Skip to contentSkip to conversation

idle Athlon with IOAPIC is 10C warmer since 2.6.3-bk1

DISCUSSION Ross Dickson 7 participants 15 messages Mar 18, 2004 – Mar 30, 2004

lore
Share & tools
lore
#0 lore raw
Quoted 15 lines
Quoted 6 lines · Maciej W. Rozycki
Quoted 4 lines · Thomas Schlichter

a few days ago I noticed that my Athlon 3000+ was relatively hot (49C) although it was completely idle. At that time I was running 2.6.3-mm3 with ACPI and IOAPIC-support enabled.

Quoted 7 lines
Quoted 3 lines

As I tried 2.6.3, the idle temperature was at normal 39C. So I did do some binary search with the -bk patches and found the patch that causes the high idle temperature. It is [email protected] aka 8259-timer-ack-fix.patch.

Interesting -- the patch removes a pair of unnecessary for your configuration PIC accesses when using an I/O APIC NMI watchdog. You have the NMI watchdog enabled, don't you?

Quoted 6 lines

No, I don't use the NMI watchdog...
So the optimization of removing these I/O accesses is bogus for my configuration. Btw. I don't know if I already mentioned it, but I use the VIA KT400 chipset. Maybe this is of interest...

The only way to cool down my CPU was to enable timer_ack.
I don't know how to help you, but of course I am willing to test patches... ;-)
Thomas

I agree with Len Brown's comments to try to examine which power saving state but if you want to try to brute force C1 state ( only works if chipset supported ) you could try this patch for process.c,
(ignore the io-apic patch as it is nforce2 specific).

http://linux.derkeiler.com/Mailing-Lists/Kernel/2004-02/6520.html
The KERNEL ARG to invoke it is "idle=C1halt".

It has an extra function pointer to prevent the power management idle routine hikjacking things after the command line arg has requested an idle routine.

These idle mods appear to assist more than just nforce2 Athlon boards. Thomas Herrmann has had success with an SIS740

Quoted 9 lines

Hi Ross,
I just want to let you know that your nforce2_idle patch does work with the SiS740 chipset too. While the current ACPI patch already routes the timer of the SiS740 to IO-APIC-edge with out the C1halt option of your nforce2_idle patch the system locked up when STPGNT was enabled. But after I applied your nforce2_idle patch to kernel 2.4.24 together with the C1halt kernel boot option, the system runs stable for hours.
Great work, thanks!
Best regards, Thomas

Craig Bradney has put it into the gentoo dev sources also.
http://linux.derkeiler.com/Mailing-Lists/Kernel/2004-03/1746.html

Hope this helps,
Ross.

#1 lore raw
Quoted 43 lines
Quoted 12 lines
Quoted 11 lines
Quoted 7 lines

a few days ago I noticed that my Athlon 3000+ was relatively hot (49C) although it was completely idle. At that time I was running 2.6.3-mm3 with ACPI and IOAPIC-support enabled.

As I tried 2.6.3, the idle temperature was at normal 39C. So I did do some binary search with the -bk patches and found the patch that causes the high idle temperature. It is [email protected] aka 8259-timer-ack-fix.patch.

Interesting -- the patch removes a pair of unnecessary for your configuration PIC accesses when using an I/O APIC NMI watchdog. You have the NMI watchdog enabled, don't you?

Quoted 6 lines

No, I don't use the NMI watchdog...
So the optimization of removing these I/O accesses is bogus for my configuration. Btw. I don't know if I already mentioned it, but I use the VIA KT400 chipset. Maybe this is of interest...

The only way to cool down my CPU was to enable timer_ack.
I don't know how to help you, but of course I am willing to test patches... ;-)
Thomas

I agree with Len Brown's comments to try to examine which power saving state but if you want to try to brute force C1 state ( only works if chipset supported ) you could try this patch for process.c,
(ignore the io-apic patch as it is nforce2 specific).

http://linux.derkeiler.com/Mailing-Lists/Kernel/2004-02/6520.html
The KERNEL ARG to invoke it is "idle=C1halt".

It has an extra function pointer to prevent the power management idle routine hikjacking things after the command line arg has requested an idle routine.

