On Tue, Jul 16, 2019 at 7:14 PM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote: > > On Tue, 16 Jul 2019, Paul Thomas wrote: > > > > > We do have trace data for a instance where cyclictest hit 5000 uS, but > > > > the file is large 11MB, and I wasn't sure if I should directly attach > > > > it to the email. What is the best way to post or send this? > > > > > > Can you upload the trace to some place? If you don't have storage, just use > > > bugzilla.kernel.org and open a bug, attach it and post a reference. > > > > Hi Thomas, > > > > I've put the last 60k line of the trace on the bug report here: > > https://bugzilla.kernel.org/show_bug.cgi?id=204189 > > > > Just looking at the timestamp time of cyclicte-2778 I can see the last > > two finish_task_switch calls are seperated by 5801 uS. Is just before > > that the correct place to be looking? > > Right. If I'm reading it correctly, I think I know what's wrong. Can you > please revert: > > 80e7f3b05365 ("arch/arm64: Add lazy preempt support") > > Either that patch is broken or it got broken on forward porting it from > 4.19. Need to look with brain awake tomorrow. Yes! this looks like it fixes it. The hash I see is slightly different: b7f42e8fb48cbe38e4c2d0cab48309aadfa6d032 But the name was the same "arch/arm64: Add lazy preempt support" I did a quick run with 5.0.19-rt11 with these results: root@xu5:/opt/rt-tests# ./cyclictest -S -m -n -p 99 -i 200 -h 400 -D 900 # /dev/cpu_dma_latency set to 0us policy: fifo: loadavg: 0.64 0.74 0.60 1/131 1931 T: 0 ( 1913) P:99 I:200 C:4499993 Min: 5 Act: 8 Avg: 7 Max: 20 T: 1 ( 1914) P:99 I:200 C:4499899 Min: 5 Act: 8 Avg: 7 Max: 16 T: 2 ( 1915) P:99 I:200 C:4499799 Min: 5 Act: 8 Avg: 7 Max: 17 T: 3 ( 1916) P:99 I:200 C:4499704 Min: 5 Act: 8 Avg: 7 Max: 16 We'll do some more testing tomorrow. -Paul
