[linux-audio-user] Re: [Jackit-devel] What is a 'delay exceeds' vs. and xrun - BIG MISSES - AND SOLUTIONS!

[Mark Knecht]

On Wed, 6 Oct 2004 04:37:17 +0200, Florian Schmidt <mista.tapas at gmx.net> wrote:
> On Tue, 05 Oct 2004 19:01:30 -0400
> Lee Revell <rlrevell at joe-job.com> wrote:
> 
> > >    The only 3 xruns happened right when I started the kernel compile.
> > >
> > >    This is 2.6.9-rc2-mm4-VP-S7
> > >
> >
> > There is absolutely _no way_ that these are kernel-induced.  The period
> > time is 5ms!!!  This kernel should _never_ produce an delay of more than
> > 200 usecs.  This _has_ to be flaky hardware, or a bug in jackd or in
> > your clients.  What happens if you run jackd with no clients?
> 
> Hi,
> 
> the kernel does weird things nowadays. I do not get any satisfactory results
> with either S8 or S9. I get xruns all over the place [under loads much
> lighter than a kernel compilation] although the VP and related settings are
> as always. T1 is working smoothly again. I have no idea what caused this
> behaviour in S8 and S9. But it feels like a scheduler problem. X11 is much
> more jumpy under these kernels and load, too than under S7 and T1
> 
> So, i would still expect some weirdnesses in VP enabled kernels, since they
> are based on the mm tree [which might hold many surprises in itself].
> 
> But anyways, back to Mark.
> 
> I talked to Mark on irc and he seems to have had SMP enabled on a UP system
> [p4 mobile w/o HT according to mark]. OTOH this worked fine on fedora,
> right? Also this is a notebook, maybe power saving features play a role.
> 
> One more thing to note. His gentoo system uses a linuxthreads glibc. The
> same hardware and kernels are working very nicely on his fedora core system
> with CCRMA and VP kernels.
> 
> I suspect some weird interplay of a linuxthreads self built glibc and a
> 2.6.x kernel, especially wrt to SCHED_FIFO scheduling.
> 
> One more thing:
> 
> The setups between fedora core and gentoo might differ in what default
> services run. Do you [Mark] have any powermanagement daemons loaded? cpu
> frequency scaling, etc? I know only little about these issues, since i
> basically do not use power management.
> 
> The two things i would recommend to try would be:
> 
> 1] Build a very very barebone kernel for, let's say, 386 architecture. no
> acpi no apm, no usb [if possible], no smp, local APIC if he likes. Only
> those devices really needed, no dri, etc [but of course with the usual set
> of VP features enabled].
> 
> 2] try a gentoo system with a nptl enabled glibc. Either build one, or maybe
> get some live cd based on gentoo which fulfills the requirements. See if the
> problems persist. Or try some third distribution with nptl libc and slap a
> VP kernel on it. If it works fine, your gentoo system is b0rk3n in some way.
> 
> flo
> 
> p.s. which jack version btw?
> 

Hi all,
   OK, Last evening and this morning I have made good headway. It
appears that the main culprit was ACPI support, or some part of it.
With most of the ACPI stuff now disabled in the kernel build I am able
to run jack with no clients attached, as a root or user with the lsm
module modprobed, at p=8, n=2 for over 1 hour with no xruns. If my
calculations are correct this would be a latency of less than 0.5mS!

   My current kernel is 2.6.9-rc2-mm4-VP-S7-lsm-UMP-noACPI:

- 2.6.8 from kernel.org
- 2.6.9-rc2 patch
- 2.6.9-rc2-mm4 patch
- voluntary_premption-S7 patch
- realtime-lsm patch

   I've built it with most everything I can find that I do not
absolutely need turned off. I'm non-SMP, ACPI mostly gone, and who
knows what else along the way. I'll send the .config file to whoever
might need it. I am then running the sound card and CDROM interrupts
non-threaded and everything else is threaded.

   In terms of better testing I now need to concentrate on finding an
application that run without creating xruns within the application
itself. At higher latency values (p=256 for instance) alsaplayer -r -o
jack runs pretty well, but as I move to lower latencies I find that
alsaplayer deterministically creates xruns at the transistion between
every song on a CD. I also tried alsaplayer playing wave files from
disk, but I see xruns as it goes from one file to the next also.

   I want to try using chrt to see if alsaplayer is really getting
real-time access, etc. Maybe there are things still not right with the
way I'm running it on my box.

   Is there any other simple app that could play a wave file from, for
instance, memory without this sort of problem?

   I'll certainly be trying out Ardour as a bigger test once I'm more
confident of the disk access issues in real-time. I think that 1394
support may not be in good shape in this kernel yet.

   So, the good news is my Gentoo system is not broken, and that I've
hopefully learned a bit and contributed just a little bit more along
the way. That said I still have a long ways to go.

Cheers,
Mark