remembering rt-sched attributes - was: [LAD] rtirq script is broken with 2.6.31

remembering rt-sched attributes - was: [LAD] rtirq script is broken with 2.6.31

[Robin Gareus]

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

Hello rt-users and -devs,

I have a question about per-device-IRQ-threads in 2.6.31-rc5-rt1.1:

After a suspend/resume cycle some IRQ-threads come up with a new PID.
The scheduling policy of those threads is reset to default values
instead of retaining the previously set values.

Is this an issue that is being worked on? Or must userspace from now on
re-init rtprio settings after each suspend/resume cycle?

As you can see from the information below (from linux-audio-dev @
linuxaudio.org) not all IRQ-threads are re-started, but for example the
HDA-Intel is. It may just as well be a specific issue with snd_hda_intel
(and sdhci, e1000e, i810/intelfb,..).

Robin Gareus wrote:
> Rui Nuno Capela wrote:
>>>  [..]
>>>
>>> Yes, I'm also baffled at the high PIDs for IRQs. I hazard a guess that
>>> those are a result of a suspend/resume cycle; and I'll check later if
>>> the chrt settings do persist after a suspend/resume.
> 
> It looks like the guess was correct. The PIDs change after
> suspend/resume and the chrt settings are retained here.
Sorry I was too quick, there. After some more suspend/resumes cycles and
a closer look: the previous statement is not correct.

RTPRIO is reset when the IRQ handler process restarts under new PID.
However not all IRQ threads are re-launched:

[from /proc/interrupts]
 17:   78393056     873204   IO-APIC-fasteoi   uhci_hcd:usb3, HDA Intel,
ohci1394
 18:     451927    2099364   IO-APIC-fasteoi   uhci_hcd:usb4, mmc0

before suspend:
  PID CLS RTPRIO  NI PRI %CPU STAT COMMAND
 1418 FF      90   - 130  0.0 S<   irq/8-rtc0	
 1450 FF      89   - 129  0.0 S<   irq/18-uhci_hcd	
32506 FF      88   - 128  0.1 S<   irq/17-HDA Inte	
32507 FF      87   - 127  0.0 S<   irq/17-ohci1394	
 1447 FF      50   -  90  0.2 S<   irq/17-uhci_hcd	
32508 FF      50   -  90  0.0 S<   irq/18-mmc0	
...

after resume:

  PID CLS RTPRIO  NI PRI %CPU STAT COMMAND
 1418 FF      90   - 130  0.0 S<   irq/8-rtc0	
 1450 FF      89   - 129  0.0 S<   irq/18-uhci_hcd	
  388 FF      50   -  90  0.0 S<   irq/17-HDA Inte	
  389 FF      50   -  90  0.0 S<   irq/17-ohci1394	
  390 FF      50   -  90  0.0 S<   irq/18-mmc0	
 1447 FF      50   -  90  0.2 S<   irq/17-uhci_hcd	
...

As you can see all IRQ17 threads get reset completely; on IRQ18 only
mmc0 gets a new PID but the usb retains it's PID and rtprio.


My first assumption - that it could correlate with the device being in
use while suspending - did also proove wrong: I tried suspending with
JACKd using an USB audio device on IRQ18 (instead of HDA-Intel on IRQ17)
and after resume IRQ18 has still high rtprio, while IRQ17 has been reset
whether in use or not.

However, after unloading the HDA-Intel module, the other IRQ-threads on
IRQ17 (ohci1394 and uhci_hcd:usb3) retain their PID and rtprio. So it
seems there's something with the HDA driver and/or hardware that causes
the kernel to re-initialize IRQ process.

<OT>
I usually unload the sd-card mmc0 driver and connect a USB-UA25 on
uhci_hcd:usb4; It then is the sole device on IRQ18 and works without
x-runs even at 64 spp * 3p / 48000sps = 4ms audio latency with JACKd.

With 2.6.29.6-rt23 I get x-runs at these low latencies when not
unloading the mmc0 driver. With 2.6.31-rc5-rt1.1 I don't. So the
per-driver IRQ priority does work.. YAY.
</OT>

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

iEYEARECAAYFAkp+4pEACgkQeVUk8U+VK0KZXACgiJCtcEfH6syPYsyoEVp19VS1
mE8AnRTyEwKD+kMI39+HSfLl4Ms8Zd58
=SEQi
-----END PGP SIGNATURE-----

Re: remembering rt-sched attributes - was: [LAD] rtirq script is broken with 2.6.31

[Rui Nuno Capela]