These idle mods appear to assist more than just nforce2 Athlon boards. Thomas Herrmann has had success with an SIS740

Quoted 9 lines

Hi Ross,
I just want to let you know that your nforce2_idle patch does work with the SiS740 chipset too. While the current ACPI patch already routes the timer of the SiS740 to IO-APIC-edge with out the C1halt option of your nforce2_idle patch the system locked up when STPGNT was enabled. But after I applied your nforce2_idle patch to kernel 2.4.24 together with the C1halt kernel boot option, the system runs stable for hours.
Great work, thanks!
Best regards, Thomas

Quoted 2 lines

Craig Bradney has put it into the gentoo dev sources also.
http://linux.derkeiler.com/Mailing-Lists/Kernel/2004-03/1746.html

Ross, your patch is pretty damn stable.. I had to shut down the PC today due to expected wiring changes (lucky, a blackout came before the wiring changes!).. the PC was at over 8 days uptime. At first I thought the box was cooler.. now my observations are as follows:
At idle and under load the motherboard temp is 1-2C lower. At idle the CPU is about the same
When compiling it gets to be a bit hotter than before but I have taken one fan offline so I'm not sure of the result but its still only reaches 45C. I'll try again soon with the 2nd fan on and see what its like.

Craig

#2 lore raw
In reply to Ross Dickson

Hi,

I'm just testing your IdleC1Halt patch (didn't reboot yet) with 2.6.4, but there is a problem if apm is enabled in the configuration:

arch/i386/kernel/built-in.o(.text+0x10b65): In function `apm_cpu_idle':
: undefined reference to `default_idle'

Your patch sets default_idle() static, so its not available in apm.c file.

Usually I compile with acpi and apm support to switch between both in case of an unsual problem, and I think many people also do so.

Quoted 3 lines

http://linux.derkeiler.com/Mailing-Lists/Kernel/2004-02/6520.html
The KERNEL ARG to invoke it is "idle=C1halt".

Thanks,
Bernd

#3 lore raw
Quoted 10 lines · Bernd Schubert

Hi,

I'm just testing your IdleC1Halt patch (didn't reboot yet) with 2.6.4, but there is a problem if apm is enabled in the configuration:

arch/i386/kernel/built-in.o(.text+0x10b65): In function `apm_cpu_idle':
: undefined reference to `default_idle'

Your patch sets default_idle() static, so its not available in apm.c file.

Apologies - Unnecessary change, I only use acpi and was cleaning things up.

It should still work fine by removing the static from it as "idle()" get first try. change
static void default_idle(void)
back to
void default_idle(void)

Quoted 2 lines

Usually I compile with acpi and apm support to switch between both in case of an unsual problem, and I think many people also do so.

Please let me know how it goes.

Ross.

Quoted 12 lines
Quoted 5 lines

http://linux.derkeiler.com/Mailing-Lists/Kernel/2004-02/6520.html
The KERNEL ARG to invoke it is "idle=C1halt".

Thanks,
Bernd

#4 lore raw
In reply to Ross Dickson

Am Donnerstag, 18. März 2004 01:19 schrieb Ross Dickson:

~~ snip ~~

Quoted 29 lines
Quoted 3 lines

The only way to cool down my CPU was to enable timer_ack.
I don't know how to help you, but of course I am willing to test
patches... ;-) Thomas

I agree with Len Brown's comments to try to examine which power saving state but if you want to try to brute force C1 state ( only works if chipset supported ) you could try this patch for process.c, (ignore the io-apic patch as it is nforce2 specific).

http://linux.derkeiler.com/Mailing-Lists/Kernel/2004-02/6520.html
The KERNEL ARG to invoke it is "idle=C1halt".

It has an extra function pointer to prevent the power management idle routine hikjacking things after the command line arg has requested an idle routine.

