Real-Time Preemption, -RT-2.6.10-rc2-mm3-V0.7.31-7

Post by Rui Nuno Capela
xruntrace1.0.tar.gz
- scripts used for xrun trace capture automation
xruntrace1-2.6.10-rc2-mm3-RT-V0.7.31-7.tar.gz
- actual xrun traces captured while running jackd on RT-V0.7.31-7

the trace buffer is too small to capture a significant portion of the
xrun - i'd suggest for you to increase it from 4096-1 to 4096*16-1, to
be able to capture a couple of hundreds of millisecs worth of traces.

also, i think it produces more stable results if the tracing is
activated and deactivated from userspace - i.e. the patch below will
measure the latency of poll() executed by the ALSA glue code within
JACK. (i also removed the xrun printf, because it can cause delays.)

this patch can be used the following way: do not activate xrun_debug via
/proc (to not create additional latencies that make xruns worse), but
otherwise set the /proc/sys/kernel/ flags the same way you do already.
If the patch is applied to jackd then /proc/latency_trace will show the
latest and highest latency that was measured.

note that this captures latencies visible in the ALSA glue, which doesnt
cover the whole jackd critical path - but it should be pretty good as an
initial start. It will definitely show all jackd-unrelated delay
sources.

to cover all latency paths, the last chunk of the patch could perhaps be
modified to do:

+ gettimeofday(1,0);
+ gettimeofday(1,1);

this will also capture latencies within jackd itself. But i think it is
better to separate 'external' from 'internal' latencies, because the two
situations are quite different and it might be hard to identify which
particular latency we are seeing in /proc/latency_trace.

Ingo

--- ./drivers/alsa/alsa_driver.c.orig 2004-11-26 14:11:26.000000000 +0100
+++ ./drivers/alsa/alsa_driver.c 2004-11-26 14:24:41.000000000 +0100
@@ -1081,9 +1081,11 @@ alsa_driver_xrun_recovery (alsa_driver_t
snd_pcm_status_get_trigger_tstamp(status, &tstamp);
timersub(&now, &tstamp, &diff);
*delayed_usecs = diff.tv_sec * 1000000.0 + diff.tv_usec;
+#if 0
fprintf(stderr, "\n\n**** alsa_pcm: xrun of at least %.3f "
"msecs\n\n",
*delayed_usecs / 1000.0);
+#endif
}

if (alsa_driver_stop (driver) ||
@@ -1187,6 +1189,7 @@ alsa_driver_wait (alsa_driver_t *driver,

poll_enter = jack_get_microseconds ();

+ gettimeofday(1,1);
if (poll (driver->pfd, nfds, driver->poll_timeout) < 0) {

if (errno == EINTR) {
@@ -1206,6 +1209,7 @@ alsa_driver_wait (alsa_driver_t *driver,
return 0;

}
+ gettimeofday(1,0);

poll_ret = jack_get_microseconds ();

Ingo Molnar

2004-11-29 11:24:23 UTC

Post by Ingo Molnar
--- ./drivers/alsa/alsa_driver.c.orig 2004-11-26 14:11:26.000000000 +0100
+++ ./drivers/alsa/alsa_driver.c 2004-11-26 14:24:41.000000000 +0100

i think the one below is a better approach - it will only trace events
that the ALSA driver has reported to be an xrun. I.e. the full latency
path from the point where poll() is called, up to the point where jackd
[after the latency has occured] considers it an xrun worth
counting/reporting. The tracing restarts at every poll(), so only the
latency is captured. If you up your tracebuffer to 40K+ entries then i'd
suggest to use the following trace settings:

echo 0 > /proc/asound/card0/pcm0p/xrun_debug

echo 1 > /proc/sys/kernel/trace_user_triggered
echo 0 > /proc/sys/kernel/trace_freerunning
echo 0 > /proc/sys/kernel/preempt_max_latency
echo 0 > /proc/sys/kernel/preempt_thresh
echo 0 > /proc/sys/kernel/preempt_wakeup_timing

i.e. dont use trace_freerunning - this will give much easier to parse
traces.

a suggestion wrt. the format of the .trc files: it would be nice if you
could dump the PIDs of all relevant tasks into it too, to make it easier
to identify who causes what latency. Ideally it would be useful to
have a more symbolic trace - i.e. instead of:

3570 00000000 254981.991ms (+0.000ms): up (ext3_orphan_del)

it would be:

jackd-3570 00000000 254981.991ms (+0.000ms): up (ext3_orphan_del)

but this is easier done in the kernel - some of the tasks involved in a
latency might be long gone by the time you detect the xrun.

Ingo

--- ./drivers/alsa/alsa_driver.c.orig 2004-11-26 14:11:26.000000000 +0100
+++ ./drivers/alsa/alsa_driver.c 2004-11-26 15:31:37.000000000 +0100
@@ -1077,13 +1077,16 @@ alsa_driver_xrun_recovery (alsa_driver_t
&& driver->process_count > XRUN_REPORT_DELAY) {
struct timeval now, diff, tstamp;
driver->xrun_count++;
+ gettimeofday(0,0);
gettimeofday(&now, 0);
snd_pcm_status_get_trigger_tstamp(status, &tstamp);
timersub(&now, &tstamp, &diff);
*delayed_usecs = diff.tv_sec * 1000000.0 + diff.tv_usec;
+#if 0
fprintf(stderr, "\n\n**** alsa_pcm: xrun of at least %.3f "
"msecs\n\n",
*delayed_usecs / 1000.0);
+#endif
}

if (alsa_driver_stop (driver) ||
@@ -1185,6 +1188,7 @@ alsa_driver_wait (alsa_driver_t *driver,
nfds++;
}

+ gettimeofday(0,1);
poll_enter = jack_get_microseconds ();

if (poll (driver->pfd, nfds, driver->poll_timeout) < 0) {

Ingo Molnar

2004-11-29 15:42:16 UTC

Post by Ingo Molnar
--- ./drivers/alsa/alsa_driver.c.orig 2004-11-26 14:11:26.000000000 +0100
+++ ./drivers/alsa/alsa_driver.c 2004-11-26 14:24:41.000000000 +0100

find yet another approach below - it aborts jackd when the first xrun
happens. Since an xrun can trigger other xruns, the most reliable trace
would be the first xrun that triggers.

Ingo

--- ./drivers/alsa/alsa_driver.c.orig 2004-11-26 14:11:26.000000000 +0100
+++ ./drivers/alsa/alsa_driver.c 2004-11-29 17:39:15.985918576 +0100
@@ -1077,13 +1077,17 @@ alsa_driver_xrun_recovery (alsa_driver_t
&& driver->process_count > XRUN_REPORT_DELAY) {
struct timeval now, diff, tstamp;
driver->xrun_count++;
+ gettimeofday(0,0);
gettimeofday(&now, 0);
snd_pcm_status_get_trigger_tstamp(status, &tstamp);
timersub(&now, &tstamp, &diff);
*delayed_usecs = diff.tv_sec * 1000000.0 + diff.tv_usec;
+#if 1
fprintf(stderr, "\n\n**** alsa_pcm: xrun of at least %.3f "
"msecs\n\n",
*delayed_usecs / 1000.0);
+#endif
+ exit(-1);
}

if (alsa_driver_stop (driver) ||
@@ -1185,6 +1189,7 @@ alsa_driver_wait (alsa_driver_t *driver,
nfds++;
}

+ gettimeofday(0,1);
poll_enter = jack_get_microseconds ();

if (poll (driver->pfd, nfds, driver->poll_timeout) < 0) {

Rui Nuno Capela

2004-11-29 13:13:40 UTC

and how I do that? Is it some /proc magic or its in the kernel configuration?

please forgive my ignorance:)

bye.

--
rncbc aka Rui Nuno Capela
***@rncbc.org

Ingo Molnar

2004-11-29 14:33:16 UTC

Post by Ingo Molnar
the trace buffer is too small to capture a significant portion of the
xrun - i'd suggest for you to increase it from 4096-1 to 4096*16-1, to
be able to capture a couple of hundreds of millisecs worth of traces.

and how I do that? Is it some /proc magic or its in the kernel configuration?

it's the MAX_TRACE define in kernel/latency.c.

but please try to the -31-10 kernel that i've just uploaded, it has a
number of tracer enhancements:

- process name tracking. The new format is:

bash- 3633 80000003 0.264ms (+0.000ms): idle_cpu (wake_idle)
bash- 3633 80000003 0.264ms (+0.000ms): idle_cpu (wake_idle)
bash- 3633 80000003 0.264ms (+0.000ms): find_next_bit (wake_idle)

this makes it easier to identify which process does what. This feature
has no significant overhead in the tracer itself, all the hard work is
done when /proc/latency_trace is read by the user.

- /proc/sys/kernel/mcount_enabled flag: if disabled then
/proc/latency_trace will only contain 'custom' events, but no
per-function entries. This can be useful to trace really long
latencies, to get a rough idea of what is going on.

- /proc/latency_trace atomicity. It was fundamentally non-atomic, due
to it being a 4K-granular file interface. Now the kernel has a second
layer of saved-trace logic, which makes sure that the only time the
trace is switched is when the header of the trace is read. I.e. when
cp-ing or cat-ing /proc/latency_trace, no new trace info will be
used. This change could solve some of the 'truncated traces'
weirdnesses reported.

Ingo

Ingo Molnar

2004-11-29 15:23:44 UTC

Post by Ingo Molnar
but please try to the -31-10 kernel that i've just uploaded, it has a

make that -31-13 (or later). Earlier kernels had a bug in where the
process name tracking only worked for the first latency trace saved,
subsequent traces showed 'unknown' for the process name. In -13 i've
also added a printk that shows the latest user latency in a one-line
printk - just like the built-in latency tracing modes do:

(gettimeofday/3671/CPU#0): new 3068 us user-latency.
(gettimeofday/3784/CPU#0): new 1008627 us user-latency.

(this should also make it easier for helper scripts to save the traces,
whenever they happen.)

Ingo

Gene Heskett

2004-11-29 23:16:43 UTC

Post by Ingo Molnar
but please try to the -31-10 kernel that i've just uploaded, it

make that -31-13 (or later). Earlier kernels had a bug in where the
process name tracking only worked for the first latency trace saved,
subsequent traces showed 'unknown' for the process name. In -13 i've
also added a printk that shows the latest user latency in a one-line
(gettimeofday/3671/CPU#0): new 3068 us user-latency.
(gettimeofday/3784/CPU#0): new 1008627 us user-latency.
(this should also make it easier for helper scripts to save the
traces, whenever they happen.)
Ingo

I just built this to see how much blood it would draw, which isn't
much. I don't have jack here, so I don't have your standard torture
test. Instead, I run tvtime, which runs at a -19 priority.

I let it run about 30 seconds (untimed), noted that the frame error
slippage wasn't improved, and got this output histogram when I quit
it.

Its (tvtime) is running here of course.
--------------------
Nov 29 18:05:45 coyote kernel: Read missed before next interrupt
Nov 29 18:05:45 coyote kernel: wow! That was a 22 millisec bump
Nov 29 18:05:45 coyote kernel: `IRQ 8'[846] is being piggy.
need_resched=0, cpu=0
Nov 29 18:05:45 coyote kernel: Read missed before next interrupt
Nov 29 18:05:45 coyote kernel: wow! That was a 22 millisec bump
Nov 29 18:05:45 coyote kernel: `IRQ 8'[846] is being piggy.
need_resched=0, cpu=0
Nov 29 18:05:45 coyote kernel: Read missed before next interrupt
Nov 29 18:05:45 coyote kernel: wow! That was a 21 millisec bump
Nov 29 18:05:45 coyote kernel: `IRQ 8'[846] is being piggy.
need_resched=0, cpu=0
Nov 29 18:05:45 coyote kernel: Read missed before next interrupt
Nov 29 18:05:45 coyote kernel: wow! That was a 21 millisec bump
Nov 29 18:05:45 coyote kernel: `IRQ 8'[846] is being piggy.
need_resched=0, cpu=0
Nov 29 18:05:45 coyote kernel: Read missed before next interrupt
Nov 29 18:05:45 coyote kernel:

And was stopped here.

Nov 29 18:05:45 coyote kernel: rtc latency histogram of {tvtime/3398,
10609 samples}:
Nov 29 18:05:45 coyote kernel: 4 11
Nov 29 18:05:45 coyote kernel: 5 1716
Nov 29 18:05:45 coyote kernel: 6 4827
Nov 29 18:05:45 coyote kernel: 7 1495
Nov 29 18:05:45 coyote kernel: 8 382
Nov 29 18:05:45 coyote kernel: 9 193
Nov 29 18:05:45 coyote kernel: 10 206
Nov 29 18:05:45 coyote kernel: 11 188
Nov 29 18:05:45 coyote kernel: 12 148
Nov 29 18:05:45 coyote kernel: 13 202
Nov 29 18:05:45 coyote kernel: 14 195
Nov 29 18:05:45 coyote kernel: 15 95
Nov 29 18:05:45 coyote kernel: 16 70
Nov 29 18:05:45 coyote kernel: 17 23
Nov 29 18:05:45 coyote kernel: 18 18
Nov 29 18:05:45 coyote kernel: 19 8
Nov 29 18:05:45 coyote kernel: 20 9
Nov 29 18:05:45 coyote kernel: 21 1
Nov 29 18:05:45 coyote kernel: 22 1
Nov 29 18:05:45 coyote kernel: 26 1
--------------------
And I note that the 1-26 column of numbers does not seem to add up to
whats being logged above there, which are all 21 and 22 ms bumps
(whatever a bump is)

Is this a helpfull report, or just noise? Subjectively, tvtime is
running with far fewer visible frame glitches than before I started
playing with these patches. A marked improvement IMO.

