nearly-working support for cpufreq on 2004 iBook G4 with 7447A cpu

nearly-working support for cpufreq on 2004 iBook G4 with 7447A cpu

[John Steele Scott]

Hi all,

I have written a patch against kernel 2.6.6 which lets the pmac cpufreq driver
support dynamic frequency scaling on the 7447A cpu in my new iBook. It also
fixes the incorrect clock speed reporting (i.e. on boot 2.6.6 reports cpu is
running at 1066 MHz when really it is 533 MHz).

Unfortunately, if I use this to set my cpu frequency up to 1 GHz, my system
locks up hard when I try and do something cpu-intensive. Christiaan Welvaart
has tested the previous version of this patch (functionally the same, but a
bit messier), and on his system it works fine. Yet even if I use the exact
same kernel as him, my system locks up. Both systems are 1 GHz, 12" iBooks.

Does anyone have any ideas on why this might not work on my system? Is there
something else that needs to be looked at when increasing the frequency after
booting?

cheers,

John

diff -ru /usr/src/linux-2.6.6-bak/arch/ppc/kernel/misc.S ../kernel/misc.S
- --- /usr/src/linux-2.6.6-bak/arch/ppc/kernel/misc.S	2004-05-13
22:59:00.000000000 +0930
+++ ../kernel/misc.S	2004-05-28 12:47:22.000000000 +0930
@@ -253,6 +253,24 @@
 	mtmsr	r7
 	blr

+_GLOBAL(choose_7447a_dfs)
+	/* Clear MSR:EE */
+	mfmsr	r7
+	rlwinm	r0,r7,0,17,15
+	mtmsr	r0
+
+	/* Calc new HID1 value */
+	mfspr	r4,SPRN_HID1
+	insrwi	r4,r3,1,9	/* insert parameter into bit 9 */
+	sync
+	mtspr	SPRN_HID1,r4
+	sync
+	isync
+
+	/* Return */
+	mtmsr	r7
+	blr
+
 #endif /* CONFIG_CPU_FREQ_PMAC && CONFIG_6xx */

 /* void local_save_flags_ptr(unsigned long *flags) */
diff
- -ru /usr/src/linux-2.6.6-bak/arch/ppc/platforms/pmac_cpufreq.c ../platforms/pmac_cpufreq.c
- --- /usr/src/linux-2.6.6-bak/arch/ppc/platforms/pmac_cpufreq.c	2004-05-13
22:59:00.000000000 +0930
+++ ../platforms/pmac_cpufreq.c	2004-05-30 22:56:07.369451720 +0930
@@ -47,6 +47,7 @@
 #warning "WARNING, CPUFREQ not recommended on SMP kernels"
 #endif

+extern void choose_7447a_dfs(int dfs);
 extern void low_choose_750fx_pll(int pll);
 extern void low_sleep_handler(void);
 extern void openpic_suspend(struct sys_device *sysdev, u32 state);
@@ -61,6 +62,7 @@
 /* Clean that up some day ... use a func ptr or at least an enum... */
 static int cpufreq_uses_pmu;
 static int cpufreq_uses_gpios;
+static int cpufreq_uses_dfs;

