efficeon and longrun
Hi,
I recently got a Sharp MP70G with a 1.6 GHz efficeon processor and have some questions regarding longrun support. I am using longrun-0.9-15 and kernel 2.6.10-rc2.
In arch/i386/kernel/cpu/proc.c the kernel seems to check bit 3 for the lrti capability, longrun.c checks bit2.
Are thermal extensions different on the efficeon compared to the crusoe? No matter what I choose (between 0 and 7) the level field in the ATM register is always 0.
Are there any new efficeon features not yet supported in the kernel or the longrun utility? Is there any register information available anywhere? I looked on the transmeta pages but did not find anything about this.
The BIOS also allows one to select C4 as possible power level, /proc/acpi/processor/CPU0/power only offers C1, C2 and C3.
I ask because under Windows, if nothing is running, the fan will stay off. Under Linux it is turning on about every two minutes and will then run for about a minute (also with everything turned off, and about every daemon shut down).
Cheers,
Ralph Metzler
Quoted 6 lines · Ralph Metzler
I recently got a Sharp MP70G with a 1.6 GHz efficeon processor and have some questions regarding longrun support. I am using longrun-0.9-15 and kernel 2.6.10-rc2.
In
arch/i386/kernel/cpu/proc.cthe kernel seems to check bit 3 for the lrti capability,longrun.cchecks bit2.
bit 3 is the correct bit for lrti... this is one of the few things we actually have in our scant public docs
<http://www.transmeta.com/crusoe_docs/Crusoe_CPUID_5-7-02.pdf>, see page 6.
the kernel appears to check the right bit, but the debian userland longrun tool doesn't... it probably hasn't mattered until now because many crusoe with LRTI also have bit 2 set (bit 2 isn't documented, sorry).
Quoted 3 lines
Are thermal extensions different on the efficeon compared to the crusoe? No matter what I choose (between 0 and 7) the level field in the ATM register is always 0.
in theory they're backward compatible -- i've actually been using a longrun executable compiled years ago on crusoe and never bothered recompiling it on efficeon. but there are definitely some quirks -- and the quirks appear to interfere poorly with the debian longrun-0.9-15 package. (i bet i'm using a longrun.c which is old enough to pre-date lrti...)
the tables defined for crusoe include information that's meaningless on efficeon -- crusoe has two memory controllers, and only one clock domain for the entire processor. when crusoe is varying the frequency it also stops/starts the memory bus and possibly runs memory at different rates -- the longrun tables include all this divisor information.
efficeon has only one memory controller, and two clock domains, one for the "northbridge" and one for the "core". only the core speed varies after boot -- the memory clocks at a fixed rate. so the divisors in the longrun table are essentially meaningless. in fact it looks like they come back as all zeros... and this causes the debian longrun-0.9-15 some angst.
below is a patch to debian longrun-0.9-15 -- it passes the "works for me" test on my development efficeon board, but i don't have a crusoe handy from remote this weekend to verify it works there. i've placed pre-built debs at <http://arctic.org/~dean/longrun/> ... if crusoe/efficeon users can verify them i'll submit my patch upstream (make sure "longrun -l", "-p" and "-s low hi" work).
Quoted 4 lines
Are there any new efficeon features not yet supported in the kernel or the longrun utility? Is there any register information available anywhere? I looked on the transmeta pages but did not find anything about this.
we suck at publishing this stuff (it makes it out to our customers, but not out for general consumption). i've no clue where it's all linked from, but google finds a bunch of stuff -- try a search for "crusoe_docs site:transmeta.com". there's nothing apparently recent there though.
aside from the efficeon kernel patch i submitted (it's in 2.6.10-rc2, and i think -rc1) there aren't any efficeon-specific features that i think need supporting. there's tweaking you could do -- but it's mostly userland (i.e. gcc -msse2 -mfpmath=sse).