These idle mods appear to assist more than just nforce2 Athlon boards. Thomas Herrmann has had success with an SIS740

Quoted 9 lines

Hi Ross,
I just want to let you know that your nforce2_idle patch does work with the SiS740 chipset too. While the current ACPI patch already routes the timer of the SiS740 to IO-APIC-edge with out the C1halt option of your nforce2_idle patch the system locked up when STPGNT was enabled. But after I applied your nforce2_idle patch to kernel 2.4.24 together with the C1halt kernel boot option, the system runs stable for hours. Great work, thanks!
Best regards, Thomas

Quoted 2 lines

Craig Bradney has put it into the gentoo dev sources also.
http://linux.derkeiler.com/Mailing-Lists/Kernel/2004-03/1746.html

OK, now I had the time to test if different C states are working with following three kernels:

1. 2.6.4-mm2 without the 8259-timer-ack-fix.patch and without the C1halt idle function.
2. 2.6.4-mm2 with the 8259-timer-ack-fix.patch and without the C1 halt idle function enabled.
3. 2.6.4-mm2 with the 8259-timer-ack-fix.patch and with the C1 halt idle function enabled.

I used following script to print the C-state counters on an complete idle machine before and after a 10second interval:

# /bin/sh
cat /proc/acpi/processor/CPU0/power
sleep 10
cat /proc/acpi/processor/CPU0/power

Now the results:

1.:

active state:            C2
default state:           C1
bus master activity:     00000000

states:
C1: promotion[C2] demotion[--] latency[000]
usage[00006280]
*C2: promotion[--] demotion[C1] latency[100]
usage[00300041]

    C3:                  <not supported>
active state:            C2
default state:           C1
bus master activity:     00000000

states:
C1: promotion[C2] demotion[--] latency[000]
usage[00006300]
*C2: promotion[--] demotion[C1] latency[100]
usage[00310045]
C3: <not supported>

2.:

active state:            C1
default state:           C1
bus master activity:     00000000

states:
*C1: promotion[C2] demotion[--] latency[000]
usage[00000000]
C2: promotion[--] demotion[C1] latency[100]
usage[00000000]

    C3:                  <not supported>
active state:            C1
default state:           C1
bus master activity:     00000000

states:
*C1: promotion[C2] demotion[--] latency[000]
usage[00000000]
C2: promotion[--] demotion[C1] latency[100]
usage[00000000]
C3: <not supported>

3.:

active state:            C1
default state:           C1
bus master activity:     00000000

states:
*C1: promotion[C2] demotion[--] latency[000]
usage[00000000]
C2: promotion[--] demotion[C1] latency[100]
usage[00000000]

    C3:                  <not supported>
active state:            C1
default state:           C1
bus master activity:     00000000

states:
*C1: promotion[C2] demotion[--] latency[000]
usage[00000000]
C2: promotion[--] demotion[C1] latency[100]
usage[00000000]
C3: <not supported>

So, as you can see, the C1halt patch does not help here... ;-(

Regards
Thomas Schlichter

#5 lore raw
Quoted 19 lines

OK, now I had the time to test if different C states are working with following three kernels:

1. 2.6.4-mm2 without the 8259-timer-ack-fix.patch and without the C1halt idle function.
2. 2.6.4-mm2 with the 8259-timer-ack-fix.patch and without the C1 halt idle function enabled.
3. 2.6.4-mm2 with the 8259-timer-ack-fix.patch and with the C1 halt idle function enabled.

I used following script to print the C-state counters on an complete idle machine before and after a 10second interval:

# /bin/sh
cat /proc/acpi/processor/CPU0/power
sleep 10
cat /proc/acpi/processor/CPU0/power

Now the results:

[snip]

Quoted 13 lines

3.:

active state:            C1
default state:           C1
bus master activity:     00000000

states:
*C1: promotion[C2] demotion[--] latency[000]
usage[00000000]
C2: promotion[--] demotion[C1] latency[100]
usage[00000000]

    C3:                  <not supported>
active state:            C1
default state:           C1
bus master activity:     00000000
Quoted 8 lines

states:
*C1: promotion[C2] demotion[--] latency[000]
usage[00000000]
C2: promotion[--] demotion[C1] latency[100]
usage[00000000]
C3: <not supported>

So, as you can see, the C1halt patch does not help here... ;-(

Hmm, I just did a cat /proc/acpi/processor/CPU0/power:

active state:            C1
default state:           C1
bus master activity:     00000000

states:
*C1: promotion[--] demotion[--] latency[000]
usage[00000000]
C2: <not supported>
C3: <not supported>

I am currently NOT using APIC mode (nforce2, as well) and using vanilla 2.6.4. It seems C1 halt state isn't used, which exlains why I am having trouble to keep my CPU cooler these day. I once started a thread suspecting acpi timer, but it is not the case. It seems to be something else. As I don't use PIC, it cannot be that 8259-timer-ack-fix.patch causin git, or can it? Maybe something broken in ACPI? I might try out older kernels to find out...

