[etherlab-users] Usage of EtherCAT-Master Userspace-library breaks hard realtime execution under RTAI/LXRT

Jürgen Kunz kunz at sim.tu-darmstadt.de
Tue Apr 6 17:50:48 CEST 2010


I am using EtherCAT Version 1.5 (rev. 1824) with RTAI 3.7.1 for
controlling a compliant robot arm. The controller is implemented in user
space because of licence issues implementing it in kernel space. As
mentioned in the feature list, the EtherCAT-Master implementation should
be usable with the different realtime implementations RTAI, Xenomai,...
and also under standard linux itself.
The controller has to work with a frequency of at minimum 1 kHz. I am
checking the real cycle times of the controller and I have found out,
that the realtime task is not being executed with the desired frequency
(1, 2 or 4 kHz). Irregularly the controller task is interrupted for more
than 25ms. Using the same controlling code with an internal software
model of the robot arm, the controller and the simulation model are
running constantly with the desired frequency. Using the
EtherCAT-library to control the real hardware leads to a break of hard
realtime execution. I commented out everything in the
realtime-step-function but these four lines mentioned in the
documentation for cyclic operation of EtherCAT:
ecrt_master_receive(), ecrt_domain_process(), ecrt_domain_queue(),
The scheduler stats of rtai (/proc/rtai/scheduler) are telling me, that
there are transitions from hard to soft to hard realtime execution
caused by syscalls and the number of syscalls is continuously
increasing, if the controller thread is executed. Compared to the
simulation model, there are no syscalls breaking hard realtime execution.
I analysed the code of the user space library and also the kernel space
implementation of the ioctl-calls of the character device
/dev/EtherCATx. There are two points which are breaking realtime
execution under rtai/lxrt: 1. using ioctl (see
http://www.rtai.dk/cgi-bin/gratiswiki.pl?LXRT_FAQs) and 2. using kernel
semaphores to lock the above mentioned kernel function calls, which are
not locked if they are used in kernel space directly. I commented out
the usage of the kernel semaphores in the relating ioctl-functions in
kernel space, but that makes no difference that the hard-realtime-thread
is pushed back to soft realtime execution.
So, why are kernel semaphores being used here? It is mentioned in the
documentation that usage of standard kernel semaphores are not
sufficient for providing a correct locking mechanism if realtime
extensions like rtai and xenomai are used and so the application which
has requested the master has to provide the appropriate locking
mechanism. Does an implementation of the kernel calls exist which can
guarantee hard realtime execution in user space? Or do I have to
implement the corresponding LXRT-calls by myself?

Best regards,
Jürgen Kunz

Dipl.-Inform. Jürgen Kunz           E-Mail: kunz(at)sim.tu-darmstadt.de
FG Simulation, Systemoptimierung    Tel.:   ++49 (0) 6151-16-70383
und Robotik                         Fax:    ++49 (0) 6151-16-6648
Technische Universität Darmstadt    http://www.sim.tu-darmstadt.de
Hochschulstr. 10                    64289 Darmstadt 

More information about the Etherlab-users mailing list