On Fri, Dec 17, 2021 at 10:57 PM Song Liu <song@xxxxxxxxxx> wrote: > > On Fri, Dec 17, 2021 at 5:42 AM Sebastian Andrzej Siewior > <bigeasy@xxxxxxxxxxxxx> wrote: > > > > On 2021-12-17 10:16:10 [+0800], Yajun Deng wrote: > > > We found an abnormally high latency when executing modprobe raid6_pq, the > > > latency is greater than 1.2s when CONFIG_PREEMPT_VOLUNTARY=y, greater than > > > 67ms when CONFIG_PREEMPT=y, and greater than 16ms when CONFIG_PREEMPT_RT=y. > > > > > > How to reproduce: > > > - Install cyclictest > > > sudo apt install rt-tests > > > - Run cyclictest example in one terminal > > > sudo cyclictest -S -p 95 -d 0 -i 1000 -D 24h -m > > > - Modprobe raid6_pq in another terminal > > > sudo modprobe raid6_pq > > > > > > This is caused by ksoftirqd fail to scheduled due to disable preemption, > > > this time is too long and unreasonable. > > > > > > Reduce high latency by using migrate_disabl()/emigrate_enable() instead of > > > preempt_disable()/preempt_enable(), the latency won't greater than 100us. > > > > > > This patch beneficial for CONFIG_PREEMPT=y or CONFIG_PREEMPT_RT=y, but no > > > effect for CONFIG_PREEMPT_VOLUNTARY=y. > > > > Why does it matter? This is only during boot-up/ module loading or do I > > miss something? > > Yes this only happens on boot-up and module loading.I don't know RT well > enough to tell whether latency during module loading is an issue. Nope. It is not. > > The delay is a jiffy so it depends on CONFIG_HZ. You do benchmark for > > the best algorithm and if you get preempted during that period then your > > results may be wrong and you make a bad selection. > > With current code, the delay _should be_ 16 jiffies. However, the experiment > hits way longer latencies. I agree this may cause inaccurate benchmark results > and thus suboptimal RAID algorithm. I explained this in the original thread. All the observed latencies are really expected. > I guess the key question is whether long latency at module loading time matters. > If that doesn't matter, we should just drop this. Again, it does not matter at all and here it is rather desired by design. Drop this, please. --nX > Thanks, > Song