Robin Gareus wrote:
> Hello rt-users and -devs,
> 
> I have a question about per-device-IRQ-threads in 2.6.31-rc5-rt1.1:
> 
> After a suspend/resume cycle some IRQ-threads come up with a new PID.
> The scheduling policy of those threads is reset to default values
> instead of retaining the previously set values.
> 
> Is this an issue that is being worked on? Or must userspace from now on
> re-init rtprio settings after each suspend/resume cycle?
> 
> As you can see from the information below (from linux-audio-dev @
> linuxaudio.org) not all IRQ-threads are re-started, but for example the
> HDA-Intel is. It may just as well be a specific issue with snd_hda_intel
> (and sdhci, e1000e, i810/intelfb,..).
> 
> Robin Gareus wrote:
>> Rui Nuno Capela wrote:
>>>>  [..]
>>>>
>>>> Yes, I'm also baffled at the high PIDs for IRQs. I hazard a guess that
>>>> those are a result of a suspend/resume cycle; and I'll check later if
>>>> the chrt settings do persist after a suspend/resume.
>> It looks like the guess was correct. The PIDs change after
>> suspend/resume and the chrt settings are retained here.
> Sorry I was too quick, there. After some more suspend/resumes cycles and
> a closer look: the previous statement is not correct.
> 
> RTPRIO is reset when the IRQ handler process restarts under new PID.
> However not all IRQ threads are re-launched:
> 
> [from /proc/interrupts]
>  17:   78393056     873204   IO-APIC-fasteoi   uhci_hcd:usb3, HDA Intel,
> ohci1394
>  18:     451927    2099364   IO-APIC-fasteoi   uhci_hcd:usb4, mmc0
> 
> before suspend:
>   PID CLS RTPRIO  NI PRI %CPU STAT COMMAND
>  1418 FF      90   - 130  0.0 S<   irq/8-rtc0	
>  1450 FF      89   - 129  0.0 S<   irq/18-uhci_hcd	
> 32506 FF      88   - 128  0.1 S<   irq/17-HDA Inte	
> 32507 FF      87   - 127  0.0 S<   irq/17-ohci1394	
>  1447 FF      50   -  90  0.2 S<   irq/17-uhci_hcd	
> 32508 FF      50   -  90  0.0 S<   irq/18-mmc0	
> ...
> 
> after resume:
> 
>   PID CLS RTPRIO  NI PRI %CPU STAT COMMAND
>  1418 FF      90   - 130  0.0 S<   irq/8-rtc0	
>  1450 FF      89   - 129  0.0 S<   irq/18-uhci_hcd	
>   388 FF      50   -  90  0.0 S<   irq/17-HDA Inte	
>   389 FF      50   -  90  0.0 S<   irq/17-ohci1394	
>   390 FF      50   -  90  0.0 S<   irq/18-mmc0	
>  1447 FF      50   -  90  0.2 S<   irq/17-uhci_hcd	
> ...
> 
> As you can see all IRQ17 threads get reset completely; on IRQ18 only
> mmc0 gets a new PID but the usb retains it's PID and rtprio.
> 
> 
> My first assumption - that it could correlate with the device being in
> use while suspending - did also proove wrong: I tried suspending with
> JACKd using an USB audio device on IRQ18 (instead of HDA-Intel on IRQ17)
> and after resume IRQ18 has still high rtprio, while IRQ17 has been reset
> whether in use or not.
> 
> However, after unloading the HDA-Intel module, the other IRQ-threads on
> IRQ17 (ohci1394 and uhci_hcd:usb3) retain their PID and rtprio. So it
> seems there's something with the HDA driver and/or hardware that causes
> the kernel to re-initialize IRQ process.
> 
> <OT>
> I usually unload the sd-card mmc0 driver and connect a USB-UA25 on
> uhci_hcd:usb4; It then is the sole device on IRQ18 and works without
> x-runs even at 64 spp * 3p / 48000sps = 4ms audio latency with JACKd.
> 
> With 2.6.29.6-rt23 I get x-runs at these low latencies when not
> unloading the mmc0 driver. With 2.6.31-rc5-rt1.1 I don't. So the
> per-driver IRQ priority does work.. YAY.
> </OT>
> 

picking on old subject, why not doing a plain `/etc/init.d/rtirq
restart` on resume?

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

Re: remembering rt-sched attributes - was: [LAD] rtirq script is broken with 2.6.31

[Robin Gareus]

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

Rui Nuno Capela wrote:
> Robin Gareus wrote:
>> Hello rt-users and -devs,
>>
>> I have a question about per-device-IRQ-threads in 2.6.31-rc5-rt1.1:
>>
>> After a suspend/resume cycle some IRQ-threads come up with a new PID.
>> The scheduling policy of those threads is reset to default values
>> instead of retaining the previously set values.
>>
>> Is this an issue that is being worked on? Or must userspace from now on
>> re-init rtprio settings after each suspend/resume cycle?
>>
>> As you can see from the information below (from linux-audio-dev @
>> linuxaudio.org) not all IRQ-threads are re-started, but for example the
>> HDA-Intel is. It may just as well be a specific issue with snd_hda_intel
>> (and sdhci, e1000e, i810/intelfb,..).
>>
>> [...]
> 
> picking on old subject, why not doing a plain `/etc/init.d/rtirq
> restart` on resume?
> 

Sure, adding it to
/usr/lib/hal/scripts/linux/hal-system-power-suspend-linux
or a script in /etc/acpi/resume.d/ (depending on the distribution)
does the trick.

However it would be necessary to be included in quite a few
distributions: Ubuntustudio, 64studio, AVlinux, JACKlab, etc. It would
be better if the problem could be solved at the source.

Anyways, can anyone explain why only a particular set of IRQ-threads are
affected by this issue?

robin
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.9 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org

iEYEARECAAYFAkqCikEACgkQeVUk8U+VK0J0eQCeNPiTVO6wLBkuNc2CaQLpPuK4
h/sAnRGExpwt4zKQ1PH7LSo2NlCF6Y55
=oGVR
-----END PGP SIGNATURE-----