--
Cheers, Gene
"There are four boxes to be used in defense of liberty:
soap, ballot, jury, and ammo. Please use in that order."
-Ed Howdershelt (Author)
99.29% setiathome rank, not too shabby for a WV hillbilly
Yahoo.com attorneys please note, additions to this message
by Gene Heskett are:
Copyright 2004 by Maurice Eugene Heskett, all rights reserved.

K.R. Foley

2004-11-30 01:50:06 UTC

Post by Ingo Molnar
but please try to the -31-10 kernel that i've just uploaded, it

make that -31-13 (or later). Earlier kernels had a bug in where the
process name tracking only worked for the first latency trace saved,
subsequent traces showed 'unknown' for the process name. In -13 i've
also added a printk that shows the latest user latency in a one-line
(gettimeofday/3671/CPU#0): new 3068 us user-latency.
(gettimeofday/3784/CPU#0): new 1008627 us user-latency.
(this should also make it easier for helper scripts to save the
traces, whenever they happen.)
Ingo

I just built this to see how much blood it would draw, which isn't
much. I don't have jack here, so I don't have your standard torture
test. Instead, I run tvtime, which runs at a -19 priority.
I let it run about 30 seconds (untimed), noted that the frame error
slippage wasn't improved, and got this output histogram when I quit
it.
Its (tvtime) is running here of course.
--------------------
Nov 29 18:05:45 coyote kernel: Read missed before next interrupt
Nov 29 18:05:45 coyote kernel: wow! That was a 22 millisec bump
Nov 29 18:05:45 coyote kernel: `IRQ 8'[846] is being piggy.
need_resched=0, cpu=0
Nov 29 18:05:45 coyote kernel: Read missed before next interrupt
Nov 29 18:05:45 coyote kernel: wow! That was a 22 millisec bump
Nov 29 18:05:45 coyote kernel: `IRQ 8'[846] is being piggy.
need_resched=0, cpu=0
Nov 29 18:05:45 coyote kernel: Read missed before next interrupt
Nov 29 18:05:45 coyote kernel: wow! That was a 21 millisec bump
Nov 29 18:05:45 coyote kernel: `IRQ 8'[846] is being piggy.
need_resched=0, cpu=0
Nov 29 18:05:45 coyote kernel: Read missed before next interrupt
Nov 29 18:05:45 coyote kernel: wow! That was a 21 millisec bump
Nov 29 18:05:45 coyote kernel: `IRQ 8'[846] is being piggy.
need_resched=0, cpu=0
Nov 29 18:05:45 coyote kernel: Read missed before next interrupt
And was stopped here.
Nov 29 18:05:45 coyote kernel: rtc latency histogram of {tvtime/3398,
Nov 29 18:05:45 coyote kernel: 4 11
Nov 29 18:05:45 coyote kernel: 5 1716
Nov 29 18:05:45 coyote kernel: 6 4827
Nov 29 18:05:45 coyote kernel: 7 1495
Nov 29 18:05:45 coyote kernel: 8 382
Nov 29 18:05:45 coyote kernel: 9 193
Nov 29 18:05:45 coyote kernel: 10 206
Nov 29 18:05:45 coyote kernel: 11 188
Nov 29 18:05:45 coyote kernel: 12 148
Nov 29 18:05:45 coyote kernel: 13 202
Nov 29 18:05:45 coyote kernel: 14 195
Nov 29 18:05:45 coyote kernel: 15 95
Nov 29 18:05:45 coyote kernel: 16 70
Nov 29 18:05:45 coyote kernel: 17 23
Nov 29 18:05:45 coyote kernel: 18 18
Nov 29 18:05:45 coyote kernel: 19 8
Nov 29 18:05:45 coyote kernel: 20 9
Nov 29 18:05:45 coyote kernel: 21 1
Nov 29 18:05:45 coyote kernel: 22 1
Nov 29 18:05:45 coyote kernel: 26 1
--------------------
And I note that the 1-26 column of numbers does not seem to add up to
whats being logged above there, which are all 21 and 22 ms bumps
(whatever a bump is)

Is this all that is in the log? For some reason there are 820 samples
not represented in the output above. The ms+ hits would have been
represented by something like:

Nov 29 18:05:45 coyote kernel: 9999 4

Not that it is related to the missing output, but does tvtime use
polling on the rtc?

kr

Post by Gene Heskett
Is this a helpfull report, or just noise? Subjectively, tvtime is
running with far fewer visible frame glitches than before I started
playing with these patches. A marked improvement IMO.

Gene Heskett

2004-11-30 03:19:06 UTC

Post by K.R. Foley

Post by Ingo Molnar
but please try to the -31-10 kernel that i've just uploaded, it

make that -31-13 (or later). Earlier kernels had a bug in where
the process name tracking only worked for the first latency trace
saved, subsequent traces showed 'unknown' for the process name.
In -13 i've also added a printk that shows the latest user
latency in a one-line printk - just like the built-in latency
(gettimeofday/3671/CPU#0): new 3068 us user-latency.
(gettimeofday/3784/CPU#0): new 1008627 us user-latency.
(this should also make it easier for helper scripts to save the
traces, whenever they happen.)
Ingo

I just built this to see how much blood it would draw, which isn't
much. I don't have jack here, so I don't have your standard
torture test. Instead, I run tvtime, which runs at a -19
priority.
I let it run about 30 seconds (untimed), noted that the frame
error slippage wasn't improved, and got this output histogram when
I quit it.
Its (tvtime) is running here of course.
--------------------
Nov 29 18:05:45 coyote kernel: Read missed before next interrupt
Nov 29 18:05:45 coyote kernel: wow! That was a 22 millisec bump
Nov 29 18:05:45 coyote kernel: `IRQ 8'[846] is being piggy.
need_resched=0, cpu=0
Nov 29 18:05:45 coyote kernel: Read missed before next interrupt
Nov 29 18:05:45 coyote kernel: wow! That was a 22 millisec bump
Nov 29 18:05:45 coyote kernel: `IRQ 8'[846] is being piggy.
need_resched=0, cpu=0
Nov 29 18:05:45 coyote kernel: Read missed before next interrupt
Nov 29 18:05:45 coyote kernel: wow! That was a 21 millisec bump
Nov 29 18:05:45 coyote kernel: `IRQ 8'[846] is being piggy.
need_resched=0, cpu=0
Nov 29 18:05:45 coyote kernel: Read missed before next interrupt
Nov 29 18:05:45 coyote kernel: wow! That was a 21 millisec bump
Nov 29 18:05:45 coyote kernel: `IRQ 8'[846] is being piggy.
need_resched=0, cpu=0
Nov 29 18:05:45 coyote kernel: Read missed before next interrupt
And was stopped here.
Nov 29 18:05:45 coyote kernel: rtc latency histogram of
Nov 29 18:05:45 coyote kernel: 4 11
Nov 29 18:05:45 coyote kernel: 5 1716
Nov 29 18:05:45 coyote kernel: 6 4827
Nov 29 18:05:45 coyote kernel: 7 1495
Nov 29 18:05:45 coyote kernel: 8 382
Nov 29 18:05:45 coyote kernel: 9 193
Nov 29 18:05:45 coyote kernel: 10 206
Nov 29 18:05:45 coyote kernel: 11 188
Nov 29 18:05:45 coyote kernel: 12 148
Nov 29 18:05:45 coyote kernel: 13 202
Nov 29 18:05:45 coyote kernel: 14 195
Nov 29 18:05:45 coyote kernel: 15 95
Nov 29 18:05:45 coyote kernel: 16 70
Nov 29 18:05:45 coyote kernel: 17 23
Nov 29 18:05:45 coyote kernel: 18 18
Nov 29 18:05:45 coyote kernel: 19 8
Nov 29 18:05:45 coyote kernel: 20 9
Nov 29 18:05:45 coyote kernel: 21 1
Nov 29 18:05:45 coyote kernel: 22 1
Nov 29 18:05:45 coyote kernel: 26 1
--------------------
And I note that the 1-26 column of numbers does not seem to add up
to whats being logged above there, which are all 21 and 22 ms
bumps (whatever a bump is)

Is this all that is in the log? For some reason there are 820
samples not represented in the output above. The ms+ hits would
Nov 29 18:05:45 coyote kernel: 9999 4

There are NO lines in the log that look like that. There are many
hundred duplicates of the 3 line pattern you see at the top of the
snip, but thats it.

AND I just discovered that I am not running the -13 kernel, but the -7
version as I had that entry set as the fallback in grub.conf. I've
since reset that to the -9 kernel as fallback and thats whats running
right now.

Which brings up something thats been an excedrin headache here since I
bought this #*&^^$^% ati 9200-SE video card to replace a failed
nvidia that took the motherboard with it when it failed. Thats the
fact that I boot blind. From the grub screen to the start of the
rc.sysinit script, this machine has no fonts! Sure, I see the curser
dancing back and forth over the bottom line of the screen, but
nothing is displayed, and this is true whether the fonts are compiled
in or as modules, whether the radeonfb is compiled in or as a module,
and with or without the VGA(16) or vesafb, compiled in or as a
module.

So its doing an automatic reboot to the kernel set as default, and
doing it during the time that I am blind from a lack of fonts.

FWIW, I've mentioned this on a couple of occasions but no one has
commented one way or the other.

One thing I haven't tried is to take all frame buffers out of the
kernel, not even as available modules, in order to force it to use
the cards own display. That would effectively convert this machine
into a 66 mhz 486, but I guess its worth the test. I'll be back
after trying that.

Post by K.R. Foley
Not that it is related to the missing output, but does tvtime use
polling on the rtc?

Sorry. All I can tell you is that I did have to put the rtc into the
kernel before tvtime would run well at all.

Post by K.R. Foley
kr

Gene Heskett

2004-11-30 04:54:05 UTC

Post by K.R. Foley

Post by Ingo Molnar
make that -31-13 (or later). Earlier kernels had a bug in where

[...]

Post by K.R. Foley
Is this all that is in the log? For some reason there are 820
samples not represented in the output above. The ms+ hits would
Nov 29 18:05:45 coyote kernel: 9999 4

Ok, I finally got -13 to run (typo in grub), and you are now correct
in that the final entry in the log after I shut tvtime down is like
this:

Nov 29 23:43:40 coyote kernel: rtc latency histogram of {tvtime/3911,
10430 samples}:
Nov 29 23:43:40 coyote kernel: 4 51
Nov 29 23:43:40 coyote kernel: 5 2058
Nov 29 23:43:40 coyote kernel: 6 3594
Nov 29 23:43:40 coyote kernel: 7 1270
Nov 29 23:43:40 coyote kernel: 8 473
Nov 29 23:43:40 coyote kernel: 9 299
Nov 29 23:43:40 coyote kernel: 10 252
Nov 29 23:43:40 coyote kernel: 11 209
Nov 29 23:43:40 coyote kernel: 12 215
Nov 29 23:43:40 coyote kernel: 13 345
Nov 29 23:43:40 coyote kernel: 14 391
Nov 29 23:43:40 coyote kernel: 15 248
Nov 29 23:43:40 coyote kernel: 16 113
Nov 29 23:43:40 coyote kernel: 17 55
Nov 29 23:43:40 coyote kernel: 18 17
Nov 29 23:43:40 coyote kernel: 19 11
Nov 29 23:43:40 coyote kernel: 20 4
Nov 29 23:43:40 coyote kernel: 21 1
Nov 29 23:43:40 coyote kernel: 23 2
Nov 29 23:43:40 coyote kernel: 28 1
Nov 29 23:43:40 coyote kernel: 4612 1
Nov 29 23:43:40 coyote kernel: 9999 820

What does this tell you now? The last 2 lines look a bit strange to
me. Particularly since the runtime was random enough that your
previous comment about the number 820 was a mssing count, and what
came out of a seperate run IS an 820.

I find that a bit hard to believe that I timed those two runs
identically.

K.R. Foley

2004-11-30 15:26:25 UTC

Post by K.R. Foley

Post by Ingo Molnar
make that -31-13 (or later). Earlier kernels had a bug in where

[...]

Post by K.R. Foley
Is this all that is in the log? For some reason there are 820
samples not represented in the output above. The ms+ hits would
Nov 29 18:05:45 coyote kernel: 9999 4

Ok, I finally got -13 to run (typo in grub), and you are now correct
in that the final entry in the log after I shut tvtime down is like
Nov 29 23:43:40 coyote kernel: rtc latency histogram of {tvtime/3911,
Nov 29 23:43:40 coyote kernel: 4 51
Nov 29 23:43:40 coyote kernel: 5 2058
Nov 29 23:43:40 coyote kernel: 6 3594
Nov 29 23:43:40 coyote kernel: 7 1270
Nov 29 23:43:40 coyote kernel: 8 473
Nov 29 23:43:40 coyote kernel: 9 299
Nov 29 23:43:40 coyote kernel: 10 252
Nov 29 23:43:40 coyote kernel: 11 209
Nov 29 23:43:40 coyote kernel: 12 215
Nov 29 23:43:40 coyote kernel: 13 345
Nov 29 23:43:40 coyote kernel: 14 391
Nov 29 23:43:40 coyote kernel: 15 248
Nov 29 23:43:40 coyote kernel: 16 113
Nov 29 23:43:40 coyote kernel: 17 55
Nov 29 23:43:40 coyote kernel: 18 17
Nov 29 23:43:40 coyote kernel: 19 11
Nov 29 23:43:40 coyote kernel: 20 4
Nov 29 23:43:40 coyote kernel: 21 1
Nov 29 23:43:40 coyote kernel: 23 2
Nov 29 23:43:40 coyote kernel: 28 1
Nov 29 23:43:40 coyote kernel: 4612 1
Nov 29 23:43:40 coyote kernel: 9999 820
What does this tell you now? The last 2 lines look a bit strange to
me. Particularly since the runtime was random enough that your
previous comment about the number 820 was a mssing count, and what
came out of a seperate run IS an 820.
I find that a bit hard to believe that I timed those two runs
identically.