Prakash

#6 lore raw
Quoted 18 lines · Prakash K. Cheemplavam

Hmm, I just did a cat /proc/acpi/processor/CPU0/power:

active state:            C1
default state:           C1
bus master activity:     00000000
states:
    *C1:                  promotion[--] demotion[--] latency[000] 

usage[00000000]
C2: <not supported>
C3: <not supported>

I am currently NOT using APIC mode (nforce2, as well) and using vanilla 2.6.4. It seems C1 halt state isn't used, which exlains why I am having trouble to keep my CPU cooler these day. I once started a thread suspecting acpi timer, but it is not the case. It seems to be something else. As I don't use PIC, it cannot be that 8259-timer-ack-fix.patch causin git, or can it? Maybe something broken in ACPI? I might try out older kernels to find out...

Quoted 1 line

Prakash

Actually I think it is that we don't _count_ C1 usage.

cheers,
-Len

#7 lore raw
Quoted 14 lines · Len Brown
Quoted 11 lines · Prakash K. Cheemplavam

Hmm, I just did a cat /proc/acpi/processor/CPU0/power:

active state:            C1
default state:           C1
bus master activity:     00000000
states:
   *C1:                  promotion[--] demotion[--] latency[000] 

usage[00000000]
C2: <not supported>
C3: <not supported>

I am currently NOT using APIC mode (nforce2, as well) and using vanilla 2.6.4. It seems C1 halt state isn't used, which exlains why I am having

[snip]

Quoted 3 lines

Actually I think it is that we don't _count_ C1 usage.

Hmm, OK, then I am really puzzled what specifically about mm sources make my idle temps hotter, as I still couldn't properly resolve it what is causing it. I thought ACPI, but no, using APM only does the same (apm only with vanilla is low temp though.)

Prakash

#8 lore raw
Quoted 19 lines · Prakash K. Cheemplavam
Quoted 15 lines · Len Brown
Quoted 12 lines · Prakash K. Cheemplavam

Hmm, I just did a cat /proc/acpi/processor/CPU0/power:

active state:            C1
default state:           C1
bus master activity:     00000000
states:
   *C1:                  promotion[--] demotion[--] latency[000] 

usage[00000000]
C2: <not supported>
C3: <not supported>

I am currently NOT using APIC mode (nforce2, as well) and using vanilla 2.6.4. It seems C1 halt state isn't used, which exlains why I am having

[snip]

Quoted 2 lines
Quoted 6 lines
Quoted 1 line

Actually I think it is that we don't _count_ C1 usage.

Hmm, OK, then I am really puzzled what specifically about mm sources make my idle temps hotter, as I still couldn't properly resolve it what is causing it. I thought ACPI, but no, using APM only does the same (apm only with vanilla is low temp though.)

Hi Prakash,

Have you seen this thread, it may be relevant?
Re: [2.6.4-rc2] bogus semicolon behind if()
http://linux.derkeiler.com/Mailing-Lists/Kernel/2004-03/4170.html

I have not looked to see which kern sources besides 2.6.4-rc2 may have it.

Regards
Ross.

Quoted 5 lines

Prakash

#9 lore raw
Quoted 23 lines · Ross Dickson
Quoted 21 lines · Prakash K. Cheemplavam
Quoted 15 lines · Len Brown
Quoted 11 lines · Prakash K. Cheemplavam

Hmm, I just did a cat /proc/acpi/processor/CPU0/power:

active state:            C1
default state:           C1
bus master activity:     00000000

states:
*C1: promotion[--] demotion[--] latency[000]
usage[00000000]
C2: <not supported>
C3: <not supported>

I am currently NOT using APIC mode (nforce2, as well) and using vanilla 2.6.4. It seems C1 halt state isn't used, which exlains why I am having

[snip]

Quoted 1 line
Quoted 8 lines
Quoted 3 lines
Quoted 1 line

Actually I think it is that we don't _count_ C1 usage.

