Package: linux-image-2.6.32-5-powerpc64
Version: 2.6.32-31
Tags: patch fixed-upstream squeeze

The 2.6.32-31 kernel from squeeze-proposed-updates (and 2.6.32-30 from squeeze) sets the frequency transition latency on PowerMac G5s' CPUs to CPUFREQ_ETERNAL:

# uname -a
Linux crash 2.6.32-5-powerpc64 #1 SMP Tue Mar 8 02:01:42 UTC 2011 ppc64 GNU/Linux

# dpkg -l linux-image-* | grep ^ii
ii linux-image-2.6-powerpc64 2.6.32+29 Linux 2.6 for 64-bit PowerPC (meta-package)
ii linux-image-2.6.32-5-powerpc64 2.6.32-31 Linux 2.6.32 for 64-bit PowerPC

# cat /proc/cpuinfo
processor : 0
cpu : PPC970FX, altivec supported
clock : 2700.000000MHz
revision : 3.1 (pvr 003c 0301)

processor : 1
cpu : PPC970FX, altivec supported
clock : 2700.000000MHz
revision : 3.1 (pvr 003c 0301)

timebase : 33333333
platform : PowerMac
model : PowerMac7,3
machine : PowerMac7,3
motherboard : PowerMac7,3 MacRISC4 Power Macintosh
detected as : 336 (PowerMac G5)
pmac flags : 00000000
L2 cache : 512K unified
pmac-generation : NewWorld

# grep . /sys/devices/system/cpu/cpu?/cpufreq/cpuinfo_transition_latency
/sys/devices/system/cpu/cpu0/cpufreq/cpuinfo_transition_latency:4294967295
/sys/devices/system/cpu/cpu1/cpufreq/cpuinfo_transition_latency:4294967295

This keeps the ondemand cpufreq governor from working:

# for x in 0 1; do echo ondemand > /sys/devices/system/cpu/cpu$x/cpufreq/scaling_governor; done

# dmesg | tail -n 2
[ 1380.988816] ondemand governor failed, too long transition latency of HW, fallback to performance governor
[ 1380.990366] ondemand governor failed, too long transition latency of HW, fallback to performance governor

This was fixed in the 2.6.32.* tree in March 2009, and the same code is still in use in 2.6.38.

Can Debian's 2.6.32 kernel pick up this patch as well?

http://git.kernel.org/?p=linux/kernel/git/longterm/linux-2.6.32.y.git;a=commitdiff_plain;h=16962e7ce1dce29e 1e92d231ac7d6844d7385d54

Alternatively, in case the above link gets mangled by wrapping:

http://bit.ly/gYKmiV


--
To UNSUBSCRIBE, email to debian-kernel-REQUEST@lists.debian.org
with a subject of "unsubscribe". Trouble? Contact listmaster@lists.debian.org
Archive: 3417259A-A31C-43B6-A4D4-4A0DF5BBA92F@vt.edu">http://lists.debian.org/3417259A-A31C-43B6-A4D4-4A0DF5BBA92F@vt.edu

> We already have all the fixes up to 2.6.32.32 (as noted in the
> changelog).
...
> Notice that this sets a finite transition latency for 'Neo2' platforms
> but not for other systems (PowerMac7,2; PowerMac7,3; and RackMac3,1).
> Unfortunately you have one of the latter.

You're absolutely right, of course. This is what I get for not looking at the patch more closely, then not looking at the source package to see which patches were already being applied. My apologies for the lame bug report.

For others who have the same problem and stumble across this bug report:

I've built a kernel that sets transition_latency to 20000 (nanoseconds) in the "pm72" code path (line 680 of arch/powerpc/platforms/powermac/cpufreq_64.c). This is enough to get the ondemand and conservative cpufreq governors working on a dual 2.7 GHz PowerMac G5, though 20000 ns is a complete guess at the transition latency and is most likely very wrong.


--
To UNSUBSCRIBE, email to debian-kernel-REQUEST@lists.debian.org
with a subject of "unsubscribe". Trouble? Contact listmaster@lists.debian.org
Archive: 3D7254FC-F2E1-4A6C-B35B-943AB36817B0@vt.edu">http://lists.debian.org/3D7254FC-F2E1-4A6C-B35B-943AB36817B0@vt.edu