When I mentioned 820 that was just what was missing from the histogram
data vs. the total samples. As for why this number showed up again, dumb
luck? What the histogram represents is usecs (left column) between
interrupt and read. The right column is the number of occurrences. What
this tells me is that you had 820 samples that were greater than 9.999ms.

I took a brief look at the tvtime code. I noticed several things:
1) tvtime sets the interrupt frequency for the rtc to 1024 Hz which
would imply that you should have had roughly 30000 samples in a 30
second run, minus initial setup time. The histogram only shows 10430
samples. I have no way of knowing if the setup time is really that long
or if you are missing some interrupts. Unfortunately the rtc histogram
stuff doesn't currently track missed interrupts except to print out the
messages you were seeing

"<some process> is being piggy... Read missed before next interrupt"

2) tvtime is probably running at a RT priority of 99. The IRQ handler
for the rtc defaults to 48-49 (I think). If you didn't already do so,
you should bump the priority up as in:

chrt -f -p 99 `/sbin/pidof 'IRQ 8'`

Try that and see if it helps at all.

kr

Gene Heskett

2004-11-30 16:24:18 UTC

Post by K.R. Foley
"<some process> is being piggy... Read missed before next interrupt"
2) tvtime is probably running at a RT priority of 99. The IRQ
handler for the rtc defaults to 48-49 (I think). If you didn't
chrt -f -p 99 `/sbin/pidof 'IRQ 8'`

[***@coyote root]# chrt -f -p 99 `/sbin/pidof 'IRQ 8'`
bash: chrt: command not found

chrt is an unknown command here. WTH? Basicly an FC2 system.

Post by K.R. Foley
Try that and see if it helps at all.
kr

K.R. Foley

2004-11-30 16:57:12 UTC

bash: chrt: command not found
chrt is an unknown command here. WTH? Basicly an FC2 system.

It's part of schedutils pkg.

kr

Zwane Mwaikambo

2004-11-30 16:52:41 UTC

bash: chrt: command not found
chrt is an unknown command here. WTH? Basicly an FC2 system.

Install the package first (from an FC2 system)

***@r3000 ~ {0:1} rpm -qif `which chrt`
Name : schedutils Relocations: (not relocatable)
Version : 1.3.0 Vendor: Red Hat, Inc.
Release : 6 Build Date: Tue 17 Feb 2004
10:16:15 MST
Install Date: Tue 13 Jul 2004 11:13:52 MDT Build Host: tweety.devel.redhat.com
Group : Applications/System Source RPM: schedutils-1.3.0-6.src.rpm
Size : 39412 License: GPL
Signature : DSA/SHA1, Thu 06 May 2004 16:36:57 MDT, Key ID b44269d04f2a6fd2
Packager : Red Hat, Inc. <http://bugzilla.redhat.com/bugzilla>
Summary : Utilities for manipulating process scheduler attributes
Description : schedutils is a set of utilities for retrieving and
manipulating process scheduler-related attributes, such as real-time
parameters and CPU affinity.

This package includes the chrt and taskset utilities.

Install this package if you need to set or get scheduler-related
attributes.

Gene Heskett

2004-12-01 07:16:11 UTC

Post by Zwane Mwaikambo

Post by K.R. Foley
"<some process> is being piggy... Read missed before next
interrupt"
2) tvtime is probably running at a RT priority of 99. The IRQ
handler for the rtc defaults to 48-49 (I think). If you didn't
chrt -f -p 99 `/sbin/pidof 'IRQ 8'`

bash: chrt: command not found
chrt is an unknown command here. WTH? Basicly an FC2 system.

Install the package first (from an FC2 system)
Name : schedutils Relocations: (not
relocatable) Version : 1.3.0
Vendor: Red Hat, Inc. Release : 6
Build Date: Tue 17 Feb 2004 10:16:15 MST
tweety.devel.redhat.com Group : Applications/System
Source RPM: schedutils-1.3.0-6.src.rpm Size : 39412
License: GPL Signature : DSA/SHA1, Thu 06 May
2004 16:36:57 MDT, Key ID b44269d04f2a6fd2 Packager : Red Hat,
Inc. <http://bugzilla.redhat.com/bugzilla> Summary : Utilities
schedutils is a set of utilities for retrieving and manipulating
process scheduler-related attributes, such as real-time parameters
and CPU affinity.
This package includes the chrt and taskset utilities.
Install this package if you need to set or get scheduler-related
attributes.

Ok, I installed it just now, and ran it twice, once before running
tvtime, and once after, but it made no visible difference in the log
input, which is still being very verbose and the delays it reports
also remain the same.

So it would appear that this isn't the nail my hammer needs. :-)

Rui Nuno Capela

2004-11-30 10:29:44 UTC

Post by Ingo Molnar
but please try to the -31-10 kernel that i've just uploaded, it has a

make that -31-13 (or later). Earlier kernels had a bug in where the
process name tracking only worked for the first latency trace saved,
subsequent traces showed 'unknown' for the process name. In -13 i've
also added a printk that shows the latest user latency in a one-line
(gettimeofday/3671/CPU#0): new 3068 us user-latency.
(gettimeofday/3784/CPU#0): new 1008627 us user-latency.
(this should also make it easier for helper scripts to save the traces,
whenever they happen.)

OK.

I did my homework and managed to grab some latency traces for you to take
a look and see if it's alright. Please.

On attach you may find the files:

xruntrace1.1.tar.gz
- scripts used to arm and capture the latency traces;

jack-0.99.14.0xruntrace-alsa.patch.gz
- the actual patch applied to jack for trace auto-triggering;

config-2.6.10-rc2-mm3-RT-V0.7.31-13.gz
- the running kernel configuration, as taken from /proc/config.gz

xruntrace1-2.6.10-rc2-mm3-RT-V0.7.31-13-*.trc.gz
- the captured traces, as dumped with xruntrace1_watch.sh script.

jack_test3.2.tar.gz
- updated my jack_test3 benchmarking suite.

The traces were captured, while running a usual KDE desktop session (from
qjackctl) on my ***@2.53Ghz laptop (Mandrake 10.1):

jackd -R -P60 -dalsa -dhw:0 -r44100 -p64 -n2 -S -P

Each trace shows only the first XRUN occurrence on a distinct jackd
session. Every other trace were triggered and captured after restarting
jackd.

Must say that XRUNs are rare but do occurr. Using the same jackd command
line, my jack_test3 suite is showing these figures by now:

*********** CONSOLIDATED RESULTS ************
Total seconds ran . . . . . . : 7200
Number of clients . . . . . . : 16
Ports per client . . . . . . : 4
*********************************************
Summary Result Count . . . . : 2 / 2
*********************************************
Timeout Rate . . . . . . . . :( 0.0)/hour
XRUN Rate . . . . . . . . . . : 13.8 /hour
Delay Rate (>spare time) . . : 13.8 /hour
Delay Rate (>1000 usecs) . . : 7.0 /hour
Delay Maximum . . . . . . . . : 1022 usecs
Cycle Maximum . . . . . . . . : 1048 usecs
Average DSP Load. . . . . . . : 62.2 %
Average CPU System Load . . . : 20.6 %
Average CPU User Load . . . . : 48.5 %
Average CPU Nice Load . . . . : 0.0 %
Average CPU I/O Wait Load . . : 0.0 %
Average CPU IRQ Load . . . . : 0.0 %
Average CPU Soft-IRQ Load . . : 0.0 %
Average Interrupt Rate . . . : 1689.2 /sec
Average Context-Switch Rate . : 19778.2 /sec
*********************************************

The IRQ process/threading priority status has been fixed as shown below:

PID CLS RTPRIO NI PRI %CPU STAT COMMAND
2 FF 90 - 130 0.0 S IRQ 0
728 FF 80 - 120 0.0 S< IRQ 8
2943 FF 70 - 110 0.0 S< IRQ 5
1500 FF 60 - 100 0.0 S< IRQ 10
841 FF 50 - 90 0.0 S< IRQ 1
736 FF 49 - 89 0.0 S< IRQ 12
20 FF 48 - 88 0.0 S< IRQ 9
814 FF 45 - 85 0.0 S< IRQ 6
821 FF 44 - 84 0.0 S< IRQ 14
823 FF 43 - 83 0.0 S< IRQ 15
2227 FF 40 - 80 0.0 S< IRQ 11

# cat /proc/interrupts

CPU0
0: 6435186 XT-PIC timer 0/35186
1: 1164 XT-PIC i8042 1/1164
2: 0 XT-PIC cascade 0/0
5: 73 XT-PIC ALI 5451 0/72
8: 1 XT-PIC rtc 0/1
9: 20602 XT-PIC acpi 0/20602
10: 0 XT-PIC ohci_hcd, ohci_hcd 0/0
11: 5 XT-PIC yenta 5/5
12: 64317 XT-PIC i8042 0/64317
14: 11772 XT-PIC ide0 0/11772
15: 19 XT-PIC ide1 0/19
NMI: 0
LOC: 0
ERR: 0
MIS: 0

Cheers.

--
rncbc aka Rui Nuno Capela
***@rncbc.org

Ingo Molnar

2004-11-30 13:19:56 UTC

Post by Rui Nuno Capela
xruntrace1-2.6.10-rc2-mm3-RT-V0.7.31-13-*.trc.gz
- the captured traces, as dumped with xruntrace1_watch.sh script.
Each trace shows only the first XRUN occurrence on a distinct jackd
session. Every other trace were triggered and captured after
restarting jackd.

interesting - the trace shows only a latency of 20-40 usecs. Perhaps the
tracing should be done over two periods?

Could you try to hack alsa_driver.c to do the gettimeofday(1,1) call for
only every second case. Something like this ought to be enough:

{
static count = 0;

if (!(count++ & 1))
gettimeofday(0,1);
}

(only one thread accesses this particular variable so there are no
threading issues.)

with this variant there's a 50% chance that we get the trace of the last
2 poll() instances. (and there's a 50% chance that we get only 1 period
covered.)

in any case, the scripts & approach seems to be almost there, i hope we
only need the above change to see the true source of the xruns.

Ingo

Rui Nuno Capela

2004-11-30 15:39:58 UTC

Ingo Molnar

interesting - the trace shows only a latency of 20-40 usecs. Perhaps the
tracing should be done over two periods?
Could you try to hack alsa_driver.c to do the gettimeofday(1,1) call for
{
static count = 0;
if (!(count++ & 1))
gettimeofday(0,1);
}
(only one thread accesses this particular variable so there are no
threading issues.)
with this variant there's a 50% chance that we get the trace of the last
2 poll() instances. (and there's a 50% chance that we get only 1 period
covered.)
in any case, the scripts & approach seems to be almost there, i hope we
only need the above change to see the true source of the xruns.

Done that.

New XRUN traces are attached, while running RT-V0.7.31-15 now. However, I
don't seem to get any notorious difference on the results, since previous
ones. All latencies traced ca. 26-27 usecs.

Bye.

--
rncbc aka Rui Nuno Capela
***@rncbc.org

Ingo Molnar

2004-11-30 16:42:07 UTC

Post by Rui Nuno Capela
Done that.
New XRUN traces are attached, while running RT-V0.7.31-15 now.
However, I don't seem to get any notorious difference on the results,
since previous ones. All latencies traced ca. 26-27 usecs.

hm, weird. Next idea: could you activate trace_freerunning and change
the alsa_driver.c count hack to do:

if (!count) {
gettimeofday(0,1);
count = 1;
}

this means that tracing will only be activated once, and the tracing
goes on non-stop from that point on, up until the first xrun, at which
point the full tracebuffer is saved. Ok?

Ingo

Ingo Molnar

2004-12-01 10:32:51 UTC

Post by Ingo Molnar
if (!(count++ & 1))
gettimeofday(0,1);
}

Done that.
New XRUN traces are attached, while running RT-V0.7.31-15 now.
However, I don't seem to get any notorious difference on the results,
since previous ones. All latencies traced ca. 26-27 usecs.

ah, found the reason why the trace length didnt improve:

+ { // Trigger off trace every other poll...
+ static unsigned int xruntrace_count = 0;
+ if ((xruntrace_count++ % 1) == 0)
+ gettimeofday(0,1);
+ }

this should either be '& 1' as i originally suggested, or '% 2'. The way
it is right now it will stop the trace every time - i.e. what we had
before.

so please disregard my freerunning suggestionsand try the two-periods
solution first.