Hmm, OK, then I am really puzzled what specifically about mm sources make my idle temps hotter, as I still couldn't properly resolve it what is causing it. I thought ACPI, but no, using APM only does the same (apm only with vanilla is low temp though.)

Have you seen this thread, it may be relevant?
Re: [2.6.4-rc2] bogus semicolon behind if()
http://linux.derkeiler.com/Mailing-Lists/Kernel/2004-03/4170.html

Hi Ross, I don't think so, as I currently don't use APIC and thus fix in above post wouldn't help me. Or should I read further?

cya,

Prakash

#10 lore raw
Quoted 12 lines · Prakash K. Cheemplavam
Quoted 11 lines · Ross Dickson
Quoted 9 lines

[snip]

Quoted 2 lines

Actually I think it is that we don't _count_ C1 usage.

Hmm, OK, then I am really puzzled what specifically about mm sources make my idle temps hotter, as I still couldn't properly resolve it what is causing it. I thought ACPI, but no, using APM only does the same (apm only with vanilla is low temp though.)

Quoted 6 lines
Quoted 3 lines

Have you seen this thread, it may be relevant?
Re: [2.6.4-rc2] bogus semicolon behind if()
http://linux.derkeiler.com/Mailing-Lists/Kernel/2004-03/4170.html

Hi Ross, I don't think so, as I currently don't use APIC and thus fix in above post wouldn't help me. Or should I read further?

Hmm Valid point.
bye
Ross.

Quoted 8 lines

cya,

Prakash

#14 lore raw
In reply to Len Brown
Quoted 1 line · Len Brown

Actually I think it is that we don't _count_ C1 usage.

Are you sure? At least it seems we do in 2.6.4:

active state:            C2
default state:           C1
bus master activity:     00000000

states:
C1: promotion[C2] demotion[--] latency[000]
usage[00067690]
*C2: promotion[--] demotion[C1] latency[090]
usage[00673373]
C3: <not supported>

Servus,
Daniel

#11 lore raw
In reply to Prakash K. Cheemplavam
Quoted 29 lines · Prakash K. Cheemplavam
Quoted 27 lines · Ross Dickson
Quoted 25 lines · Prakash K. Cheemplavam
Quoted 15 lines · Len Brown
Quoted 11 lines · Prakash K. Cheemplavam

Hmm, I just did a cat /proc/acpi/processor/CPU0/power:

active state:            C1
default state:           C1
bus master activity:     00000000

states:
*C1: promotion[--] demotion[--] latency[000]
usage[00000000]
C2: <not supported>
C3: <not supported>

I am currently NOT using APIC mode (nforce2, as well) and using vanilla 2.6.4. It seems C1 halt state isn't used, which exlains why I am having

[snip]

Quoted 2 lines

Actually I think it is that we don't _count_ C1 usage.

Quoted 9 lines
Quoted 7 lines
Quoted 1 line

Hmm, OK, then I am really puzzled what specifically about mm sources make my idle temps hotter, as I still couldn't properly resolve it what is causing it. I thought ACPI, but no, using APM only does the same (apm only with vanilla is low temp though.)

Have you seen this thread, it may be relevant?
Re: [2.6.4-rc2] bogus semicolon behind if()
http://linux.derkeiler.com/Mailing-Lists/Kernel/2004-03/4170.html

So, I seem to have found the bugger causing higher temps: It is NVidia binary driver, or rather its AGP part of the 53.36 driver. Using AGPGART and Nvidia driver leaves my system cool. Using NVAGP it seems as though C1 state isn't actually used anymore thus making the CPU hotter.

Tested with (and without) ACPI and APIC (and Ross' tack patch). Currently running in PIC mode (with ACPI) and idle temp of 44°C (instead of about 50°C...). But it was as cool in APIC mode.

