Strange behaviour on MPC5200

[Detlev Zundel <dzu@denx.de>]

[-- Attachment #1: Type: text/plain, Size: 1325 bytes --]

Hi,

during tests with the rt-preempt kernel on PowerPC, I stumbled across
this really weird phenomenon of cyclictest latencies > 30ms.  The test
system is a TQM5200 supported mainline, so the complete kernel source is
unmodified mainline + rt-patch.  The -00001 in the kernel revision is
due to the fact that I commited the results of applying the respective
rt-patch to my local git repo.

To debug the problem, I turned the latency tracers on and started
cyclictest with 'cyclictest -n -p80'.  This looks quite good, hackbench
and the cache calibrator do not do much harm, _until_ I start to do a
ping flood from outside.  Then within a few seconds the high latencies
occur.  To test for a regression, I did the same test under
2.6.33.7-rt29 and 2.6.31.12-rt21.  Both versions yield similar results.

Attached are the trace outputs for those events (gziped as they are
larger than 300KiB).  

It looks like the scheduling goes completely wrong as there is even the
idle task running before the runnable cyclictest gets scheduled in.  Can
someone give me a hint on what may be wrong here?

Thanks
  Detlev

--
DENX Software Engineering GmbH,      MD: Wolfgang Denk & Detlev Zundel
HRB 165235 Munich,  Office: Kirchenstr.5, D-82194 Groebenzell, Germany
Phone: (+49)-8142-66989-40 Fax: (+49)-8142-66989-80 Email: dzu@denx.de

[-- Attachment #2: rt-2.6.33.7-latency-39516us.gz --]
[-- Type: application/octet-stream, Size: 37319 bytes --]

[-- Attachment #3: rt-2.6.31.12-latency-33568us.gz --]
[-- Type: application/octet-stream, Size: 35313 bytes --]