Ingo

Ingo Molnar

2004-12-01 11:25:20 UTC

Post by Rui Nuno Capela
New XRUN traces are attached, while running RT-V0.7.31-15 now.
However, I don't seem to get any notorious difference on the results,
since previous ones. All latencies traced ca. 26-27 usecs.

another thing: could you try the latest jackd CVS tree too? I found
spurious xruns in 99-0 (the last official release), while with 99-11 i
dont see any xruns on my box, up until i completely saturate CPU time.

I.e. both latest -RT kernels are needed for this (earlier RT kernels,
prior the PI fixes still generated lots of xruns even with latest jack
CVS), and latest jack CVS is needed.

Ingo

Rui Nuno Capela

2004-12-01 12:49:44 UTC

another thing: could you try the latest jackd CVS tree too? I found
spurious xruns in 99-0 (the last official release), while with 99-11 i
dont see any xruns on my box, up until i completely saturate CPU time.
I.e. both latest -RT kernels are needed for this (earlier RT kernels,
prior the PI fixes still generated lots of xruns even with latest jack
CVS), and latest jack CVS is needed.

I am using the latest-as-it-could-be jack CVS snapshot, as handed by Jack
O'Quin, since sourceforge's CVS service has been broken for a while.

The latest of my last report is 0.99.15.

--
rncbc aka Rui Nuno Capela
***@rncbc.org

Rui Nuno Capela

2004-12-01 12:47:09 UTC

Post by Ingo Molnar
if (!(count++ & 1))
gettimeofday(0,1);
}

Done that.
New XRUN traces are attached, while running RT-V0.7.31-15 now.
However, I don't seem to get any notorious difference on the results,
since previous ones. All latencies traced ca. 26-27 usecs.

+ { // Trigger off trace every other poll...
+ static unsigned int xruntrace_count = 0;
+ if ((xruntrace_count++ % 1) == 0)
+ gettimeofday(0,1);
+ }
this should either be '& 1' as i originally suggested, or '% 2'. The way
it is right now it will stop the trace every time - i.e. what we had
before.
so please disregard my freerunning suggestionsand try the two-periods
solution first.

Very good eyes! It was all my mistake. Sorry, sorry. It was a sure typo
after all, the '%' key is right next to '&', at least on my on my keyboard
:)

Will correct it and bbl...

--
rncbc aka Rui Nuno Capela
***@rncbc.org

Ingo Molnar

2004-12-01 15:40:47 UTC

ok, got the latest traces from Rui offlist.

here's a summary of an xrun that went like this:

jackd-4711 00000000 0.002ms (+0.000ms): do_poll (sys_poll)
...
jackd-4711 80000002 0.012ms (+0.000ms): dequeue_task (deactivate_task)
...
IRQ 5-3052 ........ 1.946ms (+0.000ms): -> jackd-4711
[ 00000027 00000001 ]: try_to_wake_up
...
qjackctl-4719 ........ 2.362ms (+0.000ms): common_interrupt:
[ b75e9090 00000005 00000000 ]
...
IRQ 5-3052 00000000 2.473ms (+0.000ms): xrun (snd_pcm_period_elapsed)
...

this shows an interesting phenomenon: for whatever reason IRQ5's thread
didnt run until timestamp 1.946ms. There is lots of idle time between
timestamps 0.012ms and 1.946ms, so this must be some weird condition.

Rui, could you activate trace_freerunning and send me a couple of 'big'
traces that result out of it? Apparently there is some other, earlier
activity that caused this to happen to IRQ5. Nothing seems to have woken
up IRQ5 when it ran, which is weird by itself too.

if IRQ thread 5 got delayed somehow that might explain the xrun: the
hardware interrupt had no chance to run. But how IRQ5 could stay on the
runqueue while the idle thread managed to run is strange.

Ingo

Ingo Molnar

2004-12-01 16:20:34 UTC

ok, this could be ACPI CPU-sleep related. Could you disable all ACPI
options in your .config (as a workaround), and re-check whether the
xruns still occur?

i think i found the bug - it's an upstream ACPI bug. Does the patch
below (or the -31-19 kernel, which i've just uploaded) fix the xruns?

the upstream ACPI bug is this: we check need_resched() _before_
disabling preemption. This opens up the following scenario:

swapper: !need_resched()
[IRQ context]
wakes up a task
marks idle task as need-resched

swapper: acpi_processor_idle(); // sleeps until next irq

instant 1msec latency introduced...

normally default_idle() is safe because it re-checks need_resched with
interrupts disabled before it truly halts the CPU. But
acpi_processor_idle() doesnt seem to be doing this! Your trace clearly
shows a missed preemption due to ACPI. I'm wondering why no-one has
triggered this before, it's a really bad bug that should be fixed in
2.6.10.

Ingo

Signed-off-by: Ingo Molnar <***@elte.hu>

--- linux/drivers/acpi/processor.c.orig
+++ linux/drivers/acpi/processor.c
@@ -370,6 +370,15 @@ acpi_processor_idle (void)
*/
local_irq_disable();

+ /*
+ * Check whether we truly need to go idle, or should
+ * reschedule:
+ */
+ if (unlikely(need_resched())) {
+ local_irq_enable();
+ return;
+ }
+
cx = &(pr->power.states[pr->power.state]);

/*

Ingo Molnar

2004-12-01 16:31:03 UTC

Post by Ingo Molnar
i think i found the bug - it's an upstream ACPI bug. Does the patch
below (or the -31-19 kernel, which i've just uploaded) fix the xruns?
the upstream ACPI bug is this: we check need_resched() _before_
swapper: !need_resched()
[IRQ context]
wakes up a task
marks idle task as need-resched
swapper: acpi_processor_idle(); // sleeps until next irq
instant 1msec latency introduced...
normally default_idle() is safe because it re-checks need_resched with
interrupts disabled before it truly halts the CPU. But
acpi_processor_idle() doesnt seem to be doing this! Your trace clearly
shows a missed preemption due to ACPI. I'm wondering why no-one has
triggered this before, it's a really bad bug that should be fixed in
2.6.10.

i've attached the particular trace from Rui's collection of traces which
nicely shows the ACPI race in action. Here are the relevant events:

swapper-0 08000000 0.840ms (+0.000ms): preempt_schedule (cpu_idle)
swapper-0 ........ 0.913ms (+0.000ms): -> IRQ 5-3052 [ 0000001d 0000008c ]: wake_up_process
swapper-0 08000001 0.913ms (+0.920ms): acpi_processor_idle (cpu_idle)
swapper-0 ........ 1.834ms (+0.000ms): common_interrupt: [ e0022266 00000000 00000000 ]
swapper-0 88010001 1.834ms (+0.000ms): direct_timer_interrupt (do_IRQ)
IRQ 5-3052 80000002 1.848ms (+0.000ms): __switch_to (__sched_text_start)

i.e. at timestamp 0.840ms cpu_idle() got preempted (it was after the
need_resched() check but before the rcu_read_lock() which disables
preemption), then at 0.913ms a task was woken up by the hardirq context,
but despite this we called into acpi_processor_idle() which slept
dutifully until the next timer interrupt hit us. Then did only the woken
up task schedule. (which caused an xrun in Rui's setup.)

Ingo

Rui Nuno Capela

2004-12-01 18:59:25 UTC

ok, this could be ACPI CPU-sleep related. Could you disable all ACPI
options in your .config (as a workaround), and re-check whether the
xruns still occur?

i think i found the bug - it's an upstream ACPI bug. Does the patch
below (or the -31-19 kernel, which i've just uploaded) fix the xruns?

So far so good.

I have now an armed RT-V0.31-19 on my laptop, running for about 30 minutes
already, running the usual 'jackd -R -dalsa -dhw:0 -r44100 -p64 -n2 -S -P'
from qjackctl, several qsynth (fluidsynth) engines are up and running,
normal desktop janitor work (KDE) and yes, ACPI is on. Oh, wi-fi is also
pumping nice :)

Guess what? No XRUNs, not even one to show you a single latency trace.

I'll keep you posted, if anything goes less right than it is :).

--
rncbc aka Rui Nuno Capela
***@rncbc.org

Ingo Molnar

2004-12-01 21:30:23 UTC

Post by Rui Nuno Capela
Guess what? No XRUNs, not even one to show you a single latency trace.
I'll keep you posted, if anything goes less right than it is :).

great - it seems the ACPI fix made a difference. If there are no xruns
then you might want to try the # of engines at which point xruns show

^-- to increase

up. (can the soundcard period size / buffering be reduced further, to
make it more sensitive to scheduling latencies?)

Ingo Molnar

2004-12-01 21:58:37 UTC

up. (can the soundcard period size / buffering be reduced further, to
make it more sensitive to scheduling latencies?)

OK.
A couple of hours later, as I thrown in some more jack clients into
the picture, the XRUNs started to appear, but very discrete still.

just curious, what type of CPU load did this create - over 50%?

Ingo

Rui Nuno Capela

2004-12-01 22:04:46 UTC

up. (can the soundcard period size / buffering be reduced further, to
make it more sensitive to scheduling latencies?)

OK.
A couple of hours later, as I thrown in some more jack clients into
the picture, the XRUNs started to appear, but very discrete still.

just curious, what type of CPU load did this create - over 50%?

At this very moment? About 75%. Here's a top snapshot:

top - 22:03:48 up 3:36, 4 users, load average: 1.91, 2.73, 2.16
Tasks: 86 total, 1 running, 85 sleeping, 0 stopped, 0 zombie
Cpu(s): 33.9% us, 39.7% sy, 0.0% ni, 26.5% id, 0.0% wa, 0.0% hi, 0.0% si
Mem: 500900k total, 411308k used, 89592k free, 12680k buffers
Swap: 506008k total, 0k used, 506008k free, 209300k cached

PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
18300 rncbc 16 0 89128 87m 14m S 30.8 17.8 13:02.31 qsynth
18288 rncbc 20 0 28148 27m 2316 S 10.8 5.6 4:29.60 jackd
3078 root -71 -5 0 0 0 S 8.9 0.0 16:33.69 IRQ 5
16701 rncbc 15 0 26988 26m 14m S 7.3 5.4 4:05.01 qjackctl
3587 root 15 0 24288 13m 2696 S 3.8 2.8 9:27.28 X
4291 rncbc 15 0 28240 15m 12m S 1.9 3.1 1:44.69 kdeinit
4356 rncbc 15 0 21312 7832 5952 S 1.6 1.6 2:45.78 gkrellm
4346 rncbc 15 0 23744 12m 9.9m S 0.6 2.5 0:19.92 krandrtray
4347 rncbc 15 0 24260 12m 9m S 0.6 2.5 0:19.62 kdeinit
4355 rncbc 15 0 23684 11m 9504 S 0.6 2.3 0:16.15 kdeinit
4361 rncbc 15 0 24948 13m 11m S 0.6 2.8 0:38.53 kscd
4453 rncbc 16 0 27168 15m 12m S 0.6 3.1 0:28.73 kdeinit
24504 rncbc 17 0 2064 996 780 R 0.6 0.2 0:00.25 top
4323 rncbc 15 0 29280 13m 11m S 0.3 2.8 0:14.93 kdeinit
4330 rncbc 16 0 32296 19m 15m S 0.3 3.9 0:30.49 kdeinit
4333 rncbc 15 0 26880 14m 12m S 0.3 3.1 0:25.56 kdeinit
4337 rncbc 15 0 24680 13m 10m S 0.3 2.7 0:25.52 kamix
4345 rncbc 15 0 26500 14m 12m S 0.3 2.9 0:26.77 korgac

Bye.

--
rncbc aka Rui Nuno Capela
***@rncbc.org

Ingo Molnar

2004-12-01 22:09:16 UTC

Here goes attached the captured latency traces for you to check. I
guess this time xmms-jack its at stake 'coz it was the cherry on top
of the cake, for three times in a row :)

the latency does seem to be related to clients, as in the following
trace:

xruntrace1-2.6.10-rc2-mm3-RT-V0.7.31-19-20041201202926.trc

jackd-16714 goes into sys_poll() at timestamp 0.850ms, and does a pipe
related wait:

jackd-16714 00000000 0.853ms (+0.000ms): pipe_poll (do_pollfd)

and goes to sleep:

jackd-16714 00000000 0.857ms (+0.000ms): schedule_timeout (do_poll)

almost 3 milliseconds later, at timestamp 3.404ms, the pipe related
timer expires and jackd gets scheduled again:

ksoftirq-3 00000000 3.403ms (+0.000ms): wake_up_process (run_timer_softirq)
ksoftirq-3 ........ 3.404ms (+0.000ms): -> jackd-16714
[ 00000027 00000003 ]: try_to_wake_up

obviously an xrun has happened during this 2.6 msec wait:

IRQ 5-3078 00000000 2.277ms (+0.000ms): xrun (snd_pcm_period_elapsed)

but ... this brings up the question, is this a valid scenario? How can
jackd block on clients for so long? Perhaps this means that every audio
buffer has run empty and jackd desperately needed some new audio input,
which it didnt get from the clients, up until after the xrun? In theory
this should only be possible if there's CPU saturation (that's why i
asked about how much CPU% time there was in use).

One indication that this might have been the case is that in the full
3.5 msecs trace there's not a single cycle spent idle. But, lots of time
was spent by e.g. X or gkrellm-4356, which are not RT tasks - so from
the RT task POV i think there were cycles left to be utilized. Could
this be a client bug? That seems a bit unlikely because to let jackd
'run empty', each and every client would have to starve it, correct?

anyway, your plan to try jack_test3.1 again looks like the correct
approach: we first need to check the behavior of 'trivial clients',
before more complex scenarios with real clients can be considered.