Of course I have to test few more days, but at least currently I am happy again. :-)

Prakash

#12 lore raw
Quoted 45 lines · Prakash K. Cheemplavam
Quoted 38 lines · Prakash K. Cheemplavam
Quoted 36 lines · Ross Dickson
Quoted 31 lines · Prakash K. Cheemplavam
Quoted 17 lines · Len Brown
Quoted 13 lines · Prakash K. Cheemplavam

Hmm, I just did a cat /proc/acpi/processor/CPU0/power:

active state:            C1
default state:           C1
bus master activity:     00000000

states:
*C1: promotion[--] demotion[--] latency[000]
usage[00000000]
C2: <not supported>
C3: <not supported>

I am currently NOT using APIC mode (nforce2, as well) and using vanilla 2.6.4. It seems C1 halt state isn't used, which exlains why I am having

[snip]

Quoted 2 lines

Actually I think it is that we don't _count_ C1 usage.

Hmm, OK, then I am really puzzled what specifically about mm sources make my idle temps hotter, as I still couldn't properly resolve it what is causing it. I thought ACPI, but no, using APM only does the same (apm only with vanilla is low temp though.)

Quoted 5 lines
Quoted 3 lines

Have you seen this thread, it may be relevant?
Re: [2.6.4-rc2] bogus semicolon behind if()
http://linux.derkeiler.com/Mailing-Lists/Kernel/2004-03/4170.html

Quoted 4 lines

So, I seem to have found the bugger causing higher temps: It is NVidia binary driver, or rather its AGP part of the 53.36 driver. Using AGPGART and Nvidia driver leaves my system cool. Using NVAGP it seems as though C1 state isn't actually used anymore thus making the CPU hotter.

Hmmm.
Would you happen to have a copy of athcool handy - it would be interesting to see the northbridge disconnect bit status - if its been turned off by their driver?

Quoted 12 lines

Tested with (and without) ACPI and APIC (and Ross' tack patch). Currently running in PIC mode (with ACPI) and idle temp of 44°C (instead of about 50°C...). But it was as cool in APIC mode.

Of course I have to test few more days, but at least currently I am happy again. :-)

Prakash

#13 lore raw
Quoted 46 lines · Ross Dickson
Quoted 44 lines · Prakash K. Cheemplavam
Quoted 42 lines · Prakash K. Cheemplavam
Quoted 37 lines · Ross Dickson
Quoted 21 lines · Prakash K. Cheemplavam
Quoted 15 lines · Len Brown
Quoted 10 lines · Prakash K. Cheemplavam

Hmm, I just did a cat /proc/acpi/processor/CPU0/power:

active state:            C1
default state:           C1
bus master activity:     00000000

states:
*C1: promotion[--] demotion[--] latency[000]
usage[00000000]
C2: <not supported>
C3: <not supported>

Quoted 1 line
Quoted 1 line

I am currently NOT using APIC mode (nforce2, as well) and using vanilla 2.6.4. It seems C1 halt state isn't used, which exlains why I am having

Quoted 7 lines

[snip]

Quoted 1 line

Actually I think it is that we don't _count_ C1 usage.

Hmm, OK, then I am really puzzled what specifically about mm sources make my idle temps hotter, as I still couldn't properly resolve it what is causing it. I thought ACPI, but no, using APM only does the same (apm only with vanilla is low temp though.)

Have you seen this thread, it may be relevant?
Re: [2.6.4-rc2] bogus semicolon behind if()
http://linux.derkeiler.com/Mailing-Lists/Kernel/2004-03/4170.html

Quoted 6 lines
Quoted 1 line

So, I seem to have found the bugger causing higher temps: It is NVidia binary driver, or rather its AGP part of the 53.36 driver. Using AGPGART and Nvidia driver leaves my system cool. Using NVAGP it seems as though C1 state isn't actually used anymore thus making the CPU hotter.

Hmmm.
Would you happen to have a copy of athcool handy - it would be interesting to see the northbridge disconnect bit status - if its been turned off by their driver?

That is the funny thing: Athcool reports it is on. So something from their AGP code seems to prevent the CPU going "full idle", I guess. Well, now I am using 53.41 driver with agpgart and everything finally seems to be stable, cool and nice. :)

Prakash