond_resched()) as this is a voluntary preemption kernel).

Thanks.