Ingo

Rui Nuno Capela

2004-12-01 22:31:49 UTC

Post by Ingo Molnar
the latency does seem to be related to clients, as in the following
xruntrace1-2.6.10-rc2-mm3-RT-V0.7.31-19-20041201202926.trc
jackd-16714 goes into sys_poll() at timestamp 0.850ms, and does a pipe
jackd-16714 00000000 0.853ms (+0.000ms): pipe_poll (do_pollfd)
jackd-16714 00000000 0.857ms (+0.000ms): schedule_timeout (do_poll)
almost 3 milliseconds later, at timestamp 3.404ms, the pipe related
ksoftirq-3 00000000 3.403ms (+0.000ms): wake_up_process
(run_timer_softirq)
ksoftirq-3 ........ 3.404ms (+0.000ms): -> jackd-16714
[ 00000027 00000003 ]: try_to_wake_up
IRQ 5-3078 00000000 2.277ms (+0.000ms): xrun (snd_pcm_period_elapsed)
but ... this brings up the question, is this a valid scenario? How can
jackd block on clients for so long? Perhaps this means that every audio
buffer has run empty and jackd desperately needed some new audio input,
which it didnt get from the clients, up until after the xrun? In theory
this should only be possible if there's CPU saturation (that's why i
asked about how much CPU% time there was in use).
One indication that this might have been the case is that in the full
3.5 msecs trace there's not a single cycle spent idle. But, lots of time
was spent by e.g. X or gkrellm-4356, which are not RT tasks - so from
the RT task POV i think there were cycles left to be utilized. Could
this be a client bug? That seems a bit unlikely because to let jackd
'run empty', each and every client would have to starve it, correct?

Did I mention that the XRUNs seem to occur only when xmms-jack is in the
graph? AFAICT xmms-jack uses a sort of double-buffering wrapper
(bio2jack), or in other words, its not a "native" jack client.

Post by Ingo Molnar
anyway, your plan to try jack_test3.1 again looks like the correct
approach: we first need to check the behavior of 'trivial clients',
before more complex scenarios with real clients can be considered.

Yes, and I'll have to rebuild a new RT-V0.7.31-19 kernel without latency
timing stuff, just to make sure we'll compare apples to apples ;)

Bye now.

--
rncbc aka Rui Nuno Capela
***@rncbc.org

Rui Nuno Capela

2004-12-02 09:12:28 UTC

... I'll have to rebuild a new RT-V0.7.31-19 kernel without latency
timing stuff, just to make sure we'll compare apples to apples ;)

Oh, and running also an jackd 0.99.17cvs _without_ the xrun trace
auto-trigger patch, so to end up with a production-like system.

And that was it. After a total of 4 hours running free, the jack_test3
suite spitted out these results, which speak from themselves:

*********** CONSOLIDATED RESULTS ************
Total seconds ran . . . . . . : 14400
Number of clients . . . . . . : 16
Ports per client . . . . . . : 4
*********************************************
Summary Result Count . . . . : 4 / 4
*********************************************
Timeout Rate . . . . . . . . :( 0.0)/hour
XRUN Rate . . . . . . . . . . : 0.0 /hour
Delay Rate (>spare time) . . : 0.0 /hour
Delay Rate (>1000 usecs) . . : 0.0 /hour
Delay Maximum . . . . . . . . : 77 usecs
Cycle Maximum . . . . . . . . : 1029 usecs
Average DSP Load. . . . . . . : 60.0 %
Average CPU System Load . . . : 19.6 %
Average CPU User Load . . . . : 47.7 %
Average CPU Nice Load . . . . : 0.0 %
Average CPU I/O Wait Load . . : 0.0 %
Average CPU IRQ Load . . . . : 0.0 %
Average CPU Soft-IRQ Load . . : 0.0 %
Average Interrupt Rate . . . : 1689.1 /sec
Average Context-Switch Rate . : 19781.4 /sec
*********************************************

This is frankly the best performance EVER on this laptop machine, given
these jackd parameters (-R -P60 -dalsa -dhw:0 -r44100 -p64 -n2 -S -P).

Next step is really trying to increase the stress and look after when it
will break apart. It will not take too long...

First attempts, by just increasing the client count from 16 to 20, are
leading to what will be the next "horror show" :) CPU tops above 90%, and
after a couple of minutes running steadly it enters into some kind of
turbulence and hangs. Yeah, it just freezes completely.

So I guess we're having a lot more food to mangle ;)

Stay tuned.

--
rncbc aka Rui Nuno Capela
***@rncbc.org

Rui Nuno Capela

2004-12-02 12:59:41 UTC

Post by Rui Nuno Capela
Next step is really trying to increase the stress and look after when it
will break apart. It will not take too long...
First attempts, by just increasing the client count from 16 to 20, are
leading to what will be the next "horror show" :) CPU tops above 90%, and
after a couple of minutes running steadly it enters into some kind of
turbulence and hangs. Yeah, it just freezes completely.
So I guess we're having a lot more food to mangle ;)

After a bit of investigation, I've found some evidence that the "horror
show" is primarily due to XRUN fprintf stderr storm. If I simply remove
that fprintf line from jack/drivers/alsa/alsa_driver.c (around line 1084),
I can run more than 20 jack_test3_client's without hosing the system.

Most probably, the main issue is about fprintf(stderr) being called in a
RT thread context. This is a jackd issue, not of RT kernel's. I remember
there's a jack patch, somewhere in the limbo, for queing all printable
messages out of the jackd RT threads.

I think I'm urging for that patch right now, even though I'm probably
pushing all this pressure into real/physical/hard limits ;) OK. I'll look
if I can take that back from the jackit-devel archives and see what I can
do about it.

Back to a latency-trace enabled RT-0V0.7.31-19, and a patched jackd
0.99.17cvs, I've trapped some more traces (on attachment). These were
taken while running some jack_test3 with light stress (10~12 clients, 4
ports each).

Bye now.

--
rncbc aka Rui Nuno Capela
***@rncbc.org

Fernando Lopez-Lezcano

2004-12-02 16:38:06 UTC

Yes, that's what's happening. I've been seeing this for a while under
very high loads.

Post by Rui Nuno Capela
This is a jackd issue, not of RT kernel's. I remember
there's a jack patch, somewhere in the limbo, for queing all printable
messages out of the jackd RT threads.

There is one I created a while ago, I'm attaching it. It has not been
made part of jack because a proper patch is needed (this one does not
cover all possible uses of printf, the VERBOSE macro itself should be
changed, and apparently it should be made thread safe - but I'm using
this in production machines without problems and it covers the most
obnoxious uses of printf).

Post by Rui Nuno Capela
I think I'm urging for that patch right now, even though I'm probably
pushing all this pressure into real/physical/hard limits ;) OK. I'll look
if I can take that back from the jackit-devel archives and see what I can
do about it.
Back to a latency-trace enabled RT-0V0.7.31-19, and a patched jackd
0.99.17cvs, I've trapped some more traces (on attachment). These were
taken while running some jack_test3 with light stress (10~12 clients, 4
ports each).

Rui: [MANY] thanks for the amazing job of tracking stuff you are
doing...

-- Fernando

Florian Schmidt

2004-12-01 22:43:55 UTC

On Wed, 1 Dec 2004 23:09:16 +0100

Post by Ingo Molnar
but ... this brings up the question, is this a valid scenario? How can
jackd block on clients for so long? Perhaps this means that every audio
buffer has run empty and jackd desperately needed some new audio input,
which it didnt get from the clients, up until after the xrun? In theory
this should only be possible if there's CPU saturation (that's why i
asked about how much CPU% time there was in use).
One indication that this might have been the case is that in the full
3.5 msecs trace there's not a single cycle spent idle. But, lots of time
was spent by e.g. X or gkrellm-4356, which are not RT tasks - so from
the RT task POV i think there were cycles left to be utilized. Could
this be a client bug? That seems a bit unlikely because to let jackd
'run empty', each and every client would have to starve it, correct?

actually a single client doing nasty (non RT) stuff in its process()
callback can "starve" jackd. AFAIK jackd waits until the last client has
finished its process callback. So, if some client's process callback
decides to use (for example) some blocking system call (big no no) and
consequently falls asleep for a relatively long time, then it can cause
jackd to miss its deadline. I'm not sure though wether this triggers an
xrun in jackd or just a delay exceeded message.

flo

Ingo Molnar

2004-12-02 08:40:40 UTC

Post by Florian Schmidt
On Wed, 1 Dec 2004 23:09:16 +0100

actually a single client doing nasty (non RT) stuff in its process()
callback can "starve" jackd. AFAIK jackd waits until the last client
has finished its process callback. So, if some client's process
callback decides to use (for example) some blocking system call (big
no no) and consequently falls asleep for a relatively long time, then
it can cause jackd to miss its deadline. I'm not sure though wether
this triggers an xrun in jackd or just a delay exceeded message.

since the period size in Rui's test is so small (-p64 is 1.6 msecs?),
the 2.6 msec timeout in pipe_poll() already generates an xrun.

Ingo

Florian Schmidt

2004-12-02 12:22:18 UTC

On Thu, 2 Dec 2004 09:40:40 +0100

Post by Florian Schmidt
actually a single client doing nasty (non RT) stuff in its process()
callback can "starve" jackd. AFAIK jackd waits until the last client
has finished its process callback. So, if some client's process
callback decides to use (for example) some blocking system call (big
no no) and consequently falls asleep for a relatively long time, then
it can cause jackd to miss its deadline. I'm not sure though wether
this triggers an xrun in jackd or just a delay exceeded message.

since the period size in Rui's test is so small (-p64 is 1.6 msecs?),

ca. 1.3ms

Post by Ingo Molnar
the 2.6 msec timeout in pipe_poll() already generates an xrun.

yep. now the question is: why did jackd have to wait so long for the
client? what was the client doing? was it sleeping? what client was it?
probably not the simple jack_test client, right?

Short tests ((20 minutes, due to lack of time) with cvs jackd and a
number of modified jack_test clients (added a for loop in the process()
callback to burn cycles to increase the cpu load) show no xruns here
even with relatively high cpu load (around 70%) and a periodsize of 64
frames. Of course i had the mandatory additional kernel compile running
in the background :)

flo

Oh wow. Just before hitting send i got three xruns of around
0.020-0.050msec. Ok, will read up on recent emails to see what to do to
debug these.

Ingo Molnar

2004-12-02 12:29:31 UTC

Post by Florian Schmidt

Post by Ingo Molnar
the 2.6 msec timeout in pipe_poll() already generates an xrun.

yep. now the question is: why did jackd have to wait so long for the
client? what was the client doing? was it sleeping? what client was
it? probably not the simple jack_test client, right?

it's very likely not the simple jack_test client. I've attached the
trace in question. Here are the tasks that were running:

gkrellm
IRQ 0
IRQ 14
IRQ 5
jackd
kblockd
korgac
ksoftirq
qjackctl
qsynth
X
xmms

the trace doesnt show what task jackd was waiting on, and it would be
hard to establish it, the tracepoint would have to 'discover' all other
holders of the pipe fd, which is quite complex.

Post by Florian Schmidt
Oh wow. Just before hitting send i got three xruns of around
0.020-0.050msec. Ok, will read up on recent emails to see what to do
to debug these.

which jackd version is this? I saw similar small spurious xruns with
99.0, those went away in recent CVS versions.

Ingo

Florian Schmidt

2004-12-02 13:06:12 UTC

On Thu, 2 Dec 2004 13:29:31 +0100

Post by Ingo Molnar
it's very likely not the simple jack_test client. I've attached the
gkrellm
IRQ 0
IRQ 14
IRQ 5
jackd
kblockd
korgac
ksoftirq
qjackctl
qsynth
X
xmms
the trace doesnt show what task jackd was waiting on, and it would be
hard to establish it, the tracepoint would have to 'discover' all other
holders of the pipe fd, which is quite complex.

I'm not knowledgable enough to read the trace, but what was for example
the last thing qsynth was doing? Did it go to sleep? I suppose this was
Rui's 9 qsynth's test, right?

Hmm, i wonder if there's a way to detect non RT behaviour in jackd
clients. I mean AFAIK the only thing allowed for the process callback of
on is the FIFO it waits on to be woken, right? Every other sleeping is
to be considered a bug.

Post by Florian Schmidt
Oh wow. Just before hitting send i got three xruns of around
0.020-0.050msec. Ok, will read up on recent emails to see what to do
to debug these.

which jackd version is this? I saw similar small spurious xruns with
99.0, those went away in recent CVS versions.

For this test i used jackd from yesterdays CVS. With the 0.99 version i
think i saw more of these xruns, but cannot tell (still running the CVS
version right now).. They all had in common though that they are in the
0.020-0.040msec range.

Flo

Ingo Molnar

2004-12-02 13:10:02 UTC

Post by Florian Schmidt
Hmm, i wonder if there's a way to detect non RT behaviour in jackd
clients. I mean AFAIK the only thing allowed for the process callback
of on is the FIFO it waits on to be woken, right? Every other sleeping
is to be considered a bug.

there's such a feature in -RT kernels. If a user process calls:

gettimeofday(1,1);

then the kernel turns 'atomic mode' on. To turn it off, call:

gettimeofday(1,0);

while in atomic-mode, any non-atomic activity (scheduling) will produce
a kernel message and a SIGUSR2 sent to the offending process (once,
atomic mode has to be re-enabled again for the next message). Preemption
by a higher-prio task does not trigger a message/signal.

