[LAT] rtlimits (was: a working rt-kernel -- please test )

[Robin Gareus]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

krgn wrote:
> one or 2 quick addedums:
> 
> ubuntu hardy kernel works great for me. it is presumably compiled with a
> different version of gcc, 4.2.3. could this be the source of the problems?

maybe, maybe not. it depends on compiler flags. That may be a question
for gcc specialists on the linux-rt-user mailing list.

I start to think we're chasing two independent but related bugs:
/something/  inside the kernel is triggered by invalid user setup. eg.
RT-priority inversion.

> On Thu, Aug 7, 2008 at 11:56 PM, krgn <k.gebbert at gmail.com> wrote:
> 
>> I did test with a few other versions of config, patchset and kernel sources
>> but keep running into problems. I can't seem to be able to produce something
>> that works reliably, all of the kernels run around 30 to 45 minutes with
>> around 65 - 70% load on jackd and then hard lock. Since this machine I am
>> using is pretty new, and I am havn't really run rt-linux on this one
>> extensivly yet (partly because its almost not necessary with current stock
>> debian kernels anymore), it could be something about the hardware, although
>> it seems unlikely. Its a T61 T5500 with 1G Ram fwiw. With this in mind I
>> think I should ask for some advice on what people consider sane values for
>> important scripts, pam etc. My /etc/security/limits.conf looks like this:
>>
>> @audio - rtprio 99
>> @audio - nice -15
>> @audio - memlock 1024000000

i use

@audio - rtprio 90
@audio - nice -5
@audio - memlock 1400000

The man page says memlock's unit is KB not bytes.
IIRC someone said it's actually pages, but I did not check the source.

>> What values do people use, and why? Are these sane values? 

I think so.

rtprio ranges from 1 to 99 and it does not matter how big the difference
is; as long as a value is larger it will "win" against lower values -
for SCHED_FIFO that is. - many suggest to leave a little "headroom"; it
does not need to be 99.

>> What are the rtirq settings that are
>> considered safe? 

100 > timer-prio > sound-card-prio > jackd-process prio > 0

with a UA-25 on USB I use:

rtirq: reset [softirq-timer] pid=6 prio=1: OK.
rtirq: reset [softirq-timer] pid=11161 prio=1: OK.
rtirq: start [softirq-timer] pid=6 prio=99: OK.
rtirq: start [softirq-timer] pid=11161 prio=99: OK.
rtirq: start [rtc] irq=8 pid=2397 prio=90: OK.
rtirq: start [uhci_hcd:usb3] irq=22 pid=1066 prio=89: OK.
rtirq: start [i8042] irq=1 pid=330 prio=87: OK.
rtirq: start [i8042] irq=12 pid=329 prio=86: OK.

and run /usr/bin/jackd -R -P70 ...

<bit-off-topic>
@audio  /usr/bin/jackd          nice=-1 rtprio=85
@audio  /usr/bin/qjackctl       nice=-1 rtprio=84
@audio  /usr/bin/ardour         nice=-1 rtprio=83
@audio  /usr/bin/hydrogen       nice=-1 rtprio=82

as found on
http://proaudio.tuxfamily.org/wiki/index.php?title=Howto_RT_Kernel

is IMHO nonsense - in fact wrong: ONLY the jack-realtime callback should
have elevated priority. The audio-process-callback of the applications
will inherit this priority.
The above example would prefer GUI or JACK-stdio/log over disk-I/O.

There are use-cases where you may want to give an older-RTC-based
MIDI-application realtime priority. JACK-midi solves that for good.
</bit-off-topic>

>> For now, I switched it off,

makes sense to debug&trace.

>> yet I can see how it could be
>> useful/necessary for certain setups, and don't want to rule it out
>> completely. I start jackd oh hw:1 (hdsp multiface) with the following
>> parameters:
>>
>> /usr/bin/jackd -R -P89 -m -dalsa -dhw:1 -r48000 -p128 -n2 -Xseq

now it clicked in the back of my head: there was something like
subtracting or adding 10 from the rt-priority in qjackctl, or jackd.
just a shot in the dark: jackd itself starts a couple of threads with
different priorities. I never cared since it just worked (and still does
on 2.6.24-rt3). - later today I'll be looking at jackd.

Meanwhile, does anyone here know about changes in the scheduler,
priority inversion etc of 2.6.26 ?

For high-level tests it may be interesting if a 2.6.26-rtX would run
stable if it does not do any audio or RT at all - just to double-back.
same without ACPI. for all we know this could well be caused by the
keyboard or some tty driver, no? I wish I had more machines to test ;)

puh, sorry for another long email - i got carried away.

robin
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.9 (GNU/Linux)

iEYEARECAAYFAkicUs4ACgkQeVUk8U+VK0LcGwCgwR6VIOLUN7YUGvS/0vEPpn+Q
PNkAnR+c2CeekQUK89S65FTtIeAqB1BU
=ZidW
-----END PGP SIGNATURE-----