Quoted 7 lines
The BIOS also allows one to select C4 as possible power level,
/proc/acpi/processor/CPU0/poweronly offers C1, C2 and C3.I ask because under Windows, if nothing is running, the fan will stay off. Under Linux it is turning on about every two minutes and will then run for about a minute (also with everything turned off, and about every daemon shut down).
maybe try building your kernel with HZ=100 instead of default 1000? maybe at HZ=1000 it wakes up just enough to cross some unfortunate power threshold. my next suggestion beyond that would be to see if ext2 and/or noflushd do anything... i don't know if it's still true, but ext3 used to cause sync's every 5s... and didn't interact well with noflushd.
-dean
debian/changelog +12 -0
@@ -1,3 +1,15 @@ +longrun (0.9-15.dg1) unstable; urgency=low + + * LRTI is bit 3, not bit 2... the old code happens to have worked on + many crusoe because they also set bit 2, but this bit is not set on + all crusoe, nor is it set on efficeon. (i'm not sure if bit 2 is + publically documented). + * mhz * millivolt * millivolt power estimation overflows 31-bit + signed integer once the clock is fast enough... use double instead. + * io_div (and most divisors) are 0 on efficeon, avoid div-by-0. + + -- dean gaudet <[email protected]> Mon, 29 Nov 2004 00:38:13 -0800 + longrun (0.9-15) unstable; urgency=low * Fix mistake on man page. Closes: #249866
longrun.c +15 -8
@@ -32,6 +32,7 @@#include <string.h> #include <sys/io.h> #include <sys/sysmacros.h> +#include <locale.h> #define __USE_UNIX98/* for pread/pwrite */ #define __USE_FILE_OFFSET64/* we should use 64 bit offset for pread/pwrite */ #include <unistd.h>
@@ -68,7 +69,7 @@#define CPUID_TMx86_PROCESSOR_INFO 0x80860001#define CPUID_TMx86_LONGRUN_STATUS 0x80860007#define CPUID_TMx86_FEATURE_LONGRUN (x) ((x) & 0x02) -#define CPUID_TMx86_FEATURE_LRTI (x) ((x) & 0x04) +#define CPUID_TMx86_FEATURE_LRTI (x) ((x) & 0x08)/* Advanced Thermal Management */ #define LR_NORTHBRIDGE "/proc/bus/pci/00/00.0"
@@ -242,21 +243,23 @@ {int max;int junk; -int max_voltage, max_mhz, max_power; +int max_voltage, max_mhz; +double max_power;read_msr (MSR_TMx86_LRTI_READOUT, &junk, &max);write_msr (MSR_TMx86_LRTI_READOUT,0 , max);read_msr (MSR_TMx86_LRTI_VOLT_MHZ, &max_mhz, &max_voltage); - max_power = max_mhz * max_voltage * max_voltage; + max_power = ((double )max_mhz * max_voltage) * max_voltage;printf (_ ("# %% MHz Volts usage SDR DDR PCI\n" ));int i;for (i = max; i >=0 ; i--) { -int percent, mhz, voltage, power; +int percent, mhz, voltage; +double power;int ddr_mem_div, sdr_mem_div, ddr_freq, sdr_freq;int io_div;write_msr (MSR_TMx86_LRTI_READOUT, i, max);
@@ -267,9 +270,9 @@read_msr (MSR_TMx86_LRTI_VOLT_MHZ, &mhz, &voltage);printf (_ ("%5d %6.3f " ), mhz, (voltage /1000.0 )); - power = mhz * voltage * voltage; + power = ((double )mhz * voltage) * voltage; -printf (_ ("%6.3f " ), ((float ) power) / max_power); +printf (_ ("%6.3f " ), ((double ) power) / max_power);// TODO: Watch for zeros read_msr (MSR_TMx86_LRTI_DIV_MEM, &sdr_mem_div, &ddr_mem_div);
@@ -289,9 +292,13 @@printf (_ ("%4d %4d " ), sdr_freq, ddr_freq); -// TODO: Watch for zeros read_msr (MSR_TMx86_LRTI_DIV_IO, &io_div, &junk); -printf (_ ("%4d\n" ), mhz / io_div); +if (io_div) { +printf (_ ("%4d\n" ), mhz / io_div); + } +else { +printf ("%4s\n" ,"--" ); + }/* // Most of the rest of these flags don't appear too interesting
Followup to: <[email protected]> By author: Ralph Metzler <[email protected]> In newsgroup: linux.dev.kernel
Quoted 5 lines
I recently got a Sharp MP70G with a 1.6 GHz efficeon processor and have some questions regarding longrun support. I am using longrun-0.9-15 and kernel 2.6.10-rc2.