If you run the client under gdb you should be able to catch the SIGUSR2
signal and then you can see the offending code's backtrace via 'bt'.

Ingo

Florian Schmidt

2004-12-02 13:40:37 UTC

On Thu, 2 Dec 2004 14:10:02 +0100

gettimeofday(1,1);
gettimeofday(1,0);
while in atomic-mode, any non-atomic activity (scheduling) will produce
a kernel message and a SIGUSR2 sent to the offending process (once,
atomic mode has to be re-enabled again for the next message). Preemption
by a higher-prio task does not trigger a message/signal.
If you run the client under gdb you should be able to catch the SIGUSR2
signal and then you can see the offending code's backtrace via 'bt'.

Ok, so if i want to find out whether a client violates the RT
constraints for its process callback i would have to add a call to
gettimeofday(1,1) at the start of the process callback and
gettimeofday(1,0) at the end.

Everything which causes a reschedule inbetween will then cause SIGUSR2
to be sent to the client for which i could either add a signal handler
in the client or just use gdb to get notified of it.

Cool!

Flo

Ingo Molnar

2004-12-02 13:49:34 UTC

Post by Florian Schmidt
Ok, so if i want to find out whether a client violates the RT
constraints for its process callback i would have to add a call to
gettimeofday(1,1) at the start of the process callback and
gettimeofday(1,0) at the end.
Everything which causes a reschedule inbetween will then cause SIGUSR2
to be sent to the client for which i could either add a signal handler
in the client or just use gdb to get notified of it.

correct. I'd expect there to be a number of less critical reschedules
happening around startup/shutdown of a client, which one could consider
a false positive, but there should be no unexpected rescheduling while
the client is up and running.

Ingo

Florian Schmidt

2004-12-02 16:08:08 UTC

On Thu, 2 Dec 2004 14:49:34 +0100

cool, we're discussing on jackit-devel if this could be added to jackd's
libjack which would do the gettimeofday calls right before and after
calling the process() callback. This might indeed be very very very very
useful as in this case the clients themselfes wouldn't need to be
changed and jackd might then print a console message for example when a
client does something nasty.

Flo

Florian Schmidt

2004-12-02 17:44:21 UTC

On Thu, 2 Dec 2004 14:49:34 +0100

Ok,

this simple patch adds the gettimeofday calls around the calling of the
process callback:

--- libjack/client.c.orig 2004-12-02 17:55:04.000000000 +0100
+++ libjack/client.c 2004-12-02 17:56:23.000000000 +0100
@@ -1238,6 +1238,9 @@
if (control->sync_cb)
jack_call_sync_client (client);

+ // enable atomicity check for RP kernels
+ gettimeofday(1,1);
+
if (control->process) {
if (control->process (control->nframes,
control->process_arg)
@@ -1247,7 +1250,10 @@
} else {
control->state = Finished;
}
-
+
+ // disable atomicity check
+ gettimeofday(0,1);
+
if (control->timebase_cb)
jack_call_timebase_master (client);

The results i see are rather interesting though. Even with a noop jack
client (which does nothing but return 0 in the process callback) i get a
syslog report everytime i start the client. Client source attached.

Dec 2 18:39:06 mango kernel: jack_test:22743 userspace BUG: scheduling in user-atomic context!
Dec 2 18:39:06 mango kernel: [<c02a38b6>] schedule+0x76/0x130 (8)
Dec 2 18:39:06 mango kernel: [<c02a44c5>] schedule_timeout+0x85/0xe0 (36)
Dec 2 18:39:06 mango kernel: [<c016677f>] do_pollfd+0x4f/0x90 (48)
Dec 2 18:39:06 mango kernel: [<c011ceb0>] process_timeout+0x0/0x10 (8)
Dec 2 18:39:06 mango kernel: [<c016686a>] do_poll+0xaa/0xd0 (20)
Dec 2 18:39:06 mango kernel: [<c01669e2>] sys_poll+0x152/0x230 (48)
Dec 2 18:39:06 mango kernel: [<c0165db0>] __pollwait+0x0/0xd0 (36)
Dec 2 18:39:06 mango kernel: [<c01025cb>] syscall_call+0x7/0xb (32)

The atomicity check operates on a per task (thread) basis right?

Flo

Florian Schmidt

2004-12-02 21:12:51 UTC

On Thu, 2 Dec 2004 18:44:21 +0100

Post by Florian Schmidt
this simple patch adds the gettimeofday calls around the calling of the

[snip]

Post by Florian Schmidt
The results i see are rather interesting though. Even with a noop jack
client (which does nothing but return 0 in the process callback) i get a
syslog report everytime i start the client. Client source attached.

[snip]

Post by Florian Schmidt
The atomicity check operates on a per task (thread) basis right?
Flo

Well, my error. gettimeofday(1,0) != gettimeofday(0,1) :)

flo

--
Palimm Palimm!
Sounds/Ware:
http://affenbande.org/~tapas/

Rui Nuno Capela

2004-12-02 13:18:56 UTC

Post by Florian Schmidt

I'm not knowledgable enough to read the trace, but what was for example
the last thing qsynth was doing? Did it go to sleep? I suppose this was
Rui's 9 qsynth's test, right?

No, it wasn't my old 8 fluidsynths test. It was just one normal desktop
work session. At the time, IIRC, the only jack clients that were running
were:

qjackctl (obviously :)
qsynth (3 fluidsynth engines, many soundfonts loaded).
xmms (via xmms-jack output plugin)

By just taking xmms out from the graph, I can run my desktop environment
(KDE 3.2) for hours without a single XRUN or noticeable delay. And just
before RT-V0.9.31-19 had arrived, that was just a dream ;)

Bye.

--
rncbc aka Rui Nuno Capela
***@rncbc.org

Ingo Molnar

2004-12-01 21:29:25 UTC

Post by Ingo Molnar
i think i found the bug - it's an upstream ACPI bug. Does the patch
below (or the -31-19 kernel, which i've just uploaded) fix the xruns?

So far so good.
I have now an armed RT-V0.31-19 on my laptop, running for about 30
minutes already, running the usual 'jackd -R -dalsa -dhw:0 -r44100
-p64 -n2 -S -P' from qjackctl, several qsynth (fluidsynth) engines are
up and running, normal desktop janitor work (KDE) and yes, ACPI is on.
Oh, wi-fi is also pumping nice :)
Guess what? No XRUNs, not even one to show you a single latency trace.
I'll keep you posted, if anything goes less right than it is :).

great - it seems the ACPI fix made a difference. If there are no xruns
then you might want to try the # of engines at which point xruns show
up. (can the soundcard period size / buffering be reduced further, to
make it more sensitive to scheduling latencies?)

Ingo

Fernando Lopez-Lezcano

2004-12-03 01:41:58 UTC

ok, this could be ACPI CPU-sleep related. Could you disable all ACPI
options in your .config (as a workaround), and re-check whether the
xruns still occur?

i think i found the bug - it's an upstream ACPI bug. Does the patch
below (or the -31-19 kernel, which i've just uploaded) fix the xruns?

I'm trying unsuccessfully to boot my dual processor athlon into an smp
V0.7.31-20 patched kernel, with PREEMPT_DESKTOP (has not booted
correctly with realtime preemption compiled in for a long time):

(copied from the screen)

BUG: scheduling while atomic: swapper/0x10000000/1
caller is preempt_schedule_irq+0x34/0x50
[<c0105303>] dump_stack+-x23/0x30 (20)
[<c03726dc>] __schedule+0xd5c/0xdb0 (124)
[<c0372904>] preempt_schedule_irq+0x34/0x50 (12)
[<c010429b>] need_resched+0x20/0x29 (-7256)
-------
|preempt_count: 10000001
1-level deep critical section nesting
.. [<c014610d>] .... preint_tracs+0x1d/0x60
.....[<c0105303>] .. ( <= dump_stack+0x23/0x30)
BUG at kernel/latency.c: 918!
--------
invalid operand: 0000 [#1]
PREEMPT SMP

Tried booting with softirq-preempt=1 hardirq-preempt=0 with no change.

The matching uniprocessor kernel boots fine.
-- Fernando

Fernando Lopez-Lezcano

2004-12-03 02:23:00 UTC

Post by Fernando Lopez-Lezcano

ok, this could be ACPI CPU-sleep related. Could you disable all ACPI
options in your .config (as a workaround), and re-check whether the
xruns still occur?

i think i found the bug - it's an upstream ACPI bug. Does the patch
below (or the -31-19 kernel, which i've just uploaded) fix the xruns?

I'm trying unsuccessfully to boot my dual processor athlon into an smp
V0.7.31-20 patched kernel, with PREEMPT_DESKTOP (has not booted
(copied from the screen)
BUG: scheduling while atomic: swapper/0x10000000/1
caller is preempt_schedule_irq+0x34/0x50
[<c0105303>] dump_stack+-x23/0x30 (20)
[<c03726dc>] __schedule+0xd5c/0xdb0 (124)
[<c0372904>] preempt_schedule_irq+0x34/0x50 (12)
[<c010429b>] need_resched+0x20/0x29 (-7256)
-------
|preempt_count: 10000001
1-level deep critical section nesting
.. [<c014610d>] .... preint_tracs+0x1d/0x60
.....[<c0105303>] .. ( <= dump_stack+0x23/0x30)
BUG at kernel/latency.c: 918!
--------
invalid operand: 0000 [#1]
PREEMPT SMP
Tried booting with softirq-preempt=1 hardirq-preempt=0 with no change.
The matching uniprocessor kernel boots fine.

But on shutdown (up kernel):

BUG at mm/truncate.c:48!
---- ...
kernel BUG at mm/truncate.c:48!
invalid operand: 0000 [#1]
PREEMPT

-- Fernando

Ingo Molnar

2004-12-01 16:06:32 UTC

Post by Ingo Molnar
this shows an interesting phenomenon: for whatever reason IRQ5's
thread didnt run until timestamp 1.946ms. There is lots of idle time
between timestamps 0.012ms and 1.946ms, so this must be some weird
condition.

ok, this could be ACPI CPU-sleep related. Could you disable all ACPI
options in your .config (as a workaround), and re-check whether the
xruns still occur?

Ingo

Remi Colinet

2004-11-30 18:13:26 UTC

Hi Ingo,

I'm getting this error with V0.7.31-13

CC kernel/wait.o
CC kernel/kfifo.o
CC kernel/sys_ni.o
CC kernel/rt.o
CC kernel/latency.o
kernel/latency.c: In function `check_critical_timing':
kernel/latency.c:730: too few arguments to function `___trace'
kernel/latency.c:730: warning: too few arguments passed to inline
function, suppressing inlining
kernel/latency.c: In function `__start_critical_timing':
kernel/latency.c:810: incompatible type for argument 1 of `____trace'
kernel/latency.c:810: warning: passing arg 2 of `____trace' makes
pointer from integer without a cast
kernel/latency.c:810: too few arguments to function `____trace'
kernel/latency.c: In function `trace_irqs_off':
kernel/latency.c:810: warning: too few arguments passed to inline
function, suppressing inlining
kernel/latency.c: In function `add_preempt_count':
kernel/latency.c:810: warning: too few arguments passed to inline
function, suppressing inlining
kernel/latency.c: At top level:
kernel/latency.c:810: warning: too few arguments passed to inline
function, suppressing inlining
make[1]: *** [kernel/latency.o] Error 1
make: *** [kernel] Error 2
[***@tigre01 im]#

Regards
Remi

Ingo Molnar

2004-11-30 08:15:48 UTC

Post by Remi Colinet
Hi Ingo,
I'm getting this error with V0.7.31-13
CC kernel/latency.o
kernel/latency.c:730: too few arguments to function `___trace'
kernel/latency.c:730: warning: too few arguments passed to inline

fixed this in -V0.7.31-14.

Ingo

Eran Mann

2004-12-01 08:30:26 UTC

Post by Ingo Molnar
fixed this in -V0.7.31-14.
Ingo
This mail arrived via mail.mrv.com

Compiling -V0.7.31-15 without CONFIG_LATENCY_TRACE results in:
...
CC kernel/latency.o
kernel/latency.c: In function `check_critical_timing':
kernel/latency.c:765: warning: implicit declaration of function `__trace'
...
LD .tmp_vmlinux1
kernel/built-in.o(.text+0x1e407): In function `check_critical_timing':
: undefined reference to `__trace'
make: *** [.tmp_vmlinux1] Error 1

Seems to be fixed by the patch below:

--- kernel/latency.c.orig 2004-12-01 10:21:45.000000000 +0200
+++ kernel/latency.c 2004-12-01 10:11:37.000000000 +0200
@@ -762,7 +762,9 @@
tr->critical_sequence = max_sequence;
tr->preempt_timestamp = cycles();
tr->early_warning = 0;
+#ifdef CONFIG_LATENCY_TRACE
__trace(CALLER_ADDR0, parent_eip);
+#endif
}

void notrace touch_critical_timing(void)

--
Eran Mann

Ingo Molnar

2004-12-01 08:53:37 UTC

Post by Eran Mann
--- kernel/latency.c.orig 2004-12-01 10:21:45.000000000 +0200
+++ kernel/latency.c 2004-12-01 10:11:37.000000000 +0200
@@ -762,7 +762,9 @@
tr->critical_sequence = max_sequence;
tr->preempt_timestamp = cycles();
tr->early_warning = 0;
+#ifdef CONFIG_LATENCY_TRACE
__trace(CALLER_ADDR0, parent_eip);
+#endif

thanks, applied it and uploaded -V0.7.31-16.

