Undefined Runaways during Cyclictest/Latency Measurement RT-Kernel

[weissermagier]

Hello community,

I’m using a T2080 RDB with linux version 4.1.8-rt8 with following kernel options

Processor Support ->  CPU selection -> (X) Freescale e6500

Kernel options --->  Preemption Model (Fully Preemptible Kernel (RT)) --->  (X) Fully Preemptible Kernel (RT)

U-Boot SDK 1.7

QorIQ SDK (FSL Reference Distro) 2.0

Kernel config (below as attachment)

I tested the idle(!) system with the cyclictest v1.00 via SSH and had some runaways > 30 μs like you can see in the picture below.

All power saving options like CPU Idle or Dynamic Frequence Skaling(DFS) were disabled of course.

Do you have any ideas guys?

Original Attachment has been moved to: config.txt.zip

[tracysmith]

I have the same question.  These runaways occur on the LS1043ARDB with SDK 2.0 with the PREEMPT_RT full preemption.  Customer's would like to know why these runaways occur.  Have seen them as high as > 72 uS. 

NXP please debug these runaways since this behaviour impacts the soft real-time behaviour expected from the PREEMPT_RT patch for customers.  PREEMPT_RT has been successfully validated by NXP on the LS1043ARDB.

Without using ftrace, the latency is 2uS to 3uS which is fine and with ftrace avg latency is 7uS, but the runaway or outlier cases is a serious concern and difficult to catch using ftrace because the ftrace trace wraps too quickly to catch the behaviour.  Any suggestions on how to debug and catch what is causing the runaways?

[tracysmith]

Runaway latency appear to be fixed under SDK 2.0-1609 but haven't checked later releases for regression.