 static u32 voltage_gpio;
 static u32 frequency_gpio;
@@ -273,6 +275,10 @@
 		rc = pmu_set_cpu_speed(speed_mode);
 	else if (cpufreq_uses_gpios)
 		rc = gpios_set_cpu_speed(speed_mode);
+	else if (cpufreq_uses_dfs){
+		choose_7447a_dfs(speed_mode);
+		rc = 0;
+	}
 	else
 		rc = cpu_750fx_cpu_speed(speed_mode);
 	cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
@@ -346,6 +352,7 @@
  *  - iBook2 500 (PMU based, 400Mhz & 500Mhz)
  *  - iBook2 700 (CPU based, 400Mhz & 700Mhz, support low voltage)
  *  - Recent MacRISC3 machines
+ *  - iBook G4s with 7447A CPUs (experimental, unstable)
  */
 static int __init pmac_cpufreq_setup(void)
 {
@@ -371,6 +378,20 @@
 	if (machine_is_compatible("PowerBook3,4") ||
 	    machine_is_compatible("PowerBook3,5") ||
 	    machine_is_compatible("MacRISC3")) {
+
+		/*  Check for 7447A based iBook G4 */
+		if (machine_is_compatible("PowerBook6,5")) {
+			cur_freq /= 2;	/* OF reports wrong frequency */
+			hi_freq = cur_freq*2;
+			low_freq = cur_freq;
+			has_freq_ctl = 1;
+			cpufreq_uses_pmu = 0;
+			cpufreq_uses_gpios = 0;
+			cpufreq_uses_dfs = 1;
+			has_freq_ctl = 1;
+			goto out;
+		}
+
 		struct device_node *volt_gpio_np = of_find_node_by_name(NULL,
"voltage-gpio");
 		struct device_node *freq_gpio_np = of_find_node_by_name(NULL,
"frequency-gpio");
 		struct device_node *slew_done_gpio_np = of_find_node_by_name(NULL,
"slewing-done");
@@ -487,6 +508,7 @@
 		cpufreq_uses_pmu = 0;
 		has_freq_ctl = 1;
 	}
+
 out:
 	if (!has_freq_ctl)
 		return -ENODEV;
@@ -497,7 +519,7 @@
 	printk(KERN_INFO "Registering PowerMac CPU frequency driver\n");
 	printk(KERN_INFO "Low: %d Mhz, High: %d Mhz, Boot: %d Mhz, switch method:
%s\n",
 	       low_freq/1000, hi_freq/1000, cur_freq/1000,
- -	       cpufreq_uses_pmu ? "PMU" : (cpufreq_uses_gpios ? "GPIOs" : "CPU"));
+	       cpufreq_uses_pmu ? "PMU" : (cpufreq_uses_gpios ? "GPIOs" :
(cpufreq_uses_dfs ? "DFS" : "CPU")));

 	return cpufreq_register_driver(&pmac_cpufreq_driver);
 }


** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/

Re: nearly-working support for cpufreq on 2004 iBook G4 with 7447A cpu

[Benjamin Herrenschmidt]

On Mon, 2004-05-31 at 00:01, John Steele Scott wrote:
> Hi all,
>
> I have written a patch against kernel 2.6.6 which lets the pmac cpufreq driver
> support dynamic frequency scaling on the 7447A cpu in my new iBook. It also
> fixes the incorrect clock speed reporting (i.e. on boot 2.6.6 reports cpu is
> running at 1066 MHz when really it is 533 MHz).

OF is probably not reporting wrong frrequency, check your bogomips. I
suspect your clock chip isn't configured properly at boot, like with
other newer models. I need to hack the proper bits in. I'll have a look
at this asap.

> Unfortunately, if I use this to set my cpu frequency up to 1 GHz, my system
> locks up hard when I try and do something cpu-intensive. Christiaan Welvaart
> has tested the previous version of this patch (functionally the same, but a
> bit messier), and on his system it works fine. Yet even if I use the exact
> same kernel as him, my system locks up. Both systems are 1 GHz, 12" iBooks.
>
> Does anyone have any ideas on why this might not work on my system? Is there
> something else that needs to be looked at when increasing the frequency after
> booting?

You might need the proper clock chip tweaks.

Ben.


** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/

Re: nearly-working support for cpufreq on 2004 iBook G4 with 7447A cpu

[Benjamin Herrenschmidt]

On Mon, 2004-05-31 at 08:43, Benjamin Herrenschmidt wrote:

> > Does anyone have any ideas on why this might not work on my system? Is there
> > something else that needs to be looked at when increasing the frequency after
> > booting?
>
> You might need the proper clock chip tweaks.

Oh, and you may have the voltage wrong too. I'll have a look

Ben.


** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/

Re: nearly-working support for cpufreq on 2004 iBook G4 with 7447A cpu

[Benjamin Herrenschmidt]

>
> Is there any datasheet or similar which describes this chip?


No, I need to manually decode some of the platform-do-* properties in
the device-tree that initialize things properly and implement the
necessary tweaks. That or finally write an interpreter for those ;)

Ben.


** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/

Re: nearly-working support for cpufreq on 2004 iBook G4 with 7447A cpu

[John Steele Scott]

On Mon, 31 May 2004 08:13 am, Benjamin Herrenschmidt wrote:
> OF is probably not reporting wrong frrequency, check your bogomips. I
> suspect your clock chip isn't configured properly at boot, like with
> other newer models. I need to hack the proper bits in. I'll have a look
> at this asap.