Ingo

Adam Heath

2004-12-01 18:19:35 UTC

thanks, applied it and uploaded -V0.7.31-16.

Wouldn't it be better to change the definition of __trace()? Always have it
defined, but make it empty if CONFIG_LATENCY_TRACE isn't set?

It'd keep the code from being littered with ifdef.

Ingo Molnar

2004-11-29 11:05:07 UTC

Post by Rui Nuno Capela
config-2.6.10-rc2-mm3-RT-V0.7.31-7.gz
- kernel configuration as of my laptop, taken from /proc/config.gz

Here are some .config suggestion to reduce tracing costs. Here's the
relevant portion of your .config:

CONFIG_DEBUG_KERNEL=y
CONFIG_MAGIC_SYSRQ=y
# CONFIG_SCHEDSTATS is not set
# CONFIG_DEBUG_SLAB is not set
CONFIG_DEBUG_PREEMPT=y
CONFIG_WAKEUP_TIMING=y
CONFIG_PREEMPT_TRACE=y
CONFIG_CRITICAL_PREEMPT_TIMING=y
CONFIG_CRITICAL_IRQSOFF_TIMING=y
CONFIG_CRITICAL_TIMING=y
CONFIG_LATENCY_TIMING=y
CONFIG_LATENCY_TRACE=y
CONFIG_MCOUNT=y
CONFIG_RT_DEADLOCK_DETECT=y
# CONFIG_DEBUG_KOBJECT is not set
CONFIG_DEBUG_BUGVERBOSE=y
# CONFIG_DEBUG_INFO is not set
CONFIG_FRAME_POINTER=y
CONFIG_EARLY_PRINTK=y
CONFIG_DEBUG_STACKOVERFLOW=y

i'd suggest to use the following one:

CONFIG_DEBUG_KERNEL=y
CONFIG_MAGIC_SYSRQ=y
# CONFIG_SCHEDSTATS is not set
# CONFIG_DEBUG_SLAB is not set
# CONFIG_DEBUG_PREEMPT is not set
# CONFIG_DEBUG_SPINLOCK_SLEEP is not set
CONFIG_WAKEUP_TIMING=y
# CONFIG_CRITICAL_PREEMPT_TIMING is not set
# CONFIG_CRITICAL_IRQSOFF_TIMING is not set
CONFIG_LATENCY_TIMING=y
CONFIG_LATENCY_TRACE=y
CONFIG_MCOUNT=y
# CONFIG_RT_DEADLOCK_DETECT is not set
# CONFIG_DEBUG_KOBJECT is not set
CONFIG_DEBUG_BUGVERBOSE=y
# CONFIG_DEBUG_INFO is not set
CONFIG_FRAME_POINTER=y
CONFIG_EARLY_PRINTK=y
# CONFIG_DEBUG_STACKOVERFLOW is not set

especially PREEMPT_TIMING, IRQSOFF_TIMING and RT_DEADLOCK_DETECT are
quite expensive options.

Ingo

Eran Mann

2004-11-28 19:42:59 UTC

Post by Remi Colinet
Hi Ingo,
I'm getting this error with V0.7.31-13

...

Post by Remi Colinet
CC kernel/latency.o
kernel/latency.c:730: too few arguments to function `___trace'
kernel/latency.c:730: warning: too few arguments passed to inline function, suppressing inlining
kernel/latency.c:810: incompatible type for argument 1 of `____trace'
kernel/latency.c:810: warning: passing arg 2 of `____trace' makes pointer from integer without a cast

...

Post by Remi Colinet
kernel/latency.c:810: warning: too few arguments passed to inline function, suppressing inlining
make[1]: *** [kernel/latency.o] Error 1
make: *** [kernel] Error 2
Regards
Remi

I'm guessing here, but with the following patch it at least compiles:

--- kernel/latency.c.orig 2004-11-28 21:32:04.757015856 +0200
+++ kernel/latency.c 2004-11-28 21:28:28.000000000 +0200
@@ -727,7 +727,7 @@
tr->critical_end = parent_eip;

#ifdef CONFIG_LATENCY_TRACE
- ___trace(CALLER_ADDR0, parent_eip);
+ ___trace(TRACE_FN, CALLER_ADDR0, parent_eip, 0, 0, 0);
update_max_trace(tr);
#endif

@@ -807,7 +807,7 @@
tr->critical_start = eip;
#ifdef CONFIG_LATENCY_TRACE
tr->trace_idx = 0;
- ____trace(tr, eip, parent_eip, 0, 0, 0);
+ ____trace(TRACE_FN, tr, eip, parent_eip, 0, 0, 0);
#endif

atomic_dec(&tr->disabled);

--
Eran Mann
Senior Software Engineer
MRV International
Tel: 972-4-9936297
Fax: 972-4-9890430
www.mrv.com

Ingo Molnar

2004-11-30 08:58:27 UTC

yeah, this is the correct patch, included in the -31-14 and later
patches.

Ingo

Gene Heskett

2004-11-30 16:20:08 UTC

Post by Eran Mann
I'm guessing here, but with the following patch it at least

yeah, this is the correct patch, included in the -31-14 and later
patches.
Ingo

I hadn't hit this one yet, up since about midnight with -13, and just
now found -15 so I'll build it.

M***@raytheon.com

2004-11-30 14:33:05 UTC

I have results from two builds of -V0.7.31-9. The first
build is CONFIG_RT (RT) and the second is CONFIG_DESKTOP (PK or
as described in .config help - Preempt Kernel).

Both booted OK - so the SMP lockup on _DESKTOP appears to
be fixed. Both ran my test series faster than previous 2.6
kernels. I was seeing run times over 30 minutes before
(mainly due to starvation of non-RT tasks) but both completed
in about 20 minutes.

General notes:

[1] Audio duration was much more variable on the PK than the
RT kernel. The variability on PK audio duration is about 30%
of nominal time. The RT kernel was less variable but tended to
finish well before the nominal time (and you could hear the
difference). The PK results look like buffering inside ALSA
that was not present in OSS?. Not sure why the RT kernel
is almost always finishing too soon.

[2] The PK kernel has much longer latencies as measured by
the in kernel tracing code. The RT kernel basically had NO
latencies > 50 usec where the PK kernel had several over a
millisecond (usually related to disk writes). By this measure
the RT kernel is clearly better.

[3] The overhead of RT (as measured by the CPU loop timing and
by pings from remote systems) is more than the PK kernel. I
believe this is due to the IRQ threading overhead. By these
measures, the PK kernel is better. By elapsed time, the 2.4
kernel is far superior.

More specifics:
The 2.4 numbers are from 2.4.24 w/ low latency / preempt patches.

within 100 usec
CPU loop (%) Elapsed Time (sec) 2.4
Test PK RT PK RT | CPU Elapsed
X 99.69 99.34 90 70 | 97.20 70
top 99.31 98.33 30 31 | 97.48 29
neto 97.28 97.69 205 315 | 96.23 36
neti 97.76 98.11 198 325 | 95.86 41
diskw 69.16* 94.98 51 115 | 77.64 29
diskc 96.44 98.39 230 250 | 84.12 77
diskr 99.60 98.77 240 180 | 90.66 86
total 1044 1286 | 368

* several multiple millisecond latencies measured by the
tracing code. Will send traces separately.

min ave max mdev
PK ping - 0.100 / 0.176 / 1.009 / 0.053
RT ping - 0.194 / 0.322 / 527.635 / 2.263
[not sure why the high max on RT, but I did see a few
1 msec or longer ping responses and many over 400 usec]

--Mark H Johnson
<mailto:***@raytheon.com>

Ingo Molnar

2004-12-01 09:02:20 UTC

I have results from two builds of -V0.7.31-9. The first build is
CONFIG_RT (RT) and the second is CONFIG_DESKTOP (PK or as described in
.config help - Preempt Kernel).

interesting numbers. The slowdown in networking could easily be due to
IRQ and softirq threading, so it would make sense to also add a "PNT"
test (preempt, non-threaded), just to have something functionally
comparable to 2.4 lowlat+preempt.

Ingo

M***@Raytheon.com

2004-12-01 14:57:02 UTC

Post by Ingo Molnar
interesting numbers. The slowdown in networking could easily be due to
IRQ and softirq threading, so it would make sense to also add a "PNT"
test (preempt, non-threaded), just to have something functionally
comparable to 2.4 lowlat+preempt.

Unless I am mistaken, my "PK" config is the closest to 2.4 lowlat+preempt.
For the relevant differences in .config:
PK RT
CONFIG_PREEMPT_DESKTOP=y CONFIG_PREEMPT_DESKTOP is not set
CONFIG_PREEMPT_RT is not set CONFIG_PREEMPT_RT=y
CONFIG_PREEMPT=y CONFIG_PREEMPT=y
CONFIG_PREEMPT_SOFTIRQS is not set CONFIG_PREEMPT_SOFTIRQS=y
CONFIG_PREEMPT_HARDIRQS is not set CONFIG_PREEMPT_HARDIRQS=y
(though the system still creates ksoftirqd/0 and /1 on both...)
CONFIG_SPINLOCK_BKL is not set [not present]
CONFIG_PREEMPT_BKL=y CONFIG_PREEMPT_BKL=y
CONFIG_ASM_SEMAPHORES=y [not present]
CONFIG_RWSEM_XCHGADD_ALGORITHM=y [not present]
...
[not present] CONFIG_RT_DEADLOCK_DETECT=y
...

Unless you are saying that I should back off to one of the other
preempt settings (to replicate the 2.4 config on 2.6).

Also, I ran another set of tests with -15 late yesterday. The
results were not necessarily consistent with -9. I had to run
the -PK tests twice due to an error that
/dev/dsp was busy
which aborted one of the latencytest runs (not the first one...).
The line below with "dnr" is from that run.

[the -15 data follows in the same format as the -9 data yesterday]
The 2.4 numbers are from 2.4.24 w/ low latency / preempt patches.

within 100 usec
CPU loop (%) Elapsed Time (sec) 2.4
Test PK RT PK RT | CPU Elapsed
X 99.66 99.22 53 * 68 | 97.20 70
top 99.92 97.96 240+ 34 | 97.48 29
neto dnr 99.98 dnr 360 | 96.23 36
neti 98.22 98.31 270 * 350 | 95.86 41
diskw 94.11 99.57 295+* 360 * | 77.64 29
diskc 99.67 97.49 310+* 360 | 84.12 77
diskr 99.52 98.35 310+* 180 | 90.66 86
total 1478 1712 | 368
did not count neto in this total^^
* wide variation in audio duration
+ had long stretches of audio duration well below the nominal
duration

Both the -PK and -RT tests had long pauses where the sampling
scripts I use did not capture any data. The profile script had
elapsed times of 77, 192, 52, 106, 305, and 82 seconds [for a
5 second sleep]. The load average after each of these pauses
ranged from 8.8 to 16.1. It still looks like the cpu_burn program
(non RT, nice) is starving the other normal non RT tasks
(should not happen...).

I am still getting lots of truncated latency_trace results.

The latency_trace output doesn't seem to get the process name
right. Most traces look something like this [truncated example]

preemption latency trace v1.1.1 on 2.6.10-rc2-mm3-V0.7.31-15RT
--------------------------------------------------------
latency: 102 us, entries: 1(1) | [VP:0 KP:1 SP:1 HP:1 #CPUS:2]
-----------------
| task: cat/9789, uid:0 nice:0 policy:0 rt_prio:0
-----------------
=> started at: buffered_rmqueue+0x69/0x180 <c014f859>
=> ended at: buffered_rmqueue+0xa3/0x180 <c014f89e>
========>
<unknown-9789 80000000 0.000ms (+0.000ms): buffered_rmqueue (__alloc_pages)

So the name is correct in the header but not in the trace lines in
most (but not ALL) cases. I have another example where IRQ 0/2 was
traced and its lines were OK in the trace, but I had references to
unknown-2036, unknown-2044, and unknown-3802 in the trace. This
one example appeared to be a task switch from unknown-2036 to
IRQ 0-2 and the traces are OK after the line with __switch_to in it.
Let me know if you need examples (I have plenty to choose from).

--Mark H Johnson
<mailto:***@raytheon.com>

Ingo Molnar

2004-12-01 21:20:51 UTC

Post by M***@Raytheon.com
Unless I am mistaken, my "PK" config is the closest to 2.4 lowlat+preempt.

indeed, you are right.

Post by M***@Raytheon.com
PK RT
CONFIG_PREEMPT_DESKTOP=y CONFIG_PREEMPT_DESKTOP is not set
CONFIG_PREEMPT_RT is not set CONFIG_PREEMPT_RT=y
CONFIG_PREEMPT=y CONFIG_PREEMPT=y
CONFIG_PREEMPT_SOFTIRQS is not set CONFIG_PREEMPT_SOFTIRQS=y
CONFIG_PREEMPT_HARDIRQS is not set CONFIG_PREEMPT_HARDIRQS=y
(though the system still creates ksoftirqd/0 and /1 on both...)
CONFIG_SPINLOCK_BKL is not set [not present]
CONFIG_PREEMPT_BKL=y CONFIG_PREEMPT_BKL=y
CONFIG_ASM_SEMAPHORES=y [not present]
CONFIG_RWSEM_XCHGADD_ALGORITHM=y [not present]
...
[not present] CONFIG_RT_DEADLOCK_DETECT=y
...

the only thing i'd suggest to change is to also generate an RT (and
perhaps PK) result with all debugging options disabled - i.e. both
RT_DEADLOCK_DETECT and all LATENCY_TRACING/timing related options
disabled.

but your tests did trigger asserts not so long ago so it might not be
wise to run without debugging. But it's definitely a thing to try in the
future.