longrun-0.9 is hideously out of date, and was never debugged to begin with. Given that these days longrun is handled via cpufreq, there doesn't seem to be much reason for the standalone longrun program.
Quoted 2 lines
In
arch/i386/kernel/cpu/proc.cthe kernel seems to check bit 3 for the lrti capability,longrun.cchecks bit2.
Bit 3 is correct.
Quoted 3 lines
Are thermal extensions different on the efficeon compared to the crusoe? No matter what I choose (between 0 and 7) the level field in the ATM register is always 0.
You shouldn't muck with it anyway. And yes, they are different. The thermal extensions are handled via ACPI code, so as long as the ACPI is handled correctly, it shouldn't matter.
Quoted 4 lines
Are there any new efficeon features not yet supported in the kernel or the longrun utility? Is there any register information available anywhere? I looked on the transmeta pages but did not find anything about this.
The algorithms are very different, but it's all under-the-hood stuff implemented in firmware.
Quoted 2 lines
The BIOS also allows one to select C4 as possible power level,
/proc/acpi/processor/CPU0/poweronly offers C1, C2 and C3.
C4 support (DSX) is pretty keen.
Quoted 4 lines
I ask because under Windows, if nothing is running, the fan will stay off. Under Linux it is turning on about every two minutes and will then run for about a minute (also with everything turned off, and about every daemon shut down).
-hpa
Quoted 3 lines · H. Peter Anvin
longrun-0.9 is hideously out of date, and was never debugged to begin with. Given that these days longrun is handled via cpufreq, there doesn't seem to be much reason for the standalone longrun program.
the tool still has a place... for folks not using cpufreq/2.6 especially. but also the longrun cpufreq driver is lacking support for scaling_available_frequencies, and doesn't display the voltages anywhere. in most cases the ACPI P-states driver works fine instead though.
i prefer the tool -- but then my requirements are pretty specific (i don't want cpufreq doing anything i'm not expecting while doing perf/debugging work).
-dean
Hi,
Quoted 7 lines · dean gaudet
Quoted 3 lines · H. Peter Anvin
longrun-0.9 is hideously out of date, and was never debugged to begin with. Given that these days longrun is handled via cpufreq, there doesn't seem to be much reason for the standalone longrun program.
the tool still has a place... for folks not using cpufreq/2.6 especially.
There is still a cpufreq/2.4 that I try sometimes to maintain and support longrun.
Quoted 3 lines
but also the longrun cpufreq driver is lacking support for
scaling_available_frequencies, and doesn't display the voltages anywhere. in most cases the ACPI P-states driver works fine instead though.
That may be added if you justifies that enough. Apparently, you want a fixed frequency cpufreq driver for longrun?
Quoted 3 lines
i prefer the tool -- but then my requirements are pretty specific (i don't want cpufreq doing anything i'm not expecting while doing perf/debugging work).
cpufreq can help you IMHO. All the stuff you want can be done probably via cpufreq. Debugging support, governors for your perf work if this is done in kernel space, notifier interface, and last but not the least, a more generic infrastructure so that you can change processors or architectures at will.
Please ask what you need exactly but at the cpufreq mailing list.
Cheers,
Signature · 4 lines
Bruno Ducrot -- Which is worse: ignorance or apathy? -- Don't know. Don't care.