Re: [Xenomai-core] I-pipe patch for ARM S3C24xx (v4)

[Jan Kiszka <jan.kiszka@domain.hid>]

[-- Attachment #1: Type: text/plain, Size: 1636 bytes --]

Sebastian Smolorz wrote:
> Jan Kiszka wrote:
>> Sebastian Smolorz wrote:
>>> However, there is one disturbing issue. When I execute
>>>
>>> dd if=/dev/zero of=/dev/null
>>>
>>> and the latency test in userspace with a period not less than 150 us the
>>> worst case latency is about 220 us. But when I start latency with -p 100
>>> I get a softlock like the one attached. I guess this is the same problem
>>> as Detlef Vollmann described in [2]. So I think it's time for a big fat
>>> ARM-specific warning in the troubleshooting file and perhaps a
>>> modification of the testsuite so that if being compiled for ARM the
>>> default sample periods are greater than the 100 us now.
>> Something is preventing the watchdog kthread from being executed for
>> more than 10 s. Maybe this is just a sign that the systems is hopelessly
>> overloaded (what are average latencies?).
> 
> They are 120-130 us.

Which is a good sign that the board is basically only flushing caches
when running the test at this frequency. Still the question remains if
the BUG is an expectable behaviour or a sign for hidden problems.

> 
>> Maybe it is a real IRQ or 
>> scheduling issue in Linux caused by I-pipe. Maybe it is time to port the
>> I-pipe tracer... ;)
> 
> With the latest Adeos upgrade for ARM the I-pipe tracer was integrated, wasn't 
> it? I was already taking its usage into account.

The arch-independent code is merged for ARM, but it still takes to port
ipipe-mcount.S and likely ipipe_read_tsc + related services. Not much to
do, though. Take a look at the code in PREEMPT_RT for the mcount part.

Jan


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 250 bytes --]