Here's a few bits and pieces taken from an unpatched kernel:

bash-2.05b$ cat /proc/cpuinfo
processor       : 0
cpu             : 7447A, altivec supported
clock           : 1066MHz
revision        : 1.1 (pvr 8003 0101)
bogomips        : 530.43
machine         : PowerBook6,5
motherboard     : PowerBook6,5 MacRISC3 Power Macintosh
detected as     : 287 (iBook G4)
pmac flags      : 0000000a
L2 cache        : 512K unified
memory          : 384MB
pmac-generation : NewWorld

bash-2.05b$ ls /proc/device-tree/cpus/PowerPC,G4/
altivec               device_type          PowerPC,G4
available             dynamic-power-step   processor-to-bus-ratio*2
bus-frequency         existing             recalced-clock-frequency
clock-frequency       graphics             reg
config-bus-frequency  i-cache-block-size   reservation-granule-size
cpu-info              i-cache-sets         rounded-clock-frequency
cpu-version           i-cache-size         state
data-streams          l2-cache             timebase-frequency
d-cache-block-size    l2cr                 tlb-sets
d-cache-sets          linux,phandle        tlb-size
d-cache-size          name                 translations
dcba                  needs-delayAACK
dcbz                  performance-monito

(Notice there is no min/max-clock-frequency.)

bash-2.05b$ cat /proc/device-tree/cpus/PowerPC,G4/clock-frequency | xxd
0000000: 3f94 0aa8                                ?...
bash-2.05b$ python -c 'print "%d" % 0x3f940aa8'
1066666664

> You might need the proper clock chip tweaks.

Is there any datasheet or similar which describes this chip?

thanks,

John

** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/

Re: nearly-working support for cpufreq on 2004 iBook G4 with 7447A cpu

[Benjamin Herrenschmidt]

On Mon, 2004-05-31 at 10:32, John Steele Scott wrote:
>
> On Mon, 31 May 2004 08:13 am, Benjamin Herrenschmidt wrote:
> > OF is probably not reporting wrong frrequency, check your bogomips. I
> > suspect your clock chip isn't configured properly at boot, like with
> > other newer models. I need to hack the proper bits in. I'll have a look
> > at this asap.

Ok, that's weird, I don't see the same kind of clock chip on i2c as
the previous model had... Can you check the exact dividers programmed
on the CPU and redo the calculation based on the bus frequency ?

There is a do-clock-spreading property attached to the U2 ASIC, I don't
know if it may be related, I'll have a look.

Ben.


** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/

Re: nearly-working support for cpufreq on 2004 iBook G4 with 7447A cpu

[John Steele Scott]

On Mon, 31 May 2004 01:19 pm, Benjamin Herrenschmidt wrote:
> Ok, that's weird, I don't see the same kind of clock chip on i2c as
> the previous model had... Can you check the exact dividers programmed
> on the CPU and redo the calculation based on the bus frequency ?

Okay.

At boot: HID1=80414c80 => PLL_CFG=10100, i.e. 4:1 core multiplier

After clearing DFS, HID1=80018c80 => PLL_CFG=11000, i.e. 8:1 core multiplier

I decoded the PLL values using the table on page 31 of Moto's MPC7447AEC.pdf.

bash-2.05b$ cat /proc/device-tree/clock-frequency  | xxd
0000000: 07ef 4679                                ..Fy
bash-2.05b$ python -c 'print "%d" % 0x07ef4679'
133121657

Looks sensible to me. Does it give you any clues?

cheers,

John

** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/

Re: nearly-working support for cpufreq on 2004 iBook G4 with 7447A cpu

[Benjamin Herrenschmidt]

On Mon, 2004-05-31 at 18:37, John Steele Scott wrote:
>
> On Mon, 31 May 2004 01:19 pm, Benjamin Herrenschmidt wrote:
> > Ok, that's weird, I don't see the same kind of clock chip on i2c as
> > the previous model had... Can you check the exact dividers programmed
> > on the CPU and redo the calculation based on the bus frequency ?
>
> Okay.
>
> At boot: HID1=80414c80 => PLL_CFG=10100, i.e. 4:1 core multiplier
>
> After clearing DFS, HID1=80018c80 => PLL_CFG=11000, i.e. 8:1 core multiplier
>
> I decoded the PLL values using the table on page 31 of Moto's MPC7447AEC.pdf.
>
> bash-2.05b$ cat /proc/device-tree/clock-frequency  | xxd
> 0000000: 07ef 4679                                ..Fy
> bash-2.05b$ python -c 'print "%d" % 0x07ef4679'
> 133121657
>
> Looks sensible to me. Does it give you any clues?

Not sure at this point, play with DFS during boot and see what bogomips
you get maybe...


** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/

Re: nearly-working support for cpufreq on 2004 iBook G4 with 7447A cpu

[Benjamin Herrenschmidt]

Oh, and when switching to high speed, think about tweaking the
voltage too. I don't know if it's high or low when setting the
GPIO bit, you may want to check out OF code, it certainly have
methods for that. Or you can read the voltage value back from
the adt themometer chip I think, as it does voltage control.

(see cpu-vcore-select GPIO)

I suspect that the machine boots at low speed & low voltage. You
should either set it to high speed in pmac_feature (before bogomips
calculation) or assume it has booted low speed in your driver. You
need to ramp up voltage & wait before swiotching to high speed and
ramp down voltage after switching to low speed.

The delay is unknown, I suppose 100ms should be plenty enough though
(probably even too much).

Ben.


** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/

Re: nearly-working support for cpufreq on 2004 iBook G4 with 7447Acpu

[John Steele Scott]

> Oh, and when switching to high speed, think about tweaking the
> voltage too. I don't know if it's high or low when setting the
> GPIO bit, you may want to check out OF code, it certainly have
> methods for that. Or you can read the voltage value back from
> the adt themometer chip I think, as it does voltage control.

I will double-check this tonight (am at uni now), but I thought that the
core voltage for this chip does not change with frequency . . . :(

** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/

Re: nearly-working support for cpufreq on 2004 iBook G4 with 7447Acpu

[Benjamin Herrenschmidt]

On Tue, 2004-06-01 at 12:05, John Steele Scott wrote:
> > Oh, and when switching to high speed, think about tweaking the
> > voltage too. I don't know if it's high or low when setting the
> > GPIO bit, you may want to check out OF code, it certainly have
> > methods for that. Or you can read the voltage value back from
> > the adt themometer chip I think, as it does voltage control.
>
> I will double-check this tonight (am at uni now), but I thought that the
> core voltage for this chip does not change with frequency . . . :(

It probably does since there is a GPIO for controlling it on
KeyLargo. It's not the chip itself that changes it, you use the GPIO
to change it before/after the freq. switch. That's wher the real saving
comes from, more than the freq change itself in fact.

ben.


** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/

Re: nearly-working support for cpufreq on 2004 iBook G4 with 7447A cpu

[John Steele Scott]

On Tue, 1 Jun 2004 09:17 am, Benjamin Herrenschmidt wrote:
> I suspect that the machine boots at low speed & low voltage. You
> should either set it to high speed in pmac_feature (before bogomips
> calculation) or assume it has booted low speed in your driver. You
> need to ramp up voltage & wait before swiotching to high speed and
> ramp down voltage after switching to low speed.

Okay, I think I've made some progress on this. In the openfirmware device for
the PowerPC,G4 chip, there are methods called set-dfs-low and set-dfs-high.
The first one sets the DFS bit and the seconds one clears it, so the low and
high are referring to low and high CPU speed.

It looks like both methods have a simple check of the PVR, then the main code
for set-dfs-low looks like:
hid1@ 1 1f 9 - lshift or hid1! 1 ms 4 1 gpio

for set-dfs-high it looks like:
5 1 gpio! 1 ms hid1@ 1 1f 9 - lshift andc hid1!

(note that I could have mistaken the numeral '1' for the letter 'l')

So if stop at open firmware, execute set-dfs-high and then boot with a stock
2.6.6 kernel, my cpuinfo now shows a bogomips of 1060.86, and the system
seems to be stable.

I don't know much about manipulating the GPIO stuff, but I guess I will learn
it sometime in the next few days and hopefully get to a fully working patch.

cheers,

John

** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/

Re: nearly-working support for cpufreq on 2004 iBook G4 with 7447A cpu

[Benjamin Herrenschmidt]

On Wed, 2004-06-02 at 21:08, John Steele Scott wrote:

> I don't know much about manipulating the GPIO stuff, but I guess I will learn
> it sometime in the next few days and hopefully get to a fully working patch.

GPIO just means "general purpose IO", it's just IO bits hanging out of
the KeyLargo chipsets and connected to various things on the motherboard,
in that case, one is connected to the power supply controller to control
the power to the CPU.

Look at the cpufreq code for gpios_set_cpuspeed(), it's used on models
were both the clock frequency and the voltage are controller by GPIOs,
keep the voltage part. Remember to ramp up the voltage before setting
the frequency to high, and ramp it down after setting the frequency to
low.

Ben.


** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/