Post by M***@Raytheon.com
Unless you are saying that I should back off to one of the other
preempt settings (to replicate the 2.4 config on 2.6).

no, i think the PK kernel is supposed to be quite close to what
2.4+lowlat offers.

Ingo

Andrew Burgess

2004-12-02 15:46:20 UTC

Might be handy to have the option to send a SIGABRT, then you don't need
to guess which app to run under gdb and the offending code is there in
the core file.

Also, I'm cc-ing jack-devel. This could fit into libjack so no client
mods would be needed I think. After the thread_init_callback is run libjack
could run 'gettimeofday(1,1);' for each client thread. Then if any client
breaks the rules you get a core showing where.

On further thought, I suppose libjack could install a SIGUSR2 handler and
have that call abort for all the rt client threads. Still no client mods
needed, only an RT-aware libjack.

A big thank you to Ingo and everyone else involved on behalf of all the
linux audio users!

Florian Schmidt

2004-12-02 16:03:15 UTC

On Thu, 2 Dec 2004 07:46:20 -0800

Post by Andrew Burgess

Might be handy to have the option to send a SIGABRT, then you don't need
to guess which app to run under gdb and the offending code is there in
the core file.
Also, I'm cc-ing jack-devel. This could fit into libjack so no client
mods would be needed I think. After the thread_init_callback is run libjack
could run 'gettimeofday(1,1);' for each client thread. Then if any client
breaks the rules you get a core showing where.
On further thought, I suppose libjack could install a SIGUSR2 handler and
have that call abort for all the rt client threads. Still no client mods
needed, only an RT-aware libjack.

right. Or instead of aborting jackd might print a debug output (like
"client foo violated RT constraints").

But the calls to gettimeofday would need to be done right before and
after every process callback as each client's RT thread does wait on the
FIFO to get woken by jackd. This waiting would appear as RT constraints
violation if the gettimeofday would be done only once per thread
lifetime at thread startup.

Post by Andrew Burgess
A big thank you to Ingo and everyone else involved on behalf of all the
linux audio users!

BTW: i suppose pretty much every jack client except for very simple ones
do break the RT constraints.

Flo

-------------------------------------------------------
SF email is sponsored by - The IT Product Guide
Read honest & candid reviews on hundreds of IT Products from real users.
Discover which products truly live up to the hype. Start reading now.
http://productguide.itmanagersjournal.com/

Jack O'Quin

2004-12-02 16:26:55 UTC

Post by Florian Schmidt
On Thu, 2 Dec 2004 07:46:20 -0800

Post by Andrew Burgess
On further thought, I suppose libjack could install a SIGUSR2 handler and
have that call abort for all the rt client threads. Still no client mods
needed, only an RT-aware libjack.

right. Or instead of aborting jackd might print a debug output (like
"client foo violated RT constraints").

Libjack cannot assume the client has no SIGUSR2 handler of its own.

Post by Florian Schmidt

Post by Andrew Burgess
A big thank you to Ingo and everyone else involved on behalf of all the
linux audio users!

thanks++ :-)

Post by Florian Schmidt
BTW: i suppose pretty much every jack client except for very simple ones
do break the RT constraints.

True.

It would be wonderful to have a reliable mechanism for debugging them.

--
joq

Florian Schmidt

2004-12-02 16:57:56 UTC

On 02 Dec 2004 10:26:55 -0600

Post by Florian Schmidt
On Thu, 2 Dec 2004 07:46:20 -0800

right. Or instead of aborting jackd might print a debug output (like
"client foo violated RT constraints").

Libjack cannot assume the client has no SIGUSR2 handler of its own.

i see..

Post by Jack O'Quin
It would be wonderful to have a reliable mechanism for debugging them.

I suppose instead of catching the signal the user might just monitor the
syslog. I'm not sure there's printk's triggered by thisalready , but i'm
sure if not, ingo might add them. So a trivial patch for jackd would
probably look like this:

--- libjack/client.c.orig 2004-12-02 17:55:04.000000000 +0100
+++ libjack/client.c 2004-12-02 17:56:23.000000000 +0100
@@ -1238,6 +1238,9 @@
if (control->sync_cb)
jack_call_sync_client (client);

+ // enable atomicity check for RP kernels
+ gettimeofday(1,1);
+
if (control->process) {
if (control->process (control->nframes,
control->process_arg)
@@ -1247,7 +1250,10 @@
} else {
control->state = Finished;
}
-
+
+ // disable atomicity check
+ gettimeofday(0,1);
+
if (control->timebase_cb)
jack_call_timebase_master (client);

flo

-------------------------------------------------------
SF email is sponsored by - The IT Product Guide
Read honest & candid reviews on hundreds of IT Products from real users.
Discover which products truly live up to the hype. Start reading now.
http://productguide.itmanagersjournal.com/

Jack O'Quin

2004-12-02 17:32:49 UTC

Hmm, i must have missed something in jackd's source. i thought
control->process() directly calls the clients process callback..

It does.

--
joq

-------------------------------------------------------
SF email is sponsored by - The IT Product Guide
Read honest & candid reviews on hundreds of IT Products from real users.
Discover which products truly live up to the hype. Start reading now.
http://productguide.itmanagersjournal.com/

Florian Schmidt

2004-12-02 20:03:18 UTC

On 02 Dec 2004 11:32:49 -0600

Hmm, i must have missed something in jackd's source. i thought
control->process() directly calls the clients process callback..

It does.

Then either i have misunderstood how it works, or the mechanism is still
buggy.. Wrote mail to Ingo. Will report to jackit-devel when i get an
answer.

Flo

-------------------------------------------------------
SF email is sponsored by - The IT Product Guide
Read honest & candid reviews on hundreds of IT Products from real users.
Discover which products truly live up to the hype. Start reading now.
http://productguide.itmanagersjournal.com/

Florian Schmidt

2004-12-02 21:10:26 UTC

On Thu, 2 Dec 2004 21:03:18 +0100

Post by Florian Schmidt
On 02 Dec 2004 11:32:49 -0600

Hmm, i must have missed something in jackd's source. i thought
control->process() directly calls the clients process callback..

It does.

Then either i have misunderstood how it works, or the mechanism is still
buggy.. Wrote mail to Ingo. Will report to jackit-devel when i get an
answer.

PEBCAK :) Here's the corrected patch:

--- libjack/client.c.orig 2004-12-02 17:55:04.000000000 +0100
+++ libjack/client.c 2004-12-02 22:04:06.000000000 +0100
@@ -1238,6 +1238,9 @@
if (control->sync_cb)
jack_call_sync_client (client);

+ // enable atomicity check for RP kernels
+ gettimeofday(1,1);
+
if (control->process) {
if (control->process (control->nframes,
control->process_arg)
@@ -1247,7 +1250,10 @@
} else {
control->state = Finished;
}
-
+
+ // disable atomicity check
+ gettimeofday(1,0);
+
if (control->timebase_cb)
jack_call_timebase_master (client);

seems to work well with my changed jack_test client (this one sleeps in
the 1000th call of the process callback and thuis triggers this trace
and aborts as it doesn't handle SIGUSR2). test client attached.

Dec 2 22:05:10 mango kernel: jack_test:3043 userspace BUG: scheduling in user-atomic context!
Dec 2 22:05:10 mango kernel: [<c02a38b6>] schedule+0x76/0x130 (8)
Dec 2 22:05:10 mango kernel: [<c02a44c5>] schedule_timeout+0x85/0xe0 (36)
Dec 2 22:05:10 mango kernel: [<c01e2f02>] copy_from_user+0x42/0x80 (48)
Dec 2 22:05:10 mango kernel: [<c011ceb0>] process_timeout+0x0/0x10 (8)
Dec 2 22:05:10 mango kernel: [<c011cfae>] sys_nanosleep+0xde/0x170 (20)
Dec 2 22:05:10 mango kernel: [<c01025cb>] syscall_call+0x7/0xb (52)

flo

Florian Schmidt

2004-12-02 17:09:10 UTC

On Thu, 2 Dec 2004 17:57:56 +0100

Post by Florian Schmidt
On 02 Dec 2004 10:26:55 -0600

Post by Florian Schmidt
On Thu, 2 Dec 2004 07:46:20 -0800

right. Or instead of aborting jackd might print a debug output (like
"client foo violated RT constraints").

Libjack cannot assume the client has no SIGUSR2 handler of its own.

i see..

Post by Jack O'Quin
It would be wonderful to have a reliable mechanism for debugging them.

Well, i do get syslog output with this patch like this:

Dec 2 18:01:11 mango kernel: jack_test:22645 userspace BUG: scheduling in user-atomic context!
Dec 2 18:01:11 mango kernel: [<c02a38b6>] schedule+0x76/0x130 (8)
Dec 2 18:01:11 mango kernel: [<c02a44c5>] schedule_timeout+0x85/0xe0 (36)
Dec 2 18:01:11 mango kernel: [<c016677f>] do_pollfd+0x4f/0x90 (48)
Dec 2 18:01:11 mango kernel: [<c011ceb0>] process_timeout+0x0/0x10 (8)
Dec 2 18:01:11 mango kernel: [<c016686a>] do_poll+0xaa/0xd0 (20)
Dec 2 18:01:11 mango kernel: [<c01669e2>] sys_poll+0x152/0x230 (48)
Dec 2 18:01:11 mango kernel: [<c0165db0>] __pollwait+0x0/0xd0 (36)
Dec 2 18:01:11 mango kernel: [<c01025cb>] syscall_call+0x7/0xb (32)

even if the client's process callback is a noop (except for returning
0). Hmm, i must have missed something in jackd's source. i thought
control->process() directly calls the clients process callback..

hmm..

-------------------------------------------------------
SF email is sponsored by - The IT Product Guide
Read honest & candid reviews on hundreds of IT Products from real users.
Discover which products truly live up to the hype. Start reading now.
http://productguide.itmanagersjournal.com/

Jack O'Quin

2004-12-02 17:07:39 UTC

Post by Florian Schmidt
I suppose instead of catching the signal the user might just monitor the
syslog. I'm not sure there's printk's triggered by thisalready , but i'm
sure if not, ingo might add them. So a trivial patch for jackd would
--- libjack/client.c.orig 2004-12-02 17:55:04.000000000 +0100
+++ libjack/client.c 2004-12-02 17:56:23.000000000 +0100
@@ -1238,6 +1238,9 @@
if (control->sync_cb)
jack_call_sync_client (client);
+ // enable atomicity check for RP kernels
+ gettimeofday(1,1);
+
if (control->process) {
if (control->process (control->nframes,
control->process_arg)
@@ -1247,7 +1250,10 @@
} else {
control->state = Finished;
}
-
+
+ // disable atomicity check
+ gettimeofday(0,1);
+
if (control->timebase_cb)
jack_call_timebase_master (client);

The sync_cb and timebase_cb callbacks actually need to be RT-safe,
too. ;-)

Is printk() guaranteed not to wait inside the kernel? I am not
familiar with its internal implementation.

--
joq

Lee Revell

2004-12-02 20:07:15 UTC

Post by Jack O'Quin
Is printk() guaranteed not to wait inside the kernel? I am not
familiar with its internal implementation.

Yes. It just writes to a ring buffer and klogd dumps this to syslog.
So if you really start to spew printk's they don't all make it to the
log but you never get blocked.

The implementation probably looks a lot like a correct solution to fix
the printf-from-RT-context issue in JACK would.

Lee

-------------------------------------------------------
SF email is sponsored by - The IT Product Guide
Read honest & candid reviews on hundreds of IT Products from real users.
Discover which products truly live up to the hype. Start reading now.
http://productguide.itmanagersjournal.com/

Jack O'Quin

2004-12-02 20:48:23 UTC

Post by Lee Revell

Post by Jack O'Quin
Is printk() guaranteed not to wait inside the kernel? I am not
familiar with its internal implementation.

Right. That's exactly what I have in mind, whenever I find time to
work on it. :-)

There's some similar code I wrote for JAMin, which we could adapt.

Bill Huey (hui)

2004-12-02 22:34:49 UTC

Post by Andrew Burgess
A big thank you to Ingo and everyone else involved on behalf of all the
linux audio users!

This is going to have a lot of ripple effect throughout the Linux
community in that game developers will quite possibly have the ability
to do low latency OpenGL, frame accurate video and other things along
those lines. The previous Linux kernels without these mods couldn't
allow this level of temporal precious for application developers. It's
going to push applications like jackd and others in ways that will
flush out bugs and general techniques that aren't typically apart of
a proper RT applications.

We have the possibility of providing a first class gaming platform and
other things, that Longhorn can't beging to touch if the middleware
falls into place. :)

bill

-------------------------------------------------------
SF email is sponsored by - The IT Product Guide
Read honest & candid reviews on hundreds of IT Products from real users.
Discover which products truly live up to the hype. Start reading now.
http://productguide.itmanagersjournal.com/

M***@raytheon.com

2004-12-02 21:01:27 UTC